Machine Learning Nonadiabatic Dynamics: Eliminating Phase Freedom of Nonadiabatic Couplings with the...

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Machine Learning in Chemistry

Background:

• Excited-state molecular dynamics (ESMD) simulations are crucial for understanding photochemical processes.
• Machine learning potentials (MLPs) offer efficient modeling of electron-nuclear dynamics but face challenges near conical intersections (CIs).
• Discontinuities from CI singularities and double-valued coupling elements hinder the accuracy of MLPs in nonadiabatic dynamics.

Purpose of the Study:

• To develop a robust method for improving the stability and accuracy of MLPs in ESMD simulations.
• To address the challenges posed by conical intersections and double-valued coupling elements in nonadiabatic dynamics.
• To enable more reliable and efficient large-scale and long-time scale ESMD simulations.

Main Methods:

• Introduction of the phaseless coupling term, Δ², derived from the diabatic Hamiltonian within the SI-SA-REKS formalism.
• Application of the Δ² term to improve MLP training for nonadiabatic dynamics.
• Validation using excited-state molecular dynamics simulations of the penta-2,4-dieniminium cation (PSB3).

Main Results:

• The Δ²-based method significantly enhances the stability and accuracy of MLPs by mitigating issues at conical intersections.
• The improved MLPs accurately reproduce ab initio excited-state molecular dynamics simulations.
• The approach demonstrates effectiveness in training MLPs for complex nonadiabatic dynamics.

Conclusions:

• The phaseless coupling term Δ² provides a definitive solution for discontinuities in MLPs near conical intersections.
• The developed ML-ESMD method is efficient and accurate for modeling nonadiabatic dynamics.
• This advancement holds significant potential for broader applications in large-scale and long-time scale excited-state molecular dynamics simulations.