Combining SchNet and SHARC: The SchNarc Machine Learning Approach for Excited-State Dynamics

Area of Science:

• Quantum chemistry
• Photochemistry
• Computational chemistry

Background:

• Deep learning is increasingly applied in scientific research, including quantum chemistry.
• Photochemistry simulations require accurate computation of molecular energies, forces, and couplings.

Purpose of the Study:

• To demonstrate deep learning's utility in advancing photochemistry research.
• To develop a method for simplifying photodynamics simulations.

Main Methods:

• Employed SchNet, a deep learning model, extended for multiple electronic states.
• Developed a phase-free training approach and rotationally covariant nonadiabatic couplings.
• Incorporated spin-orbit couplings and utilized the SHARC molecular dynamics program.

Main Results:

• The SchNarc approach was tested on two polyatomic molecules.
• The method simplifies photodynamics simulations by learning key molecular properties.
• Achieved efficient simulations without costly preprocessing of quantum chemistry data.

Conclusions:

• The SchNarc method offers a pathway toward efficient photodynamics simulations for complex systems.
• Deep learning significantly enhances capabilities in computational photochemistry.
• This work bridges deep learning and molecular dynamics for advanced chemical simulations.