The Catalyst That Wasn’t There: How a Model Confidently Predicted the Wrong Chemical
When researchers at the Indian Institute of Science built a computational model of carbon dioxide hydrogenation on a copper catalyst, the software did not hesitate. [2] It produced a clean answer: formic acid was the dominant product. [2] The prediction was specific, internally consistent, and derived from quantum mechanical calculations — the gold standard for this kind of chemistry. It was also wrong. Experimental measurements showed methanol as the major product and a far higher rate of CO2 conversion than the model had allowed. [2] The machine had not failed loudly. It had failed quietly, in the manner of an expert who has read only part of the file and sees no reason to mention it.
This is the shape of the problem the IISc team set out to address, and it is the shape of a much broader difficulty with computational chemistry. A model trained on a curated set of reactions will report whatever those reactions imply. It has no mechanism for flagging what it was never shown. The gap between what such a system appears to know and what it actually covers is invisible from the outside — and, as the researchers found, it can be wide enough to invert the central conclusion.
The Worry That Started It
The project began with a question that anyone who builds mechanistic models will recognize. Anand Mohan Verma, who conducted the work as a CV Raman Postdoctoral Fellow at IISc and is now an Assistant Professor at MNNIT Allahabad, put it plainly: “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?” That worry is not paranoia. It is the structural condition of any reduced model. Every simplification is a bet that the discarded material does not matter, and the bet is usually placed without a way to check it.
Conventional practice in this field accepts that bet as a necessity. Simulating every possible elementary step on a catalyst surface with quantum mechanics would demand computing resources far beyond what any group can marshal. So researchers select a few dozen reactions judged most important and build their kinetics on that foundation. The selection reflects chemical intuition, prior literature, and human judgment about what is plausible. It is a reasonable procedure. It is also, by construction, a procedure that cannot detect its own blind spots.
From 152 Reactions to 9,389
The IISc group, working with collaborators at Hindustan Petroleum Corporation Limited’s Green Research and Development Centre and at the Agency for Science, Technology, and Research in Singapore, took a different route. [1] They first ran density functional theory calculations to assemble a detailed database of 152 elementary CO2 hydrogenation reactions. [1] They then trained machine learning models to estimate activation barriers for reactions outside that set, so that new steps could be evaluated without a full quantum mechanical treatment each time. Automated enumeration tools then combed through the possible transformations among 105 chemical species present on the catalyst surface, including which conversions could occur in a single step.
The result, published in Nature Communications (DOI: 10.1038/s41467-026-77080-4), was a network of 9,389 elementary reactions — roughly sixty times the size of the starting set. [1] The expansion was not a refinement of the original picture. It supplemented it. With the larger network embedded in a kinetic model, predicted CO2 conversion rose by approximately 40 times, and the product distribution shifted to match experiment: methanol and carbon monoxide as the major outputs. Ananth Govind Rajan, Associate Professor in the Department of Chemical Engineering at IISc and the corresponding author, described the discrepancy directly: “When we modeled the process using the 152 reactions considered initially, 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 Pathway Nobody Was Looking For
The expanded network also surfaced a mechanism that the smaller one could not have contained. In several important steps, hydrogen appeared to transfer to intermediate compounds as an intact H2 molecule rather than splitting first into separate atoms. Additional quantum mechanical calculations confirmed that this molecular pathway is especially favorable when the hydrogen lands on oxygen-containing intermediates. The finding challenges a long-standing assumption in the field: that catalytic hydrogenation proceeds through atomic hydrogen generated by dissociation on the metal surface.
Shivam Chaturvedi, a PhD student in the Department of Chemical Engineering at IISc and a co-author, framed the surprise in terms of training: “The idea that hydrogen can transfer as an intact molecule, without first splitting into atoms, runs against what most of us were taught.” He added: “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.” The implication for catalyst design is concrete — catalysts that interact more strongly with H2 could potentially enhance pathways leading to methanol — but it surfaced as a byproduct of scale, not as a hypothesis anyone set out to test.
What the Machine Still Cannot Tell You
The IISc approach reduces human bias in pathway construction, but it does not eliminate the need for human judgment. The enumeration of 9,389 reactions depended on choices about which species to include, which bond rearrangements to consider, and which machine-learned barrier estimates to trust. The expansion corrected a specific failure, but it cannot certify that the new network is complete. A still larger enumeration might reveal further pathways, and the same logic that justified expanding from 152 to 9,389 applies to whatever number comes next.
This is the honest limit of the method, and it is worth stating plainly because the alternative — treating the expanded model as authoritative — would repeat the original error at a higher level of sophistication. The value of the work lies not in having found the final answer but in having demonstrated that the previous answer was demonstrably incomplete, and in providing a systematic procedure for finding what was missing. That procedure is transferable: the same combination of quantum mechanics, machine-learned barriers, automated reaction enumeration, and kinetic modeling could be applied to nitrogen reduction, water splitting, or CO2 reduction on other catalysts.
The Next Question
The practical consequence for catalyst development, in this reading, is a shift in what counts as evidence. A model that agrees with experiment on product distribution is no longer sufficient validation if the reaction network behind it was assembled by hand from a shortlist. The IISc results indicate that agreement can be achieved by a network that is missing thousands of steps, and that the missing steps can change the answer qualitatively. The next step in this line of research is not simply to build larger networks but to develop ways of estimating how much a given network might still be leaving out — a measure of residual ignorance rather than a claim of completeness Machine learning made the enumeration of nearly ten thousand reactions tractable, and that enumeration overturned a prediction that had looked solid. The tool did not deceive anyone deliberately; it answered the question it was given, within the boundaries it was given. The deception, in this account, lay in the appearance of completeness, and the remedy was to widen the boundaries until the answer changed. Whether the same remedy will work at the next scale, and how researchers will know when to stop widening, remains the open question the work leaves behind.
