Elucidating the fundamental governing principles of CO2 reduction on single-atom catalysts through interpretable machine learning

Abstract

Understanding the effect of local coordination and electronic structure on catalytic activity and selectivity is crucial to designing CO₂ reduction reaction (CO₂RR) catalysts. Combining hierarchical high-throughput density functional theory (HT-DFT) screening and interpretable machine learning (ML), we aim to accelerate and rationalize the discovery of efficient SACs for the CO₂RR. Using a BC₂ monolayer as a prototypical substrate, we constructed a library of transition-metal (TM) SACs featuring diverse vacancy types and N-coordination. A four-stage HT-DFT workflow, namely assessing stability, CO₂ adsorption, key elementary steps, and hydrogen evolution reaction (HER) selectivity, was used, and 16 promising candidates with low limiting potentials ( −0.35 V) were identified. By integrating ML regression and feature-importance analysis, five fundamental physicochemical features that control activity, including TM d-electron count (N_d), ionization energy (IE₁), electronegativity (χ_TM) and local electronegativity metrics (χ_n, W_v) were extracted. Furthermore, six analytical descriptors (φ₁–φ₆) were derived to quantify atomic features with catalytic energetics; in particular, the descriptor framework is transferability to different graphene-like supports. Mechanistic analysis revealed that increasing N_d enhances C, O-related adsorption while weakening O–TM interactions, and higher IE₁ suppresses the HER by favoring CO₂ activation. This study establishes a physically interpretable ML-DFT strategy that unifies predictive efficiency with mechanistic insight, providing a blueprint for the rational design of atomically dispersed electrocatalysts for CO₂ conversion.