From 1D Physical Engine Model To Neural Network-Based One For Real-Time Simulation

The aim of this thesis is the development of a mathematical and numerical model of an automotive diesel engine, based on the use of artificial neural networks as an alternative to traditional physics-based models implemented in simulation software such as GT-Power. The proposed model, which operates as a black-box representation of complex engine phenomena, was trained using data generated by GT-Power simulations, considered in this context as the physical reference (ground truth) for validating engine behavior. This approach fits into a broader methodological innovation framework aimed at optimizing and improving the efficiency of existing engine technologies, in line with a realistic, gradual, and sustainable energy transition. In particular, the integration of machine learning techniques enables a significant reduction in the time and cost associated with simulation and calibration phases, while maintaining high predictive reliability. The results obtained demonstrate the robustness and consistency of the neural model with respect to the behavior of the physics-based engine simulations. Moreover, the main advantage of this approach lies in its ability to rapidly simulate multiple engine operating points, achieving substantial time savings compared to conventional 1D simulations, while still preserving accurate representation of engine performance. This opens promising perspectives for the adoption of data-driven methodologies in the field of advanced engine modeling.