Lead halide perovskites (CsPbI₃) have received much attention as a promising candidate for a high-efficiency solar cell. However, a limited lifespan is a critical weak point, originating in the instability of perovskite phase at room temperature. Mixing halogen elements is one of the engineering techniques to enhance the stability of the perovskite structure. In the previous researches, it was shown that mixing Br is able to alleviate the issue despite some loss in power conversion efficiency. Then, several computational studies have tried to identify thermodynamically stable mixed halides using density functional theory (DFT) calculations. In the previous work, however, the stable structures were searched based on known prototypes of CsPbI₃. This assumption excludes a possibility of finding uncharted phases in mixed-halide perovskites. The heuristic searching methods with DFT calculations is one of the choices to overcome the limitation but DFT calculation is computationally too expensive to estimate the energies of numerous candidates during crystal structure prediction (CSP).

In this study, we perform CSP to find stable mixed-halide perovskites (CsPbI_xBr_3-x, CsPbI_xCl_3-x) using SPINNER, a recently developed in-house code for CSP using a neural network potential. Using the code, we can explore the most stable mixed-halides structures at each stoichiometry beyond prototypes with much less costs than the DFT based CSP. The perovskite-like γ phases are usually less stable than the non-perovskite δ phases. Nonetheless, by mixing Cl or Br, the energy differences between the phases are reduced and in some compositions the most stable phase is changed. We also identified that the γ phase mixed-halides are expected to have good optical properties comparable to the γ-CsPbI₃. We hope that our findings can contribute to extending the lifespan of the inorganic perovskite solar cells.