Designing Donors and Nonfullerene Acceptors for Organic Solar Cells Assisted by Machine Learning and Fragment-Based Molecular Fingerprints

The molecular structures and properties of donor and acceptor materials for organic solar cells (OSCs) determine their photovoltaic performance; however, the complex relationship between them has hindered the development of OSC materials. To study this, we constructed the database comprising 544 donor and non-fullerene acceptor pairs. Based on the principle of minimal rings and molecular units, each molecule in the database is cut into different fragments and defined as a new molecular fingerprint, where each bit corresponds to a fragment number in the molecule. Accordingly, the defined molecular fingerprint length is 234 and 723 bits for donors and acceptors, respectively. Random forest and extreme tree regression (ETR) are applied to predict the photovoltaic parameters, with ETR being the most effective. Through SHapley Additive exPlanations (SHAP) importance analysis, eight (10) important donor (acceptor) fragments are identified. Furthermore, by computing the cut fragment similarities with that of the important fragments obtained from SHAP analysis, fragments with similarity exceeding 0.6 are collected in order to design new molecules. By assembling the collected fragments, we designed 21 168 D-π-A-π-type donors and 1 156 400 A-π-D-π-A-type nonfullerene acceptors, generating 24 478 675 200 donor–acceptor pairs. Based on predictions using the trained ETR model, the highest power conversion efficiency reaches 13.2%.