New carbon nanostructures—fullerenes, nanotubes and, above all, graphene—have aroused great interest in recent years, making the field of graphene research undergo phenomenal growth that comes to light in publications.

Graphene is a one-atom-thick layer of carbon with exceptional physical and chemical properties and a potential for revolutionary applications in a wide variety of fields: strong lightweight materials, new-generation electronic units, specialized coatings, novel biomaterials and sensors, and innovative medical applications (Novoselov et al., 2012). Back in 1947 (Wallace, 1947), planar graphene was theoretically presumed not to exist in a free, stable state (Odegard et al., 2002). Within the framework of carbon-based materials, however, this 2-D material suddenly aroused the attention of scientists (Lv et al., 2011). With the experimental discovery of graphene by Novoselov et al. (2004), a rapidly growing stream of scientific literature became fertile land for bibliometric studies: Barth and Marx (2008), Wan and Pan (2010), Shapira et al. (2010), Lv et al. (2011), Munoz-Sandoval (2014), Small et al. (2014), Terekhov (2015), Klincewicz (2016), and Shapira et al. (2016).

As they say, a picture is worth a thousand words, and the notion behind information visualization is to create a mental picture of that which is invisible to our eyes. This idea arises from the conviction that an image facilitates comprehension, and invites one toward analysis better than text, numbers, or a combination thereof. Information visualization emerged at the beginning of this century as a discipline of great interest at the crossroads of bibliometrics and scientometrics, providing multiple visual representations known as scientograms (Moya-Anegón et al., 2007a). These maps of science can facilitate understanding of a scientific domain by depicting the structure of research output. At large, scientograms show relationships that occur between disciplines; and they may be used more specifically as an index to comprehend to what extent certain research lines or fields are connected with others (Vargas-Quesada and Moya-Anegón, 2007). They enable one to explore the sequential evolution of research, identify research fronts, detect emerging or decadent topics, and locate areas of interdisciplinary efforts (Gómez-Núñez et al., 2016).

Still, researchers persistently search for other underlying mechanisms that may explain certain changes and patterns within scientific networks (Chen, 2004). Bibliometric mapping tools and scientometric analyses have become increasingly sophisticated in order to reach more ambitious goals. In this work, we use automatic bibliometric mapping tools to show the intellectual structure of a relatively novel field, graphene research, joining the current debate about distinct counting approaches (full/fractional), and tracing its evolution to date.

Co-word analysis (Callon et al., 1983, 1991; Leydesdorff, 1989) has been described—and later corroborated (Leydesdorff and Nerghes, 2017)— as the best approach to determine and set out the intellectual structure of a field at the level of research specialties (Braam et al., 1991a,b). It is furthermore able to bring to light new developments of a research line over time (Peters and van Raan, 1993a,b) and help policy makers understand the complex interrelations of science and their implications for effective research planning (He, 1999). Although co-word analysis has gradually been improved upon, while not all of its limitations have been overcome (Wang et al., 2012), this type of analysis can reveal the intellectual underpinnings of a wide variety of domains: chemistry (Callon et al., 1991), data retrieval (Ding et al., 2001), the fuzzy logic field (López-Herrera et al., 2010), women and health (Zulueta et al., 2011), keyword analysis (Cantos-Mateos et al., 2012), human intelligence networks (Wang et al., 2012), and stem cells (Cantos-Mateos et al., 2014). A recent study applied it to study the intellectual structure of nanotechnology and its evolution (Muñoz-Écija et al., 2017).

Information visualization has a very active role in the distribution and depiction of the intellectual structures related with scientific research lines (Chen et al., 2001). The use of scientograms to look onto/into science is not entirely new, offering a new standpoint to reveal the scientific frontiers and dynamic intellectual structure of a research line (Alcaide-Muñoz et al., in press). The pioneer efforts in this terrain go back to the 1960s, in fact. Science maps have been used to navigate around scholarly literature and the depiction of its internal relations (Garfield, 1986), to represent the spatial distribution of research areas and their relations (Small and Garfield, 1985) and to analyze and visualize the social or intellectual spry of scientific research fronts (Braam et al., 1991a,b; Noyons et al., 1999; Börner et al., 2003).

At the end of the 20th century, the lower price of computer equipment, along with an increased capacity of processing and storage components, not to mention big data and new algorithms, set the stage for a grand era in the first decade of this millennium, with the appearance of a wide variety of free software for science mapping analysis (Cobo et al., 2011a). Among many other achievements, science mapping become the keystone of visualizing and analyzing computer graphics (Chen et al., 2001), using ISI categories to represent science (Moya-Anegón et al., 2004), mapping the backbone of science (Boyack et al., 2005), evaluating large maps of disciplines (Klavans and Boyack, 2006), visualizing the citation impact of scientific journals (Leydesdorff, 2007a), mapping interdisciplinarity (Leydesdorff, 2007b), viewing the marrow of science (Moya-Anegón et al., 2007b), creating dynamic animations of journal maps (Leydesdorff and Schank, 2008), mapping the structure and evolution of chemistry research (Boyack et al., 2009), proposing a consensus map of science (Klavans and Boyack, 2009), creating a journal map using Scopus data (Leydesdorff et al., 2010), mapping the geography of science (Leydesdorff and Persson, 2010), clustering over two million biomedical publications (Boyack et al., 2011), creating more accurate document-level maps of research fields (Klavans and Boyack, 2011), detecting and visualizing the evolution of the fuzzy sets theory field (Cobo et al., 2011b), proposing a new global science map (Leydesdorff et al., 2013a,b; Boyack and Klavans, 2014), analyzing the investigation in integrative and complementary medicine (Moral-Muñoz et al., 2014), analyzing intelligent transportation systems (Cobo et al., 2014), showing the evolution of bases knowledge systems (Cobo et al., 2015), showing the scientific evolution of social work (Martínez et al., 2015), outlining animal science research (Rodriguez-Ledesma et al., 2015), studying the conceptual evolution of marketing research (Murgado-Armenteros et al., 2015) identifying and depicting the intellectual structure and research fronts in nanoscience and nanotechnology in the world (Muñoz-Écija et al., 2017), and exploring the scientific evolution of e-Government (Alcaide-Muñoz et al., in press), among other brave new initiatives.

The proposal by Leydesdorff and Rafols (2009) of creating overlay maps can be seen as a powerful contribution integrating visualization, intellectual structure, evolution, and benchmarking, for any kind of scientific domain. Indeed, overlay maps soon proved useful for research policy and library management (Rafols et al., 2010), to build interactive overlays (Leydesdorff and Rafols, 2012), to map patent data (Leydesdorff and Bornman, 2012), to gauge interdisciplinarity (Leydesdorff et al., 2013a,b), and to appraise strategic intelligence in emerging technologies (Rotolo et al., 2017).

In the wake of the above study embracing N&N, we chose to focus our attention on the application of information visualization techniques to the monolayer of sp²-bonded carbon atoms known as graphene, a topic attracting worldwide interest not only because of its unique 2-D structure, fascinating properties, and wide range of potential applications (Chen, 2013), but also due to its intriguing intellectual structure and evolution, now ripe for comparison.

This study is intended to serve as a focal point, with the main goal of testing a methodology that reveals at a glance the main research lines involving graphene, based on scientific information. The questions that became our guidelines for analysis were:

How vast and varied is graphene research output?

Perhaps the greatest novelty of this contribution lies in the use of overlay maps and fractional counting, which together allow us to map the disciplinary network structure of a research field in a way that helps identify terms which have a mediating effect, linking different developmental stages in diverse geographic domains. Our main aim is to offer the research community a methodology and tools serving to visualize the intellectual structure of any scientific domain by means of a base map, the main research lines, and the possibility of benchmarking these scientific domains with others, or show their evolution using overlay maps. These tools and this methodology can be extrapolated to any area, discipline, or research field. Given the acute and rising importance of graphene research, it appeared to be a topic of great interest to validate our approach.

The data set was obtained on June 5, 2017, from Scopus database.

Traditionally, the most commonly used databases for bibliometric studies have been those of the Web of Science (WoS). However, the appearance on the market in 2004 of the Scopus of Elsevier, also having a multidisciplinary scope and a greater coverage of journals from diverse geographic regions (Moya-Anegón et al., 2007a) made an alternative source of great value for studies of the scientific output of countries, institutions, and disciplines. Even with the unprecedented inclusion of journals in both databases as a part of expansion in their coverage (Leta, 2011; Collazo-Reyes, 2014), Scopus offers a stronger coverage of journals not exclusively written in English (22,800 as opposed to the 18,000 of WoS), which provides for a more representative geographical picture of Graphene research. Granted, WoS has strong coverage which goes back to 1990 (Chadegani et al., 2013), but this was not especially relevant for our objectives.

Moreover, bearing in mind the methodology behind this work, currently WoS is more time-consuming than Scopus due to the fact that it is no longer possible (since about a year ago) to do searches in WoS with the field Keyword (Author Keywords and Index Keywords), one of the premises emphasized here, as a necessity to properly delimit the object of study. At present and in order to execute a similar delimitation, ISIWoS “only” offers the option of doing searches with the field “Topic,” which includes searching the fields: Title, Abstract, Author Keywords, and Index Keywords, yet with considerable documental noises in the results.

The search strategy used to retrieve all the documents published worldwide on graphene was very simple: [KEY (“single layer graphit*”) OR KEY (graphene)] AND [EXCLUDE (PUBYEAR, 2016) OR EXCLUDE (PUBYEAR, 2017)].

To validate the proposed methodology of overlay maps, it is necessary that the data used to generate the intellectual structure (basemap) be as comprehensive as possible (use of Scopus) and that they be complete so that the structure will remain invariable over time. This criterion allows for the map to be used as a reference or object of comparison in future work by other researchers, whether it be to analyze further geographic domains not addressed in this study, or to analyze the implications of just how the input (or output) of journals in the Scopus database is effected and its updating policy in the representation of the different domains and research lines.

After downloading and analyzing the evolution over time of the number of documents (Figure 1), we observed a decreased growth rate in the two final years of study, as seen in the table below. Deducing that the data were incomplete, we decided to exclude the years 2016 and 2017 from this study. This table, along with the data for the top 20 graphene-producing countries, can be found below.

Keywords have as their main objective to provide rapid access to scientific works (Soos et al., 2013). Author Keywords provide the “author aboutness,” that is, contents expressed through terms in natural language, while Indexed Keywords give the interpretation of contents (Stock and Stock, 2013). The usefulness of this approach was confirmed by Zhang et al. (2016), who hold that Author Keywords are highly effective in terms of bibliometric analysis when investigating the knowledge structure of scientific fields, and Indexed Keywords are very comprehensive for representing an article’s content. Other authors argue that keyword-based analyses might be biased, suggesting some scientists could use certain keywords to seek increased visibility (Bonaccorsi, 2008). In our view, the combination of Author Keywords and Indexed Keywords, eliminating duplications, is the optimal approach, combining the robustness of both realms.

The expression “single layer graphite” has come to largely replace the word graphene (Lv et al., 2011), for which reason we decided to include it in our search equation. It was right-truncated to include the singular and the plural. In our case, this inclusion led us to obtain just one more document, however. It should be stressed that such a truncation can be very problematic, as graphen* leads to the retrieval of 386 documents containing the German root “graphen”: for instance “graphs” in English, which does not pertain to the context of study. Also excluded were the data from 2016 to 2017, as we wished to be certain that the final year of analysis, 2015, was fully registered in the database. No restriction has been made based on the document type, language of publication, country, or other such element. The search equation for the rest of the geographic domains of study was the same as applied for the worldwide level, only delimited by the geographic affiliation of the authors. Hence, for the United States [KEY (“single layer graphit*”) OR KEY (graphene)] AND [EXCLUDE (PUBYEAR, 2016) OR EXCLUDE (PUBYEAR, 2017)] AND [LIMIT-TO (AFFILCOUNTRY, “United States”)]. For China: [KEY (“single layer graphit*”) OR KEY (graphene)] AND [EXCLUDE (PUBYEAR, 2016) OR EXCLUDE (PUBYEAR, 2017)] AND [LIMIT-TO (AFFILCOUNTRY, “China”)]. Finally, the European Union (EU) was delimited by country, for each of the 28 member states: [KEY (“single layer graphit*”) OR KEY (graphene)] AND [AFFILCOUNTRY (austria) OR AFFILCOUNTRY (belgium) OR AFFILCOUNTRY (bulgaria) OR AFFILCOUNTRY (croatia) OR AFFILCOUNTRY (cyprus) OR AFFILCOUNTRY (czech AND republic) OR AFFILCOUNTRY (denmark) OR AFFILCOUNTRY (estonia) OR AFFILCOUNTRY (finland) OR AFFILCOUNTRY (france) OR AFFILCOUNTRY (germany) OR AFFILCOUNTRY (greece) OR AFFILCOUNTRY (hungary) OR AFFILCOUNTRY (ireland) OR AFFILCOUNTRY (italy) OR AFFILCOUNTRY (latvia) OR AFFILCOUNTRY (lithuania) OR AFFILCOUNTRY (luxembourg) OR AFFILCOUNTRY (malta) OR AFFILCOUNTRY (netherlands) OR AFFILCOUNTRY (poland) OR AFFILCOUNTRY (portugal) OR AFFILCOUNTRY (romania) OR AFFILCOUNTRY (slovakia) OR AFFILCOUNTRY (slovenia) OR AFFILCOUNTRY (spain) OR AFFILCOUNTRY (sweden) OR AFFILCOUNTRY (united AND kingdom)] AND [EXCLUDE (PUBYEAR, 2016) OR EXCLUDE (PUBYEAR, 2017)].

Among the great variety of freely available computer programs to perform a science mapping analysis, we looked into CiteSpace (Chen, 2004, 2006), SciMAT (Cobo et al., 2012), and VOSviewer (Van Eck and Waltman, 2010). The pros of the first two are the facility of making evolutive or longitudinal studies, the support of smart techniques to make it easier to analyze a research line by automatically categorizing its research outputs/results into different subjects, and to highlight the clusters with high impact through citation counting. One particular advantage lies in the scientific support tools to uncover and label the conceptual structure of a research line of interest. The pros of the third option would be: the quality of the displays, the ease of working directly with Scopus data (one of our research aims) without any loss of information, together with the possibility of generating the overlay maps manually (another research objective), and the fact that they could be processed and depicted easily with the software. The proximity with our research objectives led us to select the latter software alternative. It is a software tool for building and depicting networks based on bibliometric data. It features a text mining instrument that can be used to depict co-occurrence networks of terms extracted from any part of scientific literature. With our focus on the knowledge structure and main lines of graphene research, we selected Author Keywords in conjunction with Indexed Keywords as the unit of analysis; their co-occurrence was, as we mentioned before, the unit of measurement. The most popular methodologies for estimating co-occurrence are the full counting (whole counting for others) and the fractional counting (Aksnes et al., 2012). In the case of the full counting method, for instance, when one keyword co-occurs with another five within a single document, it would be assigned a full weight of one. Under the fractional counting method, the co-occurrence is assigned to each keyword with a fractional weight of 1/6. Thus, fractional counting in the latter case divides the credit among co-authors (countries, institutions, etc.), whereas fractional counting at the network level can normalize the relative weights of links and thereby clarify the structures in the network.

One of the co-authors of this study is a graphene technologist, who helped us confirm the relevance of terms used. After analyzing and carefully studying the keywords in each cluster, detected by full counting and by fractional counting, we concluded that fractional counting gave more realistic results. That is, it better reflected what is actually the subject of study, not just the field covered by the Journal at hand, as would be the case with full counting. A similar conclusion was drawn by Perianes-Rodriguez et al. (2016) regarding other units of analysis.

For the generation of the basemap, where the intellectual structure of the world of graphene would be depicted, we used a threshold of co-occurrence of over three. According to previous observations, a lower threshold produces distortion and informational noise, while a higher threshold gives rise to a loss of information. We think it is best not alter the idea of keywords as settled and picked by the authors or the automatic indexing system. Yet as it is possible for two different keywords to be used to define the same concept, their normalization is absolutely necessary. Accordingly, duplicates are eliminated, plural and singular forms are standardized to the form showing a higher figure for occurrence in the database, empty words or words out of context are deleted, abbreviations or acronyms are replaced by the complete term whenever possible, followed by the acronym or abbreviation between parenthesis, etc. In normalizing the terms, an ad hoc thesaurus containing 25.603 equivalences was created, which is available at: http://www.ugr.es/local/benjamin/frontiers/thesauro.txt.

To produce the scientograms, parameters such as layout attraction, layout repulsion, and clustering resolution were set to the default values. Only the minimum cluster size parameter was changed, here to 100. The overlay maps were built following the same procedure, although—as their name suggests—they were superimposed upon the basemap, and the words that did not coincide were eliminated. In this way, the overlay map of every domain or period provides its knowledge structure in the form of its total keywords; the 22,000 plus keywords in the basemap function, therefore, serve as a filter. Each circle represents a keyword, and its size is proportionate to the number of times that it occurs in the documents represented. The distance between two particular circles would approximately indicate the level of co-occurrence between two terms—the closer they are, the tighter or stronger their conceptual relationship, etc. and vice versa. The colors represent the clusters that were identified automatically, by means of clustering techniques, based on the level of proximity among terms (Waltman et al., 2010).

For complementary analyses, we also generated density maps, as they prove very useful for more detailed or evolutive studies of the intellectual structure put forth. In these maps, keywords are indicated by their label. Each point in a scientogram has a color that depends on the density of keywords at that point. By default, the color is somewhere in between red and blue. The greater the number of keywords in the neighborhood of a point, and the higher the weights of the neighboring items, the more reddish the color of the point. Conversely, a smaller number of items around the point, with lower weights of the neighboring items, would mean that the color of the point is closer to blue (Van Eck and Waltman, 2016). Similarly, the intensity level in terms of density, from less to more, is depicted by: blue, green, yellow, and red.

All the maps produced in the framework of this study can be visualized in high resolution by clicking on the link indicated in each case. They may also be viewed using the online version of VOSviewer. To this end, we give the link that allows the reader to see the maps as well as the networks of each domain. Please note that for very broad domains and periods, e.g., the whole world, a computer with more than 8 GB of RAM is needed. If one has only 1 or 2 GB, for instance, the second part of the link referring to the network may simply be omitted. In other words, http://www.vosviewer.com/vosviewer.php?map=http://www.ugr.es/local/benjamin/frontiers/graphene-m.txt&label_size_variation=0.3&zoom_level=1&scale=0.9 will take us to the maps alone, without any problem owing to limited memory.

The results are divided below into six sections, according to the proposed research questions.

In this work, the intellectual structure of graphene research, its main research lines and its evolution are studied by means of bibliometric and information visualization techniques. Graphene shows a continually growing attraction worldwide, an interest that translates as scientific research and output. In view of the objectives set forth at the beginning of our work, we may affirm that:

World output in graphene research began to grow exponentially in 2004, as previously reported (Lv et al., 2011). This growth was spurned by a strong pulse in production in the United States from 2004 onward, and in Europe and China from 2006 onward. Wan and Pan (2010) state that China has placed great vigilance on graphene research, but is out of the way of the driver countries. Five years later, this study shows that China became the undisputed world leader in 2012, largely owing to governmental and financial support from the National Natural Science Foundation of China, the Ministry of Science and Technology of the People’s Republic of China, and the Chinese Academy of Sciences (Chen, 2013). A very similar trend was identified in the research field of Nanoscience and Nanotechnology (Chinchilla-Rodríguez et al., 2016). In 2013, it was appraised that 2,200 patents related to graphene were rooted in China (Gao et al., 2014). Europe overtook the United States in graphene output in the year 2014. It is widely recognized that European Research Performing Organizations contribute very significantly to the overall research objectives associated with graphene, with research and innovation activities that are not directly funded through Collaborative Projects and Coordination and Support Actions (Ciubotaru and Helman, 2015). For example, the Graphene Flagship represents a new form of joint research effort on an unprecedented scale—in fact, Europe’s largest research initiative to date—with a budget of one billion euros.

Although this analysis is founded on currently available tools, a main novelty resides in mapping the disciplinary network structure of the field “graphene research,” identifying terms that have a mediating effect in different developmental stages as well as trends in geographical domains. Likewise, the findings can be shared and commented upon among colleagues who see in the online maps what is going on in the research “hot spots,” year by year, or in any time period specified.

Despite considerable efforts to represent disciplines and their evolution over time (Klavans and Boyack, 2006), unresolved issues persist on the research platform of bibliometric indicator studies (Leydesdorff and Rafols, 2009). In any case, the findings expounded and illustrated here could help refine research agendas by providing policy makers a better view of the dynamic directions of research. An immediate goal could be to improve the baselines for benchmarking exercises, especially those intended to evaluate research endeavors.

It is likely that excessive optimism participated in the definition of our objectives in the framework of this study. In the face of the results obtained, the vast dimensions of the factors involved gradually became more apparent. This work served to plant our feet more firmly on the ground amid the numerous methodological pros and cons and the magnitude of the material of study itself.

One consideration that comes up repeatedly in studies of this nature is the database used. Although we relied on Scopus in this case, it would be desirable for future efforts to add/complement data from sources such as WoS, patent databases (USPTO, SP@CENET, PATENTSCOPE, DWPI, and WIPO Gold), and specialized databases such as Chemical Abstracts Services, along with Web Content Mining techniques (Shapira et al., 2016). It would also be very interesting to replicate this work using other tools such as CiteSpace (Chen, 2004) or SciMAT (Cobo et al., 2012), to see and compare the different perspectives that these tools can bring into view.

Co-word analysis has been improved in the past two decades but still suffers from certain limitations, e.g., the normalization of keywords, indexes applied, counting methods, and the participation of experts (He, 1999; Wang et al., 2012). Here we tried to overcome some of these limitations by normalizing keywords, building a thesaurus, applying fractional counting, and integrating an expert’s knowledge in the whole process. Fractional counting proved to be beneficial, possibly because at the network level one can normalize the relative weights of links and thereby clarify the structures within the network (Leydesdorff and Park, 2017), offering highly uniform clusters in size, which reliably reflect what is actually researched, not just the subject matter of the journals where output is presented. At any rate, further analysis should be undertaken to explore alternative techniques so as to refine both methodology and results. For example, overlay maps leave room for improvement, and should be more dynamic. They have inherited the positional structure of the basemap, but their intellectual and internal structure (clusters) are determined by the number of keywords and the associated links, not the basemap.

Knowledge is never static. Its sources and patterns of growth and interaction are constantly shifting and fusing, and output may peak or disappear entirely from view over time. Thus, incorporating new data every year into the basemap in combination with of overlay maps with a series of temporal windows, preferably year by year, or in overlapping periods, would shed more light on the evolution of any type of scientific domain. The easiest way to grasp the structure and evolution of knowledge worldwide is with the help of intuition and visual aids, but adding an element of animation could make bibliometric analyses much more attractive overall.

BV-Q provided substantial contributions to the conception, design of the work, acquisition, analysis, and interpretation of data for the work. ZC-R worked drafting the work and revising it critically for important intellectual content. NR worked drafting the work and revising it critically for important intellectual content.

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. The reviewer, EY, and handling Editor declared their shared affiliation, and the handling Editor states that the process nevertheless met the standards of a fair and objective review.