Rentschler, Tobias, Tismer, Alexander et Riedelbauch, Stefan.
2025.
« Frame field prediction for quadrilateral domain partition ».
In Proceedings of the CSME-CFDSC-CSR 2025 International Congress (Montreal, QC, Canada, May 25-28, 2025)
Coll. « Progress in Canadian Mechanical Engineering », vol. 8.
Prévisualisation |
PDF
363 - Frame field prediction for quadril.pdf - Version publiée Licence d'utilisation : Tous les droits réservés aux détenteurs du droit d'auteur. Télécharger (5MB) | Prévisualisation |
Résumé
The generation of high-quality quadrilateral meshes is an essential requirement for a variety of applications, including Computational Fluid Dynamics (CFD). Conventional approaches generally compromise between speed and quality, with the computational demands increasing in response to higher quality requirements. Cross-field-based mesh generation has the potential to produce high-quality block-structured meshes. However, this approach relies on the solution of Partial Differential Equations (PDEs), which can be computationally intensive, particularly for large or high-resolution domains. This study presents a Graph Neural Network (GNN) capable of learning and approximating these PDEs, allowing the efficient partitioning of geometric domains into quadrilateral subregions. The primary benefit of this method is its capacity to partition geometric domains into quadrilateral subregions with good scalability, while eliminating the need to solve expensive PDEs. The GNN predicts a boundary-aligned frame field that serves as the foundation for the domain partitioning. Boundary conditions are defined based on normal and tangent vectors at the domain boundaries. They are propagated smoothly across the interior using the GNN, eliminating the need for computationally expensive PDEs. In a subsequent post-processing step, singularities within the frame field are identified, and streamlines originating from these singularities are computed. These streamlines are strategically aligned with the geometry of the domain, leading to the decomposition of the domain into quadrilateral regions. The resulting regions are well-suited for fast algebraic meshing techniques, such as bilinear transfinite interpolation, ensuring the generation of high-quality meshes. This study utilizes a simple 2- dimensional NACA profile as a test case to train the GNN.
| Type de document: | Compte rendu de conférence |
|---|---|
| Éditeurs: | Éditeurs ORCID Hof, Lucas A. NON SPÉCIFIÉ Di Labbio, Giuseppe NON SPÉCIFIÉ Tahan, Antoine NON SPÉCIFIÉ Sanjosé, Marlène NON SPÉCIFIÉ Lalonde, Sébastien NON SPÉCIFIÉ Demarquette, Nicole R. NON SPÉCIFIÉ |
| Date de dépôt: | 18 déc. 2025 15:21 |
| Dernière modification: | 18 déc. 2025 15:21 |
| URI: | https://espace2.etsmtl.ca/id/eprint/32487 |
Actions (Authentification requise)
 |
Dernière vérification avant le dépôt |