Comparison of probabilistic jet fuel property models for the fuel screening and design

Abstract

Model-based property prediction is considered a key enabler in the screening and design of new jet fuel candidates. Combined with two-dimensional Gas Chromatography (GCxGC) for compositional analytics, they allow the prediction of fuel properties critical for the assessment of fuel candidates already from small volumes (5 ml). As safety relevant use case, the assessment of new aviation fuel candidates makes the consideration of uncertainties due to unidentifiable isomers, measurement and model uncertainty necessary. Recent developments yield several probabilistic modeling approaches that can estimate uncertainties as part of the property prediction. In this work, we present three of those approaches: Direct correlation (DC), Mean Quantitative Structure-Property Relationship Modeling (M-QSPR) and Quantitative Structure-Property Relationship Modeling (QSPR) with sampling. The approaches use Monte-Carlo Neural Networks (MCNN), a Machine Learning algorithm for the underling regression of the properties from the GCxGC measurements. We assess their predictive capabilities for the modeling of eight safety-relevant properties like the freezing point and the cetane number for conventional and synthetic jet fuels. The predictive capability assessment is based on metrics that quantify the accuracy, reliability and precision of the models. Identifying the strengths and limitations of the different approaches, we provide recommendations for the practical application of the different models and the use of model-based property prediction for the support of jet fuel assessments in general.