Stereo-sensitive Modelling of Gas Chromatographic Retention Indices of Mono-cycloalkanes in Jet Fuel Range

Abstract

Accurate assessment of novel aviation fuels requires detailed insight about structural differences relative to conventional jet fuels. To support compositional characterization, retention behavior of cycloalkanes needs to be investigated further to evaluate fuels on an isomeric level. Current available literature models for the prediction of saturated cyclic retention indices (RI) are limited in carbon range or by neglection of stereo isomeric effects. This study addresses the challenge of comprehensive RI prediction of mono-cycloalkanes by utilizing available retention data from the NIST23 database and the combination of descriptors from PaDEL, Mordred and RDKit. Multiple model approaches (extreme gradient boosting, multi-layer perceptron regression, support vector regression and ridge regression) were tested via repeated k-fold cross-validation to evaluate robustness and reliability in prediction. As the available literature data was skewed, model tests were run for data up to carbon number C13 and the whole dataset to C20 individually. Cross-validation results showed the best error metrics for the validation sets with ridge regression (e.g. C13-models mean absolute error: MAE = 11.01 ± 1.32 RI; C20-model MAE = 13.68 ± 1.67 RI). Correct stereo sensitive retention of diastereomeric structures was achieved in 78 %, with cis/trans differentiation being the best represented stereo effect. These advancements will enable a more detailed characterization of compositional differences and will subsequently enhance evaluation of novel fuel production paths.