Hybrid modeling with supervised learning and SCAPS simulations for performance analysis of Cs0.17FA0.83PbI3−xBrx perovskite solar cells

In this study, we integrate numerical simulations with supervised learning to predict the performance of Cs_0.17FA_0.83PbI${}_{3-x}$Br${}_x$-based perovskite solar cells. A dataset of 3,240 points was generated using SCAPS by varying absorber thickness ($t$), defect density ($N_t$), and donor density ($N_d$) across six bromine compositions (x = 0, 0.5, 1, 1.5, 2, and 2.5). Four Random Forest models were trained to predict power conversion efficiency (PCE), open-circuit voltage (Voc), short-circuit current density (Jsc), and fill factor (FF), achieving test R${}^2$ values consistently above 0.99. The corresponding RMSE values were 0.197%, 0.008 V, 0.317 mA/cm${}^2$, and 1.181% for PCE, Voc, Jsc, and FF, respectively. To validate generalization, the models were tested on Cs_0.17FA_0.83PbI_2.6Br_0.4 composition, showing strong agreement with simulations. Feature correlation analysis identified $N_t$ and $N_d$ as key performance factors. This approach can be extended to other perovskite compositions, device configurations, transport layers, and alternative ML techniques for improved generalization.