Mechanism–driven machine learning screening of dual atom catalysts for N2 selective nitrate reduction and the origin of intermediate overstabilization

Electrocatalytic nitrate reduction (NO 3 RR) offers a sustainable route for removing inorganic nitrogen from water, yet selective N 2 production remains challenging. Dual-atom catalysts (DACs) offer tunable bimetallic sites, but the vast compositional space hinders exhaustive experimental or DFT–based screening. Given its practical relevance to wastewater environments, we develop a mechanism-guided machine learning framework integrating DFT calculations with regression and classification models to screen 702 graphene–supported DACs for N 2 -selective NO 3 RR via the Duca–Feliu–Koper pathway, where *NH 2 couples with adjacent *NO to form *NH 2 NO that subsequently decomposes into N 2 . Electronic, geometric, thermodynamic, and coordination descriptors are employed to predict intermediate stability and adsorption properties. A composite N 2 Selectivity Score, incorporating NO 3 - /H competition, *NO formation tendency, and *NH 2 NO selectivity, enables high–throughput candidate identification. DFT validation identifies ScN 3 PtN 3 @Gra–A, HfN 3 PtN 3 @Gra–A, and ZrN 3 IrN 3 @Gra–B as favorable catalysts with suppressed HER and enhanced N–N coupling. Further electronic structure analysis reveals that excessive metal–metal orbital hybridization over–stabilizes the *NH 2 NO intermediate (ΔG(*NH 2 NO) = -6.642 eV) on ScN 3 PtN 3 @Gra–A, creating a thermodynamic trap that hinders N 2 desorption. This uncovers an adsorption–stability trade–off in DACs, extending the Sabatier principle to dual–site cooperative catalysis. This work demonstrates a mechanism-constrained screening strategy and provides electronic insights into the role of inter-site synergy for selective DAC design.