Field inversion and machine learning based on the Rubber–Band Spalart–Allmaras Model

Machine learning (ML) techniques have emerged as powerful tools for improving the predictive capabilities of Reynolds-averaged Navier–Stokes (RANS) turbulence models in separated flows. This improvement is achieved by leveraging complex ML models, such as those developed using field inversion and ma...

Full description

Saved in:
Bibliographic Details
Main Authors: Chenyu Wu, Yufei Zhang
Format: Article
Language:English
Published: Elsevier 2025-03-01
Series:Theoretical and Applied Mechanics Letters
Subjects:
Online Access:http://www.sciencedirect.com/science/article/pii/S2095034924000758
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1849388543917424640
author Chenyu Wu
Yufei Zhang
author_facet Chenyu Wu
Yufei Zhang
author_sort Chenyu Wu
collection DOAJ
description Machine learning (ML) techniques have emerged as powerful tools for improving the predictive capabilities of Reynolds-averaged Navier–Stokes (RANS) turbulence models in separated flows. This improvement is achieved by leveraging complex ML models, such as those developed using field inversion and machine learning (FIML), to dynamically adjust the constants within the baseline RANS model. However, the ML models often overlook the fundamental calibrations of the RANS turbulence model. Consequently, the basic calibration of the baseline RANS model is disrupted, leading to a degradation in the accuracy, particularly in basic wall-attached flows outside of the training set. To address this issue, a modified version of the Spalart-Allmaras (SA) turbulence model, known as Rubber-band SA (RBSA), has been proposed recently. This modification involves identifying and embedding constraints related to basic wall-attached flows directly into the model. It is shown that no matter how the parameters of the RBSA model are adjusted as constants throughout the flow field, its accuracy in wall-attached flows remains unaffected. In this paper, we propose a new constraint for the RBSA model, which better safeguards the law of wall in extreme conditions where the model parameter is adjusted dramatically. The resultant model is called the RBSA-poly model. We then show that when combined with FIML augmentation, the RBSA-poly model effectively preserves the accuracy of simple wall-attached flows, even when the adjusted parameters become functions of local flow variables rather than constants. A comparative analysis with the FIML-augmented original SA model reveals that the augmented RBSA-poly model reduces error in basic wall-attached flows by 50% while maintaining comparable accuracy in trained separated flows. These findings confirm the effectiveness of utilizing FIML in conjunction with the RBSA model, offering superior accuracy retention in cardinal flows.
format Article
id doaj-art-e64ba91562794b30959f847ed1035e97
institution Kabale University
issn 2095-0349
language English
publishDate 2025-03-01
publisher Elsevier
record_format Article
series Theoretical and Applied Mechanics Letters
spelling doaj-art-e64ba91562794b30959f847ed1035e972025-08-20T03:42:15ZengElsevierTheoretical and Applied Mechanics Letters2095-03492025-03-0115210056410.1016/j.taml.2024.100564Field inversion and machine learning based on the Rubber–Band Spalart–Allmaras ModelChenyu Wu0Yufei Zhang1School of Aerospace Engineering, Tsinghua University, Beijing 100084, ChinaCorresponding author.; School of Aerospace Engineering, Tsinghua University, Beijing 100084, ChinaMachine learning (ML) techniques have emerged as powerful tools for improving the predictive capabilities of Reynolds-averaged Navier–Stokes (RANS) turbulence models in separated flows. This improvement is achieved by leveraging complex ML models, such as those developed using field inversion and machine learning (FIML), to dynamically adjust the constants within the baseline RANS model. However, the ML models often overlook the fundamental calibrations of the RANS turbulence model. Consequently, the basic calibration of the baseline RANS model is disrupted, leading to a degradation in the accuracy, particularly in basic wall-attached flows outside of the training set. To address this issue, a modified version of the Spalart-Allmaras (SA) turbulence model, known as Rubber-band SA (RBSA), has been proposed recently. This modification involves identifying and embedding constraints related to basic wall-attached flows directly into the model. It is shown that no matter how the parameters of the RBSA model are adjusted as constants throughout the flow field, its accuracy in wall-attached flows remains unaffected. In this paper, we propose a new constraint for the RBSA model, which better safeguards the law of wall in extreme conditions where the model parameter is adjusted dramatically. The resultant model is called the RBSA-poly model. We then show that when combined with FIML augmentation, the RBSA-poly model effectively preserves the accuracy of simple wall-attached flows, even when the adjusted parameters become functions of local flow variables rather than constants. A comparative analysis with the FIML-augmented original SA model reveals that the augmented RBSA-poly model reduces error in basic wall-attached flows by 50% while maintaining comparable accuracy in trained separated flows. These findings confirm the effectiveness of utilizing FIML in conjunction with the RBSA model, offering superior accuracy retention in cardinal flows.http://www.sciencedirect.com/science/article/pii/S2095034924000758Turbulence modelingField inversionConstrained-recalibrationMachine learning
spellingShingle Chenyu Wu
Yufei Zhang
Field inversion and machine learning based on the Rubber–Band Spalart–Allmaras Model
Theoretical and Applied Mechanics Letters
Turbulence modeling
Field inversion
Constrained-recalibration
Machine learning
title Field inversion and machine learning based on the Rubber–Band Spalart–Allmaras Model
title_full Field inversion and machine learning based on the Rubber–Band Spalart–Allmaras Model
title_fullStr Field inversion and machine learning based on the Rubber–Band Spalart–Allmaras Model
title_full_unstemmed Field inversion and machine learning based on the Rubber–Band Spalart–Allmaras Model
title_short Field inversion and machine learning based on the Rubber–Band Spalart–Allmaras Model
title_sort field inversion and machine learning based on the rubber band spalart allmaras model
topic Turbulence modeling
Field inversion
Constrained-recalibration
Machine learning
url http://www.sciencedirect.com/science/article/pii/S2095034924000758
work_keys_str_mv AT chenyuwu fieldinversionandmachinelearningbasedontherubberbandspalartallmarasmodel
AT yufeizhang fieldinversionandmachinelearningbasedontherubberbandspalartallmarasmodel