ABSTRACT Discovering next‐generation heterogeneous catalysts calls for embracing the full complexity of active site formation under realistic conditions. Here, we develop a robust machine learning potential (MLP)‐aided computational framework that integrates realistic preparation and reaction conditions to effectively track the formation of active sites and decipher structure‐activity relationships. Using syngas conversion over the Zn x Cr y O z system as a demonstration, we identified that the system preferentially segregates into ZnO and ZnCr 2 O 4 phases, with ZnO forming a monolayer on ZnCr 2 O 4 surfaces under preparation conditions. Under reaction conditions, by deploying CH─O bond dissociation as a descriptor, we found that the ZnO/ZnCr 2 O 4 (100) surface is the active surface. Crucially, we pinpoint geometrically linked oxygen vacancy pairs as the true active sites. Full microkinetic analyses conducted on these active sites yield kinetic results that align well with experimental observations. Beyond elucidating the active structure, a model for designing oxide/oxide catalysts to achieve high activity is generalized, opening new pathways for accelerating catalyst discovery across a wide range of reactions.
Han et al. (Sun,) studied this question.