<p>The development of high-efficiency perovskite solar cells (PSCs) demands a comprehensive control of multi-scale factors that influence device performance. In recent years, artificial intelligence (AI), represented by machine learning (ML), has rapidly become a key tool for the design and optimization of PSCs. However, current ML models often oversimplify the design of PSCs at the device level, making it difficult to capture the complexity of their multi-scale features. Moreover, they are constrained by relatively small and specialized datasets, which limits their generalizability across diverse device architectures and fabrication methods. In this work, we developed a full-process AI framework based on over 20,000 experimentally measured PSC samples and approximately 260 multi-scale features. This framework offers significant advantages in both sample diversity and feature richness. It combines material selection, fabrication processes, and environmental factors to provide a more accurate and comprehensive optimization solution for PSCs. We addressed challenges from data diversity and heterogeneity through feature engineering and model training, which results in a highly generalizable PSC performance prediction model with comparable prediction error to small-scale models. The framework enables precise optimization of specific features for any PSCs, and provides valuable insights for designing high-performance photovoltaic devices.</p>