Quantifying Fuel Similarity for Reliable Property Predictions of Sustainable Aviation Fuels

Kurzfassung

The accurate prediction of properties governing safety and operability is crucial for advancing the development of sustainable aviation fuels (SAFs). Two-dimensional gas chromatography (GCxGC) provides detailed chemical compositions that have proven useful for training probabilistic machine learning models, particularly when fuel volumes are limited. However, these models often struggle with extrapolation outside the training data's range, limiting their application to similar fuels. In some instances, even this interpolation can become unreliable due to overfitting, particularly when the training dataset is small relative to the large number of features extracted from the chemical composition. To address these challenges, quantifying the similarity between fuels based on composition becomes essential. A variety of metrics have been assessed, considering both the high dimensionality of the data and the interdependencies among features, such as correspondence to the same hydrocarbon family. Based on the experimentally determined property values of closely placed fuels, the adequacy of these similarity metrics is determined. By employing these metrics, we explore the conditions under which the application of predictive models is valid, offering an additional criterion - fuel similarity - beyond the uncertainty intervals of the predictions, to assess model trustworthiness. Additionally, different feature encoding strategies are evaluated to determine their compatibility with the proposed similarity metrics. These dimensionality reduction techniques provide a pathway for future machine learning models to minimize overfitting, especially in scenarios where training data is scarce.