In silico estimation of density, vapor pressure, distillation curve, and octane numbers of gasoline fuels using machine learning

Abstract

This study aimed to develop a machine learning model for accurately estimating gasoline’s mixture density, vapor pressure, distillation curve, and research and motor octane numbers (RON and MON) using detailed hydrocarbon analysis (DHA) data to design sustainable fuels. DHA data from over 2200 gasoline samples collected globally were used for model training. The DHA component names were converted into Simplified molecular-input line-entry system (SMILES) representations, a molecular formula language. These SMILES strings were further transformed into molecular descriptors, which capture various physicochemical properties of the molecules. The molecular descriptors were then weighted by the corresponding volume fractions of the DHA components and fed into the model. The model was a SuperLearner ensemble, combining the top five best-performing learners (out of 40 base learners) identified through a data-driven selection process. The model’s performance was assessed using Mean Absolute Error (MAE) to evaluate its accuracy on unseen test data. The MAE achieved for mixture density was 0.001 g/cc, vapor pressure was 0.371 psi, distillation curve ranged from 1.110 to 2.764 ${}^{\circ }$C, RON was 0.635, and MON was 0.922. Notably, the MAE achieved by the model for the distillation curve and octane numbers was within the established reproducibility tolerances.