Developing ChIMES machine-learned interatomic potentials for cesium lead halide perovskites

Organic-inorganic hybrid perovskites offer impressive photovoltaic performance but are limited by poor stability. To address these issues, research has shifted toward cesium-based inorganic lead halide perovskites, valued for their complex crystal structures, varied defect landscapes and dynamic halide transport, all central to their optoelectronic properties and stability. Understanding atomic-scale phase behavior, ion migration and segregation is essential to improving performance, yet experimental access to these processes is limited. Here we develop a machine-learned interatomic potential for cesium-lead halide perovskites using the ChIMES physics-informed modeling framework. This approach constructs accurate potential energy surfaces from binary to multi-component interactions via compact, reusable parameter "bricks." Our training set combines randomly perturbed structures for efficient model generation with distinct melted-phase configurations for accuracy near phase transitions. Large-scale molecular simulations with these models enable probing of diffusion pathways, halide segregation and framework distortions under variable conditions. The modularity of ChIMES also supports extension to solvent interactions and degradation mechanisms. This work establishes a foundation for predictive, atomistic simulations to address open questions in the fundamental physics of perovskite materials.