Speaker
Description
X-ray free-electron lasers enable femtosecond-resolution probing of photoinduced chemical dynamics. Interpreting these experiments requires atomistic simulations of excited-state systems on matching timescales - a task where ab initio molecular dynamics remains prohibitively expensive. We present a Neural Network Quantum Molecular Dynamics (NNQMD) framework that bridges this gap by training E(3)-equivariant graph neural networks on DFT calculations of hole-doped water, mimicking the ionization conditions created by intense FEL pulses at liquid interfaces. Three architectures - NequIP, MACE, and Allegro - are benchmarked on a 324-atom water system across four charge states. Molecular dynamics simulations driven by the learned potentials remain stable even under extreme 16e excitation, capturing bond-breaking and dissociation events. The resulting pair distribution function differences (ΔPDF) reproduce signatures of bond softening and solvent shell expansion consistent with FEL pump-probe measurements. This framework provides a scalable ML-driven simulation tool for predicting and interpreting ultrafast structural dynamics observed at XFEL facilities, and establishes a foundation for reinforcement-learning-guided steering of catalytic pathways in FEL experiments.
Funding Agency
This work was supported by the U.S. Department of Energy, Office of Science, Basic Energy Sciences, under Contract No. DE-SC0063.
Footnotes
This research used resources of the National Energy Research Scientific Computing Center (NERSC), a Department of Energy User Facility (project m5047-2025).
| I have read and accept the Privacy Policy Statement | Yes |
|---|