AI-expanded model maps overlooked reactions in CO2 hydrogenation
Researchers at the Indian Institute of Science combined quantum-mechanics simulations, machine learning and automated reaction discovery to expand a copper-catalyst CO2-hydrogenation network from 152 to 9,389 reactions. The larger model aligned with experimental product observations, estimated about 40 times more CO2 conversion than the limited model, and identified a possible direct molecular-hydrogen transfer route.
The story
Researchers at the Indian Institute of Science (IISc) reported a computational analysis of CO2 hydrogenation on a copper catalyst, a process used to make chemicals including methanol. They began with a quantum-mechanics database of 152 reactions, then trained machine-learning models to estimate energy barriers for additional possible transformations. Automated tools examined reactions among 105 chemical species on the catalyst surface and expanded the network to 9,389 individual reactions. When the team used the original, smaller network in a reaction-rate model, it predicted formic acid as the main product and substantially underestimated CO2 conversion. The expanded network instead identified methanol and carbon monoxide as major products, consistent with the experimental findings described by the researchers, and predicted approximately 40 times more CO2 conversion than the original model. Experimental validation involved researchers at Hindustan Petroleum Corporation Limited’s Green Research and Development Centre and the Agency for Science, Technology and Research in Singapore. The expanded analysis also identified steps in which molecular hydrogen, H2, could transfer directly to oxygen-containing intermediates without first splitting into separate hydrogen atoms. The team says further quantum-mechanics calculations found that route could be especially favourable for those intermediates.
Why it matters
Catalytic-reaction models are used to decide which chemical pathways and catalyst properties deserve experimental attention. A model that misses consequential surface reactions can point researchers toward the wrong dominant product or underestimate conversion, as the IISc team found with its smaller network. By making the search for reactions less dependent on a preselected list, the approach could improve how researchers evaluate routes to methanol and other CO2-derived chemicals. It also provides a way to test mechanistic assumptions, including how hydrogen is transferred during a surface reaction, before treating them as design rules.
Evidence and context
CO2 hydrogenation involves many species and transformations on a catalyst surface, making an exhaustive quantum-mechanics calculation computationally demanding. Researchers therefore commonly restrict models to reactions thought most important, but the source describes how this can omit steps that shape product and conversion predictions. In this study, expanding the network changed the model’s major-product prediction from formic acid to methanol and carbon monoxide and brought it into agreement with experimental observations. The authors say the combined framework—quantum-mechanics simulations, machine learning, automated reaction identification and reaction-rate modelling—could also be applied to CO2 reduction on other catalysts, nitrogen reduction and water splitting. Those are proposed research uses, not demonstrated results of this study.
Limits and unknowns
The central limitation is that the reported improvement is in a computational reaction network and its agreement with available experiments, not a newly demonstrated catalyst or operating process. The approximately 40-fold figure compares predictions from the expanded model with those from the team’s initial 152-reaction model; it is not a measured 40-fold increase in a plant’s CO2 conversion. Although the authors performed additional quantum-mechanics calculations on the molecular-H2 pathway and report experimental validation, the source does not establish whether deliberately strengthening catalyst interactions with H2 will improve methanol yields under industrial conditions.
Follow the story
Share it, watch the explainer, or continue with the reporting trail.
Further reading on this topic
2 links
AI-expanded model maps overlooked reactions in CO2 hydrogenation
A short visual explainer for readers who want the story in a different rhythm.
Highlights
- A larger map of CO2 reactions
- From 152 to 9,389 reactions
- Observed products matched the model
- A guide for future catalyst tests
Transcript
A new model maps thousands of overlooked reactions involved in converting carbon dioxide into useful chemicals.
Indian Institute of Science researchers grew their reaction map from 152 to 9,389 using machine learning.
It matched observed methanol and carbon monoxide, and suggested hydrogen molecules might transfer without splitting.
That could help catalyst researchers choose which pathways to test for making methanol from carbon dioxide.
Its fortyfold conversion estimate compares two models; any industrial improvement still needs to be demonstrated.