Physics-Informed Machine Learning model for Emission Prediction in Exhaust Aftertreatment Systems: Prediction of NOX, CO, HC and NH3 Emissions for Volvo Penta Off-Road Diesel Engines

Sammanfattning

Modern industrial diesel engines rely on advanced exhaust aftertreatment systems(EATS) to comply with increasingly stringent emission regulations such as EU StageV, reducing emissions of nitrogen oxides (NOx), hydrocarbons (HC), carbon monoxide (CO), and ammonia slip (NH3). Among these systems, selective catalytic reduction (SCR) plays a central role in NOx reduction through urea-based ammoniadosing. However, SCR development, calibration, and catalyst sizing still dependheavily on expensive and time-consuming physical engine and rig testing. This creates a strong industrial need for reliable predictive models that can complementphysical testing through early-stage virtual evaluation.This thesis investigates the use of physics-informed feature engineering combinedwith machine learning for predicting system-out emissions in Volvo Penta off-roaddiesel engine platforms. The objective is to develop a supervised regression framework capable of predicting system-out emissions based on engine-out conditions andexhaust aftertreatment parameters, while improving robustness and physical interpretability compared to purely data-driven approaches. Particular focus is placedon NOx and NH3 prediction, since these emissions are most strongly linked to SCRbehaviour and catalyst dynamics.The study uses existing experimental test data from Volvo Penta engines in the D5D13 platform range, representing industrial and off-road applications. Separate XGBoost regression models were developed for each target variable using both directlymeasured signals and derived physics-informed features related to SCR behaviour,catalyst thermal conditions, flow dynamics, and ammonia availability. Model performance was evaluated using ShuffleSplit cross-validation together with validationon unseen datasets, primarily using the coefficient of determination (R2) and MeanAbsolute Error (MAE).The results show that physics-informed feature engineering improved both model accuracy and robustness compared to baseline models using only directly measured inputs. The strongest improvements were observed for NOx prediction, where physicsinformed features improved model accuracy and robustness across different operatingconditions and engine platforms. The final models also achieved strong predictiveperformance for HC and CO, while NH3 prediction remained more challenging dueto the complexity of ammonia storage and slip behavior. Feature importance analysis further confirmed that the learned relationships were consistent with known SCRphysics, improving confidence in model interpretability and engineering relevance.The developed framework demonstrates that physics-informed feature engineeringcombined with machine learning can provide a practical and computationally efficient approach for EATS emission prediction. The results also show clear potentialfor supporting early-stage virtual evaluation, SCR catalyst sizing, and aftertreatment system optimisation within industrial engine development.