Interpolating Nonadiabatic Molecular Dynamics Hamiltonian with Artificial Neural Networks

Area of Science:

• Computational Chemistry
• Materials Science
• Quantum Mechanics

Background:

• Nonadiabatic molecular dynamics (NA MD) is crucial for simulating processes involving excited electronic states and atomic motion.
• Traditional NA MD methods require computationally expensive calculations of excitation energies and nonadiabatic couplings (NACs).
• Machine learning (ML) can efficiently compute ground-state properties but faces challenges with excited states and NACs due to complexity and cost.

Purpose of the Study:

• To develop a computationally efficient NA MD methodology for excited-state dynamics.
• To leverage machine learning for interpolating excited-state energies and NACs, reducing computational burden.
• To enable accurate simulations of complex systems, such as metal halide perovskites, undergoing molecular dynamics.

Main Methods:

• A novel NA MD approach was developed, avoiding time extrapolation of excitation energies and NACs.
• Utilized a precomputed ground-state trajectory within the classical path approximation.
• Trained neural networks using a small fraction (2%) of geometries to interpolate excited-state energies and NACs for the remaining 98%.

Main Results:

• The ML-accelerated NA MD method achieved nearly two orders of magnitude in computational savings.
• Demonstrated accuracy comparable to traditional methods, validated with complex molecular dynamics of metal halide perovskites.
• Successfully generated accurate NA MD results for systems with intricate dynamics.

Conclusions:

• The developed ML-based NA MD methodology offers a significant computational advantage for studying excited-state phenomena.
• This approach provides a viable and accurate alternative for simulating complex molecular dynamics, particularly in materials science.
• The method paves the way for more extensive investigations of excited-state processes in various chemical and physical systems.