ShakeNBreak: Navigating the defect configurational landscape

Abstract

Point defects are present in all crystalline solids, controlling the properties and performance of most functional materials, including thermoelectrics, photovoltaics and catalysts. However, the standard modelling approach, based on local optimisation of a defect placed on a known crystal site, can miss the true ground state structure. This structure may lie within a local minimum of the potential energy surface (PES), trapping a gradient-based optimisation algorithm in a metastable arrangement and thus yielding incorrect defect structures that compromise predicted properties (Mosquera-Lois & Kavanagh, 2021). As such, an efficient way to explore the defect energy landscape and identify low-energy structures is required. Statement of need To tackle this limitation, two approaches have recently been designed. Arrigoni and Madsen (Arrigoni & Madsen, 2021a) developed an evolutionary algorithm combined with a machine learning model to navigate the defect configurational landscape and identify low-energy structures. While ideal to study specific defects, its complexity and computational cost hinders its application to typical defect investigations. Alternatively, Pickard and Needs (Pickard & Needs, 2011) applied random sampling to the atoms near the defect site-with the limitation that random sampling on a high-dimensional space lowers efficiency and increases computational cost. To improve sampling efficiency, domain knowledge can be used to tailor the sampling structures towards likely energy-lowering distortions. This is the purpose of our package, which aims to serve as a simple, efficient, and affordable tool to identify low-energy defect structures.