SGMM - Enhancing Fuel Property Prediction with Compositional Subgroup Information

Abstract

The prescreening process developed by Heyne, Rauch, Le Clercq, and Colket has become an important element in supporting development of new sustainable aviation fuel production pathways and increases the likelihood of successful ASTM approval. Prescreening involves the determination of the fuel composition at a molecular level using two-dimensional gas chromatography (GCxGC). Based on the composition, probabilistic machine learning models can then predict the properties of the fuel. However, a major limitation results from the inability of GCxGC to precisely resolve molecules on the level of structural isomers, which leads to serious disadvantages when predicting fuel properties. This limitation is particularly relevant for properties that differ significantly due to the isomeric structure of the molecules, such as freezing point and viscosity. Since these properties are strongly isomer dependent, the assumption that all known isomers can be present in the fuel has a particularly negative impact on both the inaccuracy and uncertainty of fuel property predictions. To address this challenge, the German Aerospace Center (DLR) is developing a method to identify subgroups within GCxGC bins. Building on this, we introduce the novel SubGroupMeanMatrix (SGMM) model that integrates these subgroups into the property prescreening framework to improve prediction accuracy and reduce the prediction uncertainty. In this study, we explain the methodology behind this new model and evaluate its impact on the accuracy of property prediction for selected representative aviation fuels as well as limitations of the new model.