Fewest-Switches Surface Hopping with Long Short-Term Memory Networks

Area of Science:

• Computational Chemistry
• Quantum Dynamics
• Machine Learning in Science

Background:

• Nonadiabatic phenomena are crucial in photophysics and photochemistry.
• Mixed quantum-classical dynamical simulations are essential for studying these phenomena.
• Machine learning models are increasingly used to accelerate nuclear subsystem time evolution.

Purpose of the Study:

• To implement long short-term memory (LSTM) networks as a propagator.
• To accelerate the time evolution of the electronic subsystem in fewest-switches surface hopping (FSSH) simulations.
• To demonstrate the applicability of LSTM to surface hopping simulations for nonadiabatic processes.

Main Methods:

• Generated reference trajectories using the original FSSH method.
• Built LSTM networks using these reference trajectories.
• Applied the constructed LSTM networks to FSSH simulations with identical initial conditions and random numbers.

Main Results:

• Successfully implemented LSTM networks to accelerate electronic subsystem time evolution within FSSH.
• Produced trajectory ensembles to reveal mechanisms of nonadiabatic processes.
• Qualitatively reproduced collective results using Tully's three models as test systems.

Conclusions:

• LSTM networks can effectively accelerate electronic subsystem dynamics in FSSH simulations.
• This machine learning approach is applicable to popular surface hopping simulations.
• The method aids in understanding nonadiabatic processes in photochemistry and photophysics.