The Estimation of Latent Heat and Vapor Pressure of Ethanol–Gasoline Blends Using Machine Learning and Thermodynamic Relations

Abstract

Latent heat of vaporization (LH_vap) is a crucial property in internal combustion engines (ICEs). It affects the cylinder temperature (T), ignition delay, NO_x emission and other phenomenon in ICEs. With the increase in global warming, use of alternative fuels in ICEs such as gasoline–ethanol blends, gasoline–methanol blends have become evident. Another important use of LH_vap values is encountered while performing the spray combustion analysis of blended fuels using Computational Fluid Dynamics (CFD). This paper investigates the use of various Machine Learning (ML) techniques to predict the latent heat of vaporization (LH_vap) for blended fuels. The algorithms used were Linear Regression (LR), Polynomial Regression (PR), Support Vector Machine (SVM), K Nearest Neighbors (KNN), Decision Tree (DT), and Random Forest (RF). The features used were Temperature, Blend Ratio, Molecular weight, Carbon (wt%), Hydrogen (wt%), and Oxygen (wt%). For training the algorithms, data was collected from several published research papers with LH_vap values of gasoline-alcohol fuel blends and various diesel blends. The model was initially trained with data pertaining to gasoline-alcohol blends only, which showed that LR performed better than other algorithms in predicting both LH_vap, with a coefficient of determination (R² score) of 90% and Mean Absolute Percentage Error (MAPE) of 6.2%. This behavior was attributed to the linearity in the data, as most of the data points were of gasoline-alcohol blends with different blend ratios and temperatures. However, when more datapoints were included such as various oxygenated blends of diesel, it was found that RF performed much better than other algorithms, with an R² score of 95.7% and MAPE of 6.8%. Furthermore, additional features were considered such as vol% of paraffins, aromatics, olefins, iso-paraffins, cyclo-paraffins, napthenes along with original features. The results showed that predictions improved with adding these features as the R² score improved for all algorithms. It could be summarized that with availability of more data, the performance of RF algorithm could further improve. Thermodynamic relations have been attempted to correlate predictions of latent heat of vaporization with vapor pressure data of the blends.