Evaluating the Performance of Equivariant Graph Neural Network Force Fields on Point Defects in Solids

Point defects are a universal feature of crystalline materials, whose identification is often addressed by combining experimental measurements with theoretical models. Recently, the defect modelling community has made steps toward adopting machine learning approaches for defect simulations, to avoid the limits imposed by costly DFT supercell relaxations.¹ However, many challenges remain to be tackled before these ML methods can be adopted by the wider defects community, such as energy accuracies beyond coarse screening and the ability to predict forces (i.e. interatomic potentials). Indeed, many of the limitations of state-of-the-art ML force-fields (MLFFs) for modelling defects, and their underlying root causes, are not yet well established.<br/><br/>In this work, we train a state-of-the-art equivariant graph neural network model on a large dataset of oxygen vacancy supercell relaxations² across hundreds of metal oxide compounds, initialised using the ShakeNBreak defect structure-searching approach.^3,4 Oxygen vacancies are typically the most important defect species in wide-bandgap oxides, governing key properties such as conductivity and catalytic activity, making for a relevant chemical subspace for evaluating MLFF performance with defects. In doing so, we identify key limitations in the standard architectures of ML potentials for the investigation of localised perturbations (e.g. defects, polarons, interfaces) within large systems, along with strategies to reduce these shortcomings. By comparison of our model with an invariant analogue, we additionally show the performance impact of <i>equivariant</i> features within the model.<br/><br/>1 M. D. Witman, A. Goyal, T. Ogitsu, A. H. McDaniel and S. Lany, <i>Nat Comput Sci</i>, 2023, <b>3</b>, 675–686.<br/>2 Y. Kumagai, N. Tsunoda, A. Takahashi and F. Oba, <i>Phys. Rev. Materials</i>, 2021, <b>5</b>, 123803.<br/>3 I. Mosquera-Lois, S. R. Kavanagh, A. Walsh and D. O. Scanlon, <i>Journal of Open Source Software</i>, 2022, <b>7</b>, 4817.<br/>4 I. Mosquera-Lois, S. R. Kavanagh, A. Walsh and D. O. Scanlon, <i>npj Comput Mater</i>, 2023, <b>9</b>, 1–11.