Electrocatalytic nitrate reduction to ammonia (NO₃RR) offers a transformative route for pollutant remediation and green synthesis, yet rational catalyst design remains challenging. We combined density functional theory (DFT) with machine learning (ML) to study TM@M₂CO₂ systems (M = Mo, W, and Cr; TM = 3d–5d metals). Of the 150 models, 16 passed screening for stability (formation energy and dissolution potential) and selectivity (suppression of the hydrogen evolution reaction, HER). Their limiting potentials (Uₗ) were −0.38 V for CuI@Mo₂CO₂ and −0.45 V for RuII@W₂CO₂. These two systems are the most promising NO₃RR candidates among the systems studied. Mechanistic and kinetic analyses confirmed that both catalysts can activate nitrate effectively and exhibit lower barriers for the potential-determining step than for HER. Feature-importance and Shapley Additive Explanations (SHAP) analyses identified the average TM–O bond length (L_TM–O), nitrate adsorption free energy (ΔG_*NO₃), and TM electronegativity (χ_P) as influential descriptors, revealing a nonmonotonic association between (L_TM–O) and the predicted limiting potential. This work identifies two theoretically promising catalyst candidates and quantifies descriptor importance, offering a data-driven framework for the rational design of MXene-based single-atom catalysts.