A hybrid framework for cetane number prediction: integrating a novel cetane index formula with explainable machine learning

Cetane number (CN) is a key indicator for diesel fuel ignition quality and combustion performance. Accurate CN prediction is essential for ensuring fuel quality, improving engine efficiency, and reducing pollutant emissions. This study developed a high-precision CN prediction model using a dataset of 7,303 diesel samples. An automated machine learning framework was employed to screen multiple algorithms and optimize hyperparameters, with the deep learning model exhibiting the best predictive performance. Through feature selection and comparative analysis, seven readily available physicochemical parameters were identified: density at 20 °C, 50 %/90 %/95 % distillation recovery temperatures, cold filter plugging point, kinematic viscosity, and flash point. These parameters were innovatively integrated with a cubic-polynomial-enhanced cetane index formula to construct an optimal feature set, forming a hybrid prediction model. This model achieved excellent performance with an R 2 of 0.997, indicating improved prediction accuracy and good predictive stability within the investigated CN range. Interpretable techniques, including feature importance, partial dependence analysis, LOESS-based smoothing, and sensitivity analysis, systematically clarified the relationships between physicochemical properties and CN. Overall, this study proposes a high-accuracy, cost-effective framework for rapid and reliable assessment of diesel ignition performance, providing strong technical support for fuel quality control, production optimization, and operational decision-making in the petroleum industry.