Abstract
Machine Learning for Small Polaron Modeling: Stability, Transport, and Defect Interactions
Small polarons -- localized charge carriers coupled to lattice distortions -- play a key role in determining transport properties, defect chemistry, and surface reactivity in oxide materials. However, their local character and dynamic behavior pose major challenges for first-principles modeling, especially due to the large configurational space of defect-polaron arrangements and the difficulty of capturing hopping dynamics. This dissertation explores how machine learning (ML) can overcome these challenges to enable accurate and efficient modeling of small polarons in complex oxide environments. ML models trained on DFT data predict stable configurations in Nb-doped SrTiO₃ and TiO₂, revealing both known and novel polaron patterns and clarifying experimental observations. Message-passing neural networks are further used to model polaron hopping in MgO and TiO₂, capturing mobilities, activation barriers, and polaron defect correlation in long-timescale simulations. These results demonstrate that ML can bridge the gap between computational efficiency and physical accuracy in polaron modeling, extending the applicability of ML-based potential energy surfaces to localized charge carriers.
Defense committee:
Ralf Drautz, Ruhr Universität Bochum, DE (reviewer)
Julia Wiktor, Chalmers University of Technology, SE (reviewer)
Cesare Franchini (supervisor)
Thomas Pichler (chair)