Evaluation of n-component surrogate mixtures formulated for jet fuel physicochemical property predictions

Abstract

Compression ignition (CI) engines operating on aviation fuels are sensitive to fuel property variation but can maintain robust ignition with inline, feedforward sensing and control. Due to the expensive and time-consuming nature of conventional fuel property testing methods, alternate methods like Quantitative Structure Property Relationship (QSPR) models have been developed, which link fuel properties to their molecular structures, but many fail to account for the complex intermolecular interactions in multicomponent mixtures like jet fuels. Recent studies have employed machine learning (ML) models trained on surrogate mixtures of jet fuels to predict fuel properties like the derived cetane number, density, and viscosity; and they show good performance. These surrogate mixtures, composed of hydrocarbons representing real fuel components, facilitate dataset generation for ML model training. However, the sufficiency of such datasets has not been studied exhaustively. This study investigates the sufficiency of datasets composed of jet fuel surrogate mixtures in predicting fuel properties for fuels beyond the training set. While dataset sufficiency does depend on the population and the property modeled, as a heuristic, we emphasize spanning the chemical functional groups present in likely mixtures and doing this with low correlation and high diversity in the chemical functional groups. A large dataset of surrogate mixtures was generated spanning the ranges of UNIversal Functional Activity Coefficients (UNIFAC) functional groups present in real jet fuels. This study establishes criteria for dataset sufficiency, suggesting a minimum dataset size for developing a robust ML model. We also report the tools and codes developed for determining the minimum number of surrogate mixtures needed for the purpose of training ML models given user defined UNIFAC functional group compositions and a palette of suitable neat hydrocarbon components. This research contributes to developing efficient, accurate ML models for predicting the properties of complex fuel mixtures, with applications in fuel formulation, sensing, and control.