Data-Driven Insights Into Feature Engineering and Efficiency Prediction in Lead-Free Rb-Based Perovskite Solar Cells

Abstract

Lead-free perovskite solar cells (PSCs) require systematic identification of device parameters that govern their performance. Here, a machine learning (ML)-assisted SCAPS-1D framework is developed to analyze and optimize the simulated performance of RbGeI₃-based PSCs across different transport-layer configurations. Random forest (RF) and Decision Tree (DT) models exhibited the strongest 10-fold cross-validation (CV) performance, with R² values of 0.9843 ± 0.0044 and 0.9774 ± 0.0079, respectively. SHapley Additive exPlanations (SHAP) analysis identified active-layer thickness as the dominant predictor of power conversion efficiency (PCE), followed by hole-transport layer (HTL)-related parameters and operating temperature. Systematic electron-transport layer (ETL)/HTL screening identified CuSbS₂ as a promising HTL, yielding a maximum simulated PCE of 28.32% within the investigated parameter space. Importantly, held-out ETL/HTL tests revealed substantial variation in out-of-distribution (OOD) generalization, demonstrating that high random CV accuracy does not necessarily translate to reliable prediction of unseen material combinations. The results establish an interpretable computational framework for parameter screening and device optimization of RbGeI₃-based PSCs, while highlighting the need for experimental and independent computational validation.