A Machine Learning Model To Predict CO2 Reduction Reactivity and Products Transferred from Metal-Zeolites

Various reactive intermediates and Cn products from carbon dioxide reduction reaction (CO2RR) play critical roles in the chemical and fuel industry. Developing easily accessible activity descriptors to predict possible intermediates and products of CO2RR is of great importance. The free energy changes (ΔG) for all possible reaction intermediate and product probability (P) of CO2 reduction to methanol, methane, and formaldehyde on 26 single-atom catalysts (SACs) in zeolites were predicted by density functional theory (DFT) calculations and machine learning (ML) models. The adsorption free energies of ΔG*OH and ΔG*O*CH2 were highly correlated with catalytic activity. Producing methanol was favorable for metal-zeolites with early transition metals and main group elements. Methane production was more feasible for some systems such as Co-zeolite, due to the low free energy and high selectivity against the hydrogen evolution reaction. Both XGBoost and ExtraTrees algorithms could give satisfactory predictions of ΔG and P in CO2RR using descriptors of reaction pathways, metal, charge transfer (CT) between the metal and reaction intermediate, hydrogen bond interaction between the intermediate and zeolite framework, and geometry. The global electronegativity difference (δχT) and average ionization energy difference (δIE) between the metal-zeolite and intermediate were introduced as features (along with the valence electron number of metals and the atomic number of reaction species) for prediction of CT values without the need of DFT calculations. The CT feature could be replaced by some additional descriptors such as the band gap (Eg) or coordination number of metals to intermediates in training ML models for free energy prediction. ML models on an external test such as MOFs, 2D materials, and molecular complex materials indicate that the proposed descriptors are general for the reaction free energy change and product prediction of SACs in CO2RR.