A group of researchers from Brazil, the United Kingdom and Spain has developed an advanced fuel design system using artificial intelligence and machine learning methods. The system makes it possible to significantly speed up the process of creating new fuel mixtures: instead of multi-week laboratory tests, promising compositions can be selected in just a few hours. As a result of their work, the scientists have already managed to generate 157 fuel recipes that have all characteristics required for use in diesel engines and are capable of reducing soot emissions by more than 70% compared to traditional diesel fuel.
In industries where switching to electricity is impossible or extremely difficult to do (such as heavy transport, aviation and marine transportation), liquid fossil fuels remain an indispensable source of energy. However, climate challenges are driving more and more attention to alternatives, including synthetic and biological liquid fuels. Yet the development of these fuels faces a number of difficulties. New fuels must combine flammability (i.e., the appropriate cetane number), low soot formation, compatibility with existing engines and a minimal carbon footprint. At the same time, each fuel is a complex mixture of tens or hundreds of compounds, and testing the properties of each new combination in a lab or on propulsion systems is a lengthy and expensive process.
In order to streamline and speed it up, researchers from Queen Mary University of London, the Federal University of Rio de Janeiro and the Barcelona Supercomputing Center have developed a multi-stage digital fuel design system. The system is based on the use of numerical descriptions of molecules, or so-called molecular fingerprints, which allow the chemical structure of compounds to be represented in a format suitable for machine learning. These data were fed to a deep neural network capable of accurately predicting key fuel properties (including its flammability and soot formation). The trained model was used to reverse engineer results: the researchers set the target properties of the fuel, and the system selected combinations of substances that provide the required characteristics in view of technological and regulatory constraints.
To train the model, the researchers collected large databases containing the physical and chemical properties of hundreds of fuel components and their mixtures. Specifically, the cetane number database included 708 measurements for 475 substances, while the soot index database included over 400 pure compounds.
The results exceeded expectations. The trained model demonstrated high accuracy: the determination coefficient (R²) exceeded 0.9 even when predicting the properties of new, previously unseen compounds. A total of 157 fuel mixture recipes were generated, including options with additives of alcohols, oxymethylene ethers (OMEx) and fatty acid methyl esters (FAME). The best of these recipes not only met current diesel fuel requirements, but also demonstrated a significantly lower soot index, which means cleaner combustion. Calculations show that such mixtures can reduce soot emissions by up to 72% without deteriorating performance.
The new system paves the way for the rapid and targeted development of sustainable fuels that could play a key role in reducing the carbon footprint of heavy transport in the coming years.
