Deep learning accelerated high-throughput screening of organic solar cells

Abstract

Organic solar cells (OSCs) have attracted impressive interest due to their advantages of flexibility, light weight, non-toxicity, and transparency. However, it is not feasible to explore the gigantic chemical space purely through experimental approaches. Herein, a framework based on deep learning models was developed to establish a direct relationship between the molecular structure and device efficiency. Eight graph neural network models were applied to a newly established dataset consisting of 1060 realistic organic donor/acceptor (D/A) pairs to predict the power conversion efficiency (PCE). It is notable that the data fidelity and unity were enhanced by manually collecting reported upper limit values. Among these models, the graph attention network (GAT) model exhibited the best performance (r = 0.74, RMSE = 2.63), comparable to previous studies but with significantly lower computational costs. The deep learning models were then employed to predict and screen a developed dataset, consisting of 45 430 possible donor–acceptor combinations, in which the donors and acceptors were sourced from cases with PCE values exceeding 10%. The average predicted PCE range is from 3.61 to 17.43%, and only 2320 (5.1%) D/A pairs were predicted to achieve PCE values above 15%, indicating the importance of theoretical calculations. Several pairs were identified with high PCE values, which were reported in the literature but not in the dataset used in this work. Furthermore, electronic structure calculations were performed on potential candidates to gain insights into the materials, further validating the reliability of the predictions. Our work then provides an efficient workflow to accelerate high-throughput screening of OSC materials, aiding in the development of highly efficient OSCs.