Machine-learning-guided screening of organic photosensitizers accelerated by domain adaptation

While machine learning (ML) is transforming reaction-condition screening, harnessing prior knowledge as effectively as an experienced organic chemist remains a significant challenge. Here, we report an ML-based screening framework for identifying optimal organic photosensitizers (OPSs), which leverages domain-adaptation-based transfer learning (TL) to enhance surrogate-model performance. Even when applied to photoreactions that are more complex than those in the source domain, the TL-based system outperforms conventional ML-based optimization methods that do not utilize TL. These results highlight the feasibility of using chemically rational TL that enables the transfer of knowledge from simple reactions to those where mechanistic understanding or database construction is more difficult. Moreover, the proposed approach performs reliably no matter how the initial samples are selected, and it is effective in identifying a highly active OPS that is missing from the source domain, thereby addressing critical limitations of current ML-based approaches in organic synthesis.