Identifying split vacancy defects with machine-learned foundation models and electrostatics

Abstract Point defects are ubiquitous in solid-state compounds, dictating many functional properties such as conductivity, catalytic activity and carrier recombination. Over the past decade, the prevalence of metastable defect geometries and their importance to relevant properties has been increasingly recognised. A striking example is split vacancies, where an isolated atomic vacancy transforms to a stoichiometry-conserving complex of two vacancies and an interstitial (V _X → [V _X + {X}_i + V _X]), which can be accompanied by a dramatic energy lowering and change in behaviour. These species are particularly challenging to identify from computation, due to the ‘non-local’ nature of this reconstruction. Here, I present an approach for the efficient identification of these defects, through tiered screening which combines geometric analysis, electrostatic energies and foundation machine-learned (ML) models. This approach allows the screening of all solid-state compounds in the Materials Project database (including all entries in the ICSD, along with several thousand predicted metastable materials), identifying thousands of low energy split vacancy configurations, hitherto unknown. This study highlights both the potential utility of (foundation) machine-learning potentials, with important caveats, the significant prevalence of split vacancy defects in inorganic solids, and the importance of global optimisation approaches for defect modelling.