Designing highly efficient C-coordinated single-atom catalysts for CO2 reduction through modulating metal center and its second coordination sphere: A combined DFT and machine learning approach.

Electrochemical CO2 reduction (CO2RR) plays a crucial role in realizing carbon circulation and reducing CO2 emissions. In this realm, single atom catalyst (SAC) is one promising catalyst owing to its maximal atom utilization and high catalytic performance. Herein, in this study, we designed 520 surface models of TM-C3 supported on the N-doped graphene, with various combinations of heteroatom (HAs) being introduced in the second coordination sphere of TM (TM-C3-X). Subsequently, 460 catalysts pass the stability test and were employed as the research subject. Among them, we selected 92 catalysts and investigated their CO2 adsorption behavior, reaction mechanism and catalytic performance in CH4 formation through the high-throughput density functional theory (DFT). Combined with the feature variables obtained from the electronic/structural properties of TM-C3-X, these energetics data were employed as the training set for various machine learning (ML) regression and classification models. Feature importance analysis identifies Ne(TM) and χ(TM) as the key factors governing the stability of TM-C3-X, whereas its catalytic activity towards CH4 formation is primarily regulated by εd(TM), εp(C3), and εp(HAs). Through electronic structure analysis as well as SHapley Additive exPlanations (SHAP) analysis, a universal descriptor has been developed to reflect the catalytic activity of SACs. Based on these findings, SACs exhibiting both the high catalytic activity and product selectivity in CO2RR to CH4 have been determined. We hope our current work could provide useful insight into exploring highly-efficient catalysts in CO2RR.