Advanced engine modeling and simulation approaches using artificial intelligence, hybrid physics-informed method, and the integration of data-driven models in 1d-0d environments for the development of highperformance powertrains

The urgent necessity to mitigate CO_2 emissions in the transportation sector has driven the development of advanced modeling for internal combustion engines and their integration with sustainable propulsion systems. This thesis provides a detailed study on the application of Artificial Neural Networks, hybrid physics-informed approaches, and 1D–0D model integration to accurately predict key combustion, knock, and exhaust gas indices. By focusing on improving model accuracy and generalization, the research aims to reduce the experimental effort typically required for model calibration.Initially, pure Neural Network-based models were developed to estimate combustion and knock intensity. While these models achieved high accuracy under steady-state and transient conditions, they struggled with extrapolation due to the inherent bias–variance trade-off of data-driven methods. To address this, a novel hybrid methodology was proposed, combining neural networks with analytical corrective functions derived from physical principles. These hybrid models successfully captured the impact of parameters like spark advance and variable valve timing while ensuring consistency with physical trends, ultimately reducing the reliance on large datasets and saving time, cost, and fuel.A direct comparison revealed that while pure ANNs are slightly more accurate within their training domain, the hybrid approach offers superior robustness and reliability under diverse conditions. The thesis further explores integrating these 0D models with 1D GT-Suite simulations in a co-simulation environment, specifically for knock prediction. This framework ensures consistency across modeling domains and demonstrates adaptability to alternative fuels, such as hydrogen. In conclusion, these hybrid and co-simulation strategies offer a promising balance of accuracy and physical consistency, facilitating the design of reliable digital twins for faster, more sustainable engine development.