Siamese Equivariant Neural Network for Property Predictions of Point Defects in Solids

Computations of point defect properties such as formation energy using density functional theory (DFT) in defected materials is critical for the understanding of defect-property correlations and defected material growth mechanisms, yet the accurate and efficient calculation of defect properties remains a challenge in materials science. In this study, we introduce the Siamese Equivariant Neural Network (SENN) for predicting properties in defected material systems. We leverage E(3) equivariance to construct representations for both defects and their host crystal structures, and use the difference of the learned representations for property predictions, thereby forming a Siamese network structure. Our results demonstrate that the E(3) model surpasses previous invariant graph neural network models, and the proposed SENN further enhances the prediction performance on various defects-in-materials datasets. Our model can be applied for fast prediction of defect properties such as formation energies and beyond, which can be used for fast screening of functional defects and high-throughput computational study of defected material systems at a unprecedented scale.