A Hybrid Physics-Guided Residual Multi-Task Neural Network Framework for Predictive Modeling of Power Conversion Efficiency and Static Electrical Stability in Perovskite Solar Cells
This study proposed a hybrid physics-guided residual multi-task neural network framework based on a large-scale simulated J-V dataset, comprising approximately 2.47 million samples. Physics-based descriptors were constructed, and residual learning with a heuristic empirical baseline was adopted through carrier transport and interfacial contact mechanisms. The model achieved a test coefficient of determination (R2) of 0.8538 while reducing the prediction Root Mean Square Error (RMSE) by approximately 80% relative to the empirical baseline alone. For static electrical stability prediction with shunt resistance, the model obtained a test of 0.1113. However, the performance was limited by the static proxy and the lack of aging data. SHapley Additive exPlanations (SHAP) analysis indicated consistency with known device physics, especially the key role of work function difference. The proposed framework offers experimentalists a rapid and interpretable prescreening tool for device parameter optimization, and can be used to accelerate the discovery of high-efficiency and stable PSCs.