Machine Learning-Assisted Design Framework of Carbon Edge-Dominated Dual-Atom Catalysts for Urea Electrosynthesis.

Direct electrosynthesis of urea is highly desirable but is severely hindered by intricate proton-coupled electron transfer networks and competing reduction side reactions. Herein, we present a closed-loop data-driven strategy integrating high-throughput density functional theory and machine learning (ML) to systematically design edge-anchored dual-atom carbon-based catalysts. By decoding the reaction networks of 90 heteroatomic metal pairs, we demonstrate that conventional single-molecule adsorption descriptors fail under coadsorption conditions. Instead, the coadsorption energy (Eads(*CO_NO)) emerges as a robust universal descriptor (R2 = 0.72-0.91). Based on this, a quantitative selectivity phase diagram was constructed, identifying a narrow thermodynamic window (-3.57 to -3.08 eV) that favors the C-N coupling pathway against competitive CO reduction reaction and nitrogen reduction reaction. Leveraging an XGBoost regression model trained on intrinsic atomic features, we rapidly screened a chemical space of 1458 candidates. This workflow successfully narrowed the field to identify Zr_Pd@A and Zn_Pd@Z as superior catalysts, exhibiting completely downhill thermodynamic pathways. Electronic structure analysis reveals that the high d-electron density of Pd near the Fermi level optimally activates NO, while the completely empty or fully occupied d-orbitals of early (Zr) and late (Zn) transition metals weakly bind CO, preventing its deep reduction. This work establishes a scalable ML-assisted paradigm for decoupling competitive mechanisms in complex electrocatalysis.