Active Learning Enables Discovery of Asymmetric Dual-Atom Catalysts for Nitrate-to-Ammonia Electrocatalysis

Abstract

Electrochemical nitrate reduction to ammonia requires a delicate balance between nitrate activation, hydrogen delivery, suppression of competing hydrogen evolution, and stabilization of the catalytic metal centers. Such coupled requirements make the rational design of dual-atom catalysts challenging, particularly when the two metal sites and their coordination environments contribute differently to the reaction. Here, we establish an interpretable active-learning framework to navigate the chemical space of asymmetric dual-atom catalysts and uncover the physicochemical origins of their performance. Rather than learning an empirical activity score, the framework predicts individual DFT-derived properties from readily accessible elemental and structural descriptors and uses uncertainty-aware sampling to iteratively explore previously uncharted metal-pair and coordination environments. Model interpretation reveals distinct roles of the two metal sites in regulating nitrate activation and hydrogen adsorption, and identifies moderate electronic asymmetry coupled with mixed N/O coordination as a favorable design regime. Guided by these principles, active learning efficiently explores an expanded library of 1050 structures and progressively identifies candidates beyond the initial chemical space. Prospective DFT validation highlights Mo–Ta/N3O3–C, Mo–Nb/N3O3–C, and Mo–Nb/N5O1–C as promising catalysts, whose performance arises from a balanced combination of reaction thermodynamics, adsorption selectivity, and structural stability. These findings establish a transferable strategy for designing multifunctional dual-atom sites and demonstrate how interpretable, closed-loop learning can transform catalyst screening from property prediction into mechanism-guided materials discovery.