High-Throughput Molecular Design of Donors and Non-Fullerene Acceptors for Organic Solar Cells Based on Convolutional Neural Networks

Abstract

Designing novel high-performance donor and acceptor molecules is essential for improving the power conversion efficiency (PCE) of organic solar cells (OSCs). However, conventional experimental methods for developing new materials are often time-consuming, costly, and inefficient. Herein, the deep learning convolutional neural network (CNN) model, random forest, extra trees regression, gradient boosting regression tree, and adaptive boosting models were trained. The comparison indicates that the performance of the CNN model prevails over the traditional machine learning models. Furthermore, a CNN-based molecular generation model combined with transfer learning was presented to design novel donor and acceptor molecules. Consequently, 260,767 donor and 937,155 acceptor molecules were generated, forming 244,379,097,885 novel donor–acceptor pairs. Their OSC performance was predicted using the trained CNN model, identifying 12,224 donor–acceptor pairs with predicted PCE exceeding 19%, with the highest PCE reaching 19.20%. The proposed CNN approach rapidly predicts photovoltaic performance but also enables cost-effective generation of numerous candidate OSC materials.