Efficient Nonadiabatic Molecular Dynamics with Machine Learning Hamiltonian Interpolation

Nonadiabatic (NA) molecular dynamics (MD) is the method of choice for modeling far-from-equilibrium, excited state processes in molecules and materials. Machine learning (ML) can streamline all NAMD components, enabling quantum dynamics simulations of thousand-atom systems over nanoseconds. By comparing three qualitatively different ML models to interpolate the NA Hamiltonian and testing them on a metal halide perovskite, we demonstrate that bidirectional long-short-term-memory (BiLSTM) gives the best performance, since it is efficient for smaller, sequence-dependent data sets. Transformer also provides an accurate representation, although the amount of data is not sufficiently large to take full advantage of transformer capabilities. Kernel ridge regression (KRR) is simple and inexpensive, achieving rapid and robust NA Hamiltonian interpolation, although it requires smaller steps. Even with sparse training data, the models can closely replicate ab initio results while achieving 2 orders of magnitude computational savings. The reported advances allow one to accelerate the discovery and optimization of energy and optoelectronic materials.