Smagorinsky Model in LES Turbulence

• The Smagorinsky model is a subgrid-scale closure in LES that uses an algebraic eddy viscosity based on the local deformation rate to model unresolved stresses.
• It effectively spans various flow regimes with its simple formulation, though it tends to over-dissipate energy in wall-bounded and shear-driven environments.
• Enhanced variants, including dynamic, corrected, and tensor-coefficient models, adjust coefficients and incorporate backscatter to improve accuracy in complex turbulent flows.

The Smagorinsky model is a large-eddy simulation (LES) subgrid-scale closure in which unresolved stresses are represented by an algebraic eddy viscosity proportional to a filter length squared and to the magnitude of the resolved deformation rate. In filtered notation, its classical form is

$$\tau^{sgs}_{ij}-\tfrac13\tau^{sgs}_{kk}\delta_{ij}=-2\,\nu_t\,\overline S_{ij}, \qquad \nu_t=(C_S\Delta)^2\,|\overline S|,$$

while, in one periodic-box formulation, the same closure appears through the nonlinear diffusion term

$$-\nabla\!\cdot\Bigl((C_s\delta)^2\,|\nabla u|\,\nabla u\Bigr)$$

in the momentum equation (Rozema et al., 2021, Layton, 2016). The model is one of the most studied sub-grid scale representations in LES, and its simplicity has made it a reference point for analyses of dissipation, wall effects, dynamic coefficient procedures, reduced-order approximations, and structurally enriched SGS closures (Perez et al., 2024).

1. Classical closure and parameterization

In LES, filtering the incompressible Navier–Stokes equations introduces the subgrid-scale stress

$$\tau_{ij}=\overline{u_i u_j}-\overline u_i\,\overline u_j,$$

which must be modeled. The classical Smagorinsky closure adopts the Boussinesq hypothesis and assumes that the deviatoric SGS stress is aligned with the resolved rate-of-strain tensor,

$$\overline S_{ij}=\tfrac12\bigl(\partial_i\overline u_j+\partial_j\overline u_i\bigr), \qquad |\overline S|=\sqrt{2\,\overline S_{ij}\,\overline S_{ij}},$$

through the eddy viscosity

$\nu_t=(C_S\Delta)^2\,|\overline S| .$

Equivalent formulations appear in incompressible, compressible, and periodic-box settings, sometimes written with $|\nabla u|$ rather than $|S|$, but the constitutive idea is the same: unresolved motions are modeled by an effective viscosity that grows with local resolved deformation (Jadhav et al., 2023, Layton, 2016).

The filter width $\Delta$ is typically tied to the discretization. Examples in the literature include $\Delta=2\pi/N$ on an $N^3$ pseudo-spectral grid, $$-\nabla\!\cdot\Bigl((C_s\delta)^2\,|\nabla u|\,\nabla u\Bigr)$$0 on a rectangular grid, and $$-\nabla\!\cdot\Bigl((C_s\delta)^2\,|\nabla u|\,\nabla u\Bigr)$$1 in HIGRAD/FIRETEC (Jadhav et al., 2023, Rozema et al., 2021, Perez et al., 2024). Representative coefficient choices reported in the literature include $$-\nabla\!\cdot\Bigl((C_s\delta)^2\,|\nabla u|\,\nabla u\Bigr)$$2, $$-\nabla\!\cdot\Bigl((C_s\delta)^2\,|\nabla u|\,\nabla u\Bigr)$$3 for homogeneous isotropic turbulence, $$-\nabla\!\cdot\Bigl((C_s\delta)^2\,|\nabla u|\,\nabla u\Bigr)$$4 in a canopy LES implementation, and values in the range $$-\nabla\!\cdot\Bigl((C_s\delta)^2\,|\nabla u|\,\nabla u\Bigr)$$5 in many flows (Layton, 2016, Jadhav et al., 2023, Perez et al., 2024, Santos et al., 2024).

A defining feature of the classical model is that it is purely dissipative. The modeled SGS dissipation in one standard formulation is

$$-\nabla\!\cdot\Bigl((C_s\delta)^2\,|\nabla u|\,\nabla u\Bigr)$$6

so increasing $$-\nabla\!\cdot\Bigl((C_s\delta)^2\,|\nabla u|\,\nabla u\Bigr)$$7 increases dissipation and damps small scales more strongly (Rozema et al., 2021). This also means that the classical eddy-viscosity model cannot represent backscatter, i.e. energy transfer from unresolved fluctuations back into the resolved mean or resolved scales (Siddiqua et al., 2022, Jadhav et al., 2023).

2. Energy balance and dissipation scaling

For the periodic Smagorinsky model on $$-\nabla\!\cdot\Bigl((C_s\delta)^2\,|\nabla u|\,\nabla u\Bigr)$$8,

$$-\nabla\!\cdot\Bigl((C_s\delta)^2\,|\nabla u|\,\nabla u\Bigr)$$9

the exact energy equality for strong solutions is

$$\tau_{ij}=\overline{u_i u_j}-\overline u_i\,\overline u_j,$$0

The instantaneous total model dissipation is therefore

$$\tau_{ij}=\overline{u_i u_j}-\overline u_i\,\overline u_j,$$1

with long-time average

$$\tau_{ij}=\overline{u_i u_j}-\overline u_i\,\overline u_j,$$2

(Layton, 2016).

Using the large-scale quantities

$$\tau_{ij}=\overline{u_i u_j}-\overline u_i\,\overline u_j,$$3

with $$\tau_{ij}=\overline{u_i u_j}-\overline u_i\,\overline u_j,$$4, the periodic-box analysis yields the bound

$$\tau_{ij}=\overline{u_i u_j}-\overline u_i\,\overline u_j,$$5

or equivalently

$$\tau_{ij}=\overline{u_i u_j}-\overline u_i\,\overline u_j,$$6

As $$\tau_{ij}=\overline{u_i u_j}-\overline u_i\,\overline u_j,$$7 or $$\tau_{ij}=\overline{u_i u_j}-\overline u_i\,\overline u_j,$$8, the last term vanishes; as $$\tau_{ij}=\overline{u_i u_j}-\overline u_i\,\overline u_j,$$9, the $$\overline S_{ij}=\tfrac12\bigl(\partial_i\overline u_j+\partial_j\overline u_i\bigr), \qquad |\overline S|=\sqrt{2\,\overline S_{ij}\,\overline S_{ij}},$$0 term vanishes (Layton, 2016).

This result is central in the dissipation theory of the model. For body-force-driven turbulence under periodic boundary conditions, it gives the same $$\overline S_{ij}=\tfrac12\bigl(\partial_i\overline u_j+\partial_j\overline u_i\bigr), \qquad |\overline S|=\sqrt{2\,\overline S_{ij}\,\overline S_{ij}},$$1 scaling as the incompressible Navier–Stokes equations up to a vanishing $$\overline S_{ij}=\tfrac12\bigl(\partial_i\overline u_j+\partial_j\overline u_i\bigr), \qquad |\overline S|=\sqrt{2\,\overline S_{ij}\,\overline S_{ij}},$$2 correction and a vanishing $$\overline S_{ij}=\tfrac12\bigl(\partial_i\overline u_j+\partial_j\overline u_i\bigr), \qquad |\overline S|=\sqrt{2\,\overline S_{ij}\,\overline S_{ij}},$$3 correction. The stated conclusion is that, without boundary layers, the unmodified Smagorinsky model does not over-dissipate (Layton, 2016).

3. Boundary layers, mesh effects, and the over-dissipation problem

The Smagorinsky model is often reported to over-damp or over-dissipate in wall-bounded and shear-driven flows. In lid-driven shear flow on an under-resolved mesh $$\overline S_{ij}=\tfrac12\bigl(\partial_i\overline u_j+\partial_j\overline u_i\bigr), \qquad |\overline S|=\sqrt{2\,\overline S_{ij}\,\overline S_{ij}},$$4 satisfying

$$\overline S_{ij}=\tfrac12\bigl(\partial_i\overline u_j+\partial_j\overline u_i\bigr), \qquad |\overline S|=\sqrt{2\,\overline S_{ij}\,\overline S_{ij}},$$5

the computed time-averaged dissipation obeys

$$\overline S_{ij}=\tfrac12\bigl(\partial_i\overline u_j+\partial_j\overline u_i\bigr), \qquad |\overline S|=\sqrt{2\,\overline S_{ij}\,\overline S_{ij}},$$6

For $$\overline S_{ij}=\tfrac12\bigl(\partial_i\overline u_j+\partial_j\overline u_i\bigr), \qquad |\overline S|=\sqrt{2\,\overline S_{ij}\,\overline S_{ij}},$$7, the viscosity term is negligible, and the estimate indicates over-dissipation for any $$\overline S_{ij}=\tfrac12\bigl(\partial_i\overline u_j+\partial_j\overline u_i\bigr), \qquad |\overline S|=\sqrt{2\,\overline S_{ij}\,\overline S_{ij}},$$8 and $$\overline S_{ij}=\tfrac12\bigl(\partial_i\overline u_j+\partial_j\overline u_i\bigr), \qquad |\overline S|=\sqrt{2\,\overline S_{ij}\,\overline S_{ij}},$$9, consistent with numerical evidence on the effects of model viscosity without wall damping (Pakzad, 2018).

That analysis also states that, for the lid-driven problem, the turbulent boundary layer is a more important length scale than the Kolmogorov microscale. Shear flows display two small scales,

$\nu_t=(C_S\Delta)^2\,|\overline S| .$0

and the controlling scale for energy dissipation in the lid-driven problem is $\nu_t=(C_S\Delta)^2\,|\overline S| .$1, not $\nu_t=(C_S\Delta)^2\,|\overline S| .$2 (Pakzad, 2018). This is consistent with the periodic-box conclusion that the dominant over-dissipation mechanism is associated with boundary effects rather than with the turbulent cascade itself (Layton, 2016).

Several remedies have therefore been developed. One is damping of the SGS viscosity near walls. With a damping function $\nu_t=(C_S\Delta)^2\,|\overline S| .$3, the model becomes

$\nu_t=(C_S\Delta)^2\,|\overline S| .$4

and the time-averaged dissipation is

$\nu_t=(C_S\Delta)^2\,|\overline S| .$5

For a modified $\nu_t=(C_S\Delta)^2\,|\overline S| .$6 damping function $\nu_t=(C_S\Delta)^2\,|\overline S| .$7 with algebraic wall contact order $\nu_t=(C_S\Delta)^2\,|\overline S| .$8, the reported bound is

$\nu_t=(C_S\Delta)^2\,|\overline S| .$9

which restores $|\nabla u|$0 scaling independently of $|\nabla u|$1 as $|\nabla u|$2 (Pakzad, 2016).

Another remedy is spatially adaptive damping. One recent adaptive formulation replaces the constant coefficient by

$|\nabla u|$3

so that

$|\nabla u|$4

That work states existence and uniqueness of weak solutions and an energy dissipation bound for the adaptive model, with the explicit aim of reducing over-dissipation near boundaries (Santos et al., 2024).

4. Dynamic, corrected, and structurally enriched variants

The classical constant-coefficient model has generated a broad family of variants designed to reduce excessive dissipation, recover backscatter-like effects, or relax the Boussinesq alignment assumption.

Variant	Defining modification	Stated purpose or effect
Dynamic Smagorinsky	$|\nabla u|$5	Compute the coefficient on-the-fly from two filter levels
Dynamic gradient Smagorinsky	Replace $|\nabla u|$6 by $|\nabla u|$7	Remove the local dynamic singularity without averaging
Corrected Smagorinsky model	Add $|\nabla u|$8	Incorporate energy flow from unresolved fluctuations back into the mean
Dynamic tensor-coefficient model	Replace scalar coefficient by tensor $|\nabla u|$9	Relax local stress–strain alignment
Helicity SGS model	Add helicity-gradient/vorticity coupling to Smagorinsky viscosity	Incorporate structural information beyond turbulence intensity

The dynamic Smagorinsky procedure of Germano et al. compares SGS stresses at a grid filter $|S|$0 of width $|S|$1 and a test filter of width $|S|$2. In one formulation,

$|S|$3

and in rotating stratified turbulence the coefficient was stabilized by whole-domain averaging in the absence of walls (Jadhav et al., 2023). A related issue is that the widely used dynamic Smagorinsky model becomes singular if used locally without averaging. The local dynamic gradient Smagorinsky model addresses this by replacing the rate-of-strain tensor by the full resolved velocity gradient,

$|S|$4

which removes the singularity because the denominator now decays only when the full $|S|$5 vanishes (Rozema et al., 2021).

The corrected Smagorinsky model is motivated by the fact that the classical model is an eddy-viscosity approximation to a resolved mean velocity and cannot represent a flow of energy from unresolved fluctuations to the resolved mean velocity. Its governing equation is

$|S|$6

and the added Voigt-type term is designed to model the omitted $|S|$7 contribution from the Reynolds-stress energy balance (Siddiqua et al., 2022).

The dynamic tensor-coefficient Smagorinsky model replaces the scalar eddy-viscosity coefficient by a tensorial coefficient. In one reduced formulation,

$|S|$8

with four independent dynamic coefficients after imposing tracelessness constraints. That model is reported to vanish in laminar flow and at solid boundaries, and to have the correct asymptotic behavior in the near-wall region of a turbulent boundary layer (Agrawal et al., 2022).

A different structural enrichment incorporates helicity. In LES notation, the deviatoric SGS stress is modeled as

$|S|$9

with

$\Delta$0

The stated motivation is that the standard Smagorinsky framework uses only information about turbulence intensity through turbulent energy, whereas helicity supplies structural or geometrical information of turbulence (Yokoi et al., 27 Mar 2026).

5. Functional analysis, well-posedness, and regularity

The Smagorinsky model has also become a substantial object of nonlinear analysis. For the unsteady rotational Smagorinsky model, the classical viscous term is replaced by a weighted curl operator,

$\Delta$1

with $\Delta$2. In the natural case $\Delta$3, the appropriate space is

$\Delta$4

and existence of weak solutions follows from Bochner pseudo-monotone evolution equations. The threshold $\Delta$5 is identified as intrinsic, and $\Delta$6 is the minimal exponent making the wall-degenerate convective term manageable in the pseudo-monotone setting (Berselli et al., 2021).

Recent variational refinements formulate the SGS stress as

$\Delta$7

where $\Delta$8 is a small regularization parameter and $\Delta$9 ensures $\Delta=2\pi/N$0. In that framework, existence of weak solutions is obtained on bounded Lipschitz domains for $\Delta=2\pi/N$1 and $\Delta=2\pi/N$2, while uniqueness is asserted under the additional regularity $\Delta=2\pi/N$3 (Santos, 2024).

Analyses with dynamic or evolving boundaries extend the same program. One formulation couples the bulk Smagorinsky system

$\Delta=2\pi/N$4

to a Wentzell-type tangential boundary evolution and proves existence, uniqueness, an energy estimate, and higher-order Sobolev regularity (Santos et al., 2024). Related corrected-model analyses in time-dependent domains derive $\Delta=2\pi/N$5-bounds, $\Delta=2\pi/N$6-error estimates relative to Navier–Stokes, and long-time convergence statements under assumptions on forcing and domain smoothness (Santos, 2024). A further high-order extension introduces both a mild inverse-Laplacian temporal correction and an $\Delta=2\pi/N$7-order dissipation term,

$\Delta=2\pi/N$8

and proves enhanced regularity and error estimates in higher Sobolev spaces (Santos, 2024).

A common theme across these analyses is that the nonlinear SGS operator is treated through monotonicity, coercivity, compactness, or pseudo-monotonicity. This has turned the Smagorinsky closure from a purely empirical LES device into a model with a substantial functional-analytic theory.

6. Discretization, reduced models, and representative applications

The model remains prominent in applied LES because its algebraic form is easy to embed in existing solvers. In HIGRAD/FIRETEC, the Smagorinsky closure

$\Delta=2\pi/N$9

was implemented with fixed $N^3$0 and $N^3$1 equal to the local cell diagonal. In neutral, no-fire simulations of flow over a homogeneous forest edge, the modeled fraction $N^3$2 was $N^3$3 of the total Smagorinsky TKE profile, peaks in modeled TKE occurred at shear interfaces, and resolved TKE profiles, mean velocity profiles, sweep/ejection ratios, and spectral slopes were comparable to those of FIRETEC’s original Linn 1.5-order TKE closure. Against field data at mid-canopy, the reported root-mean-square errors were $N^3$4 for Linn and $N^3$5 for Smagorinsky (Perez et al., 2024).

In decaying rotating stratified turbulence, LES with constant-coefficient Smagorinsky, dynamic Smagorinsky, and dynamic Clark closures were compared against filtered DNS on a $N^3$6 grid. The reported result was that all SGS model predictions were very similar, with the classical Smagorinsky model displaying the highest deviation compared to DNS. Dynamic Smagorinsky improved the fit up to $N^3$7, but both Smagorinsky variants over-dissipated the smallest LES scales and failed to reproduce inverse-cascade behavior because $N^3$8 everywhere, so no energy flows from subgrid to resolved scales even when the DNS does (Jadhav et al., 2023).

The model has also become a target for advanced discretizations. A certified reduced-basis Smagorinsky model for steady flows used the Empirical Interpolation Method to approximate the nonlinear eddy diffusion term and reported speed-up factors $N^3$9 for a backward-facing step and $$-\nabla\!\cdot\Bigl((C_s\delta)^2\,|\nabla u|\,\nabla u\Bigr)$$00 for a lid-driven cavity in FreeFem++ (Rebollo et al., 2017). In the virtual element setting, a general-order discretization with the local eddy viscosity

$$-\nabla\!\cdot\Bigl((C_s\delta)^2\,|\nabla u|\,\nabla u\Bigr)$$01

was tested up to $$-\nabla\!\cdot\Bigl((C_s\delta)^2\,|\nabla u|\,\nabla u\Bigr)$$02. The reported conclusion was that isotropic refinement with hanging nodes enhanced accuracy relative to anisotropic meshes, used fewer degrees of freedom than uniform meshes, and yielded the most stable Newton convergence behavior (Berrone et al., 3 Oct 2025).

Across these settings, a consistent pattern emerges. The classical Smagorinsky model remains attractive because it is local, algebraic, and computationally light; it captures many mean and resolved statistics adequately in homogeneous or wall-free settings; but it also inherits structural limitations tied to positivity of eddy viscosity, local stress–strain alignment, and sensitivity to boundary layers. Much of the subsequent literature can be read as a sequence of attempts to preserve its simplicity while correcting those three liabilities through damping, dynamic coefficient evaluation, modified dissipation, tensorial coefficients, or additional structural information such as helicity (Layton, 2016, Rozema et al., 2021, Agrawal et al., 2022, Yokoi et al., 27 Mar 2026).