New Physics-Based Methods for Learning Sustainable Aviation Fuel Certification Fit-for-Purpose Properties

Abstract

The development of predictive models for key physical properties of aviation fuels using atom type mole fractions derived from 1H 13C HSQC NMR spectroscopy was investigated. The approach enables property estimation from small sample volumes, supporting the prescreening of sustainable aviation fuel (SAF) candidates. Datasets were compiled for six properties: liquid density, surface tension, flash point, distillation curve, derived cetane number, and kinematic viscosity. Experimental and literature data provided 1,241 points for density, 1,260 for surface tension, 1,074 for distillation curve, 405 for derived cetane number, 184 for flash point, and 1,373 for viscosity. Models were developed using multiple linear regression, binomial regression, random forest regression, and artificial neural networks, with validation against unseen test data. The density and surface tension models achieved the highest predictive accuracy, with mean errors of 0.47 % and 2.22 % respectively. The flash point and distillation curve models also performed well, showing mean errors below 1 % and R2 values above 0.97. The derived cetane number model showed lower general correlation (R2 < 0.95) but strong agreement for real-fuel validation. The viscosity model, developed with an artificial neural network, achieved a mean error of 3.61 %. Overall, atom type based modelling demonstrates strong potential for efficient SAF property prediction, demonstrating that the properties are physical by nature. Future improvements will rely on expanding datasets to cover more complex mixtures and compositional ranges.