Developing high-performance organic photovoltaic (OPV) materials is a time-consuming and costly process. This study introduces DeepAcceptor, a deep learning framework that accelerates the discovery of efficient small molecule acceptor materials. DeepAcceptor utilizes a custom deep learning model (abcBERT) that leverages graph representation learning and Bidirectional Encoder Representations from Transformers (BERT) to predict power conversion efficiencies (PCEs) from molecular structures. The model is pre-trained on a computational dataset and fine-tuned on an experimental dataset, achieving superior PCE prediction accuracy (MAE = 1.78, R² = 0.67). A molecular generation and screening process, integrated into a user-friendly interface, identifies three high-performance acceptor candidates, with the best achieving a PCE of 14.61% upon experimental validation.

Jinyu Sun, Dongxu Li, Jie Zou, Shaofeng Zhu, Cong Xu, Yingping Zou, Zhimin Zhang, Hongmei Lu