Comparative convolutional neural networks for perovskite solar cell PCE predictions

Abstract Imaging offers a fast and accessible means for spatial characterization of halide perovskite photovoltaic materials, yet extracting optoelectrical properties—such as power conversion efficiency (PCE)—remains challenging. This study presents a deep learning methodology that correlates optical reflective images of perovskite solar cells with their PCE by focusing on image differences rather than absolute visual features. The approach predicts relative changes in PCE by comparing images of the same device in different states (e.g., before and after encapsulation) or against a reference image. This comparative technique significantly outperforms traditional methods that attempt to directly infer PCE from a single image. Furthermore, it demonstrates high effectiveness in low-data regimes, using only 115 samples. By leveraging convolutional neural networks (CNNs) trained on small datasets, the method offers an adaptable and scalable solution for device characterization. Overall, the comparative approach enhances the accuracy and applicability of machine vision in perovskite solar cell analysis.