The urgent need to transition toward sustainable energy sources has positioned perovskite solar cells (PSCs) as a leading candidate for next-generation photovoltaics. Among them, formamidinium lead iodide ( \(\textrm{FAPbI}_3\) ) based devices have demonstrated power conversion efficiencies (PCEs) exceeding 25% with the potential for low-cost fabrication. However, structural defects such as pinholes, \(\textrm{PbI}_2\) accumulation, and grain boundary irregularities significantly compromise their efficiency, stability, and long-term reliability. Conventional defect characterization using scanning electron microscopy (SEM) is labor-intensive, subjective, and unsuitable for large-scale quality control, underscoring the need for automated, high-precision detection strategies. In this study, we propose a multi-model deep learning framework for automated defect classification in mixed-dimensionality \(\textrm{FAPbI}_3\) perovskite films. The framework targets five critical defect types: pure 3D perovskite, 3D perovskite with \(\textrm{PbI}_2\) excess, 3D perovskite with pinholes, 3D–2D mixed perovskite, and 3D–2D mixed perovskite with pinholes. Three complementary architectures are benchmarked: ResNet50V2 and DenseNet169 for high-accuracy classification, and YOLOv9 for real-time detection with computational efficiency. Extensive data augmentation and transfer learning were employed to mitigate dataset scarcity, enabling robust feature extraction from a limited set of 2,380 SEM images. The results show that ResNet50V2 and DenseNet169 achieved a test accuracy of 96.7% and weighted F1-score of 0.966, while YOLOv9, though moderate in accuracy (45.0%), demonstrated exceptional computational efficiency with an 8-minute training time. The proposed framework not only enables precise identification of morphological defects but also supports scalable quality control in PSC manufacturing. Furthermore, the deployment of the trained models as an interactive Streamlit-based web application demonstrates its practical utility for real-time laboratory and industrial adoption. These findings highlight the potential of deep learning–driven defect analysis to accelerate the optimization and commercialization of perovskite solar cell technologies.
