Machine learning finds thousands of missing reactions in a copper CO2 hydrogenation model

Quantum mechanics is the accurate way to get a reaction's energy barrier. It is also the expensive one. Running it on every step of CO2 hydrogenation over a copper surface would demand immense computing power, so modelers usually pick a limited set of reactions they believe matter most ^[3]. The IISc team ran quantum calculations on 152 reactions to build a database, then trained machine learning models to estimate barriers for additional reactions ^[4]. Automated tools enumerated the possible reactions among 105 surface species and kept those that could happen in a single step ^[5]. That added 9,237 reactions ^[12]. The final network is about 62 times the size of the starting one ^[13].

The motive, said first author Anand Mohan Verma, was a completeness check. "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?" ^[10]

The paper's central comparison is the small network against the large one. Moving to the larger network raised predicted CO2 conversion by approximately 40 times ^[7], and the expanded model identified methanol and carbon monoxide as major products, in agreement with experiment ^[8]. "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," said corresponding author Ananth Govind Rajan ^[11]. The experiments came from G Valavarasu and Santhosh Kotni at Hindustan Petroleum's Green Research and Development Centre and from Amol Amrute and colleagues at the Agency for Science, Technology and Research in Singapore ^[9].

That comparison does not say how much of the 40-fold change any single added reaction explains. It sets the new model against the old one, a model-to-model ratio ^[7]. The IISc summary describes the hydrogen step only as surprising and at odds with established assumptions ^[14], and it does not report the experimental conversion values or conditions.

Agreement on two outputs, product identity and overall conversion, fits a more complete network. It cannot by itself show that each of the 9,237 added reactions matters ^[12].

In our view the durable result is the diagnosis: on this copper system, a curated network got the main product wrong, and a machine-learned expansion fixed it. That holds for the one catalyst studied ^[1]. According to IISc, the findings, published in Nature Communications, could help scientists develop more effective catalysts for turning CO2 into methanol ^[2].