Machine Learning-Assisted Screening and Performance Prediction of Single-Atom Catalysts for Electrocatalytic CO2 Conversion to CO

To study the single-atom catalysts (SACs) for CO2 electrocatalytic reduction (CO2RR) to CO, three machine learning models [decision regression tree (DRT), random forest (RF), and gradient boosting decision tree (GBDT)] were trained and evaluated based on the database containing 333 sets of experimental data. The feature importance and Shapley additive explanation (SHAP) analysis were used for model interpretation. The partial dependency plot (PDP) was used to predict the impact of individual feature value on the Faraday efficiency of the CO2RR to CO (FEco). Moreover, the optimal carrier, optimal single atom, and relative hydrogen evolution potential were predicted. The results indicate that the GBDT model has the best predictive performance, with R2 values of 0.94 and 0.90 on the training and testing sets, respectively. The catalyst loading, specific surface area, doping atomic content, relative hydrogen evolution potential, and atomic number are important factors affecting the FECO. The ZIF-8, MTV MOFs, and graphene are the best SAC supports. Using ZIF-8 as the carrier, with a relative hydrogen evolution potential of −0.8 V and Ni-doped single atom, the best FEco is obtained with the predicted and experimental values of 91.48 and 99.70%, respectively, accompanied by an error of 8.2%. The finding provides strong guidance for the selection of carriers, doped atoms, and reaction conditions (relative hydrogen evolution potential) of SACs for the CO2RR to CO, as well as the exploration of general influencing laws.