Modelling the Viscosity-Temperature Relationship of Alternative Fuel Blends: Comparison of Empirical and Machine Learning Models

Abstract

The viscosity of fuel blends is significant in fuel injection, atomization, and engine performance. However, accurately estimating viscosity for various blend ratios and temperatures is challenging due to the nonlinear interactions between fuel components. The available models generally lack sufficient accuracy, and thus, the researchers need advanced predictive models. Therefore, this study aims to develop more accurate empirical and machine learning models to predict the viscosity of vegetable oil-biodiesel blends and vegetable oil–diesel fuel blends. For this aim, corn oil methyl ester is produced via transesterification. The dynamic and kinematic viscosities of corn oil–corn oil biodiesel blends and corn oil–diesel fuel blends are measured at various temperatures (10 °C to 70 °C) and corn oil blending ratios (10 \(\%\) to 50 \(\%\)). A rational model is developed based on 899 viscosity data points that include experimental and literature data. The accuracy of the rational model is compared with the machine learning (linear, decision trees, support vector machine, neural network, and Gaussian process regression) and empirical models previously proposed in the literature. The rational model has the best prediction ability with the lowest overall absolute relative deviations of 0.9469 \(\%\) and 0.8789 \(\%\) for the corn oil–corn oil biodiesel blends and corn oil–diesel fuel blends, respectively, outperforming machine learning and other empirical models. These findings confirm that the rational model can accurately improve viscosity prediction of fuel blends for engine modelling and optimisation studies. The model is also capable of optimising fuel formulations, improving engine performance, and reducing emissions through optimal control of fuel properties under real-world applications.