Percorrer por autor "Vieira, Beatriz Susana"
A mostrar 1 - 1 de 1
Resultados por página
Opções de ordenação
- Simulating stresses and strains in solid mechanics directly from images using convolutional neural networksPublication . Vieira, Beatriz Susana; Rodrigues, José Alberto de Sousa; Bordas, Stéphane P. A.Abstract Finite Element Method (FEM) simulations provide reliable displacement, strain and stress fields, but they become costly when many geometry–load combinations must be tested. In several practical scenarios the input is an image that leads to non-rectangular, unstructured meshes, which is not ideal for strictly grid-based models. This dissertation investigates deep learning surrogates trained on FEM solutions that can deliver the mechanical response much faster. We propose a complete processing pipeline that starts from image segmentation, builds both structured and unstructured meshes, and trains two distinct models on top of them. The first one is a grid U-Net, designed for rectangular domains. The second one is MAgNET, which operates directly on the mesh and preserves the original discretization at the boundaries and at the loaded regions. Both models are trained and evaluated with exactly the same dataset, training schedule and metrics, including tests with loads above the training range and measurements of training and inference time, which enables a fair comparison between the two approaches. Results show that both surrogates reproduce displacement accurately, and that the largest strain and stress errors remain confined to the loaded boundary and to high-gradient areas. The grid U-Net is faster and very competitive on regular meshes, while MAgNET is the better option when the geometry comes from images and the mesh is unstructured.