Integrated machine learning framework for the Co-optimization of efficiency and stability in tin-based perovskite solar cells

Tin-based perovskite solar cells (TPSCs) offer a promising alternative for sustainable and lead-free photovoltaics. However, their low power conversion efficiency (PCE) and stability hinder their commercialisation. Finding the relations between fabrication parameters is challenging, which can affect the performance of the TPSCs. This work presents a novel integrated machine learning (ML) framework that combines two ML models to identify complex relationships among fabrication process parameters, thereby enhancing TPSC performance. Selecting a suitable ML model manually is difficult. This work developed a framework for selecting the best models based on the coefficient of determination (R 2 ). Random forest models outperformed all the regression models. These random forest models were trained on two experimental datasets of >160 features of the fabrication process to predict both PCE and stability accurately. A weighted combined pipeline was introduced to rank the device, unlike traditional single metric optimization. These models achieve high R² values of 0.84 for PCE and 0.87 for stability. Shapley additive explanations (SHAP) analysis reveals key features, including the antisolvent volume, perovskite precursors, and charge-transport-layer processing conditions. The combined score of distribution analysis and t-distributed stochastic neighbour embedding (t-SNE) visualization demonstrates the clustering of high-performance devices, suggesting reproducible fabrication pathways with enhanced PCE and stability. This work establishes a reproducible ML-driven framework for multi-objective optimization in TPSCs. Shifting from a PCE-centric design to a balanced dual-objective optimization provides actionable guidelines for fabricating TPSCs with enhanced PCE and stability.