Effect of subgrid-scale anisotropy on wall-modeled large-eddy simulation of turbulent flow with smooth-body separation

This study demonstrates that incorporating anisotropic subgrid-scale stress, particularly normal stress contributions, significantly improves the prediction of flow separation in wall-modeled large-eddy simulations of smooth-body separation under favorable pressure gradients by more accurately capturing the dissipation and diffusion of Reynolds stresses during grid refinement.

The Gist

Imagine you are trying to predict how a river flows around a large, smooth rock in the middle of the stream. Sometimes, the water speeds up over the rock, and sometimes, it slows down and swirls into a chaotic whirlpool (a "separation bubble") on the back side.

In the world of engineering, predicting these whirlpools is crucial. If you get it wrong, an airplane wing might stall, or a car might lose control.

This paper is about a specific computer simulation technique called Large-Eddy Simulation (LES). Think of this technique as a high-tech weather forecast for fluid flow. Instead of trying to track every single water molecule (which would take a supercomputer a million years), the simulation tracks the big, swirling eddies and uses a "rule of thumb" (a model) to guess what the tiny, invisible swirls are doing.

Here is the breakdown of what the researchers found, using simple analogies:

1. The Problem: The "Blind" Model

The researchers were studying flow over a "Gaussian bump" (a smooth, hump-shaped obstacle). They compared two different "rule of thumb" models used to guess the behavior of those tiny, invisible swirls:

• The Old Model (Eddy-Viscosity/Smagorinsky): Imagine a model that assumes the tiny swirls are like a uniform, thick syrup. It only cares about how much energy is being lost (dissipated). It's like a chef who only knows how to stir a pot but doesn't understand that the ingredients inside are moving in different directions.
• The New Model (Anisotropic): This model is more sophisticated. It realizes that the tiny swirls aren't just a uniform syrup; they have a specific shape and direction. They are "stretched" and "squashed" differently depending on where they are. It's like a chef who understands that the ingredients are being pulled apart in one direction and squeezed in another.

The Result: When they refined their computer grid (made the simulation more detailed), the "Old Model" got confused. It predicted the whirlpool would get smaller, then bigger, then smaller again, depending on how detailed the grid was. It was inconsistent. The "New Model," however, gave a steady, reliable answer every time, no matter how detailed the grid became.

2. The Secret Location: The "Windward" Side

You might think the whirlpool forms on the back of the rock (the leeward side), so that's where you need to focus your modeling. The researchers discovered something surprising: The most important place to get the model right is actually on the front of the rock (the windward side).

• The Analogy: Imagine a runner sprinting up a hill (the front of the bump). If the runner's shoes are too slippery or too grippy, it changes how they run all the way to the top and how they stumble when they go down the other side.
• The Finding: The "New Model" correctly predicted how the tiny swirls behaved on the front slope (where the air is speeding up). Because it got the physics right there, it correctly predicted how the air would behave when it reached the back and decided to spin off into a whirlpool. The "Old Model" messed up the front, which caused it to fail at predicting the back.

3. The Mechanism: The "Back-and-Forth" Energy

Why did the new model work better? It comes down to how energy moves between the big swirls (which the computer sees) and the tiny swirls (which the computer guesses).

• The Old Model: It only lets energy flow one way: from the big swirls down to the tiny ones, where it disappears as heat. It's a one-way street.
• The New Model: It allows for backscatter. Sometimes, the tiny swirls get excited and kick energy back up to the big swirls.
• The Analogy: Think of a crowd of people (the big swirls) and a group of kids running around their legs (the tiny swirls).
  • The Old Model says the kids just trip the adults and slow them down.
  • The New Model realizes that sometimes the kids get a running start and actually push the adults forward.
  • In the specific area where the air speeds up over the bump, this "push" (backscatter) is crucial. It helps the air stay attached to the surface longer, which determines exactly where and how big the whirlpool will be on the back side.

4. The "Normal" Stress: The Hidden Hero

The researchers looked closely at the math and found that the "New Model" works because it accounts for Normal Stresses.

• Shear Stress: This is like friction (rubbing side-to-side). The old model was good at this.
• Normal Stress: This is like pressure (pushing up-and-down or side-to-side). The old model ignored this.
• The Discovery: Near the wall, the tiny swirls are being squeezed and stretched vertically. The "New Model" captures this squeezing effect. This squeezing changes the pressure distribution, which is the key to predicting whether the flow will separate (stall) or stay attached.

5. The Conclusion: Why It Matters

This paper proves that to accurately predict complex fluid flows (like air over a wing or water over a ship hull), we can't just use a simple "friction" model for the tiny, invisible parts of the flow.

We need models that understand that these tiny swirls are anisotropic (they have a shape and direction) and that they can push back (backscatter) energy to the larger flow. By getting the physics right on the front slope of the obstacle, we can finally predict the chaotic whirlpools on the back slope with much higher confidence, regardless of how detailed our computer grid is.

In short: To predict the storm on the back of the hill, you have to understand the wind on the front of the hill. And to understand the wind, you need a model that knows the wind isn't just a uniform fog, but a complex, squeezing, pushing dance of tiny eddies.

Technical Summary

1. Problem Statement

Accurate prediction of flow separation in complex turbulent flows is critical for aerodynamic and hydrodynamic performance. While Wall-Modeled Large-Eddy Simulation (WMLES) offers a computationally efficient alternative to Wall-Resolved LES (WRLES) by modeling near-wall effects, it faces significant challenges in predicting separation bubbles, particularly under non-equilibrium conditions.

The core problem addressed is the sensitivity of separation prediction to Subgrid-Scale (SGS) models. Classical eddy-viscosity models (e.g., Smagorinsky) assume isotropic turbulence and often yield non-monotonic convergence in separation bubble size as grid resolution increases. They frequently fail to capture the correct separation onset and extent in complex flows, such as flow over a Gaussian-shaped bump, because they cannot simultaneously represent both energy dissipation and the anisotropic stress contributions required at coarse resolutions typical of WMLES.

2. Methodology

Numerical Approach:

• Flow Configuration: The study simulates flow over a spanwise-uniform Gaussian-shaped bump at a Reynolds number $Re_L = 2 \times 10^6$. This configuration mimics smooth junctions (e.g., wing-fuselage) where smooth-body separation occurs under combined pressure gradients and surface curvature.
• Boundary Conditions: To isolate SGS model effects from wall-modeling errors, the authors replaced the standard no-slip condition with an idealized Neumann boundary condition. This prescribes the mean wall-shear stress directly from a reference Direct Numerical Simulation (DNS) by Uzun & Malik (2022).
• Grids: Simulations were performed on four mesh resolutions (Coarsest to Fine), ranging from ~2 million to ~1 billion cells. The finest mesh approaches WRLES resolution in the streamwise/spanwise directions but remains coarser in the wall-normal direction.

SGS Models Compared:

1. Smagorinsky Model (SM): A classical isotropic eddy-viscosity model.
2. Modified Smagorinsky Model (MSM): An anisotropic model that augments the SM with an additional stress term:
   $$\tau_{ij}^{ani} = C_a \Delta^2 (S_{ik}R_{kj} - R_{ik}S_{kj})$$
   where $S$ is the strain-rate tensor and $R$ is the rotation-rate tensor. Crucially, this term contributes zero SGS kinetic energy dissipation ($\tau_{ij}^{ani}S_{ij} = 0$), allowing the study of anisotropy's role in momentum transport and stress fluctuations independent of energy transfer.

Analysis Techniques:

• Virtual Interface Experiments: The domain was split into upstream and downstream sections with different SGS models to identify the critical region where anisotropy matters most.
• Budget Analysis: Detailed examination of the mean streamwise momentum equation and Reynolds stress transport equations to isolate the physical mechanisms driving separation.
• A Priori Study: Filtered DNS of turbulent Couette-Poiseuille flow (with favorable pressure gradient) was used to validate the physical realism of the SGS stress anisotropy observed in the WMLES.

3. Key Contributions

1. Identification of Critical Region: The study pinpointed the windward side (upstream of the bump peak) under a strong Favorable Pressure Gradient (FPG) as the critical region. Anisotropic SGS stress in this region dictates the downstream evolution of the boundary layer and the eventual separation point.
2. Mechanism of Improvement: The paper elucidates why anisotropic models improve predictions. It is not merely due to better mean stress representation, but rather the anisotropic SGS stress fluctuations which modify SGS dissipation and diffusion in the Reynolds stress transport equation.
3. Role of Normal Stresses: The study demonstrates that the improvement stems primarily from the inclusion of significant normal stress components ($\tau_{11}$ and $\tau_{22}$) in the SGS model, which are absent in isotropic models but crucial for capturing near-wall dynamics under pressure gradients.
4. Resolution Dependence: The authors clarified the transition of dominance from mean SGS shear stress (at coarse resolutions) to resolved Reynolds stresses (at finer resolutions), explaining why isotropic models fail to converge monotonically while anisotropic models provide consistent results.

4. Key Results

Separation Prediction & Convergence:

• SM (Isotropic): Exhibited non-monotonic convergence. On the medium mesh, it failed to predict any separation. On the coarsest and fine meshes, it predicted separation, but the bubble size varied erratically with grid refinement.
• MSM (Anisotropic): Produced a separation bubble consistent with DNS across all mesh resolutions. The prediction was robust and monotonic, converging smoothly toward the DNS reference as resolution increased.

Physical Mechanisms:

• Reynolds Stress Peaks: In the FPG region, the MSM successfully reproduced internal peaks in Reynolds shear ($u'_1 u'_2$) and normal ($u'_2 u'_2$) stresses near the wall, which were observed in DNS but missed by the SM.
• SGS Dissipation & Backscatter: The Reynolds stress transport analysis revealed that the SM acts purely as an energy sink. In contrast, the MSM introduces backscatter (energy transfer from subgrid to resolved scales) and specific diffusion patterns driven by the anisotropic term. This facilitates the formation of the internal stress peaks.
• Momentum Balance:
  • Coarse Mesh: Mean SGS shear stress dominates; both models behave similarly.
  • Medium Mesh: Resolved Reynolds stresses become significant. The MSM's ability to correctly shape the Reynolds stress distribution (via SGS fluctuations) alters the momentum balance, leading to correct separation prediction.
  • Fine Mesh: Most scales are resolved; model differences diminish.

A Priori Validation:

• Filtered DNS of Couette-Poiseuille flow confirmed that near-wall turbulence under FPG is highly anisotropic, with normal stress components playing a dominant role. The MSM captured these anisotropic features realistically, whereas the SM did not.

5. Significance

This work provides a fundamental physical explanation for the long-standing issue of non-monotonic convergence in WMLES separation predictions. It establishes that SGS anisotropy is not just a correction for energy dissipation but a critical mechanism for momentum transport and Reynolds stress evolution in non-equilibrium flows.

The findings suggest that future SGS model development for WMLES must move beyond simple eddy-viscosity formulations. Models must explicitly account for anisotropic stress fluctuations, particularly normal stresses, to accurately capture the history effects of favorable pressure gradients that determine separation onset. This insight paves the way for more robust, high-fidelity simulations of complex engineering flows (e.g., aircraft stall, turbine blade separation) without requiring prohibitively fine grid resolutions.