Molecular metal complexes offer opportunities for developing artificial photocatalytic systems. The search for efficient molecular photocatalytic systems, which involves a vast number of photosensitizer–catalyst combinations, is extremely time consuming via a conventional trial and error approach, while high-throughput virtual screening has not been feasible owing to a lack of reliable descriptors. Here we present a machine learning-accelerated high-throughput screening protocol for molecular photocatalytic $CO_{2}$ reduction systems using multiple descriptors incorporating the photosensitization, electron transfer and catalysis steps. The protocol rapidly screened 3,444 molecular photocatalytic systems including 180,000 conformations of photosensitizers and catalysts during their interaction, enabling the prediction of six promising candidates. Then, we experimentally validated the screened photocatalytic systems, and the optimal one achieved a turnover number of 4,390. Time-resolved spectroscopy and first-principles calculation further validated not only the relevance of the descriptors within certain screening scopes but also the role of dipole coupling in triggering dynamic catalytic reaction processes.