The structural flexibility of organic semiconductors offers vast a search space, and many potential candidates (donor and acceptor) for organic solar cells (OSCs) are yet to be discovered. Machine learning is extensively used for material discovery but performs poorly on extrapolation tasks with small training data sets. Active learning techniques can guide experimentalists to extrapolate and find the most promising D:A combination in a significantly small number of experiments. This study uses an active learning technique with a predictive random forest model to iteratively find the most optimal D:A combinations in the search space using various acquisition functions. Active learning results with five different acquisition functions (MM, MEI, MLI, MU, and UCB) are compared. Results reveal that acquisition functions that combine exploitation and exploration (MEI, MLI, and UCB) perform far better than purely exploiting (MM) and purely exploring (MU) acquisition functions. Interestingly, the proposed model can overcome the bottleneck of extrapolating small training data sets and find most promising D:A combinations in relatively fewer experiments.
Keywords: acquisition function; active learning; donor:acceptor combinations; machine learning; organic solar cells; power conversion efficiency.