What is spatial filtering in LES?

Spatial filtering is the operation of convolving the Navier-Stokes equation with a filter function G to extract only components larger than the grid width Δ. After filtering, the equations contain the subgrid-scale stress tensor τ_ij, which must be modeled. In practice, the mesh itself often acts as an implicit filter.

How do you choose between Smagorinsky and WALE models?

The Smagorinsky model is simple but tends to over-predict eddy viscosity near walls. The WALE model accounts for the rotational component of the velocity gradient tensor, so it exhibits more natural behavior near walls, making it preferable for problems where wall boundary layers are important. Dynamic Smagorinsky automatically adjusts the model constant for even greater versatility.

How fine should the mesh be for LES?

Near walls, y⁺ ≈ 1 is required, with Δx⁺ ≈ 50–100 in the flow direction and Δz⁺ ≈ 15–40 in the span direction. LES typically requires 2–10 times more cells than RANS, and with 3D unsteady computation, the overall computational cost can be 100–1000 times higher than RANS.

Large Eddy Simulation (LES) — CAE Glossary

Q: What is the difference between LES (Large Eddy Simulation) and RANS?

RANS models all turbulence scales using time averaging, whereas LES directly computes large eddies (energy-containing scales) and models only the small eddies that cannot be resolved by the mesh using subgrid-scale (SGS) models. This makes LES capable of capturing unsteady eddy structures and provides much higher accuracy for separated flows and mixing problems.

Category: Glossary | 2026-03-28

CAE visualization for les - technical simulation diagram

LES — Spatial Filtering Concept

🧑‍🎓

What is the difference between LES (Large Eddy Simulation) and RANS? Both are methods for handling turbulence, right?

🎓

Good question. RANS time-averages "all scales" of turbulence and models everything. LES, on the other hand, directly computes large eddies (the energy-containing scales) while modeling only the small eddies that the mesh cannot resolve. So LES is a "hybrid" approach: large eddies are computed directly, small eddies are modeled. That's why LES captures the unsteady eddy structure much better, and gives far superior accuracy for separated flows and mixing problems.

🧑‍🎓

How do you separate the "large eddies" from the "small eddies"?

🎓

That's done by spatial filtering. You convolve the Navier-Stokes equation with a filter function $G$ to extract only components larger than the grid width $\Delta$. The math looks like this:

$$\bar{f}(\mathbf{x}) = \int G(\mathbf{x} - \mathbf{x}', \Delta) \, f(\mathbf{x}') \, d\mathbf{x}'$$

After filtering, the NS equation contains the subgrid-scale (SGS) stress tensor $\tau_{ij}^{\mathrm{sgs}} = \overline{u_i u_j} - \bar{u}_i \bar{u}_j$, which is an unknown. You need to model it to close the system. In practice, the mesh itself often acts as the filter—this is called "implicit filtering."

SGS Models (Smagorinsky, WALE, Dynamic Models)

🧑‍🎓

What are the main approaches for modeling the SGS stress?

🎓

Here are three major ones.

1. Smagorinsky Model (1963)
The classical SGS model. The SGS eddy viscosity is defined as:

$$\nu_{\mathrm{sgs}} = (C_s \Delta)^2 |\bar{S}|, \quad |\bar{S}| = \sqrt{2 \bar{S}_{ij} \bar{S}_{ij}}$$

$C_s$ is the Smagorinsky constant, typically around 0.17 for isotropic turbulence, but often reduced to 0.1 for channel flow. It's simple and stable, but over-predicts eddy viscosity near walls. A van Driest damping function $f_d = 1 - \exp(-y^+/A^+)$ is typically applied to correct this.

🧑‍🎓

So excessive eddy viscosity near the wall means turbulence is diffused too strongly?

🎓

Exactly. Near the wall, turbulent eddies are small and weak, but Smagorinsky increases viscosity because $|\bar{S}|$ is large (steep velocity gradient). The next model fixes this.

2. WALE Model (Wall-Adapting Local Eddy-viscosity, Nicoud & Ducros 1999)
Uses the trace-free symmetric part of the velocity gradient tensor:

$$\nu_{\mathrm{sgs}} = (C_w \Delta)^2 \frac{(\mathcal{S}_{ij}^d \mathcal{S}_{ij}^d)^{3/2}}{(\bar{S}_{ij} \bar{S}_{ij})^{5/2} + (\mathcal{S}_{ij}^d \mathcal{S}_{ij}^d)^{5/4}}$$

Here $\mathcal{S}_{ij}^d$ is the trace-free symmetric part of the velocity gradient tensor squared, with standard constant $C_w \approx 0.325$. WALE naturally drives $\nu_{\mathrm{sgs}} \to 0$ at the wall—no van Driest damping needed. It also goes to zero in laminar regions, making it user-friendly.

🧑‍🎓

What about the third dynamic model?

🎓

3. Dynamic Smagorinsky Model (Germano et al. 1991, Lilly 1992)
Rather than using a fixed constant $C_s$, this method adjusts it dynamically during computation. It applies two levels of filtering—grid filter $\Delta$ and test filter $\hat{\Delta}$ (usually $\hat{\Delta} = 2\Delta$)—and derives $C_s^2$ locally from the Germano identity:

$$L_{ij} = \widehat{\bar{u}_i \bar{u}_j} - \hat{\bar{u}}_i \hat{\bar{u}}_j$$

The model constant adapts to the flow field, so it works well in separated flows, wakes, and near walls. The downside is numerical instability; you need spatial averaging or clipping to keep $C_s^2$ non-negative. OpenFOAM provides implementations like dynamicKEqn and dynamicLagrangian.

LES Mesh Requirements and Computational Cost

🧑‍🎓

How fine does the mesh need to be for LES? Very different from RANS?

🎓

Very different. LES must directly resolve eddies, so wall-normal mesh requirements are much stricter. Here's the rule of thumb:

Wall-normal: $y^+ \approx 1$ (first cell)
Streamwise: $\Delta x^+ \approx 50 \sim 100$
Spanwise: $\Delta z^+ \approx 15 \sim 40$

For example, external aerodynamics of a car at $\text{Re}_L = 10^6$, RANS might need a few million cells, but wall-resolved LES (WRLES) needs hundreds of millions. Plus, it's a 3D unsteady computation with $\text{CFL} \le 1$, so computational cost becomes 100–1000× higher than RANS.

🧑‍🎓

1000× higher! Is LES still worth it?

🎓

Absolutely. RANS gives you only a time-averaged flow field, but LES reveals eddy dynamics. Consider unsteady pressure fluctuations in a vehicle wake, peak wind loads on buildings, or flame-swirl interaction in combustors—these are "inherently unsteady" phenomena that RANS fundamentally cannot capture. Combustor CFD in aircraft engines is now almost entirely LES, and wind engineering is rapidly adopting it.

LES vs RANS vs DNS — Decision Criteria

🧑‍🎓

There's also DNS (Direct Numerical Simulation). How does that differ from LES?

🎓

DNS resolves all turbulence scales—from the largest eddies down to the Kolmogorov scale $\eta = (\nu^3/\varepsilon)^{1/4}$. No modeling. Most accurate, but the mesh requirement scales as $N \propto \text{Re}^{9/4}$, making it impossible for high-Reynolds number industrial flows.

Here's a summary:

Method	Modeled Range	Mesh Scaling	Practical Re Limit
DNS	None (all scales resolved)	$\sim \text{Re}^{9/4}$	$\sim 10^4$
LES	SGS scales only	$\sim \text{Re}^{13/7}$ (wall-resolved)	$\sim 10^6$
RANS	All turbulence scales	Weakly Re-dependent	No limit

🧑‍🎓

How do you choose in practice? When should I use LES?

🎓

Simple decision tree:

RANS is enough: Steady flows, attached flow drag/lift, pipe pressure drop calculations. Cost is minimal.
LES is needed: Large-scale separation, wake unsteadiness, mixing/combustion, aeroacoustics, wind gust response.
DNS is needed: Fundamental turbulence research, model validation, transition phenomena. Academic focus.

In practice, "try RANS first, switch to LES if accuracy is insufficient" is the realistic path. Starting with LES makes cost projection very difficult.

DES/DDES — Hybrid LES/RANS Methods

🧑‍🎓

If wall meshing is so demanding, couldn't you use RANS near the wall and LES away from it?

🎓

Exactly that idea led to DES (Detached Eddy Simulation), proposed by Spalart in 1997. It behaves as RANS (SA, SST, etc.) near walls and switches to LES away from them. The transition is automatic based on mesh size and wall distance.

But early DES had an issue: the "grey zone problem." If the mesh is moderately fine in the RANS-LES transition zone, RANS viscosity doesn't drop enough and eddies aren't resolved well either—a "neither fish nor fowl" zone develops.

🧑‍🎓

Is there a fix?

🎓

Yes: DDES (Delayed DES) and IDDES (Improved DDES). Spalart et al. (2006) introduced a shielding function $f_d$ that forces RANS mode inside the boundary layer:

$$\tilde{d} = d - f_d \max(0, d - C_{\mathrm{DES}} \Delta)$$

This prevents the boundary layer from accidentally switching to LES (the "Modeled Stress Depletion" problem). IDDES adds wall-modeled LES (WMLES) capability too—if wall mesh is sufficiently fine, it acts like WMLES. OpenFOAM offers SpalartAllmarasDDES and kOmegaSSTDDES; FLUENT supports DDES. The practical path is "RANS → DDES → wall-resolved LES" as needed.

Practical LES Operation Tips

🧑‍🎓

What should I watch out for when running LES in practice?

🎓

Here are the critical ones:

Initial conditions and wash-out time: LES is unsteady, so you need a "wash-out period" before statistics become meaningful. Run the flow through the domain 5–10 times before collecting data.
Inflow boundary conditions: Critical for LES. A uniform inflow needs long development; use Synthetic Eddy Method (SEM) or Recycling/Rescaling to inject realistic turbulent fluctuations efficiently.
Numerical scheme: Central differences are standard. Upwind introduces too much numerical diffusion and kills eddies. Use a blend (5–10% upwind) or low-dissipation schemes (LUST, Gamma).
Time stepping: Keep CFL ≤ 0.5–1.0. Even implicit methods with CFL > 5 lose temporal accuracy and undermine LES.
Statistics: Compute mean, RMS, and Power Spectral Density (PSD). Verify that statistics don't depend on the averaging window—an insufficient window just gives you a lucky snapshot.

🧑‍🎓

I get it now. LES is "high-cost but reveals eddy dynamics, essential for unsteady phenomena."

🎓

Perfect summary. One more thing: GPU computing is dropping LES costs fast. NVIDIA AmgX, PETSc GPU support, and Lattice Boltzmann GPU solvers (PowerFLOW, XFlow, ProLB) are now standard in OEM aerodynamics. In five years, LES will likely be the default in many domains where RANS dominates today.

Accurate CAE terminology is the foundation of team communication. — Project NovaSolver supports learning for practitioners.

Share your LES challenges

Project NovaSolver addresses real issues CAE engineers face daily—setup complexity, computational cost, result interpretation. Your experience drives better tool development.