Constrained re-calibration of two-equation Reynolds-averaged Navier–Stokes models

• Published in: Theoretical and applied mechanics letters Vol. 14; no. 2; p. 100503
• Main Authors: Bin, Yuanwei, Hu, Xiaohan, Li, Jiaqi, Grauer, Samuel J., Yang, Xiang I.A.
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.03.2024; Elsevier
• Subjects: Machine learning; Reynolds-averaged Navier–Stokes; Turbulence modeling; Turbulence modeling; Reynolds-averaged Navier–Stokes; Machine learning
• ISSN: 2095-0349
• DOI: 10.1016/j.taml.2024.100503

•Machine learned turbulence models do not generalize outside the training dataset because they do not preserve the calibrations of the baseline model.•This paper identifies constraints that preserve the calibrations of the baseline model.•Accounting for these constraints in model training leads to augmentations that generalize outside the training dataset. [Display omitted] Machine-learned augmentations to turbulence models can be advantageous for flows within the training dataset but can often cause harm outside. This lack of generalizability arises because the constants (as well as the functions) in a Reynolds-averaged Navier–Stokes (RANS) model are coupled, and un-constrained re-calibration of these constants (and functions) can disrupt the calibrations of the baseline model, the preservation of which is critical to the model’s generalizability. To safeguard the behaviors of the baseline model beyond the training dataset, machine learning must be constrained such that basic calibrations like the law of the wall are kept intact. This letter aims to identify such constraints in two-equation RANS models so that future machine learning work can be performed without violating these constraints. We demonstrate that the identified constraints are not limiting. Furthermore, they help preserve the generalizability of the baseline model.