Multitask Learning Reveals Shared Descriptors Governing Activity and Selectivity in Catalytic Nitrate Reduction

Catalytic nitrate reduction (CNR) is a promising technology for nitrate removal in water treatment, but catalyst development remains largely empirical and constrained by a persistent trade-off between reaction rate and ammonium formation. Here, we develop a multitask Gaussian process (MTGP) framework to better characterize the activity-selectivity trade-off by jointly modeling catalytic activity and ammonium selectivity using a large literature-curated data set of Pd- and Pt-based catalysts. By learning shared catalyst property–performance relationships, the MTGP captures the activity-selectivity coupling and improves predictive stability and generalization relative to single-task models, reducing the generalization gap by 41–45% in ΔR2 and up to 48% in ΔRMSE. To enhance mechanistic interpretability, we introduce a consensus feature-importance framework that integrates model-dependent and model-agnostic approaches to identify key catalytic descriptors. The learned relationships are consistent with established catalytic principles, including distinct kinetic trends, optimal catalyst mass and loading, and the coupled influence of hydrogen availability and pH on N2 versus NH4+ formation. External validation using newly published data and knowledge transfer to a data-limited kinetic data set further support the predictive robustness and transferability of the framework within the studied domain. These results suggest that multitask learning can capture transferable activity-selectivity relationships in CNR, providing a data-driven framework to guide catalyst and process design.