Metal-halide perovskites exhibit excellent optoelectronic properties, making them suitable for various applications. In this work, we propose a novel approach, employing the message-passing neural networks (MPNNs), to model the potential energy surface (PES) of metal-halide perovskites. The MPNN model can learn the PES of perovskites with an error of less than 1 meV per atom, which is comparable to ab initio calculations. With the accurate and fast prediction of the MPNN model, we perform large-scale molecular dynamics simulations for ${\mathrm{CsPbBr}}_3$ and ${\mathrm{Cs}}_2{\mathrm{AgBiBr}}_6$, successfully identifying their temperature-dependent octahedron tilting patterns, ion displacement, and structural phases that are crucial to obtaining their electronic structure correctly. We also observe significant Ag migration and vacancy defects in ${\mathrm{Cs}}_2{\mathrm{AgBiBr}}_6$, which are further investigated using ab initio calculations, offering insights into its ambient stability. This work provides a comprehensive understanding of the structural phase transition in metal-halide perovskites, as well as information regarding ion migration.

Close

PrevNext

View figure in article

• Structural phase transition
• Perovskites
• Density functional calculations
• Machine learning