Today one can sit at a computer and design an inorganic material from scratch, prescribing which atoms should be where, and using first-principles calculations to get quality estimates of several properties. These predictions are good enough to judge whether the material shows promise as a battery electrode or a catalyst, for example. As computing time is becoming cheaper, we see more examples of computational campaigns to find functional materials in this way. Forecasting which of these “hypothetical” materials can be realized in the laboratory, however, turns out to be a significant challenge. This challenge comes as no surprise to many, as the synthesis of inorganic materials is a process controlled not only by fundamental factors like thermodynamics and kinetics of reactions between materials, but also by circumstantial ones, such as equipment or precursor availability, or expertise in certain material chemistries.
While searching for an alternative, data-driven solution to this problem, I recognized that the phase diagrams of materials that we obtain from first-principles calculations in large computational databases of materials (such as Open Quantum Materials Database or Materials Project) can be interpreted as a network, where stable materials are the “nodes” and the tie-lines are the “edges”. Extraction of a timestamp from the published literature for each material in this network converted it to a “temporal” network of materials -- where one could see how the network evolves as materials are discovered.
This realization was a moment of excitement for me and my colleagues at the Toyota Research Institute because while the linkage rule was described purely by thermodynamics, the evolution of the network encoded complex factors that influence realization of new materials. The information to build a predictive model for synthesizability was potentially in the network. Furthermore, we could see a cross-cutting approach emerging from our work, where the field of materials science was intersecting with the science of complex networks.
We teamed up with Chris Wolverton and Vinay Hegde from Northwestern University to analyze this time-dependent, thermodynamic network of materials. We were able to connect the dots between the evolution of the network and “discovery” or “synthesis” via simple machine-learning classifiers. For stable, hypothetical materials, we are able to assign a likelihood of realization in the laboratory. Our results have recently appeared in Nature Communications.
We hope that this work will serve as an example for many new applications of network science methods in materials research.