Compression ignition engines when operated on gasoline fuels cause significant reduction in NOx and particulate emissions. In such advanced combustion strategy, the fuel-oxidiser mixing process, intensified by the prolonged ignition delays of gasoline fuels, directly affects the stability and efficiency of combustion. Thus, optimising fuel spray characteristics leads to optimisation of injector design, engine performance and subsequent decrease in emissions. Since spray development is a complex process that involves wide range of length and time scales, computationally expensive modelling techniques can be replaced with machine learning (ML) models. These ML models are employed to predict spray characteristics utilizing datasets generated from comprehensive spray studies. In this work, a dataset of about 5400 instances taken from non-evaporating gasoline fuel spray imaging experiments under Gasoline Compression Injection (GCI) engine conditions is used to train various ML models with data split of 70 % and 30 % for training and testing, respectively, with five-folds cross-validation performed within the training. The fuel injection pressure (60 – 150 MPa), chamber pressure (0.1 – 2 MPa), nozzle diameter, nozzle hole conicity and injection duration are used as input features to the models for predicting the spray tip penetration and spray angle. The performance of four ML models was evaluated and compared under default and tuned hyperparameters against experimental data and available physics-based correlations in the literature. The models include random forest, extreme gradient boosting, multilayer perceptron, and elastic-net. The results show that the hyperparameter-tuned extreme gradient boosting model performs best in predicting the spray parameters. The overall model performance was evaluated using the coefficient of determination (R2), mean absolute error (MAE), and root mean squared error (RMSE), resulting in values of 0.884, 0.651, and 1.571, respectively. This study presents compelling evidence demonstrating the effectiveness of ML as a powerful tool for isolating non-linear behaviors from physical processes. By effectively decoupling these behaviors, ML enhances the accuracy of predicting spray characteristics while significantly reducing computational costs. The application of ML in fuel injector design has the potential to revolutionize engine performance and contribute to substantial reductions in emissions.
Read full abstract