Deep Learning the Defect Landscape: Long-Timescale Vacancy Dynamics in Lead-Free Double Perovskites

Abstract Point defects strongly influence charge transport and recombination, as well as structural stability in metal halide perovskites, yet their geometric evolution and impact on dynamic optoelectronic properties are difficult to understand. In lead-free halide double perovskites (HDPs) such as Cs2AgBiBr6, lattice anharmonicity and thermal disorder can potentially introduce dynamically active defect environments. Here, we combine first-principles defect calculations with machine-learning-model-assisted long-timescale molecular dynamics (MD) to investigate vacancy defects in Cs2AgBiBr6. Static calculations determine the energetics of defect formation and charge-state stability, identifying the bromide vacancy as the dominant electronically active intrinsic defect with deep transition levels. The silver vacancy is thermodynamically accessible but electronically benign. Nanosecond-long MD simulations reveal that vacancies sample a fluctuating ensemble of local structures characterized by persistent octahedral distortions, continuous symmetry breaking, and intermittent metastable configurations. While the global metal-halide framework remains intact, halide vacancies induce localized cation rearrangements enduring over nanosecond timescales, highlighting the inherently dynamic nature of point defects in HDPs. Such fluctuating defect geometries evidently impact the transient electronic structures of Cs2AgBiBr6. The study highlights the subtle role of dynamic defect configurations in governing the optoelectronic properties of perovskites under ambient conditions.