Enhancing perovskite solar cell efficiency and stability: a multimodal prediction approach integrating...

The performances of perovskite solar cells (PSCs) are significantly influenced by material composition, processing techniques, and microstructure, all of which critically impact photovoltaic conversion efficiency (PCE). Traditional machine learning approaches often overlook multi-parameter coupling effects, leading to incomplete analyses. To tackle this challenge, we developed a multimodal model that integrates SEM-derived microstructural features, material composition, and processing parameters. Our model utilizes a feature extraction network with a Convolutional Block Attention Module (CBAM) and an adaptive feature fusion module, achieving an R² value of 0.84 (RMSE: 1.89) for PCE prediction and an R² value of 0.95 (RMSE: 0.77) for bandgap estimation. Among the tested algorithms, the Gradient Boosting Regressor demonstrated superior performance. We also used machine learning to evaluate PSC stability, an essential factor for renewable energy applications. The model classified stability categories with AUC scores of 0.76 (moderately stable), 0.81 (very stable), and 0.78 (unstable), indicating robust performance with room for refinement. This research emphasizes the significant direct relationship between larger perovskite grain sizes and higher PCE, offering actionable insights for material optimization. The integrity of our experimental validation is supported by comprehensive testing across different device sizes and mass production verification, demonstrating the scalability of our framework. By integrating materials science and machine learning, this study advances the development of efficient, durable, and scalable PSCs, contributing to the broader adoption of renewable energy technologies.