Machine Learning Expands Copper CO2 Hydrogenation Network to 9389 Steps, Correcting Product Predictions

The team started with 152 quantum-mechanics reactions, trained ML models on energy barriers, then enumerated all single-step transformations among 105 surface species. The expanded network flipped the dominant product from formic acid to methanol and carbon monoxide while lifting overall CO2 turnover by a factor of roughly 40. Validation against lab data from Hindustan Petroleum and A*STAR confirmed the larger model’s accuracy. Conventional truncations had omitted critical hydrogen-transfer steps that AI surfaced systematically.

This work shows how incomplete reaction networks systematically bias mechanistic models toward wrong products and underestimated rates. By quantifying thousands of overlooked pathways, the study reveals that hydrogen can participate via previously ignored routes, overturning textbook assumptions about rate-limiting steps in CO2 hydrogenation. The 40-fold correction underscores that model fidelity, not just catalyst choice, determines whether simulations guide or mislead experimental programs.

Next steps include integrating the expanded network into microkinetic software used by industry and testing the newly identified pathways on doped or nanostructured copper surfaces. If the additional channels prove generalizable, catalyst design cycles could shorten from years to months. High-throughput validation across multiple metals will be required before the approach becomes standard.

The main limitation remains computational: ML barrier predictions still carry extrapolation error outside the original training distribution. Larger, diverse quantum datasets and uncertainty-aware models would strengthen evidence for deployment in commercial reactor design.