Big data-driven machine learning transformation for atomic-scale heterogeneous catalyst design: a critical review

Machine learning (ML) methods have been applied to the study of chemical microstructures, such as predicting quantum properties or generating molecular conformations. From a microscopic perspective, heterogeneous metal-based catalytic reactions involve a series of elementary steps in which reactants interact with catalysts. The specificity of heterogeneous catalysis necessitates adapting general ML originally designed for molecular systems. Recent advancements in artificial intelligence have demonstrated unprecedented capabilities in processing cross-domain information. With the emergence of big data-driven catalyst design, we synthesize recent progress, identify gaps, and outline challenges in the applications of ML for both small-scale and large-scale datasets in the discovery and design of solid heterogeneous catalysts. Specifically, we first summarize and analyze current applications of ML models in predicting microstructural properties and identifying reaction mechanism by integrating optimization algorithms. We then discuss recent developments in artificial intelligence (e.g., large language models, generative models, and multimodal models) for heterogeneous catalyst design.