Fewest-Switches Surface Hopping with Combined Deep Learning Potential and Long Short-Term Memory Network Propagator for Simulating Realistic Photochemical Processes

TL;DR

This work develops a scalable machine-learning framework that couples NequIP-based adiabatic potential energy surfaces with an LSTM network to propagate electronic degrees of freedom in fewest-switches surface hopping (FSSH) dynamics. By redesigning input features and incorporating physical constraints, the approach enables realistic photochemical simulations for large systems with only a small number of reference trajectories (as few as 10). The method is validated on CH$_2$NH and azobenzene photoisomerizations, yielding excited-state lifetimes and product yields in good agreement with conventional FSSH, while substantially reducing computational cost. The results highlight both the potential and current limitations of ML-driven nonadiabatic dynamics, including PES accuracy near conical intersections and the need for targeted sampling to improve training efficiency.

Abstract

Fewest-switches surface hopping (FSSH) is the most popular method for simulating photochemical processes of molecular systems. Recently, we have constructed long short-term memory (LSTM) networks as a propagator for electronic subsystems in FSSH dynamics simulations. The collective results on Tully's three models have been reproduced satisfactorily. In the present work, we develop an extended LSTM-FSSH framework to simulate realistic photochemical reactions. The input features of LSTM as well as the training procedure are redesigned to represent high-dimensional nuclear degrees of freedom in an effective way. Equivariant neural networks are integrated with LSTM to build adiabatic potential energy surfaces in ground and excited states. Photoisomerizations of $\mathrm{CH_2NH}$ and azobenzene are simulated, showing that our new proposed LSTM-FSSH method can produce excited-state lifetimes and product yields accurately in comparison with conventional FSSH simulations as reference. Only 10 reference trajectories are required for training LSTM networks, and then a trajectory ensemble can be generated with very efficient LSTM-FSSH dynamics simulations to obtain collective results.

Figures (7)

• Figure 1: Framework (a) and unit (b) of LSTM network model.
• Figure 2: Workflow of combined LSTM and NequIP for FSSH simulations.
• Figure 3: Photoisomerizations of $\mathrm{CH_2NH}$ (a) and azobenzene (b).
• Figure 4: Comparisons between reference and NequIP-predicted potential energies for $\mathrm{CH_2NH}$ in $\mathrm{S}_0$ (orange), $\mathrm{CH_2NH}$ in $\mathrm{S}_1$ (red), azobenzene in $\mathrm{S}_0$ (green), and azobenzene in $\mathrm{S}_1$ (blue). Absolute values of potential energies have been shifted by a constant for clarity: 2,555 eV for $\mathrm{CH_2NH}$ in $\mathrm{S}_0$, 2,552 eV for $\mathrm{CH_2NH}$ in $\mathrm{S}_1$, 2,102 eV for azobenzene in $\mathrm{S}_0$, and 2,100 eV for azobenzene in $\mathrm{S}_1$.
• Figure 5: Population of $\rho_{00}$ and $x_{log\Delta E}$ as functions of time evolved in representative trajectories for $\mathrm{CH_2NH}$ (a) and azobenzene (b). Different colors represent different variables: $\rho_{00}$ (FSSH) in black, $\rho_{00}$ (LSTM) in red, and $x_{log\Delta E}$ in blue.