Publication date: 15th December 2025
Compositionally complex halide perovskites provide a lead-free platform for engineering structure–property coupling through the competition between octahedral tilting and B-site off-centering. However, their finite-temperature phase behavior is difficult to resolve with conventional first-principles approaches because it requires large supercells and long simulations to capture slow, collective distortions. Here, we develop a machine learning interatomic potential (MLIP) within the neuroevolution potential (NEP) framework [1] to model phase transitions and local symmetry breaking in mixed B-site perovskites CsGexSn1–xBr3. The MLIP is trained on density functional theory (DFT) data spanning diverse Ge/Sn chemical environments, achieving near-DFT accuracy while enabling long-time, large-scale Monte Carlo (MC) and molecular dynamics (MD) simulations.
Benchmarking the end members with MD and group-theory-based structural analysis [2] establishes distinct instability limits that define the mixed system. For CsGeBr3, heating induces a sequence from a polar, tilted monoclinic (Cc) phase to an intermediate polar rhombohedral (R3m) phase with vanishing long-range tilts, and finally to a cubic (Pm-3m) phase [3]. In contrast, CsSnBr3 follows a tilt-dominated pathway from orthorhombic (Pnma) to tetragonal (P4/mbm) and then to a cubic phase [4].
Extending to CsGexSn1–xBr3, combined MC/MD simulations across composition map how Ge/Sn mixing tunes the balance and coupling between off-centering and tilt networks. We present a composition–temperature phase diagram that identifies phase boundaries and crossover regimes between Ge-driven polar distortions and Sn-stabilized tilts. Overall, this work demonstrates that NEP-based MLIP enables predictive phase diagrams and engineering of local instabilities in compositionally complex, lead-free halide perovskites, with direct relevance to stabilizing targeted structural states in nanocrystal synthesis and tailoring optoelectronic functionality.
