Active learning and explainable machine intelligence for lead-free chalcogenide perovskite photovoltaics: Gaussian process Bayesian optimization and SHAP interpretability applied to CaZrSe 3/CdZnS solar cells

Abstract

The global transition toward lead-free photovoltaics has intensified interest in chalcogenide perovskites with the formula ABX 3, where calcium zirconium triselenide (CaZrSe 3) stands out for its direct bandgap of approximately 1.35 eV, chemical stability under ambient conditions, and complete absence of toxic elements. Yet the device physics of CaZrSe 3 heterojunction solar cells remain poorly mapped, and the conventional approach of exhaustive parametric sweeps in SCAPS-1D, while rigorous, places a prohibitive computational demand on research groups seeking rapid, systematic optimization. This work addresses this bottleneck by integrating a Gaussian process (GP) surrogate model, guided by the expected improvement acquisition function, into the SCAPS-1D simulation workflow for CaZrSe 3/CdZnS heterojunction solar cells. A Latin hypercube initial sample of 403 simulations conditioned the GP; 201 sequential Bayesian optimization iterations then identified the global power conversion efficiency (PCE) optimum of 10.4927%, with Voc=0.5388 V, Jsc=27.43 mA cm −2, and FF = 70.99%, using 604 total evaluations, representing a 95.0% reduction relative to the full 12 100-point reference grid. Concurrently, an ensemble of random forest, XGBoost, and artificial neural network surrogates trained on the complete 12 100-point dataset achieves R2>0.997 for all performance targets, and SHapley Additive exPlanations analysis decomposes the PCE response surface to reveal that absorber acceptor doping ( NA) is the primary determinant of device performance, while second-order interaction values quantify the coupled recombination dynamics of NA and ETL donor density ND. The optimal device specification an absorber thickness of 0.70 μm, NA=5.00×1015 cm −3, an electron transport layer (ETL) thickness of 0.10 μm, and an ETL donor density of ND=1.00×1018 cm −3 provides fabrication-targeted design rules grounded in quantitative physical reasoning rather than single-variable parametric intuition.