Machine Learning-Assisted performance prediction and analysis of Ammonia-Diesel Dual-Fuel engines
Year: 2026
Type: article
Abstract: The brake thermal efficiency (BTE) and the NO X , NH 3 and N 2 O emissions of a 12.5 L marine low-pressure ammonia-diesel dual-fuel engine were predicted and analyzed using machine learning (ML). The dataset comprises 462 operating conditions accumulated during engine development with four hardware configurations, and it is unevenly distributed over the eleven input variables. Seven regression algorithms were compared by nested cross-validation with two data-splitting methods. With random splitting, the six nonlinear models did not differ significantly in mean R 2 (0.77 to 0.84). When similar operating conditions were excluded from the training folds, the tree ensembles retained the highest accuracy (mean R 2 of 0.67 to 0.71), whereas the neural network and Gaussian process models deteriorated considerably. None of the models could predict a hardware configuration that was absent from the training data. For the selected XGBoost model, the prediction error increased with the distance from the training data and with the disagreement among the seven models (Spearman ρ = 0.56). The distance did not explain the large errors at low load without the main diesel injection, where the combustion was less stable·NH 3 was the least accurately predicted output because of a few conditions with very high values. SHAP analysis showed that the air excess ratio (λ) was the most important input for NH 3 and N 2 O, and its opposite contributions to NO X and NH 3 agreed with the measured trends. In contrast, the contributions of the micro-pilot injection variables and the intake and exhaust pressures largely reflected the hardware configuration or the engine load. The validation method thus had a greater influence on model selection than the type of nonlinear algorithm, and the models are applicable only within the tested hardware configurations.
Source: Fuel
Language: English
Subfield: Fluid Flow and Transfer Processes
Field: Chemical Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $4,330
OpenAlex ID: W7203768981
