The motivation behind the work stems from a fundamental tension in the race toward net-zero emissions. Electrification is advancing rapidly, but internal combustion engines remain indispensable for heavy-duty transport, from long-haul trucking to shipping and agriculture, where batteries struggle to deliver the energy density these applications demand. Renewable synthetic fuels, produced by combining renewable electricity with captured carbon dioxide through processes such as thermal and electrocatalysis, offer a compelling alternative because they are compatible with existing fuel infrastructure and can store energy over long periods. Yet if these fuels are simply dropped into engines designed around fossil diesel, much of their potential is wasted. The researchers argue that fuels and engines must be co-optimized, with the fuel’s properties deliberately shaped to match the combustion system that will burn them.

Conventional fuel design falls into two camps, and both have limits. Passive fuel design involves blending components in fixed ratios before the engine runs, while active fuel design dynamically adjusts injection ratios during operation. Both are composition-driven: they can only optimize within the limited palette of candidate components on hand. Renewable synthetic fuels such as Fischer-Tropsch synfuels break through this constraint because their composition can be tuned at the synthesis stage itself, through adjustments in feedstock, catalyst selection, and reaction conditions. But this expanded freedom creates a combinatorial nightmare. Synthetic fuels can contain dozens to hundreds of components, and the number of possible mixtures explodes beyond what trial-and-error experimentation or brute-force simulation could ever explore.

The team’s solution draws on an idea known as the central fuel property hypothesis, developed under the US Department of Energy’s Co-Optima program. The hypothesis holds that the macroscopic combustion behavior of a fuel is dictated primarily by a small set of key physicochemical properties rather than by its exact molecular makeup. If that is true, the vast space of possible fuel compositions collapses into a compact set of property variables, and those properties can serve as a bridge between what a fuel is made of and how an engine performs. The researchers selected five properties for conventional diesel combustion: density, which governs spray penetration and mixture formation; cetane number, which controls autoignition reactivity and ignition delay; lower heating value, which determines the energy released per kilogram of fuel; oxygen content, which influences oxidation rates; and the yield sooting index, a measure of a fuel’s intrinsic tendency to produce soot.

This property-based representation also solves a second, equally stubborn problem: data scarcity. Unlike biomedical science, materials science, or geophysics, engine research has no large open-source datasets, and fuel composition data from different studies cannot be meaningfully merged because the fuels themselves differ so wildly, from gasoline and diesel to alcohols, ethers, and biodiesel with various additives. By converting every fuel into the same five physicochemical properties, the framework makes heterogeneous data from different laboratories comparable. The team mined published experimental studies on conventional diesel combustion engines, standardized units, applied domain knowledge to derive missing variables, and assembled a standardized dataset of 986 samples spanning diverse engine platforms, fuel types, and operating conditions. Because the resulting distributions were skewed and imbalanced, the researchers used a structured splitting strategy combining k-means clustering with label-stratified sampling to ensure that training, validation, and test folds all faithfully represented the full dataset.

The predictive model at the heart of the framework is architecturally inventive. Engine experiments in the literature fall into three intake modes: simulated boosting, naturally aspirated operation, and exhaust gas turbocharging, and intake conditions are frequently not reported for the latter two. To handle this variability, the team built a conditional expert architecture inspired by the mixture-of-experts concept. A Transformer encoder serves as a shared backbone, learning latent representations of the common input variables, while intake-mode-specific multilayer perceptron expert modules are dynamically activated depending on which intake mode applies. Crucially, the model embeds physics directly into its learning. The work produced per cycle equals the mean effective pressure multiplied by cylinder displacement, so thermal efficiency can be independently inferred from the predicted pressure. The deviation between the model’s directly predicted efficiency and this physics-derived value is penalized during training and used during inference to select, among an ensemble of five sub-models, the prediction with the smallest physical inconsistency.

The model’s performance on an independent test set was impressive: a coefficient of determination of 0.994 for mean effective pressure and 0.921 for thermal efficiency. With this predictive engine in hand, the researchers deployed the NSGA-II multi-objective optimization algorithm to search for fuel property combinations that would maximize efficiency in a naturally aspirated, four-cylinder diesel engine running at 1,800 rpm under a specified load. The optimization produced a Pareto front of solutions, from which the team retained those with physical consistency deviations below one percent and selected the most efficient. The winning combination differed from diesel in telling ways: slightly lower density, a modestly reduced cetane number, a lower heating value, and most dramatically, an oxygen content of 14.55 percent by weight and a yield sooting index of just 13.06, compared with diesel’s 125.82.

Computational fluid dynamics simulations revealed why these properties work together so effectively. The lower cetane number lengthens the ignition delay, giving more time for the lower-density fuel to evaporate and mix with hot air, which increases the proportion of premixed combustion. The higher oxygen content accelerates oxidation, so combustion starts later but proceeds faster, concentrating heat release and shortening the combustion duration. The reduced heating value smooths the pressure rise that this faster combustion would otherwise produce. Meanwhile, the shrunken fuel-rich zones and enhanced oxidation slash soot formation, and the shorter persistence of high-temperature reaction zones curtails the thermal residence time that generates nitrogen oxides. The efficiency gain, the soot reduction, and the emissions cut all emerge from the same coordinated set of property changes, which is precisely the kind of synergistic design that composition-driven blending struggles to find.

The framework’s implications reach well beyond a single simulation. The optimized property combination can serve as a concrete target for the synthesis stage of renewable fuel production, guiding catalyst and feedstock choices rather than leaving fuel developers to grope through unguided trial and error. The researchers are candid about limitations: the dataset, though carefully constructed, remains modest; the demonstration covered one engine and one operating condition; and the optimized fuel was represented by conventional surrogate components rather than an actual synthetic fuel. Extending the framework to spark-ignition engines would require swapping cetane number for octane number, while other combustion modes and aviation fuels would demand additional property descriptors such as volatility, auto-ignition temperature, or aromatic content. Future work aims to validate the property targets in real engine experiments and to close the loop from design through synthesis to testing. If that loop can be closed, the era of fuels engineered atom-by-atom for the engines that burn them may be closer than it appears.

Subject of Research: AI-driven property-oriented design of renewable synthetic fuels for sustainable internal combustion engines

Article Title: An AI-driven framework for property-oriented fuel design toward sustainable engine combustion

Article References: Liu, Y., Li, A., Li, Y., Fan, Y., Qi, Y., Zhang, C., Zhu, L., & Huang, Z. (2026). An AI-driven framework for property-oriented fuel design toward sustainable engine combustion. iScience, 29(10), Article 117677. https://doi.org/10.1016/j.isci.2026.117677

Image Credits: AI Generated

DOI: 10.1016/j.isci.2026.117677

Keywords: artificial intelligence, fuel design, internal combustion engines, renewable synthetic fuels, diesel combustion, machine learning, thermal efficiency, emissions reduction, Fischer-Tropsch, soot, nitrogen oxides, Co-Optima