Excited State Machine Learning Molecular Dynamics Simulations for Ultrafast Scattering Experiments.

Understanding photoinduced charge and structural dynamics in complex materials requires simulation frameworks that bridge femtosecond electronic processes and atomic-scale structural evolution. Ab-inito frameworks for excited dynamics can often give good insight into ultrafast scattering experiments but are computationally expensive and typically cannot capture large spatial scale/small wavevector dynamics.  Machine learning interatomic potentials (MILPs) offer a promising route forward to understand such dynamics but are often less explored for excited dynamics due to the multiscale challenges of modeling electronic and ionic dynamics on the same footing.  Here we discuss our multiscale approaches to integrate MILPs and excited state dynamics, with recent results on examining ultrafast electron scattering experiments of plasmonically pumped solvated Pt nanoparticles.  Our ML-MD approach utilizing both foundational and fine-tuned ML potentials can reproduce experimental diffraction and pair-distribution features with near ab initio fidelity at orders-of-magnitude lower cost. Integrating such ML-MD frameworks within realtime data streams has the potential to utilize MD for adaptive feedback on human decision making timescales in next-generation ultrafast scattering experiments.