AI uncovered 9,389 chemical reactions involved in turning carbon dioxide into fuel

A reaction network describing how carbon dioxide and hydrogen turn into fuel on a copper surface has grown to 9,389 chemical reactions after researchers at the Indian Institute of Science used machine learning to fill in what earlier models left out. The result matched experiments that the smaller network got wrong. That correction is the practical test of the exercise: a model that mislabels the main product cannot guide catalyst work, while one that matches measured methanol and carbon monoxide yields can start to point at which steps are worth changing.

The tech news worth your time: 400,000+ professionals from Google, Apple and OpenAI read Techpresso, a daily email with the most interesting tech news and insights.

Subscribe to Techpresso →

We may earn a commission if you sign up.

Ananth Govind Rajan, an associate professor of chemical engineering at the institute and corresponding author of the study, said the predictions fell in line only after the network was expanded “to include thousands of additional, previously overlooked reactions.” The work was published in Nature Communications.

From 152 computed reactions to 9,389

The ScienceDaily report of Oct. 13 lays out the method in steps. Quantum mechanical simulations first built a database of 152 reactions. Machine learning models then estimated the energy barriers for many more, and automated tools examined possible transformations among 105 chemical species on the catalyst surface, identifying which could happen in a single step. The network that emerged contained 9,389 reactions. Some summaries round that figure to nearly 10,000, but the study’s own count is 9,389. A reaction-rate model built on the expanded network raised the predicted conversion of carbon dioxide by about 40 times compared with the original model.

The paper, “Data-driven massive reaction networks reveal mechanistic pathways underlying catalytic CO2 hydrogenation,” is registered under DOI 10.1038/s41467-026-77080-4.

Methanol and carbon monoxide, not formic acid

The test of any such network is whether it predicts what happens in the lab. The small network wrongly named formic acid as the main product of carbon dioxide hydrogenation over copper. The expanded one identified methanol and carbon monoxide as the major products, which is what experiments show. Those experiments were run by collaborators at the Green R&D Centre of Hindustan Petroleum Corporation Limited, represented by G. Valavarasu and Santhosh Kotni, and at the Agency for Science, Technology and Research in Singapore, where Amol Amrute and colleagues took part. Having laboratory groups in two countries check the predictions gave the modelers an independent yardstick, which matters because a network this large can fit a story that the chemistry does not support.

Anand Verma, the first author, did the work as a postdoctoral fellow at the institute and is now at MNNIT Allahabad. He framed the modeling problem as a question of how researchers can know that a reaction network has not left out the one step that matters.

Hydrogen moving as an intact molecule

One surprise came out of the larger network. In several steps, hydrogen was transferred to the reacting species as an intact H2 molecule and not as separate atoms. Quantum calculations showed this is especially favorable when hydrogen moves onto oxygen-containing intermediates. Shivam Chaturvedi, a doctoral student in chemical engineering, said the result runs against what most chemists were taught, which was the conventional picture of hydrogen splitting into atoms on the metal first.

The authors suggest this could guide catalyst design, particularly for materials that interact more strongly with H2. They also point to possible uses of the same approach in carbon dioxide reduction with other catalysts, nitrogen reduction and water splitting. An institute listing of a talk, Using machine learning and computation to unveil mechanisms for CO2-to-fuel conversion, shows the group has been presenting the line of work, and an earlier ChemRxiv preprint on massive reaction networks carries a version of the same idea.

What the model does and does not deliver

The study is a computational mechanism map checked against experiments on one catalyst, copper, and it does not describe a new fuel plant or a finished process, and the copper surface is only one of many catalysts used for carbon dioxide hydrogenation. Nothing in the report claims a cost, an efficiency or a scale at which the chemistry would be commercially useful. What it offers is a more complete account of which steps matter, a correction to a simpler model that picked the wrong product, and a hint about how hydrogen travels that catalyst designers can test. For readers tracking carbon-capture and fuel-synthesis research, the takeaway is a better map rather than a finished route: a clearer view of the reaction steps that decide whether carbon dioxide ends up as methanol, carbon monoxide or something less useful.

The authors’ broader argument is about method. Hand-built reaction networks risk leaving out the one step that matters, and automated enumeration paired with machine-learned energy estimates can catch steps that a researcher would not think to include. Whether the same approach holds up on other catalysts and reactions is the question the team says is next. The group also presented the framework at a 2025 AIChE Annual Meeting session on data-driven massive reaction networks, an early public outing for the idea before the journal paper.

This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.

The tech news worth your time: Subscribe to Techpresso →

We may earn a commission if you sign up.