Toward Discretization-Consistent Closure Schemes for Large Eddy Simulation Using Reinforcement Learning DOI Creative Commons
Andrea Beck, Marius Kurz

arXiv (Cornell University), Journal Year: 2023, Volume and Issue: unknown

Published: Jan. 1, 2023

This study proposes a novel method for developing discretization-consistent closure schemes implicitly filtered Large Eddy Simulation (LES). Here, the induced filter kernel, and thus terms, are determined by properties of grid discretization operator, leading to additional computational subgrid terms that generally unknown in priori analysis. In this work, task adapting coefficients LES models is framed as Markov decision process solved an posteriori manner with Reinforcement Learning (RL). optimization framework applied both explicit implicit models. The model based on element-local eddy viscosity model. optimized found adapt its within discontinuous Galerkin (DG) methods homogenize dissipation element adding more near center. For modeling, RL identify optimal blending strategy hybrid DG Finite Volume (FV) scheme. resulting yields accurate results than either pure or FV renders itself viable modeling ansatz could initiate class high-order compressible turbulence combining shock capturing single framework. All newly derived achieve match outperform traditional different discretizations resolutions. Overall, demonstrate proposed can provide closures reduce uncertainty LES.

Language: Английский

Reinforcement learning for wind-farm flow control: Current state and future actions DOI Creative Commons
Mahdi Abkar, Navid Zehtabiyan-Rezaie, Alexandros Iosifidis

et al.

Theoretical and Applied Mechanics Letters, Journal Year: 2023, Volume and Issue: 13(6), P. 100475 - 100475

Published: Oct. 20, 2023

Wind-farm flow control stands at the forefront of grand challenges in wind-energy science. The central issue is that current algorithms are based on simplified models and, thus, fall short capturing complex physics wind farms associated with high-dimensional nature turbulence and multiscale wind-farm-atmosphere interactions. Reinforcement learning (RL), as a subset machine learning, has demonstrated its effectiveness solving problems various domains, studies performed last decade prove it can be exploited development next generation for wind-farm control. This review two main objectives. Firstly, aims to provide an up-to-date overview works focusing schemes utilizing RL methods. By examining latest research this area, seeks offer comprehensive understanding advancements made through application techniques. Secondly, shed light obstacles researchers face when implementing RL. highlighting these challenges, identify areas requiring further exploration potential opportunities future research.

Language: Английский

Citations

18
Toward discretization-consistent closure schemes for large eddy simulation using reinforcement learning DOI
Andrea Beck, Marius Kurz

Physics of Fluids, Journal Year: 2023, Volume and Issue: 35(12)

Published: Dec. 1, 2023

This study proposes a novel method for developing discretization-consistent closure schemes implicitly filtered large eddy simulation (LES). Here, the induced filter kernel and, thus, terms are determined by properties of grid and discretization operator, leading to additional computational subgrid that generally unknown in priori analysis. In this work, task adapting coefficients LES models is thus framed as Markov decision process solved an posteriori manner with reinforcement learning (RL). optimization framework applied both explicit implicit models. The model based on element-local viscosity model. optimized found adapt its within discontinuous Galerkin (DG) methods homogenize dissipation element adding more near center. For modeling, RL identify optimal blending strategy hybrid DG finite volume (FV) scheme. resulting yields accurate results than either pure or FV renders itself viable modeling ansatz could initiate class high-order compressible turbulence combining shock capturing single framework. All newly derived achieve match outperform traditional different discretizations resolutions. Overall, demonstrate proposed can provide closures reduce uncertainty LES.

Language: Английский

Citations

12
Prediction and control of two-dimensional decaying turbulence using generative adversarial networks DOI Creative Commons
Jiyeon Kim, Junhyuk Kim, Changhoon Lee

et al.

Journal of Fluid Mechanics, Journal Year: 2024, Volume and Issue: 981

Published: Feb. 21, 2024

An accurate prediction of turbulence has been very costly since it requires an infinitesimally small time step for advancing the governing equations to resolve fast-evolving small-scale motions. With recent development various machine learning (ML) algorithms, finite-time became one promising options relieve computational burden. Yet, a reliable motions is challenging. In this study, PredictionNet, data-driven ML framework based on generative adversarial networks (GANs), was developed fast with high accuracy down smallest scale using relatively number parameters. particular, we conducted two-dimensional (2-D) decaying at finite lead times direct numerical simulation data. The model accurately predicted turbulent fields up half Eulerian integral over which large-scale remain fairly correlated. Scale decomposition used interpret predictability depending spatial scale, and role latent variables in discriminator network investigated. good performance GAN predicting attributed scale-selection scale-interaction capability variable. Furthermore, by utilising PredictionNet as surrogate model, control named ControlNet identify disturbance that drive evolution flow field direction optimises specified objective function.

Language: Английский

Citations

5
Ensemble data assimilation-based mixed subgrid-scale model for large-eddy simulations DOI Open Access
Yunpeng Wang, Zelong Yuan, Jianchun Wang

et al.

Physics of Fluids, Journal Year: 2023, Volume and Issue: 35(8)

Published: Aug. 1, 2023

An ensemble Kalman filter (EnKF)-based mixed model (EnKF-MM) is proposed for the subgrid-scale (SGS) closure in large-eddy simulation (LES) of turbulence. The coefficients are determined through EnKF-based data assimilation technique. direct numerical (DNS) results filtered to obtain benchmark LES. Reconstructing correct kinetic energy spectrum DNS (fDNS) has been adopted as target EnKF optimize coefficient functional part model. EnKF-MM framework subsequently tested LES both incompressible homogeneous isotropic turbulence (HIT) and turbulent mixing layer (TML). performance comprehensively examined predictions flow statistics including velocity spectrum, probability density functions (PDFs) SGS stress, PDF strain rate flux. structure functions, evolution energy, mean Reynolds stress profile, iso-surface Q-criterion also evaluate spatial-temporal by different models. consistently more satisfying compared traditional models, dynamic Smagorinsky (DSM), (DMM) gradient (VGM), demonstrating its great potential optimization models

Language: Английский

Citations

11
Constrained Recalibration of Reynolds-Averaged Navier–Stokes Models DOI
Yuanwei Bin,

George Huang,

Robert F. Kunz

et al.

AIAA Journal, Journal Year: 2023, Volume and Issue: 62(4), P. 1434 - 1446

Published: Dec. 20, 2023

The constants and functions in Reynolds-averaged Navier–Stokes (RANS) turbulence models are coupled. Consequently, modifications of a RANS model often negatively impact its basic calibrations, which is why machine-learned augmentations detrimental outside the training dataset. A solution to this identify degrees freedom that do not affect calibrations only modify these identified when recalibrating baseline accommodate specific application. This approach colloquially known as “rubber-band” approach, we formally call “constrained recalibration” paper. To illustrate efficacy Spalart–Allmaras log law calibration. By subsequently interfacing data-based methods with freedom, train solve historically challenging flow scenarios, including round-jet/plane-jet anomaly, airfoil stall, secondary separation, recovery after separation. In addition good performance inside dataset, trained yield similar

Language: Английский

Citations

10
An intercomparison of wall fluxes in a turbulent thermal convection chamber: Direct numerical simulations and wall-modeled large-eddy simulations enhanced by machine learning DOI Creative Commons
Aaron Wang, Silvio Schmalfuß, Kamal Kant Chandrakar

et al.

Physics of Fluids, Journal Year: 2025, Volume and Issue: 37(4)

Published: April 1, 2025

Thermal convection in a closed chamber is driven by warm bottom, cold top, and side walls at various temperatures. Although wall fluxes are the source of energy, accurately modeling these (i.e., model) challenging. In large-eddy simulations (LESs), many models traditionally derived from canonical boundary layer, which may be unsuitable for thermal bounded both horizontal vertical walls. This study conducts model intercomparison dry cubic-meter using three direct numerical (DNSs) four LESs with different models. The employ traditional models, new employing physics-aware neural networks, refined grid near experiment involves cases varying sidewall Our results show that capture main flow features trends mean fluxes. networks grids can improve temporally averaged local when large-scale circulation has preferred direction. Even without improvement fluxes, LES quantities (temperature velocities) still largely match those DNSs, provided flux matches DNSs. Additionally, DNSs reveal variation corner treatments minimal impacts on away corners. Finally, underestimate entire due to their inability resolve regions, but better DNS.

Language: Английский

Citations

0
Knowledge-integrated additive learning for consistent near-wall modelling of turbulent flows DOI
Fengshun Zhang, Zhideng Zhou, Xiaolei Yang

et al.

Journal of Fluid Mechanics, Journal Year: 2025, Volume and Issue: 1011

Published: May 13, 2025

Developing a consistent near-wall turbulence model remains an unsolved problem. The machine learning method has the potential to become workhorse for modelling. However, learned suffers from limited generalisability, especially flows without similarity laws (e.g. separated flows). In this work, we propose knowledge-integrated additive (KIA) approach wall models in large-eddy simulations. proposed integrates knowledge simplified thin-boundary-layer equation with data-driven forcing term non-equilibrium effects induced by pressure gradients and flow separations. capability each dataset is encapsulated using basis functions corresponding weights approximated neural networks. fusion of capabilities various datasets enabled distance function, way that preserved generalisability other cases allowed. demonstrated via training sequentially data gradient but no separation, data. preserve previously tested turbulent channel cases. periodic hill 2-D Gaussian bump showcase different surface curvatures Reynolds numbers. Good agreements references are obtained all test

Language: Английский

Citations

0
A priori assessment of nonlocal data-driven wall modeling in large eddy simulation DOI Open Access
Golsa Tabe Jamaat

Physics of Fluids, Journal Year: 2023, Volume and Issue: 35(5)

Published: May 1, 2023

In the present study, a priori assessment is performed on ability of convolutional neural network (CNN) for wall-modeling in large eddy simulation. The data used training process are provided by direct numerical simulation (DNS) turbulent channel flow. Initially, study carried out input choices CNN, and effect different flow parameters establishing wall model investigated. Then, influence wall-normal distance established data-driven studied choosing CNN from two regions inner layer (y+>10,y/δ<0.1) logarithmic layer. performance obtained models based inputs further investigated feeding with outside range. next step, tested under various conditions, including grid size higher Reynolds number. results show that using (excluding y+≤10) as have better accuracy compared to layer, especially when implemented After optimizing hyperparameters high correlation coefficient 0.9324 achieved between shear stress calculated filtered DNS predicted best model, which trained excluding y+≤10. also existing wall-stress models, it shown has model. Additionally, good applied or

Language: Английский

Citations

8
Large-Eddy Simulation of Flow over Boeing Gaussian Bump Using Multi-Agent Reinforcement Learning Wall Model DOI
Di Zhou, Michael P. Whitmore, Kevin P. Griffin

et al.

AIAA Aviation 2019 Forum, Journal Year: 2023, Volume and Issue: unknown

Published: June 8, 2023

Language: Английский

Citations

4
A priori screening of data-enabled turbulence models DOI
Peng E. S. Chen, Yuanwei Bin, Xiang I. A. Yang

et al.

Physical Review Fluids, Journal Year: 2023, Volume and Issue: 8(12)

Published: Dec. 21, 2023

$A$ $p\phantom{\rule{0}{0ex}}o\phantom{\rule{0}{0ex}}s\phantom{\rule{0}{0ex}}t\phantom{\rule{0}{0ex}}e\phantom{\rule{0}{0ex}}r\phantom{\rule{0}{0ex}}i\phantom{\rule{0}{0ex}}o\phantom{\rule{0}{0ex}}r\phantom{\rule{0}{0ex}}i$ validation and verification of black box machine learned turbulence models is time consuming not always fruitful. We discuss a theoretical framework that allows $a$ $p\phantom{\rule{0}{0ex}}r\phantom{\rule{0}{0ex}}i\phantom{\rule{0}{0ex}}o\phantom{\rule{0}{0ex}}r\phantom{\rule{0}{0ex}}i$ screening machine-learned are based on feed-forward neural networks. It requires no knowledge the weights bias only activation function. The method tells one whether model preserves basic calibrations like law wall.

Language: Английский

Citations

4