Phase Stability and Transformations in Lead Mixed Halide Perovskites from Machine Learning Force Fields

Abstract

Lead halide perovskites (APbX₃) offer tunable optoelectronic properties, but feature an intricate phase stability landscape. Here, we employ on-the-fly data collection and an equivariant message-passing neural network potential to perform large-scale molecular dynamics of three prototypical lead mixed-halide perovskite systems: CsPbX₃, MAPbX₃, and FAPbX₃. Integrating these simulations with the PDynA structure analysis toolkit, we resolve both equilibrium phase diagrams and the dynamic structural evolution under varying temperatures and halide mixtures. Our findings reveal that A-site cations strongly modulate tilt modes and phase pathways: MA⁺ effectively “forbids” the β-to-γ transition in MAPbX₃ by requiring extensive molecular rearrangements and crystal rotation, whereas the debated low-temperature phase in FAPbX₃ is predicted to be best represented as an Im3̅ (a⁺a⁺a⁺) cubic phase. Additionally, small changes in halide composition and arrangement, from uniform mixing to partial segregation, alter octahedral tilt correlations. Segregated domains can even foster anomalous tilting modes that impede uniform phase transformations. These results highlight the multiscale interplay between the cation environment and halide distribution, offering a route for tuning perovskite architectures toward improved phase stability and control.