Machine-Learning Dynamics of Redox-Active Materials: Polaron Diffusion on the TiO2(110) Surface

In redox-active materials, atoms dynamically change their oxidation states in response to excess charge and electronic transport, with the potential for novel technologies spanning electronics, catalysis, and energy storage and conversion. Despite the remarkable progress of modern artificial intelligence methods in accelerating atomistic simulations, capturing the quantum-mechanical nature of charge transport and dynamically varying oxidation states remains a fundamental challenge, limiting scientific and technological progress. In this work, we introduce a machine-learning-based molecular dynamics framework capable of explicitly tracking the oxidation state dynamics arising from small-polaron hopping (excess charges trapped on single lattice sites that diffuse upon thermal activation). Our model uses descriptors augmented with charge localization information at each time step and is trained on density functional theory data to capture signatures of polaronic diffusion. We apply our method to the rutile TiO2(110) surface, which is a prototypical reducible oxide. Our simulations reveal that charge-carrier diffusion is highly sensitive to the local surface environment: polarons preferentially localize near oxygen vacancies, whereas their interaction with hydroxyl groups is weaker. These results provide a microscopic interpretation of experimental observations, highlight strategies to tune charge-carrier mobility through surface engineering, and demonstrate the applicability of our approach to investigating charge transport in complex materials.