Machine learning-guided scalable manufacturing of high-efficiency perovskite solar modules

Abstract

The transition from lab-scale demonstrations to industrial production of perovskite photovoltaics faces significant challenges in scalability, particularly the persistent area-dependent efficiency loss. Traditional trial-and-error approaches struggle with the multi-dimensional parameter space involving composition, processing, and architecture variations. This study presents a machine learning framework addressing the critical challenge. By analyzing 332 experimental data points through optimized XGBoost models (achieving R² = 0.8151, RMSE = 1.9699% for PCE prediction), we identify inverted (p-i-n) architectures and Slot-Die coating deposition as optimal for scalability, demonstrating significantly reduced efficiency decay rates compared to formal structures and the other five processes. SHAP analysis reveals the dominant roles of device area, perovskite composition, and device structure in performance optimization. Furthermore, we establish formamidinium-dominant compositions with minimal halide mixing as universally favorable across scaling scenarios. The research provides both fundamental insights into crystallization dynamics and interfacial requirements for large-area fabrication, and practical guidelines for material selection and process optimization, effectively bridging data-driven prediction with experimental development to accelerate the commercialization of perovskite solar modules.