Machine Learning Accelerated Non‐Adiabatic Molecular Dynamics Elucidates Local Polarization Effects on Non‐radiative Recombination in Halide Perovskites

Non‐radiative recombination is a critical factor limiting the optoelectronic performance of halide perovskites, yet how local polarization induced by charge redistribution regulates this process remains unclear. To gain deeper insight while reducing the high computational cost of conventional non‐adiabatic molecular dynamics (NAMD), we developed Hefei‐NAMD‐S, a machine learning (ML) framework constructed using stacked models. The relative errors of the ML predicted non‐adiabatic coupling and pure‐dephasing time, calculated with respect to the first‐principles values, remain below 1.10%, while the total computational time is reduced by approximately 78%, demonstrating the accuracy and efficiency of the proposed framework. NAMD simulations further reveal that the B‐site local polarization is involved in regulating the non‐radiative recombination process. The rubidium‐substitution‐doped system (FA Rb ) and the cesium interstitial doped system exhibit recombination times of about 280 ns, nearly 2.8 times that of the pristine system, and the enhanced performance of FA Rb has been supported by previous experimental evidence. These results identify B‐site local polarization as one of the important factors in suppressing non‐radiative recombination and provide a theoretical foundation for designing perovskite materials through local polarization modulation.