Molecular Design and High-Throughput Virtual Screening of Electron Donor and Non-fullerene Acceptors for Organic Solar Cells

The complicated trilateral relationships among molecular structures, properties, and photovoltaic performances of electron donor and acceptor materials hinder the rapid improvement of power conversion efficiency (PCE) of organic solar cells (OSCs). Herein, the database of 310 donor and non-fullerene acceptor pairs is constructed and 39 molecular structure descriptors are selected. Four kinds of machine learning (ML) algorithms random forest (RF), extra trees regression, gradient boosting regression trees, and adaptive boosting are applied to predict photovoltaic parameters. The coefficient of determination, Pearson correlation coefficient, mean absolute error, and root mean square error are adopted to evaluate ML performance. The results show that the RF model exhibits the best prediction accuracy. The Gini important analysis suggests the fused ring and aromatic heterocycles are critical fragments in determining PCE. The molecular unit sets are constructed by cutting each donor and acceptor molecules in database. The 31 752 D-π-A-π type donor molecules and 5 455 164 A-π-D-π-A type acceptor molecules are designed by recombination of molecular units, and 173 212 367 328 donor–acceptor pairs are generated by combining the newly designed donor and acceptor molecules. Based on the predicted PCE using the trained RF model, 42 donor–acceptor pairs exhibit the predicted PCE > 16%, in which the highest PCE is 16.24%.