Rapid Prediction of Hot-Carrier Relaxation by Learning of Nonadiabatic Hamiltonians with Graph Neural Networks.

An electron-vibrational Hamiltonian fully encodes corresponding quantum dynamics; however, extracting the dynamics still relies on time and memory-consuming trajectory-based nonadiabatic molecular dynamics (NAMD) simulations, typically stochastic surface hopping. Here, we develop a general graph neural network, artificial intelligence ab initio NAMD (AINAMD) that establishes an end-to-end mapping from Hamiltonian to hot carrier relaxation dynamics. We validated the generality of AINAMD across multiple materials, including a zero-dimensional Si quantum dot (QD), a one-dimensional carbon nanotube (CNT), a two-dimensional twisted MoS/WS bilayer, and a three-dimensional soft-lattice MAPbI perovskite. With only 10% training data, AINAMD can rapidly and accurately generate picosecond energy decay curves for hot electron and hot hole relaxation for the remaining 90% Hamiltonians, while delivering a computational speed-up of more than 6 orders of magnitude compared to standard CPU-based NAMD simulations. Moreover, AINAMD can also map directly the Hamiltonian to the carrier relaxation time, bypassing generation of the energy decay curves and demonstrating the ability to handle complex NAMD tasks. Further, by projecting high-dimensional Hamiltonian encoding features into a two-dimensional space with unsupervised learning, we demonstrate that AINAMD can effectively distinguish Hamiltonian types, verifying its ability to identify a particular system (QD, CNT, MoS/WS and MAPbI) and a charge carrier (electron or hole). Overall, the developed AINAMD approach provides a novel computational methodology and a conceptual framework for accelerating NAMD simulations with machine learning by many orders of magnitude.