Turbulence closure modeling with machine learning: a foundational physics perspective

New Journal of Physics, Journal Year: 2024, Volume and Issue: 26(7), P. 071201 - 071201

Published: July 1, 2024

Abstract Turbulence closure modeling using machine learning (ML) is at an early crossroads. The extraordinary success of ML in a variety challenging fields had given rise to expectation similar transformative advances the area turbulence modeling. However, by most accounts, current rate progress toward accurate and predictive ML-RANS (Reynolds Averaged Navier–Stokes) models has been very slow. Upon retrospection, absence rapid can be attributed two factors: underestimation intricacies overestimation ML’s ability capture all features without employing targeted strategies. To pave way for more meaningful closures tailored address nuances turbulence, this article seeks review foundational flow physics assess challenges context data-driven approaches. Revisiting analogies with statistical mechanics stochastic systems, key physical complexities mathematical limitations are explicated. It noted that approaches do not systematically inherent approach or inadequacies forms expressions. study underscores drawbacks supervised learning-based stresses importance discerning framework. As methods evolve (which happening pace) our understanding phenomenon improves, inferences expressed here should suitably modified.

Language: Английский

---

Continuous data assimilation closure for modeling statistically steady turbulence in large-eddy simulation

Physical Review Fluids, Journal Year: 2025, Volume and Issue: 10(1)

Published: Jan. 16, 2025

A closure model is presented for large-eddy simulation (LES) based on the three-dimensional variational data assimilation algorithm. The approach aims at reconstructing high-fidelity kinetic energy spectra in coarse numerical simulations by including feedback control to represent unresolved dynamics interactions flow as stochastic processes. forcing uses statistics obtained from offline and requires only a few parameters compared number of degrees freedom LES. This modeling strategy applied geostrophic turbulence sphere enables simulating indefinitely reduced cost. method accurately recovers zonal velocity profiles three generic situations. Published American Physical Society 2025

Language: Английский

---

Numerically Consistent Data-Driven Subgrid-Scale Model via Data Assimilation and Machine Learning

AIAA SCITECH 2022 Forum, Journal Year: 2025, Volume and Issue: unknown

Published: Jan. 3, 2025

Language: Английский

---

Application of the ensemble Kalman filter to unsteady inflow reconstruction for an axisymmetric body

AIP Advances, Journal Year: 2025, Volume and Issue: 15(3)

Published: March 1, 2025

Environmental perception is a crucial issue for underwater vehicles. This study investigates the use of ensemble Kalman filter to reconstruct unsteady currents ahead these vehicles by sampling surrounding flow fields at scattered locations, which leads an inverse problem. To mitigate high computational cost associated with Monte Carlo methods, axisymmetric simulation employed data assimilation. Therefore, it important discuss influence factors such as observation noise, sample size, and covariance inflation parameters on final performance, especially when compared full three-dimensional model. The results suggest that while most error stems from model discrepancies, careful parameter selection can effectively control within acceptable limits.

Language: Английский

---

LESnets (Large-Eddy Simulation nets): Physics-informed neural operator for large-eddy simulation of turbulence

Journal of Computational Physics, Journal Year: 2025, Volume and Issue: unknown, P. 114125 - 114125

Published: May 1, 2025

Language: Английский

---

Reduced-order model and attractor identification for large eddy simulation of squirrel cage fan

Physics of Fluids, Journal Year: 2023, Volume and Issue: 35(11)

Published: Nov. 1, 2023

A large eddy simulation (LES) of a squirrel cage fan (SCF) provides precise representation turbulent flows with different degrees complexity. This study comprehensively analyzes the coherent structures in an SCF using LES, proper orthogonal decomposition (POD), dynamic mode (DMD), and multi-resolution (mrDMD). An intelligent reduced-order model is established by integrating hierarchical deep learning sparse identification nonlinear dynamics. The result shows that evolution global DMD modes attenuated due to spatial distribution variations localized high-frequency mrDMD modes, along fragmented non-steady development modal patterns. Unlike POD, quantifies quality impeller inlet environment captures antisymmetric low-dimensional associated shedding rotating vortex structures. interaction strength between stationary areas accurately represented attractors characterized petal-like trajectory faithfully maps structural attributes, quasi-periodic behavior, gradual attenuation characteristics exhibited modes. number systems their temporal oscillations are good agreement blades rotational cycles. new insight into engineering advance flow control strategies improve understanding underlying mechanisms.

Language: Английский

---

Scale-resolving simulations of turbulent flows with coherent structures: Toward cut-off dependent data-driven closure modeling

Physics of Fluids, Journal Year: 2024, Volume and Issue: 36(6)

Published: June 1, 2024

Complex turbulent flows with large-scale instabilities and coherent structures pose challenges to both traditional data-driven Reynolds-averaged Navier–Stokes methods. The difficulty arises due the strong flow-dependence (the non-universality) of unsteady structures, which translates poor generalizability models. It is well-accepted that dynamically active reside in larger scales, while smaller scales turbulence exhibit more “universal” (generalizable) characteristics. In such flows, it prudent separate treatment flow-dependent aspects from universal features field. Scale resolving simulations (SRS), as partially averaged (PANS) method, seek resolve motion model only stochastic features. Such an approach requires development scale-sensitive closures not allow for but also appropriate dependence on cut-off length scale. objectives this work are (i) establish physical characteristics dependent turbulence; (ii) develop a procedure subfilter stress neural network at different cut-offs using high-fidelity data; (iii) examine optimal incorporation consistent posteriori use. scale-dependent closure physics analysis performed context PANS approach, technique can be extended other SRS benchmark “flow past periodic hills” case considered proof concept. self-similarity parameters incorporating identified. study demonstrates when data suitably normalized, machine learning based indeed insensitive

Language: Английский

---

Artificial neural-network-based subgrid-scale model for large-eddy simulation of isotropic turbulence

Physics of Fluids, Journal Year: 2024, Volume and Issue: 36(7)

Published: July 1, 2024

This study is concerned with accurately predicting the subgrid-scale (SGS) stress using an artificial neural network (ANN) a linear eddy-viscosity term and nonlinear as input variables. A priori posteriori tests are conducted to examine prediction performance of ANN-based SGS model in decaying homogeneous isotropic turbulence. In test, present shows high correlation coefficients between true predicted stresses, excellent predictions dissipation. it found that can predict turbulence statistics more than traditional dynamic models. The generalization capabilities untrained flow conditions unstrained types turbulent have been evaluated. It proposed provide accurate under different Reynolds numbers types. comparison among several existing models variables presented, demonstrating significant advantage model.

Language: Английский

---

A wall model for separated flows: embedded learning to improve a posteriori performance

Journal of Fluid Mechanics, Journal Year: 2024, Volume and Issue: 1002

Published: Dec. 23, 2024

Developing large-eddy simulation (LES) wall models for separated flows is challenging. We propose to leverage the significance of flow data, which existing theories are not applicable, and knowledge wall-bounded (such as law wall) along with embedded learning address this issue. The proposed so-called features-embedded-learning (FEL) model comprises two submodels: one predicting shear stress another calculating eddy viscosity at first off-wall grid nodes. train former using wall-resolved LES (WRLES) data periodic hill wall. For latter, we a modified mixing length model, coefficient trained ensemble Kalman method. FEL assessed different configurations, resolutions Reynolds numbers. Overall good posteriori performance observed statistics recirculation bubble, stresses turbulence characteristics. modelled subgrid-scale (SGS) grids compared those calculated WRLES data. comparison shows that amplitude distribution SGS energy transfer obtained agree better reference when conventional model.

Language: Английский

---

A physics-informed deep learning closure for Lagrangian velocity gradient evolution

Physics of Fluids, Journal Year: 2023, Volume and Issue: 35(11)

Published: Nov. 1, 2023

The pressure Hessian tensor is entangled with the inherent nonlinearity and nonlocality of turbulence; thus, it crucial importance in modeling Lagrangian evolution velocity gradient (VGT). In present study, we introduce functional strategy into classic structural to model based on deep neural networks (DNNs). its contributions VGT are set as, respectively, learning targets. An a priori test shows that DNN-based accurately establishes mapping from adequately models physical effect invariants. posteriori verifies reproduces well principal features turbulence-like skewness vorticity strain-rate alignments obtained via direct numerical simulations. Importantly, flow topology predicted, particularly for strain-production-dominant regions invariant space. Moreover, an extrapolation generalization ability higher Reynolds number flows have not been trained.

Language: Английский

---