Cetane number prediction for hydrocarbons from molecular structural descriptors based on active subspace methodology

Abstract

Abstract Cetane number (CN) is a key factor assessing auto-ignition tendency for diesel fuels. The active subspace method and the quantitative structure-property relationship strategy were implemented to predict the cetane numbers of hydrocarbons. 110 experimentally measured CNs of hydrocarbons including n-alkanes, iso-alkanes, cycloalkanes, alkenes and aromatics were used in this study. Easily obtained topological indices and carbon-chain related descriptors from the molecular structures were used to build a one-dimensional prediction model. The contribution of each individual descriptor was provided as a byproduct using this model. The coefficient of determination of the prediction for the testing data was 0.93 and the average absolute error was 5.0, which was comparable to the level of experimental error. To further save the computational cost, a model with fewer descriptors was established based on the ranking of the contribution of each descriptor. The coefficient of determination of this reduced model for the testing data was 0.89 and the average absolute error was 5.6, indicating that the prediction performance of this reduced model was not significantly sacrificed. This as such confirms the capability of the active subspace method to simplify the descriptors selection.