The growing demand for high-performance materials is driving the development of diverse strategies to accelerate the discovery of inorganic materials. Among these, generative models, which sample candidates directly from chemical space, have been proposed as efficient and reliable tools. This trend is now extending to the field of heterogeneous catalysis, which was previously dominated by traditional high-throughput screening (HTS) approach. However, in contrast to the progress made in inorganic crystal generation, catalyst generative models have not yet achieved practical and universal catalyst discovery. In this study, we propose an active learning strategy that iteratively fine-tunes the model, generates candidates, and validates them to steer the generation process toward a desired distribution. We implemented this strategy using CatGPT, a predeveloped transformer decoder-based catalyst generative model (J. Am. Chem. Soc. 2024, 146, 49, 33712–33722). As a proof of concept, we applied this approach to the two-electron oxygen reduction reaction (2e-ORR) for hydrogen peroxide (H2O2) production. Through this method, we identified 34 promising candidates while reducing the computational cost by approximately 80% compared with HTS. Subsequent computational analysis and experimental validation further confirmed that MnPt3 is an effective 2e-ORR catalyst. These results demonstrate the potential of our framework to enable efficient discovery of catalysts.
Mok et al. (Mon,) studied this question.