Machine learning theory for acceptor molecular design of D:A1:A2 ternary organic solar cells

The exhaust and tedious experiment for optimizing third component in bulk-heterojunction and huge chemical space of molecules hider the effective development of ternary organic solar cells (OSCs). Herein, we constructed dataset of 479 D:A1:A2 ternary OSCs, and the machine learning model adopted the optimized and adjusted random forest model (RF) with good accuracy. Based on different donors and acceptors in the constructed data set, 1347,357 D:A1:A2 ternary OSCs were reconstructed, with a predicted PCE of up to 18.59 %, corresponding to PBDB-TF:L8-BO:bi-asy-YC12. Five kinds of efficient A2 molecules were selected from the constructed high PCE data set, and 76 new A2 molecules were generated by cutting and recombining them, and combined with 2 efficient donors and 3 A1 acceptors, 456 D:A1:A2 ternary OSCs were generated. Among them, there are 31 groups of ternary OSCs with PCE greater than 18.50 %. This work provides effective theoretical guidance for the development of ternary OSCs.