Transfer Learning with Graph Neural Networks for Experimental Prediction of Organic Photovoltaic Efficiency

Abstract Herein, we report a transfer learning approach that enhances power conversion efficiency (PCE) predictions in organic photovoltaics (OPVs) by resolving the critical issue of missing descriptors in real-world databases. Specifically, we address the challenge of missing frontier molecular orbital descriptors (HOMO and LUMO) within an extensive data set of 1346 experimental OPV records established by the Korea Research Institute of Chemical Technology (KRICT). To resolve this, a graph neural network (GNN) was pretrained on 100,000 computational molecules from the Harvard Clean Energy Project (CEP). Crucially, this model was utilized to impute the missing experimental HOMO/LUMO levels, effectively bridging the computational–experimental domain gap. By incorporating these successfully imputed electronic descriptors with device-level processing parameters, we achieved relative R2 gains of approximately 8–25% across most machine learning models, with Random Forest yielding the top performance (R2 = 0.837). By integrating GNN-based representation learning with transfer learning from experimental data, the proposed approach is not restricted to specific regression models and can be extended to various molecular- and device-level properties as a framework, providing a flexible and data-efficient strategy for the design of high-performance OPV systems.