Kalia, Nikhila, Mcconkey, Ryley, Yee, Eugene et Lien, Fue-Sang. 2025. « Kolmogorov-Arnold networks for turbulence anisotropy mapping ». Communication lors de la conférence : CSME-CFDSC-CSR 2025 International Congress (Montreal, QC, Canada, May 25-28, 2025).
Prévisualisation |
PDF
426 - Kolmogorov-Arnold networks for tur.pdf - Version publiée Licence d'utilisation : Tous les droits réservés aux détenteurs du droit d'auteur. Télécharger (50kB) | Prévisualisation |
Résumé
Reynolds-averaged Navier-Stokes (RANS) models remain a widely used approach for turbulence modeling in computational fluid dynamics (CFD). However, conventional turbulence closures often struggle with accurately predicting anisotropic turbulence effects in complex flow configurations. Kolmogorov-Arnold Networks (KANs) have recently been explored as an alternative to Multi-Layer Perceptrons (MLPs) for machine learning-driven turbulence closures, demonstrating their effectiveness in predicting anisotropy tensor components in flat plate boundary layers.Building on this prior work, we extend the KAN-based anisotropy mapping methodology to two additional benchmark flows: turbulent flow in a square duct and flow over periodic hills. These cases pose significant challenges for RANS models, particularly due to secondary flows induced by turbulence anisotropy and flow separation in the presence of strong pressure gradients. KANs leverage spline-based function approximations, allowing for more compact architectures than conventional deep learning models while maintaining key physics-informed constraints. However, the optimization of spline parameters introduces additional computational cost, leading to longer training times compared to MLP-based architectures.We train KAN-based models on high-fidelity direct numerical simulation (DNS) datasets and evaluate their performance against existing machine learning-based closures. The results indicate that KANs provide stable and realizability-preserving predictions while effectively capturing secondary flows and anisotropic turbulence effects. Furthermore, we assess the trade-offs between accuracy, network size, and computational efficiency, positioning KANs as a viable alternative to MLP-based turbulence models.This study further expands the scope of KANs in turbulence modeling, demonstrating their applicability across multiple canonical flows and reinforcing their potential as a physics-informed machine learning approach for RANS turbulence closure modeling.
| Type de document: | Communication (Communication) |
|---|---|
| Informations complémentaires: | Progress in Canadian Mechanical Engineering, Volume 8. Co-chairs: Lucas A. Hof, Giuseppe Di Labbio, Antoine Tahan, Marlène Sanjosé, Sébastien Lalonde and Nicole R. Demarquette. |
| Date de dépôt: | 18 déc. 2025 14:38 |
| Dernière modification: | 18 déc. 2025 14:38 |
| URI: | https://espace2.etsmtl.ca/id/eprint/32150 |
Actions (Authentification requise)
 |
Dernière vérification avant le dépôt |