Materials discovery by screening data sets

A common type of study that uses informatics is the screening of databases to find existing materials that suit a new application. Screening databases is hardly a new concept in materials engineering,^{Reference Ashby8} but the first examples of creation of “large” data sets (with hundreds to thousands of entries) with high-throughput (HT) ab initio computations appeared only in the early 2000s.^{Reference Smithson, Marianetti, Morgan, Van der Ven, Predith and Ceder9–Reference Madsen12} The first applications of these HT searches were to identify new stable materials among a small range of chemistry and structure types or to predict crystal structures via composition. Greeley et al.^{Reference Greeley, Jaramillo, Bonde, Chorkendorff and Nørskov10} identified BiPt as a promising electrocatalyst following a screening strategy based on electrochemical stability and activity of a series of binary alloys in one of the first examples of an HT search. With the upsurge in computing power and collaborative development of robust automation codes supporting first-principles calculations, such HT studies paved the way for the creation of several large high-fidelity databases of material properties, such as the Materials Project (MP),^{Reference Jain, Ong, Hautier, Chen, Richards, Dacek, Cholia, Gunter, Skinner, Ceder and Persson13} Open Quantum Materials Database (OQMD),^{Reference Saal, Kirklin, Aykol, Meredig and Wolverton14,Reference Kirklin, Saal, Meredig, Thompson, Doak, Aykol, Rühl and Wolverton15} and Automatic Flow for Materials Discovery Library (AFLOWLIB).^{Reference Curtarolo, Setyawan, Wang, Xue, Yang, Taylor, Nelson, Hart, Sanvito, Buongiorno-Nardelli, Mingo and Levy16}

With the introduction of these databases to the community, the number, scale, and complexity of material screening studies increased significantly. In a typical screening study, domain experts first identify the material features that are (1) descriptive of the functionality requirements for the target application and (2) computable using the information exiting in the databases or within their supporting computational framework.^{Reference Curtarolo, Hart, Nardelli, Mingo, Sanvito and Levy6} Many materials are then screened through a funnel of “filters” to identify the most promising ones to explore with more advanced calculations (e.g., diffusion, transport) or via experiment. For example, researchers sifted through ∼130,000 materials in the OQMD to find 30 functional cathode coatings (e.g., barriers that neutralize acids from the electrolyte) for Li-ion batteries, using descriptors that included their calculated reactivities with acids in the electrolyte.^{Reference Aykol, Kim, Hegde, Snydacker, Lu, Hao, Kirklin, Morgan and Wolverton17} Given the openness of these databases, HT-DFT data have supported similar studies by researchers unaffiliated with the database creators.^{Reference Jain, Persson and Ceder18} These studies exemplify the acceleration in materials discovery enabled by HT-DFT databases, where the initial, costly screening phase of material selection is offloaded to computers, allowing the experiments to focus on the most viable materials.

Improved understanding of materials and methods

Some important trends are only visible in large data sets. Thus, another common type of informatics study is to extract design rules from large collections of materials data. Using techniques such as visualization and automated data mining algorithms, materials engineers have employed atomic-scale calculation data to further scientific knowledge and generate useful heuristics to employ during materials design.

HT computational databases also enable better understanding of the accuracy of their underlying methods. Kirklin et al. found that DFT errors in computed formation energies are at the level of 0.136 eV/atom, but can be brought down to 0.081 eV/atom with systematic corrections of energies of the reference state, as part of the OQMD effort.^{Reference Kirklin, Saal, Meredig, Thompson, Doak, Aykol, Rühl and Wolverton15} De Jong et al. compiled a database of elastic properties of inorganic materials as part of the MP, which helped them demonstrate that the bulk and shear moduli calculated using DFT are mostly within ±15% of measured values.^{Reference de Jong, Chen, Geerlings, Asta and Persson19} Systematic analysis of the performance of a method is only possible when large resources of consistent data are available, and these studies illustrate how such databases lead to better understanding of the limitations of DFT.

Large databases of materials properties also allow for better understanding of materials, especially when combined with methods for automatically learning important descriptors. Meredig et al. used polynomial regression coupled with clustering techniques to learn that different factors affect the stability of zirconia for different groups of elements in the periodic table.^{Reference Meredig and Wolverton20} Emery et al. used a database of DFT-computed stabilities of 5000 perovskites to learn that conventional rules for predicting perovskite stability are not sufficient.^{Reference Emery, Saal, Kirklin, Hegde and Wolverton21} Ghiringhelli et al. and Kong et al. have used algorithms that produce simple, human-interpretable models (e.g., compressed sensing, decision trees) to understand which crystal structure forms at different compositions.^{Reference Ghiringhelli, Vybiral, Levchenko, Draxl and Scheffler22,Reference Kong, Luo, Arapan, Villars, Iwata, Ahuja and Rajan23} Compressed sensing has also proven useful for understanding how structure effects the dielectric properties of polymers.^{Reference Mannodi-Kanakkithodi, Huan and Ramprasad24} In general, these studies demonstrate the potential scientific gains to be achieved by aggregating data.

Accelerating discovery with machine learning

A third common application of materials informatics is the use of machine learning to further accelerate materials design. In contrast to learning design rules, the emphasis in machine-learning-driven studies is less on advancing human understanding than building a predictive surrogate model. Consequently, many machine-learning studies use learning algorithms that create models that are exceptionally accurate, but too complex to be easily interpreted by humans. After careful validation, these models can perform materials screening at rates far faster than conventional computational tools.

Machine learning in atomic-scale modeling is often used to replace expensive ab initio methods with cheaper approximations. For example, machine-learning methods are often used to create force fields,^{Reference Behler25} which predict the energy of and forces acting on an atom given its environment and can be used as in-place alternatives for ab initio calculations in many applications. Several groups have demonstrated the utility of directly integrating the construction of machine learning force fields into a calculation method (e.g., molecular dynamics).^{Reference Li, Kermode and De Vita26,Reference Jørgensen, Larsen, Jacobsen and Hammer27} Machine-learning models are also commonly used to estimate the aggregate properties of a material, such as formation enthalpies,^{Reference Faber, Lindmaa, von Lilienfeld and Armiento28,Reference Ward, Liu, Krishna, Hegde, Agrawal, Choudhary and Wolverton29} given its atomic structure, at rates several orders of magnitude faster than the methods used to create their training set. These fast surrogates can be used to guide HT searches for materials, which make it possible to explore spaces too large for conventional methods.^{Reference Hautier, Fischer, Jain, Mueller and Ceder30–Reference Seko, Hayashi, Kashima and Tanaka33}

Machine-learning methods also expand the space of properties accessible to atomic-scale modeling. Seko et al. combined experimental data and DFT calculations to quickly predict melting temperatures of inorganic binary octet materials,^{Reference Seko, Maekawa, Tsuda and Tanaka34} which takes thousands of CPU hours to compute by conventional atomic-scale modeling methods.^{Reference Hong and van de Walle35} Hutchinson et al. have also shown how to accurately estimate bandgap energies—a property known to be difficult to accurately estimate with DFT—by combining the Materials Project and experimental data with transfer learning.^{Reference Hutchinson, Antono, Gibbons, Paradiso, Ling and Meredig36} Combining machine learning, atomic-scale modeling, and experimental data may soon become a common pattern for accessing material properties currently impossible to compute from the atomic scale. These studies, among many other examples,^{Reference Ward and Wolverton37} illustrate how machine learning is becoming an integral part of performing atomic-scale calculations.

Automating experiments

Exponential increase in computing power combined with open, community-developed software enabled the automation of computational workflows to produce data sets amenable to informatics. Physical experiments are, of course, significantly more challenging to automate than calculations. There is, however, substantial progress in high-throughput experimentation (HTE), which focuses on evaluating many materials in rapid succession and is uniquely suited to creating informatics-ready data—large resources of consistently measured and consistently processed samples. Beyond the technical limitations of automating certain types of experiments, adoption of HTE is slowed by limited access to HTE equipment, such as combinatorial synthesis machines (e.g., cosputtering chambers) or automated characterization equipment (e.g., robotic x-ray diffractometers). Consequently, collaborations that allow scientists to access a distributed ecosystem of HT synthesis, characterization, and data infrastructure are necessary^{Reference Green, Choi, Hattrick-Simpers, Joshi, Takeuchi, Barron, Campo, Chiang, Empedocles, Gregoire, Kusne, Martin, Mehta, Persson, Trautt, Van Duren and Zakutayev62} along with development of new HT experimental facilities.^{63,Reference Lu64} Beyond HTE, the development of autonomous materials experimentation techniques capable of designing, executing, and analyzing their own experiments could accelerate the development of materials and the creation of useful data sets even further.^{Reference Nikolaev, Hooper, Webber, Rao, Decker, Krein, Poleski, Barto and Maruyama65} Increased use of HT and autonomous methods will produce richer materials data, which could lead to the same proliferation of informatics techniques in many fields as it has in atomic-scale modeling.

Digitizing materials data

Another issue limiting the use of materials informatics is that the metadata for most materials data is never recorded in a digital form. Unlike atomic-scale modeling data, metadata for experiments is rarely contained within the data files themselves, but instead is often placed in separate, nondigital resources (e.g., lab notebooks). A suite of materials-science-centric electronic lab notebook software and Laboratory Information Management Systems (LIMSs) that assist in the describing, managing, and sharing of data have been developed to address this problem.^{Reference Dima, Bhaskarla, Becker, Brady, Campbell, Dessauw, Hanisch, Kattner, Kroenlein, Newrock, Peskin, Plante, Li, Rigodiat, Amaral, Trautt, Schmitt, Warren and Youssef66–Reference Nguyen, Konstanty, Nicholson, O’Brien, Schwartz-Duval, Spila, Nahrstedt, Campbell, Gupta, Chan, McHenry and Paquin68} Widespread use of these tools in materials science will reduce the amount of human effort required to make data useful for informatics.

Structuring materials data

Significant amounts of useful materials data are and will continue to be embedded in the text of research articles. Natural Language Processing (NLP) will play a key role in extracting and structuring the data to make them usable by materials informatics. Olivetti et al. recently aggregated synthesis parameters from more than 640,000 journal articles using NLP and demonstrated the use of these data in machine-learning-based synthesis planning.^{Reference Kim, Huang, Tomala, Matthews, Strubell, Saunders, McCallum and Olivetti69} Others have demonstrated automated extraction of small molecule and polymer properties, respectively.^{Reference Swain and Cole70,Reference Tchoua, Chard, Audus, Ward, Lequieu, De Pablo and Foster71} The continued development of science-focused NLP tools (e.g., ChemDataExtractor^{Reference Swain and Cole70}) and analogous systems for retrieving metadata from computer file systems^{Reference Beckman, Skluzacek, Chard and Foster72} offer routes for drastically increasing the amount of machine-accessible materials data for informatics.

Improving data accessibility

Beyond being digital and structured, data must be easily sharable and accessible to software to enable informatics. As outlined by the FAIR Principles,^{Reference Wilkinson, Dumontier, Aalbersberg, Appleton, Axton, Baak, Blomberg, Boiten, da Silva Santos, Bourne, Bouwman, Brookes, Clark, Crosas, Dillo, Dumon, Edmunds, Evelo, Finkers, Gonzalez-Beltran, Gray, Groth, Goble, Grethe, Heringa, ’t Hoen, Hooft, Kuhn, Kok, Kok, Lusher, Martone, Mons, Packer, Persson, Rocca-Serra, Roos, van Schaik, Sansone, Schultes, Sengstag, Slater, Strawn, Swertz, Thompson, van der Lei, van Mulligen, Velterop, Waagmeester, Wittenburg, Wolstencroft, Zhao and Mons73} research requires data to be findable, accessible, interoperable, and reusable. Several efforts in the materials community have emerged to simplify the process of helping data meet the FAIR principles. For example, the Materials Data Facility provides a route for publishing, describing, and indexing materials data of any form or size on distributed data storage.^{Reference Blaiszik, Chard, Pruyne, Ananthakrishnan, Tuecke and Foster74} Citrination^{Reference O’Mara, Meredig and Michel75} allows scientists to publish data in a flexible, well-documented format, and provides easily usable machine-learning tools. Both services provide software interfaces to allow for automated data querying, processing, and access. The widespread adoption of materials informatics will require more databases to deploy machine-accessible interfaces and increased adoption of data publication practices. Even if not immediately useful, open data and machine-accessible databases lay the groundwork for future materials innovation via informatics.