Smagorinsky/Ladyzhenskaya LES Model

• Smagorinsky/Ladyzhenskaya LES model is a nonlinear eddy-viscosity closure for large eddy simulations that adapts viscosity based on the local velocity gradient or strain.
• It unifies classical turbulence modeling with modern dynamic, discretization-aware, and PDE regularization approaches for complex flow applications.
• The model is applied across engineering and mathematical frameworks, yet faces challenges like coefficient sensitivity and accurate wall treatments.

A Smagorinsky/Ladyzhenskaya-type LES model is a large-eddy-simulation closure in which unresolved turbulence enters through a nonlinear effective viscosity depending on the local magnitude of the resolved velocity gradient or strain. In its classical LES form, the model closes the deviatoric subgrid-scale stress by an eddy viscosity proportional to $(C_s\Delta)^2|S|$; in its Ladyzhenskaya formulation, it appears as a nonlinear diffusion of $p$-Laplacian type, typically written as $-\nabla\cdot(\bar\nu |\nabla u|^{p-2}\nabla u)$, with the Smagorinsky case corresponding to $p=3$ (Larios et al., 10 Aug 2025). Across the literature represented here, this model family appears in engineering LES, finite-element regularizations, stochastic PDEs, data assimilation, reduced-order modeling, and several solver-aware reformulations (Lan et al., 17 Oct 2025, Santos, 2024).

1. Classical constitutive structure

The standard incompressible LES starting point is the filtered Navier–Stokes system

$$\frac{\partial \bar v_i}{\partial t} + \bar v_j \frac{\partial \bar v_i}{\partial x_j} = - \frac{1}{\rho} \frac{\partial \bar p}{\partial x_i} + \nu_N \frac{\partial^2 \bar v_i}{\partial x_j \partial x_j} + \frac{\partial}{\partial x_j} M_{ij}, \qquad \frac{\partial \bar v_i}{\partial x_i}=0,$$

with subgrid stress

$M_{ij} := \bar v_i \bar v_j - \widetilde{v_i v_j}.$

The classical eddy-viscosity ansatz is

$$M_{ij} = \nu\, S_{ij}, \qquad S_{ij} := \frac12\left(\frac{\partial \bar v_i}{\partial x_j} + \frac{\partial \bar v_j}{\partial x_i}\right),$$

and the Smagorinsky choice is

$\nu = C_s^2 \delta^2 (2S_{ij}S_{ij})^{1/2}.$

In the more conventional deviatoric-stress form, this is

$$\tau_{ij}^{\rm SGS}-\frac13\tau_{kk}^{\rm SGS}\delta_{ij}=-2\nu_t S_{ij}, \qquad \nu_t=(C_s\Delta)^2|S|.$$

This is the canonical Smagorinsky closure used as the reference point in several of the cited works (Matharu et al., 2019, Santos, 2024).

In compressible LES, the same closure is written in Favre-filtered variables and usually combined with a modeled isotropic SGS contribution and a turbulent heat flux. For compressible homogeneous isotropic turbulence, the SGS stress is written as

$$\tau_{ij}-\frac{1}{3}\delta_{ij}\tau_{kk} = -2\overline{\rho}\,\nu_{\mathrm{sgs}} \left( \widetilde{S}_{ij}-\frac13\delta_{ij}\widetilde{S}_{kk} \right),$$

with

$p$0

where the cited implementation used $p$1, $p$2, $p$3, and $p$4 (Cordova et al., 2023). This compressible formulation preserves the same constitutive idea—strain-dependent eddy viscosity—while extending it to density-weighted filtering and filtered energy transport.

A finite-element realization used in reduced-basis LES replaces $p$5 by the Frobenius norm of the full gradient and defines a mesh-local viscosity,

$p$6

which is still a Smagorinsky-type nonlinear viscosity, but intrinsically tied to the discrete mesh (Moreno et al., 2023). This suggests that, even within the classical family, the exact tensor norm and the relation between filter width and discretization are solver-dependent modeling choices rather than invariant features.

2. Ladyzhenskaya interpretation and generalized nonlinear viscosity

A central theme of the later mathematical literature is that Smagorinsky closure can be recast as a Ladyzhenskaya-type nonlinear constitutive law. In the data-assimilation formulation,

$p$7

so the LES model becomes

$p$8

In this setting, the model is explicitly called “Smagorinsky/Ladyzhenskaya-type” because $p$9 yields the classical Smagorinsky nonlinear eddy viscosity (Larios et al., 10 Aug 2025). The model is simultaneously a turbulence closure and a PDE regularization.

An even more general finite-element Ladyzhenskaya model writes the added nonlinear diffusion as

$-\nabla\cdot(\bar\nu |\nabla u|^{p-2}\nabla u)$0

The corresponding weak form contains

$-\nabla\cdot(\bar\nu |\nabla u|^{p-2}\nabla u)$1

and the paper states explicitly that $-\nabla\cdot(\bar\nu |\nabla u|^{p-2}\nabla u)$2 and $-\nabla\cdot(\bar\nu |\nabla u|^{p-2}\nabla u)$3 recover the classical Smagorinsky model (Lan et al., 17 Oct 2025). In this sense, Smagorinsky appears as one distinguished member of a wider Ladyzhenskaya family parameterized by nonlinear diffusion exponents.

A related hybrid construction starts not from filtered Navier–Stokes, but from generalized Navier–Stokes with constitutive viscosity

$-\nabla\cdot(\bar\nu |\nabla u|^{p-2}\nabla u)$4

For $-\nabla\cdot(\bar\nu |\nabla u|^{p-2}\nabla u)$5, the authors state that one recovers the Smagorinsky model by properly choosing $-\nabla\cdot(\bar\nu |\nabla u|^{p-2}\nabla u)$6. After filtering, the remaining unclosed terms are approximated by Clark expansion rather than by a Boussinesq SGS ansatz (Rodríguez et al., 2015). This establishes a precise distinction between two traditions: in one, nonlinear viscosity is introduced after filtering as an SGS closure; in the other, it is already part of the constitutive PDE and filtering generates additional closure terms.

The relation is therefore structural rather than merely terminological. Smagorinsky closure is an LES eddy-viscosity model; Ladyzhenskaya regularization is a nonlinear-viscosity PDE model. The two coincide when the LES closure is written as a strain- or gradient-dependent viscosity and inserted directly into the momentum diffusion operator (Larios et al., 10 Aug 2025, Lan et al., 17 Oct 2025).

3. Coefficients, dynamic procedures, and local variants

A persistent issue in the literature is the status of the Smagorinsky coefficient $-\nabla\cdot(\bar\nu |\nabla u|^{p-2}\nabla u)$7. In plane Couette LES with wall damping, the eddy viscosity is written as

$-\nabla\cdot(\bar\nu |\nabla u|^{p-2}\nabla u)$8

with

$-\nabla\cdot(\bar\nu |\nabla u|^{p-2}\nabla u)$9

That study notes that the “typical value seems to be in the range of $p=3$0 to $p=3$1” and uses $p=3$2 both as an LES parameter and as a bifurcation parameter for unstable periodic orbits (Sasaki et al., 2021). In this wall-bounded context, the model is classical static Smagorinsky with van Driest damping.

Dynamic Smagorinsky models replace fixed $p=3$3 by coefficients determined from resolved fields at grid and test-filter levels. In the thermally driven cavity study, the dynamic model is described as allowing the Smagorinsky constant “to vary in space and time depending on an algebraic identity between the subgrid-scale stresses at two different filtered levels and the resolved turbulent stresses.” On the same coarse mesh, the dynamic model captured $p=3$4 of the reference peak vertical velocity fluctuation, whereas the standard model predicted about $p=3$5 near the wall; the dynamic simulation cost about $p=3$6 more CPU time (Sayed et al., 2021). In compressible homogeneous isotropic turbulence, dynamic Smagorinsky and Vreman were reported to be closer to DNS than static Smagorinsky for enstrophy and dilatation trends, while all models reproduced the overall kinetic-energy decay (Cordova et al., 2023).

The local applicability of the dynamic procedure is not straightforward. A detailed analysis of the local DSM showed that, without averaging in homogeneous directions, the dynamic coefficient has a singular leading-order form and can develop exceedingly large positive values. The proposed remedy is the dynamic gradient Smagorinsky model (DGSM), which replaces the strain tensor by the full resolved velocity gradient tensor. In its basic form,

$p=3$7

with a corresponding local Germano coefficient $p=3$8. The paper shows that the leading-order singularity of local DSM is removed, and the DGSM remained stable in decaying isotropic turbulence, temporal mixing layers, and turbulent channel flow without spatial or temporal averaging, while local DSM became unstable in the reported tests (Rozema et al., 2021). This does not eliminate the broader problem of coefficient sensitivity, but it shifts the model from a singular local dynamic law to a bounded one.

These developments show that “the Smagorinsky constant” is not a single question. It is alternately treated as a fixed empirical constant, a dynamic field, a bifurcation or homotopy parameter, or a learned constitutive function. The family remains unified by the nonlinear-viscosity structure, not by a universal prescription for $p=3$9.

4. Anisotropy, wall treatment, and discretization-aware reformulations

A major reformulation is the Lattice Eddy Simulation (LAES) approach. LAES starts from a node-interaction picture on a CFD mesh and derives extra terms

$$\frac{\partial \bar v_i}{\partial t} + \bar v_j \frac{\partial \bar v_i}{\partial x_j} = - \frac{1}{\rho} \frac{\partial \bar p}{\partial x_i} + \nu_N \frac{\partial^2 \bar v_i}{\partial x_j \partial x_j} + \frac{\partial}{\partial x_j} M_{ij}, \qquad \frac{\partial \bar v_i}{\partial x_i}=0,$$0

with analogous expressions for $$\frac{\partial \bar v_i}{\partial t} + \bar v_j \frac{\partial \bar v_i}{\partial x_j} = - \frac{1}{\rho} \frac{\partial \bar p}{\partial x_i} + \nu_N \frac{\partial^2 \bar v_i}{\partial x_j \partial x_j} + \frac{\partial}{\partial x_j} M_{ij}, \qquad \frac{\partial \bar v_i}{\partial x_i}=0,$$1 and $$\frac{\partial \bar v_i}{\partial t} + \bar v_j \frac{\partial \bar v_i}{\partial x_j} = - \frac{1}{\rho} \frac{\partial \bar p}{\partial x_i} + \nu_N \frac{\partial^2 \bar v_i}{\partial x_j \partial x_j} + \frac{\partial}{\partial x_j} M_{ij}, \qquad \frac{\partial \bar v_i}{\partial x_i}=0,$$2. On isotropic meshes, $$\frac{\partial \bar v_i}{\partial t} + \bar v_j \frac{\partial \bar v_i}{\partial x_j} = - \frac{1}{\rho} \frac{\partial \bar p}{\partial x_i} + \nu_N \frac{\partial^2 \bar v_i}{\partial x_j \partial x_j} + \frac{\partial}{\partial x_j} M_{ij}, \qquad \frac{\partial \bar v_i}{\partial x_i}=0,$$3, these reduce to

$$\frac{\partial \bar v_i}{\partial t} + \bar v_j \frac{\partial \bar v_i}{\partial x_j} = - \frac{1}{\rho} \frac{\partial \bar p}{\partial x_i} + \nu_N \frac{\partial^2 \bar v_i}{\partial x_j \partial x_j} + \frac{\partial}{\partial x_j} M_{ij}, \qquad \frac{\partial \bar v_i}{\partial x_i}=0,$$4

which the authors state is identical to classical Smagorinsky LES (Xu et al., 2022). On anisotropic meshes, however, LAES is not standard Smagorinsky LES: it retains directional widths explicitly and yields what the paper describes as an “asymmetric subgrid stress matrix” if interpreted in classical LES terms.

This anisotropic form is used to argue against the usual need for wall damping or dynamic adjustment. LAES sets $$\frac{\partial \bar v_i}{\partial t} + \bar v_j \frac{\partial \bar v_i}{\partial x_j} = - \frac{1}{\rho} \frac{\partial \bar p}{\partial x_i} + \nu_N \frac{\partial^2 \bar v_i}{\partial x_j \partial x_j} + \frac{\partial}{\partial x_j} M_{ij}, \qquad \frac{\partial \bar v_i}{\partial x_i}=0,$$5 in all simulations and attributes improved wall-normal fluctuations to the wall-normal filter width being set to $$\frac{\partial \bar v_i}{\partial t} + \bar v_j \frac{\partial \bar v_i}{\partial x_j} = - \frac{1}{\rho} \frac{\partial \bar p}{\partial x_i} + \nu_N \frac{\partial^2 \bar v_i}{\partial x_j \partial x_j} + \frac{\partial}{\partial x_j} M_{ij}, \qquad \frac{\partial \bar v_i}{\partial x_i}=0,$$6. Since $$\frac{\partial \bar v_i}{\partial t} + \bar v_j \frac{\partial \bar v_i}{\partial x_j} = - \frac{1}{\rho} \frac{\partial \bar p}{\partial x_i} + \nu_N \frac{\partial^2 \bar v_i}{\partial x_j \partial x_j} + \frac{\partial}{\partial x_j} M_{ij}, \qquad \frac{\partial \bar v_i}{\partial x_i}=0,$$7 is small near the wall on channel meshes, wall-normal turbulent diffusion is automatically reduced. The paper therefore claims that LAES “need not use any ad hoc damping functions or any dynamic filtering procedure for wall turbulence,” but the same text also states that the universality claim is only empirically suggested by moderate-$$\frac{\partial \bar v_i}{\partial t} + \bar v_j \frac{\partial \bar v_i}{\partial x_j} = - \frac{1}{\rho} \frac{\partial \bar p}{\partial x_i} + \nu_N \frac{\partial^2 \bar v_i}{\partial x_j \partial x_j} + \frac{\partial}{\partial x_j} M_{ij}, \qquad \frac{\partial \bar v_i}{\partial x_i}=0,$$8 channel-flow evidence and is not broadly established (Xu et al., 2022).

A different reformulation arises from time-discretization-aware LES. For forward Euler, the time difference satisfies the exact filter-swap identity

$$\frac{\partial \bar v_i}{\partial t} + \bar v_j \frac{\partial \bar v_i}{\partial x_j} = - \frac{1}{\rho} \frac{\partial \bar p}{\partial x_i} + \nu_N \frac{\partial^2 \bar v_i}{\partial x_j \partial x_j} + \frac{\partial}{\partial x_j} M_{ij}, \qquad \frac{\partial \bar v_i}{\partial x_i}=0,$$9

which yields an exact space-time residual decomposition with a new temporal term

$M_{ij} := \bar v_i \bar v_j - \widetilde{v_i v_j}.$0

Its leading-order form is

$M_{ij} := \bar v_i \bar v_j - \widetilde{v_i v_j}.$1

a Lax–Wendroff-type diffusion. The resulting augmented Smagorinsky closure is

$M_{ij} := \bar v_i \bar v_j - \widetilde{v_i v_j}.$2

In Burgers tests, this time-aware augmentation remained accurate at coarse time steps where space-only Smagorinsky closures degraded (Agdestein, 16 Jun 2026). This suggests that “Smagorinsky-type” dissipation can be made discretization-aware in both space and time, not only in the filter-width sense.

5. Mathematical analysis, stochastic formulations, and data assimilation

The Ladyzhenskaya interpretation makes the model accessible to monotone-operator and SPDE analysis. For stochastic Ladyzhenskaya–Smagorinsky equations with damping,

$M_{ij} := \bar v_i \bar v_j - \widetilde{v_i v_j}.$3

with

$M_{ij} := \bar v_i \bar v_j - \widetilde{v_i v_j}.$4

the paper proves local monotonicity for $M_{ij} := \bar v_i \bar v_j - \widetilde{v_i v_j}.$5, $M_{ij} := \bar v_i \bar v_j - \widetilde{v_i v_j}.$6, and global monotonicity for $M_{ij} := \bar v_i \bar v_j - \widetilde{v_i v_j}.$7, $M_{ij} := \bar v_i \bar v_j - \widetilde{v_i v_j}.$8, followed by existence and pathwise uniqueness of strong solutions and a small-time large deviation principle (Mohan, 2021). The Smagorinsky case is $M_{ij} := \bar v_i \bar v_j - \widetilde{v_i v_j}.$9. This places the model in a rigorous SPDE framework rather than only an LES-engineering one.

For the 3D periodic stochastic Ladyzhenskaya–Smagorinsky equations,

$$M_{ij} = \nu\, S_{ij}, \qquad S_{ij} := \frac12\left(\frac{\partial \bar v_i}{\partial x_j} + \frac{\partial \bar v_j}{\partial x_i}\right),$$0

the long-time averaged dissipation rate

$$M_{ij} = \nu\, S_{ij}, \qquad S_{ij} := \frac12\left(\frac{\partial \bar v_i}{\partial x_j} + \frac{\partial \bar v_j}{\partial x_i}\right),$$1

is shown to satisfy an upper bound of Kolmogorov order $$M_{ij} = \nu\, S_{ij}, \qquad S_{ij} := \frac12\left(\frac{\partial \bar v_i}{\partial x_j} + \frac{\partial \bar v_j}{\partial x_i}\right),$$2, and the paper emphasizes that it remains finite as $$M_{ij} = \nu\, S_{ij}, \qquad S_{ij} := \frac12\left(\frac{\partial \bar v_i}{\partial x_j} + \frac{\partial \bar v_j}{\partial x_i}\right),$$3. The conclusion drawn there is that, in a periodic domain without boundary layers, the model “does not over-dissipate” (Fan et al., 13 Oct 2025). This sharply qualifies a common criticism of Smagorinsky closure: the cited result does not deny over-dissipation in wall-bounded flows, but it does separate that phenomenon from the periodic bulk case.

In continuous data assimilation, the forecast model

$$M_{ij} = \nu\, S_{ij}, \qquad S_{ij} := \frac12\left(\frac{\partial \bar v_i}{\partial x_j} + \frac{\partial \bar v_j}{\partial x_i}\right),$$4

is nudged toward NSE observations $$M_{ij} = \nu\, S_{ij}, \qquad S_{ij} := \frac12\left(\frac{\partial \bar v_i}{\partial x_j} + \frac{\partial \bar v_j}{\partial x_i}\right),$$5. In 2D, the assimilated system is proved globally well posed for $$M_{ij} = \nu\, S_{ij}, \qquad S_{ij} := \frac12\left(\frac{\partial \bar v_i}{\partial x_j} + \frac{\partial \bar v_j}{\partial x_i}\right),$$6 under $$M_{ij} = \nu\, S_{ij}, \qquad S_{ij} := \frac12\left(\frac{\partial \bar v_i}{\partial x_j} + \frac{\partial \bar v_j}{\partial x_i}\right),$$7, and the error satisfies

$$M_{ij} = \nu\, S_{ij}, \qquad S_{ij} := \frac12\left(\frac{\partial \bar v_i}{\partial x_j} + \frac{\partial \bar v_j}{\partial x_i}\right),$$8

so that

$$M_{ij} = \nu\, S_{ij}, \qquad S_{ij} := \frac12\left(\frac{\partial \bar v_i}{\partial x_j} + \frac{\partial \bar v_j}{\partial x_i}\right),$$9

The paper interprets the residual floor as the irreducible model-observation mismatch caused by the extra LES dissipation (Larios et al., 10 Aug 2025). In this formulation, Smagorinsky/Ladyzhenskaya-type LES is not only a turbulence model but also a biased forecast model whose structural error can be quantified.

At the fully discrete level, the EMAC-Ladyzhenskaya formulation combines nonlinear viscosity with the EMAC convection form,

$\nu = C_s^2 \delta^2 (2S_{ij}S_{ij})^{1/2}.$0

which conserves energy, linear momentum, and angular momentum under weak incompressibility enforcement. The resulting EMAC-LM scheme retains the same asymptotic order as skew-symmetric discretizations but has a Gronwall factor without explicit Reynolds-number dependence, and the reported step-flow benchmarks showed better long-time preservation of momentum and angular momentum than the skew form (Lan et al., 17 Oct 2025).

6. Applications, limitations, and current directions

The cited applications span incompressible channel flow, plane Couette flow, atmospheric boundary layers, buoyancy-driven cavity flow, compressible homogeneous isotropic turbulence, supersonic jets, reduced-order models, SPH reformulations, and machine-learned non-Boussinesq closures. This breadth shows that Smagorinsky/Ladyzhenskaya-type modeling functions less as a single model than as a constitutive backbone reused across numerics and flow classes (Xu et al., 2022, Cordova et al., 2023).

At the same time, several limitations recur. Static Smagorinsky is repeatedly described as over-dissipative, especially near walls or in mean-strain-dominated regions (Sasaki et al., 2021, Junqueira-Junior et al., 2022). In coarse LES of a heated cavity at $\nu = C_s^2 \delta^2 (2S_{ij}S_{ij})^{1/2}.$1, the standard model predicted first moments reasonably well but underpredicted near-wall fluctuation levels relative to the dynamic model (Sayed et al., 2021). In compressible homogeneous isotropic turbulence, all tested SGS models captured the overall decay trends, but the same assessment reported underprediction of peak enstrophy and dilatation and overprediction of peak temperature variance, underscoring the difficulty of compressible turbulence dynamics for classical eddy-viscosity closures (Cordova et al., 2023).

Another recurring theme is that discretization can rival or dominate SGS modeling. In low-order LES of supersonic jets, the authors concluded that the characteristics of numerical discretization can be as important as the effects of the SGS models, and that static Smagorinsky, dynamic Smagorinsky, and Vreman produced similar behavior on the refined grid (Junqueira-Junior et al., 2022). This suggests that model assessment cannot be separated from numerical scheme order, built-in dissipation, and filter-width definition.

Several recent directions explicitly move beyond the strict Boussinesq assumption. In APG turbulent boundary layers, a numerically consistent machine-learned SGS stress

$\nu = C_s^2 \delta^2 (2S_{ij}S_{ij})^{1/2}.$2

was proposed to overcome “the limitations of linear eddy-viscosity closures in complex flows.” The reported a posteriori tests improved mean velocity and wall-shear stress relative to the Dynamic Smagorinsky model and achieved monotonic convergence with grid refinement (Ling et al., 28 Jan 2026). This does not discard Smagorinsky entirely—the first tensor basis term is still Smagorinsky-like—but it treats linear stress-strain proportionality as insufficient in non-equilibrium wall turbulence.

Other extensions change the constitutive law rather than the tensor basis. A PDE-constrained optimization framework on the 1D Kuramoto–Sivashinsky equation replaces the fixed Smagorinsky law

$\nu = C_s^2 \delta^2 (2S_{ij}S_{ij})^{1/2}.$3

by an optimized constitutive function $\nu = C_s^2 \delta^2 (2S_{ij}S_{ij})^{1/2}.$4. The identified optimal law is nonlinear, can become negative for small strain, and reduced trajectory divergence relative to the standard Smagorinsky model, while also illustrating the inherent limitations of the eddy-viscosity paradigm (Matharu et al., 2019). This suggests that the best model inside the Smagorinsky/Ladyzhenskaya class need not be a positive linear function of $\nu = C_s^2 \delta^2 (2S_{ij}S_{ij})^{1/2}.$5.

Finally, some reformulations shift the modeling target altogether. In SPH, a rigorous filtering derivation shows that properly smoothed SPH solves the same filtered equations as explicit LES and could, in principle, use Smagorinsky or dynamic Smagorinsky closures, although the paper argues that approximate deconvolution is more natural in that context (Chola, 2018). In reduced-order modeling, Smagorinsky viscosity has been embedded into reduced basis formulations and monitored through a Kolmogorov-spectrum indicator rather than only through residual norms (Moreno et al., 2023).

Taken together, these works portray the Smagorinsky/Ladyzhenskaya-type LES model as a durable but non-final closure class. Its enduring feature is the nonlinear viscosity law; its unresolved questions concern coefficient universality, tensor structure, wall asymptotics, stochastic consistency, solver dependence, and the extent to which eddy viscosity alone can represent backscatter, anisotropy, and non-equilibrium transfer.