Abstract
Halide perovskites (HaPs) have emerged as promising new materials for a wide range of optoelectronic applications, notably solar energy conversion. These materials are well known to exhibit significant dynamical effects even at room temperature, which affect both their electronic properties and their long-term stability. Molecular dynamics (MD) simulations can provide significant insights into such effects. However, long time scale simulations require both accuracy and scalability. The latter is an issue for first principles methods and the former is challenging for classical force fields. Machine-learned force fields (MLFF) are a promising avenue for bridging across this seeming contradiction. Here, we apply the gradient-domain machine learning approach, using CsPbBr3 as an example. We find that training based on room temperature density functional theory (DFT) data fails to generate an MLFF that provides long-term stable MD, owing to an insufficient sampling of rare events in the training set. We show that this problem is resolved by using a temperature ensemble (TE) method, which can be generated in parallel and yields a combined data set based on MD trajectories from a variety of temperatures. The MLFF model based on the TE method yields high accuracy for long-term simulations, showing remaining errors of the same magnitude of inherent errors in the DFT calculation.