Article révisé par les pairs
Résumé : The aviation industry’s dependence on liquid fossil fuels makes it one of the most challenging sectors to decarbonise. Sustainable aviation fuels (SAF) offer a promising pathway; however, their widespread deployment is constrained by high production costs and the lack of systematic design tools for identifying viable drop-in fuel candidates. To address this challenge, this work presents an AI-guided de novo fuel discovery framework based on deep kernel learning to establish a high-fidelity mapping between molecular structure and key physicochemical fuel properties. The proposed probabilistic surrogate model demonstrates strong predictive performance, achieving coefficients of determination exceeding 0.91 across all target properties while providing calibrated uncertainty estimates through probabilistic inference. Quantitative uncertainty diagnostics, including negative log-likelihood, continuous ranked probability score, interval sharpness, and predictive uncertainty calibration analyses, demonstrate robust predictive reliability across the evaluated fuel-property space. Integrated with a virtual high-throughput screening framework, the proposed methodology enables rapid exploration of high-dimensional blend-composition spaces and identification of candidate SAF formulations whose predicted properties are consistent with, or exceed, those of JP-8 and Jet A reference fuels. The identified SAF blends additionally exhibit an estimated reduction in particulate emissions exceeding 17%. Repeated optimisation and sensitivity analyses further demonstrate robustness with respect to stochastic initialisation and optimisation hyperparameters. Importantly, these results should be interpreted as prediction-guided identification of promising SAF candidates rather than direct experimental validation. Although the predicted properties align with key certification-relevant specifications, full assessment of 100% drop-in capability, operational compatibility, and certification readiness requires comprehensive experimental validation and qualification according to established aviation standards. Overall, this work establishes an AI-enabled framework that integrates predictive modelling, uncertainty quantification, and virtual high-throughput optimisation to accelerate the discovery and prioritisation of sustainable aviation fuel candidates, supporting more efficient exploration of the SAF design space and contributing to the aviation sector’s transition towards net-zero emissions.

Documents en relation

DI-fusion