Predicting The Structure and Properties of Bulk and Surface Perovskite Oxides with Machine Learning and Simulations

Ceramics based on multicomponent oxides are a promising platform for catalysis and electronics. By controlled doping of multiple elements, it may be possible to tune electronic and transport properties of oxide materials towards high performance applications. This opens an exciting and high-dimensional design space for choosing the most promising choices and stoichiometric ratios of elements that fine tune desired properties.<br/><br/>Here we will describe how machine learning models, powered by electronic structure simulations can address property prediction and design in multicomponent space. In particular, we will explore how perovskite oxides, which can support quaternary or even more complex compositions, can be engineered in silico.<br/><br/>We will describe the use of elemental and electronic descriptors to predict whether perovskite oxide materials will exhibit chemical ordering when synthesized, we will evaluate the use of per-site deep learning models to predict atomically-project properties such as magnetic moments of catalytic side activity, the use of machine learning interatomic potentials to relax perovskite structures and to create surface phase diagrams, and the ability of equivariant models to capture symmetry-breaking relaxation and properties from idealized, unrelaxed prototypes.<br/><br/>These tools put together, alongside machine learning models for synthesis planning, will support the development of bottom-up data-driven design pipelines for multicomponent oxide materials with tailored properties.