Efficient Nonadiabatic Molecular Dynamics with Machine Learning Hamiltonian Interpolation

Area of Science:

• Computational chemistry and materials science.
• Quantum dynamics simulations.
• Machine learning applications in scientific research.

Background:

• Nonadiabatic (NA) molecular dynamics (MD) is crucial for simulating excited-state processes far from equilibrium.
• Current NA-MD methods face computational limitations for large systems and long timescales.
• Machine learning (ML) offers a promising avenue to enhance the efficiency of NA-MD simulations.

Purpose of the Study:

• To investigate the efficacy of different ML models in interpolating the NA Hamiltonian for accelerated NA-MD simulations.
• To compare the performance of Bidirectional Long-Short-Term-Memory (BiLSTM), Transformer, and Kernel Ridge Regression (KRR) models.
• To demonstrate the potential of ML-accelerated NA-MD for simulating large systems and facilitating materials discovery.

Main Methods:

• Developed and compared three ML models (BiLSTM, Transformer, KRR) for interpolating the NA Hamiltonian.
• Tested the ML models on a metal halide perovskite system.
• Evaluated the accuracy and computational efficiency of each ML model against ab initio calculations.

Main Results:

• Bidirectional Long-Short-Term-Memory (BiLSTM) demonstrated superior performance for smaller, sequence-dependent datasets.
• Transformer models showed accurate representation, though requiring larger datasets to reach full potential.
• Kernel Ridge Regression (KRR) provided rapid and robust interpolation with smaller time steps, offering a simple and cost-effective solution.
• All tested ML models achieved up to two orders of magnitude computational savings, closely replicating ab initio results even with sparse training data.

Conclusions:

• Machine learning models can significantly accelerate nonadiabatic molecular dynamics simulations.
• BiLSTM, Transformer, and KRR models offer viable pathways for efficient NA Hamiltonian interpolation.
• These advancements pave the way for faster discovery and optimization of energy and optoelectronic materials through large-scale quantum dynamics simulations.