Transfer learning accelerated discovery of conjugated oligomers for advanced organic photovoltaics

Abstract

Machine learning accelerates material discovery which includes selection of candidate small molecules and polymers for high-efficiency organic photovoltaic (OPV) materials. However, conventional machine learning models suffer from data scarcity for conjugated oligomers, crucial for OPV material production. To address this challenge, transfer learning within a graph neural network was introduced to reduce the data requirement while accurately predicting the electronic properties of the conjugated oligomers. By leveraging on transfer learning using original conjugated oligomer data and pre-trained models from the renowned PubChemQC dataset, the limitations posed by insufficient data were mitigated. The models in this study achieved a low mean absolute error, ranging from 0.46 to 0.74 eV, for the HOMO, LUMO, and HOMO–LUMO gap. An original candidate dataset of 3710 conjugated oligomers was constructed for materials discovery, and a high-throughput screening pipeline was developed by integrating the models with density functional theory. This pipeline effectively identified 46 promising conjugated oligomer candidates, showcasing its effectiveness in accelerating the discovery of advanced materials for organic photovoltaics. These results demonstrated the potential of the approach used in this study to overcome data scarcity while accelerating the discovery of new innovative materials in organic electronics.