Adaptive Transfer Crystal Graph Convolutional Networks for Accelerated Discovery of Organic Perovskite Photovoltaic Materials

Perovskite solar cells (PSCs) exhibit outstanding photovoltaic performance, but the discovery of new hybrid organic-inorganic perovskites (HOIPs) remains constrained by the scarcity of high-quality data. Given the strong structure-property correlations in crystalline materials, we adopt the Crystal Graph Convolutional Neural Network (CGCNN) to learn intrinsic structural representations. To overcome the limitations of small data sets, we propose an Adaptive Transfer Learning-based CGCNN (ATL-CGCNN), which integrates transfer learning and adaptive fine-tuning within the CGCNN framework. The model is first pretrained on a large-scale crystal data set, then transferred to the HOIP domain, and subsequently fine-tuned using high-confidence samples identified based on prediction errors. Using ATL-CGCNN, we predicted the HSE06 band gaps of 160 unreported HOIPs. Seven candidates with promising photovoltaic potential were identified and validated using Density Functional Theory (DFT), showing good agreement between predicted and calculated values. Additionally, machine learning analysis confirmed that elemental features had limited contribution to bandgap prediction. This approach demonstrates a data-efficient and accurate strategy for accelerating the discovery of novel HOIP materials.