Researchers at the Indian Institute of Science combined quantum mechanics simulations with machine learning to map the chemical reactions that turn carbon dioxide into fuels and industrial chemicals. Starting from a conventional network of just 152 reactions on a copper catalyst, they used automated tools to explore thousands of additional possibilities, ultimately building a network of 9,389 individual reaction steps.
The expanded network dramatically changed the model's predictions. The original smaller network wrongly identified formic acid as the main product and underestimated how much CO2 gets converted. With thousands of additional reactions included, the predicted CO2 conversion increased by about 40 times, and the model correctly identified methanol and carbon monoxide as major products, bringing it in line with experimental observations.
The larger network also revealed a surprising chemical pathway: in several important steps, hydrogen can transfer directly as an intact H2 molecule to intermediate compounds, rather than first splitting into individual atoms. This mechanism was especially favorable when hydrogen moved to oxygen-containing intermediates, challenging a long-standing assumption in catalytic chemistry and suggesting that catalysts interacting more strongly with H2 could enhance methanol production.
The findings, published in Nature Communications, were validated experimentally by researchers at Hindustan Petroleum Corporation Limited and A*STAR in Singapore. The team notes that the discovery emerged only because the network was large enough to include reactions that conventional models tend to omit.