Interpretable machine learning framework: Predicting catalytic activities of complex photocatalysts towards CO2 conversion

The photocatalytic activity of CO₂ conversion via reduction reaction (CO₂RR) is limited by linear scaling relations of adsorption energies between intermediates. Dual-atom photocatalysts (DAPs) have been proposed to overcome this limitation and enhance catalytic efficiency. In this work, we develop an interpretable DFT-based machine learning (ML) framework to predict and rationalize the CO₂RR activity of 299 transition metal (TM)-based DAPs anchored on stoichiometric equivalent graphitic carbon nitride (gC₆N₆). Leveraging structure-sensitive (Magpie) and structure-property (Coulomb Matrix Eigenvalues) features, the AdaBoost regressor model accurately predicts the potential determining step (PDS) with R² = 0.95 and RMSE = 0.06 eV. Feature selection identified the mean electronegativity of the anchored TM atoms as the most important descriptor. The ML model highlights ScTi/gC₆N₆ and RhMo/gC₆N₆ as promising photocatalysts with limiting potentials of 0.58 eV and 0.73 eV, respectively. Further DFT calculations confirm their suitable optoelectronic properties for CO₂RR. This strategy enables accurate and interpretable predictions, potentially accelerating the discovery and design of efficient photocatalytic materials.