Evaluation of machine learning models for optimization of internal combustion engines: A comparative study

The high costs associated with Computational Fluid Dynamics (CFD) predictions limits the execution of some optimization processes of internal combustion engines. The use of machine learning algorithms instead of CFD during optimization of a spark ignition engine fueled with a biomass-derived syngas is proposed. Polynomial regression, support vector regression, Gaussian process regression, artificial neural networks, and random forest are the artificial intelligence methods considered. A general methodology for building, tuning, evaluating, and comparing machine learning models is presented. The fuel injection pressure data are utilized to estimate fuel consumption, equivalence ratio, nitrogen oxides emissions, and indicated mean effective pressure of the engine operating at 2500 and 4500 rpm. Results from previous CFD-based optimization studies are utilized to train, validate, and evaluate the models. All methods are K-fold cross-validated to determine their hyperparameters. Then, the models are evaluated by comparing their predictions accuracies for each output on a test data set. The results show that polynomial regression is the most accurate model to estimate fuel consumption and equivalence ratio with (R² ≥ 0.99999), while support vector regression demonstrates superior accuracy to estimate nitrogen oxides emissions and indicated mean effective pressure with (R² ≥ 0.99977). Overall, when averaging the accuracy of results across the four outputs, support vector regression emerges as the most accurate method (R² ≥ 0.99991), followed by Gaussian process regression (R² ≥ 0.99986). Subsequently, the optimization processes are executed using the selected learning models, leading to improved optimization outcomes with significantly reduced computational costs compared to those of the previous CFD-based optimization works.