Physics-enhanced machine learning methods for industrial process modeling and optimization

Recent advancements in digital technologies and Artificial Intelligence (AI) are increasingly impacting the manufacturing sector, enabling the development of "Intelligent Machines" that are more autonomous, flexible, and efficient. These machines support process monitoring, parameter optimization, and human-machine collaboration, aligning with Industry 4.0 goals and the human-centric vision of Industry 5.0. This thesis explores modeling and optimization strategies that combine physics-based models and expert knowledge with data-driven machine learning methods, following the physics-informed AI paradigm. Specifically, it investigates Random Forests as surrogate models suitable for small datasets, proposing an original low-complexity method for global optimization based on their tree structure. A novel physics-informed semi-supervised classification approach is also introduced to improve quality prediction by integrating prior knowledge from physics-based models. The proposed methods are applied to a centerless grinding process using synthetic data generated by high- and low-fidelity models. Results demonstrate the effectiveness of the proposed optimization and classification techniques in high-dimensional settings. However, the overall performance of hybrid models is limited by the small dataset size and the limited accuracy of the physics-based predictions.