AI Uncovers Nearly 10,000 Hidden Chemical Reactions That Could Turn CO2 into Fuel
Researchers at the Indian Institute of Science (IISc) have used machine learning combined with atomistic simulations to expand the reaction network for carbon dioxide conversion on a copper catalyst from 152 reactions to 9,389, uncovering thousands of pathways that conventional models had missed. The enlarged network raised predicted CO2 conversion roughly fortyfold and correctly identified methanol and carbon monoxide as the major products, matching experimental findings. The work is published in Nature Communications.
- Reaction network expanded from 152 to 9,389 reactions, covering 105 chemical species on the catalyst surface
- The original small network wrongly predicted formic acid, not methanol, as the main product and underestimated conversion
- Predicted CO2 conversion increased approximately 40 times with the expanded network, aligning with experiments
- The study also revealed a surprising way hydrogen participates in the reactions, challenging established assumptions
- Findings could guide more efficient catalysts for turning CO2 into fuels such as methanol
The Modeler’s Nightmare: Missing the Step That Matters
Recycling CO2 into methanol and other valuable chemicals via catalytic hydrogenation involves enormous numbers of individual reactions across a catalyst’s surface. Simulating every possible step with quantum mechanics is prohibitively expensive, so researchers typically model a limited subset of reactions believed to matter most, potentially excluding hundreds of critical ones.
“We began with a worry familiar to anyone who does mechanistic modeling: How do you know that your reaction network has not omitted the one step that matters?” said first author Anand Mohan Verma, who conducted the work as a CV Raman Postdoctoral Fellow at IISc and is now an assistant professor at MNNIT Allahabad. The team built a database of 152 reactions from quantum simulations, trained machine learning models to rapidly estimate energy barriers for new reactions, and used automated computational tools to systematically map every plausible single-step transformation among 105 chemical species on the catalyst surface.
A Fortyfold Correction, Validated by Experiment
The difference was striking. With the original 152-reaction network, the model incorrectly predicted formic acid rather than methanol as the main product and substantially underestimated how much CO2 gets converted. Once the thousands of overlooked reactions were included, predicted conversion rose about fortyfold and the major products matched what experiments observe. Experimental validation was carried out by teams at Hindustan Petroleum Corporation Limited’s Green Research and Development Centre and at the Agency for Science, Technology and Research in Singapore.
Corresponding author Ananth Govind Rajan, associate professor in the Department of Chemical Engineering at IISc, said the study also uncovered a surprising way hydrogen participates in the reactions, challenging established assumptions and pointing toward the design of more efficient catalysts for CO2 conversion.
