Demonstration of a Machine Learning-Based Approach to Predict Thermophysical Properties of Species Relevant to Aviation Fuels

Sustainable Aviation Fuels (SAF) are being considered to replace current fuels, such as Jet A, to support the effort of industry and regulatory agencies to target the decarbonization of the aviation sector by 2050. Strict regulations on fuel properties, both in terms of applicability in current engines and in emission improvements (i.e., particulate matter control towards the reduction of contrails), require extensive analysis on the fuel thermophysical and chemical characteristics. The current lack of experimental data at engine-relevant pressure and temperatures for SAF candidates, motivates the exploration of accurate and robust models to capture the behavior of hydrocarbon mixtures at engine relevant conditions to support the development and deployment of net-zero carbon propulsion. This work showcases a data-driven approach based on a novel encoder-Gaussian process, which is designed to guarantee smoothness, comes with uncertainty quantification, and can incorporate physics-guided understanding as required. These capabilities are utilized for the modeling of thermophysical properties of pure species, including transcritical regimes while reducing the need for access to the critical properties. This effort arises from the shortcomings of both input properties availability and overall performance of previously investigated cubic equations of state. This paper introduces MeGS-RFM, a machine-learning based real-fluid modeling approach, and compares its performance with available databases and a volume- translated Soave-Redlich-Kwong equation of state. MeGS-RFM uses a generative modeling approach to generalize across species not available in the training datasets. Finally, we use this to demonstrate improved characterization of iso-paraffins relevant to aviation fuels, showing better agreement with the sparse datasets in open literature.