Deep-Learning-Based Sequential Bayesian Updating and Uncertainty Quantification of Vaporization of Aviation Fuel Surrogates: A Case Study on HEFA-SPK

Abstract

Sustainable aviation fuel (SAF) surrogates developed in the literature generally target gas-phase combustion characteristics of the reference fuel, with less emphasis on two-phase behavior. Notably, although evaporation is a rate-limiting process in jet engines, SAF surrogates that emulate droplet vaporization experimental data are not available, thereby limiting the ability to predict preferential vaporization effects on ignition, lean blow-out (LBO), and cold-start behavior.

In this contribution, building on the BayeSAF framework, we propose a sequential Bayesian updating strategy that comprises a first Bayesian inference stage targeting a physicochemical surrogate and a second stage, informed by the first, addressing evaporation experiments. To limit the computational overhead, we rely on a droplet evaporation pseudo-model yielded by a long short-term memory (LSTM) deep recurrent neural network. We assess sequential Bayesian updating by formulating a HEFA-SPK physicochemical-evaporation surrogate and observe good agreement against both physicochemical targets, with relative errors < 5% for the maximum-a-posteriori (MAP) composition, and evaporation data, largely within the 95% confidence interval. Lastly, by performing uncertainty quantification (UQ) of the surrogate vaporization characteristics, we demonstrate the potential of the proposed framework to (i) provide robust predictions of SAF evaporation and (ii) highlight improvement directions in surrogate formulation via global sensitivity analysis.