AI Maps 10,000 Reactions to Turn CO2 Into Fuel

A team at the Indian Institute of Science has used machine learning to map nearly ten thousand chemical reactions involved in turning carbon dioxide into fuels and chemicals, and in the process caught an error that had been quietly skewing this kind of modelling for years. The work, published in Nature Communications and summarised by Phys.org, points toward catalysts that could convert CO2 far more efficiently than current models suggest.

The problem the researchers tackled is a practical one. Simulating a chemical reaction from first principles, using quantum mechanics, is extraordinarily expensive in computing time. So chemists usually pick a small handful of reactions they believe matter and model only those. The risk is obvious once you say it out loud. If an important step is not on the list, the model never sees it.

The Bengaluru team started with 152 reactions calculated the slow, rigorous way, then trained machine learning models to predict the energy barriers for thousands more. Automated tools generated every plausible reaction among 105 chemical species sitting on a copper surface, ending up with a network of 9,389 of them. Copper is the metal of choice for this chemistry, and the catalyst most likely to turn up in any real plant.

The missing steps

The payoff came when the fuller picture changed the answer. With only the original 152 reactions, the model predicted that the main product would be formic acid. The expanded network instead predicted methanol and carbon monoxide, which is what experiments actually produce. Fed into a kinetic model, the larger network predicted roughly forty times more CO2 conversion than the smaller one, and it surfaced an unexpected hydrogen reaction that could point toward better catalyst designs.

That last detail is the interesting one for anyone hoping to use the work. A model that merely agrees with experiment is reassuring. A model that explains why a reaction runs better than expected gives chemists something concrete to aim at.

Beyond carbon capture

The method is not tied to this one reaction. The researchers, led by Ananth Govind Rajan with Anand Mohan Verma as first author, say the same approach can be pointed at other industrially important processes, nitrogen reduction among them, where the same combinatorial explosion has forced the same corner-cutting.

It fits a pattern worth tracking, of AI being put to work on the unglamorous but genuinely hard parts of science rather than chasing headlines. China's mathematicians recently formed an alliance to do something similar for proofs. The common thread is less about machines replacing researchers and more about clearing away drudgery that human time and compute budgets could never quite cover.