Design and virtual screening of donor and non-fullerene acceptor for organic solar cells using long short-term memory model
Abstract
In organic solar cells (OSCs), electron donor-acceptor materials are key factors influencing device performance. However, traditional experimental methods for developing new, high-performance materials are often time-consuming, costly and inefficient. To accelerate the development of novel OSCs donor-acceptor materials, we constructed a database of 547 donor-acceptor pairs and derived 30 easily obtainable molecular structure descriptors through transformation screening. Using long short-term memory (LSTM) network model, belonging to deep learning, we tuned the LSTM model with grid search for optimal hyperparameters, and predicted power conversion efficiency (PCE), open-circuit voltage, short-circuit current density and fill factor. SHAP analysis revealed that the number of rotatable bonds and the presence of two or more rings in acceptor molecules positively impact PCE. We then systematically fragmented and recombined molecules in the constructed database, creating 142,560 donor molecules and 61,732 acceptor molecules. The tuned LSTM model predicted photovoltaic performance parameters for these new donor-acceptor pairs. After excluding the donor-acceptor pairs in the database, we identified 7,632 novel pairs with predicted PCE greater than 18.00%, including five pairs exceeding 18.50%, with the maximum PCE of 18.52%. This method facilitates the cost-effective design and rapid, accurate prediction of OSCs material performance, enabling efficient screening of high-performance candidates.