# THE ROLE OF NON-STOICHIOMETRY IN THE FUNCTIONAL PROPERTIES OF OXIDE MATERIALS

EDITED BY : Maria Veronica Ganduglia-Pirovano, Javier Carrasco and Claudio Cazorla PUBLISHED IN : Frontiers in Chemistry and Frontiers in Physics

### Frontiers eBook Copyright Statement

The copyright in the text of individual articles in this eBook is the property of their respective authors or their respective institutions or funders. The copyright in graphics and images within each article may be subject to copyright of other parties. In both cases this is subject to a license granted to Frontiers. The compilation of articles constituting this eBook is the property of Frontiers.

Each article within this eBook, and the eBook itself, are published under the most recent version of the Creative Commons CC-BY licence. The version current at the date of publication of this eBook is CC-BY 4.0. If the CC-BY licence is updated, the licence granted by Frontiers is automatically updated to the new version.

When exercising any right under the CC-BY licence, Frontiers must be attributed as the original publisher of the article or eBook, as applicable.

Authors have the responsibility of ensuring that any graphics or other materials which are the property of others may be included in the CC-BY licence, but this should be checked before relying on the CC-BY licence to reproduce those materials. Any copyright notices relating to those materials must be complied with.

Copyright and source acknowledgement notices may not be removed and must be displayed in any copy, derivative work or partial copy which includes the elements in question.

All copyright, and all rights therein, are protected by national and international copyright laws. The above represents a summary only. For further information please read Frontiers' Conditions for Website Use and Copyright Statement, and the applicable CC-BY licence.

ISSN 1664-8714 ISBN 978-2-88963-155-1 DOI 10.3389/978-2-88963-155-1

### About Frontiers

Frontiers is more than just an open-access publisher of scholarly articles: it is a pioneering approach to the world of academia, radically improving the way scholarly research is managed. The grand vision of Frontiers is a world where all people have an equal opportunity to seek, share and generate knowledge. Frontiers provides immediate and permanent online open access to all its publications, but this alone is not enough to realize our grand goals.

### Frontiers Journal Series

The Frontiers Journal Series is a multi-tier and interdisciplinary set of open-access, online journals, promising a paradigm shift from the current review, selection and dissemination processes in academic publishing. All Frontiers journals are driven by researchers for researchers; therefore, they constitute a service to the scholarly community. At the same time, the Frontiers Journal Series operates on a revolutionary invention, the tiered publishing system, initially addressing specific communities of scholars, and gradually climbing up to broader public understanding, thus serving the interests of the lay society, too.

### Dedication to Quality

Each Frontiers article is a landmark of the highest quality, thanks to genuinely collaborative interactions between authors and review editors, who include some of the world's best academicians. Research must be certified by peers before entering a stream of knowledge that may eventually reach the public - and shape society; therefore, Frontiers only applies the most rigorous and unbiased reviews.

Frontiers revolutionizes research publishing by freely delivering the most outstanding research, evaluated with no bias from both the academic and social point of view. By applying the most advanced information technologies, Frontiers is catapulting scholarly publishing into a new generation.

### What are Frontiers Research Topics?

Frontiers Research Topics are very popular trademarks of the Frontiers Journals Series: they are collections of at least ten articles, all centered on a particular subject. With their unique mix of varied contributions from Original Research to Review Articles, Frontiers Research Topics unify the most influential researchers, the latest key findings and historical advances in a hot research area! Find out more on how to host your own Frontiers Research Topic or contribute to one as an author by contacting the Frontiers Editorial Office: researchtopics@frontiersin.org

# THE ROLE OF NON-STOICHIOMETRY IN THE FUNCTIONAL PROPERTIES OF OXIDE MATERIALS

Topic Editors: Maria Veronica Ganduglia-Pirovano, Institute of Catalysis and Petrochemistry (ICP-CSIC), Spain Javier Carrasco, CIC Energigune, Spain Claudio Cazorla, UNSW Sydney, Australia

Citation: Ganduglia-Pirovano, M. V., Carrasco, J., Cazorla, C., eds. (2019). The Role of Non-Stoichiometry in the Functional Properties of Oxide Materials. Lausanne: Frontiers Media SA. doi: 10.3389/978-2-88963-155-1

# Table of Contents

*05 Editorial: The Role of Non-stoichiometry in the Functional Properties of Oxide Materials*

Claudio Cazorla, Maria Verónica Ganduglia-Pirovano and Javier Carrasco


Zhong-Kang Han, Lei Zhang, Meilin Liu, Maria Verónica Ganduglia-Pirovano and Yi Gao


Byung-Hyun Kim, Jolla Kullgren, Matthew J. Wolf, Kersti Hermansson and Peter Broqvist


Emiliano Poli, Joshua D. Elliott, Sergey K. Chulkov, Matthew B. Watkins and Gilberto Teobaldi


Tatyana Y. Kardash, Elizaveta A. Derevyannikova, Elena M. Slavinskaya, Andrey I. Stadnichenko, Vasiliy A. Maltsev, Alexey V. Zaikovskii, Sergey A. Novopashin, Andrei I. Boronin and Konstantin M. Neyman

*127 Activation of Water on MnOx -Nanocluster-Modified Rutile (110) and Anatase (101) TiO2 and the Role of Cation Reduction* Stephen Rhatigan and Michael Nolan

# *139 Dynamics of the Interaction Between Ceria and Platinum During Redox Processes*

Paola Luches, Gabriele Gasperi, Marc Sauerbrey, Sergio Valeri, Jens Falta and Jan Ingo Flege

# *150 Oxygen Vacancies in Oxide Nanoclusters: When Silica is More Reducible Than Titania*

Andi Cuko, Stefan T. Bromley and Monica Calatayud

# Editorial: The Role of Non-stoichiometry in the Functional Properties of Oxide Materials

Claudio Cazorla<sup>1</sup> \*, Maria Verónica Ganduglia-Pirovano<sup>2</sup> and Javier Carrasco<sup>3</sup>

*<sup>1</sup> School of Materials Science and Engineering, University of New South Wales, Sydney, NSW, Australia, <sup>2</sup> Institute of Catalysis and Petrochemistry-CSIC, Madrid, Spain, <sup>3</sup> CIC Energigune, Vitoria-Gasteiz, Spain*

Keywords: oxide materials, density functional theory, experiments, oxygen vacancies, functionality

### **Editorial on the Research Topic**

### **The Role of Non-stoichiometry in the Functional Properties of Oxide Materials**

Oxides can show a plethora of emergent complex phenomena beyond their traditional role as dielectrics, which makes them attractive materials to a very broad range of scientific and technological processes. This includes areas such as catalysis, energy conversion and storage, and magnetism, to name just a few. Furthermore, nano-engineered oxides offer the possibility of realizing hitherto unobserved phenomena as due to their unique responsiveness to external stimuli, multiple order-parameter couplings, and increased volume-to-surface ratio. Yet, controlled synthesis and thorough microscopic understanding of functional oxide materials remain challenging. This is a complex task to accomplish due to the ubiquitous stoichiometry fluctuations found in oxides, which affect considerably their structural, chemical, and magnetic properties (Venkatesan et al., 2004; Herklotz et al., 2017).

Non-stoichiometry is indeed a pivotal facet of the functioning of oxide materials. For instance, the presence of oxygen vacancies in oxide-based heterogeneous catalysts is often necessary for any activity (Campbell and Peden, 2005; Jeen et al., 2013). Oxygen vacancies can also distort significantly the equilibrium arrangement of atoms and modify the super-exchange interactions between neighboring magnetic ions, thus inducing spin-order transformations (Wang et al., 2003; Eerenstein et al., 2005). Furthermore, extrinsic vacancies enable ionic conductivity in perovskite-based solid solutions to be used in electrochemical applications such as solid oxide fuel and electrolytic cells (Adler, 2004; Chroneos et al., 2011). Engineering of chemical defects in oxide compounds, therefore, emerges as a likely avenue for the design of new materials with tailored functionality. In this Research Topic, we present a collection of original articles that highlight how state-of-the-art theoretical and experimental approaches are leading to a better atomic-level understanding of non-stoichiometric oxide systems and improved design of novel functional materials.

From a theoretical viewpoint, the accurate description of the electronic structure of nonstoichiometric oxides of technological interest is not an easy task and requires efficient and robust theory (Ganduglia-Pirovano et al., 2007). Martínez-Casado et al. thoroughly analyze this issue for defective anatase surfaces and conclude that all-electron-based hybrid density functionals are required to yield accurate trends in the position of band gap states of defects and adsorbates in large model systems. Furthermore, Poli et al. efficiently combine non-local density functional theory (DFT) and linear-response time-dependent DFT to enable the interpretation of recent experimental

Edited and reviewed by: *Bretislav Friedrich, Fritz-Haber-Institut, Germany*

> \*Correspondence: *Claudio Cazorla c.cazorla@unsw.edu.au*

### Specialty section:

*This article was submitted to Physical Chemistry and Chemical Physics, a section of the journal Frontiers in Chemistry*

> Received: *17 June 2019* Accepted: *18 July 2019* Published: *08 August 2019*

### Citation:

*Cazorla C, Ganduglia-Pirovano MV and Carrasco J (2019) Editorial: The Role of Non-stoichiometry in the Functional Properties of Oxide Materials. Front. Chem. 7:547. doi: 10.3389/fchem.2019.00547*

**5**

UV-vis spectra for cation-vacancy defects in aluminosilicate imogolite nanotubes, considering large supercells of around 1,000 atoms.

Research on non-stoichiometric oxide materials is typically focused on bulk and extended surfaces. Generally, the interaction of oxygen vacancies in such materials is poorly understood. Gao et al. address the mechanisms of aggregation of oxygen vacancies in SiO<sup>2</sup> and HfO<sup>2</sup> theoretically and suggest that electron trapping at oxygen vacancies may facilitate the creation of new neighboring vacancies and that this process becomes even more efficient with increasing vacancy cluster size. This finding is particularly relevant for improving the performance of such materials in applications where, for example, electrons can be injected via tunneling from an electrode in a device. At non-stoichiometric oxide surfaces, the structure and dynamics of the near-surface oxygen vacancies and excess charges are most relevant for technological applications such as catalysis. Wolf et al. address the heavily discussed appearance and properties of anionic defects at the (111)-oriented CeO<sup>2</sup> ceria surface in scanning tunneling microscopy (STM) images, and Han et al. investigate the effect of applied stress on the relative stability of isolated surface and subsurface oxygen vacancies at the same surface, the formation of vacancy pairs, and the localization of the excess charge. Noticeably, at moderate compressive strains the formation of first-neighbor surface oxygen vacancy pairs is energetically preferred over that of isolated vacancies, contrary to what it is predicted at null strain. Yet, little is known about nanosize effects on oxygen vacancy formation. This salient aspect of current materials design is theoretically explored by Cuko et al. considering as a case study titania, silica, and mixed TixSi1−xO<sup>2</sup> nanoclusters. Interestingly, the authors show that, in contrast to the results obtained for the bulk and for surfaces, silica nanoclusters are more reducible than titania, thus disclosing an unexpected chemical behavior at the nanoscale. Moreover, Rhatigan and Nolan model hydroxylated titania surfaces modified with dispersed Mn4O<sup>x</sup> nanoclusters and reveal important changes in light absorption and reactivity toward water splitting as a consequence of the formation of nanocluster-surface interactions that influence the system electronic structure. Furthermore, Kim et al. consider the effect that agglomeration of CeO<sup>2</sup> nanoparticles have on the stability and oxygen vacancy formation energy, applying a computational protocol based on a multiscale simulation approach. Surprisingly, for a possible {100}/{100} interface of two agglomerated Ce40O<sup>80</sup> truncated octahedral, the energy cost

# REFERENCES


to create an oxygen vacancy is negative, indicating that under reducing conditions, oxygen removal from the interface will have a stabilizing effect.

The structural changes of oxide surfaces during redox processes are essential to understanding their properties in many applications. Schöttner et al. track the evolution of the hematite a-Fe2O3(0001) surface under reducing conditions by polarization-resolved infrared reflection absorption spectroscopy (IRRAS) using CO as a probe molecule. Not only the formation of oxygen vacancies is observed, but a massive reconstruction yielding the formation of Fe3O4(111) domains. Such findings cannot be ignored when thinking about the surface chemistry of iron oxide surfaces. Furthermore, Luches et al. focus on the dynamics of the interaction between ceria epitaxial islands and films on Pt (111) during reduction and oxidation, and observe different reduced structural phases that include the formation of a Ce-Pt alloy. These results are most relevant for the understanding of Pt/CeO<sup>2</sup> catalysts, which can be used for low-temperature CO oxidation, as discussed by Kardash et al., who investigate Pt/CeO<sup>x</sup> and Pt/CeSnO<sup>x</sup> systems by means of a synergistic approach that combines experiment and theory. It is found that the Sn-doping promotes oxygen removal and increases catalytic activity. The effect of catalyst doping is also considered by Alvarez-Galvan et al., who address the production of synthesis gas from methane over Rh-doped Ni-based catalysts, supported on Al2O3, CeO2, La2O3, MgO, and ZrO2. This is an important process for converting natural gas into added value high-quality liquid products. However, current catalysis methods for achieving partial oxidation of methane are energy-intensive and hence more energy-efficient alternatives to produce synthesis gas are highly desirable. It is found that Ni/CeO<sup>2</sup> catalyst shows the best behavior per surface area and that small amount of Rh improves the catalytic performance.

We believe that the Role of Non-stoichiometry in the Functional Properties of Oxide Materials Research Topic provides recent insights into our ongoing quest for rationalizing the intriguing phenomena observed in non-stoichiometric oxide systems and how this can affect their functionality, with special focus on energy harvesting applications.

# AUTHOR CONTRIBUTIONS

All authors listed have made a substantial, direct and intellectual contribution to the work, and approved it for publication.


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Cazorla, Ganduglia-Pirovano and Carrasco. This is an openaccess article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Structural Evolution of α-Fe2O3(0001) Surfaces Under Reduction Conditions Monitored by Infrared Spectroscopy

Ludger Schöttner, Alexei Nefedov, Chengwu Yang, Stefan Heissler, Yuemin Wang\* and Christof Wöll\*

Institute of Functional Interfaces (IFG), Karlsruhe Institute of Technology, Karlsruhe, Germany

### Edited by:

Maria Veronica Ganduglia-Pirovano, Institute of Catalysis and Petrochemistry (ICP), Spain

### Reviewed by:

Shamil Shaikhutdinov, Fritz-Haber-Institut, Germany Oscar Rodríguez De La Fuente, Complutense University of Madrid, Spain Geoff Thornton, University College London, United Kingdom

### \*Correspondence:

Yuemin Wang yuemin.wang@kit.edu Christof Wöll christof.woell@kit.edu

### Specialty section:

This article was submitted to Physical Chemistry and Chemical Physics, a section of the journal Frontiers in Chemistry

> Received: 01 March 2019 Accepted: 05 June 2019 Published: 25 June 2019

### Citation:

Schöttner L, Nefedov A, Yang C, Heissler S, Wang Y and Wöll C (2019) Structural Evolution of α-Fe2O3(0001) Surfaces Under Reduction Conditions Monitored by Infrared Spectroscopy. Front. Chem. 7:451. doi: 10.3389/fchem.2019.00451 The precise determination of the surface structure of iron oxides (hematite and magnetite) is a vital prerequisite to understand their unique chemical and physical properties under different conditions. Here, the atomic structure evolution of the hematite (0001) surface under reducing conditions was tracked by polarization-resolved infrared reflection absorption spectroscopy (IRRAS) using carbon monoxide (CO) as a probe molecule. The frequency and intensity of the CO stretch vibration is extremely sensitive to the valence state and electronic environments of surface iron cations. Our comprehensive IRRAS results provide direct evidence that the monocrystalline, stoichiometric α-Fe2O3(0001) surface is single Fe-terminated. The initial reduction induced by annealing at elevated temperatures produces surface oxygen vacancies, where the excess electrons are localized at adjacent subsurface iron ions (5-fold coordinated). A massive surface restructuring occurs upon further reduction by exposing to atomic hydrogen followed by Ar+-sputtering and annealing under oxygen poor conditions. The restructured surface is identified as a Fe3O4(111)/Fe1−xO(111)-biphase exposing both, Fe3<sup>+</sup> and Fe2<sup>+</sup> surface species. Here the well-defined surface domains of Fe3O4(111) exhibit a Feoct2 termination, while the reduced Fe1−xO(111) is Fe2+(oct)-terminated. These findings are supported by reference IRRAS data acquired for CO adsorption on magnetite (111) and (001) monocrystalline surfaces.

Keywords: iron oxide, CO adsorbed infrared spectroscopy, IR spectroscopy, reduced oxides, XPS

# INTRODUCTION

Although iron oxides have been the topic of numerous studies, their surface is of complex nature and their structure depends strongly on the interaction with the ambience (Kuhlenbeck et al., 2013; Parkinson, 2016). Not only because of the comprehensive phase diagram and the ability to form polymorphs, but also because of the unique material properties including magnetism, the surface properties of iron oxide are among the most challenging research fields in inorganic chemistry and solid-state physics. In previous works, a number of structural conversions and surface reconstructions have been reported (Ketteler et al., 2001; Weiss and Ranke, 2002; Sakurai et al., 2009; Kuhlenbeck et al., 2013; Parkinson, 2016). It is therefore not surprising that the discussion on many common iron oxide facets toward surface stabilization and termination has remained controversial. For detailed insights, we refer the reader to two recent excellent review papers (Kuhlenbeck et al., 2013; Parkinson, 2016). In section Preview on Related Iron Oxide Crystal Structures, we have briefly summarized the important experimental and theoretical results reported in the literature for three relevant iron oxide surfaces [α-Fe2O3(0001), Fe3O4(111), Fe3O4(001)] in the form of both thin films and single crystals.

The chemical and physical properties of metal oxides strongly depend on the structural arrangement of atoms at their surfaces (Wang and Wöll, 2017). Furthermore, it has been shown that the chemical reactivity can be dramatically changed in the presence of defects like oxygen vacancies. Especially under reducing conditions there are numerous such defects, with their densities depending on temperature and the surrounding atmosphere (Parkinson, 2016). However, it is difficult to gain atomic-level information on metal oxide surface structures since some of the standard methods in surface science cannot be applied in a straightforward fashion, e.g., due to charging problems.

This problem can be avoided when using non-destructive surface analysis methods insensitive to charging effects. An important example is infrared (IR) spectroscopy. In combination with appropriate probe molecules like carbon monoxide (CO), this vibrational spectroscopy has been shown to provide insights into structural, electronic and reactive properties of model and nanostructured catalysts (Yang and Wöll, 2017). In addition, the interaction between CO and iron oxide (α-Fe2O3, hematite, Fe3O4, magnetite, Fe1−xO, wüstite) is of great interest in itself because of detailed knowledge about the binding of CO to iron oxides is required for a fundamental understanding of industrial catalytic processes such as CO oxidation and water gas-shift reactions (de Smit and Weckhuysen, 2008; Freund et al., 2011; Zhu and Wachs, 2016; Wei et al., 2018). The adsorption of CO on various well-defined metal surfaces has been extensively studied by infrared reflection absorption spectroscopy (IRRAS). However, the application of this approach on metal oxide singlecrystal surfaces is challenging due to the inherent experimental difficulties arising from the low reflectivity of dielectric substrates in the IR regime. Recent progress in instrumentation, however, has allowed us to obtain high-resolution data for a number of different oxide surfaces (e.g., CeO2, TiO2, ZnO, α-Fe2O3) (Buchholz et al., 2016; Yu et al., 2016; Yang et al., 2017a,b; Chen et al., 2019; Schöttner et al., 2019). These IRRAS results have demonstrated that CO is well suited as a probe molecule to monitor the atomic structure evolution of oxide surfaces under different conditions. In particular, defects like oxygen vacancies and surface restructuring could be clearly identified (Xu et al., 2012; Yang et al., 2014, 2017a,b).

To the best of our knowledge, no IRRAS data of CO adsorption on surfaces of macroscopic iron oxide single crystals have been reported so far. Several studies have been reported on related systems including CO adsorbed on α-Fe2O<sup>3</sup> powders (Zecchina et al., 1988) and Fe3O4(111) thin films grown on Pt(111) (Lemire et al., 2004, 2005; Kuhlenbeck et al., 2013; Li et al., 2018). Here, we present a thorough polarization-resolved IRRAS study of CO adsorption on differently treated iron (III/II) oxide single-crystal surfaces, focusing on α-Fe2O3(0001) that has shown to be the most abundant facet exposed by hematite nanoparticles (Hartman, 1989). Our systematic IRRAS results allow to gain detailed insight into the structural evolution of the α-Fe2O3(0001) surface under reducing conditions. The assignment of the IR bands was assisted by additional reference measurements of CO adsorption on well-defined Fe3O4(111) and Fe3O4(001) monocrystalline substrates.

# MODELS AND METHODS

# Preview on Related Iron Oxide Crystal Structures

α-Fe2O<sup>3</sup> crystallizes in a corundum structure characterized by a slightly distorted hexagonal close-packed oxide-ion lattice where 2/3 of the octahedral sites are occupied by Fe3<sup>+</sup> ions with nondegenerated energy levels on metal d-electrons due to a ligand field splitting originating from Fe-O hybridization (**Figure S1**) (de Groot et al., 1989). Oxygen ions that are located parallel to (0001) plane levels, are separated by an iron double layer, yielding a stacking sequence of O3-Fe-Fe repeat units. There are three different surface terminations possible for α-Fe2O3(0001): a single iron layer (Fe-O3-Fe-R), a double iron layer (Fe-Fe-O3-R) and an oxygen (O3-Fe-Fe-R) layer (**Figure 1A**).

It is known that the structural and electronic properties of hematite (0001) depends strongly on the interaction with the ambience and the preparation parameters (Ketteler et al., 2001; Weiss and Ranke, 2002; Kuhlenbeck et al., 2013; Parkinson, 2016). In particular, the surface termination α-Fe2O3(0001) is highly sensitive to the oxygen partial pressure and preparation temperature. Freund and coworkers have systematically studied the surface structures of α-Fe2O3(0001) thin films supported on Pt(111) by scanning tunneling microscopy (STM) and IRRAS (Lemire et al., 2005). The results revealed a ferryl (Fe=O) termination coexisting with domains of the Fe-terminated surface in 10−<sup>3</sup> -1 mbar O<sup>2</sup> at ∼1,050 K, while the preparation at higher pressures of O<sup>2</sup> (>1 mbar) yields O-terminated surfaces (Shaikhutdinov and Weiss, 1999). In contrast, an exclusive Fetermination was reported on thin films (Thevuthasan et al., 1999) as well as on single crystal substrates under O-deficient conditions (<10−<sup>5</sup> mbar) (Yamamoto, 2010; Schöttner et al., 2019). For α-Fe2O3(0001) single crystals, it has been reported that Ar+-sputtering followed by annealing in 10−<sup>6</sup> mbar of O<sup>2</sup> at about 1,000 K leads to a massive surface restructuring, forming a Fe3O4(111) layer (Condon et al., 1994, 1998; Camillone et al., 2002). A biphase structure was observed upon heating in oxygen (10−<sup>6</sup> mbar) to higher temperatures (∼1,100 K) (Condon et al., 1995, 1998; Camillone et al., 2002). The STM results revealed the formation of long-range order of α-Fe2O3(0001) and Fe1−xO(111) domains, which are characterized by floretlike spots in the low-energy electron diffraction (LEED) pattern described as (<sup>√</sup> 3 × √ 3)R30◦ reconstruction (Kurtz and Henrich, 1983; Lad and Henrich, 1988; Condon et al., 1995). In addition, the surface chemistry and physics of α-Fe2O3(0001) has been studied by a number of theoretical groups [for details see the review by Parkinson, 2016] (Wang et al., 1998; Bergermayer et al., 2004; Trainor et al., 2004; Nguyen et al., 2014; Lewandowski et al., 2016; Ovcharenko et al., 2016).

In the crystal structure of Fe3O<sup>4</sup> oxygen anions are aligned in a cubic closed packing of equal spheres, where Fe3<sup>+</sup> cations are located at tetrahedrally coordinated interstitial sites. Other cations are located at octahedrally coordinated lattice positions, among which 50% exhibit the valence state +3 while the other 50% are in the +2 oxidation state. This chemical composition is typical for inverse spinel bulk structures and can be written as Fe3+[Fe3+Fe2+]O<sup>4</sup> to illustrate that the first type is tetrahedrally coordinated (A-type), whereas the parenthesized Fe2<sup>+</sup> and Fe3<sup>+</sup> ions occupy octahedrally coordinated sites (B-type). At low-pressure and low-temperature (CO measurement conditions), Fe3O<sup>4</sup> consists of a crystal structure with monoclinic symmetry (Wright et al., 2002; Bengtson et al., 2013). Above 120 K, known as the Verwey temperature (Verwey, 1939), the octahedrally coordinated iron atoms in magnetite have an average oxidation state of Fe +2.5 due to electron hopping between equal numbers of Fe2<sup>+</sup> and Fe3<sup>+</sup> ions. Such electron movement is frozen in Fe3O<sup>4</sup> below 120 K and therefore accompanied by a significant decrease in the electrical conductivity.

For Fe3O4(111) surfaces, six terminations may occur when cleaving the (111) stacking sequence (**Figure 1B**). The different terminations exhibit Fe atoms on octahedral (Feoct1 or Feoct2) and tetrahedral (Fetet1 or Fetet2) lattice sites, as well as O layers (O<sup>1</sup> or O2). The surface structure of Fe3O4(111) has been the subject of numerous experimental and theoretical investigations. Quantum chemical calculations have predicted that the Feoct2-termination is energetically favored on Fe3O4(111) under oxygen-deficient conditions, followed closely by the Fetet1 termination (Ahdjoudj et al., 1999; Zhu et al., 2006; Grillo et al., 2008; Noh et al., 2015). Recent work combining IRRAS and DFT (Li et al., 2018) revealed that the regular Fe3O4(111) thin film is Fetet1-terminated over close-packed oxygen layer, in line with the LEED IV results (Ritter and Weiss, 1999; Sala et al., 2012). STM studies provided further evidence for the Fetet1 termination on the pristine Fe3O4(111) surface, while the Feoct2 termination was observed under oxygen-poor conditions (Lennie et al., 1996; Shimizu et al., 2010). In addition, a long-range ordered biphase structure was reported for reduced Fe3O4(111) in the form of both thin films (Shaikhutdinov et al., 1999) and single-crystals (Condon et al., 1997; Paul et al., 2007). The STM and LEED data suggested the coexistence of Fe3O4(111) and Fe1−xO(111) domains.

From the ideal Fe3O4(001) model, only two surface terminations are possible, namely Fetet and FeoctO layers (**Figure 1C**). Again, the surface structure of Fe3O4(001) depends strongly on the preparation parameters (Tarrach et al., 1993; Parkinson et al., 2011; Bartelt et al., 2013). The polar Fe3O4(001) surface is well-known to be stabilized by a ( √ 2 × √ 2)R45◦ reconstruction. The reconstructed surface was first studied using STM by Wiesendanger et al. (1992) and later by other groups [see e.g., Chambers and Joyce, 1999; Rustad et al., 1999; Stanka et al., 2000; Mariotto et al., 2002]. Recently, Parkinson et al. (Bliem et al., 2014) proposed a subsurface cation vacancy (SCV) model by using STM in conjunction with LEED IV analysis and theory. In this model, the reconstructed surface features a FeoctO termination (B-layer) and is slightly distorted by a rearrangement of the cations in the subsurface layers.

Overall, due to the great complexity, an unambiguous experimental determination of the surface atomic structure for these iron oxides under different conditions is still a challenging task.

# Experimental Details

The IR measurements were carried out in an UHV multichamber system (Prevac). The base pressure of the entire chambers is in 10−<sup>10</sup> mbar range. The IR spectra were recorded in reflection absorption mode using a state-of-the-art Fourier-Transform (FT)-IR spectrometer (Bruker Vertex 80v) that is directly coupled to the UHV chamber via differentially pumped KBr-windows. The IR-spectrometer contains an internal polarizer module, which allows to carry out polarizationresolved IRRAS experiments. All CO adsorption measurements reported here were conducted for the respective iron oxide substrate mounted in a cryo-pumped IR-chamber at low temperatures. With liquid-He cooling, we could achieve sample temperatures of 65 K, which was sufficiently low to yield CO adsorption on all substrates studied here. Exposures to CO were carried out by using a leak-valve based directional doser. Gas dosages are quoted in Langmuir (1 L = 1.33·10−<sup>6</sup> mbar·s). The IR spectra were recorded after saturating the iron oxide surfaces with CO (1 L). The polarization-dependent IRRAS measurements were conducted with p- and/or spolarized light at grazing incidence, with the angle of 80◦ with respect to the surface normal. The IRRA spectra are shown in absorbance mode. The resolution of the IR spectrometer was set to 4 cm−<sup>1</sup> and 2,048 scans were accumulated for a single spectrum.

Iron oxide single crystals were mounted on a sample holder dedicated for e-beam heating. Temperatures were measured with a NiAl-NiCr thermocouple, that is spot-welded at a position between heating plate and backside of the sample. Surface quality and cleanliness were checked by XPS and LEED. Supplementary XPS measurements on the oxidation states of iron atoms for differently prepared iron oxide surfaces were conducted at the HESGM-beamline of the BESSY II storage ring, which is a part of the Helmholtz-Center Berlin.

The iron oxide single crystals used in this study were cut from minerals and polished (SurfaceNet GmbH). The α-Fe2O3(0001) substrate was cleaned by repeated cycles of annealing with stepwise heating from 850 K to 950 K in 10−<sup>5</sup> mbar O<sup>2</sup> (Supporting information, **Figures S2A–C**) (Lübbe and Moritz, 2009; Schöttner et al., 2019). The slightly reduced Fe2O3(0001) was formed after over-annealing the sample by gradual heating from 950 K to 1,050 K in 10−<sup>5</sup> mbar O2. A highly reduced surface was prepared by exposing the clean and stoichiometric α-Fe2O3(0001) to atomic hydrogen followed by post-annealing at 1,000 K. Then the reduced α-Fe2O3(0001) surface was prepared by repeated cycles of sputtering with 1.0 keV Ar<sup>+</sup> for 15 min and subsequent annealing at 950 K in 10−<sup>6</sup> mbar oxygen for 20 min until a LEED pattern gets visible.

The Fe3O4(111) single crystal surface was prepared by sputtering with 1.0 keV Ar<sup>+</sup> for 15 min and subsequent annealing at 920 K in 10−<sup>6</sup> mbar oxygen for 20 min. The surface of Fe3O4(001) was cleaned by Ar+-sputtering at 1.0 keV for 15 min and annealing at 870 K in a background of oxygen at 10−<sup>6</sup> mbar for 20 min (**Figure S3**) (Nie et al., 2013).

# RESULTS AND DISCUSSION

# CO Adsorption on the Pristine α-Fe2O3(0001) Surface

In **Figure 2** we display polarization-resolved IRRAS data recorded after a saturating exposure of the pristine stoichiometric α-Fe2O3(0001) single crystal surface to CO at 65 K. For ppolarized light, we observe a single, symmetric band at 2,169 cm−<sup>1</sup> which is characteristic for CO bound to the coordinatively unsaturated surface Fe3<sup>+</sup> cations (Zecchina et al., 1988), indicating the presence of a well-defined surface with only one adsorption site. As mentioned above, the surface termination of iron oxides depends strongly on the preparation conditions (Kuhlenbeck et al., 2013; Parkinson, 2016). For example, the ferryl termination was observed for α-Fe2O3(0001) thin films after oxidation with O<sup>2</sup> partial pressures of up to 1 mbar, showing a typical IR band at 2,185 cm−<sup>1</sup> (Lemire et al., 2005), which is not seen in the present IR spectra (**Figure 2**). Our IRRAS data provides direct spectroscopic evidence that the pristine <sup>α</sup>-Fe2O3(0001) surface is single Fe3+-terminated.

The difference between p- and s-polarizations can be understood by considering the transition dipole moment (TDM) describing the vibrational excitation of CO. The intensity and sign of an IR peak is given by the interaction of the TDM with the vector of the E-field. Note that for metal surfaces including metalsupported oxide thin films, IRRAS data can only be obtained for IR light polarized perpendicular to the surface according to the so-called surface selection rule. In the case of dielectric bulk materials (e.g., metal oxide single crystals), however, both s- and p-polarized components can interact with adsorbate vibrations being highly dependent on the orientation of TDM, thus allowing for a determination of the adsorption geometry (Wang and Wöll, 2017). The absorbance bands excited by E<sup>s</sup> are always negative, while the vibrations excited by p-polarized components (normal Ep,<sup>n</sup> and tangential Ep,<sup>t</sup> , showing always opposite signs) can be negative or positive depending on the incidence angle Θ and the refractive index n of the substrate (Brunner et al., 1997).

For the adsorption of a full CO monolayer on α-Fe2O3(0001) at 65 K, a single IR band with negative sign was observed at 2,169 cm−<sup>1</sup> for p-polarized light but not for s-polarized light (**Figure 2A**). Since the TDM is oriented parallel to the CO molecular axis, these polarization-dependent IRRAS results indicate that CO is bound to surface Fe3<sup>+</sup> ions in an upright geometry. Furthermore, we observed a coverage-dependent frequency shift (**Figure S4**), in line with IR results reported for CO adsorbed on α-Fe2O<sup>3</sup> powder samples (Zecchina et al., 1988). This coverage-induced frequency shift can be explained in terms of lateral adsorbate-adsorbate interactions including both dynamic and substrate-mediated static effects. For CO adsorption on oxide surfaces, the substrate mediated interactions (static shift) play a predominant role, whereas the contribution from dynamic interactions originating from the dipole-dipole coupling between CO adsorbates is rather small (Mahan and Lucas, 1978; Hollins and Pritchard, 1980).

The coverage-dependent frequency shift in CO stretch frequency was also observed after heating the sample to higher temperatures, i.e., with decreasing the CO coverage (temperature-induced CO desorption). The intensity of the CO vibrational band remains stable on α-Fe2O3(0001) up to 70 K (**Figure 2B**) but strongly decreases upon further heating. The IR band has disappeared at 100 K, indicating complete desorption of CO. The thermal desorption of CO is accompanied by a blue-shift in frequency from 2,169 to 2,178 cm−<sup>1</sup> . A quantitative analysis of the temperature-dependent IRRAS results (CO coverage as a function of temperature, see **Figure 2C**) (Gottfried et al., 2003) yields an activation energy for CO desorption on α-Fe2O3(0001) of 24 kJ·mol−<sup>1</sup> (using a frequency factor of 10<sup>13</sup> s −1 ), demonstrating that CO binds only weakly to the surface Fe3<sup>+</sup> ions.

# CO Adsorption on Reduced α-Fe2O3(0001)

In the next step the stoichiometric α-Fe2O3(0001) was subjected to annealing in 10−<sup>5</sup> mbar oxygen at higher temperatures increased stepwise from 850 K to 1050 K. The corresponding structure evolution was tracked by measuring the vibrational frequency of adsorbed CO. Upon annealing to 950 K, the IRRA spectrum shows only a single CO band at 2,169 cm−<sup>1</sup> (**Figure 3**), assigned to CO adsorbed at the 3-fold coordinated (Fe-O3) Fe3<sup>+</sup> sites of the stoichiometric α-Fe2O3(0001) surface, as discussed

above. A new CO band at 2,163 cm−<sup>1</sup> appears and becomes the dominating one when increasing the sample temperature to 950– 1000 K under O-poor conditions. After heating the sample to 1,000–1,050 K, only the 2,163 cm−<sup>1</sup> band is detected, while the Fe3+-related CO vibration at 2,169 cm−<sup>1</sup> observed for the fully oxidized surface disappears completely (see **Figure 3**). Note that in the fitting process, the half width for all CO peaks was set to be equal.

Overall, the present IRRAS data indicate the presence of surface oxygen vacancies (VO) created by annealing to elevated temperatures. Formation of V<sup>O</sup> defects have been found to induce a red-shift of CO frequencies on other oxides [e.g., rutile TiO2(110) surface; Xu et al., 2012]. Here, the slightly reduced α-Fe2O3(0001) surface (Fe-O3−x-Fe-R) is characterized by a CO band position of 2,163 cm−<sup>1</sup> , which we assign to CO bound to 2-fold coordinated (Fe-O2) Fe3<sup>+</sup> ions adjacent to O-vacancy sites. Importantly, the rather small red shift of 6 cm−<sup>1</sup> stands for an only little electronic structure modification of surface iron cations, suggesting that the excess electrons are not trapped at the 2-fold coordinated surface Fe3<sup>+</sup> ions. This finding is in excellent agreement with the DFT calculations made on the reduced α-Fe2O3(0001) surface (Ovcharenko et al., 2016). Along with the removal of a surface oxygen atom, two excess electrons were predicted to occupy the easily available Fe 3d states on two neighboring subsurface iron cations (5-fold coordinated).

We turn our attention to the experimental results for the highly reduced surface states of hematite, which were created by subjecting the stoichiometric Fe2O3(0001) to atomic hydrogen exposure followed by Ar+-sputtering and annealing cycles. Again, the changes in surface structure were monitored by measuring the CO vibrational frequency. Our IRRAS data allow to gain detailed insight into the surface structure of highly reduced α-Fe2O3(0001).

After CO adsorption at 65 K, the p-polarized IRRAS data show a number of CO bands in the regions between 2,175– 2,135 cm−<sup>1</sup> and 2,125–2,025 cm−<sup>1</sup> (**Figure 4A**), which are related respectively to CO species adsorbed on surface Fe3<sup>+</sup> and Fe2<sup>+</sup> sites, revealing substantial changes of the surface structure. To achieve a thorough analysis of the data, we subjected the IR spectra to a peak-fitting procedure starting with Lorentzian curves, which were centered at local maxima in the spectra. In the deconvoluted spectrum (**Figure 4A**), four distinct CO stretch bands are resolved at 2,166, 2,150, 2,104, and 2,085 cm−<sup>1</sup> for highly reduced α-Fe2O3(0001), in full agreement with those obtained after CO adsorption on the reduced Fe3O4(111) surface [see section CO Adsorption on Fe3O4(111) and Fe3O4(001)].

Interestingly, **Figure 4A** shows that the intensity of the Fe2+ related CO bands is much stronger than that of the Fe3+ related CO vibrations at 2,166/2,150 cm−<sup>1</sup> . A higher population of CO species adsorbed to Fe2<sup>+</sup> sites could account for this

observation. However, the substantial intensity difference should be also related to a larger transition dipole moment for CO-Fe2+, resulting from the enhanced electron backdonation compared to CO/Fe3+, as confirmed by the large red-shift in frequency of CO species bound to Fe2<sup>+</sup> cations. A similar phenomenon has been reported for CO adsorption on differently charged (Cu<sup>+</sup> and Cu2+) copper ions, where the CO stretch intensity was calculated to be more than a factor of five higher for the CO-Cu<sup>+</sup> than for the CO-Cu2<sup>+</sup> complex (St. Petkov et al., 2012).

Our IRRAS data strongly suggests that the highly reduced α-Fe2O3(0001) surface reconstructs. This conclusion is further supported by the corresponding LEED pattern (**Figure 4B**), which reveal that the α-Fe2O3(0001) surface undergoes a complete structural conversion from α-Fe2O3(0001) to Fe3O4(111) termination, in agreement with literature (Condon et al., 1994, 1998; Camillone et al., 2002). Furthermore, the outer hexagon spots become much broader and more intense, which could be attributed to the presence of additional, ordered Fe1−xO(111) domains [further detailed explanations are given in section CO Adsorption on Fe3O4(111) and Fe3O4(001)]. The surface structures of α-Fe2O3(0001), Fe3O4(111) and Fe1−xO(111) are illustrated in **Figure 5**. The LEED results indicate the formation of a Fe3O4(111)/Fe1−xO(111)-biphase structure on highly reduced α-Fe2O3(0001). The approach using non-destructive IR spectroscopy and CO as a probe molecule allows to provide more reliable information on the surface termination and chemical environments because CO only interacts with the outermost surface (Wang and Wöll, 2017; Chen et al., 2019). The CO vibrations at 2,166/2,104 cm−<sup>1</sup> and 2,150 cm−<sup>1</sup> could be attributed to octahedral Fe3+/Fe2<sup>+</sup> sites as well as to tetrahedral Fe3<sup>+</sup> sites within Fe3O4(111) domains, while the CO band the 2,085 cm−<sup>1</sup> implies the coexistence of reduced Fe1−xO(111) islands terminated by octahedral Fe2<sup>+</sup> cations. These assignments are corroborated by reference experiments where CO was adsorbed on Fe3O4(111) single-crystal surfaces, as discussed below in section CO adsorption on Fe3O4(111) and Fe3O4(001). Further evidence for this massive surface restructuring of the highly reduced α-Fe2O3(0001) is provided by photoelectron spectroscopy, where Fe2<sup>+</sup> and Fe3<sup>+</sup> species are clearly identified in both Fe 2p (**Figure S5A**) and valence band regions (**Figure S5B**) (Liu et al., 2012).

As mentioned in section Preview on Related Iron Oxide Crystal Structures, the highly reduced α-Fe2O3(0001) surface may also feature an ordered α-Fe2O3(0001)/Fe1−xO(111) biphase reconstruction which is characterized by floret-like satellite spots, positioned adjacent to the peaks for the pristine (1×1)-α-Fe2O3(0001) surface (Condon et al., 1995, 1998). However, this surface reconstruction is not supported by our combined results from IRRAS and LEED (**Figure 4**). Our observation is in line with the STM studies on α-Fe2O3(0001) single crystal surfaces (Condon et al., 1998), where a Fe3O4(111) termination was created by annealing the sample at ∼1,000 K in 10−<sup>6</sup> mbar O2. The structural transformation to a <sup>α</sup>-Fe2O3(111)/Fe1−xO(111)-biphase occurs only when increasing the temperature to about 1,100 K (Condon et al., 1998). In addition, more recent DFT calculations revealed that the α-Fe2O3(0001) surface can expose O-, Fe-, and ferryl-terminated regions, whose size varies depending on the oxygen chemical potential (Lewandowski et al., 2016). The long-range order was explained in terms of the dipole-dipole interaction between domains with different work functions.

# CO Adsorption on Fe3O4(111) and Fe3O4(001)

In order to aid the assignment of the CO bands and to gain a thorough understanding of the structural evolution of the α-Fe2O3(0001) surface during reduction, we have carried out additional reference measurements on Fe3O4(111) and reconstructed Fe3O4(001)-<sup>√</sup> 2 × √ 2R45◦ (**Figure S3**) single crystal surfaces. **Figure 6a** shows p-polarized IRRAS data recorded after saturating the slightly (two sputter/annealing cycles) and strongly (five sputter/annealing cycles) reduced Fe3O4(111) single-crystal surfaces with CO at 65 K, which is much lower than the Verwey transition temperature of ∼120 K (Verwey, 1939). The deconvoluted spectra display four CO stretch bands at 2,166, 2,150, 2,104, and 2,085 cm−<sup>1</sup> (**Figure 6a**),

950–1,000 K; (C) 1,000–1,050 K.

Å. Surface atoms are shown inked; subsurface atoms below the first oxide layer are depicted hollow. Red spheres correspond to oxide lattice ions; dark blue spheres and cyan spheres represent iron ions on octahedral and tetrahedral lattice sites, respectively.

in excellent agreement with the results found for the highly reduced surface of hematite α-Fe2O3(0001) (see **Figure 4A**).

As discussed in section Preview on Related Iron Oxide Crystal Structures, the surface structure of Fe3O4(111) is extremely sensitive to the preparation parameters (Kuhlenbeck et al., 2013; Parkinson, 2016). The experimental and theoretical studies revealed consistently that the regular Fe3O4(111) surface is Fetet1 terminated over a close-packed oxygen layer, whereas this surface is stabilized by the Feoct2-termination under oxygen-deficient conditions (Lennie et al., 1996; Ahdjoudj et al., 1999; Ritter and Weiss, 1999; Zhu et al., 2006; Grillo et al., 2008; Shimizu et al., 2010; Sala et al., 2012; Noh et al., 2015). The vibrational frequencies at 2,166/2,104 cm−<sup>1</sup> and 2,150 cm−<sup>1</sup> are assigned to CO species adsorbed to octahedral Fe3+/Fe2<sup>+</sup> sites (B type) as well as to tetrahedral Fe3<sup>+</sup> sites (A type), respectively, indicating a Feoct2-terminated Fe3O4(111) structure (see **Figure 5B**).

The assignment of the CO bands is further assisted by additional IRRAS data obtained for CO adsorption on the highly reduced α-Fe2O3(0001) surface created by annealing at 950 K in a higher O<sup>2</sup> pressure of 10−<sup>4</sup> mbar (see **Figure S6**). In this condition, the Fe3O4(111) layer within the Fe3O4(111)/Fe1−xO biphase should be dominated by the Fetet1-termination. Indeed, the corresponding IR data shows a single Fe3+-related vibrational band at 2,153 cm−<sup>1</sup> , which is attributed to CO molecules bound on surface Fetet1 sites (see **Figure S6**).

Furthermore, we assign the low-lying IR band at 2,085 cm−<sup>1</sup> to CO bound to Fe2<sup>+</sup> surface sites from reduced Fe1−xO phases (see **Figure 5C**), as reported in the literature (Benziger and

Larson, 1982). Again, the Fe1−xO(111) surface can expose either Fe or O layers according to the preparation conditions. Whereas, the STM investigations suggested the presence of O-terminated Fe1−xO(111) domains formed by oxidation of the Fe3O4(111) layer on α-Fe2O3(0001) (Tang et al., 2013), the DFT calculations predicted a Fe-termination in O-poor environments (Li et al., 2005). This assignment is supported by the observation that the 2,085 cm−<sup>1</sup> band becomes more intense along with the Ar+-sputtering/annealing treatment of the Fe3O4(111) singlecrystal surface (**Figure 6a**), indicating the growth of ordered Fe1−xO(111) domains, which coexist with the Feoct2-terminated Fe3O4(111) structures on the reduced Fe3O4(111) surface. The stretch frequencies of all CO species adsorbed on different iron(III/II) oxide single-crystal surfaces studied in the course of this work are summarized in **Table 1**.

Our IR data provide direct spectroscopic evidence for the formation of a Fe3O4(111)/Fe1−xO(111) biphase structure on Fe3O4(111), in line with the results from STM and LEED (Condon et al., 1997; Shaikhutdinov et al., 1999; Paul et al., 2007). Here, the presence of a Fe3O4(111)/Fe1−xO(111) biphase is also supported by the LEED patterns shown in **Figure 6B**. In addition to the pristine Fe3O4(111) structure (Tarrach et al., 1993), the spots at the outer hexagon become much broader and more intense under reducing conditions, which is also seen for the highly reduced α-Fe2O3(0001) surface (see **Figure 4B**). This should originate from the presence of ordered Fe1−xO(111) domains. However, we did not see the floreted diffraction patterns as observed for a typical α-Fe2O3(0001)/Fe1−xO(111) biphase. This can be attributed to the rather small lattice mismatch (2%) between Fe3O4(111) and Fe1−xO(111), which makes an unambiguous observation of the floret-like satellite spots extremely difficult (Condon et al., 1997; Paul et al., 2007).

The temperature-dependent IRRAS data allows to gain deep insight into the interaction of CO molecule with various surface sites. As shown in **Figure 7A**, the Fe3+-related CO bands at 2,166 and 2,150 cm−<sup>1</sup> disappear after annealing the sample to ∼100 K, indicating a weak binding of CO to coordinatively unsaturated surface Fe3<sup>+</sup> cations. The thermal stability and the frequency (blue-shifted compared to the gas phase value, see **Table 1**) are in good agreement with those observed on pristine

TABLE 1 | CO stretch frequencies collected on regular and restructured hematite (0001) as well as on magnetite (111) and (001) single-crystal surfaces.


The frequencies are shown for saturating CO exposures at 65 K.

α-Fe2O3(0001) (**Figure 2B**) and Fe3O4(001) surfaces (**Figure 7B**, see below), in which both are terminated by single Fe3<sup>+</sup> species. In comparison, the Fe2+-related CO species at 2,104 and 2,085 cm−<sup>1</sup> are thermodynamic more stable and desorb completely only upon annealing to temperatures above 165 K (**Figure 7A**). Again, the stronger binding of CO to Fe2<sup>+</sup> is attributed to the enhanced electron back-donation to the CO 2π ∗ antibonding orbital, thus strengthening the CO-substrate interaction and weakening the internal C–O bond (redshift in frequency).

In addition, a new weak IR band at 2,205–2,209 cm−<sup>1</sup> is detected after annealing to temperatures above 85 K. The origin of this band is not clear, but a similar CO species has been reported for ZnO and Au/ZnO systems and is assigned to CO coadsorbed with carbonate species that increase the Lewis acidity of adjacent Zn2<sup>+</sup> cations (Wang et al., 2007; Noei et al., 2012). Carbonate could be produced via the oxidation of adsorbed CO with surface O atoms on Fe3O4(111) (Tinkle and Dumesic, 1987; Huang et al., 2006). In a recent study on Fe3O4(111) thin films grown on Pt(111) (Li et al., 2018), a CO vibration at 2,204 cm−<sup>1</sup> was also observed. According to the DFT calculations, this band was ascribed to CO adsorbed at Fe2<sup>+</sup> oct2 ions located at step edges (defect sites). However, the Fe3O4(111) thin film, prepared under oxidation conditions, features a Fe3<sup>+</sup> tet -termination over closepacked oxygen layer, which is quite different from the present work focusing on the structural evolution of iron oxides under reducing conditions.

We now turn to the experimental results obtained on the reconstructed Fe3O4(001)-(<sup>√</sup> 2 × √ 2)R45◦ surface. After adsorption of CO on the Fe3O4(001) surface at 65 K only one single sharp CO band at 2,169 cm−<sup>1</sup> is observed (**Figure 7B**), which is typical for CO bound to Fe3<sup>+</sup> sites. As discussed in section Preview on Related Iron Oxide Crystal Structures, the (<sup>√</sup> 2 × √ 2)R45◦ -reconstructed Fe3O4(001) surface has been studied extensively by numerous groups (Wiesendanger et al., 1992; Tarrach et al., 1993; Chambers and Joyce, 1999; Rustad et al., 1999; Stanka et al., 2000; Mariotto et al., 2002; Parkinson et al., 2011; Bartelt et al., 2013; Bliem et al., 2014). Recently, Parkinson et al. (Bliem et al., 2014) presented a thorough study combining STM, LEED IV analysis and theory, in which a subsurface cation vacancy (SCV) model was developed. According to this model, the reconstructed surface is Feoct-O terminated, which is accompanied by a rearrangement of the cations in the subsurface layers (Bliem et al., 2014).

Our IRRAS results are fully consistent with the SCV structure in that the single IR band at 2,169 cm−<sup>1</sup> is characteristic for CO bound to surface Fe3<sup>+</sup> oct sites. The assignment to Fe3<sup>+</sup> is further supported by a temperature-programmed desorption measurement, where the IRRAS data reveal a complete desorption of CO species after heating to 105 K, in excellent agreement with the TPD results (Hulva et al., 2018). This finding demonstrates again a weak interaction between CO and surface Fe3<sup>+</sup> cations as observed for CO adsorption on the Fe3<sup>+</sup> sites of α-Fe2O3(0001) and Fe3O4(001) surfaces. In addition, the thermal desorption of CO leads to a blue shift in frequency from 2,169 to 2,181 cm−<sup>1</sup> (**Figure 7B**), which is attributed to the lateral adsorbate-adsorbate interactions [for a detailed discussion see section CO Adsorption on the Pristine α-Fe2O3(0001) Surface].

Overall, the IRRAS results, acquired for CO adsorption on the macroscopic Fe3O4(111) and Fe3O4(001) surfaces, provide a solid basis for a reliable assignment of the CO bands observed for the restructured α-Fe2O3(0001) surface under different reduction conditions (**Table 1**). Importantly, the comprehensive IR data, together with XPS and LEED analysis, allow for identifying the

Fe3O4(111)/Fe1−xO(111) biphase structure on highly reduced α-Fe2O3(0001) surfaces.

# CONCLUSIONS

In summary, the structural evolution of differently treated iron(III/II) oxide single-crystal surfaces was monitored by polarization-resolved IRRAS using CO as a probe molecule. We presented a comprehensive investigation on stoichiometric and reduced α-Fe2O3(0001) surfaces as well as on Fe3O4(001) and Fe3O4(111) surfaces. The polarization-dependent IRRAS results demonstrate that the CO adsorption is well suited to monitor the atomic structure evolution of iron oxide surfaces under different conditions.

The pristine stoichiometric α-Fe2O3(0001) surface is terminated in a single-Fe3<sup>+</sup> configuration as characterized by the single sharp IR band at 2,169 cm−<sup>1</sup> (CO on 3-fold coordinated Fe3+). Annealing the hematite (0001) surface to elevated temperatures leads to the formation of surface oxygen vacancies. The reduced α-Fe2O3(0001) surface shows a typical vibrational band at 2,163 cm−<sup>1</sup> (CO on Fe3<sup>+</sup> at O vacancy). A further reduction of α-Fe2O3(0001) induced by atomic hydrogen treatment followed by Ar+-sputtering and annealing cycles under oxygen poor conditions causes a massive surface restructuring yielding a Fe3O4(111) overlayer, which coexists with Fe1−xO(111) domains. The reconstructed Fe3O4(111) surface is Feoct2-terminated and shows typical CO bands at 2,166/2,150 and 2,104 cm−<sup>1</sup> . They are assigned to Fe3+(oct)/Fe3+(tet) and Fe2+(oct) sites, respectively. The Fe1−xO(111) domains within the Fe3O4(111)/Fe1−xO(111) biphase are characterized by the intense band at 2,085 cm−<sup>1</sup> originating from CO adsorbed on Fe2+(oct) sites. The assignment is further demonstrated by reference IRRAS

# REFERENCES


data acquired for CO adsorption on Fe3O4(111) single-crystal surfaces. The surface of Fe3O4(001) is terminated with Fe3<sup>+</sup> oct-O in a SCV reconstruction as determined by the single intense IR band of CO at 2,169 cm−<sup>1</sup> .

# DATA AVAILABILITY

All datasets generated for this study are included in the manuscript and/or the **Supplementary Files**.

# AUTHOR CONTRIBUTIONS

YW and CW conceived the idea of the experiment. LS and CY carried out the experiments, with additional help by AN and SH. Data analysis was carried out by LS, CY, and YW. The manuscript was written by LS, YW, and CW.

# ACKNOWLEDGMENTS

The Helmholtz-Research-School Energy-related-catalysis is gratefully acknowledged for providing financial support and the donation of a Ph.D.-Scholarship to LS. We acknowledge funding from the Science and Technology of Nanosystems Program (432202). The authors thank Helmholtz-Zentrum Berlin for the allocation of synchrotron radiation beamtime at BESSY II.

# SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fchem. 2019.00451/full#supplementary-material

Additional XPS, LEED, and polarization-resolved IRRAS analysis.


Noei, H., Birkner, A., Merz, K., Muhler, M., and Wang, Y. (2012). Probing the mechanism of low-temperature CO oxidation on Au/Zno catalysts by vibrational spectroscopy. J. Chem. C 116, 11181–11188. doi: 10.1021/jp302723r

Noh, J., Osman, O. I., Aziz, S. G., Winget, P., and Brédas, J. L. (2015). Magnetite Fe3O<sup>4</sup> (111) surfaces: impact of defects on structure, stability, and electronic properties. Chem. Mater. 27, 5856–5867. doi: 10.1021/acs.chemmater.5b02885

Ovcharenko, R., Voloshina, E., and Sauer, J. (2016). Water adsorption and O-defect formation on Fe2O3(0001) surfaces. Phys. Chem. Chem. Phys. 18, 25560–25568. doi: 10.1039/C6CP05313K

Parkinson, G. S. (2016). Iron oxide surfaces. Surf. Sci. Rep. 71, 272–365. doi: 10.1016/j.surfrep.2016.02.001

Parkinson, G. S., Novotný, Z., Jacobson, P., Schmid, M., and Diebold, U. (2011). A metastable Fe(A) termination at the Fe3O4(001) surface. Surf. Sci. 605, L42–L45. doi: 10.1016/j.susc.2011.05.018

Paul, M., Sing, M., Claessen, R., Schrupp, D., and Brabers, V. A. M. (2007). Thermodynamic stability and atomic and electronic structure of reduced Fe3O<sup>4</sup> (111) single-crystal surfaces. Phys. Rev. B 76:075412. doi: 10.1103/PhysRevB.76.075412

Ritter, M., and Weiss, W. (1999). Fe3O<sup>4</sup> (111) surface structure determined by leed crystallography. Surf. Sci. 432, 81–94. doi: 10.1016/S0039-6028(99)00518-X


Stanka, B., Hebenstreit, W., Diebold, U., and Chambers, S. A. (2000). Surface reconstructions of Fe3O4(001). Surf. Sci. 448, 49–63. doi: 10.1016/S0039-6028(99)01182-6


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Schöttner, Nefedov, Yang, Heissler, Wang and Wöll. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# The Structure of Oxygen Vacancies in the Near-Surface of Reduced CeO<sup>2</sup> (111) Under Strain

Zhong-Kang Han1†‡, Lei Zhang2†, Meilin Liu<sup>2</sup> , Maria Verónica Ganduglia-Pirovano<sup>3</sup> \* and Yi Gao1,4 \*

<sup>1</sup> Shanghai Institute of Applied Physics, Chinese Academy of Sciences, Shanghai, China, <sup>2</sup> Center for Innovative Fuel Cell and Battery Technologies, School of Materials Science and Engineering, Georgia Institute of Technology, Atlanta, GA, United States, <sup>3</sup> Institute of Catalysis and Petrochemistry, Spanish National Research Council (CSIC), Madrid, Spain, <sup>4</sup> Zhangjiang Laboratory, Shanghai Advanced Research Institute, Chinese Academy of Sciences, Shanghai, China

### Edited by:

Zhuhua Zhang, Nanjing University of Aeronautics and Astronautics, China

### Reviewed by:

Liangzhi Kou, Queensland University of Technology, Australia Luqing Wang, Rice University, United States

### \*Correspondence:

Maria Verónica Ganduglia-Pirovano vgp@icp.csic.es Yi Gao gaoyi@sinap.ac.cn

†These authors have contributed equally to this work

### ‡Present address:

Zhong-Kang Han, Fritz Haber Institute of the Max Planck Society, Berlin, Germany

### Specialty section:

This article was submitted to Physical Chemistry and Chemical Physics, a section of the journal Frontiers in Chemistry

> Received: 03 March 2019 Accepted: 28 May 2019 Published: 18 June 2019

### Citation:

Han Z-K, Zhang L, Liu M, Ganduglia-Pirovano MV and Gao Y (2019) The Structure of Oxygen Vacancies in the Near-Surface of Reduced CeO<sup>2</sup> (111) Under Strain. Front. Chem. 7:436. doi: 10.3389/fchem.2019.00436 Strain has been widely recognized as important for tuning the behavior of defects in metal oxides since properties such as defect configuration, electronic structure, excess charge localization, and local atomic distortions may be affected by surface strain. In CeO2, the most widely used promoter in three-way catalysts and solid state electrolyte in fuel cells, the behaviors of oxygen vacancies, and associated Ce3<sup>+</sup> polarons are crucial in applications. Recent STM and AFM investigations as well as DFT-based calculations have indicated that in the near-surface of CeO<sup>2</sup> (111), at low temperatures and vacancy concentrations, subsurface oxygen vacancies are more stable than surface ones, and the Ce3<sup>+</sup> ions are next-nearest neighbors to both types of vacancies, which can be explained by the better ability of the system to relax the lattice strain induced by vacancy formation as well as by the excess charge localization. The results also revealed that the interaction between first-neighbor vacancies is repulsive. In this work, the relative stability of surface and subsurface oxygen vacancies at the CeO<sup>2</sup> (111) surface under in-plane strain is investigated by means of DFT+U calculations. The tensile strain favors isolated surface vacancies with next nearest neighbor polarons, whereas isolated subsurface vacancies with nearest neighbor polarons are energetically favored under compressive strain. In addition, the formation of both surface and subsurface dimers is favored over having corresponding isolated species under compressive strain, which implies the possibility of controlling the formation of vacancy clusters using strain. In many applications, ceria is employed as a supported thin film or within a heterostructure in which ceria can be strained, and this study shows that strain can be a useful handle to tune properties of such materials.

Keywords: CeO2, density functional theory, oxygen vacancy, strain, surface structures

# INTRODUCTION

Ceria is widely used as solid oxide fuel cell electrolyte (Inaba and Tagawa, 1996) and in catalytic applications (Trovarelli, 1996; Fu et al., 2003; Vayssilov et al., 2011; Montini et al., 2016) largely due to its facile oxygen vacancy formation and diffusion, either within the bulk or at its surfaces. It has been previously reported (Skorodumova et al., 2002), and later supported by subsequent research (Ganduglia-Pirovano et al., 2007, 2009; Li et al., 2009; Shoko et al., 2010; Jerratsch et al., 2011; Paier et al., 2013), that the formation of an intrinsic neutral oxygen vacancy Han et al. Heterogeneous Catalysis

in ceria is accompanied by the formation of two nearby polarons (Ce3<sup>+</sup> or Ce′ Ce in Kröger-Vink notation), which result from the transfer of two electrons originally residing in 2p states of the missing oxygen ion to Ce 4f states of two cations. For oxygen vacancies in the near-surface of CeO<sup>2</sup> (111), it has been predicted from Density Functional Theory (DFT) calculations that the two Ce3<sup>+</sup> polarons are not necessarily NN (Nearest-Neighbor) to the oxygen vacancies, but rather prefer to locate at NNN (Next-Nearest-Neighbor) cationic sites (Ganduglia-Pirovano et al., 2007, 2009; Li et al., 2009; Paier et al., 2013), which has been later validated in Scanning Tunneling Microscope (STM) experiments (Jerratsch et al., 2011). The preference for the NNN positions upon localization of the excess charge has been explained by the better ability of the system to relax the lattice strain induced by the presence of the vacancies as well as by the excess charge localization; a Ce3<sup>+</sup> ion is more spacious than its Ce4<sup>+</sup> counterpart and at NN sites Ce3<sup>+</sup> – O bonds would be compressed (Ganduglia-Pirovano et al., 2009). Moreover, the preferred sites for oxygen vacancies in the near-surface of CeO<sup>2</sup> (111) (Esch et al., 2005; Castleton et al., 2007; Torbrügge et al., 2007; Ganduglia-Pirovano et al., 2009; Li et al., 2009; Jerratsch et al., 2011; Murgida and Ganduglia-Pirovano, 2013; Sutton et al., 2015) as well as the type of interactions between them (Namai et al., 2003; Esch et al., 2005; Torbrügge et al., 2007; Conesa, 2009; Murgida and Ganduglia-Pirovano, 2013; Kullgren et al., 2014, 2017; Han et al., 2018), have also been assessed. Under zero applied stress, isolated subsurface vacancies are more stable than surface ones, and the interaction between nearest neighbor vacancies at the surface or in the subsurface is repulsive.

As for isolated oxygen vacancies in bulk ceria, most theoretical studies reported that the NNN Ce3<sup>+</sup> locations are also preferred under zero applied stress (Kullgren et al., 2012; Allen and Watson, 2014; Murgida et al., 2014; Wang et al., 2014; Grieshammer et al., 2016; Han et al., 2017), i.e., at the calculated equilibrium volume of CeO2. Moreover, it has been demonstrated that the ground-state configuration varies whether tensile or compressive stress is applied (Arapan et al., 2015), namely, at large volumes, both Ce3<sup>+</sup> ions are in the second coordination sphere of the vacancy, and at small volumes, they are in the first. Additionally, it has been experimentally found that the lattice parameter of reduced ceria in the fluoritetype cubic structure is expanded (Rossignol et al., 2003), which has been consistently reproduced in DFT calculations (Arapan et al., 2015) that predict that the averaged lattice parameter over different configurations for the pair of Ce3+'s is slightly larger than the equilibrium lattice parameter of perfect ceria.

Furthermore, for doped ceria with three-valent dopants (Andersson et al., 2006; Wang and Cormack, 2012; Grieshammer et al., 2016), the dependence of the location of the dopants on the dopant's ionic radius has been reported, in which larger ions such as La prefer NNN sites to vacancies, and smaller ions such as Gd prefer NN sites. It has been argued that it is the balance between repulsive elastic and attractive electronic contributions to the interaction between dopants and vacancies what determines the NN or NNN site preference (Andersson et al., 2006), and, as in the case of undoped ceria mentioned above, applied tensile or compressive stress can change the site preferences (Wang and Cormack, 2012). In addition, changes in the lattice parameter upon ceria doping have been reported, as, for example, the one of Gd-doped ceria (Artini et al., 2014; Žguns et al., 2019) that shows a maximum as a function of Gd content.

Strain in ceria has received considerable attention because oxygen vacancy migration can be modified through strain, which is most relevant for tuning ion conduction in ceria-based applications (De Souza et al., 2012; Hinterberg et al., 2013; Sun et al., 2015). It is fair to say that it is practically inevitable to avoid that ceria surfaces will experience strain effects, either induced by the lattice mismatch when created as a thin film on a substrate (Duchon et al., 2013; Luches et al., 2014), or when reduced (Grieshammer et al., 2016), particularly due to the higher stability of oxygen vacancies in the near-surface as compared to deeper layers (Duchon et al., 2013; Murgida and Ganduglia-Pirovano, 2013; Sutton et al., 2015; Olbrich et al., 2017) that are accompanied by the formation of Ce3<sup>+</sup> ions which induces a near-surface lattice expansion. For reduced ceria nanoparticles, mainly expansions of the lattice have been observed, which were associated to the presence of Ce3<sup>+</sup> cations (Rossignol et al., 2003; Wang and Cormack, 2012; Allen and Watson, 2014; Arapan et al., 2015), but also lattice contractions have been reported (Wang and Cormack, 2012), which were attributed to the additional pressure caused by the surface tension between the crystallite and the ambient atmosphere as the nanoparticle size decreases. Despite these efforts, the effect of applying a tensile or compressive stress to the reduced CeO<sup>2</sup> (111) surface has not yet been comprehensively studied.

In this article, using DFT-based methods, we systematically investigate the effect of applied stress on the reduced CeO<sup>2</sup> (111) surface and address how strain affects the relative stability of isolated surface and subsurface oxygen vacancies, the formation of vacancy pairs, and the localization of the excess charge. It turns out that, depending on strain, surface vacancies can be more stable than subsurface ones, the interaction between first nearest neighbor vacancies in the near-surface can be attractive, and different localized charge distributions can be attained.

# METHODOLOGY

First-principles spin-polarized DFT-based calculations were carried out employing the VASP (Vienna Ab-initio Simulation Package) code. The DFT+U methodology (Dudarev et al., 1998) was used with the generalized gradient approximation (GGA) of Perdew-Burke-Ernzerhof (PBE) (Perdew et al., 1996), and an effective U-value of 5.0 eV. The value of the U parameter necessary to describe localized electrons associated to Ce3<sup>+</sup> ions—lies within the range of suitable values to describe reduced ceria-based systems (Castleton et al., 2007). We used projector augmented wave (PAW) potentials (Kresse and Joubert, 1999) with Ce (4f, 5s, 5p, 5d, 6s) and O (2s, 2p) electrons as valence states, and a plane-wave cutoff of 400 eV. The CeO<sup>2</sup> (111) surface was modeled using three O-Ce-O trilayers separated by 15 Å vacuum space (i.e., a nine atomic layer slab) with both 5 × 5 and 2 × 2 periodicities. A 1 × 1 × 1 (3 × 3 × 1) Gamma-centered Monkhorst-Pack grid was used for the k-point sampling of the Brillouin zone of the 5 × 5 (2 × 2) slab. Stress ranging from −5% (compressive) to +5% (tensile) has been applied to the CeO<sup>2</sup> (111) surface, i.e., to the two vectors which define the surface unit cell. It should be mentioned that for compressive stress, experiments show that CeO<sup>2</sup> transforms from a cubic fluorite-type structure to an orthorhombic cotunnitetype (PbCl2) structure at a pressure of 31 GPa (Duclos et al., 1988). However, in this study, the underlying fluorite lattice is prevented from undergoing a phase transition. In all geometry optimizations, all atoms in the bottom trilayer were kept fixed after stress had been applied, whereas the rest of the atoms were allowed to fully relax. For all strained surfaces, the interlayer spacings within the bottom trilayer correspond to those of a bulk-truncated CeO<sup>2</sup> (111) under zero stress, and the in-plane atom positions correspond to those in the corresponding strained CeO<sup>2</sup> bulk.

To systematically investigate how strain affects the relative stability of isolated surface and subsurface oxygen vacancies in the near-surface of CeO<sup>2</sup> (111), as well as the formation of vacancy pairs, and the localization of the excess charge, reduced strained slabs were created. Isolated surface and subsurface vacancies with different configurations for the Ce3+'s were considered with both 5 × 5 and 2 × 2 periodicities. Vacancies were placed near the surface, i.e., within the outermost trilayer, and the Ce3+'s within the (two) outermost cationic plane(s) for the 5 × 5 (2 × 2) periodicity. Moreover, using the larger 5 × 5 surface unit cell, nearest and next-nearest neighbor surface and subsurface vacancy pairs with various configurations for the Ce3+'s were considered. In order to obtain distinct configurations of the reduced Ce3<sup>+</sup> sites, a two-step relaxation procedure was applied. In the first step, we replaced the selected Ce4<sup>+</sup> with Ce3<sup>+</sup> ions, i.e., for the latter we used PAW potentials for which the 4f 1 state was moved to the core, and performed non-spin polarized calculations. The so-obtained relaxed structure was further optimized using the regular Ce4<sup>+</sup> PAW potentials and spin-DFT. We limit the discussion to high-spin states because the difference between these states and any other spin state is <0.01 eV (Keating et al., 2009). The oxidation state of a given Ce atom has been estimated by considering its local magnetic moment (the difference between up and down spin on the atoms), which can be estimated by integrating the site- and angular momentumprojected spin-resolved density of states over spheres with radii chosen as the Wigner–Seitz radii of the PAW potentials. For reduced Ce ions, the occupation of Ce f states is close to one, and the magnetic moment is ∼1 µB. Hence, those ions are referred to as Ce3<sup>+</sup> (4f 1 ).

# RESULTS AND DISCUSSION

We have modeled the reduced CeO<sup>2</sup> (111) surface by removing one single surface (SSV) or subsurface (SSSV) vacancy from a supercell with 5 × 5 and 2 × 2 periodicities, as shown in **Figures 1**, **2**, respectively. Nearest (NN) and next-nearest neighbor (NNN) Ce ions to vacancies are labeled with uppercase letters (A, B, C, D, . . . ) and numerals (1, 2, 3, 4, . . . ), respectively. Employing the larger supercell, first-neighbor vacancy dimers at the surface or in the subsurface (SurDimer and SubDimer), as well as corresponding pairs with a vacancy separation corresponding to that of third-nearest neighbors in the oxygen plane (SurDimer-d and SubDimer-d), have also been considered (cf. **Figure 2**). Stress ranging from −5% (compressive) to +5% (tensile) has been applied to both supercells as mentioned in section Methodology. It is known that under zero applied stress, NNN sites in the outermost cerium layer are the energetically preferred locations of the Ce3<sup>+</sup> ions, but they would rather be in NNN sites of a deeper layer than next to a vacancy in the outermost one (Murgida and Ganduglia-Pirovano, 2013). Since under applied stress preferences may change, for the strained surfaces with 5 × 5 (**Tables S1–S7**) or 2×2 (**Table S8**) periodicity, vacancy structures corresponding to the most relevant combinations of the possible locations of the Ce3<sup>+</sup> ions have been considered.

The averaged vacancy formation energy (E<sup>f</sup> ) was calculated as:

$$E\_f = \frac{1}{n} [E\_{defect} + \frac{n}{2} E\_{O\_2} - E\_{perfect}] \tag{1}$$

where Edefect and Eperfect are the total energies of the (relaxed) defective (reduced) and perfect (clean) slabs, respectively, n is the number of oxygen vacancies, and EO<sup>2</sup> is the total energy of the isolated O<sup>2</sup> molecule in its triplet ground state.

# Low Vacancy Concentration Stability, Structural Relaxation, and Excess Charge Localization

**Figures 3**, **4** summarize the calculated vacancy formation energy for all 5 × 5 structures with a SSV and a SSSV vacancy, respectively. For the surface vacancy, configurations with both

Ce3<sup>+</sup> ions in the outermost cationic plane and either in the first coordination shell (AB), in the second (12), or one in each shell (1C) were considered (**Figure 1**). In addition, a configuration with both Ce3<sup>+</sup> ions in the second coordination shell, but one in the outermost cationic plane and the other one in the plane beneath, was considered (1S) (**Figure S1**). Under 0% tension, the 12 configuration is by 0.23 eV more stable than the AB one (**Figure 3**), in line with previous works (Ganduglia-Pirovano et al., 2009; Li et al., 2009; Murgida and Ganduglia-Pirovano, 2013; Han et al., 2018). As the tensile strain is increased from 0 to +5%, the 12 >1C >1S >AB stability trend does not change although the difference between the most stable configuration (12) and the least stable one (AB) becomes larger (0.33 eV under +5% strain, **Table S1**). In contrast, if we look at the surface under large compressive strain (≥ −5%), the trend is reversed, i.e., AB >1C>12; (the 1S configuration is not stable); under −5% strain, the configuration with both Ce3<sup>+</sup> ions in the first coordination shell (AB) is by 0.12 eV more stable than that with the Ce3<sup>+</sup> ions in the second coordination shell (12).

For the subsurface vacancy, configurations with both Ce3<sup>+</sup> ions in the outermost cationic plane and either in the first coordination shell (AB) or in the second (12), or one in each shell (1C), as well as with both Ce3<sup>+</sup> ions in the second coordination shell, but one in the outermost cationic plane and the other one in the plane beneath (1S), were also considered (cf. **Figure 1** and **Figure S1**). The behavior of the SSSV is comparable to that of the SSV, although the slopes of the energy-strain curves for the SSSV are somewhat smaller than those of the SSV (cf. **Figures 3**, **4**). Under 0% tension, the SSSV-12 configuration is by 0.32 eV more stable than the AB one, in line with previous works (**Figure 4**) (Ganduglia-Pirovano et al., 2009; Li et al., 2009; Murgida and Ganduglia-Pirovano, 2013; Han et al., 2018). The

12 >1C >1S >AB stability trend obtained under 0% tension, does not change upon increasing the tensile strain from 0 to +4% (the AB configuration is not stable under +5% strain), but the difference between the 12 and AB configurations becomes larger, for example, it is 0.42 eV under +4% strain (**Table S2**). If the surface is compressed, for example by −5%, the most stable configuration has both Ce3<sup>+</sup> ions in the first coordination shell (AB), and that with the Ce3<sup>+</sup> ions in the second coordination shell (12), is by 0.07 eV less stable.

Before analyzing the origin of the observed transitions between NNN-NNN (12) and NN-NN (AB) Ce3<sup>+</sup> configurations for both vacancy types under large compressive strain (cf. section Electronic Density of States), we first discuss how their relative stability is influenced by strain. **Figure 5** collects the results of the most stable SSV and SSSV configurations for each value of the applied stress (cf. **Figures 3**, **4**, and **Tables S1, S2**, **S7**). In agreement with the literature (Ganduglia-Pirovano et al., 2009; Li et al., 2009; Murgida and Ganduglia-Pirovano, 2013), we find that under 0% tension, the subsurface vacancy position is by 0.23 eV more stable than the surface one. Under compressive strain, from 0 to −5%, the SSSV remains more stable than the SSV, but the difference becomes larger as the compression increases, for example under −5% compression, the SSSV is by 0.56 eV more stable than the SSV. However, as the tensile tension increases from +3%, the relative stability reverses and the SSV becomes more stable.

As vacancy-induced lattice relaxations are known to play an important role in the stabilization of both a particular Ce3<sup>+</sup> configuration and the subsurface vacancy position in the nearsurface of CeO<sup>2</sup> (111) under 0% strain, (Ganduglia-Pirovano et al., 2009; Li et al., 2009) we analyze their effects on the strained surfaces. To this end, we define two contributions to

the vacancy formation energy, E<sup>f</sup> = E<sup>b</sup> + E<sup>r</sup> (cf. Equation 1), namely, the bond breaking energy, E<sup>b</sup> , as the energy cost to create a vacancy without allowing for lattice relaxations, i.e., E<sup>b</sup> = E unrelax defect <sup>+</sup> 1 2 EO<sup>2</sup> − Eperfect, and the energy gained from the structural relaxation in the presence of the vacancy, E<sup>r</sup> , i.e., E<sup>r</sup> = E relax defect <sup>−</sup> <sup>E</sup> unrelax defect . In **Figure 6** and **Table S9**, we show how these different contributions depend on strain.

TABLE 1 | O–Ce bond lengths (in Å) between a surface oxygen atom at the fully oxidized strained CeO<sup>2</sup> (111) surface (5 × 5 periodicity), Os, and its three nearest neighbor cerium atoms in the outermost cerium layer, Ce1, as well as between a subsurface oxygen atom, Oss, and its three (one) nearest neighbor cerium atoms in the first (second) cerium layer, Ce<sup>1</sup> (Ce2).


For both SSV and SSSV, E<sup>b</sup> is largest for the highest compression and smallest for the largest tensile stress, i.e., the correlation between E<sup>b</sup> and strain is negative. To create unrelaxed surface and subsurface oxygen vacancies, three Ce–O<sup>s</sup> and four Ce–Oss bonds, respectively, have to be cut. For the SSV, the Ce1-O<sup>s</sup> bond lengths (Ce<sup>1</sup> and O<sup>s</sup> are in the outermost cationic and anionic layers, respectively) vary from 2.30 to 2.43 Å as the stress changes from −5 to +5% (**Table 1**), thus, the observed E<sup>b</sup> vs. strain trend is in line with the weakening of the bonds upon increasing tensile strain. For the SSSV, the corresponding variations in the three Ce1-Oss bond lengths are practically identical to those of the SSV, but the Ce2-Oss bond (Ce<sup>2</sup> and Oss are in the second cationic and anionic layers, respectively) is contracted from 2.52 (−5%) to 2.26 (+5%). In any case, the observed E<sup>b</sup> vs. strain trend for the SSSV vacancy is in line with the weakening of the average Ce–Oss bond length upon increasing tensile strain, which varies from 2.35 (−5%) to 2.40 (+5%).

Moreover, we observe that the E<sup>b</sup> vs. strain curves for both the SSV and the SSSV resemble straight lines but with a different slope, i.e., the curves cross (**Figure 6**). Under the largest tensile strain (+5%), three Ce1-O<sup>s</sup> bonds of 2.43 Å for the SSV have to be cut (**Table 1**), whereas for the SSSV they are three Ce1-Oss bonds of 2.45 Å and one particularly strong Ce2-Oss bond of 2.26 Å; hence, creating an unrelaxed SSV is preferred. As already mentioned, upon compression the Ce2-Oss bond expands in such a way that for the largest compressive strain (−5%), three Ce–O<sup>s</sup> bonds with an average distance of 2.30 Å and four Ce–Oss ones of 2.35 Å have to be cut to create an unrelaxed SSV and SSSV, respectively (**Table 1**). The latter is energetically preferred.

In **Figure 6** we further compare the energy gained from structural relaxation in the presence of oxygen vacancies, E<sup>r</sup> (< 0), as a function of surface strain. The E<sup>r</sup> vs. strain curves also resemble straight lines with a different slope for the SSV and SSSV, but the curves do not cross within the interval of −5 to +5% strain values. The lattice relaxations are the result of both the presence of the oxygen vacancy, which for example induces the neighboring O (Ce) ions to move toward (away from) the vacancy, and the localization of the excess charge driving the Ce4<sup>+</sup> → Ce3<sup>+</sup> reduction of two cations. For isolated oxygen vacancies in bulk ceria under strain (De Souza et al., 2012), it has been shown that the energy gained due to the displacements of the atoms upon relaxation increases almost linearly with the increasing of the tensile stress, independent of the sites on which the excess charge is localized, but the gain is more pronounced when the localization occurs at NNN cationic sites to the vacancy. Similarly, for both SSV and SSSV vacancies in the near-surface of CeO<sup>2</sup> (111), the energy gained from structural relaxation is smallest for the biggest compressive strain and largest for the highest tensile stress, i.e., the correlation between the absolute value of Erand strain is positive. We note that the energy gain is always largest for the SSSV within the −5 to +5% strain interval, which is related to the fact that the numbers of neighboring O and Ce ions that will be displaced upon the creation of a SSSV are larger than the corresponding ones for a SSV. Moreover, the relaxation energy gained from the localization of the excess charge also favors the SSSV; for the unrelaxed SSSV and SSV structures, the two excess electrons are equally shared by four and three NN Ce ions, respectively, and thus, the energy gained due to the 4×Ce0.5<sup>+</sup> → 2×Ce3<sup>+</sup> charge localization in the case of the SSSV, is larger than the corresponding one for the SSV, namely, 3×Ce0.6ˆ+ <sup>→</sup> <sup>2</sup>×Ce3+. Furthermore, with respect to the above-mentioned smallest gain in the relaxation energy for the largest compressive strain for both SSV and SSSV, we recall here –and address below– that the Ce3<sup>+</sup> ions in the first coordination sphere are preferred when the surface is substantially compressed and then, it is expected that the contribution associated with the relocation of electrons will be lower when it occurs between first neighbors.

Summarizing, there is an almost linear correlation between both contributions to the oxygen vacancy formation and the surface strain. The correlation between the bond breaking energy, Eb , and strain, is negative and that between the absolute energy gained due to relaxation, –E<sup>r</sup> , and strain is positive. The slopes of the almost linear E<sup>b</sup> vs. strain and E<sup>r</sup> vs. strain relationships are not the same for the SSV than for the SSSV. The preference for the SSSV under compressive strain and for the SSV under tensile strain (>3%) is the result of the different behaviors of the contributions to create the corresponding vacancies with strain. For example, for the −5% compressed surface, both E<sup>b</sup> and E<sup>r</sup> favor the SSSV by 0.14 and 0.42 eV, respectively, and thus, the SSSV is by 0.56 eV more stable than the SSV. However, for the +5% strained surface, E<sup>b</sup> favors the SSV by 0.31 eV and E<sup>r</sup> the SSSV by 0.14 eV, and, as a result, the SSV is by 0.17 eV more stable than the SSSV.

### Electronic Density of States

The effects of strain on the distribution of the excess charge are also reflected in the calculated densities of states (DOS). In pure CeO2, all valence Ce states, including the 4f states, are empty and the system is a wide gap insulator with a measured fundamental band gap of 6 eV between the valence and the conduction bands (Pfau and Schierbaum, 1994), which are formed predominantly by the O 2p and Ce 5d states, respectively. The vacant 4f states lie in the O2p-Ce5<sup>d</sup> gap. PBE+U underestimates the O2p-Ce5<sup>d</sup> gap of bulk CeO<sup>2</sup> by ∼12% (5.3 eV) (Da Silva et al., 2007). The formation of a pair of Ce3<sup>+</sup> ions following the creation of an oxygen vacancy results in the appearance of split-off (defect) states of the initially empty Ce 4f band, lying inside the O2p-Ce5<sup>d</sup> gap below the Fermi level (Jerratsch et al., 2011). If the two Ce3<sup>+</sup> ions occupy different cationic shells around the vacancy,

the resulting variation in the chemical environment leads to a splitting of the two filled Ce3<sup>+</sup> f levels, which could be detected with STM spectroscopy (Jerratsch et al., 2011). The densities of states for the example of an SSV at the CeO<sup>2</sup> (111) surface with the Ce3+s in either the 12, 1C, or AB configuration under −5, 0, or +5% strain (**Figure 7**) are discussed in the following.

In the 12 configuration, the two next-nearest neighbor Ce3<sup>+</sup> ions have equal coordination number (CN = 7) and filled f orbitals with the same energy as they experience the same chemical environment, and hence no splitting of the f orbitals is expected; this is certainly observed for the 12 configurations within the −5 to +5% strain range (**Figure 7**; **Table S10**). Similarly, for the AB configuration with the two Ce3<sup>+</sup> ions in nearest neighbor sites, and experiencing the same chemical environment (CN = 6), no splitting of the f orbitals is expected. However, two peaks are observed in the DOS for −5% compression with a splitting of the order of 0.16 eV. The reason for the splitting is not that the Ce3+s are located in different environments but a hybridization of the polaronic states, which causes a spread of charges across both adjacent sites. That is, the polarons are no longer fully localized in one specific Ce site, rather, the polaronic charge is shared between the two sites in a sort of bonding/anti-bonding configuration (cf. isosurfaces in **Figure S2**). Finally, for the 1C configuration with one nearest and one next-nearest Ce3<sup>+</sup> (CN = 6 and 7, respectively), we find a splitting of the f orbitals of the order of 0.1 eV for −5 and 0% strain, but increases to about 0.2 eV for +5% tensile strain (**Table S10**). Before discussing the changes in the positions of the occupied f states in the O2p-Ce5<sup>d</sup> gap as a consequence of in-plane strain here below, we note that both compressive and tensile strain decrease the O2p-Ce5<sup>d</sup> gap (**Figure 7**; **Table S10**), and that the width of the valence band widens (shrinks) as the lattice is compressed (expanded).

The position of the occupied f states in the O2p-Ce5<sup>d</sup> gap is hereby given as O2p-Ce4<sup>f</sup> , i.e., the difference between the eigenvalue of the lowest occupied f state and the top of the valence band. We note that, as the in-plane strain changes from −5 to +5%, the occupied f states are shifted toward the valence band (**Table S10**), but the amount by which they are shifted depends on the actual positions of the Ce3<sup>+</sup> with respect to the

vacant site. **Figure 8** compares the total DOS of the strained SSV and SSSV with the AB and 12 configurations and shows that under surface compression, the characteristic polaron peaks of the AB configuration lie lower in energy than those of the 12 one, whereas under surface expansion, the situation is reversed. This is in line with the above observed changes in the relative stability of the NN-NN (AB) and NNN-NNN (12) Ce3<sup>+</sup> configurations for both SSV and SSSV vacancy types (cf. **Figures 3**, **4**) under varying strain, i.e., AB configurations are stable under large compressive strain whereas 12 ones are stable otherwise.

# High Vacancy Concentration

In this section we analyze the results for a SSV and a SSSV in the near-surface of CeO<sup>2</sup> (111) under strain with 2 × 2 periodicity (**Figure 9**; **Table S8**), and compare them to those obtained with the 5 × 5 unit cell. In the 2 × 2 unit cell under zero applied stress, the most stable SSV and SSSV configurations have both Ce3<sup>+</sup> ions in NNN sites, one in the outermost cationic layer, and the other in the layer beneath, i.e., both configurations are labeled as 14 (cf. **Figure 2**), in line with previous calculations (Murgida and Ganduglia-Pirovano, 2013). These SSV (14) and the SSSV (14) configurations are by 0.22 and 0.54 eV more stable, respectively, than the corresponding ones with the Ce3<sup>+</sup> ions in NN sites (AB). We note that at the lower vacancy concentration (5 × 5 unit cell) and zero applied stress, the SSV and the SSSV configurations with both Ce3<sup>+</sup> ions in the second coordination shell (12) are by 0.23 and 0.32 eV more stable, respectively, than the corresponding ones with Ce3<sup>+</sup> ions in the first coordination shell (AB). For the SSV and SSSV oxygen vacancies with 2 × 2 periodicity under 0% strain, the 14 >1B >AB and 14 > 1A ≈ 1D > AD ≈ AB stability trends, respectively, are obtained (**Table S8**) (Ganduglia-Pirovano et al., 2009; Murgida and Ganduglia-Pirovano, 2013). As the tensile strain is increased from 0 to +5%, the SSV- and SSSV-14 configurations remain more stable with respect to the corresponding AB ones. For example, for the +5% stretched surface, the SSV- and SSSV-14 configurations are by 0.23 and 0.53 eV more stable, respectively, than the corresponding SSVand SSSV-AB ones (**Figure 9**; **Table S8**). However, in line with the results for the lower vacancy concentration (5 × 5 unit cell), for both vacancy types under compressive strain, an increasing preference for configurations with both Ce3<sup>+</sup> ions in the first coordination shell (AB) is observed. For example, for −5% strain, the SSV-AB configuration is by 0.16 eV more stable than the SSV-14 (**Figure 9**; **Table S8**); for the SSSV, an even larger compressive strain would be needed to observe the crossing between the AB and 14 configurations (cf. **Figure 9**), but the tendency is clear. In summary, for both vacancy concentrations, the ground-state configuration of both vacancy types varies whether tensile or compressive stress is applied, namely, under tensile strain, both Ce3<sup>+</sup> ions do not prefer the first coordination sphere of the vacancy, but under large compressive strain they do.

With respect to the relative stability between the SSV and SSSV when stress is applied and the concentration of vacancies is larger (**Figure 9**), a behavior very similar to that already discussed for the case of low concentration is observed (cf. **Figure S3**, **Figure 5**). In agreement with the literature (Ganduglia-Pirovano

et al., 2009; Murgida and Ganduglia-Pirovano, 2013), we find that under 0% tension, the subsurface vacancy position is by 0.46 eV more stable than the surface one; at the lower vacancy concentration the difference is by a factor of two smaller. Under compressive strain, from 0 to −5%, the SSSV remains more stable than the SSV, but the difference becomes larger as the compression increases, for example, under −5% compression, the SSSV is by 0.82 eV more stable than the SSV. However, under tensile strain, the larger relative stability of the SSSV as compared to the SSV becomes smaller and smaller as the strain increases from 0 to +5%, but an even larger tensile strain would be needed to observe the crossing between the corresponding E<sup>f</sup> vs. strain curves; at the lower vacancy concentration the crossover occurs at about +3% strain (cf. **Figure 6**). For the higher vacancy concentration, we have also evaluated the bond breaking, E<sup>b</sup> , and the relaxation energy, E<sup>r</sup> , contributions to the vacancies formation energies as a function of surface strain (**Figure S3**; **Table S11**), which behave similarly to the case of the lower vacancy concentration discussed above. Both contributions resemble straight lines with a different slope for the SSV and SSSV, and the correlation between E<sup>b</sup> and strain is negative, whereas that between –E<sup>r</sup> and strain is positive. For the −5% compressed surface, both E<sup>b</sup> and E<sup>r</sup> favor the SSSV by 0.11 and 0.71 eV, respectively, and thus, the SSSV is by 0.82 eV more stable than the SSV. However, for the +5% strained surface, E<sup>b</sup> favors the SSV by 0.14 eV and E<sup>r</sup> the SSSV by 0.21 eV, and, as a result, the SSSV is only by 0.07 eV more stable than the SSV.

## Vacancy Clustering

As mentioned above (section Introduction), under zero applied stress, isolated subsurface vacancies are more stable than surface ones, and the interaction between nearest neighbor vacancies at the surface or in the subsurface is repulsive. We here consider strained reduced surfaces with either first-neighbor vacancy dimers at the surface (SurDimer), or in the subsurface (SubDimer), or pairs with a vacancy separation corresponding to that of third-nearest neighbors in the surface oxygen plane (SurDimer-d), or in the subsurface one (SubDimer-d), cf. **Figure 1**. Due to the large number of possible different configurations of the four Ce3<sup>+</sup> created upon formation of a vacancy dimer, we have selected a set that mostly involve combinations of nearest-neighbor (A, B, C, D. . . .) or next-nearest-neighbor (1, 2, 3, 4. . . .) sites to the vacancies (**Tables S3–S6**). **Figure 5** shows the averaged vacancy formation energy of the most stable structures as a function of strain (**Table S7**). As expected (Murgida and Ganduglia-Pirovano, 2013; Han et al., 2018), under zero applied stress, the interaction between first-neighbor surface and subsurface vacancies is repulsive; creating two well-separated surface and subsurface vacancies is energetically more favorable by about 0.1 [1E<sup>f</sup> = 2 E<sup>f</sup> (Surdimer)– 2 E<sup>f</sup> (SSV), cf. Equation 1] and 0.2 eV [1E<sup>f</sup> = 2 E<sup>f</sup> (Subdimer)– 2 E<sup>f</sup> (SSSV)], respectively, than the corresponding pair formation. Tensile strain not only does not help neither the formation of surface nor subsurface vacancy pairs, but makes the preference for having corresponding isolated species even more pronounced, for example, for +5% tensile strain, by about 0.6 and 0.4 eV, respectively. However, upon lattice compression, the SurDimer and the SubDimer become more stable than the corresponding isolated species, for example, for −5% compressive strain, by about 0.4 and 0.2 eV, respectively, with the SubDimer being by about 0.9 eV more stable than the SurDimer.

In order to understand why compression favors vacancy clustering, for the example of the SurDimer, we have analyzed the two contributions to the vacancy dimer formation energy, namely, the bond breaking energy, E<sup>b</sup> , and the energy gained from the structural relaxation in the presence of the vacancies, E<sup>r</sup> , and compared them to those of the SSV (**Figure 6**; **Table S9**). **Figure 6** shows that within the −5 to +5% strain interval, creating an unrelaxed SurDimer is always energetically preferred over having two isolated SSV species, for example by 0.7 [1E<sup>b</sup> = 2 E<sup>b</sup> (Surdimer)– 2 E<sup>b</sup> (SSV)] and 0.6 eV under 0 and −5% strain, respectively (**Table S9**). The reason for this is likely due to the way the four excess electrons are "delocalized" among the nearest-neighbor cations in the unrelaxed structures, i.e., for the SurDimer, the shared cation (B, **Figure 1**) is fully reduced to Ce3+, and the other four (A, C, E, and D), are partially reduced to Ce0.75+, whereas in the case of two isolated SSV, the three nearestneighbor cations of each vacancy would share two excess electrons (3×Ce0.6<sup>ˆ</sup> <sup>+</sup>).

With respect to the energy gained from lattice relaxations, it is always smaller for the SurDimer as compared to two isolated SSV, expectedly, but the difference becomes smaller and smaller as the strain decreases from +5 to −5% (**Table S9**). For instance, under 0% (−5%) strain, the gain amounts to 1.46 (0.96) and 1.86 (1.04) eV/vacancy for the SurDimer and a SSV, respectively. Consequently, under 0% (−5%) strain, 1E<sup>f</sup> = 1E<sup>b</sup> + 1E<sup>r</sup> = −0.7 + 0.8 (−0.6 + 0.16) = +0.1 (−0.44), i.e., it is only when the energy gained from lattice relaxations is considered that the preference for dimer formation under compressive strain (>−2%) can be explained.

We finally note that as for the case of the isolated species discussed above, for which compressive (tensile) strain favors NN (NNN) Ce3<sup>+</sup> configurations, for the dimeric species, a similar tendency is observed, for example, the stable SurDimer and SubDimer configurations under −5% compressive strain have all four Ce3<sup>+</sup> ions in the first cationic coordination shell of the vacancies (**Tables S3**, **S4**).

# CONCLUSION

This work shows the important effect of lattice strain on the relative stability of different types of near-surface oxygen vacancies at the CeO<sup>2</sup> (111) surface, and on their groundstate energy configurations with respect to the localization of the excess charge. We have found three situations of energy crossover: (i) isolated surface vs. sub-surface oxygen vacancy species; (ii) Ce3<sup>+</sup> within the first vs. second coordination shell of the vacancies; (iii) isolated species vs. first-neighbor vacancy dimers. Under 0% strain, isolated subsurface oxygen vacancies with both Ce3<sup>+</sup> ions within the second coordination shell of the vacancies are favored, and the interaction between firstneighbor vacancies is repulsive. However, if a large tensile stress is applied, isolated surface vacancies with both Ce3<sup>+</sup> in the second coordination shell become more stable than the subsurface ones, with both Ce3<sup>+</sup> in next-nearest neighbor sites too. Moreover, if a large compressive strain is applied, isolated subsurface vacancies are more stable than surface ones, as in the case of 0% strain, but for both vacancy types configurations with both Ce3<sup>+</sup> in the first coordination shell of the vacancies are favored over those with both Ce3<sup>+</sup> in the second coordination shell. Finally, under large compressive strain, the formation of first-neighbor vacancy dimers at the surface and in the subsurface, with the four Ce3<sup>+</sup> in nearest-neighbor sites, is more stable than having the two corresponding isolated species, with the subsurface vacancy dimer being more stable than the surface one.

We have further analyzed the effects of lattice strain on the energy cost to create a single vacancy without allowing for lattice relaxations, as well as on the energy gained due to lattice relaxations, and found that the preference for the surface or subsurface sites is the result of the different dependences of these energies on the applied stress. Regarding the Ce3<sup>+</sup> switched preference in favor of the first coordination shell of the vacancies under compressive strain, we have inspected the calculated densities of states and found that the Ce3<sup>+</sup> switched preference is related to the exchange in the relative positions of highest occupied Ce3<sup>+</sup> <sup>f</sup> states in the O2p-Ce5<sup>d</sup> gap region.

Generally speaking, the vacancy structure at ceria surfaces is of importance for technological applications. Clearly, the finding that the relative stabilities of vacancy structures do depend on the applied in-plane strain, particularly in relation to the formation of vacancy aggregates, cannot be ignored when considering that in most applications ceria is employed within a heterostructure or as a supported thin film in which ceria can be strained.

# DATA AVAILABILITY

The raw data supporting the conclusions of this manuscript will be made available by the authors, without undue reservation, to any qualified researcher.

# AUTHOR CONTRIBUTIONS

MVG-P and YG supervised the whole project. Z-KH and LZ performed the calculations. Z-KH, LZ, MVG-P, and YG wrote the paper. All authors participate in the discussions.

# FUNDING

This work has been funded by the National Natural Science Foundation of China (11574340, 21773287) as well as by the US Department of Energy SECA Core Technology Program (DE-NT0006557) and DOE ARPA-E REBELS Program (DE-AR0000502). MVG-P acknowledges support by the MINECO- and MICINN-Spain (CTQ2015-71823-R and RTI2018-101604-B-I00, respectively). The computational resources utilized in this research have been provided by the National Supercomputing Center in Tianjin, the Shanghai Supercomputer Center and the Guangzhou Supercomputer Center, as well as by the National Energy Research Scientific Computing Center, a DOE Office of Science User Facility supported by the Office of Science of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231. This work used the Extreme Science and Engineering Discovery Environment (XSEDE), which is supported by National Science Foundation grant number TG-DMR170045. We thank Gustavo E. Murgida from the CAC-CNEA and CONICET-Argentina for fruitful discussions.

# SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fchem. 2019.00436/full#supplementary-material

# REFERENCES

Allen, J. P., and Watson, G. W. (2014). Occupation matrix control of dand f-electron localisations using DFT+ U. Phys. Chem. Chem. Phys. 16, 21016–21031. doi: 10.1039/C4CP01083C

Andersson, D. A., Simak, S. I., Skorodumova, N. V., Abrikosov, I. A., and Johansson, B. (2006). Optimization of ionic conductivity in doped ceria. Proc. Natl. Acad. Sci. U.S.A. 103, 3518–3521. doi: 10.1073/pnas.0509537103

Arapan, S., Simak, S. I., and Skorodumova, N. V. (2015). Volume-dependent electron localization in ceria.

Phys. Rev. B 91:125108. doi: 10.1103/PhysRevB.91.1 25108


moduli of Ce1−xGdxO2−x/2: a theoretical study. Scr. Mater. 158, 126–130. doi: 10.1016/j.scriptamat.2018.08.034

**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Han, Zhang, Liu, Ganduglia-Pirovano and Gao. This is an openaccess article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# STM Images of Anionic Defects at CeO2(111)—A Theoretical Perspective

Matthew J. Wolf 1†, Christopher W. M. Castleton2,3, Kersti Hermansson<sup>1</sup> and Jolla Kullgren<sup>1</sup> \*

<sup>1</sup> Department of Chemistry–Ångström Laboratory, Uppsala University, Uppsala, Sweden, <sup>2</sup> School of Science and Technology, Nottingham Trent University, Nottingham, United Kingdom, <sup>3</sup> Division of Physics and Mathematics/Natural Science Didactics, Mälardalen University, Västerås, Sweden

### Edited by:

Maria Veronica Ganduglia-Pirovano, Institute of Catalysis and Petrochemistry (ICP), Spain

### Reviewed by:

Joachim Paier, Humboldt-Universität zu Berlin, Germany Niklas Nilius, University of Oldenburg, Germany

> \*Correspondence: Jolla Kullgren jolla.kullgren@kemi.uu.se

### †Present Address:

Matthew J. Wolf, Department of Physics, University of Bath, Claverton Down, United Kingdom

### Specialty section:

This article was submitted to Physical Chemistry and Chemical Physics, a section of the journal Frontiers in Chemistry

> Received: 07 January 2019 Accepted: 18 March 2019 Published: 06 June 2019

### Citation:

Wolf MJ, Castleton CWM, Hermansson K and Kullgren J (2019) STM Images of Anionic Defects at CeO2(111)—A Theoretical Perspective. Front. Chem. 7:212. doi: 10.3389/fchem.2019.00212 We present a theoretically oriented analysis of the appearance and properties of plausible candidates for the anionic defects observed in scanning tunneling microscopy (STM) experiments on CeO2(111). The simulations are based on density functional theory (DFT) and cover oxygen vacancies, fluorine impurities and hydroxyl groups in the surface and sub-surface layers. In the surface layer, all three appear as missing spots in the oxygen sublattice in filled state simulated STM images, but they are distinguishable in empty state images, where surface oxygen vacancies and hydroxyls appear as, respectively, diffuse and sharp bright features at oxygen sites, while fluorine defects appear as triangles of darkened Ce ions. In the sub-surface layer, all three defects present more complex patterns, with different combinations of brightened oxygen ion triangles and/or darkened Ce ion triangles, so we provide image maps to support experimental identification. We also discuss other properties that could be used to distinguish the defects, namely their diffusion rates and distributions.

Keywords: simulated STM images, cerium dioxide (CeO2), density functional theory, anionic defects, reducible oxide

# 1. INTRODUCTION

The functional properties of technologically relevant materials can depend critically on the properties of point defects in the lattice structure, which may be present unavoidably as contaminants or introduced deliberately as dopants. For this reason, much research effort goes into identifying and characterizing such defects. In the bulk, typically one must rely on spectroscopic probes, which provide spatially averaged information, but at the surface one can make use of spatially resolving techniques such as scanning tunneling microscopy (STM) and atomic force microscopy (AFM). Of these latter two methods, STM remains more widely used, despite the fact that its applicability is limited to materials with sufficient conductivity, and atomic resolution STM imaging of surfaces is now fairly routine. However, the interpretation of the results of STM experiments is complicated by the fact that the tunneling current depends not only on the topography of the surface, but also on the local electronic structure. Substantial success has been achieved by combining experimental data with the results of electronic structure calculations, most commonly based on density functional theory (DFT) (see e.g., Setvín et al., 2017 for a particularly relevant review of such studies on oxide surfaces).

The focus of the present contribution is the identification of anionic defects at the (111) surface facet of CeO<sup>2</sup> (ceria) observed using STM. The earliest such experimental studies were carried out on single crystals obtained commercially, and mainly with negative bias voltages, thereby imaging the filled electronic states (Nörenberg and Briggs, 1997, 1998, 1999; Nörenberg, 2002; Fukui et al., 2003; Namai et al., 2003; Esch et al., 2005). The predominant defects observed were dark "depressions" in this sublattice, which formed triangular and extended linear clusters at higher concentrations, and exhibited sites of enhanced brightness around their edges. However, despite the fact that the samples used in all of these studies were from the same commercial source, the results are not entirely consistent. For instance, apparently similar defects were observed to be mobile at room temperature in one study (Namai et al., 2003), but immobile at temperatures up to 673 K in another (Esch et al., 2005), suggesting that the defects observed in the different experiments might not all be the same species, despite their similar appearances (Campbell and Peden, 2005).

A considerably greater number of studies have been performed on thin ceria(111) films grown on metallic substrates, including Rh(111) (Castellarin-Cudia et al., 2004; Chan and Yuhara, 2015), Ru(0001) (Lu et al., 2006; Zhou et al., 2008; Weststrate et al., 2009; Jerratsch et al., 2011; Hasegawa et al., 2014; Shahed et al., 2014), Cu(111) (Szabova et al., 2012; Hu et al., 2015), Pt(111) (Berner and Schierbaum, 2001; Grinter et al., 2010; Luches et al., 2011) and Au(111) (Zhao et al., 2007; Ma et al., 2008). The coverage, thickness and degree of structural order of the ceria films, as well as the resolution with which they were imaged, vary considerably, making direct comparison between them challenging. Depressions and protrusions, in isolation and in the form of clusters, are commonly observed, both in empty and filled state images.

In **Table 1**, we have attempted to summarize the experimental appearances and assignments of the features observed in filled and empty state STM images of CeO2(111), along with the experimental conditions (temperature and bias voltage), and substrate where applicable. We note that the results of different studies are not entirely consistent, with numerous different features being assigned to oxygen vacancies, with consideration of other possibilities being limited to an assignment of bright protrusions to hydroxyl groups in Shahed et al. (2014).



Bias voltages have been rounded off to the nearest 0.5 V and room temperature is reported as 298 K. <sup>a</sup>Ru(0001) means a ceria(111) film on a Ru(0001) substrate, for example.

With the above discussion in mind, the principal purpose of the present paper is to offer a consistent set of simulated STM images of three particularly plausible candidate defects at anionic sites on CeO2(111), drawing attention to their similarities and differences, which we believe will aid in the interpretation of STM experiments on this system. We will also discuss other properties which could aid in distinguishing between them, making use of theoretical data available in the literature where available.

The three defects that we consider are oxygen vacancies, fluoride ions, and hydroxyl groups (denoted respectively OV, FI, and OH in the remainder of this paper). Our consideration of the OV defects is natural, given the well known ability and tendency of ceria to release oxygen under oxygen lean conditions, a quality which plays a central role in many of its technological applications (Trovarelli, 2013).

Surface OHs can arise due to the presence of H2O, which is a ubiquitous molecule even under ultra high vacuum conditions, and there are numerous reports that it reacts with OVs to form hydroxide ions embedded in the surface (Fronzi et al., 2009; Molinari et al., 2012; Mullins et al., 2012).

Finally, our consideration of fluoride ions is primarily due to the observation in Pieper et al. (2012) that fluorine was present in large quantities in single crystalline samples from the same source as those used in Nörenberg and Briggs (1997), Nörenberg and Briggs (1998), Nörenberg and Briggs (1999), Nörenberg (2002), Namai et al. (2003), Fukui et al. (2003), and Esch et al. (2005). Furthermore, it was observed in Zarraga-Colina et al. (2004) and Zarraga-Colina et al. (2005) that if CeO<sup>2</sup> films are grown on CaF, F diffuses into the CeO<sup>2</sup> upon annealing, replacing some of the O ions within the lattice. Taken together, these observations suggest that if CeO<sup>2</sup> is exposed to a source of F, it is likely to become contaminated, and such contamination may be difficult to avoid entirely.

Previously, we concluded that, at least in filled state STM images, fluorine impurities and hydroxyl groups should be difficult to distinguish from oxygen vacancies at CeO2(111) based on their appearances alone (Kullgren et al., 2014), although OVs and FIs should be distinguishable based on their mobilities, relative concentrations in the surface and sub-surface layers, and tendency to form clusters in the surface layer. In this paper, we expand upon that study, providing a more detailed examination and comparison of the STM appearances of these three oxygen sub-lattice defects (OVs, FIs, and OHs), as well as discussing the results of related studies by other authors. We present both filled and empty state simulated STM images of the three defect species, both when located in the surface layer itself (section 3.2) and when located in the first oxygen sub-surface layer (section 3.3), and relate those images both to the local topography and to the electronic structure. In sections 3.4, 3.5 and 3.6 we discuss other distinguishing characteristics of the three defects, namely the positions of the f electron levels of the associated Ce3<sup>+</sup> ions, the defects' mobilities and their distributions. We then conclude in section 4.

# 2. COMPUTATIONAL DETAILS

We carried out Kohn–Sham density functional theory (DFT) calculations using the projector augmented wave (PAW) formalism (Blöchl, 1994), as implemented in the Vienna Ab initio Simulation Package (Kresse and Hafner, 1993; Kresse and Furthmüller, 1996a,b; Kresse and Joubert, 1999) (VASP), version 5.3.5. We used the Perdew–Burke–Ernzerhof (PBE) functional (Perdew et al., 1996, 1997), augmented with a Hubbard-like (effective) U term in the simplified, rotationally invariant form introduced by Dudarev et al. (1998); Himmetoglu et al. (2014). The projectors that were used in the calculations of the occupation matrix were those associated with the PAW potentials, as is done by default in VASP, and we used a value of U = 5 eV (Nolan et al., 2005; Andersson et al., 2007; Castleton et al., 2007). The core electrons were described with the "standard" PBE derived PAW potentials distributed with VASP, in which the cores comprise the [Kr]4d electrons of Cerium and the 1s electrons of oxygen and fluorine. Aspherical contributions from the gradient corrections inside the PAW spheres were taken into account. The Kohn–Sham states for the valence electrons were expanded over the set of all plane waves of kinetic energy less than a cut-off value of 400 eV. Energies were converged to 1 × 10−<sup>5</sup> eV, and forces to 1 × 10−<sup>2</sup> eV/Å. We also used the Occupation Matrix Control (OMC) method in the implementation of B. Dorado and co-workers (Dorado et al., 2009) to avoid the problems associated with the existence of electronic self-consistent "local minima," or "meta-stable" states, corresponding to different shapes and/or orientations of the f orbital that is occupied on a given Ce3<sup>+</sup> ion.

For the structural model of CeO2(111), we used a periodic slab exposing a p(5 × 5) supercell of the (111) surface facet on both sides, with a thickness of three O–Ce–O triple layers; this leads to a formula for the stoichiometric slab of Ce75O150. A vacuum gap equivalent to the height of 5 bulk triple layers, which is more than 15 Å, was also used. Due to the large supercell used in our calculations, we found that it was sufficient to restrict sampling of the Brillouin zone to the Ŵ point.

Densities of states were produced by smearing the oneelectron spectrum using a normalized Gaussian function with a width of 0.1 eV. Simulated STM images were produced under the Tersoff–Hamann approximation (Tersoff and Hamann, 1983, 1985), that is to say, directly from the local density of states (LDOS) obtained from partial charge densities calculated with VASP. The charge density was smeared with a Gaussian function with a width of 0.1 Å, and then linearly interpolated onto a fine 2D grid parallel to, and at a height 2.8 Å above, the upper surface of the slab; we found that increasing the distance any further introduced considerable numerical noise into the images. The partial charge densities were generated by summing the charge densities of all states within −3.0 eV or +3.0 eV of the Fermi level, assumed to be in the middle of the gap, for filled and empty state images, respectively. We note that the simulation protocol used here, which is based on the PBE+U functional, underestimates the band gaps somewhat, yielding O-2p → Ce-4f and O-2p → Ce-5d gaps of 2.25 and 5.20 eV, respectively. However, we find the simulated STM images to be rather insensitive to the precise value of the limit for the summation of states (and therefore the precise positions of the Fermi level and the band edges) with in general, although we present additional images when this is not the case.

# 3. RESULTS

# 3.1. Preliminary Considerations

### 3.1.1. Imaging the Stoichiometric Surface

Simulated filled and empty state images of stoichiometric CeO2(111) are shown in **Figure 1**.

The (filled) valence band states of ceria derive (primarily) from 2p orbitals of the O2<sup>−</sup> ions, with a smaller contribution from 4f orbitals of the Ce4<sup>+</sup> ions in the layer beneath them, so it is natural that the topmost surface O2<sup>−</sup> ions are what is imaged at negative bias voltages.

Regarding empty state images, some authors have suggested that they are due to electrons tunneling into conduction band states derived (primarily) from Ce 5d orbitals (Esch et al., 2005), rather than states derived (primarily) from Ce 4f orbitals, due to the more compact nature of the latter. However, the available experimental spectral data, summarized in Castleton et al. (2007), place the Ce 5d and 4f band edges at approximately 6 eV and 3 eV, respectively, above the valence band maximum (Wuilloud et al., 1984; Marabelli and Wachter, 1987; Pfau and Schierbaum, 1994; Mullins et al., 1998). It seems reasonable to assume that the Fermi level is below the unoccupied Ce 4f states, as it must be unless the sample is very heavily reduced indeed. We can therefore surmise that tunneling into the Ce 5d states is rather unlikely unless very high positive biases are used.

Given the above discussion, we therefore assume that in empty state imaging, electrons tunnel into states that derive mainly

The location of the ions is indicated by superimposing the STM images over a schematic top-view of the surface with white and black circles corresponding to oxygen and cerium ions, respectively. (Lower) Density of states (DOS) plots for the stoichiometric slab. Total DOS shown (gray) with dashed lines and projections onto a surface oxygen ion and a surface cerium ion are shown by white and black filled curves, respectively.

from Ce 4f orbitals. Note that the surface oxygen ions also contribute to the unoccupied states, and since they are in the layer above the Ce ions, they are also partly visible at positive bias voltages.

# 3.1.2. Locations of Ce3<sup>+</sup> Ions

All three defects introduce excess electrons into the lattice; in ceria, these electrons localize on individual Ce sites, reducing their charge states from +4 to +3, forming polarons. In order to maintain charge neutrality, OVs introduce two such Ce3<sup>+</sup> ions, while FIs and OHs introduce one each.

The energetically preferential location of these Ce3<sup>+</sup> ions in the vicinity of OVs has been the subject of much debate in the literature. The current consensus, based on theoretical calculations, is that the Ce3<sup>+</sup> are located preferentially at nextnearest neighbor (NNN) sites to OVs (Murgida and Ganduglia-Pirovano, 2013; Sutton et al., 2015). In contrast, the single Ce3<sup>+</sup> associated with an FI is preferentially located at a nearest neighbor (NN) site (Kullgren et al., 2014; Wolf et al., 2017), while for a surface OH, the two locations have essentially identical energies (Fernández-Torre et al., 2014; Kullgren et al., 2014; Wolf et al., 2017).

In principle, the Ce3<sup>+</sup> ions (polarons) which charge compensate the substitional ions should break the threefold rotational symmetry of the lattice around the latter, but at higher temperatures, the electrons can hop between equivalent sites, so that the symmetry is restored (on average) on the time-scale required to obtain an STM image; indeed, this would appear to be the case for most of the experimental images in the literature, in which the rotational symmetry is preserved. For this reason, we present both images with the Ce3<sup>+</sup> at specific locations, and images that have been rotationally averaged around the substitional ion.

# 3.2. STM Images of Defects in the Surface Layer

### 3.2.1. Filled State Images

The simulated filled state STM images of the three surface anionic defects, with the associated Ce3<sup>+</sup> ions at either nearest neighbor (NN) or next nearest neighbor (NNN) locations, are shown in the bottom row of panels of **Figure 2**.

In the filled state images, the main feature of all three defects is a dark "depression" at the lattice site of the defect center. For the surface OV, this is simply due to the absence of the ion and the associated electron density. The reason for the reduced tunneling current in the case of the fluoride or hydroxide ion in the filled state image can be understood by examining the partial density of states projected onto the defects (see **Figure 3**). Both the OH and the FI make their main contributions to the local density of occupied states at lower energies than the surrounding surface oxygen ions. Hence, despite the fact that both defects protrude from the surface, they are essentially invisible unless the bias voltage is large enough to allow tunneling from these states. At more negative bias voltages, both OH and FI begin to appear, with the OH being more visible than the FI for a given voltage, which in turn would be more visible than an OV. Thus, in principle, very

FIGURE 2 | Simulated filled and empty state images of the three surface anionic defects, with the associated Ce3<sup>+</sup> ions at either nearest neighbor (NN) or next nearest neighbor (NNN) lattice sites. The precise locations of the Ce3<sup>+</sup> ions are shown in the schematics, for comparison with the non-rotationally averaged images.

sensitive STM measurements should be able to distinguish them by their relative apparent depth profile for a fixed bias voltage which probes states deep in the valence band.

Moving on to more subtle features of the STM images, we note that some of the oxygen ions in the vicinity of the central depressions are brighter than others. These features are correlated with the positions of the Ce3<sup>+</sup> ions, and is easy to understand based on the fact that Ce3<sup>+</sup> ions are larger than Ce4<sup>+</sup> ions, and they are negatively charged with respect to the lattice. Both of these effects will cause the negatively charged O2<sup>−</sup> ions to be repelled out of the surface. However, in some publications, an enhanced brightness was reported between certain pairs of the surrounding surface O ions (see e.g., Esch et al., 2005; Grinter et al., 2010). Such features have been explained as being due to a relaxation of pairs of surface O ions toward one another Esch et al. (2005). We found that we were only able to reproduce such effects using a p(2×2) surface supercell, as was done in Esch et al. (2005), which corresponds to a much higher defect concentration than that observed in the corresponding experiments. When we use larger p(4 × 4) or p(5 × 5) supercells, corresponding more closely to the experimental concentrations, we find a much weaker pairing, and furthermore it is rotated by 120◦ , i.e., it is between ions which are symmetrically inequivalent to those for which pairing was observed in the experiments in Esch et al. (2005). However, we note that our simulations are performed under the Tersoff–Hamann approximation (as were those in Esch et al., 2005), and thus (a) do not fully treat orbital directionality and (b) do not include the effects of the tip. Given the very low tunneling current of the measurements, it is possible that there are angular effects and/or non-negligible tip–surface interactions, which may be responsible for these features; in any case their origin remains unexplained. Furthermore, the pairing is not observed in all experiments (see Shahed et al., 2011 for example), which could be due to weaker tip–surface interactions.

### 3.2.2. Empty State Images

The empty state images of the three surface defects are given in the top row of **Figure 2**. At positive bias, the OV appears bright compared to its surroundings, due to a large, diffuse state centered on the vacancy, in agreement with images in Jerratsch et al. (2011). At +3.0 V bias the diffuse state spans the six closest cerium ions at the surface, while at a lower bias of +2.0 V, this diffuse state is transformed into a single or triple protrusion of Ce4<sup>+</sup> ions neighboring the OV, as shown in **Figure 4**. The simulated empty state image of OH at positive bias voltages is also bright. Here, we find a rather localized feature at the OH, as compared to the more diffuse feature of the OV in the same bias voltage range. In contrast to the OV and OH, the FI is actually somewhat dark in the empty state image, which would make it more readily distinguishable from the other two defects than they would be from each other. In agreement with Jerratsch et al. (2011) we also find that Ce3<sup>+</sup> ions appear dark due to their empty states being shifted toward higher energies.

# 3.3. STM Images of Defects in the Sub-Surface Layer

### 3.3.1. Filled State Images

The filled state STM appearances of the defects in the first oxygen sub-surface (third atomic) layer are shown in the bottom row of **Figure 5**. In contrast to the surface defects, their appearance is due primarily to the effect the defects have on the surface layer ions above them, rather than imaging the defect site itself. Hence, the apparent topography is a closer reflection of the actual surface topography (i.e., the relative heights of the surface ions).

Triangular protrusions in experimental filled state images have been ascribed to sub-surface oxygen vacancies (Esch et al., 2005; Grinter et al., 2010; Jerratsch et al., 2011). Our simulated images of a sub-surface OV do produce such a protrusion, although we also find that it should be accompanied by a significant depression of the three closest surface oxygen ions, which relax into the surface due to the presence of the vacancy beneath them. The magnitude of this effect depends upon the locations of the Ce3<sup>+</sup> ions; in the lowest energy NNN configuration, the ions relax downwards by 0.19 Å, while with NN localization, this is reduced to 0.09 Å.

For the sub-surface FI, the main effect for NN localization is the outward relaxation by 0.05 Å. This is also seen if the electrons are at NNN positions, but then it is accompanied by a slight inward relaxation of the three nearest surface O ions, by 0.04 Å. Both of these effects are similar for the OV, but more subtle due to the smaller degree of relaxation.

The sub-surface OH is somewhat different from the other two defects. The triple protrusion is again present, and the associated displacement of the ions is of similar magnitude to the surface

FI with NN localization, namely 0.05 Å. However, the more striking feature is the apparent protrusion of the three nearest neighbor oxygen ions, which are pushed out of the surface by the OH by 0.05 Å. This produces a small, bright triangle with the opposite orientation relative to the underlying lattice, as compared to the sub-surface oxygen vacancy. A careful analysis of the partial density of states suggests that the reason why these ions appear brighter, despite being at a similar height, is that the states associated with them are pushed up slightly in energy, likely due to the presence of the proton between them.

# 3.3.2. Empty State Images

The empty state images, shown in the top row of **Figure 5**, all follow similar patterns. Again, the Ce3<sup>+</sup> ions appear as dark depressions in all cases, due to the unoccupied states being pushed upwards in energy. For the OV, with NNN localization, we also see a triangular protrusion associated with the Ce4<sup>+</sup> ions being repelled upward by the defect. We note that this pattern was suggested to be the fingerprint of a surface OV in the combined experimental and theoretical work of Jerratsch et al. (2011). While their observations in empty state images were matched to a single protrusion at negative bias, identification based solely on empty state imaging could lead to mis-interpreting sub-surface OVs as surface OVs, especially at low positive bias where the two appear very much alike (see **Figure 4**). There is a similar, but more subtle protrusion for the FI.

FIGURE 5 | Simulated filled and empty state images of the three sub-surface anionic defects, with the associated Ce3<sup>+</sup> ions at either nearest neighbor (NN) or next nearest neighbor (NNN) lattice sites. The precise locations of the Ce3<sup>+</sup> ions are shown in the schematics, for comparison with the non-rotationally averaged images.

# 3.4. f-Electron Spectral Features

Scanning tunneling spectroscopy (STS) has been performed on CeO2(111) thin films to determine the energy levels of localized f electrons on Ce3<sup>+</sup> ions located near defects. This provides another possible way to distinguish between defects that share the same gross appearance. Therefore, we have plotted in **Figure 6** the eigenvalues of the f-electron states associated with all of the defects studied herein. They fall in a range from 0.72 to 0.86 eV with respect to the valence band edge. In the combined experimental and theoretical work by Jerratsch et al. (2011) such spectral features were used to discriminate between defects with different Ce3<sup>+</sup> localization patterns.

# 3.5. Diffusion Rates

The diffusion barrier for OVs at CeO2(111) has been calculated by a number of authors (Plata et al., 2013; Su et al., 2016). The values are generally quite small, i.e., below 0.5 eV, which is consistent with the experimentally observed high ionic conductivity of the material; indeed, this is a property of the material which is exploited in many of its technological applications.

In Kullgren et al. (2014), a diffusion mechanism of the FI was considered which involves it moving out of the lattice position to a site on top of a neighboring Ce4<sup>+</sup> ion; with the OV thus formed moving simultaneously to a sub-surface position. The barrier height was calculated to be 1.47 eV, which we re-calculate using our current computational set-up to be 1.42 eV.

In Fernández-Torre et al. (2014), a diffusion barrier was presented for surface OH groups. The mechanism involves the transfer of the proton between a surface and sub-surface oxygen site, with a reported activation energy of 1.8 eV. Here, we considered an alternative mechanism similar to that of FI diffusion, in which the entire OH group moves up on to a surface Ce4<sup>+</sup> ion. We find the barrier height for such a process to be 1.50 eV, i.e., 0.3 eV lower than the barrier reported in Fernández-Torre et al. (2014), and much more similar to the barrier for FI diffusion.

We used these values to calculate diffusion rates from an Arrhenius expression at 77 K, 300 K and 600 K, along with a

prefactor of 1 × 1013s −1 . These values, are collected in **Table 2**. Furthermore, given that the diffusion path for the latter two defects considered here requires the concerted motion of both the defects themselves, and a sub-surface oxygen ion, the actual rate is likely to be lower than that calculated here. We therefore conclude that OVs are too mobile to be imaged by STM, for which a single scan takes on the order of minutes, unless experiments are conducted at temperatures significantly lower than room temperature. FIs and OHs on the other hand, should be immobile for sufficiently long for them to be imaged within the full range of temperatures used in experiments reported in the literature.

We note that it has been demonstrated that migration of charged defects on the ceria surface could proceed via a diabatic channel, in which case the activation energy barriers may be drastically increased, by as much as a factor of three (Lustemberg et al., 2016). However, this only applies to migrating species that change their charge state during the transition. For the defects considered here the migrating species remain in the same charge state in the initial, final and transition states.

# 3.6. Defect Distributions and Interactions

For non-interacting defects of a given species, the proportion of surface to sub-surface defects at thermal equilibrium can be determined by evaluating an appropriate Boltzmann factor. The energy difference between an isolated surface and sub-surface defect of the same type is +0.12 eV for the OV, −0.48 eV for an FI and −1.45 eV for an OH (with a negative value indicating greater stability in the surface layer), according to our calculations. Using these values to calculate the Boltzmann factors at 600 K (within the range of temperatures used in Nörenberg and Briggs, 1999; Nörenberg, 2002; Namai et al., 2003; Esch et al., 2005) suggests that the number of surface OVs for each sub-surface OV is at most 1 × 10−<sup>1</sup> , for FIs at least 1 × 10<sup>4</sup> and at least 1 × 10<sup>12</sup> for OHs. These values, along with analogous ones calculated for T = 77 K and T = 300 K, are collected in **Table 3**, and can be summarized by saying that under most conditions one would expect to see OVs in the sub-surface layer, but FIs and OHs in the surface layer.

TABLE 2 | Defect diffusion rates of isolated defects (in s−<sup>1</sup> ) from the Arrhenius equation, using a pre-factor of 1 × 10<sup>13</sup> s −1 .


TABLE 3 | Relative proportions of isolated surface to sub-surface defects of a given species from Boltzmann factors.


However, at higher defect concentrations, interactions between defects will necessarily become important; indeed, one of the enduring points of interest regarding the defects observed in experimental images is their distribution, and in particular, their tendency to form linear and triangular clusters at the surface. DFT simulations have shown that surface OVs do not exhibit any tendency toward NN clustering (Conesa, 2009; Zhang et al., 2009; Sutton et al., 2015), based on comparisons of total energies. On the other hand, both surface FIs (Kullgren et al., 2014) and OHs have been calculated to be stable as constituents of NN dimers (Fernández-Torre et al., 2014) although the binding energies are small (< 0.1 eV/dimer).

Furthermore, combinations of surface OVs, along with subsurface OVs or sub-surface OHs have also been considered in the DFT study in Wu and Gong (2016), although all of those combinations were shown subsequently to be unstable with respect to decomposition into their isolated constituents (Wolf et al., 2016).

The above conclusions are based solely on differences in total energies, but more recently, the importance of configurational entropy in determining distributions of OVs and FIs has been studied. In Kullgren et al. (2017), DFT-based Monte Carlo simulations showed that OHs have far less tendency to form NN clusters than FIs, and that OHs instead tend to form clusters in which they are located at next-nearest-neighbor sites. For FIs, large and compact clusters are most abundant at 300 K while at 600 K straight and hooked linear clusters are abundant.

DFT based Monte Carlo calculations have also been performed including both surface and sub-surface OVs at a variety of temperatures and levels of reduction (Han et al., 2018). At lower concentrations (10% and 15%) and/or temperature (80 K) the number of NN clusters at the surface was found to be very small, although the numerous sub-surface vacancies do exhibit p(2×2) ordering, consistent with the interpretation of low temperature AFM experiments (Torbrügge et al., 2007). Meanwhile at higher temperatures and levels of reduction, numerous linear NN clusters were observed.

However, standard Monte Carlo simulations such as those described above do not take into account kinetic effects, which must be significant due to the high mobility of OVs. Indeed, the surface OV clusters are reported by Han et al. to be abundant only in simulations performed at temperatures at which OVs are expected to be mobile on the time-scale of STM experiments (see **Table 2**). Although it is conceivable that OVs which comprise parts of clusters could be significantly less mobile than isolated OVs, calculations that address this possibility have not been reported.

# 4. CONCLUSION

STM images of the ceria (111) surface exhibit a number of point defects in an otherwise regular hexagonal pattern. These features are commonly assigned to surface oxygen vacancies, although alternative interpretations have been proposed. In this contribution, we have used Density Functional Theory (DFT) with the PBE+U functional to examine three likely candidates for anionic defects on CeO2(111), namely oxygen vacancies (OV), fluorine impurities (FI) and hydroxyl groups (OH), in both the surface and the first sub-surface layers. We have presented and analyzed a consistent set of simulated STM images for them under the Tersoff–Hamann approximation.

Based on our results we suggest that missing spots in the filled state images at the surface oxygen lattice could be caused by any of the three defects considered here. However, the corresponding empty state images of the three defects are rather different to one another. The clean surface in this case shows spots at Ce sites, and against this background, surface OVs appear as large, bright features centered at oxygen sites, and OHs give smaller but more intense features also at oxygen sites. The FI signature is also centered on an oxygen site, but consists of a triangle of darkened Ce ions surrounding it, relative to the background lattice.

# REFERENCES


The appearance of the sub-surface defects in filled state images is mainly a reflection of topographical changes at the surface (relaxation of ions) induced by those defects. In filled state images, OVs produce triangular protrusions of next nearest neighbor surface oxygen ions similar to those observed in some experiments (cf. Esch et al., 2005; Grinter et al., 2010), but they also produce a rather striking triangular depression inside the protruding triangle, to an extent that may in fact render it difficult to distinguish from a triangular surface defect cluster of either OV, OH or FI. The FI reproduces both of these features, but the contrast variations are more subtle, while the OH produces only protrusions. In the empty state images we note that Ce4<sup>+</sup> which are nearest neighbors to a sub-surface OV appear bright. A similar but much more subtle effect is also seen for sub-surface FI.

In general, our simulations suggest that missing spots in the empty state images, at the cerium lattice, can be caused by Ce3<sup>+</sup> that are not nearest or next nearest neighbor to a surface vacancy, or nearest neighbor to a surface hydroxide.

Regarding their physical properties other than appearances, calculations indicate that OVs are most stable at the sub-surface and that they diffuse rapidly unless the temperature is kept much below room temperature. OVs do not have a strong energetic tendency to cluster, although DFT-based Monte Carlo simulations suggest that, given sufficient vacancy concentration and temperature, clusters can emerge. However, their amounts seem to be appreciable only when the temperature is too high for them to be imaged. In contrast, both FI and OH are quite stable at the surface and diffuse slowly. FI's show a tendency for clustering in the surface layer over a wide range of temperatures and concentrations, with OH representing an intermediate case.

# AUTHOR CONTRIBUTIONS

MJW and JK performed the calculations and prepared the first version of the manuscript. All authors were involved in the planning of the work, the analysis and interpretation of the results and in the finalization of the text.

# ACKNOWLEDGMENTS

This work was supported financially by the Swedish Research Council (VR), the Swedish National Strategic e-Science programme eSSENCE, and Åforsk. Some of the calculations were performed on resources provided by the Swedish National Infrastructure for Computing (SNIC) at UPPMAX and NSC.

Castellarin-Cudia, C., Surnev, S., Schneider, G., Podlucky, R., Ramsey, M. G., and Netzer, F. P. (2004). Strain-induced formation of arrays of catalytically active sites at the metal–oxide interface. Surf. Sci. 554, L120–L126. doi: 10.1016/j.susc.2004.01.059

Blöchl, P. E. (1994). Projector augmented-wave method. Phys. Rev. B 50:17953. doi: 10.1103/PhysRevB.50.17953

Campbell, C. T., and Peden, C. H. F. (2005). Oxygen vacancies and catalysis on ceria surfaces. Science 309, 713–714. doi: 10.1126/science.1113955


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Wolf, Castleton, Hermansson and Kullgren. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Multiscale Modeling of Agglomerated Ceria Nanoparticles: Interface Stability and Oxygen Vacancy Formation

Byung-Hyun Kim1,2, Jolla Kullgren<sup>1</sup> \*, Matthew J. Wolf <sup>1</sup> , Kersti Hermansson<sup>1</sup> \* and Peter Broqvist <sup>1</sup> \*

<sup>1</sup> Department of Chemistry—Ångström Laboratory, Uppsala University, Uppsala, Sweden, <sup>2</sup> Platform Technology Laboratory, Korea Institute of Energy Research, Daejeon, South Korea

### Edited by:

Maria Veronica Ganduglia-Pirovano, Institute of Catalysis and Petrochemistry (ICP), Spain

### Reviewed by:

Michael Nolan, University College Cork, Ireland Sergey M. Kozlov, King Abdullah University of Science and Technology, Saudi Arabia

### \*Correspondence:

Jolla Kullgren jolla.kullgren@kemi.uu.se Kersti Hermansson kersti@kemi.uu.se Peter Broqvist peter.broqvist@kemi.uu.se

### Specialty section:

This article was submitted to Physical Chemistry and Chemical Physics, a section of the journal Frontiers in Chemistry

> Received: 09 January 2019 Accepted: 18 March 2019 Published: 22 May 2019

### Citation:

Kim B-H, Kullgren J, Wolf MJ, Hermansson K and Broqvist P (2019) Multiscale Modeling of Agglomerated Ceria Nanoparticles: Interface Stability and Oxygen Vacancy Formation. Front. Chem. 7:203. doi: 10.3389/fchem.2019.00203 The interface formation and its effect on redox processes in agglomerated ceria nanoparticles (NPs) have been investigated using a multiscale simulation approach with standard density functional theory (DFT), the self-consistent-charge density functional tight binding (SCC-DFTB) method, and a DFT-parameterized reactive force-field (ReaxFF). In particular, we have modeled Ce40O<sup>80</sup> NP pairs, using SCC-DFTB and DFT, and longer chains and networks formed by Ce40O<sup>80</sup> or Ce132O<sup>264</sup> NPs, using ReaxFF molecular dynamics simulations. We find that the most stable {111}/{111} interface structure is coherent whereas the stable {100}/{100} structures can be either coherent or incoherent. The formation of {111}/{111} interfaces is found to have only a very small effect on the oxygen vacancy formation energy, Evac. The opposite holds true for {100}/{100} interfaces, which exhibit significantly lower Evac values than the bare surfaces, despite the fact that the interface formation eliminates reactive {100} facets. Our results pave the way for an increased understanding of ceria NP agglomeration.

Keywords: multiscale modeling, density functional theory, self-consistent charge density functional tight binding, reducible oxides, cerium dioxide, nanoparticles, agglomeration

# INTRODUCTION

Synthesized ceria samples often appear in the form of agglomerates of nanoparticles (NPs) (Wang and Feng, 2003; Mai et al., 2005; Du et al., 2007; Lin et al., 2012; Liu et al., 2015; Schlick et al., 2016; Lykaki et al., 2017). In most cases, agglomeration degrades the total catalytic performance due to the elimination of reactive surfaces (Li et al., 2003; de Diego et al., 2005). However, in some cases, agglomeration has been found to enhance the catalytic performance due to favorable properties of the interface itself, e.g., by facilitating the formation of critical defects (Maillard et al., 2005). For ceria, only a small fraction of the numerous published studies in recent years have scrutinized the nature of native interfaces. On the experimental side, Wang and Feng (2003) synthesized ceria NPs of truncated octahedral shape and of sizes of 3–10 nm, and noted that the agglomeration of their ceria NPs could be correlated with the disappearance of accessible {100} facets during growth. The resulting agglomerates consisted of NPs in the range of ∼1µm in diameter and with each consecutive NP pair forming a coherent interface, i.e., perfectly matching their lattices across the interface by sharing a common {111} facet. Reactivity was not discussed in that study, but in a series of papers, Hojo et al. (2010) and Feng et al. (2012, 2016) reported that the atomic structure and local oxygen dynamics are modified by the formation of grain boundaries in thin-films of ceria. They also concluded that grain boundaries can be stabilized by the presence of oxygen vacancies.

On the theoretical side, Sayle and co-workers have used forcefield simulations to investigate structure-property relationships for ceria nanostructures and have for example investigated nanocubes, nanorods, nanochains, and mesoporous structures (Sayle et al., 2005, 2011; Jeyaranjan et al., 2018). In one of their studies, ceria nanochains were generated via oriented attachment (OA) and used as model systems to explain the mechanism for ceria nanorod formation (Sayle et al., 2011). Sk et al. (2014) performed density functional theory (DFT) calculations for selfassembled one and two dimensional nanoceria networks. Thus, they studied nanowires and nanogrids built from corner-sharing Ce ions or from sharing of oxygen terminated {100} facets; the latter requiring an energy cost to reduce one side of the {100} facets of the NPs prior to agglomeration. For future reference, we note that the periodic boundary conditions used in Sk et al. (2014) induce constraints on the possible NP interfaces that can be modeled. To the best of our knowledge, detailed theoretical investigations at the atomic scale of aperiodic ceria NP interfaces and their redox activity are lacking in the scientific literature. The present study primarily deals with the interface between two ceria NPs and the specific oxygen chemistry that may accompany it.

In this work, we will show that stable ceria-ceria interface structures formed between agglomerated ceria NPs display oxygen redox properties that are quite different from those of either the bare surface or the bulk material, and that this is true both for the "rim" region and the "internal" region of the interface, and both for coherent and incoherent interfaces. Here the term "rim" refers to interface regions where the constituent atoms are exposed while the term "internal" refers to regions where they are not. Our model systems are ceria NP pairs, with an outlook toward chains at the end of the article.

Our method is a sequential multiscale simulation approach linking three levels of theory along the multiscale ladder, namely Hubbard augmented DFT (DFT+U) (Perdew et al., 1996; Dudarev et al., 1998), the self-consistent-charge density functional tight-binding approach (SCC-DFTB+U) (Elstner et al., 1998; Aradi et al., 2007; Kullgren et al., 2017), and finally, for stoichiometric NPs, also the reactive force-field (ReaxFF) method (van Duin et al., 2001; Broqvist et al., 2015). The SCC-DFTB method is an approximate DFT method, which combines an electronic structure description at the minimal basis-set level with a Coulombic expression for long-range interactions. The charges are of Mulliken type and are calculated self-consistently (SCC) in the simulation depending on the elements and the geometric structure. The ReaxFF method, finally, is an advanced bond-order-dependent force-field method, which allows for the description of dissociation and formation of chemical bonds in various kind of compounds.

Moving from DFT to SCC-DFTB to ReaxFF implies a scheme taking us from fine-grained toward more coarse-grained materials models, namely from electrons to atoms in the prevailing MODA language's model entities (Anne de Baas, 2018), as promoted within the European Materials Modeling Council (EMMC, 2019). We use a consistent parameterization of the SCC-DFTB and ReaxFF interaction models which are both trained against DFT+U data. These parameterization efforts were described in detail in previous publications (Broqvist et al., 2015; Kullgren et al., 2017). In the present paper, however, we mostly move in the opposite direction, from SCC-DFTB to DFT, using the former method to screen many NP assembly structures, the most stable of which are subsequently scrutinized using DFT. Finally we make use of the ReaxFF model for ceria to perform long MD simulations during the initial stages of nanoceria agglomeration. In the following, we will often omit the U term in our notation for DFT and SCC-DFTB not to burden the text, but it is in fact present in all our electronic calculations of (partially) reduced ceria systems.

We will present the stable ceria interface structures found and their effects on the oxygen vacancy formation energy, which we will use as a key descriptor of oxygen chemistry. We find that while agglomeration of course lowers the NPs' surface areas, the interface-specific defects might lead to an activation of agglomerated NPs toward redox processes. Finally, we also discuss how to extend our findings into larger interface systems using the ReaxFF method. Here we will briefly address the mechanism of NP agglomeration and relate to the concept of OA which has been amply discussed in the literature (Si et al., 2006; Du et al., 2007; Lin et al., 2012; Raju et al., 2014; Zhang and Banfield, 2014), and where nano-crystallites collide and interact favorably, e.g., by sharing some common crystallographic orientations, resulting in the formation of interfaces between the NPs.

# INTERFACE CONSTRUCTION PROTOCOL AND COMPUTATIONAL METHODS

In this work, as mentioned, we use a computational protocol based on a combination of electronic structure methods, namely the SCC-DFTB and DFT methods, to study ceria NP interface structures. SCC-DFTB is roughly 2 orders of magnitude faster than standard (semilocal, i.e., "GGA"-type) DFT calculations. With a proper parameterization, the accuracy with respect to the training set will be close to that of the underlying DFT method. This agreement will be reported in the section Consistency Between the Methods That We Have Used. The considerable speedup compared to DFT makes SCC-DFTB a valuable tool for the investigaton of a large number of interface structures, which is exactly the way we have used it here.

The SCC-DFTB calculations were performed using the DFTB+ code (Elstner et al., 1998; Aradi et al., 2007). The parameters describing the Slater-Koster tables and repulsive potentials for ceria have been generated and thoroughly tested previously (Kullgren et al., 2017). DFT calculations in the implementation with plane waves and pseudopotentials were performed using the Vienna ab-initio simulation package (VASP) (Kresse and Hafner, 1993, 1994; Kresse and Furthmüller, 1996a,b). The exchange correlation energy was described within the generalized gradient approximation (GGA) following the work of Perdew-Burke-Ernzerhof (PBE) (Perdew et al., 1996).

Pseudopotentials of the Projector Augmented Wave (PAW) type were used to describe the core electrons (Blöchl, 1994; Kresse and Joubert, 1999). The Kohn-Sham single-electron wave functions were expanded by plane waves with a cut-off energy of 408 eV. To accurately treat the strongly correlated Ce 4f states, the rotationally invariant Hubbard correction scheme from Dudarev et al. (1998) with a U-value of 5 eV on the Ce 4f states was applied. Previous studies have shown that this U-value gives a good description of the electronic properties of reduced ceria (Castleton et al., 2007; Loschen et al., 2008; Kullgren et al., 2013; Du et al., 2018). In all calculations, the Brillouin zone was sampled at the Ŵ-point. All structures were pre-optimized using SCC-DFTB. Some were selected and subsequently geometry optimized at the DFT level of theory. Geometry optimizations for all modeled structures were carried out until the maximal force acting on any one atom was <0.02 eV/Å.

We have primarily studied NP pairs, (NP)2, formed by two stoichiometric truncated octahedral ceria NPs with the chemical formula Ce40O80. The two NPs were connected via low-index facets, forming {111}/{111} or {100}/{100} interfaces. The free solid-vacuum {111} surfaces of stoichiometric ceria are known to be the most stable ones and they make up the largest surface area on the isolated Ce40O<sup>80</sup> particles (see **Figure 1**). Free {100} surfaces are the most reactive low-index surfaces of ceria and have been found to be present on truncated octahedral NPs in UHV (Zhang et al., 2011; Lin et al., 2014).

To generate representative NP interfaces we placed two identical NPs, facet to facet at a distance of 2.7 Å in a periodic cell of the dimensions, 4.5 × 3 × 3 nm<sup>3</sup> . Reference calculations were performed for the isolated Ce40O<sup>80</sup> NP, placed in a 3 × 3 × 3 nm<sup>3</sup> box. One such particle is approximately 15 Å in diameter. We used rather large cell sizes to prevent image-image interactions between particles in neighboring supercells. The same supercell sizes were used in both the DFT and the SCC-DFTB calculations. The final optimized structures were analyzed in terms of stability, structural parameters and electronic properties, as defined in the Results and Discussion section. The calculations for the O<sup>2</sup> molecule in its ground triplet electronic state were also performed, using a large periodic cell, 3 × 3 × 3 nm<sup>3</sup> .

In summary, the following stoichiometric systems are examined by electronic structure methods in this study: an isolated truncated Ce40O<sup>80</sup> labeled NP (**Figure 1**), and a number of (Ce40O80)<sup>2</sup> systems labeled (NP)2. Reduced systems, where one neutral O atom was removed from (various locations on) the ceria particle to form half an oxygen molecule, were also studied for both the NP and the (NP)<sup>2</sup> systems.

# RESULTS AND DISCUSSION

# Consistency Between the Methods That We Have Used

The DFTB parameters (the entries in the Slater-Koster table and the Vrep parameters) used here were trained against DFT calculations (Kullgren et al., 2017). In Kullgren et al. (2017), the DFTB model was subjected to an extended validation process, and passed the test of reproducing the reference DFT results with respect to NP structures as well as oxygen vacancy formation energies in ceria bulk and at the low-index surfaces. In the present paper, the DFTB model was further validated for some of the interface structures, and we find that calculated energies, such as the interface formation energy, Einterface, and the oxygen vacancy formation energy, Evac, from SCC-DFTB are in reasonably good agreement with those from DFT (see the correlation in **Figure 2** and **Table S-I**). Moreover, the optimized DFT and SCC-DFTB structures are found to differ by <1% in terms of bond lengths. This consistency is a great advantage in our approach to generating interface models, as it reduces the number of computationally expensive DFT energy evaluations. We find that the computational approach used here, i.e., pre-screening using SCC-DFTB followed by full structural optimisations using DFT for selected systems is an efficient and useful protocol.

# Identifying Stable Coherent and Incoherent Interfaces

For the single NP, we use a prototypical model structure (**Figure 1**), namely a stoichiometric Ce40O<sup>80</sup> particle, which is a truncated octahedron exposing eight large {111} facets and six small {100} facets, two of which are decorated with an apex Ce ion. Ceria NPs exhibiting truncated octahedral shapes have been reported in many experiments (Wang and Feng, 2003; Mai et al., 2005; Du et al., 2007; Tan et al., 2011; Wang and Mutinda, 2011; Lin et al., 2012; Cordeiro et al., 2013; Florea et al., 2013; Liu et al., 2015) and previous theoretical studies have shown the truncated octahedral NPs to be stable also under vacuum conditions [see e.g., (Fronzi et al., 2009)].

SCC-DFTB calculations were performed to examine the stabilities of the NP pairs with respect to variation of the angle of rotation of "the second" NP with respect to "the first," where the axis of rotation is the direction joining the two centers-of-mass of the constituent NPs. Shifting one NP with respect to the other was also considered, leading to an interface which is a continuation of the bulk fluorite structure, i.e., a coherent interface. In this way, {111}/{111} and {100}/{100} interfaces were formed. After the SCC-DFTB pre-screening, four final structures emerged from our search for stable (NP)<sup>2</sup> agglomerates: one with the two NPs joined by {111} faces, and three which were joined by {100} faces; they were subjected to subsequent DFT optimizations. All figures and tables in the following are based on DFT results, except the very last figure which will show results from ReaxFF MD simulations.

# The {111}/{111} Interface

Here we note that shifting one NP with respect to the other NP, parallel to the interface, is essential in order to avoid the Coulomb repulsion between ions of the same type; see the cross-sectional view to the right in **Figure 3**. The most stable structure of the {111}/{111} interface is obtained by shifting one NP to match the stacking sequence of the other NP. As can be seen in **Figure 3**, the interface structure formed by attaching {111} to {111} facets is coherent. Hereafter, we use the label "(NP)2-111-shift" to refer to agglomerated NPs with the {111}/{111} interface structures obtained by shifting.

The interface formation energy, Einterface, is calculated by the following equation:

$$E\_{\text{interface}} = E\_{\text{tot}} \left[ (\text{NP})\_2 \right] - 2 \times E\_{\text{tot}} \left[ \text{NP} \right] \tag{1}$$

where Etot[(NP)2] is the total energy of agglomerated NPs with an interface and Etot[NP] is that of the isolated NP. To aid in the comparison of interface stabilities, especially for interfaces with oxygen vacancies, we will primarily use Einterface values but also discuss the role of the interface area with normalized interface energies. Our calculated Einterface of the (NP)2-111-shift interface is −8.12 eV, as listed in **Table 1**. Note that a hexagonal close packed type {111}/{111} interface, which can be considered as a stacking fault, was also tested and found to be far less stable (by 2.72 eV) than the (NP)2-111-shift interface.

# The {100}/{100} Interface

The protocol for finding stable {100}/{100} interfaces included a scan of the angle of rotation of one NP with respect to the other from 0 to 45◦ with an increment of 1◦ . Following this pre-screening, three distinctly different stable structures were found, c.f. **Figure 4** and **Table 1**. They are either rotated by 15 or 45◦ with respect to each other, or by shifting one NP in a direction perpendicular to the interface normal by half of the lattice parameter so as to match the bulk fluorite structure. We will refer to them as "(NP)2-100-rot15," "(NP)2-100-rot45,"

TABLE 1 | Results for the stoichiometric systems, namely for the four stable DFT-optimized NP pairs found in this work, based on the truncated octahedral Ce40O80 NP.


Data for the single NP itself is given as references. The table lists the notation, the interfacial energy, Einterface, calculated according to Equation (1), the number of surface/interface oxygen ions with certain coordination numbers (CNs), the number of new Ce–O bonds created on interface formation, and the figures where the DFT-optimized structures are shown. The interface notation NP(111)/NP(-1-1-1) refers to all symmetry-equivalent combinations, which is of course also true for the (111) surface notation.

<sup>a</sup>Number of surface or interface O ions with CN=2, CN=3, CN=4 (Thus OCN=2, for example, means that all these O atoms in the interface region have exactly two nearest-neighbor Ce ions each).

<sup>b</sup>Sum of the coordination numbers of all O ions in the interface region (they are depicted in Figures 3, 4) minus the sum of the coordination numbers of the same atoms in the isolated NPs. Neighbors within a distance limit of 2.6 Å were counted.

and "(NP)2-100-shift" in the following. Thus, our results show that the {100}/{100} interface structures can be either coherent or incoherent: the {100}/{100} interface obtained by shifting is coherent while the {100}/{100} interfaces obtained by rotating by 15 or 45◦ are incoherent. In the literature, it has been suggested (Sk et al., 2014) that the formation of {100}/{100} interfaces requires that one side of a truncated ceria surface be ceriumterminated (having oxygen vacancies at a {100} surface) while the opposite side needs to be oxygen-terminated in order to match the crystal ordering when two particles agglomerate. However, this is a restriction following the use of periodic boundary conditions, where one NP formed an interface between a ceriumterminated surface and an oxygen-terminated surface of the same NP. When we give full freedom to the relaxation and formation of the {100}/{100} interface, such constraints do not exist.

The calculated Einterface values for the agglomerated NPs with the 15 and 45◦ rotation is −7.19 and −5.13 eV, respectively, and for "(NP)2-100-shift", it is −6.82 eV. Thus, if the NPs are stoichiometric, the {111}/{111} interface (−8.12 eV) is more stable than the {100}/{100} interface structures, likely due to the denser and larger contact areas of the former. It is important to note that for the reported interface energies, the contact areas are very different. The approximate area of the (111) facet on the free Ce40O<sup>80</sup> NP is 43.8 Å<sup>2</sup> and that of the (100) facet is only 13.3 Å<sup>2</sup> . This implies that the normalized interface energy per interface area will be much larger in case of the {100}/{100} interface. For the most stable {111}/{111} and {100}/{100} interfaces on the (NP)<sup>2</sup> agglomerate, we get −2.97 J/m<sup>2</sup> (−0.19 eV/Å<sup>2</sup> ) and −8.66 J/m<sup>2</sup> (−0.54 eV/Å<sup>2</sup> ), respectively. It is interesting to compare these values with the calculated surface energies of 0.71 J/m<sup>2</sup> and 1.54 J/m<sup>2</sup> for the extended (111) and (100) surfaces, respectively (Kullgren et al., 2017). The calculated surface energy of the (100) surface in Kullgren et al. (2017) was obtained for a reconstructed (100) surface where half of the terminal O ions on the (100) surface were moved to the opposite face to compensate the polarity of the surface. The normalized interface energy in the limit of an infinite interface will be twice the negative surface energy, and thus becomes −1.42 and −3.08 J/m<sup>2</sup> , respectively. We note that the calculated normalized nano-interface energies reveal stronger adhesion by at least a factor of 2 compared to the limiting values. This reflects the fact that the surfaces on the small Ce40O<sup>80</sup> NP mainly consist of rim atoms. Thus, even though the {111}/{111} interface on a total energy scale is more stable than the {100}/{100} interfaces, the latter has a stronger adhesion per unit area, and it may very well form in experiments through an initial first approach via the OA mechanism, and then become kinetically trapped (Si et al., 2006; Tan et al., 2011; Florea et al., 2013).

The interatomic distance within each nearest neighbor (NN) pair, Ce—O, Ce—Ce, and O—O at the rim of the interface of the three {100} systems, (NP)2-100-rot15, -rot45, and -shift, are further analyzed here. First, let us discuss the NN distances of all constituent pairs at the {100} facets that are not involved in the interface formation. They are found to be identical with those of the isolated single Ce40O<sup>80</sup> NP; this indicates that the effect of interface formation on the structural distortion is localized at the interface. Intuitively, one could imagine that the interface stability will correlate with the number of bonds formed upon agglomeration. We have therefore calculated coordination numbers and coordination number differences, 1(bonds), for the all studied interfaces, c.f. **Table 1**. We note that while this hypothesis can be used to explain why (NP)2-100-rot15 is the most stable interface among the three {100} systems, it fails to discriminate between the different {100}/{100} interfaces. The calculated 1(bonds) for (NP)2-100-rot15, -rot45, and -shift is +4, +8, and +8, respectively. Here, the NN distances of Ce—Ce and O—O, which will cause repulsive forces and thereby result in a less negative Einterface (less stable interface), give a better correlation, c.f. **Figure 5**, which shows the calculated Einterface vs. average O—O NN distance. The average O—O NN distance at the {100} facet in the isolated Ce40O<sup>80</sup> NP is 2.56 Å, and at the interfaces the resulting distances are 2.86, 2.54, and 2.74 Å for (NP)2-100-rot15, -rot45, and -shift, respectively. For (NP)2- 100-rot45 which is the least stable interface structure among our model interfaces, all the distances of O—O NN pairs at the interface are shorter than 2.56 Å. On the other hand, the interface formation by rotating 15◦ eliminates these short O—O

NN pairs at the interface. We conclude that the larger stability of (NP)2-100-rot15 is attributed to the stretching of the short O—O pairs at the interface, which mitigates repulsion.

The presence of O—O pairs with short distances affects the electronic properties of agglomerated ceria NPs. This is seen in the electronic projected density of states (PDOS) for the three {100}/{100} interface systems when compared to the isolated NP (see **Figure 6**). Here, most notably, we find a clear peak at the valence band maximum in the PDOS for the isolated NP and (NP)2-100-rot45 interface. Interestingly, the magnitude of the peak is found to be smaller in case of the (NP)2-100-shift and -rot15 interfaces (see the black arrows in **Figure 6**). From the PDOS, it is clearly seen that this peak primarily originates from terminal O ions at the (100) facets (see the dashed red lines in **Figure 6**). We find that they are preserved for (NP)2-100 rot45 whereas in the (NP)2-100-shift and -rot15 interfaces, they are effectively quenched (compare solid green lines in **Figure 6**). Thus, in a sense, the (NP)2-100-rot45 interface structure may be regarded as electronically analogous to the free {100} facets of isolated Ce40O<sup>80</sup> NPs. For (NP)2-100-rot15 and -shift, on the other hand, the surface states are significantly reduced by the interface formation.

## Oxygen Chemistry

Oxygen vacancies play a central role in the redox chemistry of ceria and we have investigated whether and how the oxygen vacancy formation energy, Evac, is affected by the creation of an interface between the NPs. Evac is here calculated in the usual way, i.e., as

$$E\_{\rm vac} = E\_{\rm tot} \left[ \rm NP + nV\_{\rm O} \right] - E\_{\rm tot} \left[ \rm NP \right] + n \times \frac{1}{2} E \left[ \rm O\_2(g) \right] \tag{2}$$

where Etot[NP + nVO] and Etot[NP] are the total energies of the system with and without the number of oxygen vacancies, nVO, respectively. E[O2(g)] is the energy of an O<sup>2</sup> molecule in its ground triplet electronic state.

First we give a general remark, namely that we find that in all cases, for the interfaces as well as for the single NP, the oxidation states of the apex Ce ions (determined from the 4felectron occupancy) change from 4+ to 3+ after the oxygen

vacancy formation. The apex Ce ions are marked in **Figure 1** for the isolated NP and are present in all the (NP)<sup>2</sup> systems as well but are not so easy to discern in the figures. The locations of the excess electrons created at the interface is consistent with several previous single-NP studies in the literature, which reported that the two Ce3<sup>+</sup> ions are preferentially located at the apex sites where the coordination numbers of the Ce ions are the lowest (see Loschen et al., 2007, 2008; Kullgren et al., 2013; Sk et al., 2014).

Let us now consider oxygen vacancy formation at different surfaces and sub-surface positions on an isolated NP as shown in **Figure 1**. The vacancy will be labeled according to the facet involved and the label of the surface of sub-surface O removed, as listed in the second column in **Table 2**. The resulting Evac values are also given in the table. The smallest Evac value, i.e., the energetically most stable site for an oxygen vacancy to be formed in an isolated NP is the 2-coordinated O2<sup>−</sup> site at the {100} facet ("O(100)" in **Figure 7A**) with Evac = +0.55 eV, which is in good agreement with the previous DFT work published by the group of Neyman et al., showing that a low-coordinated oxygen vacancy formation costs the least energy (Migani et al., 2010). Oxygen vacancies at the {111} facet (see **Figure 7A**) fall in the range from +1.50 to +1.91 eV depending on their vicinity to the {100} facets; the further from the {100} facet, the higher the Evac value.

We find that the most preferable site for an oxygen vacancy to form at the {111}/{111} interface is at a 3-coordinated O2<sup>−</sup> site (labeled O<sup>A</sup> interface in **Figure 7B**) with a formation energy of +0.59 eV which is similar to the corresponding site in the isolated NP. For O<sup>B</sup> interface, <sup>E</sup>vac is <sup>+</sup>1.31 eV. When an oxygen vacancy was formed at O<sup>∗</sup> interface, the vacancy was found to be healed by O′ which was originally located behind the adjacent Ce, which

TABLE 2 | Vacancy formation energies, Evac, for the formation of one and two vacancies at the surface/interface according to Equation (2).


For the (NP)<sup>2</sup> systems, a 2nd vacancy was created by removal of each of the remaining interface O ions in turn, but only one of them at a time; the lowest value found is reported as the 2nd Evac in the table. <sup>a</sup>Figure 7A. <sup>b</sup>Figure 7B. <sup>c</sup>Figure 8A. <sup>d</sup>Figure 8B. <sup>e</sup>Figure 8C.

results in the same structure as when the vacancy was created at O<sup>A</sup> interface. We conclude that for a pair of truncated octahedral NPs joined by an {111}/{111} interface, which is coherent, the interface formation has none (or a very small) effect on the oxygen vacancy formation capability at the interface.

For the {100}/{100} interfaces, the structures and oxygen site labeling schemes are shown in **Figure 8** and their coordination numbers and Evac values are given in **Table 2**. We find that here the oxygen vacancy formation behavior is different from both that of the isolated NP and that of the {111}/{111} interface. For the most stable configuration among the {100}/{100} interface structures, (NP)2-100-rot15, the lowest Evac value (+0.28 eV) is about half of the lowest value for the isolated NP, which was already small (+0.55 eV) (c.f. **Figure 8A**). In fact, all the {100}/{100} interfaces display very small Evac values despite the fact that the interfaces themselves are quite stable. The reason is no doubt connected to the fact that new bonds are formed between the NPs but the NN coordination numbers are modest in these structures: enough to hold the particles strongly together but not enough to hang on to O ions if the opportunity for reduction arises. This is true not only at the rims of these

FIGURE 7 | Atom labeling scheme used in the text and in Table 2 for discussion of oxygen vacancy formation at (A) the {111} facet of the isolated Ce40O80 NP and at (B) the (NP)2-111-shift interface formed from it. The same color scheme as in Figure 3 has been used, except that the sub-surface oxygens in (A) are colored dark red. In both (A) and (B), the left-most Ce atom in the upper row and the right-most Ce atom in the bottom row in (B) are the "apex Ce" marked in Figure 1.

interfaces, but throughout the cross-section, as the coordination number is maintained also in the middle of the cross-section.

The ultimate oxygen lability is demonstrated by the (NP)2- 100-rot45 interface which has a large negative Evac, −1.53 eV, meaning that, thermodynamically speaking, this oxygen vacancy would be formed spontaneously at the interface (and have a stabilizing effect on the interface). A second oxygen can also easily be removed, at hardly any cost (c.f. **Table 2**). The Evac values of the second oxygen vacancy lies in the range from +0.11 to +0.56 eV for all the {100}/{100} interface structures. Under alternating reducing and oxidizing conditions, i.e., when the chemical potential of oxygen changes, Evac will change accordingly. At some point, Evac will turn negative and vacancies will stabilize the interface. Thus, reducing conditions may promote the stability of the interface significantly by oxygen removal from the interface.

The decreased Evac value at {100}/{100} interfaces can be understood by examining the average NN O—O distance at the interface. We find that the occurrence of short O—O pairs at the interface (discussed above) diminishes after the oxygen vacancy formation. For example, the average NN O—O distance of (NP)2-100-rot45 increases significantly, from 2.54 to 2.85 Å, when an oxygen vacancy is formed at the interface. It is then also observed that the two Ce3<sup>+</sup> ions are generated at the apexes of the agglomerated NPs. We have found (not shown here) that these generated Ce3<sup>+</sup> ions can be further utilized as adsorption sites for incoming oxygen molecules from the surroundings, similar to the phenomenon of supercharged oxygen storage capacity discussed earlier in the literature for ceria NPs (Kullgren et al., 2013; Renuka et al., 2015).

In summary, we find that {100}/{100} interface formation can lower Evac significantly. In other words, the interface formation eliminates the areas of reactive (100) surfaces but promotes a more facile oxygen extraction, which in turn may promote an increased low-temperature oxygen storage capacity of ceria via the so called supercharging mechanism (Kullgren et al., 2013). **Figure 9** summarizes the relative energies of the interface structures considered in this work as a function of increasing number of oxygen vacancies. The reference system is two isolated NPs at infinite distance, and their oxygen vacancies are created by

FIGURE 8 | Atom labeling scheme used in the text and in Table 2 for discussion of oxygen vacancy formation at (A) the (NP)2-100-rot15, (B) the (NP)2-100-shift, and (C) the (NP)2-100-rot45 interface structures.

removing neutral O atoms from a {100} facet of each NP. While the {111}/{111} interface is more stable for the stoichiometric cases, we find that the {100}/{100} interface promotes a lower Evac compared to the {111}/{111} interface. Interestingly, in the same comparison, oxygen vacancy formation at the {111}/{111}

interface slightly increases Evac. As a result, in the presence of two oxygen vacancies, the stabilities of {100}/{100} interface structures, here (NP)2-100-rot15 and -rot45, become comparable with that of the {111}/{111} interface (see **Figure 9**). This result paves the way not only for understanding the engineering of the interface structure of nanoceria for improved low-temperature redox activity, but also for suggesting experimental guidelines to control a growth direction of ceria nanowires. For example, our results suggest that under oxidizing conditions, a preferential growth direction of agglomerated NPs would be h111i while the {100}/{100} interface can be formed dominantly under reducing conditions.

# Extrapolation to Larger Systems

One might raise a question whether the interface construction procedure used in this work will also apply to larger systems and what role the interface area plays. To answer these questions, we constructed interfaces between two Ce132O<sup>264</sup> NPs as shown in **Figure 10**. The {111}/{111} interface can still be formed following a continuation of bulk fluorite (see **Figure 10A**), and one might expect that the interface formation will be similar to that of the smaller system, Ce40O80, albeit with a larger area of overlap. However, for {100}/{100} interfaces, the interface construction by rotation does not work since it requires a large rearrangement of constituent ions at the interface which will induce a distortion of the structure. For example, since the interface structure of (Ce132O264)2-100-rot15 (**Figure 10B**) contains overlapping Ce and O ions at the center of the interface, it is impossible to avoid the repulsion between these unless the whole interface structure is distorted, which obviously requires a huge amount of energy. "Overlapping ions" here means that if the interface region is seen in a projection perpendicular to the interface, then the Ce ions in the two constituent NPs would (almost) overlap with each other, and similarly for the O ions. Interface construction by shifting also suffers from O ion overlaps as the interface area becomes larger. For a (Ce132O264)2-100-shift NP dimer, constructed as shown in **Figure 10C**, there are 36 overlapping O ions at the interface, but only two sites are available for Ce—O bonding (dashed circle in **Figure 10C**). To prevent this overlapping and produce a coherent bulk fluorite structure at the interface as was achieved for (Ce40O80)2-100-shift, at

FIGURE 10 | Cross-sectional views of the interface structures of (A) (NP)2-111-shift, (B) (NP)2-100-rot15, (C) -shift, and (D) -rot45 of two unrelaxed agglomerated Ce132O264 NPs (taken from truncated DFT bulk structures).

isolated NPs constitute the reference system.

least 16 O ions need to be removed at the interface. Needless to say, the stoichiometric interface structure of (Ce132O264)2- 100-rot45 as shown in **Figure 10D** would be very unstable due to the overlapped ions at the interface. However, they could be stabilized at reducing conditions by oxygen removal from the interface.

For the large NPs as well as for the smaller ones, we have also performed ReaxFF molecular dynamics (MD) simulations addressing the dynamics of the agglomeration process. The parameter set of ReaxFF used here was derived by us from DFT data (Broqvist et al., 2015). Further technical details about the validation of the parameters of ReaxFF are presented in the **Supplemental Material**. Two simulations extending over 5 ns at 400 K using the NVT ensemble, one with 10 Ce40O<sup>80</sup> NPs and one with 10 larger Ce132O<sup>264</sup> NPs, were performed in the current study. The former system consists of particles exposing four small {100} facets each (c.f. **Figure 1**), while the latter represents a system of particles exposing only one, albeit larger, {100} facet per particle. One might say the MD simulations are too short to capture the entire agglomeration process. However, they are long enough to include the coalescence and initial restructuring of the interfaces. In the case of a low particle density (as we have here) we expect that the effective barrier to transform a metastable {100}/{100} interface into the energetically more stable {111}/{111} interface must be bound by the interface energy of the former. The magnitude of the interface energy is larger than 5 eV which is large enough to permit long-lived interfaces, even if they happen to be meta-stable ones. The real barrier could obviously be quite different since particles forming the interface may slide or rotate while still maintaining contact. One aim with these MD simulations is therefore to look for indications of dynamical changes in the interface structures formed.

**Figures 11A,B** compare representative snapshots from the two simulations. We first note that the Ce40O<sup>80</sup> NPs form an isolated agglomerate while the Ce132O<sup>264</sup> NPs form an extended two-dimensional web. In both cases, we observe that various kinds of interfaces are formed, not only the energetically stable {111}/{111} interfaces as we described in the earlier section, but also {111}/{100}, apex/{100}, and partially matched {111}/{111} interfaces. This observation is in line with the theoretical findings of Fichthorn and co-workers (Alimohammadi and Fichthorn, 2009; Raju et al., 2014) who showed that titania NPs prefer to aggregate in the direction of approach in vacuum, which suggests that the interface formation may not always follow the thermodynamically most stable pathway. For example, in the system of the smaller particles (Ce40O80), we do not observe any formation of the energetically stable {100}/{100} interfaces, but instead find the coherent {111}/{111} interfaces, incoherent {111}/{100} interfaces, and apex/{100}. We attribute this fact to the strong variation in interface energy with respect to particle rotation, which implies that particles have to meet at "just the right" angle to form the {100}/{100} interface. We observe two interfaces where the apex ion of one particle binds to the center of a {100} facet of another particle (apex/{100}). We note the similarity of this structure to the interface presented in the theoretical work by Sk et al. (2014), although, in our case the two particles are rotated with respect to each other. In this structure (**Figure 11A**), a positively charged Ce ion at one apex is bonded to a square of negatively charged O ions on the {100} facet. The interface is seen to readily rotate during the MD simulation. For comparison, we calculated Einterface of the apex/{100} interface energy using Equation (1) at the DFT level to be −1.87 eV. This value is significantly smaller in magnitude than that of the {100}/{100} interfaces which range from −5.1 to −7.2 eV. Nevertheless, the energy is not small enough to rule out that apex/{100} interfaces, once formed, could have a substantial life time, clearly longer than the short simulation time used here. The snapshot from the MD simulation of the Ce132O<sup>264</sup> system shown in **Figure 11B** dominantly displays the formation of {111}/{111} interfaces. This is in good agreement with an experimental finding (c.f. Figure 7B in Florea et al., 2013). As discussed for overlapping ions of larger NPs above, the fact that {100}/{100} interfaces are not found in our agglomerate of Ce132O<sup>264</sup> NPs is likely due to the repulsion between overlapping Ce and O ions, respectively. However, under reducing conditions such interfaces might be more likely to form, which is an interesting topic for further study.

configuration of an apex/{100} interface.

# CONCLUSIONS

In this work, we addressed the effects that agglomeration may have on the stability and redox activity of ceria NPs. We applied a computational protocol based on a multiscale simulation approach combining three levels of theory, namely DFT, SCC-DFTB and ReaxFF. In particular, we have studied the formation of Ce40O<sup>80</sup> pairs using DFT and SCC-DFTB, and larger agglomerates of Ce40O<sup>80</sup> or Ce132O<sup>264</sup> NPs in MD simulations with ReaxFF.

The most stable {111}/{111} interface structure was found to be coherent, i.e., it exhibits a continuation of the bulk structure, whereas the stable {100}/{100} structures can be either coherent or incoherent. A systematic study of the implications of interface formation on the oxygen chemistry revealed that (i) {111}/{111} interface formation has only a very small effect on the redox activity, and (ii) Evac decreases significantly for oxygen atoms in the vicinity of {100}/{100} interfaces. We conclude that while interface formation eliminates reactive {100} facets, it promotes an enhanced low-temperature redox activity involving the extraction of lattice oxygen; the net result might be an overall increased oxygen storage capacity of ceria.

# AUTHOR CONTRIBUTIONS

B-HK has performed the simulations and drafted the first version of the manuscript. All other authors have been equally

# REFERENCES


involved in the planning of the work and in the finalization of the text.

# FUNDING

This work was supported by the Swedish Research Council (VR) and the Swedish strategic e-science research programme eSSENCE, and the Sweden-Korea program within The Swedish Foundation for International Cooperation in Research and Higher Education (STINT). B-HK acknowledges support by Nano-Material Technology Development Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Science and ICT (NRF-2016M3A7B4024138). JK acknowledges support from Åforsk. The calculations were performed on resources provided by the Swedish National Infrastructure for Computing (SNIC) at UPPMAX and NSC. This work was also supported by the National Institute of Supercomputing and Network/Korea Institute of Science and Technology Information with supercomputing resources including technical support (KSC-2018-S1-0012).

# SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fchem. 2019.00203/full#supplementary-material


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Kim, Kullgren, Wolf, Hermansson and Broqvist. This is an openaccess article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# First Principles Calculations on the Stoichiometric and Defective (101) Anatase Surface and Upon Hydrogen and H2Pc Adsorption: The Influence of Electronic Exchange and Correlation and of Basis Set Approximations

Ruth Martínez-Casado1,2 \*, Milica Todorovic´ 3 , Giuseppe Mallia<sup>4</sup> , Nicholas M. Harrison<sup>4</sup> and Rubén Pérez 2,5 \*

### Edited by:

Javier Carrasco, CIC Energigune, Spain

## Reviewed by:

Annabella Selloni, Princeton University, United States Jose Julio Gutierrez Moreno, Shenzhen University, China

### \*Correspondence:

Ruth Martínez-Casado mariarum@ucm.es Rubén Pérez ruben.perez@uam.es

### Specialty section:

This article was submitted to Physical Chemistry and Chemical Physics, a section of the journal Frontiers in Chemistry

Received: 30 November 2018 Accepted: 20 March 2019 Published: 16 April 2019

### Citation:

Martínez-Casado R, Todorovic M, ´ Mallia G, Harrison NM and Pérez R (2019) First Principles Calculations on the Stoichiometric and Defective (101) Anatase Surface and Upon Hydrogen and H2Pc Adsorption: The Influence of Electronic Exchange and Correlation and of Basis Set Approximations. Front. Chem. 7:220. doi: 10.3389/fchem.2019.00220 <sup>1</sup> Departamento de Física de Materiales, Universidad Complutense de Madrid, Madrid, Spain, <sup>2</sup> Departamento de Física Teórica de la Materia Condensada, Universidad Autónoma de Madrid, Madrid, Spain, <sup>3</sup> Department of Applied Physics, Aalto University, Espoo, Finland, <sup>4</sup> Department of Chemistry, Imperial College London, White City, London, United Kingdom, <sup>5</sup> Condensed Matter Physics Center (IFIMAC), Universidad Autónoma de Madrid, Madrid, Spain

Anatase TiO<sup>2</sup> provides photoactivity with high chemical stability at a reasonable cost. Different methods have been used to enhance its photocatalytic activity by creating band gap states through the introduction of oxygen vacancies, hydrogen impurities, or the adorption of phthalocyanines, which are usually employed as organic dyes in dye-sensitized solar cells. Predicting how these interactions affect the electronic structure of anatase requires an efficient and robust theory. In order to document the efficiency and accuracy of commonly used approaches we have considered two widely used implementations of density functional theory (DFT), namely the all-electron linear combination of atomic orbitals (AE–LCAO) and the pseudo-potential plane waves (PP–PW) approaches, to calculate the properties of the stoichiometric and defective anatase TiO<sup>2</sup> (101) surface. Hybrid functionals, and in particular HSE, lead to a computed band gap in agreement with that measured by using UV adsorption spectroscopy. When using PBE+U, the gap is underestimated by 20 % but the computed position of defect induced gap states relative to the conduction band minimum (CBM) are found to be in good agreement with those calculated using hybrid functionals. These results allow us to conclude that hybrid functionals based on the use of AE–LCAO provide an efficient and robust approach for predicting trends in the band gap and the position of gap states in large model systems. We extend this analysis to surface adsorption and use the AE–LCAO approach with the hybrid functional HSED3 to study the adsorption of the phthalocyanine H2Pc on anatase (101). Our results suggest that H2Pc prefers to be adsorbed on the surface Ti5<sup>c</sup> rows of anatase (101), in agreement with that seen in recent STM experiments on rutile (110).

Keywords: density functional theory, oxides, anatase, hybrid functionals, defects, phthalocyanine

# 1. INTRODUCTION

TiO<sup>2</sup> is an important technological material with widespread applications in solar cells and photocatalysis (Fujishima and Honda, 1972; Fujishima et al., 2008; Schneider et al., 2014; Stetsovych et al., 2015). Although rutile is the stable bulk phase, anatase nanoparticles are often found to be the most active components in these applications.

Among the possible terminations, the (101) facet is found to be the most stable anatase surface (Labat et al., 2008) and represents a significant portion of the exposed surface area in the equilibrium crystallites that are characteristic of polycrystalline surfaces (Lazzeri et al., 2001; Barnard and Curtiss, 2005). The large anatase band gap (3.2 eV) limits its photocatalytic activity to the small ultraviolet portion of the solar spectrum (Hagfeldt and Grätzel, 1995). This has led to the use of a wide variety of approaches to engineer the band gap in order to enhance the activity. These include the incorporation of metallic and non-metallic ion impurities (Hoffmann et al., 1995). Hydrogen impurities are also usually introduced during processing and have significant influence on material properties, and therefore on the device performance. It is well established that hydrogen defects affect the ionic conductivity, electronic, and optical properties of both the bulk crystal and its surfaces (Amano et al., 2016). Hydrogen binds to lattice O2<sup>−</sup> ions to form OH<sup>−</sup> and this n-dopes the crystal modifying its electronic structure and, at elevated temperatures, proton transport makes a significant contribution to the ionic conductivity (Amano et al., 2016). Oxygen vacancies also play an important role in determining the surface chemistry and electronic properties of TiO<sup>2</sup> (Morgan and Watson, 2010; Pan et al., 2013; Hongfei et al., 2015). Under typical processing conditions the dominant intrinsic defect is the oxygen vacancy, which also n-dopes the crystal and significantly affects photocatalytic properties (Liborio and Harrison, 2008). Previous DFT calculations have shown that TiO<sup>2</sup> typically accommodates n-doping in localized states within the band gaps approximately 1eV below the CBM. These states could extend optical absorption to the visible region (Justicia et al., 2002; Liborio and Harrison, 2008). An alternative, and potentially complementary, approach is to enhance light absorption through surface species as happens in dye-sensitized solar cells (Graetzel, 2001). The details of the binding of the dye molecules to the surface are crucial for the improvement of device performance, with phthalocyanines being the most widely studied sysetm (Godlewski and Szymonski, 2013).

The quantitative analysis, and correct interpretation, of these processes has been assisted by the use of ab initio calculations of the surface structure and electronic properties. Anatase, as other reducible oxides, represents a challenge for standard implementations of DFT. Generalized gradient approximations (GGA) functionals often underestimate the band gap (Reshak et al., 2010) and provide a poor description of the localized gap states which are essentially lattice Ti4+(3d 0 ) ions reduced to Ti3+(3d 1 ) ions (Wang and Doren, 2005). It has been previously shown that the use of hybrid exchange functionals (where a proportion of Fock exchange is included in the exchange functional) provides a qualitatively correct description of the structure, energetics and electronic properties for many different materials, and in particular, for transition metal oxides (Patel et al., 2012, 2014; Sanches et al., 2014). However, in planewave codes, the use of hybrid exchange functionals increases dramatically the computational cost to an extent that prevents studies of complex geometries. For this reason, it is common to use the so-called DFT+U (Anisimov et al., 1993; Stetsovych et al., 2015) approximation, where LDA or GGA functionals are supplemented by an on-site local repulsion term U, representing the Coulombic self-interaction of the Ti 3d orbitals. This term partially corrects for the electronic self-interaction error inherent in local approximations to DFT and thus provides a more reliable, if somewhat ad hoc, description of the electron localization around Ti3<sup>+</sup> ions.

In this work, firstly we document the performance of two widely employed DFT implementations (AE–LCAO and PP– PW) for computing the electronic structure of the defective (101) anatase surface using different exchange-correlation functionals. The defects considered are the dominant intrinsic ones for this system: single-atom oxygen sub-surface vacancies and hydrogen adsorbates (Setvin et al., 2013). The implementation of hybridexchange functionals using local atomic basis sets, in the CRYSTAL code (Dovesi et al., 2014), is very efficient, even when a high quality all-electron basis sets are used. CRYSTAL has therefore been used to investigate the stoichiometric and defective (101) anatase surface and upon hydrogen and H2Pc adsorption using the Perdew-Burke-Ernzerhof (PBE Perdew et al., 1996a) functional and the hybrid [PBE0 (Perdew et al., 1996b), B3LYP (Becke, 1988), and HSE (Heyd et al., 2003)] functionals. For the PP–PW calculations, only the PBE and PBE+U levels have been considered, rather than hybrid exchange functionals, due to the high computational cost of using hybrid functionals for systematic studies (Wu et al., 2009). In our experience, for PW calculations of the anatase bulk crystal, the computational time increases by three orders of magnitude when using hybrid exchange functionals compared to the AE– LCAO approximation. The comparison of the two approaches allows us to document the effects of the choice of basis set, the pseudopotential and the validity of the empirical PBE+U approach in this system. The PBE+U formalism has been shown to describe qualitatively correctly the defective rutile phase of TiO2, both in the bulk and on the surface, but the results for anatase surface remain ambiguous (Haa and Alexandrova, 2016). A similar analysis to the one presented here have been reported for bulk anatase (Finazzi et al., 2008; Valentin et al., 2009), but the complexity of the surface chemistry makes clear the value of the present study.

A judicious combination of the PP–PW and AE–LCAO approaches can provide an accurate description of complex defective and hybrid organic/oxide structures needed to understand and to improve the chemical reactivity and the photocatalytic properties. Once we have shown that the AE– LCAO approach with the HSE functional is able to accurately describe this surface, we go one step further by studying the adsorption of the phthalocyanine H2Pc on anatase (101). In the molecule-surface interaction, the effect of the Van der Waals forces is more important and has been taken into account by adding the D3 dispersion correction (HSED3) (Grimme and Krieg, 2010). We will show an analysis of the optimized structure and stability of this interaction, which will be a valuable tool in order to improve the efficiency of dye-sensitized solar cells. The article is structured as follows. In section 2 details are presented. In sections 3.1 and 3.2 the results for the bulk and stoichiometric anatase (101) surface are shown as a reference. The main features of the oxygen defective anatase (101) surface, together with the adsorption of both hydrogen and phthalocyanine on it are presented in sections 3.3, 3.4, and 3.5, respectively.

# 2. COMPUTATIONAL DETAILS

The calculations carried out with all-electron linear combination of atomic orbitals (AE–LCAO) have been performed using the CRYSTAL software (Dovesi et al., 2014). In the AE–LCAO formalism implemented in CRYSTAL (Dovesi et al., 2014), the crystalline orbitals are expanded as a linear combination of atom centered Gaussian orbitals with s, p, or d symmetry (the basis set). AE calculations were performed in which there is no shape approximation to the potential or density.

DFT has been applied with the PBE (Perdew et al., 1996a), PBE0 (Perdew et al., 1996b), B3LYP (Becke, 1988), and HSE (Heyd et al., 2003) functionals. The main numerical approximation in these calculations is the choice of the basis set (BS). In order to systematically approach the BS limit, a hierarchy of all-electron basis sets, labeled as BS1, BS2, and BS3 (see **Table 1**), has been selected for O, Ti. The most complete basis set (BS3), which includes extended polarization d orbitals for both the Ti and O atoms, has been used for the calculations reported below if not indicated. Apart from the obvious consequences on the structural and electronic properties, the subtle effect of the basis set on the description of the gap states will be discussed at the end of section 3.3. For the H2Pc molecule, we used BS4 in order to reduce the BS superposition error. In the H2Pc-anatase analysis we perform only AE–LCAO calculations with the HSE functional by adding the D3 dispersion correction (Grimme and Krieg, 2010), which is crucial for this kind of interactions.

In the CRYSTAL (Dovesi et al., 2014) calculations, integration over the reciprocal space was carried out using Monkhorst-Pack (MP) meshes of 8 × 8 × 8 for the anatase bulk and 4 × 4 × 1 for the periodic slab representing the surface. The Coulomb and exchange series are summed directly in direct space and truncated using overlap criteria with thresholds of [7,7,7,7,14] (Pisani et al., 1988). The self-consistent field (SCF) algorithm was set to converge at the point at which the change in energy was < 10−<sup>7</sup> Hartree per unit cell. Structural optimisation


of both the bulk cell parameters and internal coordinates was performed using the Broyden-Fletcher-Goldfarb-Shanno scheme. Convergence was determined from the root-meansquare (rms) and the absolute value of the largest component of the forces. The thresholds for the maximum and the rms forces (the maximum and the rms atomic displacements) have been set to 0.00045 and 0.00030 (0.00180 and 0.0012) in atomic units. Geometry optimisation was terminated when all four conditions were satisfied simultaneously.

Calculations performed using the VASP code (Kresse and Furthmueller, 1996) include projected augmented wave (PAW) pseudopotentials (Bloechl, 1994; Kresse and Joubert, 1999) and a plane wave basis set with a cutoff of 500 eV. The SCF cycle stopping criterion was 10−<sup>6</sup> eV. The PBE exchange-correlation functional was supplemented with an onsite U = 4eV terms on the Ti 3d orbitals to describe better the electronic structure of the (101) anatase surface and the electron localization on Ti4<sup>+</sup> sites associated with the creation of its most common point defects. Ti PP is described by 4 valence electrons and O by 6 valence electrons. Previous studies on the anatase surface showed this U value provides the best description for the position of the defect-induced Ti gap states with respect to the conduction band minimum of TiO<sup>2</sup> (Cheng and Selloni, 2009; Aschauer et al., 2010; Stetsovych et al., 2015). We included 15 Å of vacuum in order to isolate the two sides of the slab. Defect geometries were accessed using the robust Conjugate Gradient algorithm. Structural optimisation was halted when the value of the largest component of the forces reached a threshold value of 0.01 eV/Å. The Brillouin zone was sampled by a 6 × 6 × 2 MP mesh in bulk anatase calculations, and at Ŵ point in surface calculations in order to reduce the computational cost. Surface electronic properties were extracted from static calculations featuring a 4 × 4 × 1 MP mesh.

# 3. RESULTS AND DISCUSSION

# 3.1. Bulk

The anatase structure belongs to the I41/amd tetragonal space group and the unit cell is defined by the lengths of lattice vectors a and c and the oxygen internal coordinate u. The primitive cell contains two atoms in the asymmetric unit: a Ti ion at (0,0,0) and an O ion at (0,0,u), in fractional coordinates. Anatase is an indirect band-gap semiconductor with an optical band gap of approximately 3.2 eV (measured by using UV adsorption spectroscopy) (Reddy et al., 2003). Anatase bulk lattice parameters have been relaxed with AE–LCAO and the DFT functionals: PBE, PBE0, B3LYP, HSE and HSED3. The computed structures and eigenvalue gaps (here, referred to as band gaps for convenience) are presented in **Table 2**, and compared to the structure deduced from X-ray diffraction and the optical band gap measured using UV adsorption spectroscopy (Reddy et al., 2003). It is clear from **Table 2** that the best agreement, for both lattice parameters and band gap, is obtained with the HSE functional. The results obtained within the HSED3 approach are also reported in **Table 2**, showing no significant changes with respect to pure HSE. The B3LYP lattice parameters and band gap are very similar to the ones previously

TABLE 2 | AE–LCAO and PP–PW relaxed lattice parameters and band gap for bulk anatase with respect to the experimental value measured by using X-ray diffraction and the adsorption spectrum (Reddy et al., 2003), respectively.


The percentage error with respect to the experiment is reported in parenthesis.

obtained in the literature (Labat et al., 2008; Sanches et al., 2014). The eigenvalue gap is an estimate of the fundamental band gap. When using HSE and HSED3, the 13% overestimation of the optical band gap is not unreasonable as the optical gap is expected to be less than the fundamental gap due to the exciton binding energy. A comparison of the DOS calculated with PP– PW (PBE and PBE+U) and AE–LCAO (PBE, B3LYP and HSE) is presented in **Figure 1**. All DOS reported are plotted relative to the CBM to aid comparison. It is important to highlight the good agreement between AE–LCAO and PP–PW approaches with the same functional (PBE). Both methods provide a very similar band gap and lattice parameters suggesting that the PP–PW and local basis set approaches provide a very similar description of the electronic structure in both the valence and conduction bands. The projected density of states (DOS) for bulk anatase performed with AE–LCAO (HSE) shows that the conduction band is largely made up of Ti 3d orbitals and the upper valence band of O 2p orbitals, as one expects from a simple ionic model of the bonding.

# 3.2. Stoichiometric (101) Surface

The (101) surface has two possible terminations, both resulting in stoichiometric, nonpolar surfaces. In this study we consider the low-energy termination (Sanches et al., 2014), where, as shown in **Figure 2A**, one of each of the following atoms are exposed per periodic cell: O2<sup>c</sup> ,O3<sup>c</sup> ,Ti5<sup>c</sup> and Ti6<sup>c</sup> producing the characteristic saw-tooth (101) surface structure. The 2c, 3c, 5c, and 6c indexes represent the two, three, 5 and 6 fold– coordinated atoms, respectively. The surface was modeled with a 2D periodic slab that includes four TiO<sup>2</sup> trilayers with the in– plane periodicity given by a 1×3 supercell of the conventional surface cell.

The fact that both the AE–LCAO and PP–PW approaches provide a very similar description at the PBE level suggests that the plane-wave and local Gaussian orbital basis sets used are both well converged and that the pseudopotentials are describing the valence-core interactions adequately. This also facilitates the comparison of the results of PP–PW-PBE+U calculations with AE–LCAO-HSE and -B3LYP calculations, as we can be confident that the numerical approximations used are equivalent. This allows us to assess the accuracy of the widely used DFT+U approach.

An analysis of the DOS for the stoichiometric anatase (101) calculated with AE–LCAO and HSE (**Figure 3**) exhibits a similar structure to that seen in the bulk crystal. The conduction band is mainly composed of Ti 3d orbitals and the upper valence band of O 2p orbitals. In order to separate the surface contribution from that of the bulk, the DOS have been plotted in four different ways: projected (PDOS) on the most important atoms of the top layer (O2<sup>c</sup> , O3<sup>c</sup> , Ti5<sup>c</sup> , and Ti6c) (**Figure 3A**), PDOS on the two top trilayers (**Figure 3B**), total DOS (**Figure 3C**), and PDOS on the two bottom trilayers (**Figure 3D**).

The computed DOS of the stoichiometric surface using the PP–PW (PBE+U) and AE–LCAO (HSE and B3LYP) approximations are compared in **Figure 3**. It can be seen that the global hybrid (B3LYP) and screened exchange (HSE) approaches produce very similar band width, shape and band gap. The empirical PBE+U method yields similar band widths but a significantly smaller band gap.

In **Table 3** the bond distances and atom height (with respect to the position of the first oxygen atom of the second trilayer) using PP–PW (PBE+U) and AE–LCAO (HSE and B3LYP) are reported for the selected atoms of the surface layer. The surface structure is found to be relatively insensitive to both the numerical approximation and to the exchange correlation functional used.

# 3.3. Oxygen Vacancy on (101) Anatase

A simple model of the reduced anatase surface can be generated by oxygen desorption induced by heating. Although, as the least coordinated, the outermost O2<sup>c</sup> oxygen atoms might be expected to be the first to leave the surface, previous work (Setvin et al., 2013) has shown that these surface vacancies are unstable and quickly evolve to form subsurface vacancies. We therefore consider a subsurface oxygen vacancy model where an oxygen atom is removed from the second trilayer of the surface (see **Figure 2C**).

The extra two electrons left after the vacancy formation could form an F-center at the vacant lattice site or enter the Ti 3d band and, potentially, localize to form Ti3<sup>+</sup> species. A number of previous theoretical studies have described the formation of these localized states (Lindan et al., 1997; Muscat et al., 1999; Chen et al., 2001; Deak et al., 2015). The ground state considered here corresponds to the occupation of a localized 3d orbital on each of two next-nearest neighbors, a Ti5<sup>c</sup> atom in the first trilayer and a Ti6<sup>c</sup> in the third trilayer (cyan atoms in **Figure 2C**). A previous PP–PW (PBE+U) study also suggested that this configuration for the reduced titanium ions is the most stable (Stetsovych et al., 2015). This is also the case in calculations based on hybrid functionals. For example, the configuration where two nearest neighbor Ti ions are reduced (magenta atoms in **Figure 2C**) is a local energy minimum but has an energy 0.82 eV higher than the one described above. The reduction of two Ti4<sup>+</sup> ions is confirmed by the spin–polarized DOS presented in **Figure 4**. The two Ti3<sup>+</sup> ions have a magnetic moment of 0.90 and 0.91 µ<sup>B</sup> for AE–LCAO (3LYP), 0.92 and 0.91 µ<sup>B</sup> for AE–LCAO (HSE) and 0.91 and 0.89 µ<sup>B</sup> for PP– PW (PBE+U). The new localized states, mainly formed by Ti

represented in gray, the O2<sup>c</sup> bonded to H in green and the spin polarized Ti5<sup>c</sup> in cyan, (C) O monovacancy, where the spin polarized Ti5<sup>c</sup> atoms are represented in cyan (minimum energy configuration) and yellow (alternative configuration of the spin).

3d orbitals from the Ti6<sup>c</sup> in the third trilayer and the Ti5<sup>c</sup> in the first trilayer, appear at 1.0 and 1.3 eV below the conduction band for AE–LCAO (HSE). PP–PW (PBE+U) gives a very similar result (1.1, 1.3 eV), while AE–LCAO–B3LYP predicts those states to be deeper in the band gap (1.2, 1.5 eV). The AE–LCAO (HSE) calculation is in good agreement with photoemission

FIGURE 3 | (Color online) Stoichiometric anatase (101) : (A) PDOS on the most important atoms (O2<sup>c</sup> ,O3<sup>c</sup> ,Ti5<sup>c</sup> and Ti6<sup>c</sup> ) of the top layer by using AE–LCAO (HSE), (B) PDOS on the two top trilayers by using AE–LCAO (HSE), (C) DOS by using AE–LCAO (HSE), (D) PDOS on the two bottom trilayers by using AE–LCAO (HSE) (bulk-like region), (E) DOS comparison of AE–LCAO (B3LYP) and PP–PW (PBE+U), and (F) DOS comparison of AE–LCAO (HSE) and PP–PW (PBE+U).

TABLE 3 | Bond distance and height of relevant atoms in the top layer (Å) for AE–LCAO and PP–PW in the anatase (101) stoichiometric surface, hydrogen adsorption and oxygen monovacancy.


The height of the atoms is defined with respect to the position of the first oxygen atom of the second trilayer in order to allow a direct comparison.

experiments, which identify a state at around 1 eV bellow the CBM (Thomas et al., 2007).

The relaxed structure for the subsurface vacancy can be seen in **Figure 2C**, where the Ti3<sup>+</sup> ions depicted in cyan. The charge localization has a profound influence on the structure, as shown by the bond distances and relative height of the key atoms in the top layer in **Table 3**. A significant decrease of the Ti6c-O2<sup>c</sup> bond is seen. The Ti3<sup>+</sup> 6c -O2<sup>c</sup> bond is reduced by 0.1 Å, with respect to the Ti4<sup>+</sup> 6c -O2<sup>c</sup> bond. There are also significant relaxations in the reduced top Ti6<sup>c</sup> , that changes its coordination from 6 to 5 and moves up by 0.5 Å. In the presence of two polarons and a vacancy, the polaronic distortions might interact and therefore affect the total energy, geometry and position of the gap state. This structural feature appears consistently in all of the calculations presented here and can thus be considered to be insensitive to the approximation of the electronic exchange and correlation.

Previous combined STM/AFM experiments show a strong distortion of the O2<sup>c</sup> site above the vacancy which is lifted up by 0.3 Å (Stetsovych et al., 2015). PP–PW (PBE+U) and AE– LCAO (B3LYP) support this result with the O2<sup>c</sup> moving up by 0.28 and 0.31 Å, respectively. AE–LCAO (HSE) reproduces this trend with a normal displacement of 0.38 Å, identical to the PP–PW–PBE+U result.

The four–layer slab used so far is clearly suitable to describe the relaxations in the top layers associated with the creation of the surface and the H–defect. In the case of the subsurface vacancy, where one of the reduced Ti3<sup>+</sup> ions is located in the third trilayer, one may wonder if an additional trilayer is needed

FIGURE 4 | (Color online) O monovacancy on anatase (101) : (A) PDOS on the most important atoms (O2<sup>c</sup> ,O3<sup>c</sup> ,Ti5<sup>c</sup> , and Ti6<sup>c</sup> ) of the outermost layer by using AE–LCAO (HSE), (B) PDOS on the two top trilayers by using AE–LCAO (HSE), (C) DOS by using AE–LCAO (HSE), (D) PDOS on the two bottom trilayers by using AE–LCAO (HSE) (bulk-like region), (E) DOS comparison of AE–LCAO (B3LYP) and PP–PW (PBE+U), and (F) DOS comparison of AE–LCAO (HSE) and PW (PBE+U).

in order to relax the local strain induced by the larger ion, as the fourth trilayer has been kept fixed. This strain can be more easily accommodated by atomic relaxations when the reduced ion is located in the first trilayer. This is consistent with the fact that the gap states associated with the two Ti3<sup>+</sup> ions have different energy, the one associated with the Ti3<sup>+</sup> in the top trilayer being lowered in energy. Thus, we have optimized the structure of a five–trilayer slab using AE–LCAO–HSE and keeping fixed the fifth (bottom most) trilayer. The structural analysis shows no significant differences, with very similar bond distances among the key atoms and the same surface relaxation pattern, where the O2<sup>c</sup> above the vacancy moves up again by 0.4 Å. Regarding the electronic properties, a comparison of the DOS calculated with four and five trilayers (see **Figure 5A**) shows that they are nearly identical. The position of the new localized states and local spin magnetic moment (0.94 and 0.93 µB) are also very similar to the ones calculated with four trilayers. These results confirm that when the donated electrons are accommodated in the described positions, a four–trilayer slab provides an adequate description.

The localized states associated with the Ti3<sup>+</sup> ions provide a rigorous test of the flexibility of the local Gaussian basis sets used. The nature of the localized d-state is very different to that of the delocalized d-band of the undoped system. We therefore expect some sensitivity to the flexibility of the basis set used to describe the Ti-d derived orbitals. In **Figure 5B** the DOS in the band gap is displayed for both the oxygen subsurface vacancy with the two basis sets, BS2 and BS3. Although the calculations using BS2 reproduce the surface structure very accurately, the computed position of the defects state relative to the CBM differs by 6% to that predicted by the more flexible BS3. The computational time increases by 33% by using BS3 instead of BS2.

In **Table 4**, AE–LCAO (B3LYP and HSE) calculations show an increase of the surface band gap of around 2% when an oxygen monovacancy is present. This result can be due again to the approximation of considering few layers to simulate the real surface or to the interaction of polaronic distortions as explained above. AE–LCAO (HSE) also describes very similar peaks for the gap state corresponding to the Ti3<sup>+</sup> ion on the outermost layer in both defects, as it can be seen in **Figure 5C**.

TABLE 4 | Band gap for the stoichiometric anatase (101) and the defective surface by using AE–LCAO and PP–PW.


The heterogeneous photocatalysis is based on the ability of photocatalysts to absorb the light energy required to generate electron–hole pairs for a surface reaction. TiO<sup>2</sup> can only absorb ultraviolet (UV) light, but its optical properties can be enhanced by defect engineering. Both oxygen vacancies and hydrogen adsorption give rise to local states below the CB edge, which can extend the light absorption of TiO<sup>2</sup> from the UV to the visible ranges. The formation energy is calculated by using the formula:

$$E\_f = E\_{def} + 1/2E\_{\text{O2}} - E\_{model},\tag{1}$$

where Edef and Enodef are the energies of the defective and non defective anatase(101), respectively, and EO<sup>2</sup> is the energy of the oxygen molecule. The calculated formation energy with AE– LCAO (HSE) is 2.6 eV, in good agreement with the value 2.9 eV, given in the literature (Haa and Alexandrova, 2016).

# 3.4. H Adsorption on (101) Anatase

The adsorption mechanism for H investigated here corresponds to the R arrangement in the DFT study by Leconte et al. (2002), since this was found to be the most favorable adsorption mode. A hydrogen atom is adsorbed onto a surface oxygen atom, upon which a reduction of a nearby titanium atom takes place. In our calculation, a neutral H atom is adsorbed to O2<sup>−</sup> resulting in the formation of OH<sup>−</sup> and the donation of an electron to the surface CB.

This electron is transferred to the oxide lattice, and occupies one of the localized 3d orbitals of the nearest neighbor Ti5<sup>c</sup> ions. As a result, this Ti4<sup>+</sup> ion is reduced and converted to a Ti3<sup>+</sup> ion. The analysis of the spin (unpaired electrons) population confirms the transfer of an electron onto the Ti ion. The magnetic moment is 0.920 µ<sup>B</sup> for PP–PW

FIGURE 6 | (Color online) H adsorption on anatase (101) : (A) PDOS on the most important atoms (O2<sup>c</sup> ,O3<sup>c</sup> ,Ti5<sup>c</sup> , and Ti6<sup>c</sup> ) of the top layer by using AE–LCAO (HSE), (B) PDOS on the two top trilayers by using AE–LCAO (HSE), (C) DOS by using AE–LCAO (HSE), (D) PDOS on the two bottom trilayers by using AE–LCAO (HSE) (bulk-like region), (E) DOS comparison of AE–LCAO (B3LYP) and PP–PW (PBE+U), and (F) DOS comparison of AE–LCAO (HSE) and PP–PW (PBE+U).

FIGURE 7 | (Color online) (A) side view and (B) top view of the optimized structure for the adsorption of H2Pc on anatase (101) by using AE–LCAO (HSE) and centering the molecule on the O2<sup>c</sup> row. We represent Ti in blue, O in red, N in magenta, C in green, and H in orange.

FIGURE 8 | (Color online) (A) side view and (B) top view of the optimized structure for the adsorption of H2Pc on anatase (101) by using AE–LCAO (HSE) and centering the molecule on the Ti5<sup>c</sup> row. We represent Ti in blue, O in red, N in magenta, C in green, and H in orange.

(PBE+U), 0.957 µ<sup>B</sup> for AE–LCAO (HSE) and 0.944 µ<sup>B</sup> for AE–LCAO (B3LYP). This localization of the electron is due to on–site electronic correlation and therefore one could expect it to be much less when compared to the LDA or GGA approximations.

In **Figure 2B**, the final relaxed structure using the AE– LCAO method and the HSE functional is displayed. The bond distances and atom heights for the key atoms in the top layer calculated with different exchange-correlation functionals are compared in **Table 3**. All the functionals consistently predict the O2<sup>c</sup> atom bonded to H that moves outwards by 0.13 Å. The height of the H atom above the O2<sup>c</sup> plane is directly observable in non–contact atomic force microscopy (NCAFM) images (Stetsovych et al., 2015) resulting in a value of 0.11 Å, which is in good agreement with the one predicted here.

The calculated DOS displayed in **Figure 6** –analogous to **Figure 3** for the stoichiometric surface– shows the presence of a localized state 1.2 eV below the conduction band minimum (CBM) for AE–LCAO with HSE (and 1.4 eV for AE–LCAO with B3LYP and PP–PW with PBE+U). This new peak in the band gap, mainly formed by Ti 3d states coming from the CBM, is the signature of the reduction of one Ti4<sup>+</sup> ion of the surface layer to Ti3+. This behavior has been observed in UV absorption spectra and documented in previous calculations of the K atom adsorbed on TiO<sup>2</sup> surface, where the donation of the electron to TiO<sup>2</sup> was localized in a similar state (Muscat et al., 1999). It is remarkable that, in spite of the different position of the valence band maximum, PBE+U predicts a position of that peak with respect to the CBM in agreement with hybrid functionals. In previous infrared spectroscopy experiments and theoretical calculation based on GGA+U on rutile reduced by hydrogen, these gap states have been reported to be in the range 0.75-1.18 eV below the CBM (Cronemeyer, 1959; Finazzi et al., 2008; Islam et al., 2011; Antila et al., 2015), which is consistent with the data presented here. Therefore, both the AE–LCAO and PP–PW implementations predict a similar position of the band gap state due to the localization of the extra electron on the 3d orbital of a Ti5<sup>c</sup> of the top layer.

It has been stated before that the introduction of defects on TiO<sup>2</sup> can lead to a modification of the band gap value (Naldoni et al., 2012; Mehta et al., 2016). In **Table 4**, a comparison of the computed electronic structure of the stoichiometric and defective surface is presented and the band gap is indicated. The adsorption of hydrogen on anatase (101) leads to a small reduction of the band gap of around 2 % in both the HSE and B3LYP approximations (this result agrees with the value of 1.8% found by diffusion reflectance spectroscopy Mehta et al., 2016). In calculations based on PBE+U the inclusion of defects produces no change in the band gap.

The formation energy is calculated by using the formula:

$$E\_f = E\_{slab} + E\_{H2}/2 - E\_{H/slab}.\tag{2}$$

where Eslab and EH/slab are the energies of the slabs without and with hydrogen adsorbed on the surface, respectively, and E<sup>H</sup> is the energy of atomic hydrogen. The calculated formation energy with AE–LCAO (HSE) is 2.26 eV, in good agreement with the value 2.31 eV, given in the literature (Islam et al., 2011).

# 3.5. Adsorption of Unsubstituted Phthalocyanine on Anatase(101)

The data presented above suggested that the AE–LCAO-HSE approach is generally the most satisfactory for describing the electronic structure of the stoichiometric and reduced anatase surface. We therefore use this approach to consider the adsorption of H2Pc on anatase (101). In this calculation we also include the effects of London dispersion interactions through the empirical Grimme D3 correction, as these are likely to play a role in the molecule-surface binding. Initial geometries with the molecule placed centered on O2<sup>c</sup> or Ti5<sup>c</sup> rows of the anatase surface have been used in order to explore alternative binding sites on the anatase surface. The optimized structures for both adsorption sites are shown in **Figures 7**, **8**. Local minima are found at both sites. It is perhaps surprising to observe the significant changes in the molecular structure of H2Pc induced when it is centered on O2<sup>c</sup> . The molecule becomes non–co–planar and the C-H rings move away from the surface by 0.3–0.4 Å with respect to the central atoms of the molecule, which makes clear the strong repulsion between these rings and the O2<sup>c</sup> . The reason for this behavior can be the strength of the C-H bond which makes it relatively unreactive. However, the N atoms from the molecule interact strongly with the surface Ti atom, resulting in the N atom placed just on top of the Ti5<sup>c</sup> (N1 in **Figure 7B**) moving closer to the surface by 0.3 Å , with respect to the rest of the N in the molecule.

The central hydrogens of the molecule form a hydrogen bond of 1.8 Å with the neighboring nitrogens, while one of them (H1) tries to attach to the O2<sup>c</sup> of the surface. These effects make the O2<sup>c</sup> (O1) to move up by 0.1 Å . On the contrary when the molecule is centered on Ti5<sup>c</sup> (**Figure 8**), its remains relatively flat with just variations of 0.1 Å in height between the different atoms. For both sites, H2Pc is adsorbed with a board altitude that reaches approximately 2.7 Å , in agreement with the result obtained by using STM on rutile (Godlewski and Szymonski, 2013). The calculated adsorption energy is of –3.18 eV with the molecule centered on the oxygen row and of −3.54 eV with the molecule centered on the titanium row. Therefore, the adsorption of H2Pc is more favorable if the molecule is centered on Ti5<sup>c</sup> by 0.36 eV. It is clear that when the molecule is adsorbed centered in the O2<sup>c</sup> the repulsion between the surface and the C-H rings is stronger than the bonds formed.

# 4. CONCLUSIONS

The AE–LCAO approach is a robust and efficient ab initio method to correctly describe the defects and adsorption of molecules on oxide surfaces. In this work, we have documented the performance of two widely used methods based on Density Functional Theory: linear combination of atomic orbitals (AE– LCAO) and plane waves (PP–PW) to describe the properties of the stoichiometric and defective TiO2(110) anatase surface. Full relaxation of stoichiometric and defective anatase (101) surfaces has been carried out using the PBE functional and the hybrid PBE0, B3LYP and HSE functionals. For the PP–PW calculations, only the PBE and PBE+U levels of theory have been considered due to the high computational cost of using hybrid functionals for systematic studies. Hybrid functionals, and, in particular, HSE, lead to a computed band gap in excellent agreement with the experimental data. PBE+U underestimates the gap by 20 % but the predicted gap states induced by the defect with respect to the bottom of the conducting band are in good agreement with the hybrid functionals. An understanding of the defective anatase surface is very important for many technological applications. It has been argued that vacancies are very important in the photocatalytic property of surfaces. The formation of mid-gap states, as the ones described here, could improve the photoactivity of the system in the visible light region. Once we have proved that the AE–LCAO approach with the HSE functional is able to accurately describe this surface, we went one step further by studying the adsorption of the phthalocyanine H2Pc on anatase (101). H2Pc was placed both on top of the oxygen and titanium rows of anatase (101), with the latter lower in energy by 0.36 eV. This work presents an ab initio study of the adsorption of H2Pc on anatase (101), and makes clear the influence of the adsorption site in the phthalocyanine-anatase interaction. The AE–LCAO method presented here opens a way to correctly describe the defects and adsorption of molecules on oxide surfaces by using hybrid exchange functionals at a reasonable computational cost, which will be a valuable tool in order to improve the efficiency of dye-sensitized solar cells.

# AUTHOR CONTRIBUTIONS

All authors listed have made a substantial, direct and intellectual contribution to the work, and approved it for publication.

# REFERENCES


Graetzel, M. (2001). Nature 414:338. doi: 10.1038/35104607

Grimme, S., Antony, J., Ehrlich, S., and Krieg, H. (2010). J. Chem. Phys. 132:154101. doi: 10.1063/1.3382344

# FUNDING

We acknowledge financial support from the Spanish MINECO through projects MDM-2014-0377, MAT2014-54484-P, and MAT2017-83273-R and a Juan de la Cierva contract (RM-C). Computer time was provided by the Spanish Supercomputing Network (RES, Spain) at the Magerit Supercomputer (Madrid, Spain) and via the membership of the UK's HEC Materials Chemistry Consortium (GM and NH), which is funded by EPSRC (EP/L000202). This work used the ARCHER UK National Supercomputing Service (http://www.archer.ac.uk).


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Martínez-Casado, Todorovi´c, Mallia, Harrison and Pérez. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# The Role of Cation-Vacancies for the Electronic and Optical Properties of Aluminosilicate Imogolite Nanotubes: A Non-local, Linear-Response TDDFT Study

Emiliano Poli <sup>1</sup> \*, Joshua D. Elliott 2,3,4, Sergey K. Chulkov <sup>5</sup> , Matthew B. Watkins <sup>5</sup> \* and Gilberto Teobaldi 6,7,8 \*

<sup>1</sup> The Abdus Salam Center for Theoretical Physics, Condensed Matter and Statistical Physics Department, Trieste, Italy, <sup>2</sup> Dipartimento di Fisica e Astronomia "Galileo Galilei", Università degli Studi di Padova, Padova, Italy, <sup>3</sup> CNR-IOM DEMOCRITOS, Consiglio Nazionale delle Ricerche-Istituto Officina dei Materiali, Trieste, Italy, <sup>4</sup> School of Chemical Engineering and Analytical Science, The University of Manchester, Manchester, United Kingdom, <sup>5</sup> School of Mathematics and Physics, University of Lincoln, Brayford Pool, Lincoln, United Kingdom, <sup>6</sup> Daresbury Laboratory, Scientific Computing Department, Science and Technology Facilities Council, Warrington, United Kingdom, <sup>7</sup> Beijing Computational Science Research Centre, Beijing, China, <sup>8</sup> Stephenson Institute for Renewable Energy and Department of Chemistry, University of Liverpool, Liverpool, United Kingdom

### Edited by:

Javier Carrasco, CIC energigune, Spain

### Reviewed by:

Rosendo Valero, University of Barcelona, Spain Hélio Anderson Duarte, Federal University of Minas Gerais, Brazil

### \*Correspondence:

Emiliano Poli epoli@ictp.it Matthew B. Watkins mwatkins@lincoln.ac.uk Gilberto Teobaldi gilberto.teobaldi@stfc.ac.uk

### Specialty section:

This article was submitted to Physical Chemistry and Chemical Physics, a section of the journal Frontiers in Chemistry

Received: 15 November 2018 Accepted: 18 March 2019 Published: 10 April 2019

### Citation:

Poli E, Elliott JD, Chulkov SK, Watkins MB and Teobaldi G (2019) The Role of Cation-Vacancies for the Electronic and Optical Properties of Aluminosilicate Imogolite Nanotubes: A Non-local, Linear-Response TDDFT Study. Front. Chem. 7:210. doi: 10.3389/fchem.2019.00210 We report a combined non-local (PBE-TC-LRC) Density Functional Theory (DFT) and linear-response time-dependent DFT (LR-TDDFT) study of the structural, electronic, and optical properties of the cation-vacancy based defects in aluminosilicate (AlSi) imogolite nanotubes (Imo-NTs) that have been recently proposed on the basis of Nuclear Magnetic Resonance (NMR) experiments. Following numerical determination of the smallest AlSi Imo-NT model capable of accommodating the defect-induced relaxation with negligible finite-size errors, we analyse the defect-induced structural deformations in the NTs and ensuing changes in the NTs' electronic structure. The NMR-derived defects are found to introduce both shallow and deep occupied states in the pristine NTs' band gap (BG). These BG states are found to be highly localized at the defect site. No empty defect-state is modeled for any of the considered systems. LR-TDDFT simulation of the defects reveal increased low-energy optical absorbance for all but one defects, with the appearance of optically active excitations at energies lower than for the defect-free NT. These results enable interpretation of the low-energy tail in the experimental UV-vis spectra for AlSi NTs as being due to the defects. Finally, the PBE-TC-LRC-approximated exciton binding energy for the defects' optical transitions is found to be substantially lower (up to 0.8 eV) than for the pristine defect-free NT's excitations (1.1 eV).

Keywords: inorganic nanotubes, imogolite nanotubes, defects, DFT, TD-DFT, optical properties, photo-catalysis

# INTRODUCTION

By altering the local structure and potential experienced by electrons in insulating or semiconducting materials, point-defects such as atomic-vacancies or interstitial atoms can dramatically affect the physical and chemical properties of the defective host material. Control and tuning of the properties brought about by the presence of defects rest on accurate understanding of

**67**

their properties at the atomic-scale, which has long motivated strong efforts toward the development and application of accurate, yet computationally viable, approaches for the simulation of defects in solids (Shluger et al., 2003; Ganduglia-Pirovano et al., 2007; Lany and Zunger, 2008; Neugebauer et al., 2014; Walsh and Zunger, 2017; Yin et al., 2018).

The ever-increasing demand for sustainable energy keeps driving research in the development of alternative sustainable technologies for energy production and its efficient use. Photocatalytic materials (photo-catalysts, PCs) are central to these efforts since they can exploit solar light to access alternative, highly selective, reaction routes for fuels and chemicals production (Kamat, 2007; Sastre et al., 2011, 2012a,b,c; Dietl et al., 2012; de Richter et al., 2013; Maeda, 2013; Baltrusaitis et al., 2014; Fresno et al., 2014; Ismail and Bahnemann, 2014), or environmental remediation (Fujishima et al., 2008; Pelaez et al., 2012; Habisreutinger et al., 2013; Navalón et al., 2013). In this context, the search for affordable high performance, highly selective PCs is an ongoing quest, with growing interest in and appreciation of the potential of aluminosilicate and silicate substrates for selective photo-catalysis (Sastre et al., 2012a,b; Alarcos et al., 2017; Murcia-López et al., 2017). As for any other material, also the ground- and excited-state properties of aluminosilicate and silicate-PCs can be significantly affected by the unavoidable presence of defects (Barrer, 1982; Novatski et al., 2008; Tandia et al., 2011).

The growing research in the photo-catalytic potential of 3D nano- and meso-porous silicates (Sastre et al., 2012a,b; Alarcos et al., 2017; Murcia-López et al., 2017), and parallel substantial progress in the synthesis and characterization of 1D-structured aluminosilicate (AlSi) nanotubes (NTs) based on the imogolite (Imo) structure (Wada et al., 1979; Barron et al., 1982; Theng et al., 1982; Mukherjee et al., 2005, 2007; Levard et al., 2008, 2010; Kang et al., 2010, 2011, 2014; Bottero et al., 2011; Yucelen et al., 2011, 2012b; Zanzottera et al., 2012a,b; Bonelli et al., 2013; Amara et al., 2015; Monet et al., 2018), have prompted recent exploratory computational studies of the potential of Imo-NTs for photo-catalytic applications (Teobaldi et al., 2009; Zhao et al., 2009; Poli et al., 2015, 2016; Elliott et al., 2016).

AlSi Imo-NTs are structurally analogous to the naturally occurring hydrous-aluminosilicate imogolite (Cradwick et al., 1972). Their walls consist of a single layer of octahedrally coordinated aluminum hydroxide with pendant tetrahedral silanol (Si–OH) groups facing the tube cavity (**Figure 1**). The stoichiometric formula of the unit cell is (Al2SiO7H4)N, with N being the number of radially non-equivalent aluminum atoms along the NT circumference (an even number for symmetry reasons). In recent years, control over the structure and functionalization of Imo-NTs has improved greatly. Solutionbased synthetic routes to produce single-walled AlSi NTs of controllable radius and length have been defined (Wada et al., 1979; Barron et al., 1982; Theng et al., 1982; Mukherjee et al., 2005, 2007; Levard et al., 2008, 2010; Kang et al., 2010; Yucelen et al., 2011, 2012b) and important post-synthetic, selective functionalization of the outer or inner surface of AlSi NTs (Levard et al., 2010; Bottero et al., 2011; Kang et al., 2011, 2014; Zanzottera et al., 2012a,b; Bonelli et al., 2013; Amara et al., 2015) reported. Further progress has been made in the direct synthesis of methylated (AlSi-Me) (Bottero et al., 2011; Bonelli et al., 2013) or aminated (AlSi-Am) (Kang et al., 2014) AlSi NT derivatives, with methyl (–CH3) or amine (–NH2) moieties in the NT cavity as well as in the creation of hybrid methylated Al(Si/Ge)-Me NTs with tuneable Si/Ge ratio (Amara et al., 2015). Selective amination of the outer surface of AlSi-Me NTs has also been reported (Zanzottera et al., 2012a,b). This synthetic flexibility in turn entails the possibility of widely tuning the NT's properties. Work along this line is ongoing and has led to observation of very encouraging performances of Imo-NTs for chemical separation (Levard et al., 2010; Bottero et al., 2011; Kang et al., 2011, 2014; Zanzottera et al., 2012a,b; Bonelli et al., 2013; Amara et al., 2015), as support for PCs (Katsumata et al., 2013) and as component in hybrid nanocomposites (Marzan and Philipse, 1994). In addition, recent progress in the synthesis of Fe-doped AlSi (and aluminogermanate) NTs (Ma et al., 2012; Arancibia-Miranda et al., 2014; Avellan et al., 2014; Castro et al., 2016; Shafia et al., 2016) has succeeded in introducing absorption of visible (and UV) light, opening new avenues for exploration of the potential of Imo-NTs for both UV and visible light photo-catalytic applications. Recent advances in quantitative structural resolution for (single-walled) Imo-NTs by X-ray scattering (Monet et al., 2018) is expected to further accelerate progress in the development of technological applications based on Imo-NTs. A recent review on Imo-NTs can be found in Paineau (2018).

In spite of the progress in the synthesis and sample-averaged characterization of Imo-NTs (Wada et al., 1979; Barron et al., 1982; Theng et al., 1982; Mukherjee et al., 2005, 2007; Levard et al., 2008, 2010; Kang et al., 2010, 2011, 2014; Bottero et al., 2011; Yucelen et al., 2011, 2012b; Zanzottera et al., 2012a,b; Bonelli et al., 2013; Amara et al., 2015; Monet et al., 2018; Paineau, 2018), the current understanding of point-defects in Imo-NTs is far from complete. To the best of our knowledge, the only experimental investigation of point defects in AlSi Imo-NTs reported to date appears in Yucelen et al. (2012a). By combining different NMR techniques, the authors were able to identify five dominant defect-structures in the AlSi NTs. As explained in Yucelen et al. (2012a), these are based on cation (Al or Si) vacancies, cation (Al→Si) substitution, additional nonstoichiometric Si atoms, and combinations of vacancies with additional atoms (see **Figure 1** for the atomistic models of these defects). Analysis of the experimental results led to the estimation that roughly 16% of the Si atoms in the AlSi NTs are involved in a defect structure. Given the size of Imo-NTs and the challenges in accurate modeling of the proposed defects at the experimental concentrations, to date, these defects have never been studied at Density Functional Theory (DFT) level. Although force-field studies of these defects have appeared in the literature (Liou and Kang, 2016), elucidation of their role for the (defective) AlSi NTs' electronic and optical properties is yet to be accomplished. To this end, and to sustain the growing interest in the optical properties of Imo-NTs and their Fe-doped variants (Arancibia-Miranda et al., 2014; Avellan et al., 2014; Shafia et al., 2015, 2016; Castro et al., 2016), here we present a computational study of

the role of the NMR-inferred defect structures for the optical and electronic properties of AlSi NTs.

Compromising between AlSi NTs' challenging size, the necessity of simulating models sufficiently large to enable full relaxation of the defects and their coexistence with unperturbed lattice regions in the NT-models as well as the far from solved challenges in accurate simulation of excited states properties for solid-state systems (Gonze et al., 1995; Ghosez et al., 1997; Onida et al., 2002; Bernasconi et al., 2011, 2012; Tomic et al., 2014; Ullrich and Yang, 2014; Casida and Huix-Rotllant, 2016; Strand et al., 2019), here we turn to a recent implementation (Strand et al., 2019) of a non-local [PBE0-TC-LRC (Guidon et al., 2009)], linear-response [LR (Casida, 2009)] time-dependent DFT (TDDFT) approach in the Tamn-Dancoff approximation [TDA (Hirata and Head-Gordon, 1999)], that enables computationally viable LR-TDA-TDDFT simulation for model systems of roughly 1,000 atoms. Although of finite accuracy, especially for the optical properties of defect-free NTs (vide infra), the approach provides an arguably reasonable and timely first step in the elucidation of the optical properties of Imo NTs and defects therein, contributing to research in Imo-NTs and, more generally, to the ongoing work on accurate yet viable simulation of excitestate properties in defective solids, NTs included.

The paper is organized as follows: after presentation of the atomic composition and structure for the NMR-inferred defects and computational methods used (section Methods), we present our results in section Results and Discussion. First, we numerically determine the smallest simulation cell capable of accommodating the defect-induced distortions in the presence of effectively unperturbed lattice regions in the NT model (section Determination of the Supercell Size). Structural analysis of the optimized models is presented in section Optimized Defect Structures, followed by characterization of their electronic properties (section Electronic Structure of the Defective NTs). The calculated optical absorption properties are presented in section Optical properties of defects and discussed with respect to available experimental UV-vis spectra for AlSi NTs. We finally present our conclusions in section Conclusions.

# METHODS

# The Atomic Structure of the NMR-Inferred Defects

For our computational study, we started from the atomic structure of the five main defects inferred from NMR experiments on the AlSi Imo-NTs (Yucelen et al., 2012a). They can be seen in **Figure 1** and are described and labeled as follows:

• **D1** is a single Al-vacancy. By disrupting the gibbsite Al(OH)<sup>3</sup> backbone of the NT, D1 introduces additional Al-OH dangling groups in the NTs framework, which would otherwise not be present in the defect-free NTs.

TABLE 1 | Stoichiometric formula and resulting charge for the considered defect models.



**Table 1** reports the stoichiometry of the defect models for the adopted x3 periodic repeat of the NT (vide infra, section Determination of the Supercell Size). Given the water-phase synthesis of the AlSi NTs (Mukherjee et al., 2005, 2007; Levard et al., 2008), it is inevitable that counter-cations or protons (in the water solution) will compensate the negative charge of D3, D4, and D5. These considerations prompted us to model three additional structures, one for each negatively charged defect, neutralizing the original negative excess charge by protonation of the defect-site. These structures, labeled **h-D3**, **h-D4** and **h-D5** are also displayed in **Figure 1**. Inclusion in the simulation of larger counter-cations [possibly Keggins ions as experimentally observed during the NT synthesis (Casey, 2006)] was not considered due to the prohibitive computational cost of this approach.

# Computational Methods

### Geometry Optimization

Given the extended dimension of the systems considered (∼1,000 atoms) all the defect structures were optimized using the Linear Scaling DFT program ONETEP (Skylaris et al., 2005; Haynes et al., 2006; Hine et al., 2009, 2011) with the PBE (Perdew et al., 1996) approximation to the exchange-correlation (XC) functional, and separable (Kleinman-Bylander) normconserving pseudopotentials (Gonze et al., 1991) for the atomic core levels. The adopted kinetic energy cutoff was 1,000 eV, and 4 (9) valence Non-Orthogonal Wannier Functions (NGWFs) were used for the O (Al, Si) atoms. 1 NGWF was used for the H atoms. In all cases, no truncation of the density kernel (K αβ) was enforced. The localization radius for the valence NGWFs was 8 Bohr. All simulations were performed with periodic boundary conditions ensuring at least 15 Å vacuum separation between replicated images along the non-periodic directions. Along the periodic direction of the NT, the simulation cell contained threefold (x3) replicas of the NT along its axis (optimized 8.665 Å period), yielding models of ∼1,000 atoms. This choice is extensively discussed and motivated in section Determination of the Supercell Size. Geometry-relaxations were performed via the quasi-Newton optimization scheme based on the Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm (Pfrommer et al., 1997). All the atoms of the defective NT-models were left free to relax. The geometry optimization threshold for the atomic forces was 0.05 eV/Å. Consistent with the absence of any dangling bonds (vide infra), singlet spin-states were tested to be energetically favored over alternative spin-polarized solutions.

### Hubbard-Corrected, Non-local DFT, and LR-TDDFT Simulations

Following geometry relaxation, the optimized defect models were used for single-point non-local hybrid DFT and LR-TDDFT simulations using the CP2K package (VandeVondele et al., 2005). CP2K uses a mixed Gaussian and plane wave basis set, where atom-centered Gaussian orbitals are used to represent the wavefunction, while the electronic density is represented by an expansion of plane waves.

Through all our calculations, we used Goedecker-Teter-Hutter (GTH) norm-conserving pseudopotentials (Goedecker et al., 1996) to model the electron-ion interactions. We adopted the molecularly optimized (MOLOPT) DZVP-SR basis set (VandeVondele and Hutter, 2007) for all atoms and the auxiliary density matrix method (ADMM) (Guidon et al., 2010) to compute the exchange integrals, using the cpFIT3 basis set for all the elements present in our systems.

Hubbard-corrected (PBE+U) simulations were also performed with the CP2K code applying an isotropic U = 7 eV correction on the 2p subspace of the O-atoms. The U-value was taken from earlier benchmarked studies of Al-doped bulk SiO2 (Nolan and Watson, 2006; Mao et al., 2017).

For hybrid DFT and LR-TDDFT simulations, we used the PBE0-TC-LRC XC-functional (Guidon et al., 2009) with 25% Hartree-Fock (HF) exact-exchange mixing and a real-space cutoff of 8 Å, beyond which a long-range correction [based on the spherically averaged PBE exchange hole (Ernzerhof and Perdew, 1998)] is used.

Optical spectra were computed both at the independentparticle level making use of the Fermi Golden Rule (Read and Needs, 1991; Motta et al., 2010) and via the linear-response (LR) time-dependent DFT (TDDFT) formalism (Casida, 2009) recently implemented, in the Tamn-Dancoff approximation [TDA (Casida, 2009)], in CP2K by two authors of the present paper (Strand et al., 2019). In all cases, the first 60 virtual states were converged for each defect structure using a cutoff of 500 Ry. The absorption spectra were reproduced using a Gaussian broadening of 0.3 eV. Consistent with the closedshell ground-state for the considered systems, and based on the expected extremely low probability of a singlet to triplet transition due to spectroscopic selection rules, only singlet excitation were considered in the LR-TDA-TDDFT calculations. The real space cutoff of 8 Å was numerically tested to best compromise between (i) convergence in excitation energies (50 meV) and absorption intensities (5%) sufficient to discriminate between the main features in the simulated optical spectra, and (ii) practical viability of the simulations on Tier0 High Performance Computing resources (**Figures S1, S2** in the Supporting Information).

Depending on the choice of the method used to simulate optical absorption, different quantities were calculated and are accordingly displayed. For the FGR approach, we computed the imaginary part of the dielectric function (ε2). Conversely, for the LR-TDA-TDDFT approach, we computed the absorbance (e, L mol−<sup>1</sup> cm−<sup>1</sup> ) from the calculated oscillator strengths for each optical excitation following the standard procedure reported in http://gaussian.com/uvvisplot/.

# RESULTS AND DISCUSSION

# Determination of the Supercell Size

When simulating defects by means of periodic DFT simulation approaches as used here, the size of the periodic cell containing the defect is one of the critical aspects to be numerically checked. To reduce the computational cost without introducing (finitesize) errors in the simulations, the simulation cell should be as small as possible yet capable to preserve bulk-like, unperturbed regions around the defect site following structural relaxation.

For this initial benchmark, we focused on D1. The optimized structure of the AlSi NT in Yucelen et al. (2012a) (24 Al-atoms in the circumference, 336 atoms in total) was modified to recover the same D1 composition and structure derived from the NMR study in Yucelen et al. (2012a). The same procedure was repeated for two larger systems made up by two (672 atoms) and three (1,008 atoms) NT-repeat units, respectively.

The structure of these three differently sized models of D1 was optimized by fully relaxing all the atoms in the simulation cells. We then analyzed the defect-induced atomic-displacements as a function of the distance from the defect site (**Figure 2A**). For a suitably sized model, the regions furthest from the defect site should maintain the same atomic structure as in the perfect, defect-free, system.

To facilitate the analysis, the NT-atoms were divided in different groups based on their distance from the defect-center (**Figure 2**). The atoms directly bonded to a missing atom in the defect structures are referred to as nearest neighbors **(NN)**. The atoms separated by one bond from the NN atoms are marked second nearest neighbors **(SNN)**. The atoms within a 10 Å distance from the defect-sites (but further from the defect center than the SNN atoms) have been inserted in a group called first

FIGURE 2 | (A) Computed average atomic-displacements from initial bulk-like positions (h1ri, Å) for the optimized D1 model as a function of the number of NT-repeat units in the simulation cell (supercell dimension). See text for the definition of the NN, SNN, FS, SS, and TS labeling for the different group of atoms as a function of their distance from the center of the D1 defect. (B) Front and side view of the x3 AlSi NT model hosting D1. The atoms have been colored as a function of their h1ri value. Same color labeling as in (A). (C) Computed h1ri for the D1–D5 models (x3 supercell).

shell **(FS)**. Accordingly, the FS group does not include NN and SNN atoms. The same approach has been used for atoms within a 15 Å (second shell, **SS**) and 20 Å (third shell, **TS**) distance from the center of the defect-site. The computed displacements from the initial bulk-like positions were then group-averaged, obtaining the h1ri values shown in **Figure 2**.

The simulation-cell containing two AlSi NT repeat units shows small but non-negligible displacements for the SS and TS atoms (h1ri ∼10−<sup>2</sup> Å. Conversely, the supercell containing three AlSi NT repeat-units presents negligible h1ri ≤10−<sup>3</sup> Å values for all the (FS, SS and TS) atoms further than the SNN group. The extremely small displacements for the TS atoms demonstrates that the simulation cell with three NT repeat units suffer from very limited (i.e., negligible) finite-size effects and enable modeling of regions of the defective NT with a pristine unperturbed structure.

The same analysis was repeated also for all the others defects in **Table 1** using the largest x3 simulation cell. The results are shown in **Figure 2B**. Exception made for D3, the average relaxationdisplacements for the atoms further than the SNN (i.e., >15 Å away from the defect site) are in the order of 10−<sup>2</sup> Å or less for all defects D2 to D5. The larger h1ri values for D3 can be ascribed to its more extended structure (**Figure 1**), leading to a longer-range perturbation of the host lattice. The computed h1ri < 10−<sup>2</sup> Å value for the TS atoms (further than 20 Å from the defect center) of D3, nevertheless suggest rather contained finitesize effects also for this defect, when modeled in a x3 supercell of the AlSi NT.

Overall, the results in **Figure 2A** demonstrates that, for atoms belonging to the TS groups (hence furthest from the defectsite within the periodic simulation cell), the defect-induced perturbation are very effectively screened, leading to negligible displacements from bulk-like positions. On the basis of these results, a x3 simulation cell with three NT repeat units was used for the electronic structure characterization. This choice permits to simulate defects concentrations involving ∼10% of the Si-atoms in the cells (for D1), which slightly underestimates the experimentally determined value of 16% (Yucelen et al., 2012a). Accordingly, and pending limitations of the adopted PBE XC-functional and neglect of solvent in the simulations, the considered models should provide a reasonable representation of the defects in AlSi NTs.

# Optimized Defect Structures

The atomic structure for defects D1 to D5 in Yucelen et al. (2012a) were obtained from a combination of several NMR techniques ("1H–29Si and <sup>1</sup>H–27Al FSLG-HETCOR, <sup>1</sup>H CRAMPS, and <sup>1</sup>H–29Si CP/MAS"), without further refinement via energy-based geometry optimization. In the following, we accordingly analyse in detail the role of (DFT) energy-based geometry optimization (in vacuo) for the atomic structure of the defects D1 to D5 both qualitatively and quantitatively.

The optimized structures for the D1-5 and h-D3-5 defects are shown in **Figure 3**. The D1 optimized structure presents a water (H2O) molecule coordinated to an Al-atom adjacent to the defect site. The formation (condensation) of one Al-coordinated H2O molecule is due to the transfer of one H-atom from a dangling –OH group to an adjacent one. As for D1, also optimization of D5 leads to condensation of one H2O molecule that remains coordinated to one of the Al-atoms closest to the center of the defect-site.

Optimization of the D2 structure leads to local re-organization of the local hydrogen-bond network both inside and outside the NT-cavity with no H2O condensation. A qualitatively similar relaxation (rearrangement of inner and outer hydrogen bonding network with no H2O condensation) is found also for D4.

Also for D3, the optimized geometry presents noticeable differences with respect to the NMR proposed structure in Yucelen et al. (2012a). Upon DFT-relaxation, the stretched Si-O-Si bridge breaks and the O-atom relaxes toward the closest Si-atom leaving a negatively charged deprotonated O-atom. The computed barrier-less breaking of the originally elongated Si-O-Si bridge is accompanied by extensive reorganization of the AlO<sup>6</sup> octahedron closest to the defect site. This relaxation in turn leads to condensation of a H2O molecule, which remains coordinated to one Al-atom with additional hydrogen bonding coordination (1.53 Å distance) to one deprotonated O-atom from a neighboring Al-octahedron. This geometry was found to be energetically favored by at least 0.4 eV over the alternative geometries studied (**Figure S3** in the Supporting Information).

The final structure for the protonated defects (h-D3, h-D4, h-D5) are relatively similar to their charged correspondent. However, the presence of additional hydrogen-atoms around the defect sites leads to condensation of an additional H2O molecule. In all cases, the formed H2O molecules remain coordinated the Al-atoms nearest to the center of the defects.

Given the strong sensitivity to proton environments in the experiments in Yucelen et al. (2012a), it is possible the deviation between the NMR proposed structure (**Figure 1**) and DFToptimized one for D1 and D3 (**Figure 3**) may be due to the lack of solvent (H2O) and/or counterions in the simulations. The prohibitive computational (memory) cost of simulating the present models in the presence of explicit solvation, and the unavailability of DFT-benchmarked force-fields for defects in AlSi NTs, prevent us from further investigating this aspect, that is however worth of specialist method development work.

To provide a more quantitative analysis of the optimized defect structure, we next consider their effects on the three types (Si-O, Al-O, O-H) of bonds in AlSi NTs. Given the computed small average-displacement (h1ri) for atoms further than the FS group (**Figure 2**), we limit the present analysis to the atoms belonging to the NN and SNN groups. **Figure 4** reports the computed average change in Al-O, Si-O, and O-H bond-lengths ( 1d ) for the atoms in the NN and SNN groups i.e., those closest to the defect-center. The raw data for this analysis can be found in the Supporting Information (**Table S1**).

In general, the presence of cation vacancies in the AlSi NTs framework causes larger changes in the external Al-O octahedral layer than in the inner Si tetrahedrons. The change in length of the Al-O bonds for the NN, SNN atoms varies from case to case. Whereas, the computed deviations for the Al-O bond-lengths in the NN and SNN groups lie in the 3–10% range (∼ 0.02 − 0.15 Å), those calculated for the stronger Si-O bonds are consistently smaller than 0.01 Å.

Owing to the creation of alternative local hydrogen-bonding patterns and/or condensation of H2O (D2, D4, h-D4. h-D5 in **Figure 3**) the occurrence of these defects in the NTs induces substantially larger relaxations (up to 1d >0.25 Å values in **Figure 4**).

It is useful to note that the modeled barrier-less H2O condensation is found to take place only in defect structures for which the O-O distance between adjacent Al-OH dangling group or between an Al-OH dangling group and a pristine Al-O(H)- Al bridge is <2.5 Å. In D2, D4, and h-D4 the presence of an extra Si atom increases the distances between the closest dangling

the defect-model studied. The adopted reference for each bond type is reported in Table S1 in the Supporting Information.

Al-OH groups. Provided this condition is not fulfilled, no barrier-less H2O condensation takes places during the geometry optimization. However, and interestingly, the absence of H2O condensation during the relaxation of D2, D4, and h-D4 is shown in **Figure 4** to lead to an overall larger structural relaxation around the defect. Altogether, these results suggest that H2O condensation in non-stoichiometric defective AlSi NTs is an effective mechanism to reduce the spatial-extent of perturbation to the NT-lattice, thence strain.

# Electronic Structure of the Defective NTs

As noted above, to the best of our knowledge, the electronic properties of the NMR-inferred defects in Yucelen et al. (2012a) have not been previously considered in the available literature on imogolite, which motivates our interests in the subject and this section. Based on (i) the negligible defect-induced structural distortion computed for the NT-regions farthest from the defect-sites (**Figure 2**), suggesting coexistence of both defectsites and nearly unperturbed NT-regions in our models, and (ii) the encouraging accuracy of PBE0-related XC-functionals in calculating Kohn-Sham BGs (Guidon et al., 2009, 2010), we expect our results to provide a reasonable approximation of the real NTs (in vacuo).

**Figure 5** reports the computed total Density of States (DOS) and defect-site resolved local DOS (LDOS) at both PBE and PBE0-TC-LRC levels. Depending on the defect and its protonation, shallow (h-D5) or deep (D1, D2, D3, D4, D5,h-D3, h-D4) occupied states are introduced in the NT's BG. The ensuing reduction of the energy gap between the highest-energy occupied state and lowest-energy empty state is found to be highly sensitive to the defect composition and structure, with values ranging from 0.25 eV (h-D5) to up to 4.12 eV (D3).

Comparison between the results for the D3-5 defects and their hydrogenated counterparts (h-D3-5) indicates that protonation is effective in turning the deep defect-states of (D3-5) into shallower ones (h-D3-5), with changes as large as 1.5 eV going from D5 to h-D5. For D4, protonation of the defect site leads to shift of the defect states closer to the Valence Band edge and on the SNN atoms. These results hint at the possibility of controlling the energy of the defect states, thence the NT's electronic properties, by altering the pH of the solution the NTs are dissolved in. Although, the absence of implicit or explicit solvent (and control of pH conditions) in our simulations prevents us from drawing firm conclusions on this subject, this result may motivate further experimental research in the pHdependence of Imo-NTs' optical spectra [not considered in Shafia et al. (2016) and Shafia et al. (2015)].

In all cases, the computed LDOS traces indicate dominant contributions to the occupied defect-states from the NN and SNN atoms closest to the center of the defects. Thus, as shown in **Figure 6**, the occupied defect-states turn out to be highly localized at (or around) the defect site. Conversely, the computed (L)DOS indicate that, for all cases, the CB-edge remains delocalized on the unperturbed NT, being dominated by contributions from the atoms furthest from the defect-sites.

In spite of quantitatively different BG values (**Figure 7**), comparison between the PBE, PBE0-TC-LRC (25 and 12.5% HF exact-exchange mixing) and PBE+U calculated LDOS (**Figure 5** and **Figures S4–S7** in the Supporting Information) reiterates a rather weak dependence of the contributions to the VB-edge and defect-states on the given approximation to the XC-functional. A partial exception to this trend is represented by application of 7 eV Hubbard corrections on the O-atoms (2p subspace) of D2 and h-D5, which results in a slightly enhanced localization of the defect-states. In spite of these small differences, all the computational approaches used agree in predicting the defect states localized at the defect-site.

To further support this statement, we re-optimized the structure for two selected systems, namely the single Alvacancy (D1) and the mixed Al-vacancy/additional Si system (D4) at both PBE0-TC-LRC and PBE+U=7 eV level. New **Figure S8** in the Supporting information provides a visual comparison between the PBE, PBE+U and PBE0-TC-LRC optimized geometries for D1 and D4. **Figure S9** in the Supporting Information provides a quantification of the deviations of the PBE+U and PBE0-TC-LRC optimized geometries from the PBE-optimized one.

With average deviations smaller than 0.15 Å even for the Nearest Neighbors (NN) atoms (that are directly bonded to a missing/additional atom), the PBE0-TC-LRC geometries are found to be in good agreement with PBE results. As shown in **Figure S8** the deviations are mostly due to small changes in bondorientations as well as hydrogen bonding patterns. Conversely, use of a rather aggressive U=7 eV correction for the 2p subspace of the O-atoms is found to induce somewhat larger differences between PBE and PBE+U optimized geometries for D1 and D4 (up to ∼0.25 Å average displacement for the NN atoms of D4). The substantially larger (∼0.35 Å) average displacement for the NN atoms of D1 at PBE+U level is due to the absence of water condensation for the PBE+U optimized geometry, in contrast with PBE and PBE0-TC-LRC results.

As the U= 7 eV O(2p) correction was parameterized on experimental data for hole band-gap states in Al-doped SiO<sup>2</sup> (Nolan and Watson, 2006; Mao et al., 2017), a system not containing Al-OH or Al-OH<sup>2</sup> fragments, we are inclined to consider the agreement between PBE and PBE0-TC-LRC results (leading to water condensation) as a robust benchmark of the initially proposed PBE results. Regardless of these specialist and perhaps secondary differences, the good agreement between the differently calculated (L)DOS (**Figure 5** and **Figures S4, S5, S10**) and real-space localization plots (**Figure 6** and **Figures S6, S7, S11**) for D1 and D4 reiterates the conclusion that, pending minor quantitative difference, the physical picture provided by PBE0-TC-LRC refinement of PBEoptimized geometry offer a good compromise between absence of major artifacts and computational costs, at least for the systems considered here.

Altogether, the marked localization of the occupied defectstates, and the computed delocalization of the CB-edge (**Figures 5**, **6**) hint to the possibility of defect-mediated separation of photo-generated e-h pairs via energetically advantageous relaxation of holes to different defect-sites in the NTs. Thus, and pending the actual electron (hole) transfer kinetics which may hinder such separation process, the present

results suggest that cation-vacancies in Imo-NTs could be highly reactive photo-oxidation centers. In addition, the presence of occupied defects states in the BG for defective NTs strengthens earlier suggestions (Levard et al., 2010; Yucelen et al., 2012b) regarding the potential of Imo-NTs as hole-scavengers for molecular photo-catalysts grafted on the NT's wall.

Finally, for the specialist reader we quantify the deviations between PBE and PB0-TC-LRC results for the BGs of the considered systems. As shown in **Figure 7**, not unexpectedly, the PBE-calculated BGs are substantially smaller than the PB0-TC-LRC ones with deviations in the 2.43-2.73 eV range depending on the defect. However, and in spite of the expected underestimation of BGs, the calculated relative energies and amplitudes of the PBE LDOS traces, especially for the occupied states, are in very good agreement with the PB0-TC-LRC ones (**Figure 5**). This result points to a reasonable performance of the PBE functional for the simulation of the relative energy- and realspace localization of occupied cation-vacancy states in Imo-NTs. Not unexpectedly, the results for 12.5% HF mixing are found to lie in between the PBE and PBE0-TH-LRC (25% HF mixing) values. Interestingly, BG values for 12.5% HF mixing appear to be very similar to the PBE+U=7 eV results with relatively limited (300 meV) system-dependent deviations between the two datasets.

As shown in **Figure 7**, for the pristine NTs, the calculated BG is 5.3 eV and 8.0 eV at PBE and PB0-TC-LRC level, respectively. The calculated PBE value of 5.3 eV is in quantitative agreement with earlier results at a similar level of theory (Zhao et al., 2009), and underestimated with respect to the expectedly more accurate PB0-TC-LRC value. Experimentally, aluminosilicate NTs [with 16% of Si atoms involved in defects (Yucelen et al., 2012a)] are observed to start to absorb light at roughly 4.1 eV (300 nm) with a low-energy absorbance shoulder at roughly 5 eV (250 nm) preceding a marked increase in the absorbance at 5.5–6.2 eV (225–200 nm) (Shafia et al., 2015). Tauc-plot processing of the experimental spectra leads to suggestion of an "optical gap" of 4.9 eV for the pristine (direct gap Zhao et al., 2009, Elliott et al., 2016) NTs (Shafia et al., 2015). The marked overestimation (3.1 eV) of the PB0-TC-LRC BG with respect to the low-energy features in the experimental absorption spectra point to possibly crucial roles of both existing defects (Yucelen et al., 2012a) and excitonic relaxations (Ullrich and Yang, 2014) for the optical properties of aluminosilicate NTs. In the following Section we investigate both these points.

# Optical Properties of Defects

In spite of the growing interest in imogolite NTs, characterization of their optical properties has not been as systematic as the study of their synthesis and potential for chemical separation. To the best of our knowledge, the only UV-vis optical characterization of purified (pristine and Fe-doped) imogolite NTs' samples to date have been published in Shafia et al. (2015) and Shafia et al. (2016). Also to the best of our knowledge, simulation of the optical properties of imogolite nanotubes beyond the independent-particle Fermi Golden Rule (FGR) approximation [with the PBE functional (Elliott et al., 2016; Poli et al., 2016)] is yet to be explored in the literature. To contribute to this knowledge gap, we next investigate the optical properties of the defective NTs-models at different level of theories, analyzing in detail the role of the XC-functional (PBE vs. PBE0-TC-LRC) and optical excitation approach (FGR vs. LR-TDA-TDDFT) for the simulated results.

**Figure 8** reports a comparison between the LR-TDA-TDDFT and FGR spectra computed at both PBE0-TC-LRC and PBE level for all the considered defects. We start our analysis from the LR-TDA-TDDFT spectra at PBE0-TC-LRC level that have been computed with that the least aggressive approximations among all the different combinations presented. All the neutral system (pristine NT, D1, D2, h-D3, h-D4, and h-D5) are computed to present a main absorption peak between 6.5 and 7.5 eV. Conversely, for the negatively charged defects (D3, D4, D5), we compute absorbance peaks at lower energy, with D5 presenting a ∼4.5 eV absorption peak. Interestingly, depending on the chemical composition (and charge) of the defect, the computed absorbance peak can be both red-shifted (D1, D2, D3, D4, D5) and blue-shifted (h-D3, h-D4, h-D5) with respect to the defectfree NT's absorbance.

Comparison between the calculated spectra suggests a larger low-energy optical absorbance with respect to the pristine NTs for all the defective structures but D3, whose low-energy absorbance is computed to be strongly suppressed with respect to the defectfree NT (**Tables S5, S11** in the Supporting Information). Thus, the simulations suggest defective NTs sections to absorb more light than defect-free sections. As shown in **Figure 8A**, the calculated absorbance for D1, h-D3, h-D4, h-D5 is 10–20 times

larger than for the defect-free NTs. Instead, for D2, D5, and D4 the absorbance is twice or comparable to the defect-free NT. Although an immediate connection between the local chemical bonding in the defects and the value of the absorbance cannot be easily drawn, it is interesting to note that protonation of D3, D4, and D5 defects (likely for the defective NTs in water solution at slightly acidic pH as the defect is negatively charged) results in increased optical absorbance. We believe these results motivate further experimental research in the pH-dependence of Imo-NTs's optical properties, an aspect not considered in Shafia et al. (2016) and Shafia et al. (2015).

Before discussing the results of the simulations against published experiments, we recall that as per the NMR study of Yucelen et al. (2012a), up to 16% of Si atoms are involved in a cation vacancy defect and that defects D1 to D5 should be simultaneously present in the NTs. That is, defect-free NTs are not expected to exist, at least for the currently available synthetic routes. We also recall that, experimentally, AlSi NTs start to absorb light at roughly 4.1 eV (300 nm) with a lowenergy absorbance shoulder at roughly 5 eV (250 nm) preceding a marked increase in the absorbance for energies larger than 5.5–6.2 eV (225–200 nm) (Shafia et al., 2015).

Based on the computed PBE0-TC-LRC, LR-TDA-TDDFT spectra in **Figure 8A**, it is tempting to tentatively interpret the shape of the experimental spectra in the 300–200 nm (4.1–6.2 eV) region as being due to convolution of the progressively larger absorbance peaks of D5, D4, D2, and D1, with the defectfree NTs contributing -but not dominating- the absorbance around 7 eV. As a direct consequence of the present results, the Tauc-plot derived value of 4.9 eV should likely be reassigned to a strong defect absorbance (D5), rather than the intrinsic optical-gap of pristine nanotubes (calculated to be a ∼7.0 eV with an approximated –likely underestimated– exciton binding energy of ∼1.1 eV, see below). Given the computed sensitivity to protonation for the absorbance of defects D3 to D5, further experimental research in the dependence of the NTs' UV-vis spectra on the pH conditions would be highly desirable to validate (or confute) the PBE0-TC-LRC, LR-TDA-TDDFT predictions. Recording of spectra below 200 nm would be also highly desirable and strongly contribute to the ongoing work for development of TDDFT strategies to excitons in solids (Gonze et al., 1995; Ghosez et al., 1997; Bernasconi et al., 2011; Tomic et al., 2014; Ullrich and Yang, 2014).

We have also quantified the role of HF mixing (25% as per original PBE0-TC-LRC and 12.5%) for the calculated LR-TD-TDDFT optical properties. To this end, **Figure 9** reports a comparison between the PBE, PBE0-TC-LRC (12.5% HF mixing), and PBE0-TC-LRC (25% HF mixing). In general, reduction of the HF mixing from the original 25–12.5% leads to a decrease of both the energy and intensity of the optical absorption peaks. The relative energy and intensity ranking for the defect considered remains qualitatively unaltered by reduction of the HF mixing. The only exceptions to this trend are the h-D3 and h-D4 systems that, for 12.5% mixing, are not computed to have the 2nd and 3rd largest absorption peak (as for 25% HF mixing). Decrease of HF mixing from 25 to 12.5% is found also to lead to disappearance of a well-defined lowenergy shoulder for D1 and D2, indicative of a different energy distribution for the optically active electronic excitations in the system. In spite of the quantitative dependence of the calculated spectra on the HF mixing in the hybrid-DFT approach, the results of this benchmark strengthen our earlier assignment of the experimental low-energy absorption onset [4.1-6.2 eV, (Shafia et al., 2015)] being due to the NMR-inferred defects in the NTs.

We next consider the difference between the calculated PBE0- TC-LRC Kohn-Sham BGs and the lowest energy LR-TDA-TDDFT excitation as a first approximation to the exciton binding energy in the defective NTs. As explained in Ullrich and Yang (2014) (**Figure 2** and related discussion), the approximation stems from (i) the PBE0-TC-LRC BG being used instead of the true quasi-particle gap, and (ii) the approximated PBE0- TC-LRC description of the many-body e-h interactions leading to stabilization of the exciton (as measured by the exciton binding energy). Therefore, although encouraged by previous successful applications of unscreened non-local XC-functionals (e.g., B3LYP) with non-local long-range terms in the xc-kernel (fxc) (Bernasconi et al., 2011, 2012; Tomic et al., 2014; Ullrich and Yang, 2014), a note of caution is in place, especially in the absence of many-body benchmark results on defective Imo-NTs' true quasi-particle gap, a quantity not available in the literature and well beyond our computational resources. Moreover, due to the use of real-space cut-off for the exchange integrals in our simulations at PBE0-TC-LRC (Guidon et al., 2009; Strand et al., 2019), the use of a long-range (semi-local) correction may prevent quantitative description of diffuse excitations leading to highly delocalized electron-hole pairs (Gonze et al., 1995; Ghosez et al., 1997; Ullrich and Yang, 2014). In spite of this limitation, the localized nature of the defect occupied states (dominating the low energy optical excitations, vide infra) and the inclusion of exact-exchange to within a 8 Å, should jointly contribute to contain the deficiencies intrinsic to the adopted approach. As noted in the methods section (see also **Figures S1, S2**), the adopted 8 Å cutoff was numerically checked to yield excitation

energies (∼7 eV defect-free NT absorbance-peak included) converged to within 0.05 eV for D1, which, at least to our minds, vouches for a favorable compromise between accuracy and computational viability of the present approach for the system studied.

As shown in **Figure 10**, the computed energies differences (approximated exciton binding energies) are consistently larger than 0.3 eV and can be as large as ∼1.1 eV for the pristine NT and D1. By comparing the results in **Figure 10** with the computed measures of defect-induced structural relaxation (**Figure 2**), it transpires that, at least for the system studied, the exciton binding energy decreases as a function of the defect-induced structural reorganization (D1: ∼1.1 eV, D2: ∼0.9 eV, D3-5, and h-D3-5: < 0.6 eV). Thus, disruption of the cylindrical symmetry for the NT, appears to reduce the excitonic binding energy, at least for the modeled systems in vacuo.

Reduction of the HF mixing from 25 (standard PBE0-TH-LRC) to 12.5% leads to a parallel substantial (up to nearly 400 meV for the pristine NT) decrease in the approximated exciton binding energy for all the cases apart from the negatively charged defects (D3, D4, and D5). For the latter cases, the computed differences in approximated exciton binding energies are smaller than 20 meV. This result in turn suggests a reduced role for nonlocal interactions (as captured by the HF contribution to the XCfunctional) for the low-energy excitations of these systems. As also evident from **Figure 10**, in spite of quantitative differences, use of both 25 and 12.5 % HF mixings consistently leads to smaller (approximated) exciton binding energy for the systems characterized by larger structural relaxation, strengthening the earlier conclusions based on original PBE0-TH-LRC (25% HF mixing) results.

In an attempt to further investigate the nature of the absorbance peaks and low-energy shoulders in **Figure 8**, we next analyse, for each considered case, the five excitations with the largest oscillator strength. **Tables S3–S11** and **Figures S12–S20** in the Supporting Information reveal the presence of at least one secondary excitation with non-negligible oscillator strength for D1, D2, and h-D4. These secondary excitations are, respectively 0.62 eV (D1), 0.87 eV (D2), and 1.34 eV (h-D4) lower in energy with respect to the main peak. In spite of the complex composition of calculated absorbance spectrum, the largest oscillator strength excitations in **Tables S3–S10** and **Figures S12–S19** consistently involve transitions from defectlocalized orbitals to delocalized CB orbitals (in general far from the defective part of the NTs). Combined with the substantial reduction (>0.5 eV) in the approximated exciton binding energy for all the NMR-inferred defects but D1 and D2 (**Figure 9**), it is unavoidable to note that altogether these results point to the possibility of defect-mediated exciton-separation in AlSi NTs, with creation of highly-localized oxidative states

at the defect-sites. Given the ongoing experimental work in development of (UV-light) sustainable photo-catalytic strategies for selective production (oxidation) of high-value chemicals (Sastre et al., 2012a,b; Alarcos et al., 2017; Murcia-López et al., 2017), and the absence of experimental results on the excitonic binding energies in Imo-NTs, these results prompt for interest, further research (and validation opportunities) by the experimental community. The current unavailability of excited state gradients in the used implementation of LR-TDA-TDDFT (Strand et al., 2019), prevents us from further exploring this point by simulation.

Finally, for the specialist reader we briefly discuss the dependence of the computed spectra on the simulation protocol. We start by considering the role of the xc-functional for the computed LR-TDA-TDDFT spectra. Due to both the BG underestimation (**Figure 7**) and the lack of non-local terms in the xc-kernel (fxc) (Gonze et al., 1995; Ghosez et al., 1997; Bernasconi et al., 2011, 2012; Tomic et al., 2014; Ullrich and Yang, 2014), the PBE LR-TDA-TDDFT spectra (**Figure 8C**) deviates from the PBE0-TC-LRC ones both in terms of computed absorbance values (reduced for PBE) and relative energy of the absorbance peaks. Apart from D5, the absorption peak for all the defects and the pristine NT is now computed between 4.2 and 5 eV. The largest (smallest) absorbance is computed for the D1 and D4 (D3 and D5). That is, at odds with the PBE0-TC-LRC results, protonation is not modeled to increase optical absorbance at PBE-level, in contrast with PBE0-TC-LRC results. **Table S2** in the Supporting Information reports a comparison between the approximated (LR-TDA-TDDFT) exciton binding energies at PBE0-TC-LRC (see also **Figure 10**) and PBE level. Apart from the pristine nanotubes, D1 and D2 the approximated (LR-TDA-TDDFT) exciton binding energies are found not to

excitation energy.

majorly change depending on the use of the semi-local PBE or non-local PBE0-TC-LRC (the computed changes are smaller than ∼0.1 eV). This results suggest that depending on the structural reorganization of the defects (smaller for D1 and D2, larger for the other defects in **Figure 2**), long-range effects become progressively more screened and less important for the electronic and optical properties of Imo-NTs. The agreement between 25 and 12.5% HF mixing results for D3-D5 in **Figure 10** strengthens this conclusion.

Regardless of the use of PBE0-TC-LRC or PBE, simulation of the spectra at the independent-particle level via the FGR consistently lead to an absorbance maximum at ∼9 (PBE0- TC-LRC) and 6.2 eV (PBE) with qualitatively similar computed spectra. It thus emerges that, regardless of the XC-functional used, application of the FGR prevents the eventual use of the computed absorbance spectra to distinguish between defects and pristine NTs. Accordingly, and not unexpectedly, it transpires that use of the LR-TDA-TDDFT approximation at PBE0-TC-LRC level appears to offer the most favorable accuracy-viability compromise for the study of the optical properties of pointdefects in Imo-NTs, at least out of the simulation protocols explored. Given the availability of experimental spectra for Fedoped aluminosilicate NTs (Shafia et al., 2015, 2016) future work will investigate the performance of the present PBE0-TC-LRC, LR-TDA-TDDFT approach on these systems to sustain research both in the Imo-NTs area and simulation of optical properties in (1D) solids.

# CONCLUSIONS

Simulation of the five NMR-inferred defect structures (D1 to D5) for AlSi NTs (Yucelen et al., 2012a) and their protonated counterparts (h-D3 to h-D5) indicates that:


# REFERENCES

Alarcos, N., Cohen, B., Ziółek, M., and Douhal, A. (2017). Photochemistry and photophysics in silica-based materials: ultrafast and single molecule spectroscopy observation. Chem. Rev. 117, 13639–13720. doi: 10.1021/acs.chemrev.7b00422


Based on the current lack of experimental results on the electronic and optical properties of defects in imogolite NTs, we believe the present results will be a positive complement to research in the field.

# AUTHOR CONTRIBUTIONS

EP executed the simulations and analyzed the results in collaboration with JE, MW, and GT. SC and MW implemented and tested the used TDDFT formalism in CP2K. EP and GT wrote the initial draft of the manuscript. All the authors contributed to the final version of the manuscript.

# ACKNOWLEDGMENTS

EP, JE, and GT acknowledge support from EPSRC UK (EP/I004483/1, EP/K013610/1, EP/P022189/1, and EP/P020194/1). EP acknowledges support from CINECA (ICT-CMSP-2018). Via membership of the UK's HEC Materials Chemistry Consortium (EP/L000202) and UKCP Consortium (EP/K013610/1 and EP/P022189/1), this work made use also of the ARCHER UK and UK Materials and Molecular Modeling Hub (EP/P020194/1) High-Performance Computing facilities. MW and SC acknowledge support from EPSRC UK (EP/P022235/1) and an ARCHER eCSE (eCSE03-11).

# SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fchem. 2019.00210/full#supplementary-material

Amara, M. S., Paineau, E., Rouzière, S., Guiose, B., Krapf, M. E. M., Tache, O., et al. (2015). Hybrid, tunable-diameter, metal oxide nanotubes for trapping of organic molecules. Chem. Mater. 27, 1488–1494. doi: 10.1021/cm503428q

Arancibia-Miranda, N., Escudey, M., Pizarro, C., Denardin, J. C., García-González, M. T., Fabris, J. D., et al. (2014). Preparation and characterization of a single-walled aluminosilicate nanotube-iron oxide composite: its applications to removal of aqueous arsenate. Mater. Res. Bull. 51, 145–152. doi: 10.1016/j.materresbull.2013.12.014


Zhao, M., Xia, Y., and Mei, L. (2009). Energetic minimum structures of imogolite nanotubes: a first-principles prediction. Phys. Chem. C 113, 14834–14837. doi: 10.1021/jp9056169

**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Poli, Elliott, Chulkov, Watkins and Teobaldi. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Mechanisms of Oxygen Vacancy Aggregation in SiO<sup>2</sup> and HfO<sup>2</sup>

David Z. Gao1,2,3, Jack Strand<sup>1</sup> \*, Manveer S. Munde1,4 and Alexander L. Shluger <sup>1</sup>

<sup>1</sup> Department of Physics and Astronomy, University College London, London, United Kingdom, <sup>2</sup> Nanolayers Research Computing LTD., London, United Kingdom, <sup>3</sup> Department of Physics, Norwegian University of Science and Technology, Trondheim, Norway, <sup>4</sup> Materials Science Center, Faculty of Physics, Philipps University Marburg, Marburg, Germany

Dielectric oxide films in electronic devices undergo significant structural changes during device operation under bias. These changes are usually attributed to aggregation of oxygen vacancies resulting in formation of oxygen depleted regions and conductive filaments. However, neutral oxygen vacancies have high diffusion barriers in ionic oxides and their interaction and propensity for aggregation are still poorly understood. In this paper we briefly review the existing data on static configurations of neutral dimers and trimers of oxygen vacancies in technologically relevant SiO<sup>2</sup> and HfO<sup>2</sup> and then provide new results on the structure and properties of these defects in amorphous SiO<sup>2</sup> and HfO2. These results demonstrate weak interaction between neutral O vacancies, which does not explain their quick aggregation. We propose that trapping of electrons, injected from an electrode, by the vacancies may result in creation of new neutral vacancies in the vicinity of pre-existing vacancies. We describe this mechanism in a-SiO<sup>2</sup> and demonstrate that this process becomes more efficient as the vacancy clusters grow larger.

### Edited by:

Maria Veronica Ganduglia-Pirovano, Instituto de Catálisis y Petroleoqumica (ICP), Spain

### Reviewed by:

Cesare Franchini, University of Vienna, Austria Alejandro Strachan, Purdue University, United States

> \*Correspondence: Jack Strand jack.strand.14@ucl.ac.uk

### Specialty section:

This article was submitted to Physical Chemistry and Chemical Physics, a section of the journal Frontiers in Physics

Received: 16 November 2018 Accepted: 07 March 2019 Published: 29 March 2019

### Citation:

Gao DZ, Strand J, Munde MS and Shluger AL (2019) Mechanisms of Oxygen Vacancy Aggregation in SiO2 and HfO2. Front. Phys. 7:43. doi: 10.3389/fphy.2019.00043 Keywords: dielectric oxides, oxygen vacancies, Density Functional Theory, SiO2 , HfO2

# 1. INTRODUCTION

Crystalline and amorphous SiO<sup>2</sup> and HfO<sup>2</sup> are important materials for microelectronics as dielectric layers in transistors and memory cells, fiber optics, mirror coatings, resistive random access memory (ReRAM), and neuromorphic cells (see [1–3]). Under an applied stress bias, field-effect transistors (FETs) experience a gradual increase in gate leakage, noise signal, threshold voltage shift, decrease of trans-conductance, and other degradation effects, which are accelerated by strong bias and may lead to dielectric breakdown [1, 4]. These effects have generally been attributed to defect generation and aggregation in oxide layers [1, 4, 5]. The creation of new defects is associated with structural changes in the oxide and requires significant energy [1, 6, 7]. In particular, the aggregation of oxygen vacancies as a result of electrically stressing sub-stoichiometric amorphous SiO<sup>x</sup> (x = 1.32) films is thought to facilitate the dielectric breakdown of complementary metal-oxide-semiconductor (CMOS) devices [8–10] and electroforming in resistive random access memory cells [3, 11, 12]. However, the fundamental atomistic mechanisms behind these processes are still poorly understood. Both the creation of additional oxygen vacancies and the interaction of diffusing vacancies have been proposed as mechanisms for vacancy aggregation.

Clustering of O vacancies is known to occur in the bulk and at surfaces of reducible oxides, such as e.g., TiO<sup>2</sup> [13, 14], CeO<sup>2</sup> [15, 16], and SrTiO<sup>3</sup> [17]. It is much less understood in nonreducible wide band gap oxides, such as MgO, SiO2, and HfO2. The formation of pairs of O vacancies has been suggested in heavily thermochemically reduced MgO [18]. Recent simulations [19, 20] have shed some light on the formation of O-vacancy clusters in SiO<sup>2</sup> and HfO2. Computational modeling was used to investigate the structures and binding energies of vacancy dimers and trimers in amorphous (a)-SiO2, the energy barriers for the diffusion of individual vacancies, and whether this diffusion can be stimulated by trapping of injected electrons at vacancies [19]. The calculations of di- and tri-vacancy clusters demonstrate that there are favorable sites for vacancy aggregation in a-SiO<sup>2</sup> with maximum binding energies of approximately 0.13 and 0.18 eV, respectively. However, an average neutral vacancy diffusion barrier was calculated to be 4.6 eV [19], rendering the clustering of randomly distributed neutral vacancies via diffusion unfeasible.

The calculations performed in Bradley et al. [20] demonstrate that the formation of neutral oxygen vacancy dimers, trimers and tetramers in monoclinic HfO<sup>2</sup> is accompanied by an energy gain with respect to the separated vacancies, which depends on the size and shape of the aggregate. In the most strongly bound configurations, vacancies are unscreened by Hf cations and form voids within the crystal, with the larger aggregates having larger energy gain per vacancy (0.11–0.18 eV). The negatively charged di-vacancy was found to have similar binding energies to the neutral one, while the positively charged divacancy was found to be unstable. Thus, the aggregation process of either neutral or negatively charged oxygen vacancies in m-HfO<sup>2</sup> is energetically feasible. However, the energy barrier for the diffusion of neutral O vacancies is about 2.4 eV [21], again suggesting that aggregation of vacancies via diffusion processes is too slow to explain e.g., electroforming in HfO<sup>2</sup> films.

The realization that the formation of oxygen vacancy clusters in a-SiO<sup>2</sup> and m-HfO<sup>2</sup> via thermally activated diffusion is inefficient implies that an alternative mechanism should be responsible for the aggregation of oxygen vacancies under electrical stress. In particular, O-vacancy aggregation could result from creation of new vacancies near the pre-existing vacancies. To achieve that, an O atom should be displaced from its lattice position to an interstitial site, creating an O-vacancy at the original site and an interstitial O ion, i.e., a Frenkel defect pair.

To put this mechanism in the context of previous studies, we briefly review the structure and characteristics of vacancy dimers and trimers in a-SiO<sup>2</sup> and crystalline and a-HfO2. We then discuss the mechanisms of Frenkel defect creation in SiO<sup>2</sup> and HfO<sup>2</sup> facilitated by electron injection and show how similar mechanisms can produce new O-vacancies in the vicinity of preexisting O-vacancies on the example of a-SiO2. We demonstrate that two electrons trapped at a pre-existing O-vacancy facilitate the formation of a new O-vacancy and an interstitial O2<sup>−</sup> ion and that this process becomes more probable as the vacancy clusters grow larger.

# 2. METHODS OF CALCULATIONS

The initial non-defective amorphous structures of SiO<sup>2</sup> and HfO<sup>2</sup> were obtained using classical molecular dynamics and a melt and quench procedure, which is described in detail in prior publications [22, 23]. DFT was then used to optimize starting amorphous geometries, calculate their electronic structures, and study the properties of neutral and charged O defects in the material. Ten periodic models of a-SiO<sup>2</sup> containing a total of 216 atoms each were produced using the ReaxFF forcefield [24, 25] implemented in the LAMMPS code [26] and an NPT ensemble. ReaxFF was selected because it was originally parameterized to reproduce the properties of small clusters and silica polymorphs [24, 25] and reproduces the structural characteristics of silica glass well [27]. The initial models of a-HfO<sup>2</sup> structures were created using the pair potentials [28] and the LAMMPS package [26]. Nine periodic models of a-HfO2, each containing 324 atoms, have been generated using classical molecular dynamics and a melt and quench procedure in an NPT ensemble. The Berendsen thermostat and barostat were used to control the simulations. The melt cooling rates were 8 and 0.75 K/ps for SiO<sup>2</sup> and HfO2, respectively. We note that using an NPT ensemble creates distribution of densities of amorphous samples. In this work we used densities corresponding to experimental samples, as discussed in El-Sayed et al. [22] and Kaviani et al. [23].

Most DFT calculations were performed using the Gaussians and Plane Waves [29] method as implemented in the CP2K code [30]. The PBE0-TC-LRC [31] nonlocal functional was selected to accurately reproduce the band gap of a-SiO<sup>2</sup> and used in conjunction with the auxiliary density matrix method (ADMM) [32] to mitigate the computational expense. Finally, we used the MOLOPT basis set [33] along with the corresponding pseudopotentials (GTH) [34] and a 400 Ry plane wave cutoff. This method predicts an average band gap of 8.1 eV for amorphous SiO2. The densities, radial distribution functions, distributions of Si-O bonds and Si-O-Si angles, as well as neutron structure factors of a-SiO<sup>2</sup> models used in this work have been evaluated in El-Sayed et al. [22] and showed good agreement with prior theoretical [27, 35] and experimental [36] studies. This analysis demonstrated that these amorphous models are able to produce disordered continuous random networks that accurately represent the structure of a-SiO2.

The densities of the DFT optimized a-HfO<sup>2</sup> structures used in this work are in the experimentally determined range of 9.6 – 9.7 g cm<sup>3</sup> [23]. Contrary to a-SiO2, a-HfO<sup>2</sup> samples are characterized by distributions of coordinations of Hf and O ions and bond lengths. In particular, around 56% of the Hf atoms are bonded to 6 oxygen atoms, 38% are bonded to 7, and 6% to 5 oxygen atoms. On the other hand, 83% of the oxygen atoms are 3 coordinated, and there are about 6% of 2 coordinated oxygen atoms. These are in good agreement with previous studies, as discussed in detail in Kaviani et al. [23]. The one-electron band gap of initial structures does not contain localized states due to under-coordinated atoms and the band gap determined as a difference between the energies of HOMO and LUMO Kohn-Sham (K-S) energies is equal to 6.0 eV on average (ranging from 5.8 to 6.2 eV in different structures).

The formation energy of a vacancy was calculated using the standard formula

$$E\_{\rm F} = E\_{\rm NaCl} - E\_{\rm Bulk} + \mu\_{\rm O} + q(E\_{\rm Fermi} + E\_{\rm VBM}), \tag{1}$$

where EVac and EBulk are the total energies of the vacancy and pure cell, respectively. µ<sup>O</sup> is the chemical potential of oxygen, q is the charge state of the vacancy, and EFermi is the Fermi level referenced to the energy eigenvalue of the valence band maximum, EVBM. When modeling charged vacancies, charge corrections due to the interaction of periodic images were evaluated using the Lany and Zunger method [37, 38], with a dielectric constant of 3.9 for a-SiO<sup>2</sup> and 20 for a-HfO2, and a Madelung constant of 2.8373 for the cubic cells. The population analysis of electronic states was carried out using the Bader method [39]. For the sake of convenience, all charges will be given in units of electron charge.

# 3. NEUTRAL DI- AND TRI-VACANCIES IN A-SIO<sup>2</sup> AND A-HFO<sup>2</sup>

# 3.1. Vacancy Clusters in a-SiO<sup>2</sup>

As observed consistently in previous work, the removal of an oxygen atom results in the formation of a Si-Si bond at the vacancy site ranging from 2.20 to 2.80 Å in length [19, 40–43]. This creates a doubly occupied defect state within the band gap between 0.39 and 1.34 eV above the valence band maximum (VBM), with the LUMO also positioned within the band gap between 7.04 and 8.20 eV above the VBM [40, 42]. However, such data with regards to di- and tri-vacancy clusters has not been previously published.

In the current study, we have analyzed a total of 25 di-vacancy structures and 12 tri-vacancy clusters using the sampling method described in detail in Munde et al. [19]. Typical configurations of di- and tri-vacancies in a-SiO<sup>2</sup> are shown in **Figure 1**. As illustrated, an oxygen tri-vacancy can exhibit three possible configurations: (i) the three vacancies are situated within a 3 member ring; (ii) the three vacancies diverge from a central Si atom to give a "Y" configuration; and (iii) the three vacancies form a chain. Our previous results have suggested that the 3 member ring and Y-geometry are energetically favorable when compared to the vacancy chain [19]. In all cases, the removal of additional oxygen atoms results in further network distortion. Si-Si bond lengths range from 2.20 to 3.00 Å and 2.40 to 3.00 Å in di- and tri-vacancy clusters, respectively. One can attribute the increase in the lower limit to bond length to a decrease in atomic density around the vacancy cluster site.

In all cases considered, occupied states are observed within the band gap. For the di-vacancy case, a single state is observed between 1.27 and 1.97 eV above the VBM with a mean of 1.52 eV. For the tri-vacancy case, two states are observed at 0.35 to 2.04 eV and 1.40 to 2.47 eV above the VBM, with mean values of 1.23 and 1.97 eV, respectively. These states result from the pairing up of electrons from the remaining Si atoms when an oxygen atom is removed. The LUMO is positioned 5.64 to 7.68 eV and 6.10 to 6.92 eV above the VBM for di- and tri-vacancies, with mean values of 6.86 and 6.54 eV, respectively. **Figure 2** illustrates how the LUMO state consists of an electron localized around a cluster site, with the shape of the orbital being dependent on the cluster geometry. In the di-vacancy and 3-member ring tri-vacancy cases (**Figures 2A,B**), the orbital lobes are situated above and below the plane of three Si atoms. This is different for the "Y" geometry and vacancy chain cases (**Figures 2C,D**), where orbital lobes are shared between the four Si atoms making up the vacancy. Typical density of states plots are shown in **Figure 3** for a monovacancy and its corresponding di-vacancy and tri-vacancy. The formation of these states on average ever deeper in the band gap suggests that larger clusters may facilitate conduction processes such as trap-assisted tunneling.

# 3.2. Vacancy Clusters in a-HfO<sup>2</sup>

To study the interaction between vacancies in a-HfO<sup>2</sup> we performed calculations for 15 different vacancy configurations. To produce di-vacancy models we take a subset of 9 single

vacancies, and for each vacancy in this subset produce 3 divacancies by removing adjacent oxygen atoms. This produces 27 di-vacancy models in total. The choice of the original set of parent vacancies reflected the distribution of coordination numbers of oxygen atoms in our a-HfO<sup>2</sup> models. In this approach,

FIGURE 2 | Typical LUMOs from (A) Oxygen di-vacancy in a-SiO2, (B) 3-member ring type tri-vacancy in a-SiO2, (C) Y-geometry type tri-vacancy in a-SiO2, (D) Vacancy chain type tri-vacancy in a-SiO2. Positive and negative iso-surfaces are shown in blue and red with iso values of ±0.05, respectively. The Si atoms belonging to each vacancy cluster are labeled.

the second vacancy probes the local environment around the preexisting vacancy. Vacancy dimers can form as a result of vacancy diffusion, which is inefficient for neutral vacancies, or as a result of formation of new vacancies as a result of electron trapping by pre-existing vacancies discussed below. Efficiency of these processes is governed by energy barriers.

The interaction between neutral vacancies in a di-vacancy can be quantified by comparing the energy required to produce the first vacancy and then the second vacancy nearby. The first formation energy is calculated using Eq. 1. The second formation energy, E 2nd F , is calculated using EVac in place of EBulk, that is E 2nd <sup>F</sup> <sup>=</sup> <sup>E</sup>Divac <sup>−</sup> <sup>E</sup>Vac <sup>+</sup> <sup>µ</sup>O. If <sup>E</sup> 2nd F is lower than EF, the vacancies are said to be "attractive," If E 2nd <sup>F</sup> <sup>&</sup>gt; <sup>E</sup>F, the vacancies are said to be "repulsive." We note that the notion of "attractive," and "repulsive" is blurred in amorphous structures as defect formation energy strongly depends on the local environment. We have checked that all geometric configurations of vacancy dimers are fully converged and hence variations in second vacancy formation energy are caused by the difference in the local environment of each site as well as by the overlap of the vacancy wavefunctions forming bonding states (see **Figure 4**). We have also carried out "reversibility" tests where the O atom is placed back into the vacancy and the geometry is optimized again. In all cases the system returns to the original geometry and total energy, indicating that the formation of the second vacancy does not cause topological changes of the local environment.

The standard deviation of vacancy formation energies for our set of vacancies is 0.2 eV. Therefore we investigate the shift of the formation energy distribution by comparing average formation energies. On average, the second vacancy formation energy is 0.1 eV lower than that of the first. The standard deviation, however, is 0.4 eV. This means that some di-vacancy configurations will be of the "repulsive" type, even though the

average case is "attractive." A typical di-vacancy configuration in a-HfO<sup>2</sup> is shown in **Figure 4**. Interaction between the vacancies is visualized by the plot of the di-vacancy doubly occupied electron states. Instead of two localized states adjacent to one another, the plot shows two states which are extended across the vacancies in a bond-like fashion. This type of molecular bonding type configuration corresponds to the majority of cases (approximately 80 %).

Further stabilization is possible for tri-vacancy clusters. The interaction in tri-vacancies can be quantified in a similar manner to di-vacancies by comparing the third vacancy formation energy with that of the second. It is found that the third vacancy formation energy is, on average, 0.2 eV lower than the second. Oxygen tri-vacancies have two general configurations, they can be organized in a linear or triangular fashion (see **Figure 5**). Similar to a-SiO2, the interaction between vacancies in a more compact triangular configurations is stronger-the formation energy of triangular clusters is, on average, 0.2 eV lower than that of linear clusters. The standard deviation of the tri-vacancy formation energy distribution is 0.5 eV. This means that, similar to the di-vacancy case, there exists tri-vacancy aggregates that are of the repulsive type, despite the fact that, on average, trivacancies are attractive.

Therefore, the vacancy aggregation mechanism is the one in which the presence of a vacancy (or vacancies) lowers the formation energy required to produce an adjacent vacancy. This increases the rate of multi-vacancy generation relative to isolated vacancy generation.

# 4. ROLE OF ELECTRON INJECTION IN O-VACANCY CREATION

In many applications, amorphous SiO<sup>2</sup> and HfO<sup>2</sup> films are sandwiched between electrodes with often strong (of the order of 1V/nm) electric field applied across them. Under such conditions, extra electrons can tunnel into the oxide conduction band or into defect states inside the oxide by TAT. These effects are known to cause degradation of performance of oxide layers as dielectric in devices [1, 4]. In particular, it has recently been suggested that electrons injected into a-SiO<sup>2</sup> can be trapped in the amorphous matrix at structural precursor sites and form deep electron states in the band gap [44]. These intrinsic trapping sites have been identified as wide O–Si–O angles (> 132◦ ) in the continuous random network structure and can accommodate up to two electrons (see **Figure 6**). The formation of these bielectron states lowers the energy barrier to create a neutral Ovacancy in a-SiO<sup>2</sup> from about 8 eV to an average of 0.7 eV [45]. As a result of this process, an interstitial O2<sup>−</sup> ion is created in the amorphous network, as shown in **Figures 6A–C**. The concentration of precursor sites for intrinsic electron traps in a-SiO<sup>2</sup> has been estimated at 4 × 10<sup>19</sup> [44]. More information on the structure of electron traps, their optical absorption, EPR signatures, and a detailed description of electron injection facilitated Frenkel defect creation in a-SiO<sup>2</sup> can be found in prior publications [44, 45].

We note that neutral O-vacancies and negatively charged O2<sup>−</sup> interstitial ions created by this mechanism have an average recombination barrier of 0.8 eV [45]. The formation energy diagram for O vacancies in a-SiO<sup>2</sup> is shown in **Figure 7**. The structural disorder of the amorphous material results in a wide range of local structures and hence a wide range of formation energy values. The charge transition levels of these vacancies exhibit a similar spread with an average 0/–1 charge transition level of 6.9 eV and an average –1/–2 charge transition level of 7.0 eV. Our results indicate that the charge transition levels which correspond to negatively charged vacancies are well positioned in relation to the conduction band of Si [46] and Fermi level position of some metal electrodes [47]. Thus, the newly created O vacancies can support trap assisted electron tunneling through a-SiO<sup>2</sup> in Si/SiO2/metal stacks. The O2<sup>−</sup> ions can diffuse away through the oxide via a previously studied pivot mechanism [45] which is characterized by a low 0.3 eV energy barrier. Under some conditions, oxygen has been observed to release into the gas phase [11].

A qualitatively similar mechanism for the formation of pairs of neutral O-vacancies and interstitial O2<sup>−</sup> ions in monoclinic and amorphous HfO<sup>2</sup> has been proposed in Bradley et al. [48] and Kaviani et al. [49]. Extra electrons injected into monoclinic (m)-HfO<sup>2</sup> can form relatively shallow polaron and bi-polarons with trapping energies of about 0.3 eV [50]. Simulations [48] demonstrate that the formation of stable pairs of neutral oxygen vacancies and interstitial oxygen ions assisted by extra electrons is thermodynamically feasible and requires overcoming activation barriers of about 2.0 eV. This barrier is reduced to <1.3 eV if two extra electrons are initially trapped at a pre-existing O-vacancy. This results in the formation of another O-vacancy nearby and a mobile interstitial O2<sup>−</sup> ion. The created O di-vacancy is stabilized by weak attraction between neutral vacancies, as described above. This leads to further lowering of the formation energy of the defect pair. The binding energy per vacancy in larger oxygenvacancy aggregates increases as the aggregate grows, facilitating the formation of defect pairs next to larger vacancy aggregates.

In amorphous HfO2, extra electrons injected from an electrode form deep polaron and bi-polaron states [23]. Extra electrons localize in energetically deep states at intrinsic structural traps in a-HfO<sup>2</sup> such as under-coordinated Hf ions or Hf ions with elongated Hf-O bonds, both of which are associated with a lowering of the electrostatic potential for an electron. The estimated density of such sites [23] in a-HfO<sup>2</sup> is about 2×10<sup>21</sup> cm−<sup>3</sup> . The structure and spectroscopic properties of these electron traps are described in detail in Kaviani et al. [23] and Strand et al. [51]. The bi-polaron formation causes strong Hf-O bond weakening manifested in significant (about 0.8 Å) ionic displacements. Similar to a-SiO2, thermal activation of these bi-polarons may cause the formation of Frenkel defects in a-HfO<sup>2</sup> [49].

How electron trapping at pre-existing O-vacancies could facilitate the formation of new O-vacancies in a-SiO<sup>2</sup> is discussed below.

# 5. CREATION OF VACANCIES NEAR EXISTING VACANCIES

As discussed above, O-vacancies in amorphous SiO<sup>2</sup> and HfO<sup>2</sup> create deep states in the band gap where electrons can tunnel from Si and other electrodes. One of the possible mechanisms of O-vacancy aggregation is via creation of new vacancies near the pre-existing ones. Below we consider whether this process can be facilitated by electron trapping by O-vacancies in a-SiO2. The cost of creating an oxygen vacancy depends greatly on the local environment. Not only is there a spread of formation energies owing to the disorder, but pre-existing vacancies can affect both the position and the barrier for forming a new vacancy. The presence of an existing vacancy close to the bi-polaron may either facilitate or hinder the next vacancy-creation process.

This can be manifested in several ways: (i) Electron trapping in a vacancy can facilitate the formation of a new vacancy nearby. (ii) Distortion of the surrounding network caused by electron trapping in a vacancy can create another precursor site for electron trapping nearby. Further electron trapping at these precursor sites can facilitate the creation of new vacancies. (iii) Aggregation of several vacancies can distort the surrounding network and enhance local electric field under electrical stress, facilitating creation of new vacancies. Below we consider mechanisms of some of these processes on the example of a-SiO2.

# 5.1. Correlated Vacancy Creation in SiO<sup>2</sup>

In a-SiO2, an O-vacancy can trap up to two electrons [52]. When two electrons localize onto an O-vacancy, the defect distorts as the two 3-coordinated Si atoms displace to point away from each other, as shown in **Figure 8**. As a result of this distortion, there is a small chance that one of the lobes of the HOMO localized on Si atoms of the vacancy will point directly at a neighboring 4-coordinated Si atom, facilitating the Si-Si bond formation by expelling an O atom. The HOMO of a negatively charged O vacancy is shown in **Figure 8**.

To determine the likelihood of creating another O vacancy near an existing one, 144 possible vacancy geometries were investigated. In 97% of these geometries, the vacancy simply trapped two electrons by distorting, as shown in **Figures 8**, **9B**.

FIGURE 7 | O vacancy formation energies in a-SiO2 at various charge states as a function of Fermi level. 144 vacancy geometries were calculated for each charge state due to the amorphous nature of our material. The average formation energy of each charge state is plotted using solid lines, while the minimum and maximum values for each charge state is plotted using dotted lines. The crossing point between positive and neutral vacancies is shown to be between 0.5 and 1.0 eV while the crossing point between neutral and negatively charged vacancies is between to be 6.2–7.5 eV.

In 2 cases the HOMO happened to point at a neighboring 4 coordinated Si atom, as shown schematically in **Figure 9C**. This configuration facilitates the creation of another O vacancy and O interstitial ion through a process with an average barrier of 1.1 eV. A schematic of formation of another vacancy correlated with the pre-existing vacancy is shown in **Figures 9A–D**. However, it is important to note that this process is relatively rare and was only observed in 2 out of 144 cases considered here.

Another possibility is that new wide O–Si–O angle precursor sites for electron trapping are created when O vacancies are occupied by electrons. Trapping two electrons at these new precursors can lead to creation of O vacancies by the mechanism illustrated in **Figures 6A–C**. The creation of these traps is directly related to the amount of local strain present in the system. Both the creation of new vacancies and the localization of electrons at existing vacancies result in an increase of local network distortion. After the geometry of each doubly negatively charged O vacancy was relaxed, another electron was added into the system to probe whether precursor sites for electron trapping were created. Out of a total of 142 geometries (excluding the two examples of correlated defect creation) a new wide O–Si–O angle intrinsic trapping site was found 38 times. In these cases new wide O–Si–O bond angle sites, which did not exist before, were created by the network distortion induced by the charged vacancy relaxation, as shown in **Figure 10**. Due to flexibility of a-SiO<sup>2</sup> network, wide angle precursors created in this way are not necessarily at neighboring sites to the vacancy responsible for distorting the local geometry. The probability of creating these new precursor sites is about 25%. Trapping two electrons at these new precursor sites generates new O vacancies and interstitial O2<sup>−</sup> ions, as described above.

Since creation of additional wide O–Si–O angle intrinsic electron traps is associated with the local strain due to the presence of defects, one can expect that increasing the concentration of such defects should increase the probability of generating new precursor sites for electron trapping. As additional vacancies are created near pre-existing ones, the amount of local strain produced increases as well. To investigate this effect, we examined 40 di-vacancy geometries using 214 atom periodic cells. As electrons were added into these simulations, they can either localize in one vacancy, creating an equivalent of one negatively charged vacancy and one neutral vacancy or, if the di-vacancy exhibits a high degree of symmetry, they can be shared between the two, as shown in **Figure 2**. When up to 2 electrons are localized into the di-vacancy, creation of new precursor sites was observed 10 times, the same as for single vacancies. However, when the defect states are fully occupied by 4 electrons, it was observed that in 20 cases out of 40 wide O–Si–O bond angle sites were created due the network distortion caused by electron trapping at existing vacancies. This demonstrates that indeed clustering of O vacancies and electron trapping in them increase the probabilities of creating new vacancies nearby and these of growing new vacancy clusters.

We further investigated this trend by examining whether O tri-vacancies will also induce wide O–Si–O angle precursor sites. To accommodate such large vacancy clusters, we created a library of 648-atom a-SiO<sup>2</sup> cells using the melt and quench procedure described above. These larger amorphous structures were then screened for existing defects or wide O–Si–O angle intrinsic traps and 5 defect-free geometries were selected. Using these large cells, three libraries of tri-vacancy geometries were created to account for the more variable structure of these larger clusters. The first set of tri-vacancy geometries consisted of examples where the three defects were tightly clustered. The second set was composed of examples where the three defects were arranged in a linear array. Finally, in the third set of geometries each contained three loosely associated vacancies within two nearest neighbor sites of each other. Ten geometries from each set were selected for analysis at random and each of them was simulated with 1–6 extra electrons in the system.

In 23 out of 30 geometries, strain-induced wide O–Si–O angle electron traps were observed after all the pre-existing vacancies were charged. These newly created intrinsic traps were evenly distributed between the three types of tri-vacancy geometries, indicating that their exact configuration does not have a large effect on the correlated creation of new defects. When the trivacancy structures were charged with up to 2 electrons a new intrinsic trap was observed in 8 cases and when they were charged with up to 4 electrons a new intrinsic trap was observed in 14 geometries.

These results suggest that, as the size of O vacancy clusters increases, the likelihood of creating another vacancy nearby increases as well. Single vacancies have a 25% chance of inducing a wide O-Si-O angle intrinsic electron trap when charged, divacancies have a 50% chance, and tri-vacancies are the most likely to create a nearby intrinsic trap with a 77% chance. It also indicates that a significant portion of the network distortion that is responsible for the newly formed traps can be attributed to the localization of injected electrons rather than the presence of the vacancy cluster itself.

The mechanism of creation of a second O-vacancy facilitated by an electron trapping at a pre-existing vacancy in m-HfO<sup>2</sup> has been described in Bradley et al. [48]. The qualitative similarities in vacancy formation and aggregation processes in a-SiO<sup>2</sup> and a-HfO<sup>2</sup> suggest that such mechanisms can be operational also in a-HfO2. They will be discussed in a separate publication.

# 6. CONCLUSIONS

Understanding the mechanisms of aggregation of oxygen vacancies in dielectric oxides is important for improving their performance in electronics, sensors, catalysis and other applications under electrical stress and different types of irradiation. We reviewed the initial stages of O-vacancy clustering in two prototype wide bandgap oxides, SiO<sup>2</sup> and HfO2, and demonstrated that this is energetically favorable in some vacancy configurations. However, barriers for neutral O-vacancy diffusion in both materials are high and their aggregation via thermally activated diffusion is unfeasible at room temperature. As an alternative mechanism of vacancy aggregation, we considered how new O-vacancies can be created as a result of electron trapping by pre-existing ones on the example of a-SiO2. These electrons can be injected via tunneling from an electrode in a device or produced under irradiation process. The results demonstrate that trapping two electrons at O vacancy sites in a-SiO<sup>2</sup> induces local distortion in the amorphous network such that new O–Si–O precursor sites for electron trapping are created with a probability of about 25%. These sites can trap two extra electrons, producing a new vacancy nearby. As the size of the O vacancy clusters increases, the likelihood of

FIGURE 9 | Schematic illustrating the correlated defect creation mechanism. (A) O Vacancies can trap up to two electrons; and (B) distort the amorphous structure. (C) In rare cases, the HOMO of the charged vacancy points toward a neighboring Si atom and forms a bond to (D) create a new vacancy and interstitial O.

wide O-Si-O angle intrinsic electron trap. In about 25% of geometries the third electron injected into a system which already contains a doubly negatively charged O vacancy localized onto a newly created intrinsic trap. Si atoms are shown in yellow while O atoms are shown in red. An iso-surface of the HOMO is shown in blue and pink.

creating another vacancy nearby increases as well. Di-vacancies have a 50% chance, and tri-vacancies are the most likely to create a nearby intrinsic trap with a 77% chance.

These results suggest that electron trapping at oxygen vacancies in amorphous and crystalline oxides may facilitate the creation of new vacancies and this process becomes more efficient as vacancy clusters grow larger because this requires less activation. Previous results suggest that a similar mechanism can be also operational in m-HfO2. We hope that these results will stimulate further studies of oxygen vacancy aggregation in other functional oxides, such as Al2O<sup>3</sup> and Ga2O<sup>3</sup> as well as in nitrides.

# DATA AVAILABILITY

The datasets generated for this study are available on request to the corresponding author.

# AUTHOR CONTRIBUTIONS

DG and MM have performed and analyzed calculations for oxygen vacancy aggregation in a-SiO<sup>2</sup> and wrote the corresponding parts of the paper. JS has performed and analyzed calculations for oxygen vacancy aggregation in a-HfO<sup>2</sup> and wrote the corresponding parts of the paper. AS coordinated the research, participated in analyzing the results and wrote and edited the paper.

# REFERENCES


# FUNDING

We acknowledge funding provided by the UK Engineering and Physical Sciences Research Council (EPSRC) under grants No. EP/K01739X/1 and EP/P013503/1 and by the Leverhulme Trust RPG-2016-135. JS is funded by EPSRC grant no. EP/G036675/1 to the Center for Doctoral Training. Computer facilities on Archer service have been provided via the UKs HPC Materials Chemistry Consortium (EPSRC Grant No. EP/L000202).

# ACKNOWLEDGMENTS

The authors are grateful to L. Larcher, A. Padovani, and G. Bersuker for valuable and stimulating discussions. Part of the results presented in this paper were previously published in The 2018 IEEE Reliability Physics Symposium [53].


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Gao, Strand, Munde and Shluger. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Partial Oxidation of Methane to Syngas Over Nickel-Based Catalysts: Influence of Support Type, Addition of Rhodium, and Preparation Method

### Edited by:

Claudio Cazorla, University of New South Wales, Australia

### Reviewed by:

Benjaram M. Reddy, Indian Institute of Chemical Technology (CSIR), India Tim Schäfer, University of Göttingen, Germany

### \*Correspondence:

Consuelo Alvarez-Galvan c.alvarez@icp.csic.es

### †Present Address:

Mahdi Ahmadi, Cornell University, Ithaca, NY, United States

### Specialty section:

This article was submitted to Physical Chemistry and Chemical Physics, a section of the journal Frontiers in Chemistry

Received: 16 November 2018 Accepted: 11 February 2019 Published: 13 March 2019

### Citation:

Alvarez-Galvan C, Melian M, Ruiz-Matas L, Eslava JL, Navarro RM, Ahmadi M, Roldan Cuenya B and Fierro JLG (2019) Partial Oxidation of Methane to Syngas Over Nickel-Based Catalysts: Influence of Support Type, Addition of Rhodium, and Preparation Method. Front. Chem. 7:104. doi: 10.3389/fchem.2019.00104 Consuelo Alvarez-Galvan<sup>1</sup> \*, Mayra Melian<sup>1</sup> , Laura Ruiz-Matas <sup>1</sup> , Jose Luis Eslava<sup>1</sup> , Rufino M. Navarro<sup>1</sup> , Mahdi Ahmadi 2†, Beatriz Roldan Cuenya2,3 and Jose Luis G. Fierro<sup>1</sup>

<sup>1</sup> Structure and Reactivity Department, Instituto de Catálisis y Petroleoquímica, CSIC, Madrid, Spain, <sup>2</sup> Department of Physics, University of Central Florida, Orlando, FL, United States, <sup>3</sup> Department of Interface Science, Fritz Haber Institute of the Max Planck Society, Berlin, Germany

There is great economic incentive in developing efficient catalysts to produce hydrogen or syngas by catalytic partial oxidation of methane (CPOM) since this is a much less energy-intensive reaction than the highly endothermic methane steam reforming reaction, which is the prominent reaction in industry. Herein, we report the catalytic behavior of nickel-based catalysts supported on different oxide substrates (Al2O3, CeO2, La2O3, MgO, and ZrO2) synthesized via wet impregnation and solid-state reaction. Furthermore, the impact of Rh doping was investigated. The catalysts have been characterized by X-ray diffraction, N<sup>2</sup> adsorptiondesorption at <sup>−</sup>196◦C, temperature-programmed reduction, X-ray photoelectron spectroscopy, O2-pulse chemisorption, transmission electron microscopy, and Raman spectroscopy. Supported Ni catalysts were found to be active for CPOM but can suffer from fast deactivation caused by the formation of carbon deposits as well as via the sintering of Ni nanoparticles (NPs). It has been found that the presence of Rh favors nickel reduction, which leads to an increase in the methane conversion and yield. For both synthesis methods, the catalysts supported on alumina and ceria show the best performance. This could be explained by the higher surface area of the Ni NPs on the alumina surface and presence of oxygen vacancies in the CeO<sup>2</sup> lattice, which favor the proportion of oxygen adsorbed on defect sites. The catalysts supported on MgO suffer quick deactivation due to formation of a NiO/MgO solid solution, which is not reducible under the reaction conditions. The low level of carbon formation over the catalysts supported on La2O<sup>3</sup> is ascribed to the very high dispersion of the nickel NPs and to the formation of lanthanum oxycarbonate, through which carbon deposits are gasified. The catalytic behavior for catalysts with ZrO<sup>2</sup> as support depends on the synthesis method; however, in both cases, the catalysts undergo deactivation by carbon deposits.

Keywords: syngas, methane, partial oxidation, nickel, rhodium, catalyst

# INTRODUCTION

The production of synthesis gas from methane is an important process for converting natural gas, one of the most abundant and cleanest-burning fossil fuels, into value-added high-quality liquid products (Gas-to-liquid technology, GTL). Steam reforming (SRM) has been the preferred technology for the industrial production of synthesis gas from methane to produce ammonia or methanol (Rostrup-Nielsen et al., 2002; Navarro et al., 2007a,b). Nevertheless, SRM is a highly energy-intensive process [Equation 1], and more energy-efficient alternatives to produce synthesis gas are sought.

$$\text{CH}\_4 + \text{H}\_2\text{O} \leftrightharpoons \text{CO} + \text{3H}\_2 \quad \Delta \text{H}\_{298\text{K}}^0 = +206 \text{ kJ/mol} \tag{1}$$

Among the alternatives, the catalytic partial oxidation for the production of synthesis gas from methane (CPOM) is more energy efficient since it has fast kinetics and is exothermic, thus avoiding the need of large reactors and large amounts of superheated steam (Bharadwaj and Schmidt, 1995). In addition, the stoichiometry of the CPO [Equation 2] produces a synthesis gas with an H2/CO ratio of 2:1, which enables its direct utilization for methanol or Fischer-Tropsch synthesis without additional adjustment.

$$\rm CH\_4 + \rm \%O\_2 \leftrightharpoons CO + 2H\_2 \tag{2}$$

The catalytic partial oxidation of methane to syngas is challenging due to the difficulty in controlling the reaction selectivity toward total combustion. Several studies have been performed in the literature to describe the reaction mechanism involved in the CPOM and two reaction mechanisms have been proposed (Dissanayake et al., 1991; Hickman and Schmidt, 1992): the "direct mechanism," in which CH<sup>4</sup> and O<sup>2</sup> react on the catalyst surface to yield CO and H2, and the "combustion-reforming mechanism," in which CH<sup>4</sup> and O<sup>2</sup> react first to form H2O and CO<sup>2</sup> and then, with the excess methane, the dry reforming of CO<sup>2</sup> [Equation 3] and steam reforming [Eq.4] produce the final CO and H2.

$$\rm{CH}\_4 + \rm{CO}\_2 \leftrightharpoons \rm{2CO} + \rm{2H}\_2\tag{3}$$

$$\text{CH}\_4 + \text{H}\_2\text{O} \overset{\longleftrightarrow}{\rightleftharpoons} \text{CO} + \text{3H}\_2\tag{4}$$

Due to the excessive temperature gradients at high conversion rate, the exothermic nature of the reaction and fast deactivation due to carbon deposition on the catalyst's surface, the development of efficient catalysts for the CPOM has been challenging. NPs of noble metals [Pt (Ji et al., 2001), Rh (Ruckenstein and Wang, 1999; Puolakka and Krause, 2007), Ru (Ashcroft et al., 1990), Pd (Vernon et al., 1990)] and nonnoble metals [mainly Ni and Co (Wang and Ruckenstein, 2001)] supported on various oxide substrates have been studied in the CPOM reaction. Nickel is one of the most widely used active phases for CPOM (Miao et al., 1997; Ostrowski et al., 1998; Takehira et al., 2004; Wang et al., 2004). As compared to noble metals, nickel is inexpensive but suffers from deactivation onstream as a consequence of several processes such as sintering, carbon deposition (Claridge et al., 1993) or solid-state reactions of nickel with the substrate. From the large body of work developed on the CPOM reaction, it is clear that the activity and stability of nickel catalysts depend on both the active phase and the support. Metal particle size was proven to be an important factor for the initial intrinsic activity and for the rate of deactivation, with both decreasing with increasing active metal particle sizes (Barbier and Marecot, 1986; Barbier, 1987). The influence of the support on the performance of Ni-based catalysts has been widely studied in the literature (Tsipouriari et al., 1998). Non-reducible Al2O<sup>3</sup> is one of the most studied oxides as support for Ni catalysts because of its thermal stability and high ability to disperse Ni nanoparticles (NPs) (Hu and Ruckenstein, 1998; Ostrowski et al., 1998; Zhang et al., 2000); however, its application for CPOM reaction is limited because of the relatively high deactivation of Ni NPs by sintering and the formation of coke deposits (Lu et al., 1998). Magnesium oxide is another nonreducible support widely studied to disperse stable Ni particles (Choudhary et al., 1998a; Ruckenstein and Hu, 1999; Nishimoto et al., 2004). In this case the formation of a solid solution between nickel and magnesia (Mg1−xNixO) only allows for the reduction of a small fraction of the nickel that remains in close interaction with the basic MgO substrate, favoring this structure for the production of syngas by CPOM with high activity (Requies et al., 2005). Lanthanum oxide has also been used as a support for Ni catalysts (Tsipouriari et al., 1998; Nishimoto et al., 2004). For the Ni/La2O<sup>3</sup> catalysts, good stability was reported and attributed to the increased metal-support interface because the nickel NPs are decorated by La2O2CO<sup>3</sup> species that promote the gasification of coke. Reducible supports (CeO2, ZrO2) have been also studied as systems to disperse active and stable nickel particles for CPOM. CeO<sup>2</sup> is known for its ability to improve the dispersion and stabilization of small nickel metal NPs and for its high oxygen storage/transport capacity, which allows for continuous removal of carbonaceous deposits from active sites (Choudhary et al., 1993; Diskin et al., 1998). In addition, under reducing conditions, the SMSI (Strong Metal-Support Interaction) effect could be observed on ceria, which in turn affects the stability and activity of the dispersed nickel particles (Trovarelli, 1996). Zirconia is another support that shows interesting properties for the dispersion of active and stable Ni NPs. However, the application of ZrO<sup>2</sup> for CPOM reaction is debatable, since this support decreases the availability of the oxygen that participates in the direct CPOM to synthesis gas, resulting in a decrease in activity (Pompeo et al., 2005).

The incorporation of a second metal to the Ni-based catalysts is a common practice designed to improve catalyst stability. The beneficial effect of adding small amounts of precious metals such as Ru, Pt, Pd, Ir, and Rh to a Ni catalyst was previously demonstrated (Tomishige et al., 2002). Rh is one of the most promising metals (Tanaka et al., 2010a). The improvement was explained in terms of the H-spillover from the noble metal towards the non-noble metal, helping the non-noble metal surface to stay metallic (Chen et al., 1997).

As stated above, efficient Ni catalysts for the CPOM reaction require control over the electronic and structural properties of the nickel NPs, which could be achieved by the careful selection of the support and synthesis procedure. In this context, the main objective of this work is to study the influence of different supports (Al2O3, CeO2, La2O3, MgO, ZrO2) with different textural and surface (basicity, reducibility) properties on the activity, selectivity and stability of Ni-based catalysts for the CPOM reaction at atmospheric pressure. Two different synthesis methods have been used: (I) impregnation over different commercial supports and (II) solid state reaction. The influence of a small amount of Rh over the reducibility and reactivity of the catalysts has been tackled. Using various characterization techniques, we will establish structure-activity relationships that indicate which catalyst properties determine its reactivity and thus give us a way to improve the catalytic performance of the systems.

The Ni catalysts have been characterized by X-ray diffraction, nitrogen adsorption-desorption, temperature programmed reduction, X-ray photoelectron spectroscopy, oxygen-pulse chemisorption, transmission electron microscopy, and Raman spectroscopy. The evolution of the structure and morphology of the catalysts is reported and correlated to its catalytic performance.

# MATERIALS AND METHODS

# Preparation of Catalysts Catalysts Prepared by Impregnation of Commercial Supports

Ni (5% weight) catalysts supported on commercial Al2O<sup>3</sup> (Johnson Matthey, 99.97%), CeO<sup>2</sup> (Johnson Matthey, 96%), La2O<sup>3</sup> (Fluka Chemika, 99.98%), MgO (Fluka Chemika, >98%) and ZrO<sup>2</sup> (Johnson Matthey, 99.5%) were prepared by wet impregnation. First, the supports were thermally stabilized by calcination at 850◦C over 3 h. Then, they were sieved between 212 and 425µm. The corresponding amount of Ni(NO3)2·6 H2O (Aldrich Chemie) was dissolved in distilled water. Each support was impregnated with this solution by rotary evaporation (70◦C, 1 h). Then, the obtained solid was dried at 110◦C for 2 h and calcined at 500◦C for 3 h.

# Catalysts Prepared by Solid State Reaction

The RhNi (0.1% Rh, 10% Ni, weight) catalysts, supported on Al2O3, CeO2, La2O3, MgO, and ZrO2, were prepared by mixing and grounding in an agate mortar the corresponding amounts of Ni(NO3)2·6H2O (assay > 98.5%, Aldrich) and the nitrates of the support cations (Mg(NO3)2·6H2O (Scharlau, 98%), La(NO3)2·6H2O (Johnson Matthey, 99.9%), Al(NO3)3·9H2O (99.997% Sigma Aldrich), Ce(NO3)3·6H2O (Johnson Matthey, 99.5%), and ZrO(NO3)2·xH2O, 99.9%), respectively. Since the Rh amount is very small, the method to incorporate this element was based on dissolving the corresponding amount of dicarbonyl(acetylacetonate) Rh(I) (Rh(CO)2acac) (Sigma Aldrich, purum) in acetone and adding this solution dropwise to the powder mixture of nickel nitrate and the corresponding support cation nitrate, thoroughly removing it with a spatula. Then, each mixture was dried at 110◦C for 2 h and calcined at 500◦C for 3 h. Finally, the samples were ground and sieved between 212 and 425 µm.

# Characterization Techniques X-Ray Diffraction

XRD patterns were recorded on a Seifert 3000 powder diffractometer using Cu Kα radiation (λ = 0.15418 nm) generated at 40 kV and 40 mA. Scans were recorded at a rate of 0.02◦ /s for 2θ diffraction angles between 10 and 90◦ .

# N<sup>2</sup> Adsorption-Desorption Isotherms

Textural properties were evaluated by N<sup>2</sup> adsorption-desorption isotherms of the samples recorded at liquid nitrogen temperature with a Micromeritics ASAP2000 apparatus. Samples were degassed at 150◦C under vacuum overnight. Specific areas were calculated by applying the BET method.

# Temperature Programmed Reduction

Temperature-programmed reduction (TPR) experiments were carried out using a semiautomatic Micromeritics TPD/TPR 2900 apparatus equipped with a TC detector. Prior to the reduction experiments, the samples (ca. 30 mg) were thermally treated under air stream at 300◦C to remove moisture. TPR profiles were obtained by heating the samples under a 10% H2/Ar flow (50 mL/min) from 25 to 800◦C with a linear rate of 10◦C/min.

# Oxygen Chemisorption Capacity

Oxygen chemisorption capacity was determined by O2-pulse chemisorption. First, the catalyst sample (60–70 mg) was inserted in a U-quartz reactor and heated to 500◦C in argon, maintaining this temperature for 15 minutes. Then, the temperature was decreased to room temperature and the sample was reduced under a H2/Ar flow (10% H2) up to 750◦C over 60 minutes. Finally, the carrier gas was changed to helium, and once the baseline was stabilized, O<sup>2</sup> pulses were injected until the O<sup>2</sup> detected peaks showed the same area.

# X-Ray Photoelectron Spectroscopy

To extract information about the chemical state and composition of the Ni-supported samples, XPS measurements were acquired using a monochromatic X-ray source (Al Kα, 1486.6 eV) operating at 200 W and a hemispherical electron analyzer (Phoibos 100, SPECS GmbH). The high-resolution data were acquired with a pass energy of 18 eV. CasaXPS software was used to analyze the data.

X-ray photoelectron spectra of the RhNi supported samples were recorded on a VG Escalab 200R spectrometer equipped with a hemispherical electron analyzer and Mg Kα (h·ν = 1253.6 eV) X-ray source (12 kV and 10 mA). The powder samples were degassed at 150◦C for 1 h before being transferred into the analysis chamber. Charge effects on the samples were corrected by fixing the binding energies of the C1s peak at 284.9 eV due to adventitious carbon. This reference gave binding energy values with an accuracy of ±0.1 eV. The data were treated with the "XPS peak" software. The spectra were decomposed with the least squares fitting routine using Gaussian/Lorentzian functions after subtracting the Shirley background.

# Raman Spectroscopy

Raman spectra of the samples were recorded in air, under ambient conditions (being the samples hydrated by air humidity), using a single monochromator Renishaw inVia 1000 system equipped with a thermoelectrically cooled CCD detector and holographic super-Notch filter. The samples were excited with 535 nm (1800 lines/mm).

# Catalytic Activity Tests

The catalytic behavior of the different catalyst precursors for the partial oxidation of methane to syngas was studied under atmospheric pressure at 750◦C using a stainless-steel fixed bed reactor (length = 150 mm, internal diameter = 9 mm) placed inside a hinged oven. The catalysts (100 mg) were subjected to pretreatment under 50 mLN/min of H2/N<sup>2</sup> (10% H2, molar) for the Ni-supported samples or under N<sup>2</sup> flow at 750◦C for 1 h for the RhNi-supported samples, prior to the reaction. Then, the reactants were fed to the reactor (22.6 mLN/min N2, 6 mLN/min O2, and 12 mLN/min CH4). The reaction stream was analyzed on-line by gas chromatography (Varian 45-GC) with a thermal conductivity detector, equipped with a 5A molecular sieve (CP7538) to separate H2, N2, CH<sup>4</sup> y CO and a PoraBOND Q (CP7354) column to separate CO<sup>2</sup> and H2O with a thermal conductivity detector. Argon was used as a carrier gas in order to increase the sensitivity for H<sup>2</sup> detection.

Methane conversion, CO selectivity, H<sup>2</sup> yield and H2/CO ratio were defined as follows:

$$= \frac{\text{CH}\_4 \text{ molar flow (inlet)} - \text{CH}\_4 \text{ molar flow (outlet)}}{\text{CH}\_4 \text{ molar flow (inlet)}} \cdot 100$$

$$H\_2 \text{ yield (\%)} = \frac{H\_2 \text{ molar flow (outlet)}}{2 \cdot \text{CH}\_4 \text{ molar flow (inlet)}} \cdot 100$$

$$= \frac{\text{CO molar flow (outlet)}}{\text{CO} \cdot \text{molar flow (outlet)}} \cdot 100$$

$$\begin{array}{c} \text{CO} \stackrel{\text{moon}}{\text{molar flow (mlet)}} \text{C} \\ \begin{array}{c} \text{CH}\_{4} \stackrel{\text{molar flow (mlet)}}{\text{molar flow (outlet)}} \end{array} \text{100} \\ \begin{array}{c} \text{..} \\ \text{..} \end{array} \end{array} \text{101}$$

$$\frac{H\_2}{CO} \text{(molar ratio)} = \frac{H\_2 \text{molar flow (outlet)}}{CO \text{ }molar flow \text{ (outlet)}}$$

# RESULTS AND DISCUSSION

The catalysts have been analyzed before reaction (section Physicochemical Characterization of Calcined and Reduced Samples) by XRD, adsorption-desorption of N<sup>2</sup> at <sup>−</sup>196◦C, TPR, and XPS and after reaction (section Physicochemical Characterization of Used Catalysts) by XRD, Raman spectroscopy and XPS. Finally, the results from the activity tests are presented and discussed (section Activity Tests).

# Physicochemical Characterization of Calcined and Reduced Samples Structural Properties

### X-ray diffraction patterns of nickel-supported catalysts prepared by the wetness impregnation method and acquired after calcination (before reduction) are depicted in the first line of **Figure 1**. These diagrams show the diffraction lines corresponding to each support (Al2O3: 00-048-0367; CeO2, cubic phase: 01-075-0076; La2O3, hexagonal: 01-074-2430;

MgO, cubic phase: 01-075-0447 and monoclinic ZrO2: 01-078- 1807). The crystallinity of the supports is very different, being higher for the MgO-supported catalyst. Only for the ceria- and zirconia-supported catalysts is it possible to distinguish the main diffraction line related to NiO crystallites (cubic phase (01-073- 1519), corresponding to the (200) diffraction plane. By applying the Scherrer equation, an average domain of NiO crystallites was measured (**Table 1**). On the other hand, the crystalline domains differ among the different supports, although being smaller and quite similar for the alumina and lanthana-supported catalysts.

For the Ni/MgO and Ni/Al2O<sup>3</sup> catalysts, the peaks corresponding to nickel oxide overlap with those corresponding to the respective supports. SimilarQ16ly, in the Ni/La2O<sup>3</sup> catalyst, the incorporated NiO reacts on the La2O<sup>3</sup> surface to form a LaNiO<sup>3</sup> phase (00-012-0751), whose peaks overlap with those corresponding to La2O3. On the other hand, for the Ni-MgO sample, the formation of a NiO-MgO solid solution cannot be discarded since their diffraction lines coincide with those of MgO.

The Rh-Ni/Al2O<sup>3</sup> catalyst prepared by impregnation shows a diffractogram similar to that of the Ni/Al2O<sup>3</sup> catalyst (not displayed). In order to determine the particle size of nickel and rhodium particles, the Rh-Ni/Al2O<sup>3</sup> catalyst was reduced according to the same activation procedure used for the activity tests and studied by STEM, **Figure 2**. Rh NPs appear as finely dispersed particles with an average size less than 1 nm. Ni NPs are also observed and present larger size (average around 8 nm).

In addition, the XRD patterns of Rh-promoted Ni-supported catalysts, prepared by solid state reaction, were recorded. The XRD patterns of the calcined ones (before reduction) are shown in **Figure 1** (line 2). As for the catalysts prepared by impregnation, the diffraction lines are mainly ascribed to the corresponding supports. In the case of the lanthanasupported catalyst, the formation of the LaNiO<sup>3</sup> phase cannot be discarded. For the catalyst supported on magnesia, the diffraction lines corresponding to MgO and a Ni-Mg-O solid solution cannot be distinguished because of their strong overlap (Arena et al., 1996). The formation of the Ni-Mg-O solid solution is due to the relatively high calcination temperature used in the catalyst preparation. Concerning the zirconia-supported catalyst, for this preparation method, zirconia crystallizes in the tetragonal phase.

Rh reflections were not observed, as expected from the very small crystallite size. Concerning the presence of crystalline nickel species, only in the case of the catalyst supported on CeO<sup>2</sup> was it possible to observe a small diffraction line of the (2 0 0) plane of NiO. The crystallinity of the catalysts considerably changed; being higher for the MgO-supported catalyst and lower for the Al2O3-supported one. The XRD diagram of this last catalyst indicates its amorphous nature, since the calcination temperature was not high enough to get a crystalline structure of alumina. The temperature range needed to form the gamma phase of alumina is 600–875◦C (Sathyaseelan et al., 2013).

The Scherrer equation was applied to these X-ray diffraction diagrams, and the average domain sizes of the different crystalline phases are shown in **Table 2**. For these calcined samples, prepared by solid-state reaction, the average domain size among the

TABLE 1 | Average domain sizes for calcined samples (before reduction) prepared by impregnation (determined by the Scherrer equation).


different supports did not change as much as that which occurred for the calcined samples prepared by impregnation of commercial supports. Moreover, it is noteworthy that the crystalline domains of NiO, if they could be observed, were smaller than those obtained for the samples prepared by impregnation of the commercial supports (**Table 1**).

This figure also shows the XRD patterns of the reduced catalysts (**Figure 1**, line 3). The diffraction lines correspond


TABLE 2 | Average domain sizes for calcined samples (before reduction), fresh reduced catalysts and catalysts after reaction prepared by solid state reaction.

basically to that of the supports. In the case of the lanthanasupported catalyst, the XRD profile consists of overlapping lines of La2O3, La2NiO<sup>4</sup> (tetragonal, 01-079-0953) and La(OH)<sup>3</sup> (hexagonal, 01-083-2034) as a result of the strong interaction between nickel and lanthana (Pantaleo et al., 2015). Only for the catalyst supported on alumina and lanthana, low intensity peaks assigned to Ni metal [at 51.8◦ , plane (200)] can be distinguished. On the other hand, diffraction lines assigned to NiO are observed in the ceria-supported samples, indicating that at least part of the nickel phase cannot be reduced.

### Textural Properties

Nitrogen adsorption-desorption isotherms of Ni- and Rh-Nisupported catalysts are displayed in **Figure 3** and the surface areas are summarized in **Table 3**. These isotherms, which are of type IV, are assigned to mesoporous materials. All isotherms display a type H3 hysteresis loop, indicating that the catalysts contain a mesoporous network consisting of slittype pores. In the relative pressure range of 0.7–1, the loop originates from larger and usually disordered interparticle pores (Sing et al., 1985).

For the Ni catalysts prepared by impregnation, it can be seen that the adsorbed amount of N<sup>2</sup> is higher for the catalysts supported on alumina and magnesia and smaller for the ceria- and zirconia-supported ones. This is in accordance with the large crystallite size of these commercial supports and their low surface area (see **Table 3**), which in turn decreases the dispersion of the nickel phase, as can be derived from the X-ray diffraction diagrams (**Figure 1**) and the average crystalline domains determined for supported NiO NPs (**Table 1**).

For the catalysts supported on alumina and magnesia, a decrease in surface area is observed, which is explained by the blockage of part of the porous lattice by nickel oxide NPs. On the contrary, for the samples supported on ceria, lanthana, and zirconia, an increase in surface area is observed, which might be attributed to the contribution of these NiO surface NPs. The catalyst with the largest surface area is, by far, the one supported on alumina, followed by that supported on magnesia. The other catalysts show BET surface area values one order of magnitude lower, being the lowest that supported on ceria.

On the contrary, the Rh-Ni calcined catalyst, prepared by solid state reaction and supported on ceria, presents larger porosity as derived from the higher extent of nitrogen adsorption and the larger hystheresis loop shifted to high relative P/P<sup>0</sup> pressure, indicating the formation of larger pores.

The alumina-supported Rh-Ni calcined catalyst prepared by solid state reaction, presents a H<sup>2</sup> hysteresis loop in the range of small mesopores at relative pressures of 0.4–0.6, indicating that the catalysts contain complex mesoporous networks consisting of pores with ill-defined shapes. From the nitrogen adsorption values, it can be seen that this preparation method gives catalysts with more uniform textural properties than those prepared by impregnation of commercial supports. This is reflected in the surface areas that are between 6 and 39 m<sup>2</sup> /g for the samples used in the reaction, the trend being as follows: Rh-Ni/Al2O<sup>3</sup> > Rh-Ni/CeO<sup>2</sup> > Rh-Ni/ZrO<sup>2</sup> > Rh-Ni/MgO ∼ Rh-Ni/La2O3. The highest surface area was obtained for the ceria-supported catalyst as a consequence of the formation of small nanocrystals (Kundakovic and Flytzani-Stephanopoulos, 1998).

The reduction of these Rh-Ni catalysts produced a change in surface area with respect to the corresponding calcined counterparts, increasing for lanthana and magnesia-supported catalysts and decreasing for the other ones. The used catalysts showed a decrease in surface area due to sintering, with the exception of those supported on alumina and zirconia, for which the surface area slightly increased.

FIGURE 3 | N<sup>2</sup> adsorption-desorption isotherms of the calcined samples (Ni catalysts, prepared by impregnation and Rh-Ni catalysts prepared by solid-state reaction).

### Redox Properties

The reduction profiles of the different catalysts prepared by impregnation and by solid state reaction are shown in **Figure 4**. The reduction profiles of Ni catalysts prepared by impregnation are depicted in **Figure 4A**. The H<sup>2</sup> consumption profile corresponding to the alumina-supported sample shows a very small contribution around 300◦C due to the reduction of free NiO particles and wide consumption between 400 and 850◦C, which indicates the formation of nickel oxide particles with different interactions with the support. The consumption at temperatures higher than 700◦C is related to the reduction of nickel aluminate (Poncelet et al., 2005).

For the calcined Ni/CeO<sup>2</sup> sample, two peaks are observed: the first one is centered around 320◦C and is ascribed TABLE 3 | BET surface areas of commercial supports, of calcined Ni samples prepared by impregnation, and Rh-Ni samples prepared by solid state reaction (calcined, reduced and after reaction).


to the reduction of free NiO NPs, while the second and wider one, centered around 780◦C, is associated with the reduction of surface ceria and NiO NPs strongly interacting with the support. The consumption at the highest recorded temperature is explained by the reduction of some bulk ceria (Zhang et al., 2008).

The reduction profile of Ni-La2O<sup>3</sup> calcined sample is characterized by two consumption peaks, the first one centered around 330◦C and the second one around 580◦C, both due to the 2-step reduction of the perovskite LaNiO<sup>3</sup> (see Equations 5, 6), formed during the calcination step (Barbero et al., 2003) .

$$LaNiO\_3 + 1/2H\_2 \to LaNiO\_{2.5} + 1/2H\_2O \tag{5}$$

$$LaNiO\_{2.5} + H\_2 \to \ 1/2La\_2O\_3 + Ni + H\_2O \tag{6}$$

In the first step at lower temperature, Ni3<sup>+</sup> species stabilized in the perovskite structure are reduced to Ni2+, and the resulting compound, LaNiO2.5, is an oxygen-deficient structure. In the second reduction step, a complete reduction of nickel to metallic nickel is achieved, and a system based on finely dispersed Ni<sup>0</sup> particles supported on a La2O<sup>3</sup> matrix is obtained.

The reduction profile of the Ni/MgO calcined sample is characterized by very low H<sup>2</sup> consumption, indicative of an extensive formation of a Ni-Mg-O solid solution (Kirillov et al., 2011), which is difficult to reduce under the conditions used in TPR (Temperature Programmed Reduction) analysis. The peak around 330◦C is related to the reduction of free NiO particles. The consumption between 400 and 800◦C is assigned to the reduction of Ni2<sup>+</sup> species in the outer or subsurface layers of the MgO lattice. For higher temperatures, H<sup>2</sup> consumption might be due to the reduction of Ni2<sup>+</sup> taking place in the Ni-Mg-O solid solution (Wang et al., 2009a). For Ni/ZrO2, a main reduction peak centered at 320◦C, with a shoulder around 350◦C can be

found. The main H<sup>2</sup> consumption is due to the reduction of free NiO NPs and the minor one, at some higher temperature, to the reduction of NiO strongly interacting with the support (Wang et al., 2009b).

**Figure 4B** depicts the reduction profiles of the calcined samples Ni/Al2O3, Rh/Al2O3, and Rh-Ni/Al2O3, prepared by impregnation of commercial alumina. The reduction profile of Ni/Al2O<sup>3</sup> has already been described above. The reduction profile of the Rh/Al2O<sup>3</sup> sample shows two small contributions ascribed to the reduction of Rh2O<sup>3</sup> to Rh<sup>0</sup> . As reported in the literature, it is expected that calcination at 500◦C causes a fraction of the Rh phase to interact strongly with alumina, with Rh becoming incorporated into the surface layer of this support and, thus, being difficult to reduce. Therefore, the peak around 125◦C is attributed to the reduction of rhodium oxide particles containing low interaction with the support, and the second one, around 280◦C, is due to the reduction of rhodium oxide species with strong interaction with alumina (Burch et al., 1996). The profile of the Rh-Ni/sample reflects the reduction of both Rh oxide and Ni oxide species. However, a shift to lower temperatures as compared to that of Ni/Al2O<sup>3</sup> can clearly be observed. This is because the presence of noble metals facilitates the reduction of nickel oxide via a hydrogen-spillover mechanism (Tanaka et al., 2010b; Berger-Karin et al., 2011).

**Figure 4C** displays the TPR profiles of the Rh-Ni calcined samples prepared by solid-state reaction. The profile of the sample supported on alumina shows two contributions, a minor one around 310◦C, attributed to the reduction of free NiO particles, and a second one, very wide and centered around 550◦C, assigned to the reduction of Ni species strongly interacting with the support. As commented above, the reduction of nickel aluminate occurred at temperatures greater than 800◦C. For the Rh-Ni/CeO<sup>2</sup> sample, the profile shows three different peaks: the lowest temperature peak, around 125◦C, is assigned to the reduction of free Rh2O<sup>3</sup> NPs. As the area of this peak is much higher than that corresponding to the total reduction of Rh in the sample, the consumption of hydrogen by ceria must be considered (Li et al., 2016).

The second one would be an overlapping peak comprising the reduction of Rh species strongly interacting with the support and the reduction of free NiO particles, while the third one, in the range of 650–800 ◦C, would be due to the reduction of complex NiOx species strongly interacting with the support, as well as surface ceria reduction. Finally, at temperatures higher than 800◦C, the TCD signal is attributed to the reduction of bulk CeO<sup>2</sup> species. Comparison of the reduction profiles of Rh-Ni/CeO<sup>2</sup> and Ni/CeO<sup>2</sup> indicates that the presence of Rh promotes CeO<sup>2</sup> reduction, since the reduction profile of the Rh-Ni sample is shifted to lower temperatures, which can be attributed to hydrogen spillover during RhO<sup>x</sup> reduction (Ocsachoque et al., 2016).

The reduction profile of the lanthana-supported Rh-Ni also showed two components—similar to what occurred in the Rhfree counterpart—due to the reduction of LaNiO<sup>3</sup> which is known to occur in two stages. As commented above, the first was due to the reduction of Ni3<sup>+</sup> to Ni2+, which is shifted to lower temperatures in comparison to Ni/La2O3, and the second saw LaNiO2.5 be reduced to finely dispersed metallic nickel NPs supported on La2O3.

The TPR profile of Rh-Ni/MgO is characterized by very low H<sup>2</sup> consumption, as was the case for the Ni/MgO, which is due to the formation of Ni-Mg spinel in a larger extent during the calcination step. The small contribution found that around 300◦C is attributed to the reduction of free NiO species. A small and wide peak was observed in the range of 400– 750◦C, which is associated to the reduction of Ni2<sup>+</sup> interacting with the MgO lattice in external and subsurface layers. H<sup>2</sup> consumption at temperatures higher than 750–800◦C is likely associated to the reduction of Ni2<sup>+</sup> in NiO-MgO solid-solution (Wang et al., 2009a). The reduction profile of Rh-Ni/ZrO<sup>2</sup> shows three contributions: the first one, centered around 275◦C, is assigned to the reduction of Rh oxide interacting with the support and also to free NiO particles; the second one, around 400◦C, is due to surface nickel species interacting with the support, whereas the high temperature and wide reduction peak is due to the reduction of Ni2<sup>+</sup> species inserted into the bulk zirconia lattice (Singha et al., 2016).

The interaction of nickel species with the support is one of the factors that influence the reactivity of the catalysts. In principle, a strong interaction among both phases is beneficial to stabilize the supported metal NPs, decreasing the deactivation by sintering (Ruckenstein and Wang, 1999).

# Surface Composition

The binding energies of the Ni 2p, Al 2p, Ce 3d5/2, La 3d5/2, Mg 2p, and Zr 3d5/<sup>2</sup> core levels, surface atomic ratios of Ni/(support cation) and metallic Ni and nickel oxide proportions have been determined by XPS. For the samples supported on ceria and lanthana, the analysis of the Ni 2p3/<sup>2</sup> level is difficult since this level overlaps with the La 3d3/<sup>2</sup> and Ce 3d5/<sup>2</sup> levels and the quantification was done after careful deconvolution of both spectra.

The Ni 2p XPS core level region was fitted with three doublets assigned to metallic Ni (2p3/2, 852.5 eV), Ni2<sup>+</sup> (2p3/2,855– 856.2 eV), and satellite features (861, 878.2 eV). An example of a fitted spectrum for the Ni/MgO sample is shown in **Figure 5A**. The percentage of phase content for each metallic and metal oxide species was calculated by integrating the fitted XPS data of each sample before and after the reaction. The evolution of the Ni oxidation state is shown in **Figure 5B**. Thus, for the calcined Ni samples, the concentration of surface metallic nickel shows the following trend: La2O<sup>3</sup> > CeO<sup>2</sup> > MgO > Al2O<sup>3</sup> > ZrO2. The highest metallic nickel exposure was extracted for the sample supported on lanthana, related to the formation of LaNiO3, in which nickel is atomically distributed in the perovskite lattice. In the other extreme is the sample supported on zirconia, with the lowest metallic nickel proportion. For this series, the trend found for the surface Ni/substrate ratio is the following: La2O<sup>3</sup> > MgO > ZrO<sup>2</sup> > CeO<sup>2</sup> > Al2O3. The highest value found for the catalyst supported on lanthana is explained by the atomic nickel dispersion achieved in LaNiO<sup>3</sup> perovskite. The others follow the same trend of the surface areas of the substrates, with lower dispersion being found with a smaller surface area of the support, with the exception of the Ni/Al2O<sup>3</sup> sample, for which the formation of some proportion of nickel aluminate is not discarded.

For the calcined Rh-Ni samples, the binding energy found for the Ni 2p3/<sup>2</sup> level, around 856 eV and accompanied by a satellite peak around 6 eV higher, is characteristic of Ni2<sup>+</sup> species. Rh was not analyzed since its low amount makes it difficult to be detected. The binding energies found for the support cation indicate that the surface support phases are as follows: Al2O3, CeO2, La2O3, MgO, and ZrO2. The C 1s spectra show a peak around 290 eV, corresponding to surface carbonates for the lanthana and magnesia-supported catalyst precursor, being much intense for the lanthana-supported one. The Ni/M (M: Al, Ce, La, Mg, Zr) surface ratios are compiled in **Table 4**. These ratios change considerably depending on the support type. The trend found is CeO<sup>2</sup> > La2O<sup>3</sup> > ZrO<sup>2</sup> > Al2O<sup>3</sup> > MgO. The higher ratios obtained for ceria and lanthana-supported catalyst precursors are characteristic of a highly dispersed nickel phase. The low Ni/Mg ratio is due to the formation of a Ni-Mg-O solid solution (Arena et al., 1996; Barbero et al., 2003). The C 1s core-level spectrum of the lanthana-supported catalyst precursor shows a component around 289–290 eV due to the presence of carbonates, which is in accordance with the strong basic character of this oxide.

# Physicochemical Characterization of Used Catalysts Structural Properties

Rh-Ni catalysts after reaction have been structurally characterized by X-ray diffraction. X-ray diffraction patterns, depicted in **Figure 1**, showed that the reaction lead to structural changes in all the catalysts, as observed by comparing with the XRD patterns of fresh reduced samples. An increase in crystalline domain size of the support can be observed for the catalysts Rh-Ni/CeO2, Rh-Ni-MgO, and Rh-Ni/ZrO<sup>2</sup> (**Table 2**). For the catalyst supported on lanthana, the support is based on lanthanum oxycarbonate (Requies et al., 2005; Navarro et al., 2007a), being formed by a reaction between La2O<sup>3</sup> and CO2. Except for the RhNi/MgO catalyst, the other diffractograms show lines corresponding to metallic nickel, with its crystalline domain size being higher for the catalyst supported on ceria and smaller for the one supported on alumina. On the other hand, for the catalyst supported on ceria, diffraction lines corresponding to the formation of oxygen defective ceria (CeO2−x) can be observed (peak around 26.2◦ , ascribed to the diffraction plane (2 2 2), PDF card 00-049-1415).

### Carbon Formation

For CPOM, carbon deposition can take place by the Boudouard reaction (2 CO 1 C + CO2) and by the direct decomposition of methane (CH<sup>4</sup> 1 C + 2 H2), (Pena et al., 1996). The characterization of carbon deposits on the spent catalysts was carried out by Raman spectroscopy (spectra are shown in **Figure 6**). For each sample, at least three Raman spectra were recorded in different areas to assure the homogeneity of the composition.

Raman spectra of carbon exhibited two quite sharp modes, the G (graphitic) peak around 1590 cm−<sup>1</sup> and the D (disordered) peak around 1350 cm−<sup>1</sup> , which is related to the formation of carbon nanoparticles, amorphous carbon, or defective carbon filaments (Ozdemir et al., 2010; Lopez-Fonseca et al., 2012). The ceria-supported catalysts showed low intense peaks of carbon deposits. This is due to the widely reported oxygen mobility of the ceria surface (Dong et al., 2002). The low carbon formation in lanthana-supported catalysts is ascribed to the surface La2O2CO<sup>3</sup> species, well known as gasifier agents of C precursors.

As observed in **Figure 6**, in the samples in which peaks corresponding to carbon deposits are found, with the exception of Ni/ZrO2, the graphitic band is higher than that of the disorder carbon one, particularly in the catalyst Rh-Ni/ZrO<sup>2</sup> (Song et al., 2008). This spent catalyst presents an additional band around 1570 cm−<sup>1</sup> as a consequence of the splitting of the G band, which is characteristic of carbon nanotubes (Kogler et al., 2014). The formation of different types of carbon seems to be related to the crystal phase of zirconia. As reported in the literature, a

TABLE 4 | Binding energies (eV) and atomic surface ratios of calcined RhNi samples (M: Al, Ce, La, Mg, Zr) prepared by solid state reaction.


monoclinic phase favors the formation of encapsulating carbon and, for the tetragonal phase, the formation of carbon nanotubes is more likely (Zhang et al., 2015).

### Surface Composition

Ni-supported catalysts were analyzed after reaction by XPS. **Figure 5** shows the content of metallic Ni (%) and the nickel oxide (%) for each one. If these values are compared to those obtained for the catalysts before reaction (fresh calcined), a change of the surface metallic nickel exposure concentration can be observed depending on the catalyst type. Thus, for the samples supported on alumina and zirconia, an increase in the surface metallic nickel % is obtained, contrary to our observation for the other samples. On the other hand, in order to compare the evolution of the surface composition during reaction, another parameter has been compared for all the catalysts, which is depicted in **Figure 7** and represents the ratio between the total Ni 2p area in the catalyst after and before reaction. These results point out that the total surface Ni area of the Ni catalyst supported on lanthana does not evolve significantly during reaction; on the contrary, the catalyst supported on alumina undergoes the greatest decrease in total surface nickel concentration.

# Activity Tests

Ni-based catalysts prepared by impregnation were tested for the partial oxidation of methane to syngas and/or hydrogen.

CH<sup>4</sup> conversion, H<sup>2</sup> yield, CO selectivity and H2/CO molar ratio versus reaction time are depicted in **Figure 8**. The catalytic behavior highly depends on the type of support. The most active, selective and stable catalyst is the one on alumina. The counterpart supported on magnesia, which suffers a drastic deactivation, shows the worst catalytic behavior. The beginning of the reaction depends on the catalyst support.

The reactivity trends observed with the other catalysts are as follows: Ni/CeO<sup>2</sup> > Ni/ZrO<sup>2</sup> > Ni/La2O3. However, these catalysts suffer a different deactivation rate, but similar values for methane conversion, H<sup>2</sup> yield and CO selectivity were found after 6 h of reaction. The better performance showed by the Ni/Al2O<sup>3</sup> catalyst is mainly attributed to a relatively high level of nickel dispersion, favored by the higher surface area of the alumina support.

Although the metal-to-oxide ratio for Ni/La2O<sup>3</sup> (before and after reaction) (**Figure 5**) is much higher than the analogous ratios of the other catalysts, and the Ni 2p area barely decreases during reaction (**Figure 7**), in agreement with earlier findings (Requies et al., 2005), this catalyst does not appear to be very active for CPOM. Since the analysis of the Raman spectra rules out the carbon formation and therefore carbon deposition on nickel species, the observed deactivation should be due to sintering and/or the coverage of nickel species by lanthana carbonates, which are produced by the adsorption of CO<sup>2</sup> formed during the reaction (Requies et al., 2005). The deactivation of this catalyst is explained by its metal/support ratio (**Figure 5**), far from the optimum to favor the selectivity to methane partial oxidation, for this type of support. Furthermore, the oxidation of nickel particles at the beginning of the reaction would lead to lower selectivity toward syngas or hydrogen formation, increasing the selectivity toward total oxidation and, consequently, catalyst deactivation.

As reported in the literature, the catalytic performance of a catalyst based on Ni/MgO depends on its composition, preparation conditions, and even the properties of the MgO support (Hu and Ruckenstein, 2004; Nguyen et al., 2016). The fast deactivation observed for the Ni/MgO catalyst is due to the low fraction of metallic Ni on the catalyst surface (∼7%). Indeed, this Ni-loading is far below the optimal range (10-35%). As a consequence, this catalyst shows very low activity for CPOM (Ruckenstein and Hu, 1999), with only CO<sup>2</sup> and carbon as products after 4 h on-stream. Moreover, the nickel fouling leads to a substantial decrease in the Ni-support ratio (**Figure 5**). The deactivation of the Ni/ZrO<sup>2</sup> catalyst is mainly related to the formation of carbon deposits (**Figure 6**), which agrees with the results reported in the literature (Pengpanich et al., 2004; Larimi and Alavi, 2012).

The H2/CO ratios are higher than 2 and lower than 2.7, except for the catalyst supported on magnesia, because this catalyst behaves as a combustion catalyst. For the other catalysts, the trend found for the H2/CO ratio mainly depends on the extent of the Boudouard reaction (2 CO 1 CO<sup>2</sup> + C) and reverse water gas shift reaction, rWGS (CO<sup>2</sup> + H<sup>2</sup> 1 CO + H2O), with this ratio being higher when the Boudouard reaction occurs at a large extent and the rWGS reaction at a low extent (Albarazi et al., 2013). Thus, it appears that a direct correlation between the Raman peaks related to carbon formation (**Figure 6**) and the H2/CO ratio can be observed.

**Figure 9** shows activity tests with monometallic Ni and Rh catalysts supported on alumina and with the bimetallic counterpart. Methane conversions and H<sup>2</sup> yields clearly show that adding a small amount of Rh to the nickel catalyst results in an improvement in the catalytic performance of the Ni-based catalyst. This is reasonably explained by the higher reduction degree of surface nickel particles by the H<sup>2</sup> spillover originating on Rh NPs (see **Figure 4**). The Rh monometallic catalyst suffers from fast deactivation produced as a consequence of the low amount of active phase (Hohn and Schmidt, 2001; Berger-Karin et al., 2011).

For the catalyst series prepared by the solid-state reaction based on Rh and Ni as active phases (**Figure 10**), the best performance is also obtained with the catalyst supported on alumina. The ceria-supported catalyst showed also good performance with a H<sup>2</sup> yield similar after 6 h on stream, even having a lower surface area (**Table 3**). The low deactivation of the RhNi/CeO<sup>2</sup> catalyst by carbon formation is a consequence of the oxygen mobility and storage of the ceria support promoted by the redox pair Ce(IV)/Ce(III) (Pantaleo et al., 2016), which is known to play a key role in this reaction (Ding et al., 2015). The results obtained by O<sup>2</sup> pulses at reaction temperature point out that the extent of O<sup>2</sup> chemisorbed in this catalyst is 1.344 mmol/g, a much higher amount than on the other catalysts (RhNi/Al2O3: 0.127; RhNi/La2O3: 0.144; RhNi/MgO: 0.299; RhNi/ZrO2: 0.546 mmol/g).

The other catalysts undergo fast deactivation with time onstream. The deactivation of the catalyst supported on zirconia is explained by the fouling of active phases, since they underwent extensive carbon formation. The deactivation rate of the zirconiasupported bimetallic catalysts is higher than that observed for the monometallic counterpart. The different crystalline phases of zirconia (monoclinic for the monometallic catalyst, prepared by impregnation, and tetragonal, for the bimetallic one, prepared by solid state reaction), may have an influence on carbon formation. As reported in the literature, the ability of CO<sup>2</sup> activation, which is essential to remove the surface carbon species, is related to the crystal phase of ZrO<sup>2</sup> since different active sites are expected to be present on the different surfaces (Zhang et al., 2015). Thus, it has been reported that CO<sup>2</sup> reacts by the reverse Boudouard reaction, with carbon species resulting from methane dissociation to CO (Baerns et al., 1997). The higher CO<sup>2</sup> adsorption capacity of m-ZrO<sup>2</sup> is attributed to the

higher concentration and basicity of the hydroxyl groups on this polymorph, as well as to the stronger Lewis basicity of O2-anions (Pokrovski et al., 2001). This fast deactivation during CPOM has also been reported in a study using a monoclinic zirconiasupported Ni catalyst, also prepared by solid state reaction (Choudhary et al., 1998b).

Another factor influencing the reactivity of this catalyst is the sintering of nickel produced by hot spot formation in the catalyst bed, because the space velocity may not be high enough (Tang et al., 1998; Hohn and Schmidt, 2001). Since the interaction of metallic nickel particles on ZrO<sup>2</sup> is relatively weak, the catalysts are prone to deactivate by sintering of nickel particles (Barbero

et al., 2003). Such species can be observed in the XRD diffraction lines of the spent catalyst (**Table 2**).

For the catalyst supported on magnesia, in which mainly an unreduced Ni-Mg-O solid solution was found that was inactive for the CPOM reaction, the decrease in selectivity toward syngas or H<sup>2</sup> was influenced by the low surface proportion of metallic nickel (Requies et al., 2005). Therefore, this catalyst deactivated gradually due to its oxidation (Nurunnabi et al., 2006). Although this mixed oxide phase inhibits carbon deposition, especially from CO disproportionation (Tang et al., 1998), the surface content of nickel produced from Ni-Mg-O reduction is too low (or the MgO surface proportion too large). The analysis by Raman spectroscopy of this catalyst after reaction revealed the near absence of carbon, which runs contrary to the result obtained for the monometallic catalyst (Ni/MgO). This is explained by the role of Rh, which enhances the resistance to carbon deposition (Nurunnabi et al., 2006).

In the case of the catalyst supported on lanthana, it displays the worst catalytic performance, with an initial methane conversion of 50%, which is almost half the conversion achieved by the other catalysts. Its deactivation is not due to an extensive carbon formation, since only a small peak around 1,600 cm−<sup>1</sup> , attributed to graphitic carbon, is observed (**Figure 6**). In this catalyst, carbon is gasified by lanthanum oxycarbonate species (**Figure 1**) that limits deactivation to some extent.

In summary, the most active and selective catalysts are those supported on alumina and ceria. The comparison between Ni/Al2O<sup>3</sup> and RhNi/Al2O3, as well as between Ni/CeO<sup>2</sup> and RhNi/CeO2, points out that the bimetallic ones are more active, selective and stable. This is influenced by the greater metalsupport interface achieved in these catalysts, prepared by solid state reaction, by the higher intrinsic activity of Rh and its role favoring the reduction of Ni.

The catalysts that have showed better activity per surface area for CPOM under the reaction conditions used in this research are those supported on ceria (Ni/CeO<sup>2</sup> and RhNi/CeO2). In these catalysts, methane is dissociated on nickel particles, and C species migrate to the interface with the support to form CO (Dong et al., 2002). The bimetallic catalyst showed higher methane conversion, higher selectivity to H<sup>2</sup> and syngas and higher stability as a consequence of the larger metal-support interface. One of the properties that improved the catalytic behavior of this system is the non-stoichiometry of ceria, in which there is more oxygen mobility as compared to the other supports, which plays a key role in oxygen adsorption and carbon gasification (Pengpanich et al., 2004).

# CONCLUSIONS

Catalysts based on Ni, Rh, and Rh-Ni supported on Al2O3, CeO2, La2O3, MgO, and ZrO<sup>2</sup> have been prepared by wet impregnation of commercial supports and by solid state reaction. The catalysts have been tested for CPOM and it was found that their performance depends on the support type and on the preparation method. It can also be observed that a small amount of Rh promotes the reduction of Ni species, improving the catalytic performance. Regardless of the preparation method, the most active, selective and stable catalysts are those supported on alumina, which is mainly due to the higher dispersion of the nickel particles. The Ni/CeO<sup>2</sup> catalyst shows the best behavior per surface area, in which the lack of stoichiometry of the CeO<sup>2</sup> is a key property influencing the reactivity and stability. The catalyst composition of RhNi-CeO<sup>2</sup> prepared by solid state reaction

yields higher activity and stability than that of Ni-CeO2, prepared by impregnation. The fast deactivation of the catalysts supported on magnesia is explained by the low amount of active Ni sites, which leads to the combustion of methane, rather than to its selective oxidation. For the catalysts supported on lanthana, the deactivation is ascribed to the nickel content being lower than the optimum required amount on this type of support, resulting in the oxidation of nickel. The deactivation of the zirconiasupported catalysts is ascribed to the extensive formation of carbon, being less pronounced in the Ni/ZrO<sup>2</sup> catalyst due to the higher basicity of monoclinic zirconia in comparison to the tetragonal phase of RhNi/ZrO2.

# DATA AVAILABILITY

All datasets generated for this study are included in the manuscript and/or the supplementary files.

# AUTHOR CONTRIBUTIONS

CA-G, MM, LR-M, and JE contributed to the preparation, characterization of catalysts and catalytic activity tests. RN

# REFERENCES


contributed to tune up the reaction system and GC. MA, BR, and JF discussed the XPS results. CA-G discussed the results obtained by the other characterization techniques and the catalytic activity tests data.

# FUNDING

CA-G gratefully acknowledges financial support from CSIC (Consejo Superior de Investigaciones Científicas) by a grant (PA1003284) through the Program Programa de Ayudas a Investigadores del CSIC para la realización de estancias en centros de investigación extranjeros.

# ACKNOWLEDGMENTS

CA-G would like to thank all the members of Roldan Group for their hospitality during her stay at University of Central Florida (Orlando, USA). We acknowledge support toward the publication fee by the CSIC Open Access Publication Support Initiative through its Unit of Information Resources for Research (URICI).


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Alvarez-Galvan, Melian, Ruiz-Matas, Eslava, Navarro, Ahmadi, Roldan Cuenya and Fierro. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Pt/CeO<sup>2</sup> and Pt/CeSnO<sup>x</sup> Catalysts for Low-Temperature CO Oxidation Prepared by Plasma-Arc Technique

Tatyana Y. Kardash1,2 \*, Elizaveta A. Derevyannikova<sup>1</sup> , Elena M. Slavinskaya1,2 , Andrey I. Stadnichenko1,2, Vasiliy A. Maltsev <sup>3</sup> , Alexey V. Zaikovskii <sup>3</sup> , Sergey A. Novopashin<sup>3</sup> , Andrei I. Boronin1,2 \* and Konstantin M. Neyman4,5,6 \*

<sup>1</sup> Boreskov Institute of Catalysis, Siberian Branch of the Russian Academy of Sciences, Novosibirsk, Russia, <sup>2</sup> Novosibirsk State University, Novosibirsk, Russia, <sup>3</sup> Kutateladze Institute of Thermophysics, Siberian Branch of the Russian Academy of Sciences, Novosibirsk, Russia, <sup>4</sup> Departament de Ciència dels Materials i Química Física, Universitat de Barcelona, Barcelona, Spain, <sup>5</sup> Institut de Química Teòrica i Computacional, Universitat de Barcelona, Barcelona, Spain, <sup>6</sup> ICREA (Institució Catalana de Recerca i Estudis Avançats), Barcelona, Spain

### Edited by:

Javier Carrasco, CIC energigune, Spain

### Reviewed by:

Arturo Martinez-Arias, Instituto de Catálisis y Petroleoquímica (ICP), Spain Jose A. Rodriguez, Brookhaven National Laboratory (DOE), United States

### \*Correspondence:

Tatyana Yu. Kardash kardash@catalysis.ru Andrei I. Boronin boronin@catalysis.ru Konstantin M. Neyman konstantin.neyman@icrea.cat

### Specialty section:

This article was submitted to Physical Chemistry and Chemical Physics, a section of the journal Frontiers in Chemistry

Received: 26 November 2018 Accepted: 13 February 2019 Published: 12 March 2019

### Citation:

Kardash TY, Derevyannikova EA, Slavinskaya EM, Stadnichenko AI, Maltsev VA, Zaikovskii AV, Novopashin SA, Boronin AI and Neyman KM (2019) Pt/CeO2 and Pt/CeSnOx Catalysts for Low-Temperature CO Oxidation Prepared by Plasma-Arc Technique. Front. Chem. 7:114. doi: 10.3389/fchem.2019.00114 We applied a method of plasma arc synthesis to study effects of modification of the fluorite phase of ceria by tin ions. By sputtering active components (Pt, Ce, Sn) together with carbon from a graphite electrode in a helium ambient we prepared samples of complex highly defective composite PtCeC and PtCeSnC oxide particles stabilized in a matrix of carbon. Subsequent high-temperature annealing of the samples in oxygen removes the carbon matrix and causes the formation of active catalysts Pt/CeO<sup>x</sup> and Pt/CeSnO<sup>x</sup> for CO oxidation. In the presence of Sn, X-Ray Diffraction (XRD) and High-Resolution Transmission Electron Microscopy (HRTEM) show formation of a mixed phase CeSnO<sup>x</sup> and stabilization of more dispersed species with a fluorite-type structure. These factors are essential for the observed high activity and thermic stability of the catalyst modified by Sn. X-Ray Photoelectron Spectroscopy (XPS) reveals the presence of both Pt2<sup>+</sup> and Pt4<sup>+</sup> ions in the catalyst Pt/CeOx, whereas only the state Pt2<sup>+</sup> of platinum could be detected in the Sn-modified catalyst Pt/CeSnOx. Insertion of Sn ions into the Pt/CeO<sup>x</sup> lattice destabilizes/reduces Pt4<sup>+</sup> cations in the Pt/CeSnO<sup>x</sup> catalyst and induces formation of strikingly high concentration (up to 50% at.) of lattice Ce3<sup>+</sup> ions. Our DFT calculations corroborate destabilization of Pt4<sup>+</sup> ions by incorporation of cationic Sn in Pt/CeOx. The presented results show that modification of the fluorite lattice of ceria by tin induces substantial amount of mobile reactive oxygen partly due to affecting geometric parameters of ceria by tin ions.

Keywords: platinum, ceria, tin, CO oxidation, plasma arc synthesis, DFT calculations, Pt/CeO2 catalyst

# INTRODUCTION

Pollution of biosphere with toxic emissions is becoming increasingly urgent due to the rapid increase in electric and thermal power production, the production growth in metallurgical and chemical industries as well as growing number of vehicles. Among the most effective methods of solving this problem is efficient catalytic oxidation of CO and hydrocarbons (Farrauto and Bartholomew, 1997; Muraki and Zhang, 2000). The increasing restrictions of the standards for the automotive emissions require eitherimprovement of existing catalysts or creation of fundamentally new ones (Trovarelli, 1996; Montini et al., 2016; Trovarelli and Llorca, 2017).

Broadening temperature ranges of catalysts action and maintaining their activity are important challenges therefore. In this regard, the use of conventional three-component formulations based on PGM-CeO2-MO<sup>x</sup> (where PGM are platinum group metals and MO<sup>x</sup> are modifying metal oxides) is still promising for catalytic oxidation of CO and hydrocarbons. Palladium is the most commonly used PGM. It is characterized as a PGM very strongly interacting with CeO2. According to Fornasiero et al. (Cargnello et al., 2012; Monai et al., 2015), PdO nanoparticles decorated with cerium oxide (core-shell structures Pd@CeO2) are promising materials for the efficient oxidation of methane and other hydrocarbons. Such Pd@CeO<sup>2</sup> structures are capable of oxidizing methane at 200–250◦C. However, Farrauto (Farrauto, 2012) pointed out that due to the phase transition PdO-Pd at ∼800◦C the creation of thermally stable (T > 800◦C) catalysts on the basis of this Pd form is difficult. Such a conclusion emphasizes the importance of the strong metalsupport interaction (SMSI) effect for palladium, platinum and other PGMs for creation of thermally stable catalysts (Figueroba et al., 2016, 2017; Lykhach et al., 2017a).

The realization of the SMSI effect in PGM-CeO<sup>2</sup> systems depends crucially on the synthesis procedure. Different synthetic approaches are used to obtain PGM-CeO<sup>2</sup> catalysts with desired improved characteristics (Boronin et al., 2009; Gulyaev et al., 2012, 2014b; Slavinskaya et al., 2015, 2016, 2018; Vasilchenko et al., 2016; Derevyannikova et al., 2017; Kardash et al., 2017; Kibis et al., 2017). Novel physical preparation methods that utilize the action of thermal and electrical fields, such as laser ablation, magnetron sputtering, "solution-combustion" and plasma methods attracted attention of many research groups (Bera et al., 2003; Baidya et al., 2009; Hegde et al., 2009; Gupta and Hegde, 2010; Hinokuma et al., 2014; Slavinskaya et al., 2016, 2018; Vasilchenko et al., 2016; Stadnichenko et al., 2018, 2019). These methods allow obtaining of highly dispersed PGM forms (ions and clusters) on the surface and inside ceria, which ensures the maximal SMSI effect and appearance of new highly active surface species.

Recently we have developed and applied a plasma-arc method (PAS) for synthesis of Pd/CeO<sup>2</sup> catalysts (Gulyaev et al., 2014b; Kardash et al., 2017). The PAS method is based on the sputtering of a specific anode material in a vacuum chamber filled with an inert gas under action of a plasma arc discharge. The anode material consists of a graphite rod, in the center of which a mixture of dispersed carbon and PGM metal powders, cerium and/or other components is placed. Sputtering of the anode leads to the appearance of atomic components of carbon and atomically dispersed metals in the arc. Diffusion and convection of the components in a buffer gas leads to their heterogeneous condensation. The condensation products are deposited on the cooled screen. The synthesized material consists of nanoparticles embedded in a carbon matrix. Subsequent high-temperature annealing in oxygen converts the synthesized carbon-based composite into an active oxide catalyst for low-temperature CO oxidation (LTO CO).

The aim of this work was to establish, weather the PAS method also allows obtaining joint phase of Pt and CeO2, where Pt is in a highly dispersed ionic state similar to that of Pd in the Pd-doped CeO<sup>2</sup> joint phase PdxCe1−xO2−<sup>δ</sup> (Primavera et al., 1998; Priolkar et al., 2002; Scanlon et al., 2011; Gulyaev et al., 2014a; Neitzel et al., 2016). Assuming that stable Pt2<sup>+</sup> and Pt4<sup>+</sup> (Bruix et al., 2014) species are formed in the PAS synthesis, it was important to determine concentration of these in the prepared samples and their influence on the catalytic activity

In order to increase the thermal stability and redox ability of the catalysts, Sn was used as a modifying element of the catalysts. Sn4<sup>+</sup> is assumed to substitute Pt4<sup>+</sup> and Ce4<sup>+</sup> in the fluorite structure (Matolín et al., 2008; Zeng et al., 2012; Ayastuy et al., 2014). The formation of the joint Ce1−xSnxO<sup>y</sup> phase was expected to promote the generation of defects and oxygen vacancies in the prepared samples of catalysts, thus increasing their oxygen storage capacity (OSC) and the catalytic activity in LTO CO (Tolla et al., 1999; Nguyen et al., 2003; Gupta et al., 2010). Our previous studied of the Pd/CeO2-SnO2−<sup>x</sup> catalysts showed that introduction of Sn into PdCeC formulations during the PAS synthesis formed nano-heterogeneous composites after calcination of the PdCeSnC composite. The obtained catalysts were stable against sintering and crystallization to large particles (Kardash et al., 2017). As a result, LTO CO activity was preserved even after calcination at 1,000◦C. One of the objectives of this study was to determine the thermal stability of composite Pt/CeO2-SnO<sup>2</sup> catalysts and to compare their catalytic properties with those of the unmodified by Sn Pt/CeO<sup>2</sup> catalysts prepared under the same PAS conditions.

Also in this paper we present density functional (DFT) calculations of platinum species stabilized either by pure ceria nanoparticles or ceria nanoparticles modified by tin ions. The calculations of these models allowed establishing the most stable structures and the effect of tin substitution on the properties of Pt/CeO<sup>2</sup> nanoparticles. Our DFT calculations corroborate the effect of embedded tin on the state of platinum ions in the CeO<sup>2</sup> lattice.

Obtained experimental and theoretical data showed that the PAS method allows obtaining highly defective ceria particles and ionic platinum species. Modification with tin leads to increased thermal stability of platinum catalysts, higher amount of reactive oxygen in the catalysts and enhanced catalytic activity of the latter in the low-temperature regime of CO oxidation.

# EXPERIMENTAL

# Catalysts Synthesis

PAS method was applied to obtain PtCeC and PtCeSnC composite materials. We refer to the detailed description of the procedure in our previous work (Kardash et al., 2017). To obtain oxide catalysts, the synthesized PtCe(Sn)C composites were calcined in air at 600, 800, and 900◦C. **Table 1** presents the chemical composition and characteristics of the obtained Pt/Ce and Pt/CeSn catalysts.

# Methods

# The Specific Surface

The specific surface (SBET) of the samples was determined by BET method (Lowell et al., 2006), using argon thermal desorption with a Sorbtometr-M adsorption analyzer.


\*The rest is carbon and water.

# Analysis of the Chemical Composition

Analysis of the chemical composition was performed by X-ray fluorescence (XRF) method on an ARL PERFORM analyzer with a Rh anode of an X-ray tube. Chemical compositions of tin and cerium oxides were determined using calibrations; the determination of Pt was carried out using the external standard method.

### X-Ray Diffraction (XRD)

X-ray diffraction (XRD) patterns were recorded on a Bruker D8 Advance instrument using CuKα radiation and the Bragg-Brentano focusing geometry. The aperture of the Soller slits on the primary and reflected beams was 2.5◦ . LynxEye (Bruker) multi strip detector was employed for intensity measurements. Data acquisition was performed in the 2θ range of 15–100◦ , at a 0.05◦ step and counting time of 2 s. Phase analysis was carried out using the ICDD PDF 2 database (Powder Diffraction File PDF-2. International Center for Diffraction Data. USA. 2009). Structural data were taken from the structural database ICSD (Hellenbrandt, 2004). Rietveld refinement for quantitative analysis was carried out with the help of the software package Topas V.4.3 (Coelho, 2005). The instrumental broadening was described with metallic silicon as a reference material. The diffraction line profiles were analyzed using the fundamental parameter approach. The lengths of coherent scattering domain were calculated from LVol-IB values (i.e., volume weighted mean column lengths based on integral breadth).

### Electron Microscopy

Electron microscopy investigation was performed using JEM-2010 (JEOL Ltd., Japan) and JEM-2200FS (JEOL Ltd., Japan) electron microscopes operated at 200 kV for obtaining High Resolution Transmission Electron Microscopy (HRTEM) images. High-Angle Annular Dark-Field mode in a Scanning Transmission Electron Microscope (STEM HAADF) was employed together with Energy Dispersive X-ray (EDX) spectroscopy. Samples for the TEM study were prepared on the perforated carbon film mounted on a copper grid.

## Temperature Programmed Reaction With CO (TPR-CO)

The reaction mixture containing 1.0 vol.% CO, 0.5 vol.% Ne in flowing He was fed at a rate of 100 cm<sup>3</sup> /min to the catalyst sample (0.2 g) preliminary cooled in the reactor to −10◦C. As the steady-state concentrations of CO and CO<sup>2</sup> were established, the sample was heated from −10 to 450◦C at 10◦C/min heating rate. The concentrations of CO, CO2, O2, H2, and H2O were measured during the reaction. Before each TPR-CO experiment the catalysts were pretreated by 20%O2/He gas mixture at 450◦<sup>C</sup> during 2 h with subsequent cooling in this gas mixture. After the cooling the catalysts were purged in helium.

## Investigation of Catalytic Properties

Investigation of catalytic properties of the synthesized samples was performed in an automated setup with a flow reactor and mass-spectrometric analysis of the gas mixture using the temperature-programmed reaction (TPR-CO-O2). A sample with the particle sizes of 0.25–0.5 mm was mounted in a stainless steel reactor. The reaction mixture containing 0.2 vol.% CO, 1.0 vol.% O2, 0.5 vol.% Ne in flowing He was fed at a rate of 1,000 cm<sup>3</sup> /min to the initial catalyst cooled to −10◦C. For the Pt/Ce catalysts calcined at 600, 700, 800, and 900◦C weight of samples employed for catalytic testing was 0.3, 0.36, 0.4, and 0.62 g, respectively; for the Pt/CeSn catalysts calcined at 600, 700, 800, and 900◦C −0.3, 0.34, 0.35, and 0.34 g. The catalyst was heated in the reaction mixture from −10 to 450◦C at a rate of 10◦C/min with subsequent cooling and repeated heating in the reaction mixture. The concentrations of CO, O2, and CO<sup>2</sup> were monitored in the course of reaction at a frequency of 0.34 Hz.

### X-Ray Photoelectron Spectroscopy (XPS)

X-ray photoelectron spectroscopy (XPS) measurements used an ES-300 (KRATOS Analytical) photoelectron spectrometer equipped with a MgKα (hν = 1253.6 eV) radiation source. The spectrometer was calibrated using the Au4f7/<sup>2</sup> (84.0 eV) and Cu2p3/<sup>2</sup> (932.7 eV) lines of pure metallic surfaces of Au and Cu, respectively. The U′′′ component of the Ce3d spectral line (E<sup>b</sup> = 916.7 eV) served as a standard for calibration of XPS spectra. The program of processing spectral data created by us, XPS-Calc, was used for the mathematical treatment of the XPS spectra. This software has previously been applied for different systems, including Pt/CeO<sup>2</sup> and Pt/Al2O<sup>3</sup> catalysts (Ivanova et al., 2010; Svintsitskiy et al., 2015; Stadnichenko et al., 2017, 2019). The Shirley model and the Gauss–Lorentz functions were used for background subtraction and curve fitting, respectively.

# DFT Calculations

DFT calculations were carried out employing PW91 exchangecorrelation functional (Perdew et al., 1992) in the plane-wave VASP code (Kresse and Hafner, 1993) for periodic boundary conditions. Plane-wave basis sets with a cut-off energy of 415 eV were used. Core-valence electron interactions were described by projector-augmented wave method (Kresse and Joubert, 1999). In order to properly localize Ce4f electrons in partially reduced Ce3<sup>+</sup> ions, an on-site Coulombic correction (Ueff <sup>=</sup> 4.0 eV) (Anisimov et al., 1997; Dudarev et al., 1998; Loschen et al., 2007) was applied to 4f electrons in all Ce atoms. Our Pt- and Sn-containing models (see **Supplementary Material**) were based on a model stoichiometric ceria nanoparticle Ce40O<sup>80</sup> ca. 1.5 nm large (Migani et al., 2010). Our previous calculations of this type of models resulted in several predictions subsequently confirmed experimentally (Bruix and Neyman, 2016, 2019). Each model nanoparticle was positioned in a rectangular periodically repeated cell of 2.2 × 1.9 × 1.9 nm. Spin-polarized Ŵ-point calculations of all model nanoparticles under scrutiny were performed with the single-point total energy convergence tolerance of 2 × 10−<sup>5</sup> eV. Optimization of positions of all atoms constituting a given model was continued until the maximum forces acting on each atom were <0.02 eV/Å.

# RESULTS

# TEM Study of the As-Prepared by Plasma-Arc Synthesis PtCeC and PtCeSnC Composites

The composite material PtCeC obtained by PAS consists of highly dispersed nanoparticles embedded within a matrix of amorphous carbon. **Figure 1** shows typical TEM images of this material with 1 to 10 nm large nanoparticles. The analysis of interatomic distances features that nanoparticle structure corresponds to Ce2O<sup>3</sup> (ICDD PDF-2 #00-023-1048). The reflections on the Fourier image displayed in **Figure 1B** reveal the interplanar distances of 0.32 and 0.3316 nm corresponding to the (002) and (100) reflections of a Ce2O<sup>3</sup> trigonal structure. We detected a similar microstructure for a PdCeC composite prepared by PAS (Gulyaev et al., 2014b).

Morphology of the as-prepared PtCeSnC composite is characterized by the presence of two types of nanoparticles: a highly dispersed fraction with sizes 1–10 nm and a fraction of larger particles with sizes from 10 to 40 nm. **Figure 2A** shows typical images of the PtCeSnC composite.

Larger particles (see **Figure 2B**) exhibit a core-shell-shell structure. The outer shell has lower density than the core and is built of amorphous carbon. The core of the particles contains tin. The analysis of the crystalline structure of the inner particle core shows the distance 0.3 nm, which could be assigned to different phases: SnO<sup>2</sup> (Pbcn) [0.30 nm (113)], SnO (P4/nmm) [0.299 nm (101)], Sn (I41/amd) [0.29 nm (200)], see **Figure 2C**. The next crystalline shell of the particles shows the distances of 0.31 and 0.33 nm, which are unambiguously related to the (113) and (112) reflections of SnO<sup>2</sup> (Pbcn). The observation of higher density of the particle core suggests that the core structure is related to Sn. Hence, a large fraction of the core-shell-shell particles appears to feature a composition Sn-SnO2-C, where C is amorphous carbon.

Structural analysis of the dispersed particles of the as-prepared PtCeSnC composite showed the distances of 0.3 nm and 0.29 nm, which could be related to (002) and (101) reflections of the Ce2O<sup>3</sup> correspondently, see **Figure 2D**.

# Characterization of the Calcined Pt/Ce and Pt/CeSn Samples

Our earlier studies of catalysts prepared by PAS (Kardash et al., 2017) showed that calcination at temperatures above 600◦C causes burning of carbon in PtCeC and PtCeSnC composites and transforms them into oxide materials. The chemical composition analysis was performed for the calcined samples (see **Table 1**). One can see that after calcination at 600–800◦C the Pt/Ce sample exhibits a higher specific surface area than the Pt/CeSn sample. However, after calcination at 900◦C the specific surface area of the Pt/CeSn sample becomes higher than that of the Pt/Ce sample, which indicates a stabilizing effect of Sn preventing particle sintering.

**Figure 3** presents XRD patterns of the calcined Pt/Ce and Pt/CeSn samples. **Table 2** shows the results of a quantitative phase analysis obtained by the Rietveld refinement of the XRD data.

CeO<sup>2</sup> phase with a fluorite-type structure is a major component in the Pt/Ce sample. The lattice parameter of the fluorite phase and mean particle size estimated from the XRD data are shown in **Table 2**. For the Pt/CeSn sample, the oxides SnO<sup>2</sup> (cassiterite, ICDD PDF-2 #00-034-0394) and CeO<sup>2</sup> (ICDD PDF #00-029-1484) are the major phases. Note, that the CeO<sup>2</sup> fluorite phase exhibits smaller particles in the Pt/CeSn sample compared to the Pt/Ce sample for all calcination temperatures. This observation implies that addition of SnO<sup>2</sup> to the composite allows stabilizing the CeO<sup>2</sup> particle size against sintering.

Additional peaks were detected in the Pt/CeSn sample, corresponding to the unstable orthorhombic SnO<sup>2</sup> phase (ICDD PDF-2 #04-015-3275). According to the literature data, this form of SnO<sup>2</sup> can be built at high pressures and temperatures (Gracia et al., 2007). The fraction of the orthorhombic SnO<sup>2</sup> phase is decreased upon sample calcination. We assume that this unstable phase can result from a rapid oxidation of the PtSnCeC composites, when they are exposed to the atmosphere upon taking out of the synthesis chamber. This phase could be stabilized by the carbon matrix within core-shell-shell nanoparticles. However, the subsequent calcination of the composite transforms it into a stable cassiterite phase.

Metallic Pt phase was not detected in the samples calcined at 600◦C. Peaks corresponding to metallic Pt appear on the XRD patterns only after calcination of the Pt/CeSn sample at 700◦C and of the Pt/Ce sample at 900◦C. However, the Rietveld refinement shows that the content of metallic Pt is lower than the analytical amount of Pt in the sample. This indicates the presence of Pt in highly dispersed states in the samples calcined at 600◦C and even higher temperatures.

Total amount of SnO<sup>2</sup> in the Pt/CeSn sample is less than detected by the XRF analysis. It implies that a part of Sn is embedded in the CeO<sup>2</sup> lattice. This assumption is supported by decreasing of the fluorite lattice parameter from 5.411 Å for CeO<sup>2</sup> in the Pt/Ce sample to 5.401 Å in the Pt/CeSn sample. Relying on the difference in the radii of the Ce4+/Ce3<sup>+</sup> cations and Sn4<sup>+</sup> cations, it is possible to estimate the composition of the Ce1−xSnxO<sup>y</sup> joint fluorite phase. For the Pt/CeSn-600 sample, the estimated composition of the fluorite phase is

Ce0.93Sn0.07O2−δ. However, the so determined tin content in the fluorite structure is underestimated, since the calculation did not take into account that the fluorite lattice parameter can increase with decreasing particle size to 1–3 nm (Tsunekawa et al., 2004; Baranchikov et al., 2010).

In line with these considerations, the nearest 8 Sn-O and 8 Sn-Ce distances calculated by DFT for the model tindoped nanoparticles SnCe39O<sup>80</sup> and Pt/SnCe39O<sup>80</sup> with Sn4<sup>+</sup> dopant located in an inner Ce4<sup>+</sup> position are shorter than the corresponding Ce-O and Ce-Ce distances in the pristine

core-shell-shell particles Sn@SnO2@C (C). HRTEM image and Fourier image of the crystal lattice of small Ce2O3 particles within amorphous carbon matrix (D).

FIGURE 3 | XRD patterns of the Pt/Ce (A) and Pt/CeSn (B) samples calcined at different temperatures. The graph at the bottom shows the positions of CeO2 fluorite phase reflections and corresponding Miller indexes. Symbol (\*) indicates an unstable orthorhombic form of SnO2, the symbol (v) indicates reflections of a cassiterite SnO2 phase.

TABLE 2 | Phase composition of the Pt/Ce and Pt/CeSn samples calcined at different temperatures Tcalc, lattice parameter (a) and particle mean size (D) of the fluorite phase.


The standard deviations of the calculated values are given in parentheses.

Ce40O<sup>80</sup> particle (see **Figure 4** and **Supplementary Material**). Namely, the average distances Ce-O and Ce-Ce in Ce40O<sup>80</sup> are 2.326 Å and 3.741 Å, respectively, whereas the corresponding average distances Sn-O and Sn-Ce are 2.229 Å and 3.707 Å in the SnCe39O<sup>80</sup> particle and 2.222 Å and 3.706 Å in the Pt/SnCe39O<sup>80</sup> particle. Our calculations of the latter with the doping Sn occupying various cationic positions of the CeO<sup>2</sup> lattice showed (see **Supplementary Figure 1**) that the location of the Sn4<sup>+</sup> dopant inside the nanoparticle is moderately stabilized, by 14–36 kJ/mol, compared to its surface location on the (111) terrace of the nanoparticle. Interestingly, a notably more stable location was calculated for a surface corner position of the dopant, where it substitutes a partially reduced Ce3<sup>+</sup> cation and acquires a quite unusual Sn3<sup>+</sup> state. We note in passing that such special corner sites O<sup>4</sup> exposed by nanostructured ceria are able to bind very strongly atoms of various metals in the form of cations (Figueroba et al., 2016, 2017; Neitzel et al., 2016).

**Figures 5**, **6** show TEM images of the Pt/CeSn and Pt/Ce samples calcined at 600◦C. The TEM data corroborate that calcination of the composites causes removal of the amorphous carbon matrix and crystallization of the formed oxide structures. Two particle types are observed in the Pt/CeSn-600 sample: larger ones with the sizes of 20–50 nm and smaller ones with the sizes of 5–15 nm (**Figure 5A**). A detailed analysis of the structure of the material showed that the larger particles belong to SnO<sup>2</sup> (P42/mnm) (**Figure 5B**). The crystal structure analysis of the smaller particles revealed that they correspond to a cubic fluoritetype structure (**Figure 5C**). These findings agree with the XRD analysis data showing that fluorite-type particles are smaller than those of tin oxides type.

The cerium and tin distribution in the Pt/CeSn-600 sample was investigated by means of EDX. **Figure 6** displays HAADF-STEM images and the corresponding EDX maps for cerium and tin. As can be seen from the TEM images, cerium and tin are distributed over the entire volume of the sample, and no

FIGURE 4 | Geometric structures of CeO2-based nanoparticles calculated to analyze effects of doping Sn ions on the nearest distances from the latter to O and Ce ions. Color coding for atoms: Ce—beige, O—red, Sn—dark gray, Pt—light gray. The nearest distances to the Ce4<sup>+</sup> ion marked by <sup>+</sup> in the Ce40O<sup>80</sup> model: Ce-O-2.292, 2.292, 2.293, 2.295, 2.358, 2.358, 2.361, 2.361 (average 2.326); Ce-Ce-3.732, 3.734, 3.737, 3.738, 3.744, 3.746, 3.747, 3.752 (average 3.741). The nearest distances to the Sn4<sup>+</sup> ion in the SnCe39O<sup>80</sup> model: Sn-O-2.181, 2.184, 2.187, 2.189, 2.264, 2.269, 2.276, 2.280 (average 2.229); Sn-Ce-3.687, 3.687, 3.688, 3.693, 3.717, 3.724, 3.726, 3.730 (average 3.707). The nearest distances to the Sn4<sup>+</sup> ion in the Pt/SnCe39O<sup>80</sup> model: Sn-O-2.175, 2.176, 2.205, 2.210, 2.238, 2.246, 2.260, 2.266 (average 2.222); Sn-Ce-3.682, 3.691, 3.691, 3.700, 3.714, 3.717, 3.725, 3.726 (average 3.706). All the distances are in Å. Atomic coordinates and total energies of the displayed structures are given in the Supplementary Material.

image of the CeO2 fluorite phase (C).

individual CeO<sup>2</sup> particles were detected there. This is consistent with the XRD data for the Pt/CeSn-600 sample showing the formation of composites built from SnO<sup>2</sup> and Ce1−xSnxO<sup>2</sup> nanoparticles. The EDX data on the sample composition agree with the chemical analysis data.

**Figure 7** displays TEM images for the Pt/Ce-600 sample. Nanocrystalline blocks with sizes 5–10 nm are observed. The detected distances correspond to the cubic fluorite-type ceria.

XPS spectra of the Pt/Ce and Pt/CeSn samples calcined at 600 and 900◦C are shown in **Figures 8**, **9**. The quantitative analysis data and the Pt4f binding energy values determined from the XPS spectra for the Pt/Ce-600 and Pt/CeSn-600 samples are presented in **Table 3**. According to these spectroscopic data all samples contain significant amounts of carbon, concentration of which varies from 27 to 35 at. %. From the analysis of the C1s spectra (not shown) it follows that the detected carbon is represented mainly by amorphous carbon or hydrocarbon groups. The share of oxygen-containing carbon groups, including carbonates, is at most 5–10% (**Table 3**). Such notable content of carbon in the samples obtained by sputtering in an electric arc discharge is not surprising, since the sputtering of atomically dispersed metals and species from graphite carbon rod occurs simultaneously.

Analysis of the charge states of platinum in a series of samples was carried out using the spectra curves fitting. Fitted Pt4f spectra are presented in **Figure 8**. Upon decomposition into individual components, metallic platinum (E<sup>b</sup> = 71.1 eV) could not be detected in samples calcined at 600◦C. In the Pt/Ce-600 sample platinum is present in two states characterized by Eb(Pt4f7/2) = 72.7 and 74.5 eV. Formally, these states can be referred to as Pt2<sup>+</sup> and Pt4+. In the remaining samples in the presence of tin, only the state Pt2<sup>+</sup> with Eb(Pt4f7/2) <sup>=</sup> 72.8 eV was detected, but not the state Pt4+.

After calcination at 900◦C, the Pt2<sup>+</sup> species are preserved. However, a new state with Eb(Pt4f7/2) = 70.9 eV appears. This state is referred to as Pt<sup>0</sup> , and its appearance is in a good agreement with XRD data. After calcination at 900◦C, the surface Pt concentration decreases, which could be attributed to sintering of Pt particles.

**Figure 9** shows Ce3d spectra for the Pt/Ce-600 and Pt/CeSn-600 samples, and Sn3d for the Pt/CeSn-600 sample. The Sn3d spectra show only one main doublet with Eb(Sn3d5/2) = 486.9 eV attributable to Sn4<sup>+</sup> (Moulder et al., 1992). Ce3d spectra were decomposed into individual components of Ce4<sup>+</sup> species (V, V ′′, V′′′ and U, U′′, U′′′) and Ce3<sup>+</sup> species (V<sup>0</sup> , V′ and U<sup>0</sup> , U′ ) in accordance with literature data (Anandan and Bera, 2013; Zhu et al., 2014). The analysis shows that the quantity of Ce3<sup>+</sup> species is increased significantly in the Pt/CeSn-600 sample. In the Pt/Ce-600 sample, only 6 at.% of Ce3<sup>+</sup> is detected, whereas 49% of cerium is in the Ce3<sup>+</sup> state in the Pt/CeSn-600 sample.

Such a high amount of Ce3<sup>+</sup> species in the Pt/CeSn-600 sample could be explained by the XRD and TEM data, which show the formation of the Sn-doped fluorite phase Ce1−xSnxO2−δ. When Sn4<sup>+</sup> substitutes cerium in the fluorite phase, the Ce3<sup>+</sup> in the position of Ce4<sup>+</sup> could be generated together with an oxygen vacancy according the electroneutrality requirement: CeO2→SnCe + V •• <sup>O</sup> <sup>+</sup> 2Ce'Ce, where V•• – is an oxygen vacancy with <sup>−</sup>2 charge, Ce'Ce–is a Ce3<sup>+</sup> in the Ce4<sup>+</sup> position.

Sn atoms were shown in a combined experimental and DFT study to easily reduce ceria nanoparticles, both bare and doped by Pt atoms acquiring the oxidation state +2 (Lykhach et al., 2017b). First, only Ce4<sup>+</sup> cations were partially reduced upon interactions of atomic Sn with Pt/ceria nanoparticles, but already a moderate increase of Sn content resulted in the reduction of cationic Pt2<sup>+</sup> species to essentially neutral atomic ones prone to form Pt clusters. Using the same computational approach and similar model Pt/ceria nanoparticles we computationally investigated in the present work stability of Pt dopants in a much more labile +4 state in ceria nanostructures in the presence of minimal amounts of Sn. **Figure 10** shows that the interaction

with a model particle Pt2+/Pt4+Ce39O<sup>80</sup> of just one Sn atom can already slightly favor partial reduction of Pt4<sup>+</sup> remote from Sn by 8 Å. In the presence of two doping Sn atoms (see **Figure 10**, bottom panels) we also succeed to optimize a local-minimum structure Sn2Pt2+/Pt4+Ce39O<sup>80</sup> with the inner Pt atom in the highest oxidation state +4. This structure is over 40 kJ/mol less stable than the corresponding structure Sn2Pt2+/Pt2+Ce39O80, where Pt4<sup>+</sup> is reduced to Pt2+. These calculated data substantiate and explain our experimental observations on the absence of the Pt4<sup>+</sup> XPS signals, when Sn component is added to the Pt-ceria samples.

# Catalytic Properties and TPR-CO Data

**Figure 11** presents temperature dependencies of the CO conversion over the prepared Pt/Ce and Pt/CeSn catalysts and **Table 4** shows temperatures for the 10% (T10) and 50% (T50) CO conversion.

The Pt/Ce and Pt/CeSn catalysts, calcined at 600◦C (**Figure 11A**), are characterized by unusual behavior of CO conversion at varying temperature. Up to 100 ◦C the observed CO conversion either quite weakly depends on temperature (Pt/CeSn) or even decreases with temperature increase (Pt/Ce). The conversion of CO at −10◦C is 63 and 26% for Pt/Ce and Pt/CeSn, respectively. The temperature of the 10% CO conversion is significantly below room temperature (**Table 4**).

The presence of several regions in the CO conversion curve with different temperature dependencies, i.e., with different observed activation energies, indicates the presence of active centers of different nature triggering different CO oxidation mechanisms.

TABLE 3 | Quantitative composition of the catalyst surface and the XPS binding energy values Eb of individual states of the elements in the Pt/Ce-600 and Pt/CeSn-600 catalysts.


The calcination of catalysts at temperatures above 700◦C leads to the disappearance of the low-temperature activity. The T<sup>10</sup> values increase by more than 150◦C. The 50% conversion of CO (T50) increases by 215 and 110◦C for Pt/Ce and Pt/CeSn, respectively after calcination of the catalysts at 800◦C. Catalysts calcination at 900◦C results in the increase of T<sup>50</sup> by 10◦C for Pt/Ce and by 20◦C for Pt/CeSn catalysts.

The TOF values at 150◦C are 0.005 and 0.020 for the Pt/Ce-700 and Pt/CeSn-700 samples, respectively (see **Table 5**). Furthermore, after calcination at 800 and 900◦C the Pt/CeSn samples show higher TOF values than the Pt/Ce samples. These results indicate that the promotion by tin leads to an increase in both thermal stability and catalyst activity in the medium temperature range.

**Table 5** also provides activation energy values calculated for CO oxidation reactions on Pt/Ce and Pt/CeSn catalysts calcined at 700, 800, and 900◦C. The finding that the activation energy values are quite close to each other indicates that the limiting stages of the CO oxidation catalytic process on different samples under scrutiny are the same and, therefore, the oxidation process is carried out at essentially the same active centers. The data given in **Table 5** was obtained at the temperature range of 100– 200◦C. It was not possible to get such quantitative data for the low-temperature range of 0–100◦C. Nevertheless, it can be assumed that the reaction mechanism is similar in the lowtemperature range and is triggered by active centers of the fluorite phase.

The catalysts after CO+O<sup>2</sup> reaction were analyzed by XPS. **Figure 8** (curves 2) shows Pt4f lines for the Pt/Ce-600, Pt/Ce-900, Pt/CeSn-600, and Pt/CeSn-900 catalysts after reaction. For the most active catalysts, calcined at 600◦C, XPS shows that the fraction of Pt2<sup>+</sup> ionic species in both Pt/Ce-600 and Pt/CeSn-600 catalysts does not decrease significantly during the catalytic tests. However, we observe some reduction of Pt after the tests: Pt4<sup>+</sup> species are reduced to Pt2<sup>+</sup> in the Pt/Ce-600 catalysts; a small fraction of Pt metal is observed in the Pt/CeSn-600 catalysts. This is an indication that some Pt cations are reduced in CO+O<sup>2</sup> mixture. Yet, most of Pt2<sup>+</sup> is preserved after the catalytic tests.

Furthermore, no changes in the Pt states were detected in the samples calcined at 800 and 900◦C. **Figures 9D–F** shows that Ce and Sn oxidation states of the catalysts are not changed too. This observation was expected, since the catalytic reaction CO+O<sup>2</sup> is carried out not in a stoichiometric mixture, but in an excess of oxygen (O2: CO ratio = 5: 1).

**Figure 12** presents results for the catalysts obtained by the method of temperature-programmed reaction with CO (TPR-CO). The study of our catalysts by the TPR-CO method in the absence of the gas-phase oxygen allowed us to characterize the reactivity of oxygen present in the catalysts to carbon monoxide. Spectra of both CO absorption and CO<sup>2</sup> emission during TPR-CO are presented for the Pt/Ce, Pt/CeSn and pristine CeO<sup>2</sup> samples calcined at 600◦C.

For the pristine CeO2, CO consumption and CO<sup>2</sup> release begin above 200◦C. Addition of Pt leads to a strong activation of oxygen present in ceria revealed by lowering initial temperatures of the CO consumption and the CO<sup>2</sup> release to just 20 and 45◦C, respectively. The promotion by tin causes a further decrease of the reactions begin temperature by 20◦C, i.e.,

TABLE 4 | Catalytic characteristics in CO oxidation of the samples calcined at different temperatures Tcalc.


TABLE 5 | TOF and activation energy Ea values for CO oxidation over the Pt/Ce and Pt/CeSn samples calcined at 700, 800, and 900◦C and featuring specific surface area SBET.


\*TOF(s−<sup>1</sup> ) = W/CPt, where W—reaction rate (molecules/cm<sup>2</sup> s), CPt–Pt concentration calculated from Pt content in the catalyst (XPS).

due to a reaction of CO with surface hydroxo-groups of the catalyst: CO(ads) <sup>+</sup> 2OH−(support) <sup>→</sup> CO2(gas) <sup>+</sup> <sup>H</sup>2(gas) + O2−(support).

Addition of Sn to the catalysts leads to a decrease in the area of the peak 1 due to a decrease in the CeO<sup>2</sup> content. The areas of the peaks 2 and 3 are increased due to increased amount of reactive oxygen species. **Table 6** shows that the amount of consumed CO and released CO<sup>2</sup> for the Pt/CeSn-600 catalyst is almost 3 times higher than that for the Pt/Ce-600 catalyst. This corroborates that the modification of our Pt/Ce catalyst by tin noticeably increases the amount of oxygen reactive at low temperatures in the Pt/CeSn catalyst. Notably, the catalysts calcined at temperatures above 600◦C lose their high low-temperature oxidative activity. However, the observation of higher TOF values for Pt/CeSn catalysts compared to the corresponding Pt/Ce catalysts prepared by calcination at the

to 0 and 25◦C, respectively. These data suggest that the addition of tin to the catalysts leads to weaker binding of the lattice oxygen in them. TPR-CO spectra of the catalysts under study reveal three peaks at the temperatures of 20– 200◦C (peak 1), 200–350◦C (peak 2) and above 350◦C (peak 3). The first two peaks appear to result from CO interactions with reactive oxygen activated during formation of the Pt—Ce catalyst due to appearance of ionic platinum in the (PtCe)O<sup>2</sup> mixed phase.

At temperature above 350◦C the profiles of CO consumption and CO<sup>2</sup> release match. According to Giordano et al. (2000), diffusion of oxygen from the support volume to the surface and interaction with gas-phase CO begins at these temperatures. H<sup>2</sup> release is observed in our experiments during such a process. It was suggested (Zhu et al., 2004) that H<sup>2</sup> and CO<sup>2</sup> release could be observed during the TPR-CO experiments

TABLE 6 | CO consumption and CO2 evolution during the TPR-CO.


\*CO/Pt and \*CO2/Pt are molar ratios of the CO consumed and CO<sup>2</sup> evolved during the TPR-CO to the Pt content in the catalyst, respectively.

same temperatures indicates increased thermal stability of the Sn-containing catalysts.

# DISCUSSION

To achieve effective low-temperature oxidation of CO over ceriabased catalysts particle size in them should not exceed 10 nm (Boronin et al., 2009; Kurnatowska et al., 2012). In such cases a SMSI effect between dispersed ceria and ionic forms of the active component is possible, which can result in the formation of a large number of oxygen vacancies and mobile oxygen species with increased diffusion and reactivity.

In this work we used plasma arc synthesis to obtain highly dispersed ceria particles and mixed phases with tin dioxide. TEM data show that the as-prepared condense PtCeC and PtCeSnC substances are nanocomposites, in which the Ce and Sn components are in the form of semi-oxidized particles with Ce2O3, SnO, and SnO<sup>2</sup> crystalline structures coated with amorphous carbon. However, according to the reducing conditions of the synthesis, the initial particles should be in metallic state. We assume that the semi-oxidized nanoparticles Ce2O<sup>3</sup> and SnO are formed, when samples are exposed to the atmosphere after opening the chamber, in which plasma arc synthesis was performed. Considering high reactivity of cerium and tin metal particles their partial surface oxidation occurs even despite the diffusion difficulties of oxygen molecules into the carbon matrix. As a result, formation of core-shell structures Ce2O3@C in the PtCeC composite and of coreshell-shell structures Sn@SnO@C in the PtCeSnC composite takes place.

Subsequent calcination of the nanocomposites in an oxygen atmosphere at T > 600 ◦C results in burning out of the amorphous carbon matrix and conversion of the components into oxidized forms. Our TEM and XRD data show that phases with a fluorite structure (CeO<sup>2</sup> and Sn-doped CeO2) and cassiterite (SnO2) are formed, with fluorite particles being much smaller (3–4 nm) than SnO<sup>2</sup> particles (9–12 nm). We note that so small CeO<sup>2</sup> particles were not obtained in our previous studies of samples prepared by co-precipitation and solutioncombustion methods (Slavinskaya et al., 2015; Vasilchenko et al., 2016). Our XRD data reveal formation of a mixed Ce1−xSnxO2−<sup>δ</sup> phase. Furthermore, XPS data show that the state of Pt in the Pt/CeSn-600 catalysts is characterized only by Pt2<sup>+</sup> ions, which significantly differs from the state of platinum in the Pt/Ce-600 catalyst (Pt2+/Pt4+) and also indicates the influence of tin ions on the surface state of the fluorite phase in Sn-modified catalysts. Our DFT calculations showed that the presence of Sn4<sup>+</sup> ions in ceria nanoparticles facilitates Pt4<sup>+</sup> reduction to Pt2+. In addition, very stable square planar structures PtO<sup>4</sup> were formed (Bruix et al., 2014; Figueroba et al., 2017) when platinum interacted with nanostructured CeO<sup>2</sup> surface. It can be speculated that the presence of embedded tin ions boosts formation of such stable structures. Along with the effect of tin on the state of platinum, the Ce3d spectra show a significant increase in the concentration of Ce3<sup>+</sup> ions. Namely, the concentration of Ce3<sup>+</sup> ions in the Pt/CeSn catalysts reaches 50 % at., while in the Pt/Ce catalysts the Ce3<sup>+</sup> concentration is 15–20% at. These data indicate that the number of oxygen vacancies in the Sn-modified catalysts is significantly higher.

Key factor for the implementation of efficient CO oxidation in PGM-ceria systems is the formation of mobile highly reactive oxygen species. TPR-CO experiments allowed us to establish the effect of mobile oxygen formation in the catalysts synthesized by PAS. For these catalysts, the starting temperature of the interaction with CO is 20–50◦C. At temperatures up to 200◦C a significant amount of oxygen is released due to interaction with CO. This low-temperature mobile oxygen is generated in both Pt/Ce and Pt/CeSn catalysts, but there is a difference. In fact, a significant part of mobile oxygen is released from the Pt/Ce-600 catalyst with a peak with Tmax = 100◦C, while from the Pt/CeSn-600 catalysts mobile oxygen is released in the whole temperature range from 20 to 450◦C, and its amount is more than 2.5 times higher than in the case of Pt/Ce-600 catalyst. Obviously, this effect is directly related to the modification of the CeO<sup>2</sup> fluorite phase by Sn4<sup>+</sup> ions.

Our DFT modeling allows estimating effects of Sn4<sup>+</sup> doping ions in near-surface Ce4<sup>+</sup> positions of ceria nanoparticles on the reducibility of the latter (see **Figure 13**). In line with experimental observations (vide supra), the substitution of a Ce4<sup>+</sup> cation in the lattice of ceria by Sn4<sup>+</sup> greatly facilitates formation of oxygen vacancies O<sup>v</sup> in the immediate vicinity of the dopant. Indeed, the lowest vacancy formation energy EOv (calculated with respect to ½ O2) for removal of the first O atom located near Sn4<sup>+</sup> is almost 80 kJ/mol smaller than when the vacancy

is created near Ce4<sup>+</sup> cation occupying the same site. Calculated <sup>E</sup>Ov value for creating the second O<sup>v</sup> around the Sn4<sup>+</sup> cation decreases even somewhat more with respect to the Ce4<sup>+</sup> site. Notably, both the first and the second EOv values for the Sn4<sup>+</sup> dopant case are very low, 103 and 133 kJ/mol, respectively. These EOv values are close to those calculated using the same DFT scheme for Pt-ceria nanocomposites, where predicted high oxygen mobility was confirmed experimentally (Vayssilov et al., 2011). Thus, the present low EOv values corroborate very high oxygen mobility in nanostructured ceria induced by doping with Sn. Our calculations indicate that at least two mobile reactive O atoms can be easily generated by the presence of one Sn4<sup>+</sup> cation in ceria lattice, with concomitant creation of four reduced Ce3<sup>+</sup> ions, in full agreement with our experimental observations. The enhanced mobility of oxygen is related to strongly distorted oxygen coordination of Sn4<sup>+</sup> compared to Ce4<sup>+</sup> most favorably surrounded by cubes of eight O2<sup>−</sup> anions. The calculated eight nearest Ce-O and Sn-O distances (in Å) are: for Ce4<sup>+</sup> in Pt2+/Ce40O80−2.28×2, 2.31×2, 2.33, 2.34, 2.35, 2.36 and for Sn4<sup>+</sup> in Pt2+/SnCe39O80−2.11, 2.12, 2.14, 2.15, 2.22, 2.27, 2.35, 2.54 (see sketches of these models in the upper panels of **Figure 13** and atomic coordinates in the **Supplementary Material**).

Although the amount of mobile oxygen is generally higher for the Pt/CeSn-600 catalyst, this is probably not the only important factor for low-temperature oxidation reactivity. Our CO conversion data show that the catalytic activity in the range of 0–100◦C is slightly lower for the Pt/CeSn-600 catalysts than for the Pt/Ce-600 ones. Most likely, the amount of mobile oxygen affects the catalytic activity above 100◦C, when the TOF values are higher for the Pt/CeSn-800 and Pt/CeSn-900 catalysts. The six-fold difference in TOF values for the Pt/Ce and Pt/CeSn catalysts calcined at 900◦C reinforces notably higher activity of the latter at temperatures above 100◦C. According to the experimental data, Pt2<sup>+</sup> and Pt4<sup>+</sup> ions are present in catalytically active centers of the Pt/Ce-600 samples. In the Pt/CeSn-600 samples, doping Sn species can replace Pt4<sup>+</sup> ions and also induce additional reduction of Ce4<sup>+</sup> ions.

Application of plasma arc synthesis makes it possible to mix catalysts components at the cluster and nanoscale level. Due to the mixing of components takes place in the carbon matrix, which suppresses sintering at the sputtering stage, formation of highly dispersed oxide structures during annealing in oxygen atmosphere is possible. In the processes of annealing and burning of carbon, nanoparticles of CeO<sup>2</sup> and SnO<sup>2</sup> interact in the contact area.

# CONCLUSIONS

We prepared complex PtCeC and PtCeSnC composites in the form of highly defective oxide particles stabilized in carbon matrices using a plasma arc synthesis method. High-temperature annealing of the prepared samples in oxygen removes the carbon matrix and results in the formation of Pt/CeO<sup>x</sup> and Pt/CeSnO<sup>x</sup> samples that actively catalyze CO oxidation. In the presence of Sn XRD analysis shows formation of a mixed phase CeSnOx and stabilization of more dispersed species with a fluoritetype structure, the factors significantly contributing to the high activity and thermal stability of the catalyst modified by Sn.

XPS analysis indicates the presence of both Pt2<sup>+</sup> and Pt4<sup>+</sup> cationic states in the catalyst Pt/CeOx, whereas only the state Pt2<sup>+</sup> could be detected in the catalyst Pt/CeSnOx. Insertion of Sn ions in the fluorite lattice of Pt/CeO<sup>x</sup> does not stabilize Pt4<sup>+</sup> state in the Pt/CeSnO<sup>x</sup> catalyst, but induces formation of remarkably high concentration (50%) of lattice Ce3<sup>+</sup> ions. Results of our DFT calculations corroborate destabilization of Pt4<sup>+</sup> ions by incorporation of Sn cations into the fluorite structures of Pt/CeO<sup>x</sup> nanoparticles.

Our TPR-CO experiments show the presence of loosely bound active oxygen in the Pt/CeO<sup>x</sup> and Pt/CeSnO<sup>x</sup> catalysts in contrast with the similarly prepared pristine CeO<sup>x</sup> catalyst. CO<sup>2</sup> evolution in the gas phase upon interaction of the Pt/CeO<sup>x</sup> and Pt/CeSnO<sup>x</sup> catalysts with CO takes place in the temperature range 0–200◦C, while no reaction of the CeO<sup>x</sup> catalyst with CO could be detected at these temperatures. The TPR-CO results that the ignition temperature of CO oxidation over the Pt/CeSnO<sup>x</sup> catalyst is lower than that over the Sn-free Pt/CeO<sup>x</sup> catalyst strongly suggest the presence of more reactive oxygen species in the Sn-modified catalyst. Moreover, incorporation of Sn causes increased amounts of the loosely bound highly reactive oxygen (enhanced oxygen storage capacity). According to our DFT calculations, doping Sn cations are somewhat more energetically stable inside ceria nanoparticles than on their surface. The calculations also provide an atomic-level explanation for the observed low-temperature generation of surprisingly abundant reactive oxygen and reduced Ce3<sup>+</sup> species in the presence of Sn in ceria, which is related to strongly distorted Sn-O coordination sphere compared to that of Ce-O.

Along with the effect of Sn incorporation in the fluorite lattice of CeO<sup>2</sup> we also document interactions of tin and cerium oxides at the nanoscale, which stabilize highly dispersed cerium oxide species crucial for the observed enhanced stability of the modified by tin Pt/CeSnO<sup>x</sup> catalysts.

# DATA AVAILABILITY

The data that support the findings of this study are included in this published article and the **Supplementary Material.** Extra data are available from the corresponding authors upon reasonable request.

# AUTHOR CONTRIBUTIONS

TK, ED, ES, AS, VM, AZ, and SN performed experiments. KN performed DFT modeling, analyzed the calculated data, and was involved in the preparation of the manuscript. TK, ED, ES, and AS were involved in the analysis of the experimental data and the preparation of the manuscript. AB supervised the experimental work and was involved in the analysis of the experimental data and the preparation of the manuscript.

# FUNDING

This work was conducted within the framework of the budget project #AAAA-A17-117041710084-2 for Boreskov Institute of Catalysis. This work was also supported by the Russian Foundation for Basic Research (grants no. 17-03-00754 and 17-43-540738). KN gratefully acknowledges a support by the Spanish grants PRX17/00348, CTQ2015-64618-R, and MDM-2017-0767 as well as by the 2017SGR13 grant of the Generalitat de Catalunya.

# REFERENCES


# ACKNOWLEDGMENTS

Computer resources and technical support for the DFT modeling have been partly provided by the Red Española de Supercomputación (Barcelona, Spain).

# SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fchem. 2019.00114/full#supplementary-material


catalysts synthesized using the coprecipitation method. Appl. Catal. B 166–167, 91–103. doi: 10.1016/j.apcatb.2014.11.015


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Kardash, Derevyannikova, Slavinskaya, Stadnichenko, Maltsev, Zaikovskii, Novopashin, Boronin and Neyman. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Activation of Water on MnOx-Nanocluster-Modified Rutile (110) and Anatase (101) TiO<sup>2</sup> and the Role of Cation Reduction

Stephen Rhatigan and Michael Nolan\*

*Tyndall National Institute, University College Cork, Cork, Ireland*

Surface modification of titania surfaces with dispersed metal oxide nanoclusters has the potential to enhance photocatalytic activity. These modifications can induce visible light absorption and suppress charge carrier recombination which are vital in improving the efficiency. We have studied heterostructures of Mn4O<sup>6</sup> nanoclusters modifying the TiO<sup>2</sup> rutile (110) and anatase (101) surfaces using density functional theory (DFT) corrected for on-site Coulomb interactions (DFT + U). Such studies typically focus on the pristine surface, free of the point defects and surface hydroxyls present in real surfaces. In our study we have considered partial hydroxylation of the rutile and anatase surfaces and the role of cation reduction, via oxygen vacancy formation, and how this impacts on a variety of properties governing the photocatalytic performance such as nanocluster adsorption, light absorption, charge separation, and reducibility. Our results indicate that the modifiers adsorb strongly at the surface and that modification extends light absorption into the visible range. MnOx-modified titania can show an off-stoichiometric ground state, through oxygen vacancy formation and cation reduction spontaneously, and both modified rutile and anatase are highly reducible with moderate energy costs. Manganese ions are therefore present in a mixture of oxidation states. Photoexcited electrons and holes localize at cluster metal and oxygen sites, respectively. The interaction of water at the modified surfaces depends on the stoichiometry and spontaneous dissociation to surface bound hydroxyls is favored in the presence of oxygen vacancies and reduced metal cations. Comparisons with bare TiO<sup>2</sup> and other TiO2-based photocatalyst materials are presented throughout.

Keywords: TiO2, water activation, reduction, oxygen vacancies, photocatalysis

# INTRODUCTION

Photocatalysts are semiconductor materials which absorb photons of energies in excess of the bandgap to produce electron-hole pairs. These charge carriers separate and migrate to the surface of the catalyst where they drive chemical reactions via reduction and oxidation of adsorbed species. Photocatalysis has a variety of applications, including, but not limited to, the solar production of hydrogen from water splitting (Ni et al., 2007; Fujishima et al., 2008; Maeda and Domen, 2010; Jiang et al., 2017).

Edited by:

*Javier Carrasco, CIC energigune, Spain*

### Reviewed by:

*Jose J. Plata, Universidad de Sevilla, Spain Li-Min Liu, Beihang University, China*

> \*Correspondence: *Michael Nolan michael.nolan@tyndall.ie*

### Specialty section:

*This article was submitted to Physical Chemistry and Chemical Physics, a section of the journal Frontiers in Chemistry*

> Received: *12 October 2018* Accepted: *24 January 2019* Published: *12 February 2019*

### Citation:

*Rhatigan S and Nolan M (2019) Activation of Water on MnO*x*-Nanocluster-Modified Rutile (110) and Anatase (101) TiO*2 *and the Role of Cation Reduction. Front. Chem. 7:67. doi: 10.3389/fchem.2019.00067*

A practical photocatalyst must meet a number of criteria, such as visible light absorption, efficient charge carrier separation, stability, and active surface sites for the adsorption of feedstock species. Reducible metal cations in the catalyst can be important for enhancing the activity of the catalyst toward the difficult step of water dissociation. The development of metal oxide photocatalysts is of interest as these materials are cheap, earth abundant and, in many instances, non-toxic. Indeed, the most widely studied photocatalyst is titanium dioxide (TiO2) (Ni et al., 2007; Fujishima et al., 2008; Dimitrijevic et al., 2011; Pelaez et al., 2012; Habisreutinger et al., 2013; Tada et al., 2014; Etacheri et al., 2015) which was first demonstrated as a photoanode for water splitting by Fujishima and Honda (1972). The large bandgap (>3 eV) means that photoactivity is restricted to the UV and has limited the real-world application of TiO2-based photocatalyst technologies. As a result, significant scientific effort has focused on extending the light absorption edge of TiO<sup>2</sup> to longer wavelengths.

Substitutional doping of TiO<sup>2</sup> with cations and/or anions is a widely studied approach to inducing visible light absorption through the emergence of impurity-derived energy levels in the TiO<sup>2</sup> bandgap (Di Valentin et al., 2007; Czoska et al., 2008; Haowei et al., 2008; Ikeda et al., 2008; Gai et al., 2009; Nie et al., 2009; Valentin et al., 2009; Xu et al., 2009; Yang et al., 2009; Yu et al., 2009; Zhu et al., 2009; Long and English, 2010a,b; Zhang et al., 2010; Zheng et al., 2010; Herrmann, 2012; Etacheri et al., 2015; Li, 2015; Na Phattalung et al., 2017). First principles studies of such doped systems typically focus on bandgap reduction (Cui et al., 2008; Yang et al., 2009; Zhu et al., 2009; Long and English, 2010a,b; Zhang et al., 2010, 2014; Chand et al., 2011; Guo and Du, 2012; Na Phattalung et al., 2017) and questions of charge localization and surface reactivity are often overlooked. These are important considerations as dopant-derived defect states have been shown to act as recombination centers (Herrmann, 2012; Etacheri et al., 2015; Li, 2015) and photocatalysis is generally a surface mediated phenomenon.

Studies of the chemistry and electronic properties of surfaces and interfaces are key to understanding and screening materials for photocatalysis. The enhanced performance of the benchmark material, P25, which consists of chemically interfaced rutile and anatase phases, has been attributed to the favorable alignment of the conduction and valence bands at the interface which facilitates charge transfer between phases and the suppression of charge carrier recombination (Scanlon et al., 2013). In addition, the interface can promote the formation of active catalytic sites. This effect can be tuned by considering heterostructures with metal oxides of different compositions. Such heterostructures have been realized experimentally and shown to exhibit enhanced photocatalytic activity (Boppana and Lobo, 2011; Boppana et al., 2013; Chae et al., 2017; Sotelo-Vazquez et al., 2017; Wang et al., 2017). Nanostructuring of metal oxides has been investigated as an approach to enhancing charge transfer kinetics and increasing surface area while providing low-coordinated metal and oxygen sites for the adsorption of feedstock species (Gordon et al., 2012; Bhatia and Verma, 2017; Zhang et al., 2017a; Ong et al., 2018). Further, nanostructuring can also facilitate the reduction of metal cations.

Surface modification of metal oxide surfaces with dispersed metal oxide nanoclusters combines the properties of heteroand nano-structuring. Sub-nm nanoclusters of iron oxide were deposited on TiO<sup>2</sup> surfaces via chemisorption-calcination cycle (CCC) (Jin et al., 2011; Tada et al., 2014) and atomic layer deposition (ALD) (Libera et al., 2010). FeOx-modified TiO<sup>2</sup> exhibited bandgap reduction and enhanced visible light photocatalytic activity. The modification was shown to suppress carrier recombination as indicated by photoluminescence spectroscopy (Jin et al., 2011). The red-shift in light absorption was attributed to cluster-derived states above the valence band maximum (VBM) which were identified by X-ray photoelectron spectroscopy (XPS) and density functional theory (DFT) simulations (Jin et al., 2011; Nolan, 2011a; Tada et al., 2014).

These studies, and the subsequent development of similar systems, (Jin et al., 2012; Boppana et al., 2013; Iwaszuk et al., 2013; Bhachu et al., 2014; Fronzi et al., 2016a) mean that a multitude of nanocluster-surface composites can be investigated. Considerations for tuning these systems for optimal performance include composition, surface termination, nanocluster size, and stoichiometry; all of which contribute to the light absorption properties, charge carrier mobility and surface reactivity. DFT simulations can be used to illuminate the properties underpinning experimental observations (Nolan, 2011a; Jin et al., 2012; Nolan et al., 2012; Iwaszuk et al., 2013) and to screen candidate materials worthy of further investigation (Park et al., 2009; Graciani et al., 2010; Nolan, 2011b, 2012, 2018; Iwaszuk and Nolan, 2013; Lucid et al., 2014; Nolan et al., 2014; Fronzi et al., 2016b; Rhatigan and Nolan, 2018a,b).

In the present study we use first principles DFT calculations to examine the photocatalytic properties of manganese oxide modified TiO2, using model systems of Mn4O6-nanoclusters modifying the rutile (110) and anatase (101) surfaces and consider the role of partial surface hydroxylation in the interfacial chemistry. Our analysis includes an assessment of the stability of the composite surfaces, their ground state stoichiometry and reducibility via oxygen vacancy formation. Point defects, such as oxygen vacancies, are active sites at metal oxide surfaces and can be produced thermally. A more reducible surface will lose oxygen more readily and be more active in solar thermal (Muhich et al., 2016) or Mars and van Krevelen processes 1954, (Ganduglia-Pirovano et al., 2007). Computed density of states plots elucidate the impact of modification on the light absorption properties and a model for the photoexcited state (Di Valentin and Selloni, 2011) is used to examine charge separation and localization. Finally, we study the interaction of water with the modified surfaces and focus particularly on the role of oxygen vacancies and reduced cations on water adsorption. We identify the characteristics of activation, such as dissociation, geometry distortions and charge transfer to the adsorbed species. The importance of oxygen vacancies as active sites for water dissociation at the rutile (110) surface (Schaub et al., 2001; Henderson et al., 2003) and ceria surfaces (Mullins et al., 2012) has been widely discussed and reduced Ti3<sup>+</sup> ions have been shown to be active in the chemistry at titania surfaces (Lira et al., 2011; Xiong et al., 2012). For anatase TiO2, oxygen vacancies have been shown to be more stable at subsurface and bulk sites than on the surface (He et al., 2009; Scheiber et al., 2012). However the surface can be reduced by electron bombardment (Scheiber et al., 2012; Setvin et al., 2016) and the reaction of these vacancy sites with water and O<sup>2</sup> results in water dissociation. These studies highlight the necessity of engineering photocatalytic surfaces for which vacancies can be produced with moderate energy costs.

MnO<sup>x</sup> is an interesting modifier as manganese is a multivalent, reducible element which crystallizes in oxides with a variety of oxidation states; (Franchini et al., 2007) this will have implications for the light absorption properties and reducibility of sub-nm nanoclusters of MnO<sup>x</sup> dispersed at the titania surfaces. We have previously studied similar systems of MnOx-modified TiO2, in collaboration with experiment, to interrogate their activity for CO<sup>2</sup> capture and reduction (Schwartzenberg et al., 2017). In the present study, we focus on the potential for these catalysts to be active toward water activation. Furthermore, we investigate the impact of surface hydroxylation on the reduction of the heterostructures via oxygen vacancy formation and apply a model for photoexcitation to examine the associated energetics and charge localization. In reference Schwartzenberg et al. (2017), the Mn4O6-TiO<sup>2</sup> composites were found to be stoichiometric in the ground state for both modified rutile and anatase, albeit with moderate costs to produce reducing oxygen vacancies (+0.59 eV for rutile and +1.1 eV for anatase). However, the impact of surface hydroxyls on the formation of oxygen vacancies was not investigated; in this paper we show that vacancy formation is in fact promoted with hydroxyls already present at the TiO<sup>2</sup> surfaces. The photoexcited state model, which examines localization of electrons and holes at nanocluster metal and oxygen sites, sheds light on experimental observations which suggest that the MnOx-modifiers may facilitate recombination (Schwartzenberg et al., 2017). In addition, active oxygen vacancy sites play a crucial role in the subsequent interaction of water molecules and their adsorption modes. In particular, dissociation is favored for the reduced systems; this is an important step in the water oxidation reaction.

# METHODOLOGY

Periodic plane wave DFT calculations are performed using the VASP5.4 code (Kresse and Hafner, 1994; Furthmüller et al., 1996) with an energy cut-off of 400 eV. The core-valence interaction is described with projector augmented wave (PAW) potentials, (Blöchl, 1994; Kresse and Joubert, 1999) with 4 valence electrons for Ti, 6 for O, 13 for Mn and 1 for H species. The Perdew-Wang (PW91) approximation to the exchange-correlation functional is used (Perdew et al., 1996).

The TiO<sup>2</sup> rutile (110) and anatase (101) substrates are modeled as 18 and 12 atomic layer slabs, respectively. The bulk lattice parameters for rutile were computed as a = 4.639 Å and c = 2.974 Å, and the rutile (110) surface was modeled with a (2×4) surface expansion. For anatase the bulk lattice parameters are a = 3.814 Å and c = 9.581 Å and a (1 × 4) expansion was used for the anatase (101) surface. These parameters correspond to surface areas per supercell of 13.120 × 11.896 Å and 10.312 × 15.255 Å for rutile (110) and anatase (101), respectively. The surfaces are separated from their periodic images by vacuum gaps of 20 Å, as used in our previous studies (Fronzi and Nolan, 2017; Schwartzenberg et al., 2017; Nolan, 2018; Rhatigan and Nolan, 2018a,b). Γ -point sampling is used and the convergence criteria for the energy and forces are 10−<sup>4</sup> eV and 0.02 eVÅ−<sup>2</sup> , respectively. All calculations are spin polarized.

A Hubbard U correction is implemented to consistently describe the partially filled Mn 3d states and reduced Ti3<sup>+</sup> states (Anisimov et al., 1991; Dudarev et al., 1998). The values of U used are U(Ti) = 4.5 eV and U(Mn) = 4 eV and these have been chosen based on previous work on TiO<sup>2</sup> (Morgan and Watson, 2007; Nolan et al., 2008; Iwaszuk and Nolan, 2011; Nolan, 2011a; Fronzi et al., 2016a; Fronzi and Nolan, 2017; Rhatigan and Nolan, 2018a) and manganese oxides (Franchini et al., 2007; Kitchaev et al., 2016).

To model surface hydroxylation (before the nanoclusters are adsorbed) and the impact on the heterostructure chemistry, four water molecules are dissociatively adsorbed at the clean rutile (110) and anatase (101) surfaces which gives a partial coverage of 50%. The computed energy gain when the TiO<sup>2</sup> surfaces are hydroxylated at half coverage is −1.03 eV per water molecule for rutile (110) and −0.8 eV for anatase (101), referenced to the total energy of four gas phase water molecules. These indicate favorable water adsorption and surface hydroxylation and these models have been used in our previous studies (Fronzi et al., 2016a; Fronzi and Nolan, 2017; Schwartzenberg et al., 2017; Rhatigan and Nolan, 2018a). The nature of water molecules adsorbed at metal oxide surfaces, and in particular TiO<sup>2</sup> surfaces, is widely investigated both experimentally and computationally (Valdés et al., 2008; Fronzi and Nolan, 2017; Rhatigan and Nolan, 2018a) and readers are referred to reference (Mu et al., 2017) for a review of the state of the art. These models are representative of hydroxylated rutile and anatase surfaces, while we are not attempting to describe the most stable solutions for water or dissociative water adsorption at these titania surfaces (Fronzi et al., 2016a; Fronzi and Nolan, 2017; Rhatigan and Nolan, 2018a). The hydroxylated surfaces are denoted by OH-r110 and OH-a101. For the O2<sup>−</sup> ions of the pristine titania surfaces, computed Bader charges are in the range of 7.3–7.4 electrons and this is our reference. After hydroxylation, Bader charges for those oxygen atoms of the surface to which H atoms are adsorbed increase to values in the range 7.6–7.7 electrons, with similar values for oxygen ions of the water-derived hydroxyls.

The Mn4O<sup>6</sup> nanocluster (see **Supporting Information**) was adsorbed in different configurations at the hydroxylated rutile (110) and anatase (101) surfaces and the adsorption energies are computed using:

$$E\_{ads} = E\_{surf+A} - E\_{surf} - E\_A \tag{1}$$

where Esurf <sup>+</sup>A, Esurf and E<sup>A</sup> are the energies of the adsorbatesurface composite system, the hydroxylated titania surface and the gas phase nanocluster, respectively.

For the reduction of the composite surface, each of the six O sites of the supported nanocluster is considered for the formation of an oxygen vacancy, OV. One oxygen ion is removed from the Mn4O<sup>x</sup> cluster and the vacancy formation energy is calculated as:

$$E\_{\rm vac} = E\left(Mn\_4O\_{\rm x-1}\right) + 1/2E(O\_2) - E\left(Mn\_4O\_{\rm x}\right) \tag{2}$$

where the first and third terms on the right hand side are the total energy of the cluster-surface composite with and without an oxygen vacancy and the energy is referenced to half the total energy for molecular O2. Having identified the most stable structure with a single OV, the calculation is repeated for each of the five remaining O sites to determine the most stable structure with two OV. Oxidation states are investigated with Bader charge analysis and computed spin magnetizations.

We apply a model for photoexcitation to the ground state configuration of each modified surface and to the unmodified OH-r110 and OH-a101 surfaces for comparison. This model involves imposing a triplet electronic state on the system (Di Valentin and Selloni, 2011) to promote an electron to the conduction band, with a corresponding hole in the valence band. The analysis of the energies and charge localization is discussed in more detail in the **Supporting Information**.

For the interaction of water with the modified surfaces, H2O molecules are adsorbed in various configurations at the oxygen deficient systems and the adsorption energies are calculated as:

$$E\_{ads} = E\_{surf+H\_2O} - E\_{surf} - E\_{H\_2O} \tag{3}$$

where Esurf <sup>+</sup>H2O, Esurf and EH2<sup>O</sup> refer to the energies of the H2O molecule and modified surface in interaction, the modified surface, and the gas phase H2O, respectively.

Oxygen atoms of the surface, cluster and surface-bound hydroxyls are denoted OS, O<sup>C</sup> and OOH, respectively, and similar notation is adopted for OH groups. For the interaction of water with the modified surfaces, water-derived oxygen and hydroxyls are denoted O<sup>W</sup> and OHW.

# RESULTS

# Stoichiometric Mn4O6-Modified TiO<sup>2</sup> OH-Rutile (110) and OH-Anatase (101)

**Figures 1A,B** show the adsorption energies and relaxed atomic structures of the stoichiometric Mn4O6-nanocluster modifying the OH-r110 and OH-a101 surfaces. The large, negative adsorption energies indicate that the nanocluster-surface interaction is favorable and that the nanoclusters will be stable against desorption and aggregation (Fronzi et al., 2016b; Nolan et al., 2016; Fronzi and Nolan, 2017; Nolan, 2018; Rhatigan and Nolan, 2018a,b). For Mn4O6-OH-r110 (**Figure 1A**), three Mn ions are 4-fold coordinated and to each of these is bound a terminal OH. Of these OH groups, one has migrated from a Ti site in the rutile surface to an Mn ion of the cluster (OHOH) and two OH groups result from the migration of hydrogen from surface hydroxyls to O<sup>C</sup> atoms (OHC). The fourth Mn ion is 5-fold coordinated and is bound to three O<sup>C</sup> and two O<sup>S</sup> ions (one bridging O<sup>S</sup> and one in-plane OS). Five O ions of the OHr110 surface bind with Mn of the nanocluster (three O<sup>S</sup> and two OOH) and two O<sup>C</sup> ions bind to Ti of the surface. Mn-O distances are in the range 1.8–2.1 Å; the shorter distances involve 2-fold coordinated O ions and for Mn bound to the in-plane O<sup>S</sup> ion the Mn-O distance is 2.2 Å. Ti ions which bind to the nanocluster migrate out from the surface by 0.1 Å, however, distortions to the geometry of the rutile (110) surface are minimal.

For Mn4O6-OH-a101 (**Figure 1B**), three Mn ions are 4-fold coordinated and one Mn is 5-fold coordinated. Five O<sup>C</sup> sites are 2-fold coordinated with one O<sup>C</sup> ion binding to three Mn ions and a H atom which has migrated from a bridging O<sup>S</sup> site. Of the six interfacial bonds between the Mn4O<sup>6</sup> nanocluster and OHa101, three involve Mn and OHOH groups; two involve Mn and bridging O<sup>S</sup> sites and the sixth is a Ti-O<sup>C</sup> bond. Mn-O distances are in the range 1.7–2.1 Å.

For Mn4O<sup>6</sup> adsorbed at OH-r110 and OH-a101, the computed Bader charge for each of the Mn ions is 11.3 electrons, which are typical of Mn3<sup>+</sup> ions (see **Table 1**).(Schwartzenberg et al., 2017) The spin magnetizations for these sites are each 3.9 µB, which reflects the 3d<sup>4</sup> configuration of the Mn3<sup>+</sup> ion.

For the Mn4O<sup>6</sup> nanocluster adsorbed at OH-a101, there is an accumulation of positive charge at those O<sup>C</sup> sites which are doubly-coordinated to Mn ions of the nanocluster. Computed Bader charges of 7.0 electrons for these O<sup>C</sup> sites compare with 7.3–7.7 electrons computed for O2<sup>−</sup> anions of the OH-a101 surface. The nanocluster-surface interaction is not as strong at the OH-a101 surface as indicated by the smaller adsorption energy. The consequence of this is that the supported nanocluster retains characteristics of the gas phase, for which the O<sup>C</sup> ions have computed Bader charges in the range 7.0–7.1 electrons.

# Reduction of Mn4O6-Modified TiO<sup>2</sup> OH-Rutile (110) and OH-Anatase (101) Via Oxygen Vacancy Formation

The most stable modified surfaces with a single O<sup>V</sup> are shown in **Figure 1C** for Mn4O5-OH-r110 and **Figure 1D** for Mn4O5-OHa101. For the modified OH-r110 surface the formation energy of a single O<sup>V</sup> is −0.26 eV and this formation energy indicates that O<sup>V</sup> will form spontaneously. The next three most stable vacancy sites have formation energies in the range 0.60–0.82 eV. After formation of the most stable vacancy, two Mn ions are 3-fold coordinated and the third and fourth Mn cations are 4- and 5 fold coordinated. Two bridging and one in-plane surface oxygen are bound to Mn ions of the nanocluster. Two O<sup>C</sup> ions bind to surface Ti sites while three O<sup>C</sup> ions are bound only to Mn and H ions.

The formation of the neutral oxygen vacancy releases two electrons. Bader charge analysis reveals that the electrons localize at the 3-fold coordinated Mn sites of the nanocluster. The computed Bader charges on these sites increase from 11.3 to 11.5 electrons; see (**Table 1**) for computed Bader charges of reduced Ti and all Mn sites. The computed spin magnetizations are 4.6 µ<sup>B</sup> for these Mn sites; this is typical of the formation of reduced Mn2<sup>+</sup> ions which has an electronic configuration of 3d<sup>5</sup> .

The most favorable structure with one O<sup>V</sup> is more stable than the second most favorable by 0.9 eV. However, the relaxed atomic structures of these configurations are very similar (compare

shown in panels (C) for Mn4O5-OH-r110 and (D) for Mn4O5-OH-a101. The atomic structures of the most stable composites with two OV are shown in panels (E) for Mn4O4-OH-r110 and (F) for Mn4O4-OH-a101. The energy costs to produce OV are included and computed relative to the structure with one less OV. Atomic species and oxidation states are indicated by the colors in the legend on the right hand side.



*Also included are Bader charges for titanium ions of the support which are reduced after vacancy formation. Reduced Mn*2<sup>+</sup> *and Ti*3<sup>+</sup> *are highlighted in bold.*

**Figure 1C** with **Figure S3A**). The difference in energy arises from the distibution of excess charge. For the O<sup>V</sup> structure shown in **Figure S3A**, one excess charge localizes at a 5-fold coordinated surface Ti site for which the Bader charge increases from 1.3 to 1.7 electrons. A computed spin magnetization of 1.0 µ<sup>B</sup> reflects the 3d<sup>1</sup> configuration of reduced Ti3+.

For the modified OH-a101 surface, the most stable O<sup>V</sup> has a formation energy of −0.42 eV which indicates that it will spontaneously form, so that the ground state is offstoichiometric (vacancy formation energies for other sites of the nanocluster were in the range 0.5–1.3 eV). This compares with Mn4O<sup>6</sup> modifying bare anatase (101) which was found to be stoichiometric in the ground state (Schwartzenberg et al., 2017). After the formation of this OV, two Mn ions relax toward the titania surface and bind with bridging O<sup>S</sup> sites so that, in this configuration, each Mn ion is 4-fold coordinated. Bader charge analysis and computed spin densities indicate that two Mn ions are reduced to Mn2+, having computed Bader charges of 11.5 electrons and computed spin magnetisations of 4.6 µB. The next most stable structure with one O<sup>V</sup> is shown in **Figure S3B**; in this configuration three Mn ions are reduced to Mn2<sup>+</sup> and this is accompanied by an accumulation of postive charge on 2-fold coordinated O<sup>C</sup> ions for which the Bader charges were computed as 7.0 electrons.

The formation of the second O<sup>V</sup> has a moderate energy cost for both MnOx-modified TiO<sup>2</sup> surfaces, however the modified anatase surface is reducible at a lower energy cost. Given that the anatase surface is more easily hydroxylated, (Mu et al., 2017) which these results indicate promotes vacancy formation, one would expect more O<sup>V</sup> present on modified anatase. That O<sup>V</sup> formation is more facile for modified anatase corroborates previous experimental work on MnOx-TiO<sup>2</sup> (Schwartzenberg et al., 2017). The most stable configurations of the heterostructures with two O<sup>V</sup> are shown in **Figure 1E** for Mn4O4-OH-r110 and **Figure 1F** for Mn4O4-OH-a101. For the Mn4O4-OH-r110 surface, the two most stable O<sup>C</sup> sites for formation of a second O<sup>V</sup> had similar formation energies. One such configuration is described here and the other is included in the **Supporting Information**. For the structure shown in **Figure 1E**, the removed O<sup>C</sup> ion was 2-fold coordinated to a cluster Mn and surface Ti ion. After vacancy formation the Mn ion binds to a bridging O<sup>S</sup> ion and remains 3-fold coordinated. In this configuration three Mn ions are reduced; the Bader charges and spin magnetizations for these sites are 11.5 electrons and 4.6 µB, respectively. Similarly, for the Ti site to which the removed O<sup>C</sup> was bound, the Bader charge and spin magnetization are 1.7 electrons and 1.0 µB. Hence, the Mn4O4-OH-r110 heterostructure with two oxygen vacancies has one Ti3<sup>+</sup> and three Mn2<sup>+</sup> ions.

For the modified OH-a101 surface, a 3-fold coordinated O<sup>C</sup> site, which forms a hydroxyl group bridging two Mn ions, has the lowest cost to produce a second OV. One Mn ion that was bound to the removed O<sup>C</sup> atom is 2-fold coordinated, having been originally coordinated to three O<sup>C</sup> ions and one OOH ion. The second Mn ion is 3-fold coordinated, having been 4 fold coordinated prior to vacancy formation. The H ion which was bound to the removed O<sup>C</sup> migrates to another O<sup>C</sup> ion. In this Mn4O4-OH-a101 configuration, there are four Mn2<sup>+</sup> ions, with computed Bader charges of 11.5–11.6 electrons and spin magnetizations of 4.6 µB.

Additional structures with two O<sup>V</sup> are presented in **Figure S3C** for Mn4O4-OH-r110 and **Figure S3D** for Mn4O4- OH-a101; these are close in energy to the configurations described above, and differ in the distribution of excess charge over Mn and Ti sites. Hence, Mn and Ti sites should be present at the surface in a variety of oxidation states.

The localization of electrons at Ti and Mn sites is also accompanied by localized geometry distortions. The cation-O distances increase by ∼0.1 Å after reduction, reflecting the larger ionic radii of Mn2<sup>+</sup> and Ti3<sup>+</sup> compared to Mn3+and Ti4<sup>+</sup> (Shannon and Prewitt, 1969).

# Electronic Properties of Mn4Ox-Modified TiO<sup>2</sup> OH-Rutile (110) and OH-Anatase (101)

The projected electronic density of states (PEDOS) for the heterostructures are presented in **Figure 2**. Since the heterostructures are off-stoichiometric in the ground state, the PEDOS plot for Mn4O6-OH-r110 and Mn4O6-OH-a101 have been omitted from this figure and are included in the **Supporting Information** for completeness. The top panels of **Figure 2** show the PEDOS of modified OH-r110 for (A) the ground state with one O<sup>V</sup> and (B) the reduced state with two OV. The PEDOS plots show that occupied nanocluster-derived states (Mn 3d and O<sup>C</sup> 2p) extend to 0.3 and 0.8 eV above the VBM of the rutile support for Mn4O5- and Mn4O4-OH-r110, respectively. Unoccupied Mn 3d-derived states also emerge in the titania band gap at 0.1 and 0.3 eV below the conduction band minimum (CBM) for the ground state with one O<sup>V</sup> and the reduced state with two OV. Additional states emerge in the band gap due to occupied Ti3<sup>+</sup> states (see inset of **Figure 2B**), for the heterostructure with two OV.

The bottom panels of **Figure 2** display the PEDOS of the modified OH-a101 surface for (C) the ground state, with one OV, and (D) reduced state with two OV. The PEDOS plot for the ground state, with one OV, shows that occupied Mn 3d- and O<sup>C</sup> 2p-derived states extend to 1.3 eV above the titania derived VBM, while unoccupied Mn 3d states emerge at 1 eV below the CBM, leading to a significant reduction in the computed energy gap relative to TiO2. For the reduced structure, with two <sup>O</sup>V, each of the Mn ions is reduced to Mn2+, and the highest occupied of these states is 1 eV above the VBM. The lowest energy, unoccupied state is Mn-derived and is 1 eV below the CBM. For Mn4O4-OH-a101 the energy gap is 0.6 eV, with our DFT+U set-up showing a reduction over unmodified anatase.

These features of the PEDOS for Mn4Ox-TiO<sup>2</sup> can be attributed to formation of interfacial bonds, the presence of low-coordinated Mn and O<sup>C</sup> sites and the facile formation of O<sup>V</sup> in the supported metal oxide nanocluster. Modification pushes the VBM to higher energy and results in the emergence of empty states below the CBM; these effects, and the consequent red shift, are greater for modified anatase, consistent with previous reports (Schwartzenberg et al., 2017). These metal oxide nanocluster-modified surfaces are of interest for the oxygen evolution half reaction (OER) of the water splitting process and in this context raising the VBM from that of TiO<sup>2</sup> and toward the water oxidation potential is a desirable effect. Lowering of the titania CBM from its favorable position straddling the water reduction potential is detrimental to the hydrogen evolution reaction (HER) activity. However, as H adsorbs too strongly at metal oxide surfaces, such heterostructures will in any case not be suitable photocathodes for water splitting.

# Photoexcitation Model

We apply the model for the photoexcited state to the ground state systems, Mn4O5-OH-r110 and Mn4O5-OH-a101. **Table 2** presents the computed vertical, singlet-triplet and electron-hole trapping energies, as discussed in the **Supporting Information**. As can be seen from the values listed in **Table 2**, the underestimation of the bandgap inherent in approximate DFT is present in our DFT+U computational set-up. Our goal in choosing +U corrections is to consistently describe the localization of electrons and holes rather than reproduce the bandgap of bulk TiO2, which is not advised. Comparison of these computed energies across different structures nonetheless yields useful qualitative information about the effect of surface modification and results for the unmodified OH-r110 and OHa101 surfaces are included for reference. In particular, Evertical is analogous to the optical band gap, and a reduction in this value

computed valence band max is set to 0 eV and the Fermi energy is indicated with a dashed line. The top half of each panel displays Ti and Mn 3d derived states. The bottom halves of the panels show contributions to the DOS from oxygen 2p states of the surface (OC), surface bound hydroxyls (OOH) and nanocluster (OC). Inset in panel (B) shows the mid-gap occupied Ti 3d states in the range [0 eV, 1 eV].

TABLE 2 | Vertical singlet-triplet energy difference (Evertical), the relaxed singlet-triplet energy difference (Eexcite) and the relaxation energy (Erelax) for Mn4O5-OH-r110 and Mn4O5-OH-a101.


*Values for hydroxylated rutile (110) and anatase (101) surfaces have been included for reference.*

for a heterostructure relative to unmodified titania implies that modification leads to a red shift in light absorption.

When comparing Mn4O5-OH-r110 with unmodified OHr110 we can see that the values for Evertical are similar, however E excite is reduced by 0.93 eV for the modified surface. Comparing Mn4O5-OH-a101 with unmodified OH-a101, decreases in E vertical and Eexcite by 0.34 and 0.57 eV, respectively, indicate that modification leads to a significant red shift in light absorption. These results corroborate the analysis of the PEDOS. Erelax is the energy gained by the system after structural relaxation in response to the triplet electronic state and is related to the stability of the trapped electron and hole. The relaxation energy is larger for Mn4O5-OH-r110 than that computed for unmodified OH-r110 (1.31 vs. 0.46 eV) and reflects the greater flexibility of the modified system in accommodating the triplet electronic state. The relaxation energies for Mn4O5-OH-a101 and unmodified OH-a101 are comparable. The mixture of Mn oxidation states and the proximity of the Mn ions to each other at the anatase surface (neighboring Mn-Mn distances are in the range 2.9–3.2 Å for Mn4O5-OH-a101 and 3.0–3.9 Å for Mn4O5-OH-r110) restricts the degree to which the nanocluster can respond structurally to the localization of photoexcited charges.

Through analysis of Bader charges and spin magnetizations we can determine the electron and hole localization sites and the results of this analysis are represented graphically in **Figure 3**. For Mn4O5-OH-r110, in **Figures 3A,B**, the electron localizes at an Mn site; the Bader charge and spin magnetization for this site are 11.5 electrons and 4.6 µ<sup>B</sup> after electron localization, which are typical of Mn2<sup>+</sup> formation. The hole localizes at an O<sup>C</sup> site which is 2-fold coordinated to the Mn2<sup>+</sup> ion and a surface Ti. In this instance the Bader charge is 6.8 electrons and the spin magnetization is 0.8 µB, which are consistent with formation of O−. The Mn2+-O<sup>−</sup> distance increases by 0.2 Å, relative to the ground state. The Ti-O<sup>−</sup> distance decreases by 0.1 Å.

For Mn4O5-OH-a101 (**Figures 3C,D**), the photoexcited electron localizes at an Mn site of the nanocluster, as confirmed by a computed Bader charge of 11.5 electrons and spin magnetization of 4.5 µB. The hole state localizes predominantly at an O<sup>C</sup> site which bridges Mn2<sup>+</sup> and Mn3<sup>+</sup> ions. After hole localization the Bader charge for the O<sup>−</sup> ion is 6.7 electrons and the spin magnetization is 0.8 <sup>µ</sup>B. The Mn2+-O distances increase by 0.2–0.3 Å.

These results show that the electron localizes at an Mn site of the supported nanocluster and the hole state localizes at a neighboring O<sup>C</sup> site. Based on this model for the photoexcited state, we can conclude that modification does not necessarily promote the spatial separation of photoexcited charges. However, both electrons and holes will be available at the modified surface for transfer to adsorbed species.

# H2O Adsorption at Mn4Ox-Modified OH-Rutile (110) and OH-Anatase (101)

For the interaction of water at the modified surfaces, only those composites with O<sup>V</sup> present were considered, as such vacancies are known to be active sites at metal oxide surfaces (Schaub et al., 2001; Wang et al., 2016; Ruiz Puigdollers et al., 2017; Zhang et al., 2017b). Water adsorption is favorable at multiple sites of both modified surfaces and the geometries of the most stable adsorption configurations are displayed in **Figure 4**, while the **Supporting Information** shows other, less stable, water adsorption structures.

We adsorb water in molecular form at the heterostructures and relax the geometry. For Mn4O5-OH-r110, shown in **Figure 4A**, water is adsorbed exothermically in molecular form with a computed adsorption energy of −0.75 eV. In this instance, the water molecule binds to a 3-fold coordinated Mn2<sup>+</sup> site with a Mn-O<sup>W</sup> distance of 2.3 Å. **Figure S5A** shows dissociative water adsorption at the Mn4O5-OH-r110 surface, which has an adsorption energy of −0.31 eV. Upon dissociation, an H atom migrates to a bridging O<sup>S</sup> site and the water-derived hydroxyl (OHW) is singly coordinated to an Mn site with an Mn-O<sup>W</sup> distance of 1.9 Å. The dissociation is accompanied by a transfer of charge from O<sup>W</sup> to the nanocluster modifier, indicated by a decrease of 0.4 electrons in the computed Bader charge for the O<sup>W</sup> ion. The Bader charges and spin magnetizations of cation sites are unchanged by the adsorption and dissociation.

Water adsorbs molecularly at Mn4O5-OH-a101, as shown in **Figure 4B**, with an adsorption energy of −0.74 eV. The H2O binds to a 4-fold coordinated Mn3<sup>+</sup> ion with a Mn-O<sup>W</sup> distance of 2.2 Å. Since Mn4O5-OH-r110 and Mn4O5-OH-a101 are the ground states of the systems, the single O<sup>V</sup> having formed spontaneously, these composites favor non-stoichiometry so that the strength of interaction with the water molecule is not sufficient to promote spontaneous dissociation and adsorption in molecular form is favored.

The surfaces with two O<sup>V</sup> show higher reactivity to water, as indicated by the larger adsorption energies in **Figures 4C,D**. Water adsorbs and spontaneously dissociates at both Mn4O4- OH-r110 (**Figure 4C**) and Mn4O4-OH-a101 (**Figure 4D**). For Mn4O4-OH-r110, the water molecule adsorbs at an O<sup>V</sup> site. An H atoms migrates to a bridging O<sup>S</sup> site and the OH<sup>W</sup> group is doubly coordinated to an Mn and a surface Ti site. The Mn-O<sup>W</sup> and Ti-O<sup>W</sup> distances are 2.2 Å. Bader charge analysis reveals that 0.3 electrons are transferred from the O<sup>W</sup> to the surface. Despite this charge transfer, the Ti ion which binds to OH<sup>W</sup> and which was reduced to Ti3<sup>+</sup> due to O<sup>V</sup> formation prior to water adsorption, remains in the Ti3<sup>+</sup> state. This agrees with work by Henderson et al. in which no charge transfer was observed between Ti3<sup>+</sup> and bridging hydroxyls bound at oxygen vacancy sites at the TiO<sup>2</sup> rutile (110) surface (Henderson et al., 2003). The reduced Ti site was only oxidized after interaction of O<sup>2</sup> with the Ti3+-OH group. After water adsorption and dissociation the distribution of cation oxidation states is unchanged so that there are three Mn2<sup>+</sup> ions and one Ti3+. The Bader charge for the bridging O<sup>S</sup> site to which the H ion binds increases from 7.3 to 7.7 electrons which, as discussed in the methodology, is typical of hydroxyl formation.

For Mn4O4-OH-a101, the water molecule adsorbs at an O<sup>V</sup> site and after dissociation an H atom migrates to an O<sup>C</sup> ion which shows an increase in Bader charge, from 7.1 to 7.6 electrons. The OH<sup>W</sup> group binds to three Mn2<sup>+</sup> ions; the Bader charges and spin magnetizations for cation sites are unchanged so that these ions are not involved in the charge transfer. However, for the water adsorption configuration shown in **Figure S5D**, an Mn2<sup>+</sup> ion is re-oxidized to Mn3<sup>+</sup> after dissociation of the water molecule. In this instance the adsorption energy is −1.89 eV and the OH<sup>W</sup> group is singly-coordinated to the re-oxidized Mn ion.

# CONCLUSIONS

The properties of Mn4Ox-modified hydroxylated titania surfaces and their reduction and interaction with water depend on the phase of the TiO<sup>2</sup> substrate. For Mn4O<sup>6</sup> adsorbed at the hydroxylated anatase (101) surface, one interfacial bond is established between a cluster oxygen ion and the surface and Mn ions bind mostly to oxygen ions of the surface bound hydroxyls. Conversely, for Mn4O<sup>6</sup> at hydroxylated rutile (110), the nanocluster-surface interaction is more intimate, with Mn ions binding to bridging and in-plane oxygen ions of the rutile surface.

Our results indicate that both Mn4Ox-OH-r110 and Mn4Ox-OH-a101 favor non-stoichiometry, in contrast to unhydroxylated modified TiO<sup>2</sup> surfaces, as oxygen vacancies form spontaneously and both composites can be considered highly reducible with moderate energy costs for subsequent oxygen vacancy formation. Bader charge analysis shows that Mn ions are present in a mixture of oxidation states of at the hydroxylated surfaces. Both Mn and Ti ions are reduced in response to vacancy formation.

Modification with Mn4O<sup>x</sup> has a significant impact on the light absorption properties. Occupied Mn 3d states extend the

VBM of the composites to higher energies relative to that of the titania support and empty states emerge below the CBM. The consequent red shift in the light absorption edge is confirmed by our model for the photoexcited state. In particular, the vertical energy, analogous to the optical energy gap, decreases significantly for Mn4O5-OH-a101 relative to that computed for the unmodified, hydroxylated anatase (101) surface. Analysis of this model shows that electrons and holes localize at Mn and neighboring O<sup>C</sup> sites, respectively, so that modification may not promote separation of photoexcited charges, but the trapping energies of the electron and hole are quite high, suggesting high stability.

With regard to water adsorption and activation, the formation of oxygen vacancies has an impact on the strength of interaction and the most favorable adsorption mode of H2O at the modified surfaces. For Mn4O5-OH-a101, with a spontaneously formed OV, water adsorbs only in molecular form. With formation of reducing oxygen vacancies, water adsorption becomes more exothermic and leads to spontaneous dissociation to surface bound hydroxyls, similar to observations made for water interacting at reduced TiO<sup>2</sup> (Schaub et al., 2001; Henderson et al., 2003) and CeO<sup>2</sup> (Mullins et al., 2012) surfaces.

# AUTHOR CONTRIBUTIONS

MN devised the research. SR performed the DFT modeling. Both authors analyzed the results and prepared, reviewed and approved the text of the paper.

# REFERENCES


# ACKNOWLEDGMENTS

We acknowledge support from Science Foundation Ireland through the US-Ireland R and D Parntershop program, Grant number SFI/US/14/E2915 and the ERA.Net for Materials Research and Innovation (M-ERA.Net 2), Horizon 2020 grant agreement number 685451, SFI Grant Number SFI/16/M-ERA/3418 (RATOCAT). We acknowledge access to SFI funded computing resources at Tyndall Institute and the SFI/HEA funded Irish Centre for High End Computing. We are grateful for support from the COST Action CM1104 Reducible Metal Oxides, Structure and Function.

# SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fchem. 2019.00067/full#supplementary-material

and experiments. J. Photochem. Photobiol. C: Photochem. Rev. 25, 1–29. doi: 10.1016/j.jphotochemrev.2015.08.003


photocatalyst for the degradation of methyl orange and gaseous benzene. Appl. Catal. B. 200, 19–30. doi: 10.1016/j.apcatb.2016.06.070


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Rhatigan and Nolan. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Dynamics of the Interaction Between Ceria and Platinum During Redox Processes

Paola Luches <sup>1</sup> , Gabriele Gasperi 1,2, Marc Sauerbrey 3,4, Sergio Valeri 1,2, Jens Falta3,4 and Jan Ingo Flege3,4,5 \*

1 Istituto Nanoscienze, Consiglio Nazionale delle Ricerche, Modena, Italy, <sup>2</sup> Dipartimento di Scienze Fisiche, Informatiche e Matematiche, Università degli Studi di Modena e Reggio Emilia, Modena, Italy, <sup>3</sup> Institute of Solid State Physics, University of Bremen, Bremen, Germany, <sup>4</sup> MAPEX Center for Materials and Processes, University of Bremen, Bremen, Germany, <sup>5</sup> Applied Physics and Semiconductor Spectroscopy, Brandenburg University of Technology Cottbus-Senftenberg, Cottbus, Germany

The work is focused on understanding the dynamics of the processes which occur at the interface between ceria and platinum during redox processes, by investigating an inverse catalytic model system made of ceria epitaxial islands and ultrathin films supported on Pt(111). The evolution of the morphology, structure and electronic properties is analyzed in real-time during reduction and oxidation, using low-energy electron microscopy and spatially resolved low-energy electron diffraction. The reduction is induced using different methods, namely thermal treatments in ultra-high vacuum and in H<sup>2</sup> as well as deposition of Ce on the oxide surface, while re-oxidation is obtained by exposure to oxygen at elevated temperature. The use of two different epitaxial systems, continuous films and nanostructures, allows determining the influence of platinum proximity on the stabilization of the specific phases observed. The factors that limit the reversibility of the observed modifications with the different oxidation treatments are also discussed. The obtained results highlight important aspects of the cerium oxide/Pt interaction that are relevant for a complete understanding of the behavior of Pt/CeO<sup>2</sup> catalysts.

Keywords: cerium oxide, platinum, LEEM, LEED, reduction

# INTRODUCTION

The importance of cerium oxide in catalysis is based on its reducibility, namely on the possibility for cerium ions to reversibly switch between the 4+ and the 3+ oxidation state depending on the external conditions (Trovarelli, 1996). This property is closely linked to the material ability to reversibly store and release oxygen ions in redox reactions. In most catalytic applications cerium oxide is combined with metal nanoparticles, which complete the system functionality (Trovarelli, 1996; Cargnello et al., 2012; Trovarelli and Fornasiero, 2013). In particular, a large number of studies have demonstrated that the combination of platinum and cerium oxide allows exploiting the synergy between the two materials to obtain unprecedented activities with a minimal amount of Pt (Pierre et al., 2007; Bruix et al., 2012; Fiala et al., 2015; Jones et al., 2016). The interaction, which influences the catalytic activity, is rather complex and it involves, for example, the stabilization of ionic Pt species at specific defective ceria surface sites in catalysts with ultra-low Pt loading (Bruix et al., 2014) and a relevant charge transfer between Pt nanoparticles and cerium oxide (Lykhach et al., 2015). To identify the functionality of different atomic sites, to clarify some aspects of the interaction and to optimize the activity and/or the selectivity of catalysts, several studies made use of

### Edited by:

Maria Veronica Ganduglia-Pirovano, Instituto de Catálisis y Petroleoquímica (ICP), Spain

### Reviewed by:

Juan De La Figuera, Instituto de Química Física Rocasolano (IQFR), Spain Vladimir Matolin, Charles University, Czechia

\*Correspondence:

Jan Ingo Flege flege@b-tu.de

### Specialty section:

This article was submitted to Physical Chemistry and Chemical Physics, a section of the journal Frontiers in Chemistry

Received: 12 October 2018 Accepted: 21 January 2019 Published: 08 February 2019

### Citation:

Luches P, Gasperi G, Sauerbrey M, Valeri S, Falta J and Flege JI (2019) Dynamics of the Interaction Between Ceria and Platinum During Redox Processes. Front. Chem. 7:57. doi: 10.3389/fchem.2019.00057

**139**

inverse model systems composed of well-controlled epitaxial oxide nanostructures or ultrathin films supported on metal single crystals (Suchorski et al., 2008; Senanayake et al., 2013; Grinter et al., 2016). Such systems are well-suited in particular to focus on the interplay between cerium oxide and the metal, using surface science techniques with high sensitivity and resolution.

The different properties of cerium oxide epitaxial films and nanostructures have been recently reviewed (Luches and Valeri, 2015). In particular, when Pt(111) is used as a substrate the films grow with a good epitaxial quality (Berner and Schierbaum, 2002; Grinter et al., 2010; Luches et al., 2011, 2013), with surface and subsurface oxygen vacancies (Grinter et al., 2010), step edges and domain boundaries (Grinter et al., 2010; Luches et al., 2011) as the dominant defects. The thermal stability of the oxide films indeed depends on the thickness considered, and different surface reconstructions are stabilized by thermally induced reduction (Berner and Schierbaum, 2002; Luches et al., 2014; Grinter et al., 2016). An interfacial charge transfer has been identified (Luches et al., 2015) and proposed to determine the higher reducibility of films of sub-nanometer thickness as compared to the surface of thicker deposits (Gasperi et al., 2016).

To understand the mechanisms that take place in cerium oxide during redox reactions, several recent studies have specifically focused on understanding the reduced phases stabilized on different cerium oxide surfaces (Wilkens et al., 2013a,b; Duchon et al., 2014a,b; Luches et al., 2014; Höcker et al., 2015, 2017; Grinter et al., 2016; Olbrich et al., 2017; Murgida et al., 2018). A large number of different surface reconstructions have been observed, some of which represent the truncation of a metastable bulk phase, while some others, without a bulk counterpart, are stabilized at the surface or at reduced dimensionality. The reducibility of cerium oxide, at the basis of its catalytic activity, is therefore certainly influenced by a number of factors. For example, Kozlov et al. simulated different reduced ultrathin ceria structures, and they showed that their relative stability depends on the surface lattice parameter (Kozlov et al., 2015). Indeed, epitaxialstrain is only one of the parameters which may come into play, and also the electronic interaction with the substrate, the oxide surface morphology, the island size, and the temperature may play an important role (Dvorak et al., 2011; Stetsovych et al., 2013; Wilkens et al., 2013a; Luches and Valeri, 2015; Olbrich et al., 2017).

Low-energy electron microscopy (LEEM) has been shown to be able to provide useful information in this context (Flege and Grinter, 2018), allowing to access the dynamic evolution of the morphology during reduction and oxidation cycles with a lateral resolution below 10 nm and simultaneously providing structural information from electron micro-diffraction with a high spatial resolution (∼250 nm) (Höcker et al., 2015, 2017; Grinter et al., 2016). Moreover, the energy-dependent electron reflectivity can provide important fingerprints of the local surface geometric and electronic properties of the investigated systems and in particular of the surface oxidation state of ceria at about 10 nm spatial resolution (Flege et al., 2013).

We report here a spatially resolved dynamic study of the formation of reduced phases at different sites in ceria/Pt systems. We identify in particular the phases which specifically originate from the direct cerium oxide—platinum interaction and the ones which originate at sites which are further from the interface. The obtained results provide a basis for a better understanding of the mechanisms that can occur in ceria-platinum catalysts, which always comprise a high density of interfacial sites.

# MATERIALS AND METHODS

A Pt(111) single crystal was used as a support for the cerium oxide epitaxial nanostructures. The Pt surface was cleaned by multiple cycles of sputtering with Ar<sup>+</sup> ions and annealing at 900 K in an oxygen background pressure of 5 × 10−<sup>7</sup> Torr, followed by a final heating step at 1,200 K in ultra-high vacuum (UHV). Cerium was deposited by reactive molecular beam epitaxy in oxygen background pressure of 5 × 10−<sup>7</sup> Torr using a home-made electron beam evaporator. The nominal thickness of deposited cerium oxide is estimated by considering the cerium flux from the evaporator and the deposition time and it is expressed in monolayer equivalents (MLE), where 1 MLE corresponds to the thickness of an O-Ce-O trilayer, 3.12 Å, fully covering the Pt surface (Sauerbrey et al., 2017). The flux of the evaporator was calibrated using our results from previous works (Luches et al., 2011), in which the thickness of a closed ceria layer on Pt(111) was estimated using a quartz microbalance coupled with scanning tunneling microscopy results. The reduction was performed using three methods, differing for the reducing agent and for the reduction temperature used: thermal treatments at 1,040 K in UHV, thermal treatments at 680 K in H<sup>2</sup> background pressure of 5 × 10−<sup>7</sup> Torr, and deposition of cerium at 680 K in UHV on the surface of the film. The thickness of metallic cerium is also expressed in MLE, where 1 MLE CeO<sup>2</sup> ∼1 MLE Ce since the atomic densities of cerium in its metal structure and in the oxide are almost the same. Reduction by cerium deposition in UHV was applied only to closed cerium oxide films, to avoid direct interactions between Ce and Pt, which are known to give origin to ordered alloyed phases (Baddeley et al., 1997; Berner and Schierbaum, 2002; Kemmer et al., 2014). The re-oxidation is performed by thermal treatments at 1,040 K in O<sup>2</sup> background pressure of 1 × 10−<sup>7</sup> Torr. Similar thermal treatments were shown to be effective in reducing and oxidizing cerium oxide low dimensional structures in previous studies (Duchon et al., 2014a,b; Luches et al., 2014; Höcker et al., 2015, 2017).

The morphological and structural evolution was investigated in-situ using a commercial Elmitec LEEM III system, equipped with multiple apertures in order to perform micro-illumination low-energy electron diffraction (µLEED) from selected areas down to ∼250 nm in diameter. LEEM images and µLEED patterns were acquired in real-time during reduction and oxidation cycles. I-V LEEM curves were also acquired on selected surface areas. The µLEED and LEEM data were processed with the Gxsm software package (Zahl et al., 2010).

# RESULTS

The dynamic modification of the structure and morphology of cerium oxide overlayers with reduction and oxidation cycles was

at 1,040 K. (D) I(V)-LEEM spectra of the 5 MLE film after growth at 720 K(black line) and after different reducing treatments: annealing at 690 K in UHV (yellow line), annealing at 690 K in H2 (green line), annealing at 1,040 K in H2 (purple line) and Ce deposition in UHV at 680 K on the surface (blue line).

investigated by LEEM and µLEED. I-V curves provide local information on the surface composition. Although the obtained information is only qualitative, the comparison of the curves with the ones obtained on cerium oxide samples with known stoichiometry provides relevant information on the investigated processes (Flege et al., 2013; Höcker et al., 2015). The present study focusses on two different systems: cerium oxide films completely covering the Pt substrate and well-separated cerium oxide islands with part of the substrate surface left uncovered. The continuous cerium oxide films provide essential information to interpret the more complex results obtained on cerium oxide islands. On the other hand, the islands allow for a direct observation of the interactions between the oxide and the metal at the edge sites (see, e.g., Schaefer et al., 2018).

# Continuous Cerium Oxide Films

Cerium oxide films fully covering the Pt(111) surface were obtained by depositing nominal amounts of cerium oxide above 5 MLE on the clean substrate kept at 720 K, as in Sauerbrey et al. (2017). **Figure 1** shows the LEEM images of the clean Pt substrate (a) and of the film surface after the growth (b). The film morphology does not show any relevant feature within the resolution of the technique. The I-V curve acquired on the film, shown in **Figure 1D** (black line), has the typical shape observed for stoichiometric cerium oxide films, with a well-resolved feature at 4.5 eV (T1), three partially overlapping features at 9, 12, and 15 eV (T2–T4) and a minor feature at 22.5 eV (T5) (Flege et al., 2013; Höcker et al., 2015). Thermal treatments at 690 K in UHV or in hydrogen background pressure for 50 min did not lead to relevant modifications of the I-V curves (yellow and green lines in **Figure 1D**). On the contrary, a thermal treatment at 1,040 K for 185 min in hydrogen (purple line in **Figure 1D**) leads to significant modifications of the I-V curve, namely to a change in the energy position and relative weight of the features, which are compatible with the formation of a reduced cerium oxide phase (Flege et al., 2013; Höcker et al., 2015). A stronger reducing treatment was performed by depositing approximately 0.5 MLE of Ce in UHV on the surface at 680 K. In this case the I-V curve (blue line in **Figure 1D**) showed a rather marked peak at approximately 8 eV (D1), a contribution at 4.5 eV (T1) and two minor features at 11 and 19 eV (D2, D3), which are well compatible with the formation of a fully reduced Ce2O<sup>3</sup>

bixbyite phase (Flege et al., 2013; Höcker et al., 2015). The LEEM image acquired after reduction in hydrogen background at 1,040 K (**Figure 1C**) does not show pronounced morphological modifications compared to the image acquired before oxidation (**Figure 1B**).

The evolution of the structure with reducing treatments was followed by LEED. **Figure 2** shows selected patterns acquired during Ce deposition on the film surface, while the full LEED movie is included as **Supplementary Video 1**. Before the treatment the LEED shows only the cerium oxide related spots and no substrate related spot (**Figure 2A**), which indicates that the substrate is fully covered by an epitaxial film with a good crystalline quality. After the deposition of 0.3 MLE of Ce a clear (3 × 3) superstructure appears (**Figure 2B**). This surface periodicity in reduced cerium oxide films was already observed to be formed after Ce deposition on the surface of a 3 nm thick cerium oxide films on Cu(111) followed by annealing at 870 K by Duchon et al. and it was ascribed to the CeO1.67 phase with a specific surface structure (Duchon et al., 2014b). Subsequently, the same phase was observed to appear also on ceria films grown on Ru(0001) using thermal reducing treatments (Duchon et al., 2014a; Höcker et al., 2015). By depositing larger amounts of metallic cerium the superstructure-related spots gradually move toward the cerium oxide related spots, reaching a periodicity of approximately (3.5×3.5) after the deposition of 0.5 MLE of Ce. The LEED intensity profiles along the [101] direction, reported in **Figure 2D**, show the shift of the (3×3)-related spots toward the (11) and (10) ceria spots. A previous work by Höcker et al. (2015) demonstrated that if two coexisting phases with slightly different surface periodicities (e.g., 3 × 3 and 4 × 4) gradually transform from one into the other in a non-correlated way, the lattice parameter of the resulting periodicity in LEED patterns progressively changes between the two values determined by the individual phases (Höcker et al., 2015). It seems therefore reasonable to assume that the (3.5 × 3.5) periodicity originates from an incomplete transformation of the (3 × 3) phase into a phase with a (4 × 4) periodicity. The latter corresponds to the bixbyite Ce2O<sup>3</sup> phase, which was observed to be stabilized by reduction of ceria films on Cu(111) (Duchon et al., 2014b) and on Ru(0001) (Duchon et al., 2014a; Höcker et al., 2015).

In order to investigate the effect of interface proximity on the reduced phases that are stabilized, reduction was induced by cerium deposition on the surface of a much thinner film of 1.5 MLE. **Figure 3** reports a time-lapse LEED pattern sequence during the deposition of selected amounts of cerium, while the full LEED movie is included as **Supplementary Video 2**. Before the treatment (**Figure 3A**) the LEED pattern shows both the spots related to cerium oxide and those related to the Pt(111) surface plus some extra spots (brown circles in **Figure 3A**), probably originating from a non-complete oxidation of the cerium oxide film. After the deposition of ∼0.05 MLE of Ce a well-ordered (3 × 3) phase appears (**Figure 3B**), in analogy with the phase obtained on thicker continuous films after the deposition of 0.3 MLE of Ce (**Figure 2B**). After the deposition of 0.09 MLE of cerium (**Figure 3C**) the (3 × 3) phase disappears and the surface shows a pattern similar to the one observed before reduction (**Figure 3A**), which might be due to random creation of further oxygen vacancies. With further reduction a phase with a (7 × 7) periodicity with respect to the Pt surface becomes clearly visible, which corresponds to an approximate (5 × 5) periodicity with respect to cerium oxide (**Figure 3D**). As the reduction proceeds (**Figures 3E,F**) some of the superstructure spots gradually move toward the main cerium oxide spots finally forming a 5/2(<sup>√</sup> 3 × √ 3)R30◦ reconstruction after the deposition of ∼0.5 MLE of Ce. Also in this case the gradual movement of some of the LEED spots is interpreted as due to two kinds of spatially uncorrelated domains with the (5 <sup>×</sup> 5) and the 5/2(<sup>√</sup> 3 × √ 3)R30◦ reconstruction which gradually transform one into the other, as in Höcker et al. (2015).

# Cerium Oxide Islands

To have more detailed information on the interaction between cerium oxide and Pt, we investigated the processes occurring with thermal reduction on a sample obtained by growing 1.5 MLE cerium oxide on the Pt surface at 1,040 K to allow for the formation of well-separated cerium oxide islands, using the procedures identified in one of our previous studies (Sauerbrey et al., 2017). **Figure 4A** reports a typical LEEM image of the sample surface. The bright features can be ascribed to cerium oxide, which has a higher reflectivity than the bare Pt surface at 12 eV (see **Figure 1D** and Sauerbrey et al., 2017). The lateral size of the ceria features varies between ∼50 and 200 nm and their shape appears mainly irregular, being partially constrained by the substrate step edges. The islands that nucleate on flat areas of the Pt substrate, far from step edges, have a more regular triangular shape as they have grown from a single or only very few nuclei

(highlighted by white arrows in **Figure 4**). The average height of the islands, estimated using the Pt surface coverage (25 ± 1%) and the cerium oxide equivalent thickness, amounts to approximately 6 ML.

2,495 s of deposition (0.50 ML Ce).

The evolution of the morphology of the sample during thermal treatments at 1,040 K in UHV was followed by LEEM in real-time (see also **Supplementary Video 3**). **Figures 4B–D** represent selected LEEM images during reduction. The image acquired after 440 s (**Figure 4B**) evidences significant morphological modifications of the sample. Most of the islands are surrounded by an area characterized by a very low reflectivity, which appears darker than the substrate. The low-reflectivity phase partially extends also along the step edges, partially decorating them with an almost continuous profile. The extent of the bright part of the islands decreases with reduction, as shown for example by comparing the areas highlighted in green in **Figures 4A,B**. Simultaneously, some of the bright islands increase in size and acquire a more regular triangular shape (blue arrows). With longer times, the low-reflectivity phase surrounding the ceria structures further extends on the substrate surface, while the bright part of the islands decreases in size (areas highlighted in red and yellow in **Figures 4C,D**). The morphology and contrast of the larger triangular islands (white arrows) are almost unchanged, except for a mild decrease in reflectivity. The latter can be rationalized by considering the modifications of the I-V curves induced by the different reduction treatments in **Figure 1D**. The I-V curve T3 peak at 11.6 eV observed in stoichiometric cerium oxide film does not change significantly in intensity after mild reduction treatments, while a pronounced decrease in reflectivity at ∼12 eV is observed after strong reduction. We therefore interpret the mild decrease of reflectivity of the ceria islands as originating from an intermediate reduction, in accordance with previous findings from spatially resolved I(V) analysis for chemical reduction of cerium oxide islands on Ru(0001) (Kozlov et al., 2015). The nature of the phase with a lower reflectivity than Pt at 12 eV is instead less clear, although a similar behavior was observed also on reduced ceria films on a Ru(0001) support (Höcker et al., 2017).

In order to have more information on the nature of the different phases obtained with increasing reduction µLEED images were acquired on selected areas of the fully reduced sample. Regions including mainly the brighter islands (**Figure 4** L1) show dominant spots with hexagonal symmetry ascribed to the Pt(111) surface, together with a pattern of lower intensity and smaller size in reciprocal space, compatible with a reduced cerium oxide phase without specific ordering of oxygen vacancies, considering the decreased reflectivity compared to fully oxidized phases (**Figures 4A,C**). The cerium oxide related spots are 5◦ rotated with respect to the Pt ones, possibly due to the higher stability of the rotated orientation (Luches et al., 2011; Sauerbrey et al., 2017). On the contrary, on the areas comprising mainly the phase with a reflectivity lower than Pt µLEED shows superstructures with a 5/2(<sup>√</sup> 3 × √ 3)R30◦ periodicity (**Figure 4** L2). In the literature a similar superstructure was observed by LEED at the early stages of Pt-Ce alloy formation and ascribed to a metallic Pt5Ce alloy (Baddeley et al., 1997; Kemmer et al., 2014). This suggests that there might be a non-negligible tendency for interatomic exchange between cerium oxide and Pt induced by the thermal reduction.

**Figure 5** reports the I-V curves in selected areas of the 1.5 MLE sample as-prepared, after reduction and after re-oxidation. The spectra obtained on areas between the cerium oxide islands correspond to the one expected for clean platinum, with features at 14 eV (P1) and at 18 eV (P2) (red curve in **Figure 5C**). The asgrown cerium oxide islands exhibit an I-V curve that contains both the cerium oxide related features at 3.5, 8, and 11 eV, T1- T3 in analogy with **Figure 1**, and a weak Pt-related P2 feature at 18 eV (yellow curve in **Figure 5C**). After reduction the I-V curve on the bare substrate areas is not significantly altered (red curve in **Figure 5C**), apart from an overall increase of intensity possibly originating from the desorption of oxygen from the Pt surface. The dark areas have a reflectivity that is lower than that of the Pt at all energies (purple curve in **Figure 5C**). The islands, which still appear bright after reduction, have an I-V curve with a dominant R1 peak at low energies around 4 eV, slightly shifted to higher energies compared to the T1 peak before reduction (yellow curve in **Figure 5C**). A broad feature, R2, appears at 7 eV together with attenuated Pt-related P1 and P2 features at 14 and 18 eV. The regular islands which show a mild reflectivity decrease at ∼12 eV with the reducing treatment, like the one highlighted in green in **Figure 4B**, have an I-V curve which appears somewhat intermediate between the ones observed on bright islands and on the low-reflectivity phase (green curve in **Figure 5C**).

A very similar phase with low reflectivity surrounding the brighter islands was also observed after exposure of the ceria islands to hydrogen at a pressure of 1 × 10−<sup>7</sup> Torr at 680 K (**Figures 6A,B**). The I-V curves of the different regions, bright islands, dark island areas and substrate (**Figure 6D**), are compatible with the ones observed after the thermal treatment in UHV (**Figure 5**). The LEED pattern acquired on the low reflectivity areas (**Figure 6C**) shows a reconstruction very similar to the one observed after reduction by thermal treatments in UHV (**Figure 4** L2).

The evolution of the morphology of fully reduced cerium oxide islands during oxidation in oxygen background pressure at 1,040 K was also investigated. **Figure 7** reports a time-lapse sequence of LEEM images acquired during the treatment (see also **Supplementary Video 4**). Already after the first few minutes of oxygen exposure the low-reflectivity phase, initially covering a large part of the surface, starts to vanish while the bright areas increase in size and their reflectivity is increased (area highlighted in red in **Figure 7**). New bright islands also appear during the treatment, like for example the one indicated with a yellow arrow in **Figure 7B**. With further oxygen exposure the low-reflectivity phase completely disappears, and it is replaced by bright islands with a non-uniform contrast and an irregular shape. The LEED patterns after re-oxidation are very close

to the ones acquired before reduction, demonstrating the reversibility of the structure (**Figures 7E,F**). In analogy, also the I-V curves acquired before reduction and after re-oxidation appear very close, demonstrating the reversibility of the surface stoichiometry, while the I-V curves on bare Pt surface areas appear unaltered (**Figure 5C**).

# DISCUSSION

The present study points at understanding the dynamics of the processes which occur at the interface between ceria and platinum during redox processes.

The main result of the work is the evidence for a strong interaction between the two materials involving non-negligible mass transport already at the early stages of isothermal reduction starting at the cerium oxide island edge sites (**Figure 3**). At these sites the formation of oxygen vacancies is likely favored by reduced coordination and by the proximity of Pt which possibly induces a decrease in the oxygen vacancy formation energy by charge transfer (Luches et al., 2015). At the earlier reduction stage the stability of cerium oxide islands with irregular shape is significantly lower than the one of islands with regular triangular shape (**Figure 4**), probably because the reduction is more pronounced where the island edges are less regular, which typically implies the presence of kinks and other kinds of defects, and a low structural order in the long range. The thicker and more regular islands are not significantly altered at the investigated temperature at this earlier stage. Some of the thicker islands increase their lateral size at this stage because some of the neighboring islands with smaller lateral size get some mobility and possibly merge with them. At a later stage most of the islands are partially converted into the alloy and their lateral size is decreased. The low-reflectivity phase, which appears and expands with reducing treatments, is ascribed to a surface alloy based on its LEED pattern although it has a reflectivity which is very low at all energies, as expected for a rough surface with a low crystal quality (Sauerbrey et al., 2016). This may indicate that the surface mobility at the used temperature (1,040 K) is not high temperature (1,040 K) is not high enough to reach a good surface ordering, although the longrange ordered Ce-Pt surface alloy. We note that the appearance of a low-reflectivity phase with reduction low-reflectivity phase with reduction was observed also on the Ru(0001) surface, although it did not show any diffraction pattern (Höcker et al., 2017), probably due to the different miscibility of cerium and Ru, given the alloys (Colinet, 1995). A phase with the same surface periodicity (and LEED dynamic) is observed also on a 1.5 MLE cerium oxide films fully covering the Pt surface on which the reduction is induced by cerium deposition at 680 K (**Figure 3F**), and also after reduction of a 1.8 MLE film (grown at 1,040 K) induced by H<sup>2</sup> reduction of a 1.8 MLE film (grown at 1,040 K) induced by H<sup>2</sup> exposure of at 680 K (**Figure 6**). The formation of a surface alloy on thermally reduced ceria films on Pt(111) was also hypothesized based on atomic and Kelvin probe force microscopy experiments (Gasperi et al., 2018). These evidences indicate that surface alloy formation is a process

FIGURE 6 | (A) LEEM image at E = 12 eV, (B) LEEM image at E = 17 eV, (C) LEED pattern at E = 40 eV and (D) I-V LEEM curves, acquired on selected areas of a 1.8 MLE cerium oxide film grown at 1,040 K and reduced by H<sup>2</sup> exposure (PH2 <sup>=</sup> <sup>1</sup> <sup>×</sup> <sup>10</sup>−<sup>7</sup> Torr) at 680 K: triangular island (orange and green), substrate (red), low-reflectivity (dark) phase (purple).

which also occurs at intermediate temperatures in reducing conditions.

An important aspect is that the extended low reflectivity phase can be transformed back to confined cerium oxide islands by oxidation at elevated temperature (**Figure 7**). The partially reduced bright islands increase in size, suggesting that reoxidation is more favorable at the step edges between reduced ceria and the Pt-Ce alloy, inducing a preferential oxide formation at the sides of the existing islands accompanied by a progressive shrinking of the low-reflectivity phase. Simultaneously also the reflectivity at 12 eV of partially reduced cerium oxide islands increases, evidencing a change in oxidation state and suggesting a filling of surface oxygen vacancies. Moreover, new bright islands appear in isolated sites of the low reflectivity phase, possibly where some cerium oxide nanostructures below the LEEM resolution were left by the reduction treatment or at some other defect sites created during alloy formation. The resulting different morphology of the ceria islands after reoxidation is probably due to kinetic barriers which limit the diffusion and originate less regular structures, compared to the ones formed when Ce adatoms are co-deposited with O<sup>2</sup> on the clean, very smooth Pt surface. At the investigated oxygen pressures and temperatures, both the closed films and the ceria islands were fully re-oxidized in relatively short times, of the order of tens of minutes. Indeed, it would be interesting to follow also the dynamics of re-oxidation in different conditions. The efficiency for molecular oxygen dissociation on the Pt surface for example is expected to contribute to the kinetics of reoxidation at milder oxidizing conditions. We suggest this as a relevant issue to be investigated in future studies, to obtain a complete picture of the redox process. Indeed, a non-negligible and reversible interatomic exchange is likely to occur also in Pt-ceria real catalysts during redox reactions, likely on a shorter length scale determined by the working temperature and by pressure of the reactant. The process needs to be considered in studies devoted to the optimization of cerium oxide/Pt catalysts, for example by minimizing or substituting platinum.

The present work also allows to draw some important considerations on the factors which induce the stabilization of specific reduced phases in cerium oxide. A phase diagram summarizing the phases observed for the different ceria film

FIGURE 7 | Time-lapse LEEM image sequence at E = 12.0 eV acquired during re-oxidation of a fully reduced ceria sample (1.5 MLE) by annealing in <sup>O</sup><sup>2</sup> background pressure (<sup>P</sup> <sup>=</sup> <sup>1</sup> <sup>×</sup> <sup>10</sup>−<sup>7</sup> Torr) at 1,040 K: (A) before the re-oxidation; (B) after 130 s; (C) after 150 s; (D) after 280 s. The red ellipse allows to follow the evolution of the morphology in a selected surface region. The orange arrow in (B) evidences a bright feature that was absent before the treatment and that is growing with time (C,D). (E,F) µLEED patterns (E = 40 eV) acquired on two islands of different shape after full reduction and subsequent re-oxidation by exposure to oxygen at PO2 <sup>=</sup> <sup>1</sup> <sup>×</sup> <sup>10</sup>−<sup>7</sup> Torr for 5 min at 1,040 K.

thicknesses and at the different reducing conditions is shown in **Figure 8**. The (3 × 3) phase was observed both on 1.5 MLE and on 5 MLE films, although in the first case a much smaller amount of cerium was required for its stabilization. The Pt substrate possibly favors the formation of the (3 × 3) phase which has a surface size compatible with the coincidence cell at the ultrathin limit (Luches et al., 2013). The CeO1.67 (3 × 3) phase was already observed in cerium oxide ultrathin films supported on different metal substrates, e.g., Cu(111) (Duchon et al., 2014b; Höcker et al., 2015), while it was not observed on bulk surfaces (Olbrich et al., 2017), although a recent theoretical study showed that it can result from the (111) termination of a metastable bulk Ce3O<sup>5</sup> phase (Murgida et al., 2018). The reduced dimensionality in the out of plane direction probably plays a role in its stabilization, given the higher structural flexibility of

thin films compared to bulk surfaces. The Ce2O<sup>3</sup> (4 × 4) phase was shown to coexist with the (3 × 3) phase on 5 ML films at the maximum degree of reduction accessed, while on thinner films it was not observed, being probably less stable than other reduced phases. The phase was also observed in previous studies on ultrathin films on different substrates (Duchon et al., 2014a,b; Höcker et al., 2015), as well as on bulk-like surfaces (Olbrich et al., 2017).The reduced (5 × 5) phase, observed in this work on 1.5 MLE films at intermediate reduction (**Figure 3D**), was never observed in previous studies. We speculate that its stabilization may also be ascribed to epitaxy with the Pt substrate, since a 5:7 coincidence between cerium oxide and platinum is possible and it implies a slight surface expansion of the cerium oxide surface lattice (Luches et al., 2013), indeed favored by reduction since Ce3<sup>+</sup> ions have a larger radius than Ce4<sup>+</sup> ions. The (5 × 5) phase gradually evolves with further reduction toward the 5/2(<sup>√</sup> 3 × √ 3)R30◦ phase which is related to the initial stages of an alloy formation as discussed above. The alloy was observed at low thickness and moderate degrees of reduction, while for thicker films the onset for alloy formation was found at higher degrees of reduction. The plethora of surface reconstructions, the fast transformation between the different phases and the partially different behavior of cerium oxide layers close to the interface compared to the surface of thin films highlight the complexity of the processes that occur at the interface between these two technologically important materials during redox processes.

Finally, this work proves that for a thorough understanding of the complex phenomena which occur during reduction and oxidation a real-time investigation not only of the evolution in morphology, but also of the structure and electronic properties is extremely important to fully understand the system behavior.

# CONCLUSIONS

We have investigated the dynamic changes in morphology, structure, and electronic properties in cerium oxide films and confined islands with reducing and oxidizing treatments. We relate the observed modifications to the formation of different reduced structural phases, and we discuss the effect of temperature, of dimensionality and of the interaction with platinum for their stabilization. In particular, the CeO1.67 (3 × 3) phase gradually transforms into the Ce2O<sup>3</sup> (4 × 4) bixbyite phase, with increasing reduction in sites which are not in close proximity with the Pt interface. This is in contrast to the very early stage of the reduction process, more prominent where the lateral oxide-metal interface is less smooth and more defective. The resulting (3 × 3) phase is stabilized at milder conditions in ultrathin films due to an important role of the Pt substrate. At higher degrees of reduction the interaction between cerium oxide and Pt induces the formation of more complex phases, and it finally leads to a partial interdiffusion between substrate and overlayer atoms. The structural modifications are reversible with oxidizing treatments, whereas a complete restoration of the original morphology is partially hindered by kinetic barriers.

# AUTHOR CONTRIBUTIONS

PL, JIF, JF, and SV conceived and designed the experiment. GG and MS did the sample growth, the measurements and the data analysis. JIF and JF supervised the experiment and the data

## REFERENCES


analysis. PL wrote the manuscript draft. All authors contributed to data interpretation, discussion of results and manuscript critical revisions.

# ACKNOWLEDGMENTS

The activity was performed with the support of the COST Action CM1104 Reducible oxide chemistry, structure and functions.

# SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fchem. 2019.00057/full#supplementary-material

Supplementary Video 1 | LEED movie (E = 40 eV) during reduction of a 5 ML cerium oxide film by Ce deposition at 680 K on the film surface.

Supplementary Video 2 | LEED movie (E = 40 eV) during reduction of a 1.5 ML cerium oxide film by Ce deposition at 680 K on the film surface.

Supplementary Video 3 | LEEM movie (E = 12 eV) of cerium oxide islands of nanometric thickness during thermal reduction in UHV at 1,040 K.

Supplementary Video 4 | LEEM movie (E = 12 eV) of reduced cerium oxide islands of nanometric thickness during thermal reoxidation at 1,040 K in oxygen background pressure (PO2 <sup>=</sup> <sup>1</sup> <sup>×</sup> <sup>10</sup>−<sup>7</sup> Torr).


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Luches, Gasperi, Sauerbrey, Valeri, Falta and Flege. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Oxygen Vacancies in Oxide Nanoclusters: When Silica Is More Reducible Than Titania

### Andi Cuko1,2, Stefan T. Bromley 1,3 and Monica Calatayud<sup>2</sup> \*

<sup>1</sup> Departament de Ciència de Materials i Química Física, Institut de Química Teòrica i Computacional (IQTCUB), Universitat de Barcelona, Barcelona, Spain, <sup>2</sup> CNRS, Laboratoire de Chimie Théorique, LCT, Sorbonne Université, Paris, France, 3 Institució Catalana de Recerca i Estudis Avançats (ICREA), Barcelona, Spain

Oxygen vacancies are related to specific optical, conductivity and magnetic properties in macroscopic SiO<sup>2</sup> and TiO<sup>2</sup> compounds. As such, the ease with which oxygen vacancies form often determines the application potential of these materials in many technological fields. However, little is known about the role of oxygen vacancies in nanosized materials. In this work we compute the energies to create oxygen vacancies in highly stable nanoclusters of (TiO2)N, (SiO2)N, and mixed (TixSi1−xO2)<sup>N</sup> for sizes between N = 2 and N = 24 units. Contrary to the results for bulk and surfaces, we predict that removing an oxygen atom from global minima silica clusters is energetically more favorable than from the respective titania species. This unexpected chemical behavior is clearly linked to the inherent presence of terminal unsaturated oxygens at these nanoscale systems. In order to fully characterize our findings, we provide an extensive set of descriptors (oxygen vacancy formation energy, electron localization, density of states, relaxation energy, and geometry) that can be used to compare our results with those for other compositions and sizes. Our results will help in the search of novel nanomaterials for technological and scientific applications such as heterogeneous catalysis, electronics, and cluster chemistry.

Keywords: TiO2 , SiO2 , titanosilicates, nanoclusters, oxygen vacancy, reducibility

# INTRODUCTION

Metal oxides are an important class of materials in many scientific and technological fields. Silica and titania are abundant compounds and play crucial roles in geology (Howie, 1964; Lakshtanov et al., 2007; Lyle et al., 2015), astrochemistry (Lambert et al., 2012; Lee et al., 2015; Komatsu et al., 2018), environmental and chemical sciences (Joo et al., 2014; Kapilashrami et al., 2014), biology and medicine (Mal et al., 2003; Nguyen et al., 2005; Das et al., 2009; Bawazer et al., 2012; Lucky et al., 2016; Saroj and Rajput, 2018). Their ability to exchange oxygen drastically determines their properties. In the case of TiO2, the loss of lattice oxygen results in a change of color from white to blue, the appearance of magnetism, and is connected with photocatalytic properties (Houlihan et al., 1975; Diebold, 2003; Di Valentin et al., 2006; Na-Phattalung et al., 2006; Vásquez et al., 2016; Li et al., 2018). On the other hand, oxygen vacancies present in amorphous silica have been reported to modify its electronic and optical properties (Rudra and Fowler, 1987; Skuja, 1998; Manveer et al., 2017). Oxygen substoichiometry in silica-based materials is linked to dielectric breakdown and

### Edited by:

Javier Carrasco, CIC Energigune, Spain

### Reviewed by:

Gianfranco Pacchioni, Università degli Studi di Milano Bicocca, Italy Alexander Shluger, University College London, United Kingdom

\*Correspondence:

Monica Calatayud calatayu@lct.jussieu.fr

### Specialty section:

This article was submitted to Physical Chemistry and Chemical Physics, a section of the journal Frontiers in Chemistry

Received: 09 November 2018 Accepted: 14 January 2019 Published: 07 February 2019

### Citation:

Cuko A, Bromley ST and Calatayud M (2019) Oxygen Vacancies in Oxide Nanoclusters: When Silica Is More Reducible Than Titania. Front. Chem. 7:37. doi: 10.3389/fchem.2019.00037

**150**

electroforming in resistive random access memory devices (McPherson and Mogul, 1998; Yao et al., 2010; Wang et al., 2014; Mehonic et al., 2018). Whereas, the effect of creating an oxygen vacancy in bulk and surfaces of both titania and silica is well-reported, little is known about the role these defects may play in nanosized materials and clusters. By means of density functional theory (DFT) based calculations, we characterise the energetics of creating an oxygen vacancy in silica, titania and titanosilicate clusters recently derived from global optimization studies. Low energy clusters can be generated with cluster beam experiments and stabilized by soft landing techniques. Such studies are increasingly looking into the reactivity of the cluster species formed for technologically important reactions (Yin and Bernstein, 2014; Vajda and White, 2015; Sun et al., 2016). With our work we hope to elucidate the reactivity of such nanoclusters as estimated through their reduction by oxygen removal.

From a computational point of view, the energetic cost of removing a neutral oxygen atom from a stoichiometric metal oxide can be calculated with the following equation:

$$\begin{aligned} \mathrm{M\_nO\_{2n}} &\rightarrow & \mathrm{M\_nO\_{2n-1}} + 1/2 \,\mathrm{O\_2} \\ \mathrm{E\_{vac}} &= & \mathrm{EM\_nO\_{2n-1}} + 1/2 \mathrm{EO\_2} - \mathrm{EM\_nO\_{2n}} \end{aligned} \tag{1}$$

A reducible oxide like titania tends to have small (positive) Evac values, while the energy to form an oxygen vacancy in an irreducible oxide like silica is relatively higher. Evac values can thus be used to compare different systems and provide an initial ranking of reducibility (Ganduglia-Pirovano et al., 2007; Deml et al., 2015; Helali et al., 2017). Typical values for bulk and surface Evac values in rutile and anatase TiO<sup>2</sup> are 3.7–4.5 eV whereas for SiO<sup>2</sup> the values are much higher, 5–8 eV (see Sushko et al., 2005). In addition, the localization of the electrons remaining after the removal of the oxygen in the lattice provides crucial information related to the electronic structure of the defective material (Helali et al., 2017):

$$M^{\Pi^{+}} - O^{2-} - M^{\Pi^{+}} \rightarrow M^{\Pi^{+}} - 2e^{-} - M^{\Pi^{+}} + \forall O\_{2} \tag{2}$$

It is commonly accepted that removing an oxygen atom from TiO<sup>2</sup> leads to the formation of paramagnetic Ti3<sup>+</sup> sites in both bulk and surfaces (Houlihan et al., 1975; Di Valentin et al., 2006; Li et al., 2018). The electronic state of reduced titania systems is typically open-shell. In amorphous SiO<sup>2</sup> and quartz the removal of a lattice oxygen results in the formation of a Si-Si covalent bond where the electron pair is stabilized by two Si sites (Sushko et al., 2005). In the quartz surface the removal of an oxygen atom leads to an electron pair localized on a surface silicon site (Silvi and D'Arco, 2006; Causà et al., 2015). This typically leads to electronic states that are closed-shell.

We focus on the effects that the system size reduction to the nanoscale has on the type of oxygen defects exhibited and their resultant Evac values. As has been previously reported for ZrO<sup>2</sup> (Ruiz Puigdollers et al., 2016a,b), CeO<sup>2</sup> (Kozlov et al., 2015), TiO<sup>2</sup> (Morales-García et al., 2018), this may result in a significant change in the materials' properties. Specifically, in this work we investigate the energetic, structural and electronic properties of silica, titania and mixed titanosilicates as regards their tendency to lose neutral oxygen atoms. These properties are then compared between the different oxides and with respect to the system size, including up to the bulk. At the nanoscale level, we consider a set of globally optimized silica and titania clusters with sizes between 2 and 24 MO<sup>2</sup> units. Moreover, we compare our results for these pure oxides with a comparative set of titanosilicate clusters for the size of 10 MO<sup>2</sup> units. At the bulk level, we consider alpha quartz and rutile phases, respectively, for silica and titania, which are the two most stable phases under ambient conditions. Such systems are then reduced by a single and neutral oxygen vacancy and the results are analyzed to elucidate the role of cluster size, structure and oxide type with respect to the formation of oxygen vacancies.

# METHODOLOGY

All nanocluster structures selected in this study correspond to the lowest energy global minima candidates currently reported in the literature as derived by us and/or other authors: SiO<sup>2</sup> systems are derived from Flikkema and Bromley (2003, 2004), Lu et al. (2003), Bromley and Flikkema (2005), TiO<sup>2</sup> systems from Hamad et al. (2005), Woodley et al. (2006), Calatayud et al. (2008a), Syzgantseva et al. (2011), Chen and Dixon (2013), Bhattacharya et al. (2015), Aguilera-Granja et al. (2016), Lamiel-Garcia et al. (2017), and finally mixed titanosilicates are taken from Cuko et al. (2018). A two-step procedure is followed to compute Evac (Equation 1): (i) a selected oxygen atom is removed from a structurally frozen stoichiometric nanocluster and the pre-removal/post-removal energy difference is defined as Eunrel, (ii) the post-removal non-stoichiometric nanocluster structure is then fully optimized, with the pre-optimized/post-optimized energy difference defined as Evac. The relaxation energy Erel is then derived as the difference between Eunrel and Evac. For the smaller sizes, N = 2–10, every symmetrically distinct oxygen site was tested and only the most representative ones were reported. For larger sizes, an exhaustive approach was found to be too computationally demanding and a sampling of approximately 10 oxygen sites was performed, always including all symmetrically non-equivalent terminating oxygens. All DFT based calculations involving reduced nanoclusters were carried out using a spinpolarized formalism, employing the PBE0 functional (Adamo and Barone, 1999) and a tier-1/tight numerical basis set as implemented in the FHI-AIMS code (Blum et al., 2009). In most of the calculations, we did not force the spin multiplicity of the system. However, in a few cases, especially for TiO<sup>2</sup> systems, in order to overcome convergence problems, we forced the spin multiplicity to be either an open shell singlet or a triplet. Recently, room temperature ferromagnetism has been reported for zirconia nanoparticles (Rahman et al., 2016; Albanese et al., 2018). This effect might also take place in titania and could be studied by considering multiple vacancies in future works.

# RESULTS AND DISCUSSION

# Pure (SiO2)<sup>N</sup> Nanoclusters

**Figure 1** shows the calculated values of Evac for the lowest energy SiO<sup>2</sup> nanoclusters together with some selected nanocluster structures. **Figure S1** shows the structures of all stoichiometric silica nanoclusters used in this work and **Table S1** includes the corresponding values of Evac, Eunrel, and Erel for the most stable reduced systems. Generally, stoichiometric silica global minimum nanoclusters tend to possess quite open symmetric structures and typically have between 2 and 4 unsaturated (terminal) oxygen sites. The majority of these structures present silanone centers, which are triply-coordinated planar Si centers involving a terminal oxygen atom (i.e., formally >Si=O, but arguably more accurately >Si+-O−) (Avakyan et al., 2006; Zwijnenburg et al., 2009). For nearly all the silica nanoclusters considered, the terminal oxygen atoms at these sites are found to cost the least energy to be removed, 2.6–2.8 eV, with Evac increasing slightly and monotonically with the size. The (SiO2)<sup>12</sup> nanocluster is the only exception as it only displays non-bridging oxygen (NBO) terminal species (i.e., ≡Si–O species where Si is 4 fold coordinated) rather than silanones. Removing the terminal oxygen atom of an NBO center results in a higher Evac value (3.60 eV) as compared with silanone defects (2.6–2.8 eV). We note that the values of Evac for the removal of oxygen atoms at both silanone and NBO defects are much lower than those for the removal of 2-fold coordinated oxygen atoms in the corresponding bulk systems. From our calculations, for example, we obtained Evac values between 5.06–5.39 eV for a-quartz with a range of vacancy concentrations between 1/24 and 1/3 vacancies per SiO<sup>2</sup> units. These values are also consistent with other work where an Evac value of 5.15 eV was obtained using plane-wave based DFT calculations at the PBE level for the alpha-quartz structure with a vacancy concentration of 1/48 (Helali et al., 2017). We note that, removing a 2-fold oxygen atom in our nanoclusters typically costs between 4.2 and 5.7 eV (see **Figure S2** which shows selected values for Si10O20), which is comparable to the Evac values obtained for bulk alpha quartz. This confirms that the terminal silanone oxygens are least energetically costly in silica nanoclusters. From a structural point of view, reducing a silanone defective site in a nanocluster leads to a 2-fold Si site displaying a Si-O distance of 1.71 Å. This is slightly longer than the Si-O bond length of 1.62 Å as calculated by Pacchioni and Ferrario (1998) for a 2-fold coordinated Si defect in alpha quartz using a cluster approach at both wave function theory level employing unrestricted Hartree Fock, and DFT level employing the hybrid B3LYP functional. Further comparing our results with these bulk cluster calculations, in our clusters the O-Si-O angle of a reduced silanone site is found to be between 87 and 111◦ , depending on whether the defect is part of a two or three membered ring. This compares with a value of 102◦ found from bulk cluster calculations. Also, the vicinal Si-O-SiII angle (where SiII corresponds to the 2-fold coordinated silicon atom) is found to be between 90 and 133◦ in our nanoscale clusters against the corresponding values for the same defective site of 141–161◦ from bulk cluster calculations.

We note that for silica nanoclusters, relaxation is found to have little impact on the energetics of oxygen removal. An analysis of **Figure 1** (see **Table S1** for the values) shows that Erel is generally between 0.06 and 0.12 eV, and is thus <5% of the corresponding Evac. The structures of the relaxed O-deficient nanoclusters are very similar to the original stoichiometric structures. The only exception is the nanocluster structure

for N = 12 that possesses two NBO sites. Oxygen vacancies originating from 2-fold coordinated oxygen atoms, which at nanoscale are less energetically favored than those generated from mono-coordinated oxygen sites, are also well-studied for bulk silica in the literature (Skuja et al., 1984; Skuja, 1998; Sulimov et al., 2002). Specifically, the ≡Si-Si≡ equilibrium distance is reported to be 2.3–2.5 Å in α-quartz (Skuja et al., 1984; Skuja, 1998; Sulimov et al., 2002). At the nanoscale, after removal of two-coordinated oxygens we find the ≡Si-Si≡ equilibrium distance to be between 2.12–2.37 Å, slightly lower with respect to the range reported for the bulk. Moreover, bulk oxygen removal is known to induce a long range structural distortion and asymmetric relaxation of atoms surrounding the vacancy (Sulimov et al., 2002). Such long-range effects are more difficult to evaluate for clusters for which the sizes range between 5 and 20 Å. However, generally we observe a relatively strong local relaxation for the neighboring atoms, whereas no significant long-range relaxation is observed (usually <0.1 Å of total displacement for distances >3 Å from the vacancy).

The electronic structure of the oxygen-deficient nonstoichiometric silica clusters is consistently found to be a singlet, as expected. The electron density of the highest occupied state is mainly located on the O-deficient silicon site and has the shape of a sp orbital, see **Figure S3**.

# Pure (TiO2)<sup>N</sup> Nanoclusters

with respect to nanocluster size.

**Figure 2** shows the calculated values of Evac for the lowest energy (TiO2)<sup>N</sup> nanoclusters together with selected nanocluster structures. **Table S1** shows the corresponding values of Evac, Eunrel, and Erel obtained and **Figure S4** shows the (TiO2)<sup>N</sup> structures used. It can be observed that the (TiO2)<sup>N</sup> nanocluster structures are more compact and less symmetric than silica clusters for the range selected, with titanium atoms being in 4-, 5-, and 6-fold coordinated environments. Many of these nanocluster structures, especially for sizes smaller than N = 20, possess terminal oxygen sites bonded to a Ti atom in a tetrahedral environment (≡Ti–O species). If present, these terminal oxygens are generally the least energetically costly to remove. In some cases (e.g., N = 4, 9, 11, 12), although the removal of a terminal oxygen does not have the lowest associated Eunrel value, removing a 2-fold oxygen atom nearby can lead to a reduced nanocluster with the same topology as the one that results from removing the terminal oxygen itself. Evac values for (TiO2)<sup>N</sup> nanoclusters are found to oscillate between 2.48 and 5.01 eV and display no clear size-dependent trend, contrary to the case of silica nanoclusters. The energetic cost to remove an oxygen atom seems to depend critically on the structure of each cluster. In very few cases (i.e., N = 17, 22), the Evac values for TiO<sup>2</sup> nanoclusters was found to be comparable with that found for SiO<sup>2</sup> nanoclusters. In such cases, after the relaxation, the TiO<sup>2</sup> nanocluster structure was found to have neither terminal oxygen sites (usually the site of the oxygen atom removal) nor 3-fold coordinated Ti defective centers, known to be highly destabilizing when left after the removal of a terminal site. We were not able to obtain a Evac value for bulk systems with the computational setup used for the nanoclusters due to severe convergence problems associated with the presence of Ti3<sup>+</sup> centers. Other authors have calculated Evac for bulk and surfaces of anatase and rutile using DFT-based calculations employing hybrid functionals incorporating between 20 and 25% Hartree Fock exchange, obtaining values between 4.8 eV (Finazzi et al., 2008; Yamamoto and Ohno, 2012; Deák et al., 2014) and 5.4 eV (Islam et al., 2007; Janotti et al., 2010; Lee et al., 2012). We also note that the Evac value for a terminal oxygen in an anatase bulk cut (TiO2)<sup>35</sup> nanoparticle was calculated to be 3.95 eV, using the same methodology as employed herein (Kim et al., 2016). This value is similar to the average value of 3.65 eV, calculated for the lowest Evac values over our set of nanoclusters. Interestingly, for the (TiO2)<sup>35</sup> nanoparticle, the removal of an internal bulk-like three-coordinated oxygen resulted in a lower Evac (i.e., 3.65 eV) with respect to that for removing a terminal oxygen atom. This behavior probably derives from the strong geometrical relaxation induced from the vacancy to the already metastable anatase bulk cut structure.

The relaxation of the TiO<sup>2</sup> nanoclusters after oxygen removal is found to involve more energy than in the case of silica nanoclusters. Specifically, the average Erel value for the lowest energy clusters is found to be 1.50 eV, and in some cases it accounts for more than 2.5 eV (N = 21, 23). With a few exceptions, such as for N = 4, 17, 22, the removal of terminal oxygens induces a moderate degree of local relaxation. However, when 2-fold coordinated oxygen atoms are removed, significant structural rearrangements take place (i.e., N = 10, 20, 21, 23, 24). In **Figure S3** we show the O-deficient structure for N = 11 where a 2-fold oxygen atom is removed in the presence of a terminal

oxygen, which then subsequently moves during the relaxation to fill the vacancy.

The electronic structure of the O-deficient clusters is either a triplet with two unpaired electrons occupying two titanium sites or an open shell singlet with unpaired electrons equally delocalized around three Ti centers neighboring the vacancy. **Figure S3** displays the electron density associated with the highest occupied state of the (TiO2)<sup>10</sup> and (TiO2)<sup>11</sup> nanocluster after the removal of an oxygen atom, which is localized on Ti3<sup>+</sup> sites that carry the electrons left in the vacancy.

# Mixed (SixTi1–xO2)<sup>N</sup> Nanoclusters

nanocluster size.

In order to study the effect of mixing SiO<sup>2</sup> and TiO<sup>2</sup> on Evac values, we have selected (TixSi1−xO2)<sup>10</sup> nanoclusters and a Ti content x between 0 and 1. The nanocluster structures used are taken from the global optimized structures reported in Cuko et al. (2018) (see also **Figure S5**). The nanoclusters exhibit different isomers for pure and mixed structures. All the structures possess terminal ≡Ti-O oxygen sites, with the exception of the pure (TiO2)<sup>10</sup> nanocluster. The pure silica (SiO2)<sup>10</sup> nanocluster displays only silanone terminations, whereas mixed systems display typically two terminal oxygen centers bonded with

titanium atoms. The presence of titanium thus seems to allow for relatively more energetically stable terminal oxygen sites than silicon (Cuko et al., 2018). **Figure 3** shows that the Evac values obtained for the mixed nanocluster systems oscillate around 3.7 eV, with the lowest value being found for x = 0.1 (Evac = 3.29 eV) and the highest value for x = 0.3 (Evac = 4.54 eV). The mixed nanocluster structure with x = 0.1 possesses two terminal oxygen sites bound to a Si and Ti atom. The removal of the terminal oxygen is more energetically facile (i.e., Evac is relatively lower) when it is bound to a silicon atom rather than a titanium atom. As the content of TiO<sup>2</sup> increases, the Evac values are found to initially increase up to x = 0.30, and afterwards decrease and oscillate with no obvious trend.

Over the range of (TixSi1−xO2)<sup>10</sup> nanoclusters, relaxation effects are relatively less significant for clusters with titania contents of x = 0.0, 0.1, 0.3. In the first two cases, indeed, the less energetically costly oxygen to be removed is from a silanone or a Si-NBO site, and therefore, the local structural relaxation effects are similar to those of pure silica systems. The system with x = 0.2 is a special case since lowest Evac value we could find is related to the removal of a two coordinated oxygen as shown. The removal of this oxygen atom leads to a significant structural change during the geometry optimization with a consequent relatively large associated Erel value (see **Figure 3**). For higher Ti content, Erel

energies are found to be in the range of 0.6–1.5 eV which is typical for pure TiO<sup>2</sup> nanoclusters. Overall, the mixed (TixSi1−xO2)<sup>10</sup> nanocluster systems tend to behave more like titania nanoclusters rather than silica nanoclusters in terms of reduction through O-vacancy formation.

# DISCUSSION

The results obtained point to a (sub)nanosize dependency of the formation of oxygen vacancies in silica- and titania-based structures. Firstly, the removal of singly-coordinated terminal oxygen atoms is often associated with low Evac values. This is explained by the low energetic cost of breaking an M-O bond compared with the energy required to break two M–O bonds for oxygen atoms in 2-fold coordinated sites. This is observed in the three types of nanoclusters studied in this work and has been reported in the literature for other oxides, especially in systems exhibiting structural stable dangling bonds. For instance, V2O<sup>5</sup> in bulk, surface and clusters exhibit structurally stable terminal oxygen sites. The coordination of oxygen sites in vanadia is commonly associated with chemical reactivity (Calatayud and Minot, 2006; Ganduglia-Pirovano et al., 2007; Calatayud et al., 2008b). The lower Evac found for nanoclusters compared to bulk values indicates higher reducibility of the nanosized materials as has been reported for other oxide systems like CeO<sup>2</sup> and ZrO<sup>2</sup> (Bruix and Neyman, 2016; Ruiz Puigdollers et al., 2016b).

It is, however, worth noting that the lowest Evac values found for silica clusters, 2.3–3.6 eV, are generally lower than for titania clusters, 2.48–5.01 eV as also shown from **Figures 1**, **2** and **Table S1**. A direct comparison between bulk SiO<sup>2</sup> and bulk TiO<sup>2</sup> was not possible in this study due to the high computational cost and convergence problems with TiO<sup>2</sup> systems. However, there are several studies on O vacancy defects formation in metal oxides (Pacchioni, 2008; Deml et al., 2015; Helali et al., 2017) that agree on associating a higher reducibility to titania (either rutile and anatase phases) with respect to silica bulk (mainly α-quartz). Our results suggest that silica becomes more reducible than titania at the nanoscale, which is contrary to the properties observed for respective bulk materials. The reason for such inverted behavior seems to be connected to the presence of terminal silanone oxygen atoms in low energy silica nanoclusters. The electronic structure (projected density of states) of this site is displayed in **Figure 4**, showing that the energy levels are highly localized in two sharp peaks (one between −9 and −10 eV, and the other between −10.5 and −11.5 eV) associated to higher reactivity. In comparison, the oxygen levels of the 2-fold oxygen in (TiO2)<sup>10</sup> are distributed over a 3 eV range of energies (between −9 and −12 eV). In the (TiO2)<sup>11</sup> cluster, the electronic energy levels associated with the dangling oxygen lay closer to the Fermi level, between −9 and −11 eV, and those associated with the 2-fold coordinated site levels spread between −9 and −13 eV. The electronic structure of the dangling silanone oxygen in silica is consistent with an enhanced reactivity of this site and with the low Evac values shown. Interestingly, the oxygen vacancy in bulk silica has been reported to exhibit specific electronic transitions associated with the electron pair localized on the

FIGURE 4 | Projected density of states (PDOS) on for oxygen atoms for (SiO2)10 (left), (TiO2)10 (middle), and (TiO2)11 (right). Ox indicates the projection on the oxygen center marked with a cross (blue). Dangling O indicates the projection on all single terminal oxygen sites (green). Vertical lines indicate the position of the Fermi level. The PDOS plot has been smeared for clarity.

silicon atom after reduction (Cannizzo and Leone, 2004; Sushko et al., 2005). We also observe the highest occupied electronic level of O-deficient (SiO2)<sup>10</sup> cluster localized on the resulting Si< site (**Figure S3** top right), which is consistent with the presence of an electron pair localized on the silicon site.

A second important point is the structure-dependency of the Evac values. In the case of silica nanoclusters, the Evac value is almost constant and related to the presence of Si sites with dangling silanone oxygen sites. However, titania nanoclusters display a greater range of Evac values, depending on the structure near to the removed oxygen atom. Relaxation effects can also be significantly stronger in titania nanoclusters than in silica nanoclusters, which is consistent with the greater variety of Evac sites and values found. The size dependency of Evac values is also different for each nanomaterial. For the system sizes considered here, the majority of silica nanoclusters exhibit a near constant value of Evac ∼ 2.6 eV associated with silanone sites, and exceptionally an Evac value of 3.6 eV associated with the NBO in the (SiO2)<sup>12</sup> nanocluster. For larger sizes, silica nanoclusters are predicted to become fully-coordinated (Flikkema and Bromley, 2009), which will result in a jump of Evac to near bulk values ∼5 eV. As can be seen in **Figure S2**, 2-fold coordinated oxygens cost 4.4–5.5 eV. In the case of pure titania nanoclusters there seems to be no clear trend with size.

Finally, mixing nanosilica and nanotitania results in chemically and structurally rich nanocluster systems displaying a variety of oxygen sites. Interestingly the presence of titanium induces the disappearance of silanone sites in favor of pyramidal ≡Ti–O species. Whereas terminal oxygen sites are relatively unstable when bonded to silicon, they are stabilized when bonded to titanium. As for pure titania nanoclusters, the values of Evac are found to strongly depend on the structure, the coordination of the oxygen removed, and the degree of relaxation after oxygen removal.

In order to further study the role of the chemical nature of Si and Ti in the oxygen vacancy energy, we consider a set of (TixSi1−xO2)<sup>10</sup> structural isomers with the same topology (**Figure 5**), with selected compositions between x = 0–1. Such isomer structures are global minima for titanosilicates with x = 0.1, 0.2, and 0.3, and are slightly less stable than the global minima for pure systems. They have been geometrically optimized in their stoichiometric composition. We then remove

FIGURE 5 | Top: a set of (TixSi1−xO2)10 cluster structures exhibiting the same structural isomer for pure silica (x = 0), pure titania (x = 1) and all intermediate compositions. Bottom: the corresponding oxygen formation energy without relaxation (Eunrel) in eV for all non-equivalent oxygen sites. Vacancies are generated after removing a 2-fold coordinated oxygen (gray), terminal NBO1 (blue), and terminal NBO2 (green).

each non-equivalent oxygen atom keeping the initial geometry frozen. In **Figure 5** we report the resulting unrelaxed oxygen vacancy formation energies (Eunrel).

**Figure 5** shows Eunrel as a function of the mixing composition xTiO2 of the cluster. In the plot there are three families of oxygen vacancies according of the type of oxygen removed: the vacancies originating from the removal of two-coordinated oxygen atoms are denoted in gray, whereas those originating from removing the terminal oxygen sites labeled NBO<sup>1</sup> and NBO2, respectively, are denoted in blue and green. In most of the cases, vacancies at terminal position are found to provide the lowest Eunrel value. For the pure silica cluster isomer, the Eunrel values for the NBO<sup>1</sup> and NBO<sup>2</sup> are 3.38 and 3.45 eV, respectively, which values are significantly lower than the corresponding values for pure titania systems, 4.69 and 5.38 eV. This result further confirms that the oxygen vacancy formation energy is driven by the presence of unsaturated O centers and, particularly when such centers are bonded to Si atoms, are more easily removed than when they are bonded to Ti atoms. This is clearly seen in the x = 0.1 isomer that possesses one terminal oxygen on Si and one on Ti: the Eunrel for the Si-bonded oxygen NBO<sup>1</sup> is 3.50 eV whereas for the Tibonded oxygen NBO<sup>2</sup> it is 5.65 eV. Although relaxation effects have not been considered, based on the results discussed above, they are not expected to be significant. Therefore, the nature of the cationic site, Si or Ti, seems to critically determine the stabilization of the terminal oxygen deficient systems.

# CONCLUSIONS

In this work we use DFT-based calculations employing the PBE0 hydrid functional to explore oxygen vacancy formation in globally optimized TiO2, SiO2, and mixed TixSi1−xO<sup>2</sup> nanoclusters for sizes between 2 and 24 stoichiometric oxide units. The properties computed (oxygen vacancy formation energies, electronic structure, nanocluster structure, relaxation effects) are found to critically depend on the nature of the oxide. The behavior of silica nanoclusters is found to be rather constant and dominated by the presence of dangling oxygens in silanone >Si=O sites. In titania and titanosilicate nanoclusters; however, we find a strong dependence of the properties with respect to the local geometry. When unsaturated oxygen centers are present in the system, removing such centers is energetically less costly than removing two-coordinated oxygen atoms. Since at the nanoscale such unsaturated centers are naturally present in low energy systems, contrary to the bulk and extended surfaces, we can predict that the reduction of nanosized silica clusters is energetically more favorable than in the bulk. Furthermore, we found that the oxygen vacancy formation in silica nanoclusters is also more favorable than from similar-sized low energy titania nanoclusters. This demonstrates the emergence of an unexpected chemical behavior induced by the small size of the systems considered. We hope that our results will help in the search of novel

# REFERENCES


materials for applications in scientific and technological fields such as heterogeneous catalysis, electronics, and cluster chemistry.

# AUTHOR CONTRIBUTIONS

AC performed all the calculations, figures, tables, analyzed the data, discussed results, and participated in the elaboration of the manuscript. SB and MC directed the work, discussed the data, and wrote the manuscript.

# FUNDING

This research was supported by the H2020 project ITN-EJD-642294 (TCCM: Theoretical Chemistry and Computational Modeling). The Spanish MINECO/AEI-FEDER grant: CTQ2015-64618-R and, in part, Generalitat de Catalunya grants: 2017SGR13 and XRQTC are acknowledged.

# ACKNOWLEDGMENTS

We also acknowledge support from the Spanish MCIU Maria de Maeztu award MDM-2017-0767. The GENCI-CINES/IDRIS (Grant 2016-x2016082131, 2017-x2017082131, 2018-x2018082131) and the Red Española de Supercomputación provided valuable computing time.

# SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fchem. 2019.00037/full#supplementary-material

Supplementary Data Sheet 1 | Oxygen vacancy formation energies with and without relaxation (Evac, Eunrel, Erel), nanocluster structures employed, Evac values for different O-vacancies, electron density plots of selected O-deficient structures.

algorithms: tuning the electronic properties of (TiO2)<sup>n</sup> clusters. Phys. Rev. B 91, 241115–241115. doi: 10.1103/PhysRevB.91.241115


for renewable energy applications. Chem. Rev. 114, 9662–9707. doi: 10.1021/cr5000893


and \$H(mathrm{III})\$ centers in pure and doped silica from ab initio calculations. Phys. Rev. B 58, 6090–6096. doi: 10.1103/PhysRevB.58.6090


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Cuko, Bromley and Calatayud. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.