Machine learning small polaron dynamics

Polarons are charged quasiparticles that form in polarizable materials. These particles play a significant role in important phenomena such as charge transfer, electron-hole recombination, and catalytic processes. Different methods have been used to study the equilibrium properties of these states while their dynamics continues to be accessible only through standard ab initio molecular dynamics (AIMD). However, the limited time scales and system sizes of such simulations makes the adequate sampling of infrequent polaron hopping events a challenge. In recent years, machine learning potentials (MLPs) have helped bridge the gap between time- and length-scales in AIMD. In this study, we develop a MLPs architecture using an equivariant graph neural network that successfully accounts for small polaron hopping dynamics at the nanosecond timescale. We apply the proposed ML framework to study hole and electron polaron dynamics in MgO and (F-doped) TiO2, enabling the estimation of the polaron (anisotropic) mobility across varying temperatures and in the presence of dopants.