Advancement, performance assessment and application of machine learning-based models for internal combustion engines

This PhD thesis presents research conducted in collaboration with the University of Bologna, Ferrari S.p.A., and AVL List GmbH, focusing on advancing Artificial Intelligence (AI)-based modelling methodologies for high-performance Internal Combustion Engine Vehicles. The primary objectives are to enhance AI-driven models for engine emissions and combustion metrics, apply them to real industrial scenarios, and benchmark their performance against commercial semi-physical software. The research explores data-driven approaches for two main tasks: modelling engine-out and tailpipe gaseous emissions, and predicting combustion metrics and exhaust gas temperatures. For emissions modelling, the study refines machine learning solutions—specifically, a Light Gradient Boosting Regressor (LGBR) that incorporates the time history of input features. Key advancements include a semi-automatic pipeline for managing large training datasets in engine-out emissions, improving accuracy and expanding the range of operating conditions. For tailpipe emissions, the integration of a third λ sensor and a custom experimental campaign enhances model sensitivity to non-standard conditions, boosting accuracy and generalization. An innovative Principal Component Analysis-based epistemic uncertainty evaluation algorithm is also developed to assess model reliability in the absence of experimental data. For combustion metrics and exhaust temperature, a physics-enhanced data-driven methodology is introduced, improving extrapolation and physical consistency. Validation under steady-state and transient conditions confirms the models’ accuracy and industrial applicability. The research demonstrates that combining data-driven and analytical approaches is essential for robust generalization, especially under non-standard conditions, and must be supported by targeted experimental campaigns. Finally, a comprehensive benchmark compares the developed AI-based engine model with AVL CRUISE™ M, evaluating accuracy, calibration time, and computational efficiency. Results highlight the readiness of AI-driven solutions for industrial integration. Overall, this work demonstrates the potential of AI-based modelling to accelerate engine development, reduce costs, and support the automotive industry’s transition towards cleaner and more efficient propulsion systems.