AI-assisted search rapidly pinpoints high-performance catalyst for direct CO2-to-jet fuel conversion.

Captured CO₂ can be directly converted into a sustainable jet fuel, potentially closing the loop on aircraft carbon emissions. A multidisciplinary team of catalyst researchers and AI experts at KAUST have combined experiments with machine learning to quickly identify catalysts with record-breaking performance in CO₂-to-jet fuel conversion.

Transport is among the biggest contributors to global carbon emissions. “While the electric vehicle transition has helped ensure that road transport decarbonization is already well underway, aviation presents a far more complex challenge,” says José Luis Santos, a research scientist in Jorge Gascon’s lab.

Aircraft’s strict weight and volume limitations rule out current battery- and biofuel-based low-carbon propulsion systems. The most credible pathway to aviation decarbonization is sustainable aviation fuels (SAF), Santos says. “SAF made from captured CO₂ and green hydrogen would provide a drop-in fuel compatible with existing aircraft, while helping to close the carbon cycle,” he adds.

High performance CO₂-to-jet fuel conversion catalysts are lacking, so the team initiated a search. “We used a high-throughput parallel reactor platform, capable of simultaneously testing 15 catalysts under identical conditions, to screen more than 300 catalysts with systematically varied chemical compositions,” Santos explains. They compared the CO2 conversion rate of the catalyst, and yield of hydrocarbons at least five carbon atoms in size.

Even with the high-throughput research platform, it was still difficult to cover the vast potential catalyst search space. Gascon’s group therefore teamed up with AI experts, using data from the first 300 catalysts to train a machine learning model based on Bayesian optimization to guide their search.

“Bayesian optimization sits at the intersection of statistics, machine learning and artificial intelligence,” explains KAUST AI engineer Dmitrii Khizbullin. The algorithm applies Bayes’s rule of mathematical probability to optimize the search for new catalysts based on prior knowledge of catalyst performance.

Based on their capacity to test 15 catalysts at once, the researchers used batched Bayesian optimization to automatically generate the 15 catalyst compositions that most efficiently cover the search space in each round of modelling. The experimental data from each new batch of catalysts was fed back into the model to further improve its predictive capability.

“What proved genuinely striking in this study was the speed at which the Bayesian optimization guided the search toward high-performing regions of the compositional space,” Gascon says. The team obtained the best-performing catalyst after just four rounds of catalyst iteration.

Read more at KAUST Discovery.