Kalia, Nikhila, Mcconkey, Ryley, Yee, Eugene et Lien, Fue-Sang.
2025.
« Bayesian optimization of the GEKO turbulence model for predicting flow separation over a smooth surface ».
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
293 - Bayesian optimization of the GEKO .pdf - Version publiée Licence d'utilisation : Tous les droits réservés aux détenteurs du droit d'auteur. Télécharger (2MB) | Prévisualisation |
Résumé
This paper applies the tuning parameters using Bayesian-optimization-RANS (turbo-RANS) methodology to enhance the predictive accuracy of Reynolds-averaged Navier-Stokes (RANS) turbulence models for flow over a converging-diverging channel, a benchmark case characterized by adverse pressure gradients and flow separation. Using Bayesian optimization, the Generalized k-? (GEKO) turbulence model was calibrated by tuning key coefficients (CSEP and CNW) with sparse reference data from direct numerical simulation (DNS) studies at Re = 12,600. The calibration process was guided by the Generalized Error Distribution-based Calibration Procedure (GEDCP), which optimized the coefficients based on pressure recovery (Cp) and skin friction (C??) data.The optimized model was then evaluated on its predictive capability beyond the calibration dataset. Specifically, the streamwise velocity (U) predictions at Re = 12,600 were compared against DNS results to assess whether improvements in pressure and friction coefficients translated to enhanced velocity field accuracy. Furthermore, to test the robustness of the optimized model across different Reynolds numbers, additional comparisons were performed against large-eddy simulation (LES) data at Re = 20,580. In this higher Reynolds number case, both the streamwise velocity and skin friction coefficient were analyzed to determine the generalizability of the optimized turbulence model.Results demonstrate that the optimized GEKO (turbo-RANS) model significantly improves the prediction of wall quantities, particularly in capturing flow reattachment. The improvements in streamwise velocity profiles at both Reynolds numbers suggest that Bayesian-optimized coefficients provide a more reliable representation of adverse pressure gradient effects. Additionally, the ability of the optimized model to maintain accuracy across different Reynolds numbers highlights the potential of turbo-RANS in developing turbulence model corrections that generalize across similar flow regimes.While notable improvements were observed in velocity field predictions, the optimized model showed only marginal enhancement in skin friction predictions, which can be attributed to the fundamental limitations of two-equation turbulence models in resolving near-wall stress distributions. Nevertheless, this study provides further insight into the role of machine learning-assisted RANS calibration in improving predictive accuracy for complex flows. The findings suggest that optimized coefficients obtained from a single training dataset can be effectively applied across moderate variations in Reynolds number, enhancing the practical applicability of turbo-RANS for the re-calibration of RANS turbulence closure model coefficients for improved predictive performance in industrial and scientific turbulence modeling applications.
| 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 15:15 |
| Dernière modification: | 18 déc. 2025 15:15 |
| URI: | https://espace2.etsmtl.ca/id/eprint/32432 |
Actions (Authentification requise)
 |
Dernière vérification avant le dépôt |