Numerical modeling and machine learning insight into potassium-based lead-free $$\text {K}_{2}\text {AgSbBr}_{6}$$ perovskite solar cells using SCAPS-1D

Lead-based perovskite solar cells face significant challenges related to toxicity and long-term stability despite achieving high power conversion efficiencies (PCEs). This study investigates the lead-free, potassium-based double perovskite \\(\\text {K}_{2}\\text {AgSbBr}_{6}\\) as an absorber for thin-film solar cells using SCAPS-1D simulations integrated with a machine-learning-assisted optimization framework. A planar device architecture (ITO/ \\(\\text {C}_{60}\\) / \\(\\text {K}_{2}\\text {AgSbBr}_{6}\\) / \\(\\text {Cu}_{2}\\text {NiSnS}_{4}\\) /Pt) was evaluated across a dataset of 1,050 configurations generated by systematically varying absorber thickness, acceptor density, bulk defect density, and interface defect density. Nine machine learning algorithms were evaluated on a 75/25 train/test split. Tree-based ensemble methods yielded the highest predictive accuracy, with XGBoost achieving a test \\(R^2\\) of 0.9995, an RMSE of 0.0617%, and a 5-fold cross-validation \\(R^2\\) of 0.9990%. Combining hyperparameter tuning with multi-objective optimization identified optimal device parameters, which were independently verified in SCAPS-1D. The optimized configuration yielded a simulated increase in PCE from 21.14% to 24.49%. A comprehensive Shapley Additive exPlanations (SHAP) analysis quantified parameter contributions, identifying acceptor density and bulk defect density as the primary drivers of device efficiency. These computational results reflect theoretical bounds under the modeled physical assumptions and material parameters. Overall, this work offers a physically interpretable framework for optimizing lead-free \\(\\text {K}_{2}\\text {AgSbBr}_{6}\\) solar cells.