Interpretable tree-based ensemble models for macroscopic characterization of biofuel sprays

Abstract

The growing energy demand and strict emission policies are driving the exploration of sustainable fuels for internal combustion engines. Biofuels from vegetable oils and waste food are promising candidates for future advanced combustion technologies. Since fuel spray significantly impacts engine performance and emissions, optimizing fuel injection systems requires a thorough understanding of spray characteristics. This study leverages machine learning (ML) techniques to characterize biofuel sprays, addressing the complexities and costs of experimental setups. Using a dataset from direct imaging of sprays, tree-based models like extreme gradient boosting (XGBoost) and random forest (RF) are trained with fuel properties and operating conditions. Spray characterization focuses on spray penetration, cone angle, area, and velocity, with fuels varying by density, viscosity, cetane number, and caloric value. Injection pressure is fixed at 1800 bar, with analysis at chamber temperatures of 25°C and 100°C over a 600 µs injection duration. The dataset of 850 samples is split 4:1 for training and testing, with model performance assessed using coefficient of determination (R²) and mean squared error (MSE). The optimized XGBoost model achieves the highest performance, with R² and MSE values of 0.961 and 37.308, respectively. This study demonstrates ML’s effectiveness in analyzing biofuel spray variations, paving the way for injector designs that enhance engine efficiency, reduce emissions, and support environmental sustainability.