Turning Carbon Dioxide Into Fuel Gets a Far More Complete Chemical Map

Here is the doubt that nags at a chemist building a better catalyst for turning carbon dioxide into fuel. Every design choice rests on a model of what actually happens on the metal surface, step by step. Most of those models, a new study says, carry only a few dozen reactions, when the real chemistry runs to thousands.

A team at the Indian Institute of Science in Bengaluru set out to count the rest. Working on copper, they mapped 9,389 elementary reactions. The study was published September 17, 2026 in Nature Communications, about a month ago, and the institute announced it on October 13.

Lead author Anand M. Verma put the worry plainly in the institute’s release: “How do you know that your reaction network has not omitted the one step that matters?”

Carbon dioxide as a raw material

The process is CO2 hydrogenation on copper: adding hydrogen to carbon dioxide over a copper catalyst to make products such as methanol, a liquid fuel and chemical feedstock, and carbon monoxide. It treats carbon dioxide as a raw material rather than waste.

The team, with colleagues at Hindustan Petroleum’s HP Green R&D Centre in Bengaluru and the Institute of Sustainability for Chemicals, Energy and Environment (ISCE2) at A*STAR in Singapore, first built a quantum chemistry dataset of 152 elementary reactions, one of the largest of its kind, and measured CO2 conversion on a copper catalyst in the lab. Even that network turned out to be too small. According to the institute, a model limited to those 152 reactions wrongly picked formic acid as the main product.

The full 9,389-reaction network gave about 40 times higher CO2 conversion than the original model and followed the experimental trends of methanol and carbon monoxide production. It also turned up something nobody had asked for: a crucial role for hydrogen passed between surface species, including hydrogenation by intact hydrogen molecules rather than only by single atoms. The authors call it “a surprising data-driven discovery validated post-facto,” meaning they confirmed it with further calculations after the network pointed at it. Co-author Shivam Chaturvedi said in the release that the idea that hydrogen can move “as an intact molecule, without first splitting into atoms” runs against what most people were taught.

Thousands of steps, each needing a barrier

An automated enumeration program listed the possible steps among the species on the copper surface, and a second program filtered out the ones that were not genuine elementary reactions. That alone does not make a model. Each reaction needs an energy barrier, the hill it has to climb, and computing thousands of those with full quantum chemistry would take, by the authors’ estimate, decades on a single workstation.

Machine learning closed that gap. Trained on the team’s own quantum chemistry calculations, the models estimate barriers with a mean absolute error of 0.13 electronvolts, the authors told reviewers, which is fast enough to fill in the whole network. The paper says the approach reduces “human bias in the reaction pathway.”

You cannot design for a mechanism you have wrong

Turning captured carbon dioxide into methanol at a price anyone would pay depends on catalysts that steer the reaction toward the product you want, and nobody designs well for a mechanism they have described wrongly. This work does not deliver a better catalyst. It hands the people who design them a far more complete map, plus a method the institute says could be pointed at nitrogen reduction for fertilizer and at splitting water for hydrogen.

That is a chemist’s kind of gift: not a product, but the end of a nagging doubt. For a graduate student in Bengaluru or anywhere else who has wondered whether the reactions in their model were the right ones, there is now a way to check.