The most mature technology for CO₂ capture is absorption by amines, such as monoethanolamine (MEA) in water (Rochelle, 2009; Luis, 2016; Bernhardsen and Knuutila, 2017; Wu et al., 2020). Proprietary solvents ⁴ , water-lean solvents (Jiang et al., 2023), and ionic liquids (Wang et al., 2011; Ramdin et al., 2012; Mota-Lima et al., 2021) are also in use. Chemical Informatics techniques have been applied to search for better solvents before (Li et al., 2014; Papadopoulos et al., 2016; Yang et al., 2017; Puxty and Maeder, 2020; Orlov et al., 2022).

For our analysis, MDLab begins with an identification of the most important and relevant structural characteristics of the known solvents and uses these to identify other molecules that may show promise for carbon capture. We identified 167 unique amines with CO₂ capture properties in the literature (McDonagh et al., 2022) and extracted string representations of these molecules from PubChem (Kim et al., 2021) and ChemSpider (Pence and Williams, 2010). From this set of molecules, we carried out a chemical space analysis (McDonagh et al., 2022). This analysis applied tools from chemical informatics, such as sub-structure searching with RDKit (Rational Discovery Kit, Landrum, 2014) and topological data analysis (Wasserman, 2018; Zhou et al., 2021), to identify similarities and differences in the molecular graphs of carbon capture amines compared to a set of 20,938 commercially available amines from the ZINC database (Irwin et al., 2012). From this analysis we identified 72 structural features, which we codified as a binary chemical “CCS” fingerprint related to the capability of a molecule to act as a carbon capture solvent. The code to generate these fingerprints can be accessed at github ⁵ . Datasets references in this paper are listed in Table 1.

Noting that information on carbon capture amines was sparse in the literature, we generated a new dataset from a single experimental source. We applied computational methods to select molecules that make up this data set and predict molecular performances for carbon capture. The data set is 130 molecules in total. We have provided these data to the community with the molecules represented by IUPAC (International Union of Pure and Applied Chemistry) name, InChI (International Chemical Identifier), InChIKey (a 27 character digital representation of the InChI) and SMILES (Simplified Molecular-Input Line-Entry System), making the set amenable to chemical informatics modelling. This dataset can be found at Zenodo ⁶ .

We further utilized our CCS fingerprint as a representation for Machine Learning (ML) models. To do this, we separated our data into binary classes and trained models to predict whether the molecules are promising or not. The concept here is to provide a high throughput virtual screen that enables prioritization of molecules for experimental tests. We trialed a range of ML models that are computationally inexpensive to train and are in widely available Python packages such as scikit-learn and xgboost. Promising performance was found with the Extra Trees Classifier (Geurts et al., 2006), the Ada Boost Classifier (Freund and Schapire, 1997; Hastie et al., 2009), Logistic regression (Yu et al., 2011; Defazio et al., 2014; Schmidt et al., 2017) and the Gaussian Process Classifier (Williams and Rasmussen, 2006). We judged the ability of these classifiers using a range of metrics including overall accuracy and the accuracy of classification of the positive and negative classes separately. Our initial work compares the performance of the CCS fingerprint to other commonly applied fingerprints such as MACCS keys (Molecular ACCess System; Durant et al., 2002) and 2D chemical descriptors from the Mordred engine (Moriwaki et al., 2018). Our most recent work outlines a combined computational and experimental discovery cycle that identifies several new candidate molecules (McDonagh et al., 2023; private comm). The development of these methods is shown in McDonagh et al. (2022).

Nano-porous solids are another class of materials that capture CO₂, mostly by physisorption (Oschatz and Antonietti, 2018; Khandaker et al., 2020) but also by chemisorption if the surfaces are coated with reactants like amines (Emerson et al., 2018). Nano-porous solids include zeolites, metal-organic frameworks (MOF), covalent-organic frameworks (COF), zeolitic imidazolate frameworks (ZIF), and porous polymer networks (PPN), among other things. There are many varieties of these materials, both hypothetical and real, amounting to 10⁵–10⁶ of them in databases (Moghadam et al., 2017; Boyd et al., 2019) that allow machine learning tools for screening (Wilmer et al., 2012a). The MDLab results reported here use three materials from the database CoRE MOF 2014 (Computation-Ready, Experimental Metal–Organic Frameworks; Chung et al., 2014), and place these data, the screening workflows, and a Jupyter Lab interface in containers on a cloud-based computer cluster (Neumann et al., 2022) to facilitate user interaction and workload management. A more extensive study than what is reported here is in Oliveira et al. (2023).

The screening workflow starts by querying the database for a material or class of materials to get the corresponding Crystallographic Information File (CIF) files. These files are input to applications like Zeo++ (Willems et al., 2012) and RASPA (Dubbeldam et al., 2016), which return geometric and molecular adsorption figures-of-merit. From the material structure, several geometric properties are calculated, such as accessible and non-accessible surface area, volume, and volume fraction, along with density, and the diameters of the largest free sphere, the largest included sphere, and the largest included sphere along a free path. The code used to generate the geometric figures-of-merit can be accessed at GitHub ⁷ . Figure 1 illustrates these diameters.

The CIF file for the unit cell is also used to calculate partial atomic charges in the method of EQeq (Wilmer et al., 2012b), which is saved as a new CIF file for input to RASPA to calculate an energy grid for CO₂, N₂ and H₂O molecules (Luan et al., 2022). With this energy grid, several Grand Canonical Monte Carlo (GCMC) simulations are run for 10,000 cycles, one for each temperature and pressure under consideration. The result, added to the database, is molecular uptake capacity as a function of pressure at each temperature, i.e., the standard isotherm for adsorption. The code used to generate the adsorption figures-of-merit can be accessed at GitHub ⁸ . A schematic of the GCMC method is shown in Figure 2.

Adsorption properties were determined from a time-dependent, one-dimensional dynamic model (see Eq. 1) in which multi-component gas enters one end of an adsorption bed and exits the other end. The bed model was run with solid adsorption capacity (q _i ) and gas concentration (c _g,i ) for component i as a function of time (t) and space (x,z) using the molecular-level adsorption isotherms as input to model the uptake as a function of temperature. The axial dispersion coefficient is represented by D _ax , the gas interstitial velocity by u _g , the density of the solid by ρ_s and the bed void fraction by ε. The model assumed local equilibrium between gas and solid phases from the GCMC results, and at the end of the absorption bed recorded the CO₂ purity and recovery fraction, the specific energy consumption in units of MJ per kg CO₂, and the volumetric productivity, in units of kg CO₂ per day per Liter of sorbent. The result was a rank ordering of nano-porous solids for CO₂ separation. Water was assumed to be an equal contaminant for all the sorbents and to not affect the rank order. Thus, water was not included in the input stream, which was 4.1% CO₂ and the rest N₂, but is assumed to be adsorbed with the same affinity as CO₂. Although this simplification neglects the relative water affinity between materials, it does capture the competitive nature of CO₂ and H₂O adsorption, which affects the final CO₂ capacity metrics. See Neumann et al. (2022) for details. ∂ c g , i ∂ t = D a x ∂ 2 c g , i ∂ x 2 − ∂ c g , i u g ∂ z − 1 − ϵ ρ s ϵ ∂ q i ∂ t (1)

Liquid electrolytes in modern energy storage devices typically contain one or more organic solvents, one or more inorganic salts, and may also include supplementary additives (Jie et al., 2020). Discovery of electrolyte constituents for new battery technologies using advanced computational techniques has been a key research topic. Accelerated electrolyte discovery workflow by JCESR (Joint Center for Energy Storage Research; Qu et al., 2015; Cheng et al., 2015) directs the selection of electrolyte constituents by a high-throughput screening process that focusses on a set of essential domain properties such as stability in an electrochemical window, solubility of salts, and ionic conductivities (Cheng et al., 2015). However, the subsequent development of optimum electrolyte compositions usually requires extensive experimental follow-up because the problem has a large chemical space (Benayad et al., 2022). Our electrolyte discovery framework extends beyond the computational funneling of electrolyte molecules and incorporates advanced deep learning models to optimize the composition of new electrolytes for improved battery performance. Our simulation-AI experiment synergistic framework for battery electrolyte discovery emphasizes the three aspects shown in Figure 3: (a) identification of component molecules, (b) optimization of electrolyte formulations, and (c) in silico characterization of the solid electrolyte interface.

The electrolyte discovery toolkit stack consists of AI and simulation modules that facilitate constituent-to-performance and end-to-end application of the battery electrolyte discovery framework. IBM’s AI enriched simulation workflow (Vassiliadis et al., 2022), which enables high throughput computation of molecular properties, allows screening of electrolyte constituents with desirable properties for targeted battery chemistries. In subsequent discovery steps, we utilize a formulation-based deep learning prediction model to shortlist electrolyte compositions made of candidates that predictively demonstrate target performance (Sharma et al., 2023). These models are infused with domain physical-chemical knowledge and are trained on a small set of data (representing electrolyte formulations and their battery performances – about 100 data points). Here we demonstrate the use-case of the AI models for discovering new electrolyte solvent candidates for the oxygen-assisted lithium Iodine (OALI) battery we have recently developed (Giammona et al., 2023).

Grand Canonical Monte Carlo simulations were used to calculate isotherms for adsorption of CO₂, N₂, and H₂O in three nano-porous materials, HKUST-1, Mg-MOF74, and ZIF-8.

The isotherms were calculated for single component adsorptions and fit to multi-component Langmuir models, as in Eqs 2, 3. Here, ΔH _i,k and ΔS _i,k are enthalpy and entropy of adsorption for species i on site k, q _sat,k is the saturation capacity of site k, independent of species, p _i is the partial pressure of species i, T is the temperature and R is the gas constant. The fits were used to determine the enthalpy of adsorption, which was then used in the dynamic adsorption model to determine optimum process cycle times and the corresponding purity and recovery fractions of CO₂ at the output, along with the energy requirements and volumetric rates. q e q , i = ∑ k = 1 3 q s a t , k b i , k p i 1 + b i , k p i (2) b i , k = e − ∆ H i , k − T ∆ S i , k / R T (3)

The results for Mg-MOF-74 are shown in Figure 7. The molecular structure is on the left, the GCMC isotherms with multi-component Langmuir fits are in the middle (the fit is better for CO₂ and N₂ than for H₂O, not shown). On the right is the calculated purity versus recovery fraction for CO₂.

The process model shows residual H₂O at the end of the desorption cycle for Mg-MOF-74 because of the high affinity for both CO₂ and H₂O. Nevertheless, the CO₂ adsorbed and desorbed for Mg-MOF-74 is much higher than for the others, amounting to about 1.19 kg CO₂ per day per liter of sorbent at 95% purity for Mg-MOF-74 compared to 0.41 and 0.21 kg CO₂ per day per liter for HKUST-1 and ZIF-8, respectively. The specific energy is also much less for Mg-MOF-74: 11.7 MJ/kg compared to 79.1 MJ/kg and 196.4 MJ/kg for HKUST-1 and ZIF-8, again at 95% purity. Figure 7 also shows higher recovery fractions for Mg-MOF-74. These trends are in agreement with literature expectations for dry CO₂/N₂ separation (Liu et al., 2012).

Next-generation batteries based on lithium (Li) metal anodes typically use ether-based electrolytes due to their outstanding compatibility with Li metal. Our OALI battery uses an electrolyte based on 1,3-dioxalane (DOL) and 1,2-dimethoxyethane (DME) mixed solvent system that contains two or more inorganic lithium salts such as lithium bis (trifluoromethyl) sulfonylimide (LiTFSI) and lithium nitrate (LiNO3) (Giammona et al., 2023). While the OALI battery utilizing this DOL/DME-based electrolyte has shown excellent rate capability (<50 mA/cm2) and areal capacity (>12 mAh/cm2), there is room for further enhancing cyclic stability, especially at higher active cathode loading (>60 w/w%). To achieve this improvement, a superior electrolyte design is a must.

For discovering a high-performance electrolyte, we begin by screening solvent candidates based on key properties such as low molecular weight (molecules containing <20 non-H atoms), commercial availability, toxicity, thermal properties, salt solubilities and electrochemical stability toward Li metal. Low molecular weight solvents would form a favorable solvation structure with Li ions, facilitating efficient ionic transport. Special emphasis is given to hazard labels and thermal properties of solvents to avoid any “regrettable” substitution in terms of sustainability. For thermal properties such as boiling point (b.p.), flash point (f.p.), and melting point (m.p.), a three tier criteria is defined as described in Figure 8A to make an informed decision for solvent selection.

United States Environmental Protection Agency (EPA) has established a Program for Assisting the Replacement of Industrial Solvent (PARIS-III) tool and a database which contains approximately 4,000 commercially available solvents that have low environmental impact (Harten et al., 2014). PARIS-III database solvents were extracted, followed by screening using data mining and simulation toolkits. It is important to highlight here that criteria priority in the down selecting process can impact the resulting outcomes. Here, we demonstrate the advantage of using simulation and AI synergistic workflows to discover new mixed materials such as electrolytes for batteries by following two screening Routes (Figure 9): Route-1 demonstrates application of simulation toolkits for quick screening of molecular candidates with target properties; and Route-2 shows that machine learning models trained with battery electrolyte datasets can better assist in selecting molecules for formulation by considering complex interactions between all its constituent components rather than their individual behavior.

In Route-1 screening, we first screened 2,871 solvents based on their molecular weight (<20 non-H atoms) from the PARIS-III dataset. Extracting thermal properties like boiling point (b.p.), flash point (f.p.), and melting point (m.p.) from PUBCHEM using API, we further down selected solvents to 107 count in Tier-3 criteria (b.p. > 50°C; f.p. > 30°C; m.p. < 0°C). Next, we deployed virtual experiments based on Density Functional Theory (DFT) to calculate energy levels (HOMO-LUMO levels) of shortlisted solvents and their solvation free energy for primary salts in electrolyte (LiNO3, LiI and LiBOB) using GAMESS software with functional B3LYP and basis set 6–311 G (d,p). For electrochemical stability toward Li metal, solvent LUMO level should be higher than Li/Li + potential. LUMO of DOL and DME are only slightly above Li/Li + potential and are therefore used as comparative standard for current simulation dataset. Solubility trends from the solvation simulations were validated experimentally using a high throughput solubility assessment setup as shown in Figure 8B. Passing through these screening steps, the potential electrolyte solvent candidates were further narrowed down to 4 solvents: tetraglyme, diglyme, 1,4-dioxane and tetrahydrofuran. LiNO3 solubility was the major limiting factor in the down selection process. For experimental validation, all 4 solvents were tested as solvents and co-solvents for OALI electrolytes in coin cells. Among them, tetraglyme (TG) mixed with DOL demonstrated the best cycling stability at 60 wt% active cathode loading, however, the specific capacity decreased by 10% in comparison with the previously reported value (Giammona et al., 2023).

We next developed a deep learning model to map molecular structures, compositions of constituent molecules, to battery performance such as specific capacity and Coulombic efficiency (Sharma et al., 2023). As mentioned in the experimental section, the model was trained with the dataset of OALI battery’s electrolyte formulations and their corresponding battery performance, which was generated in our lab. The trained model was used to screen solvent candidates in Route-2 based on the specific capacity prediction of electrolyte formulations. As a result, 32 solvents emerged as potential high performing electrolyte solvents from PARIS-III dataset. The solvents were now further shortlisted based on electrochemical stability over Li metal and thermal properties. Upon further testing in the lab, the results showed that adding 3-Methoxybutyraldehyde dimethyl acetal (MBDA) as a co-solvent to the DOL:TG system enhances the specific capacity to the highest value (ca. 148 mAh/g) that was observed by DOE:DME system, while demonstrating even longer cycling stability (>250 cycles). Meanwhile, this new electrolyte system (DOL:TG:MBDA) exhibits better safely quotient than DOL:DME system since both TG and MBDA have higher flash points than DME (f.p. = 136°C, 46°C and 21°C, respectively).