Learning Interpretable Activity Descriptors Directly from Measurement Features in Heterogeneous Catalysis

Conventional catalyst informatics largely relies on composition-based descriptors for catalytic activity prediction. In this study, heterogeneous characterization data are integrated into a multidimensional measurement space and transformed into measurement-derived descriptors for machine-learning analysis. This shift from composition-based to measurement-centered catalyst informatics enables characterization information that is conventionally interpreted individually to be evaluated systematically and quantitatively in combination, providing a framework for describing the activity landscape across catalytic states. Importantly, combining complementary measurement-derived descriptors enables highly accurate prediction of CO2 conversion while revealing complex factors and interactions associated with catalytic activity. Thus, measurement informatics provides a rational framework for extracting interpretable structure-activity relationships from experimental measurement data.