Fewest-Switches Surface Hopping with Combined Deep Learning Potential and Long Short-Term Memory Network Propagator...

Area of Science:

• Computational Chemistry
• Photochemistry
• Machine Learning in Chemistry

Background:

• Fewest-Switches Surface Hopping (FSSH) is a standard method for simulating molecular photochemical processes.
• Long Short-Term Memory (LSTM) networks have been previously explored as propagators in FSSH dynamics.

Purpose of the Study:

• To develop an extended LSTM-FSSH framework for simulating realistic photochemical reactions.
• To improve the representation of high-dimensional nuclear degrees of freedom in FSSH simulations.
• To accurately predict excited-state lifetimes and product yields for molecular photoisomerizations.

Main Methods:

• Redesigned LSTM input features and training procedures for effective high-dimensional nuclear dynamics.
• Integrated equivariant neural networks with LSTM to construct adiabatic potential energy surfaces.
• Simulated photoisomerizations of CH2NH and azobenzene using the developed LSTM-FSSH method.

Main Results:

• The enhanced LSTM-FSSH framework accurately reproduced collective results on Tully's three models.
• Simulations of CH2NH and azobenzene photoisomerizations showed accurate excited-state lifetimes and product yields.
• The method requires only 10 reference trajectories for training, enabling efficient dynamic simulations.

Conclusions:

• The proposed LSTM-FSSH method offers an accurate and efficient approach for simulating photochemical reactions.
• This framework effectively handles complex molecular dynamics and provides reliable predictions.
• The integration of machine learning with FSSH significantly advances the simulation of photochemistry.