Molecular design of organic photovoltaic donors and non-fullerene acceptors: a combined machine learning and genetic algorithm approach

Abstract

In organic solar cells (OSCs), electron donors and acceptors significantly influence photovoltaic performance. However, the complex and variable organic materials make traditional experimental exploration of ideal donor–acceptor combinations inefficient and challenging. Therefore, developing computational methods to guide molecular design and precisely screen/optimize donors and acceptors is essential. Here, a database of 480 donor and non-fullerene acceptor (NFA) pairs was created. Based on Pearson correlation coefficients, 43 molecular structure descriptors were chosen as inputs. The machine learning (ML) algorithms random forest (RF) and extra trees regression were applied to predict photovoltaic parameters. The RF model was demonstrated to be relatively better. Subsequently, to design new molecules effectively, molecular unit libraries were built by cutting donors and NFAs in the database. Genetic algorithm (GA) was employed to generate high-performance D–π₁–A–π₂ type donors and A₁–π₁–D–π₂–A₂ and A₁–D–A₂ type NFAs. The power conversion efficiency (PCE) of the designed donor–NFA pairs predicted by the trained RF model was used as the fitness function value of GA. After 100 iterations, novel donor–NFA pairs were obtained, with the highest predicted PCE reaching 16.85%. This work offers an innovative way to efficiently screen and optimize OSC donor–acceptor materials. The combined ML–GA approach can be extended to other molecular design areas.