Nonadiabatic Dynamics in Two-Dimensional Perovskites Assisted by Machine Learning

An exploration of the on-the-fly non-adiabatic couplings (NAC) for nonradiative relaxation and recombination of excited states in 2D Dion-Jacobson (DJ) lead halide perovskites (LHP) is accelerated by a machine learning approach to ab initio molecular dynamics. Molecular dynamics (MD) of nanostructures composed of heavy elements is performed with use of machine learned force-fields (MLFF), as implemented in Vienna Ab initio Simulation Package (VASP). The force field parameterization is established using on-the-fly learning, which continuously builds a force field using ab initio MD (AIMD) data. At each time step of MD simulation, the total energy and forces are predicted based on the MLFF and if the Bayesian error estimate exceeds a threshold an ab initio calculation is performed, which is used to construct a new force field. Model training and evaluation were performed for a range of DJ-LHP models of different thickness and halide composition. The MLFF-MD trajectories were evaluated against AIMD trajectories to assess level of discrepancy and error accumulation. To examine the practical effectiveness of this approach we have used the MLFF-based MD trajectories to compute NAC and excited-state dynamics. At each stage, results based on machine learning are compared to traditional ab initio based electronic dissipative dynamics. We find that MLFF-MD provides comparable results to AIMDs when the MLFF is trained in a NpT ensemble.