Integrating deep learning and symbolic regression for molecular design and virtual screening of organic solar cells

The photovoltaic performance of organic solar cells (OSCs) is significantly determined by the electron donor and acceptor materials in active layers. Traditional trial-and-error experiments for exploring high-performance materials suffer from long development cycles, high experimental costs, and low screening efficiency. Herein, the established database includes 547 donor-acceptor pairs, integrating photovoltaic parameters and molecular representations. The 30 molecular structure descriptors that closely relate power conversion efficiency (PCE) were extracted. Long short-term memory networks (LSTM), convolutional neural networks (CNN), and symbolic regression (SR) were trained to predict the PCE of OSCs. After hyperparameter optimization via grid search algorithm, the metrics indicate the trained models achieved high-precision for PCE prediction, and the performance of LSTM model prevail over than that of other models. Through dual validation by SHapley Additive exPlanations(SHAP) interpretability analysis and SR formulas, it was revealed that the number of structural units with double rings or more in acceptor molecules showed the significant correlation with PCE. Based on the dataset constructed using molecular fragment recombination strategy, the developed LSTM generative model successfully generated 210,660 novel donor molecules and 878,268 acceptor molecules. Following screening of 185,015,936,880 donor-acceptor pairs by the LSTM prediction model, 5753 donor-acceptor pairs with the predicted PCE exceeding 18.50% were identified, among which the highest predicted PCE reached 18.66%. This approach provides theoretical guidance for the discovery of organic photovoltaic materials and may accelerate the development of high-performance OSCs, but also can be generalized to functional molecular design.