Integrating domain knowledge into direct correlation modeling to enhance property prediction for sustainable aviation fuels

Kurzfassung

Accurately predicting the properties of sustainable aviation fuel (SAF) is an enabler for its rapid market ramp-up, as targeted by the European Green Deal. Two-dimensional gas chromatography (GC×GC) can perform detailed fuel composition analysis on small samples of less than 0.5 mL, which enables feedback at an early stage for producers of new SAF candidates. The prescreening approach developed by Heyne & Rauch et al. utilizes machine learning (ML) models to predict the properties of a fuel based on its detailed composition. These models range from mixture-based approaches, which directly incorporate the knowledge about properties of single chemical compounds into the prediction, to direct correlation (DC) models, which can learn composition to property relationships from data. While for example, mixture-based models are dependent on the availability of precise mixing rules for each property, DC models can learn these independently. This study explores how existing knowledge of the properties of chemical compounds can be integrated into DC modeling to enhance the predictive capabilities. A major challenge in DC modeling is the extensive amount of data required, which is limited by the high cost of measuring fuel composition and properties. In addition, data collection is further restricted by the limited availability of fuels with significantly varying compositions. We are investigating different strategies to address this limitation with the aim of improving the predictive capabilities of DC-based models in the context of fuels.