Machine learning-based design of target property-oriented fuels using explainable artificial intelligence

This study presents a machine learning (ML) and explainable artificial intelligence (XAI) integrated framework to guide the experimental investigation of engine combustion toward the most promising fuel compositions. Specifically, the experimental study evaluated nine diesel fuels with differing levels of paraffin, cycloparaffins, and aromatics in a heavy-duty engine, providing key data for subsequent model development. Building on this data, we introduce a multidimensional neural network methodology for feature optimization and performance assessment, incorporating 31 distinct features including engine variables, fuel properties, and components. The methodology innovatively integrates tree-based models with shapley additive explanations (SHAP) for detailed feature importance ranking. Features are sequentially added to subsets based on their importance for multilayer perceptron (MLP) model training, allowing for precise regression performance metrics for each subset. The comprehensive assessment of these best subsets revealed robust regression capabilities, with coefficient of determination (R2) values ranging from 0.9918 to 0.9999, and both root mean square error (RMSE) and mean absolute percentage error (MAPE) maintained below 0.4974 and 0.1388, respectively. Through SHAP and partial dependence plots (PDP), it was demonstrated that optimizing diesel fuel compositions by increasing paraffin levels, reducing aromatics, and moderately increasing cycloparaffins can significantly enhance combustion efficiency and reduce emissions.