AI-Built Chemical Map Finds the Step Catalysis Models Were Missing

Carbon dioxide hydrogenation, the reaction that turns the greenhouse gas into methanol or other usable chemicals when combined with hydrogen over a metal catalyst, has long been treated by chemists as a reaction with a short list of important steps. That assumption, it turns out, was the problem.

A team at the Indian Institute of Science has built a machine-learning-driven model of the reaction that tracks nearly 10,000 individual chemical steps on a copper catalyst surface, roughly 60 times more than researchers typically bother to calculate. The expanded map, described this month in Nature Communications, corrects a basic error in how scientists have modeled this chemistry for years and turns up a reaction pathway that textbooks do not describe.

What the smaller models got wrong

Converting CO2 into methanol, carbon monoxide, formic acid and other products is considered one of the more promising routes for recycling industrial carbon emissions into useful materials instead of simply storing them underground. But modeling precisely how the reaction unfolds on a catalyst's surface is expensive. Each individual chemical step requires computationally intensive quantum-mechanical calculations, so researchers have conventionally restricted their models to a few dozen reactions they judge most likely to matter.

The IISc group started there too, building a curated set of 152 reactions from quantum-mechanical simulations. Trained on that data, the smaller model made a prediction that does not match reality: it identified formic acid as the dominant product and substantially underestimated how much CO2 actually gets converted.

"We began with a worry familiar to anyone who does mechanistic modelling: how do you know that your reaction network has not omitted the one step that matters?" said Ananth Govind Rajan, the chemical engineering associate professor at IISc who led the research, describing the concern that motivated the project.

The numbers behind the bigger map

To answer that worry, the researchers trained machine-learning models to rapidly estimate the energy barriers of reactions they had not explicitly calculated, then used automated enumeration tools to generate every plausible single-step reaction among 105 chemical species that can exist on the catalyst's surface. That process expanded the network from 152 to 9,389 elementary reactions.

• 152 reactions in the original, hand-curated model, versus 9,389 in the expanded network
• 105 distinct chemical species considered on the copper catalyst surface
• An approximately 40-fold increase in predicted CO2 conversion once the larger network was used
• Methanol and carbon monoxide, not formic acid, emerged as the major products, matching laboratory results

The difference was not a matter of fine-tuning. Rajan said that when his team ran the smaller network, "the network wrongly predicted formic acid, not methanol, as the major product, and underestimated how much CO2 gets converted. Only when we expanded the network to include thousands of additional, previously overlooked reactions did the predictions fall in line with what we and others see experimentally."

A hydrogen shortcut that defies the textbook

Sorting through the larger set of reactions also surfaced a mechanistic detail that cuts against standard chemistry training. In several key steps, the model found that hydrogen gas can transfer directly to reaction intermediates as an intact H2 molecule, rather than first splitting into two separate hydrogen atoms before attaching one at a time, which is the pathway most catalysis researchers assume. Follow-up quantum-mechanical calculations showed the shortcut is especially favorable when the intermediate molecule contains oxygen.

"The idea that hydrogen can transfer as an intact molecule, without first splitting into atoms, runs against what most of us were taught. This surfaced only because the network was large enough to allow for it, and the observation held up when we went back and computed those steps explicitly," said Shivam Chaturvedi, a PhD student in IISc's Department of Chemical Engineering and a co-author of the study.

Chaturvedi added that the finding has practical implications: catalysts engineered to bind hydrogen gas more strongly could, in principle, be designed to favor the pathways that lead to methanol rather than less useful byproducts.

Who stands to benefit, and how it was checked

The modeling work did not stand alone. Experimental validation came from Hindustan Petroleum Corporation's Green Research and Development Centre in Bengaluru, where researchers G. Valavarasu and Santhosh Kotni tested the model's predictions against laboratory results, and from Amol Amrute and colleagues at Singapore's Agency for Science, Technology and Research. Machine-learning components of the framework drew on work from Ambedkar Dukkipati, a computer scientist also at IISc, underscoring how the project leaned on expertise spanning chemistry, computing and industrial catalysis.

That combination matters beyond one laboratory's curiosity, and it is part of why the paper, since its publication, has been picked up by outlets such as ScienceDaily that track peer-reviewed research for a general audience. Refiners and chemical manufacturers exploring carbon capture and utilization, an industry racing to find economically viable uses for captured CO2, depend on accurate models to avoid years of trial-and-error catalyst design. A model that gets the dominant product wrong, as the smaller network did, could send an expensive research program down the wrong path for years before the error is caught in the lab.

What happens next

The IISc team, whose first author Anand Mohan Verma carried out the work as a postdoctoral fellow at the institute before moving to a faculty post at the Motilal Nehru National Institute of Technology Allahabad, says the same combination of quantum-mechanical simulation, machine learning and automated reaction enumeration is not specific to CO2-to-methanol chemistry. Rajan's research group says the framework could be adapted to map reaction networks for other catalytic processes central to industrial decarbonization, including CO2 reduction on different metal catalysts, nitrogen reduction for fertilizer production, and water splitting to generate hydrogen fuel.

None of that is an immediate commercial product. The paper is a modeling advance, not a new catalyst ready for a refinery floor, and the authors are careful to frame the 40-fold conversion increase as a correction to a flawed prediction rather than a demonstrated improvement in an actual reactor. But by exposing how badly an incomplete reaction network can mislead researchers, the work gives catalysis labs a cautionary data point and a new computational tool as they chase cheaper, more selective ways to recycle carbon dioxide into fuel.