Mixed Quantum-Classical Dynamics with Machine Learning-Based Potentials via Wigner Sampling.

Area of Science:

• Computational Chemistry
• Quantum Dynamics
• Machine Learning Applications

Background:

• Machine learning (ML) enhances surface hopping (SH) simulations for accurate quantum dynamics.
• Challenges exist in ML-based SH due to fitting nonadiabatic coupling near conical intersections.
• Prior methods used indirect hopping probability calculations or adaptive sampling.

Purpose of the Study:

• To develop a novel ML approach for surface hopping simulations.
• To improve the efficiency of training data generation for ML-based SH.
• To reduce the computational cost of accurate quantum dynamics simulations.

Main Methods:

• Introduced a modified Wigner sampling technique for training data generation.
• Applied the method to test surface hopping simulations.
• Utilized the two-state spin-boson Hamiltonian system for validation.

Main Results:

• The modified Wigner sampling significantly reduces the required training dataset size.
• Achieved up to a 7.5-fold reduction in dataset size per degree of freedom compared to linear sampling.
• Demonstrated the effectiveness of the new sampling strategy on a model system.

Conclusions:

• Modified Wigner sampling offers an efficient strategy for generating training data in ML-based SH.
• This approach simplifies the implementation of accurate and computationally feasible SH simulations.
• The findings pave the way for more accessible quantum dynamics studies.