Machine Learning Polarons

Understanding the behavior of small polarons is essential for elucidating fundamental material properties and enabling functionalities in catalysis, electronics, and energy conversion. Yet, the highly localized and nonlinear nature of polaronic states poses significant challenges for their accurate characterization using first-principles methods alone. In this seminar, we present recent advances that integrate density functional theory (DFT) with machine learning (ML) techniques to predict the conformational and dynamical properties of small polarons at the nanoscale. We first introduce ML-accelerated frameworks that efficiently explore the polaron configurational landscape and identify stable charge-localization patterns at oxide surfaces. By designing defect-sensitive descriptors and training supervised models on DFT datasets, we dramatically reduce the computational cost of determining ground-state configurations. We demonstrate how coupling ML with global optimization strategies disentangles the interplay between defects and polarons, providing predictive insight into the inhomogeneous defect distributions observed in STM and AFM measurements. We then present a novel ML-enhanced molecular dynamics scheme, where a message-passing neural network learns the polaronic oxidation states and hopping trajectories, enabling nanosecond-scale simulations of polaron hopping dynamics. This approach delivers accurate estimates of activation energies and anisotropic mobilities, bringing theoretical predictions into close agreement with experiment.

Altogether these developments, and ongoing extensions, highlight the capability of machine learning to overcome the spatiotemporal limitations of traditional ab initio methods, opening new pathways to understand and engineer defect-mediated phenomena in semiconductors and insulators.