Machine learning-driven screening of g-C3N4 supported single-atom-catalysts for CO2 conversion and potential syngas production

The renewable energy-driven electrochemical CO2 reduction reaction (CO2RR) and syngas production provide a sustainable pathway for achieving carbon neutrality and facilitating energy transition. Developing efficient catalysts is crucial for advancing these technologies. This study systematically screened g-C3N4-supported single-atom catalysts (SACs@g-C3N4) using density functional theory calculations and machine learning (ML), evaluating their catalytic activity in CO2RR as well as their potential for syngas production. By employing the XGBoost regression ML model, we accurately predicted the formation energies (Ef) of the catalysts, achieving high reliability with R2 values of 0.9664 and 0.9120 for the training and testing sets, respectively. Furthermore, by combining SHapley Additive Explanations (SHAP) with ML predictions, we developed a highly interpretable model for predicting the Gibbs free energy change (ΔG) in CO2RR, achieving R2 values of 0.9822 and 0.9349, for the training and testing sets, respectively. Consequently, we identified 11 catalysts with excellent CO2RR performance and evaluated their potential for hydrogen evolution reaction (HER) and syngas co-production. Specifically, Al-doped catalysts exhibited preferable performance for syngas generation, whereas Zr-, Ti-, Co-, Si-, and Ni-doped catalysts were more suitable for selective CO2RR. Overall, this study provides valuable theoretical guidance for the application of SACs@g-C3N4 catalysts in CO2 conversion and syngas production.