Publication date: 22nd July 2026
Halide perovskites are often described through their average crystal structures, although many of their important properties arise from fluctuations that are absent from this static picture. In this talk, I will discuss how atomistic simulation can resolve their structural dynamics and connect them to electronic behavior. First-principles molecular dynamics provides microscopic insight into many important effects. These include octahedral tilting, local distortions, and can also account for their temperature dependence. However, this approach is restricted to relatively small systems and short trajectories. Machine-learning force fields overcome much of this limitation and make it possible to sample realistic structural ensembles at substantially larger scales. Structural dynamics alone, however, do not provide the electronic information needed to predict optoelectronic properties. I will therefore introduce HAMSTER, a physics-informed Hamiltonian-learning approach that combines an approximate physical model with a machine-learned corrections for fluctuating atomic environments. With only modest first-principles input, our HAMSTER model remains transferable across temperatures and compositions and can treat systems containing tens of thousands of atoms. This combination of learned atomic dynamics and learned electronic Hamiltonians offers a route toward interpretable finite-temperature predictions for structurally complex halide perovskites and other materials.
