Ultrafast Dynamics at Surfaces with Machine Learning Surrogates

Ultrafast dynamics at surfaces (driven by light, electrons, or hyperthermal scattering) involve the concerted motion of electrons and atoms at comparable energy and time scales, giving rise to nonadiabatic effects. Excited electrons drive chemical conversions, induce phase transitions, and mediate energy transfer between adsorbates and surfaces. To reliably predict such effects with scalable, state-of-the-art nonadiabatic dynamics simulations requires the use of accurate and data-efficient high-dimensional machine learning (ML) surrogate models. This includes representations of energy landscapes, but also nonadiabatic couplings or excited-state properties that are required for nonadiabatic simulations. I will present recent strategies to construct high-dimensional ML surrogate models of electronic structure, including active learning and fine-tuning of foundation models that allow us to reduce the required electronic structure data to a few hundred data points per gas-surface dynamics model and even transfer learn across density functional approximations. Electronic properties such as electron-phonon coupling tensors or electronic Hamiltonians can be efficiently represented by encoding physical equivariance properties in the model. I will showcase the utility of the introduced models with recent dynamics applications on reactive molecular scattering and light-driven structural dynamics.