Publication date: 16th July 2025
Large-scale, finite-temperature molecular dynamics of metal halide perovskites are critical for understanding their structural and dynamical properties under realistic operating conditions. While ab initio molecular dynamics (AIMD) provides reliable insights, its computational versatility is limited by system size and timescale. Machine-learning force fields have recently emerged to enable large-scale finite-temperature simulations with impressive accuracy, offering promising alternatives. The performance of machine-learning force fields has been well established for various systems. However, their reliability in capturing complex dynamics in metal halide perovskites remains an open question and requires careful benchmarking.
We compare AIMD simulations of a prototypical perovskite system with machine learning force fields built by a commonly used implementation of a Gaussian Process regression in the Vienna Ab initio Simulation Package (VASP) and the message-passing neural network with the atomic cluster expansion (MACE). We focus on structural distribution functions as indicators of local configuration rigor and discuss their differences and implications for practical applications among the different models.