Techniques avancées pilotées par l’IA pour le moulage des procédés et l’optimisation structurelle

As scientific progress accelerates, the demand for models with higher accuracy, faster performance, and more reliable results continues to grow. This work develops and applies deep learning methods to the analysis and optimization of mechanical systems, focusing on both processing molding, and structural design, with tailored architectures for the problems addressed in this thesis.The first part of the thesis focuses on the Compression Molding (CM) process. While there are various types of compression molding methods, our emphasis is specifically on the Sheet Molding Compounds (SMC) process. A numerical model based on Computational Fluid Dynamics (CFD) is used to describe squeezing flow. A feedforward neural network (FFNN) is trained using boundary evaluation data by applying Model Order Reduction (MOR). Then, an inverse analysis is performed with the neural network to determine the optimal precharge geometry. Building on this, a surrogate model is introduced that can provide instant predictions of the best precharge shape for a given mold and compression time.The second part of this thesis focuses on topology optimization in Computational Structural Mechanics (CSM), where the objective is to minimize compliance. A hybrid strategy is introduced that combines data-driven methods, utilizing Convolutional Neural Network (CNN) encoder-decoder structures, with classical analytical formulations. This approach improves accuracy and provides real-time solutions to topology optimization problems, even under randomly varying boundary conditions and applied loads.As a result, this study shows that combining deep learning with established computational mechanics can significantly accelerate the analysis and optimization of mechanical systems by enabling real-time predictions. At the same time, numerical models and computational mechanics remain essential to provide the physical grounding and accuracy needed to validate and guide data-driven approaches.