Machine Learning Approaches to Optimize the Performance of the Novel Stable Lead-free Heterojunction CsGeI₃/CsSn(I_1−xBr_x)₃-based Perovskite Solar Cell

This work aims to design and predict the performance of a novel heterojunction perovskite solar cell (PSC) based on CsGeI₃/CsSn(I_1−xBr_x)₃ using machine learning (ML). Electrical parameters including open-circuit voltage (V_OC), short-circuit current (J_SC), fill factor (FF), and power conversion efficiency (PCE) are generated using SCAPS-1D due to its high accuracy in matching experimental results, and are used to build two ML models based on polynomial regression (PR) and XGBoost (XGB) by considering the thickness H₁ (CsSn(I_1−xBr_x)₃) and H₂ (CsGeI₃) and the bromine composition (x) as parameters. The polynomial degree and maximum depth are considered as parameters for PR and XGB. Performance metrics including training and testing scores, correlation (r), and mean squared error (MSE) are calculated to investigate the accuracy, while the cross-validation (CV) scores are used to determine the stability and the overfitting. The results reveal the best performance for \(R^{2}\) $R^2$ , \(r\) $r$ , \(MSE\) $\mathit{MSE}$ , and CV for PR of 0.9969, 0.9915, 0.0245, and 0.9676, respectively, while those for XGB are 0.9990, 0.9991, 0.0005, and 0.9864, respectively. Moreover, the best polynomial degree and maximum depth are 9 and 5, respectively. We also reveal that the dependence of the optoelectronic parameters in terms of bromine composition in CsSn(I_1−xBr_x)₃ creates a powerful internal electrostatic field between the two absorbers for low bromine composition, which favors the extraction of the carriers out of the device. Considering both stability and performance, this prototype can achieve PCE of 29.62%, 29.54%, and 29.61% for actual, PR, and XGB, respectively, when H₁, x, and H₂ are 1 μm, 0.6, and 0.2 μm, respectively, compared to 21.91% and 17.61% for existing CsSn(I_1−xBr_x)₃ and CsGeI₃ homojunction.