Experiment-Oriented Machine Learning of Polymer:Non-Fullerene Organic Solar Cells

Despite the capacity of conjugated materials for enhanced power conversion efficiency (PCE) of organic photovoltaics (OPV), a comprehensive survey of unexplored materials is beyond the reach of most researchers’ resources. In such instances, a data-driven approach using machine learning (ML) is an efficient alternative; however, bridging the gap between experimental observations and data science requires a number of refinements. In this investigation, using a random forest model based on an experimental dataset, a high correlation coefficient of 0.85 is achieved for the ML of polymer and non-fullerene small molecule acceptor OPVs and performed virtual screening of 200,932 conjugated polymers generated by the combinatorial coupling of donor and acceptor units. Further, to evaluate the effectiveness of the ML model, a series of conjugated polymers (based on benzodithiophene and thiazolothiazole) were designed, synthesized, and characterized with different alkyl chains. Among these, PBDTTzEH:IT-4F showed a PCE of 10.10%, which is in good correspondence with ML predictions with respect to the choice of alkyl chains. Thus, the current study demonstrates how ML can be utilized for developing OPVs using a relatively small number of experimental data points (566) and screening numerous molecular structures.