Machine Learning for Exploring Small Polaron Configurational Space

Abstract

Polaron defects are ubiquitous in materials and play an important role in many processes involving carrier mobility, charge transfer and surface reactivity. Determining the spatial distribution of small polarons is essential to understand materials properties and functionalities. This requires an exploration of the configurational space, which is computationally demanding when using standard first principles methods, and technically prohibitive for many-polaron systems. Here, we propose a machine-learning (ML) accelerated search that compares the energy stability of different polaron patterns and determines the ground state configuration. The kernel-regression based ML model is trained on databases generated by density functional theory (DFT) calculations on a minimal set of initial polaron patterns, obtained by using either molecular dynamics simulations or a random sampling approach. To establish an efficient mapping between training data and configuration stability we designed simple descriptors that model the interactions among polarons and charged point defects. The proposed DFT+ML protocol is used here to explore millions of polaron configurations for two different systems, oxygen defective rutile TiO$_2$(110) and Nb-doped SrTiO$_3$(001). Our data shows that the ML-aided search correctly individuates the ground-state polaron patterns, proposes polaronic configurations not visited in the training and can be used to efficiently determine the optimal distribution of polarons at any charge concentration.