Estimation of cetane number using machine learning

This study investigates the application of machine learning for accurate cetane number prediction of single-component fuels, a crucial parameter in compression ignition engine fuels. A comprehensive dataset of cetane numbers was compiled from the literature. Chemical SuperLearner (ChemSL), a framework designed for the automated selection of molecular representations and SuperLearner model construction, was employed. The performance of various molecular representations, including Morgan fingerprints, molecular descriptors, Mol2Vec, and Coulomb matrices, was evaluated. The most favorable results were achieved for molecular descriptors, coupled with a SuperLearner model constructed by weighting the top-performing tree-based learners. The model demonstrated an overall Mean Absolute Error (MAE) of 1.454, Root Mean Squared Error (RMSE) of 4.382, and R-squared (R2) value of 0.976. This ChemSL-generated model surpassed existing literature models in cetane number prediction accuracy. Furthermore, a detailed analysis across Paraffins, Iso-paraffins, Olefins, Naphthenes, Aromatics, and Oxygenates (PIONA-Ox) chemical groups revealed high prediction accuracy, with the lowest errors observed for Iso-paraffins and Olefins. Finally, SHapley Additive exPlanations (SHAP) analysis was employed to analyze model interpretability. This analysis provided insights into the critical molecular descriptors influencing the predicted cetane number, offering valuable knowledge regarding the underlying chemical principles governing cetane number variation. •Machine learning model for predicting cetane number with high accuracy.•ChemSL framework automates model building with improved generalizability.•Key descriptors influencing cetane number were identified by SHAP analysis.•Rotational bonds ratio and molecular complexity influence cetane number.