Data-driven massive reaction networks reveal new pathways underlying catalytic CO2 hydrogenation

Heterogeneous catalytic pathways for clean energy conversion involve thousands of elementary steps, but most quantum-mechanical models involve only a few dozen reactions. We combine extensive density functional theory (DFT) calculations, machine learning (ML) for activation barrier prediction, and human intelligence-inspired reaction enumeration and elementary reaction identification. This enables automated kinetic modeling of CO2 hydrogenation on copper, a key process to produce fuels and chemicals. We construct the largest dataset of 152 elementary CO2 reduction reactions and experimentally determine CO2 conversion, finding that even large networks with 100+ reactions are insufficient. In contrast, our approach reveals 9389 elementary reactions, reducing human bias in the reaction pathway. We unravel 40-fold higher CO2 conversion rates, following experimental trends of methanol and CO production. We establish the crucial role of intermolecular hydrogen transfer and hydrogenation by molecular hydrogen, a surprising ML-enabled discovery validated post-facto. The proposed strategy to comprehensively model complex catalytic mechanisms will significantly advance catalysis research and carbon conversion processes.