BayeSAF: Emulation and design of sustainable alternative fuels via Bayesian inference and descriptors-based machine learning

As part of mid-century net-zero carbon neutrality pathways, sustainable aviation fuels (SAFs) constitute the only viable decarbonization strategy for commercial aviation in the next two decades, with hybrid-electric propulsion architectures emerging exclusively in short-range regional aircraft operation scenarios. Still, reliable prescreening of alternative jet fuels is crucial in facilitating the de-risking and scale-up of SAF production technologies. In the present study, we illustrate an innovative methodology to support the physicochemical characterization of emerging SAFs and to understand the impact on combustor operability figures of merit (FoM). First, we develop a hydrocarbon property database built upon experimental measurements and descriptor-based machine learning (DB-ML) for individual components of candidate SAFs, encompassing significant isomeric variance for seven hydrocarbon molecular groups and 992 chemical compounds. Then, we build the MATLAB©-based BayeSAF framework, which leverages Bayesian inference and global sensitivity analysis (GSA) to formulate surrogate mixtures emulating the physicochemical properties of candidate SAFs or driving fuel design processes. Lastly, to assess the suitability of the BayeSAF algorithm, we develop ad-hoc surrogate mixtures emulating the alcohol-to-jet C-1 POSF-11498 test fuel and the conventional Jet A-1 POSF-10264 fuel. While the maximum-a-posteriori (MAP) surrogates accurately replicate the chemical composition and physicochemical properties of the reference fuels, the comprehensive statistical description of the surrogate composition delivered by BayeSAF valuably accounts for the intrinsic uncertainty arising in jet fuel emulation and design. This way, we offer the community a robust methodology to facilitate multi-fidelity numerical analyses addressing combustor operability FoM, including (i) zero- and one-dimensional reduced-order models (ROMs) for vaporization and combustion characterization and (ii) three-dimensional reacting computational fluid dynamics (CFD) targeting preferential vaporization effects in realistic aeronautical combustion chambers.