Construction of Sabatier Volcanoes for CO₂ Hydrogenation to C1-2 Oxygenates Using Data-Efficient Machine Learning

The development of new technologies for CO₂ hydrogenation to valuable chemicals, such as alcohols and sustainable aviation fuels, is prioritized globally due to their potential for large-scale abatement of CO₂ emissions. However, the rational design of catalysts for this reaction is undermined by insufficient understanding of their reaction networks and the critical role of coordinatively unsaturated sites for the catalyst activity. Here, we address this gap by developing a data- and computation-efficient framework for exploring CO₂ hydrogenation to C1 and C2 oxygenates on nanoparticles of transition metals, namely, Au, fcc-Co, Cu, Ni, Pd, Pt, and Rh. Our approach integrates DFT and a neural network model to predict activation barriers with high accuracy (MAE < 0.23 eV) and generalizability even in the low-data regime. The NN leverages Coulomb matrices for molecular structures, one-hot encoding for metal identity and transition state geometries, enabling robust predictions of activation barriers even for underrepresented classes of elementary steps. Brønsted-Evans-Polanyi relationships reveal that electropositive metals (Co and Ni) favor C–O cleavage, while electronegative metals (Cu and Au) are highly active in C–C coupling. By evaluating catalyst activity based on the energetic span model implemented in an automatic fashion to efficiently analyze extended reaction networks, we construct Sabatier volcano plots and identify Rh as the best monometallic catalyst for CO₂ hydrogenation into C2 products. The proposed methodological framework opens the way for the computational discovery of CO₂ hydrogenation catalysts by overcoming data limitations and the complexity of the underlying reaction mechanisms.