Abstract This work investigates the potential of inorganic perovskites AgBiSCl₂ and Al₂Cu₂Bi₂S₃Cl₈ as absorber layers in perovskite solar cells, followed by the application of supervised machine learning models. Extensive exploration and optimization of device architectures FTO/SnO₂/AgBiSCl₂/Spiro-OMeTAD/Au and FTO/SnO₂/Al₂Cu₂Bi₂S₃Cl₈/Spiro-OMeTAD/Au are conducted, involving variations in absorber layer thickness (d), bulk defect density (N_t), and carrier mobility (μ _n,p). The AgBiSCl₂-based device achieves an optimized conversion efficiency of 10.06%, while the Al₂Cu₂Bi₂S₃Cl₈-based device achieves 12.27%. To train different machine learning models, 1600 datasets are collected for each device, and Neural Networks (NN), Random Forests (RF), and XGBoost (XGB) models are employed. The performance parameters, evaluated using mean squared error (MSE) and high R-squared (R2) values, demonstrate that XGB performs the best, achieving an MSE of 0.210 and R2 of 97.1% for AgBiSCl₂ and 0.671 and 90.6% for Al₂Cu₂Bi₂S₃Cl₈. Additionally, the impact of each variable (d, N_t, and μ _n,p) on the output is analyzed using Shapley Additive Explanations (SHAP) plots for each model. The results presented in this study pave the way for the advancement of perovskite material-based solar cells without relying on complex optoelectronic semiconducting equations and device simulators.