Materials discovery for high-temperature, clean-energy applications using graph neural network models of vacancy defects and free-energy calculations

We present a graph neural network modeling approach that fully automates the prediction of the DFT-relaxed vacancy formation enthalpy of any crystallographic site from its DFT-relaxed host structure. Applicable to arbitrary structures with an accuracy limited principally by the amount/diversity of the data on which it is trained, this model accelerates the screening of vacancy defects by many orders of magnitude by replacing the DFT supercell relaxations required for each symmetrically unique crystal site. It can thus be used off-the-shelf to rapidly screen 10,000s of crystal structures (which can contain millions of unique defects) from existing databases of DFT-relaxed crystal structures. This modeling approach therefore provides a significant screening and discovery capability for a plethora of applications in which vacancy defects are the primary driver of a material's utility. For example, by high-throughput screening the Materials Project's metal oxides, we rapidly "re-discover" and identify new high potential candidate materials for hydrogen generation via solar thermochemical water splitting and energy storage, for CO2 conversion via reverse water gas shift chemical looping, and for cathodes in solid oxide fuel cells. Thermodynamic modeling on the basis of the high-throughput screening results allows us to connect the predicted defect energies to high temperature process conditions relevant to the different application areas, and we extract the reduction entropies as an additional selection criterion for high-performance materials. Further model development and accumulation of additional training data will only serve to expand the significant utility of this generalizable defect model to solving materials discovery problems in clean energy applications and beyond.