Historically, materials discovery is driven by a laborious trial-and-error process. However, with the growth of materials databases, emerging informatics approaches o er an opportunity to transform this practice into data- and knowledge-driven rational design - accelerating discovery of novel materials exhibiting desired properties. Using data from the Aflow repository for high-throughput ab initio calculations, we have generated Machine Learning (ML) models to predict several critical material properties, namely the metal/insulator classification, elastic tensor, Fermi energy, and band gap energy. The prediction accuracy obtained with these ML models approaches that of GGA-DFT functionals for virtually any stoichiometric inorganic material. We attribute the success and universality of these models to the construction of new material descriptors - referred to as the universal property-labeled fragments (PLMF). This representation affords straightforward model interpretation in terms of simple heuristic design rules that could guide rational materials design. This proof-of-concept study demonstrates the power of materials informatics to dramatically accelerate the search for new materials.
Back to Workshop III: Collective Variables in Quantum Mechanics