Multi-feature optimization of QSPR model for fuel octane number

Abstract

Octane number is one of the most important indicators in gasoline, and the standard method for determining octane number is time-consuming and expensive. In this study, a set of quantitative structure-property relationship (QSPR) descriptors of fuel molecules were used, and a combination method of variance filter and recursive feature elimination was applied to screen the descriptor subset that influences the prediction accuracy of octane number. The motor-octane numbers (MON) of 82 hydrocarbons were collected, and five different molecular descriptor tools: E-dragon, ChemoPy, CDK, RDKit, and PaDEL, were used to calculate five molecular descriptor libraries (MDLs). These MDLs were then integrated into a large comprehensive molecular descriptor library (CMDL), and the optimal subset was selected using the above-mentioned method. Screening models were established between these five MDLs, CMDL, and MON, and the molecular descriptor subset that contributed the most to MON prediction was found using evaluation models. The results showed that the globally optimal descriptor subset was from CMDL, while the optimal descriptor subsets in E-dragon and PaDEL of the MDLs have similar contributed to MON prediction and better than ChemoPy, CDK, RDKit.