Machine Learning-Accelerated Simulation of Ultrafast Dynamics in Complex Systems

Photochemical reactions and photophysical processes occur on ultrafast timescales ranging from femtoseconds to picoseconds. While conventional data-driven models have proven effective for ground-state trajectory predictions, their accuracy in capturing multi-electronic state dynamics remains limited due to insufficient incorporation of physical insights. To overcome these challenges, we developed two machine learning-accelerated frameworks: a neural network-based potential energy surface (PES) model, enabling precise simulation of photoinduced isomerization in confined spaces for supramolecules;(1-2) and a physics-informed time-series architecture combining neural ordinary differential equations with Hamiltonian dynamics,(3) specifically optimized for nonadiabatic molecular dynamics (NAMD) to accurately capture electron-nuclear coupling effects. These approaches reduce computational costs by 2-3 orders of magnitude while maintaining quantum mechanical accuracy. Based on this foundation, we further implemented an external field co-regulation strategy using phase-matched ultraviolet-infrared dual-frequency radiation to actively control photochemical reaction pathways.(4) This work provides an efficient computational toolkit for investigating excited-state dynamics in complex systems while opening new avenues for the rational design of photochemical reactions.

 

[1] Xu, H.; Zhang, B.; Tao, Y.; Xu, W.; Hu, B.; Yan, F.; Wen, J. J. Phys. Chem. A 127, 7682 (2023).
[2] Xu, W.; Xu, H.; Zhu, M.; Wen. J. Phys. Chem. Chem. Phys. 26, 25994 (2024).
[3] Xu, H.; Zhu, L.; Niu, L.; Yan, F.; Zhu, M.; Wen. J. DOI: 10.26434/chemrxiv-2025-jrjz5.
[4] Xu, H.; Zhu, L.; Yan, F.; Zhu, M.; Wen, J. submitted.