Machine Learning Across Metal and Carbon Support for the Screening of Efficient Atomic Catalysts Toward CO₂ Reduction

Developing efficient atomic catalysts (ACs) for the CO₂ reduction reaction (CO₂RR) still requires ultrahigh experimental resources and a long research period due to the complicated reaction mechanisms and abundant active sites. Herein, this work presents the energy-based first principles machine learning (FPML) method for the first time based on over 15 000 datasets to directly predict the reaction trends of the CO₂RR. The unique scaling relationship of the hydrogenation steps is revealed in ACs for the CO₂RR, which is correlated with the active sites instead ofelectron transfer numbers. Based on machine learning (ML) predictions, this work reports that the standard electrode potential is affected by the pH values, and proposes a zero-point calibration strategy to realize more accurate predictions of electrocatalysis reactions to supply meaningful references to experiments. The formation of electroactive regions constructed by mixing active sites is revealed, which confirms the neighboring effects for the activation of active sites. In addition, the prediction of C₃ intermediates indicates the potential of multicarbon coupling processes on the carbon active sites of graphdiyne. This work supplies an effective method to predict chemical reaction trends of different ACs in the CO₂RR by ML, which is expected to accelerate the rational design of novel ACs for broad electrocatalysis.