Optimization of CsPbI₃-Based All-Inorganic Perovskite Solar Cells Using Machine Learning: A Framework Supported by SCAPS-1D and SHAP

Abstract

This paper introduces a machine learning (ML)-aided optimization framework aimed at improving the performance of all-inorganic perovskite solar cells (PSCs) utilizing CsPbI₃. More than 55,000 high-throughput simulations were performed with SCAPS-1D to assess how layer thickness and defect density influence device performance. Through the application of XGBoost and SHAP (SHapley Additive explanations) analysis, it was determined that the thickness of the perovskite layer and the intrinsic trap density are the key parameters that significantly impact power conversion efficiency (PCE). An ideal perovskite thickness of around 2 μm improved photon absorption and current generation, while keeping defect densities below 1 × 10⁵ cm⁻³, which significantly enhanced carrier diffusion. The XGBoost model demonstrated outstanding predictive capabilities (R² = 0.999, RMSE = 0.0010), allowing for precise and swift PCE estimation. With this machine learning framework, the predicted PCE rose from 15.15% to 19.16%. The suggested method promotes efficient device design, minimizes experimental iterations, and showcases the potential of data-driven techniques for wider applications in photovoltaics, batteries, and energy systems.