Matthew Witman1,Anuj Goyal2,Tadashi Ogitsu3,Stephan Lany2,Anthony McDaniel1
Sandia National Laboratories1,National Renewable Energy Laboratory2,Lawrence Livermore National Laboratory3
Matthew Witman1,Anuj Goyal2,Tadashi Ogitsu3,Stephan Lany2,Anthony McDaniel1
Sandia National Laboratories1,National Renewable Energy Laboratory2,Lawrence Livermore National Laboratory3
Graph convolutional networks (GCNs) provide a powerful technique to perform deep learning-based predictions on crystal structures and have therefore seen rapid, widespread adoption in materials science applications. Here we demonstrate a generalized GCN that can predict vacancy formation enthalpies of any site in the crystal structure by properly utilizing local node attributes following the graph convolutions. Using only the DFT relaxed <i>compound</i> as the model input to accurately predict the final vacancy formation enthalpy of the DFT relaxed <i>defected structure</i>, we greatly accelerate the computational and man-power intensive process of computing relaxed defect formation enthalpies. Various embedding and convolution strategies are investigated to extend and quantify the “extrapolative capabilities” of the model as well its potential for accurate materials discovery predictions, with model performance being evaluated by careful cross validation. Achieving mean absolute errors on oxygen vacancy formation enthalpies well below 500 meV across a diverse chemical and structural space of complex metal oxides allows us to directly screen new materials for solar thermochemical water-splitting and rapidly identify top potential candidates that have not yet been experimentally investigated.