Comparison of quantitative structure-property relationship, sequential SMILES-property and graph based property models for the prediction of jet fuel compound properties

Jet fuels are complex mixtures of hundreds of possible molecular components. The modelling of properties for jet fuels represented as mixtures often requires the prediction of physical properties of the contained fuel components e.g. derived cetane number or net heat of combustion, if respective measurements are unavailable. The underlying algorithms used to predict the properties of molecular components need to correlate the molecular structure of the components with the desired property. There exist various modelling methods that utilize different kind of structural representations and algorithms for the correlation of a compound structure with the desired property. This work presents and compares three methods that have been recently developed or are often used for the modelling of jet fuel properties: 1) The Quantitative Structure-Property Relationship (QSPR) method, which approximates the molecular structure with structural descriptors, 2) the SMILES-Property method, that uses the Simplified Molecular Input Line Entry Specification as fingerprint and a sequential model for the correlation and 3) the Graph Based Property method, were the molecular structure is directly represented by a graph, which is correlated with the property of interest. We compare the three modelling methods with respect to their predictive capability for physicochemical properties like the density, kinematic viscosity, freezing point, cetane number and sooting tendency. The comparison is performed for compounds relevant in conventional and synthetic jet fuels.