Predictive Capability Assessment of Probabilistic Machine Learning Models for Density Prediction of Conventional and Synthetic Jet Fuels

Abstract

Machine Learning (ML) models are increasingly applied in the field of jet fuel property predictions due to their ability of modelling a high number of complex composition-property relationships directly on measurement data. Their applicability is still limited as for safety relevant use cases like synthetic fuel approval or aircraft design consequences of prediction errors might be too severe to be acceptable. For Machine Learning algorithms the predictive capability strongly depends on the data utilized for the training for the models. Predictions for fuels that that differ from the training data might have uncertainties that need to be systematically considered. We present an approach of utilizing the probabilistic ML algorithm Gaussian Process Regression to model jet fuel density over a range of -40 to 140 °C and estimating the uncertainty that results from limited training data. To assess the influence of synthetic fuels on the predictive capability two models are studied, one trained exclusively on conventional fuels data and the other one trained on the same conventional fuels and additional synthetic fuels. The predictive capability of the models is assessed using metrics to rate the accuracy and precision of the prediction, as well as the validity and reliability of the estimated prediction interval. Results show that prediction intervals can correctly be estimated by both models and a valid estimation of the predictive capability is possible. Furthermore, the addition of synthetic fuels data drastically improves the accuracy, reduces the uncertainty, and is necessary to achieve adequate predictions for the considered hold-out fuels. The presented method is transferable and can be applied for different probabilistic models and jet fuel properties

A proposed Digital Twin concept for the smart utilization of Sustainable Aviation Fuels

“…• Although a single ML model was utilized for each physical property in the Digital Twin, the predictive capabilities of the models might differ depending on the application domain [9]. Thus, it would be advisable to dynamically select adequate models for each physical property.…”

“…In addition to the data collected, predictions from computer models enable the evaluation of SAF usage over the life cycle prior to the actual deployment in an aviation mission. These models cover, inter alia, the prediction of unknown fuel properties [9], the sensitivity of the combustion process on varying fuel composition in the engine [10] or the expected climate impact, e.g. from the resulting soot emissions.…”

Advanced Fuel Property Data Platform: Overview and Potential Applications

“…This dataset is already in use within the JETSCREEN project to develop and validate machine-learning and other tools to predict important fuel specifications and performance characteristics as shown by the example in Figure 6, showing the prediction of fuel density from the GCxGC compositional results alone for a wide range of different aviation fuels (Hall et al, 2021). The term holdout is used to indicate data used for assessing the machine learning model after the training step has been completed.…”

“…FIGURE 6 | Unity plot displaying the predictive capability of machine learning models for synthetic fuels compared with data from the fuels database(Hall et al, 2021).…”

Fuel sorption into polymers: Experimental and machine learning studies

“…20 vol%). Note that Hall et al recently considered these conventional and synthetic fuels [46], and representations for these fluids are proposed hereafter.…”