First principles and interpretable machine learning aided the exploration of defective MXene for nitrogen reduction reaction

Electrocatalytic nitrogen reduction to ammonia is an efficient and green energy conversion technology, emerging as a promising alternative to the Haber-Bosch process. However, the design and exploration of highly efficient catalysts continues to pose a formidable challenge. Herein, the catalytic performance including stability, activity, and selectivity for the nitrogen reduction reaction of 132 defective 2D MXenes are systematically discussed using first-principles calculations and machine learning models. Several high-performance catalysts with ultralow limiting potential are identified such as Ti₂CSe₂ (0.30 eV), Ti₂NSe₂ (0.32 eV), Nb₂CO₂ (0.41 eV), and Zr₂CSe₂ (0.43 eV). Furthermore, our findings highlight the exceptional predictive accuracy of the Ridge Regression (RDG) and Support Vector Regressor (SVR) models for ∆G₁ (*N₂→*NNH) and ∆G₂ (*NH₂→*NH₃), respectively, with coefficients of determination (R²) scores of 0.969 and 0.933. Importantly, the Shapley Additive exPlanation analysis has elucidated the significance of descriptors influencing the free energy changes, thereby revealing the underlying origins of the catalytic activity in defective MXenes. The current work offers deeper insights by intricately correlating structure and performance, which could serve as a valuable guide for the design and application of high-performance catalysts. © 2025