Probabilistic Mean Quantitative Structure–Property Relationship Modeling of Jet Fuel Properties
Résumé
We present a novel probabilistic mean quantitative structure–property relationship (M-QSPR) method for the prediction of jet fuel properties considering two-dimensional gas chromatography measurements. Fuels are represented as one mean pseudo-structure that is inferred by a weighted average over structures of 1866 molecules that could be present in the individual fuel. The method allows training of models on both data of pure components and of fuels and does not require mixing rules for the calculation of the bulk property. This drastically increases the number of available training data and allows the direct learning of the mixing behavior. For the modeling, we use a Monte-Carlo dropout neural network, a probabilistic machine learning algorithm, that estimates prediction uncertainties due to possible unidentified isomers and dissimilarity of training and test data. Models are developed to predict the freezing point, flash point, net heat of combustion, and temperature-dependent properties such as density, viscosity, and surface tension. We investigate the effect of the presence of fuels in the training data on the predictions for up to 82 conventional fuels and 50 synthetic fuels. The results of the predictions are compared on three metrics that quantify accuracy, precision, and reliability. These metrics allow a comprehensive estimation of the predictive capability of the models. For the prediction of density, surface tension, and net heat of combustion, the M-QSPR method yields highly accurate results even without the presence of fuels in the training data. For properties with nonlinear behavior over temperature and complex fuel component interactions, like viscosity and freezing point, the presence of fuels in the training data was found to be essential for the method.