Lie Symmetry-Preserving Turbulence Models

• Lie symmetry-preserving turbulence models are frameworks that construct closures and discretizations ensuring that turbulent formulations inherit the governing equations' invariance properties.
• They employ invariant parameterizations, symmetry-compatible LES closures, and equivariant neural architectures to maintain frame-independence, proper scaling, and conservation laws.
• Applications span wall-turbulence modeling, subgrid-stress representations, and structure-preserving discretizations that enhance numerical robustness and physical fidelity.

Lie symmetry-preserving turbulence models are closures, parameterizations, learned mappings, and numerical discretizations constructed so that they transform compatibly with symmetry groups inherited from the governing equations. In the turbulence literature, this idea appears in several distinct but related forms: invariant parameterization of unresolved terms in reduced PDEs, symmetry-compatible Reynolds-averaged and LES closures, equivariant neural operators for subgrid stress or super-resolution, and structure-preserving discretizations that retain Lie-Poisson geometry in long-time simulations. Across these strands, the central requirement is that modeling or numerical approximation should not introduce frame dependence, violate admissible scalings, or destroy physically decisive invariants unless such symmetry breaking is explicitly justified by the flow configuration, forcing, or boundaries (Cosserat et al., 5 Sep 2025).

1. Symmetry as a constitutive constraint

For incompressible Navier–Stokes, the symmetry content relevant to closure design includes time translations, generalized Galilean boosts, spatial rotations, pressure translations, and scale dilations. In LES, after spatial filtering, the resolved equations retain exactly the same point transformations if and only if the SGS model $\tau_{ij}[\tilde u]$ transforms equivariantly under those groups. This turns symmetry preservation into a constitutive restriction on admissible stress tensors rather than a purely formal property of the unfiltered PDE (Cosserat et al., 5 Sep 2025).

An analogous program exists for reduced geophysical equations. For the barotropic vorticity equation on the $\beta$-plane,

$$\zeta_t + \psi_x \zeta_y - \psi_y \zeta_x + \beta \psi_x = 0,$$

the maximal Lie invariance pseudogroup contains scalings, time translations, $y$-translations, generalized Galilean boosts in $x$, and gauge transformations in $\psi$. Bihlo et al. used moving frames to obtain an explicit functional basis of differential invariants and invariant differentiation operators. In that setting, symmetry preservation is not an afterthought: it organizes the entire space of admissible parameterizations (Bihlo et al., 2011).

The practical significance is that turbulence closures are then built from invariant tensors, invariant differential operators, or invariant combinations of order functions. This prevents spurious dependence on absolute position, orientation, or observer velocity, and it sharply reduces constitutive arbitrariness. A plausible implication is that symmetry acts as a regularizer for both analytic and data-driven models by restricting the hypothesis space to forms compatible with the parent equations.

2. Invariant parameterization in analytic turbulence modeling

A canonical example is invariant hyperdiffusion for two-dimensional turbulence on the $\beta$-plane. Classical hyperviscosity,

$D_n = (-1)^{n-1}\nu \Delta^n \zeta,$

fails to preserve the full symmetry group because under the scaling symmetry it transforms with the wrong weight unless $n=2$, and if $n\neq 2$ or $\beta$0 this spoils full $\beta$1-invariance. By invariantizing the operator through the moving frame, Bihlo et al. obtained

$\beta$2

which is fully $\beta$3-invariant. They also constructed conservative invariant hyperdiffusion terms compatible with circulation, momentum, and energy conservation. In freely decaying turbulence tests, the invariant hyperdiffusion scheme was reported to be close to but not exactly reproducing the $\beta$4 shape of the energy spectrum in the enstrophy inertial range (Bihlo et al., 2011).

A second analytic line is the dilation-based wall-turbulence theory of She et al. Instead of imposing symmetry directly on the mean velocity, the model introduces order functions such as the stress length

$\beta$5

with $\beta$6 and $\beta$7, and requires them to satisfy a dilation-group invariance. The corresponding group invariants are

$\beta$8

From invariant ansätze of the form $\beta$9, $$\zeta_t + \psi_x \zeta_y - \psi_y \zeta_x + \beta \psi_x = 0,$$0, or generalized relations between $$\zeta_t + \psi_x \zeta_y - \psi_y \zeta_x + \beta \psi_x = 0,$$1 and $$\zeta_t + \psi_x \zeta_y - \psi_y \zeta_x + \beta \psi_x = 0,$$2, the framework derives power laws for the viscous sublayer, buffer layer, and log layer, as well as defect-power-law forms in the bulk/core region. The reported exponents are $$\zeta_t + \psi_x \zeta_y - \psi_y \zeta_x + \beta \psi_x = 0,$$3 in the viscous sublayer, $$\zeta_t + \psi_x \zeta_y - \psi_y \zeta_x + \beta \psi_x = 0,$$4 in the buffer layer, and $$\zeta_t + \psi_x \zeta_y - \psi_y \zeta_x + \beta \psi_x = 0,$$5 in the log layer; for the outer defect law, $$\zeta_t + \psi_x \zeta_y - \psi_y \zeta_x + \beta \psi_x = 0,$$6 for channel and TBL and $$\zeta_t + \psi_x \zeta_y - \psi_y \zeta_x + \beta \psi_x = 0,$$7 for pipe. The resulting composite stress-length formula yields a RANS closure that is explicitly assembled from dilation-invariant building blocks (She et al., 2011).

These examples illustrate two distinct uses of symmetry preservation. Invariant parameterization on the $$\zeta_t + \psi_x \zeta_y - \psi_y \zeta_x + \beta \psi_x = 0,$$8-plane preserves the Lie symmetries of the governing PDE itself. Dilation-based wall modeling instead treats symmetry as an organizing principle for order functions entering a closure of the mean-momentum equation. Both approaches use symmetry to constrain model form before calibration.

3. Symmetry-preserving LES closures and stability constraints

For incompressible LES, Cosserat, Razafindralandy, and Selçuk derived the general class of Lie-symmetry-preserving SGS stresses depending on the filtered strain-rate tensor $$\zeta_t + \psi_x \zeta_y - \psi_y \zeta_x + \beta \psi_x = 0,$$9 and filtered vorticity tensor $y$0. The deviatoric stress is expanded in an isotropic tensor basis,

$y$1

where the scalar coefficients depend on the independent scalar invariants of $y$2 and $y$3. Imposing also scale invariance reduces the representation to a normalized form in which the coefficients $y$4 depend only on scale-free invariants $y$5 (Cosserat et al., 5 Sep 2025).

The same work adds a stability criterion based on pointwise nonnegativity of the total energy-dissipation density

$y$6

Requiring the modeled stress to derive from a convex scalar potential $y$7 forces the commutator-type coefficients $y$8 to vanish and yields the reduced three-term form

$y$9

with $x$0 a scalar function satisfying $x$1. The result is a symmetry-preserving and stability-compatible family of SGS models (Cosserat et al., 5 Sep 2025).

A complementary LES development concerns scalar transport. The scalar minimum-dissipation model and symmetry-preserving finite-volume discretization of Rozema-type minimum dissipation were extended to active or passive scalar transport. The discrete construction enforces a skew-symmetric convective operator and a symmetric positive semi-definite diffusive operator, so that the semi-discrete scalar transport system satisfies

$x$2

The scalar-QR closure takes the form

$x$3

and therefore switches off in wall and laminar regions. In a differentially heated cavity, the symmetry-preserving discretization improved prediction of wall shear stress, local Nusselt profiles, fluctuation profiles, and turbulent kinetic energy on highly stretched meshes, specifically for stretching ratio $x$4 on a coarse $x$5 grid (Sun et al., 17 Feb 2025).

Taken together, these LES studies show that symmetry preservation is often inseparable from boundedness and numerical robustness. In this branch of the literature, preserving Galilean, rotational, translational, and scaling symmetries is not only a matter of formal consistency but also a route to controlled dissipation and mesh-independent behavior.

4. Data-driven equivariance and symmetry-aware learning

Recent work extends symmetry preservation from analytic closures to learned models. In forced homogeneous isotropic turbulence, three data-driven LES closures were compared: a tensor-basis neural network (TBNN), a group-convolutional neural network (G-conv), and an unconstrained convolutional network (Conv). All three factor out $x$6 to preserve dimensional and scaling invariance, and all use the filtered velocity gradient so that generalized Galilean invariance is automatic. Rotation and reflection equivariance are then enforced either through invariant tensor bases or through weight projection onto the commutant of the octahedral group representation (Agdestein et al., 5 Mar 2026).

The reported a-priori tensor errors and equivariance errors are:

Model	A-priori tensor error	Equivariance error
No-model	1.0000	—
Smagorinsky	0.9398	—
Clark	0.4544	—
TBNN	0.4093	$x$7
G-conv	0.4196	$x$8
Conv	0.4225	$x$9 a-priori, $\psi$0 a-posteriori

All three learned closures achieved relative LES-solution error of approximately $\psi$1, compared with $\psi$2 for Smagorinsky and $\psi$3 for No-model, while Clark diverged at $\psi$4. TBNN and G-conv eliminated equivariance error to machine precision, whereas Conv retained a measurable violation despite comparable aggregate prediction error. The same study reports that symmetry-preserving models yield more physically consistent velocity-gradient statistics, including more faithful reproduction of the teardrop-shaped $\psi$5–$\psi$6 joint PDF, and that G-conv is about $\psi$7 slower in inference than Conv, while TBNN remains much closer to Conv in cost (Agdestein et al., 5 Mar 2026).

A related but distinct result concerns turbulence super-resolution. For the learned mapping

$\psi$8

rotation equivariance was studied under the discrete octahedral group using the normalized equivariance error

$\psi$9

Standard CNNs were found to acquire part of this symmetry even without explicit augmentation, because turbulence itself provides implicit rotational augmentation in time and space. Models trained on more isotropic midplane channel data had much lower equivariance error than models trained on boundary-layer data, and increasing the number of timesteps $\beta$0 or the number of nonoverlapping boxes per timestep $\beta$1 reduced the averaged error systematically. The study further reports that when anisotropy is strong and sample diversity is small, specifically $\beta$2 or $\beta$3, explicit rotational augmentation or equivariant layers produce a $\beta$4–$\beta$5 reduction in averaged equivariance error and a $\beta$6–$\beta$7 drop in MAE; when data are close to isotropic or sample diversity is high, specifically $\beta$8 or $\beta$9, implicit augmentation drives the averaged equivariance error below $D_n = (-1)^{n-1}\nu \Delta^n \zeta,$0 and the marginal MAE benefit of explicit enforcement falls below $D_n = (-1)^{n-1}\nu \Delta^n \zeta,$1 (Balla et al., 25 Sep 2025).

The data-driven literature therefore distinguishes two mechanisms. One is exact architectural enforcement of symmetry. The other is acquisition of approximate symmetry from distributional properties of the training data. A plausible implication is that symmetry preservation in learned turbulence models depends jointly on architecture and on the anisotropy structure of the dataset.

5. Structure preservation beyond point symmetries

A related line of research extends the symmetry-preservation program from Lie point symmetries to Lie-Poisson structure. In two-dimensional turbulence on the sphere, the Hoppe–Zeitlin finite-mode approximation replaces the continuum Poisson algebra by the finite-dimensional Lie algebra $D_n = (-1)^{n-1}\nu \Delta^n \zeta,$2 and yields the matrix system

$D_n = (-1)^{n-1}\nu \Delta^n \zeta,$3

Because the inviscid dynamics are isospectral, the discrete Casimirs

$D_n = (-1)^{n-1}\nu \Delta^n \zeta,$4

are exactly conserved. Cifani et al. reported robust evidence for the double energy cascade, including formation of the $D_n = (-1)^{n-1}\nu \Delta^n \zeta,$5 scaling of the direct cascade, and emphasized that this could be achieved at modest resolutions, with the Casimir-preserving method showing both $D_n = (-1)^{n-1}\nu \Delta^n \zeta,$6 and $D_n = (-1)^{n-1}\nu \Delta^n \zeta,$7 ranges already at $D_n = (-1)^{n-1}\nu \Delta^n \zeta,$8–$D_n = (-1)^{n-1}\nu \Delta^n \zeta,$9, compared with classical pseudo-spectral studies on periodic domains that required $n=2$0 up to $n=2$1 in each direction before the $n=2$2 slope became apparent (Cifani et al., 2022).

The same philosophy has been extended to multilayer quasi-geostrophic turbulence on the sphere. With layerwise PV matrices $n=2$3 satisfying

$n=2$4

the discrete Lie-Poisson bracket preserves all monomial Casimirs

$n=2$5

exactly to round-off. The IsoSyRK time integrator retains isospectrality and second-order symplecticity. In six-layer unforced runs at $n=2$6 over $n=2$7 days, the first sixteen Casimirs were preserved to machine precision for even orders and to about $n=2$8 for odd orders; in forced three-layer turbulence at $n=2$9, the method reproduced top-layer jets, polar-vortex complexity, and kinetic-energy spectra displaying an $n\neq 2$0 range and Rossby-radius peaks without additional regularization (Franken et al., 2024).

An even broader extension appears in reduced magnetohydrodynamics, Hazeltine’s model, and CHM. Using matrix hydrodynamics and a Poisson integrator, the discrete dynamics preserve coadjoint orbits and Casimirs such as $n\neq 2$1 to machine precision. Long-time runs at spatial resolution $n\neq 2$2 showed magnetic dipole formation in RMHD and Hazeltine’s model, inverse cascade signatures in magnetic energy and mean-square magnetic potential, and model-dependent differences in vorticity dynamics: RMHD developed thin vortex filaments with rapidly growing vorticity values, whereas Hazeltine and CHM exhibited bounded vorticity variation and inverse kinetic-energy cascade signatures (Modin et al., 18 Sep 2025).

These geometric models are not Lie symmetry-preserving closures in the narrow point-symmetry sense. They are, however, part of the same broader program of enforcing the exact transformation or conservation structure of the parent equations at the modeling or discretization level. This suggests a conceptual continuity between invariant constitutive modeling and noncanonical Hamiltonian discretization.

6. Physical admissibility, closure limits, and controversy

The symmetry-preserving literature also contains strong methodological criticism. Khujadze and Frewer argue that not every invariance admitted by an unclosed statistical system is physically admissible. In their critique of symmetry-based planar-jet modeling, they identify a nonphysical statistical scaling,

$n\neq 2$3

and statistical translations of mean and correlation fields, and show that these violate integral constraints or causality requirements. Imposing the physical momentum-flux invariant

$n\neq 2$4

forces restoration of the classical self-similar Reynolds-stress law

$n\neq 2$5

with no extra exponent. Their proposed criterion is that Lie analysis should be performed on the full closed modeled system and restricted to generators descending from genuine deterministic Navier–Stokes symmetries and consistent with experiment or DNS (Khujadze et al., 2020).

The same authors make a parallel argument for symmetry-based predictions of channel-flow moments. They show that fitting full-field moments $n\neq 2$6 can be misleading because these are dominated by the mean profile in strongly sheared flow, while the implied fluctuation moments $n\neq 2$7 fail to match DNS. In particular, the claimed symmetry-induced solutions do not fit Reynolds stress in the center region and break down for higher moments in the log region. Their broader claim is that Lie reduction of an unclosed statistical hierarchy does not bypass the closure problem; it merely shifts closure into arbitrary functions or unclosed admitted symmetries. They therefore emphasize several requirements: separate mean and fluctuations, use unbiased diagnostics, enforce causality by demanding a realizable fine-scale cause for any statistical symmetry, and validate on fluctuation statistics rather than on mean or full-field moments alone (Frewer et al., 2022).

This controversy marks an important boundary condition for the subject. Symmetry preservation is powerful when it is applied to closed equations, to admissible constitutive classes, or to exact geometric structure. It becomes problematic when formal invariances of unclosed statistical equations are reified into physical symmetries without deterministic origin, realizability, or constraint compatibility. A plausible synthesis is that successful Lie symmetry-preserving turbulence models are those that combine three ingredients: a symmetry inherited from the governing dynamics, a closed or explicitly modeled equation set, and validation on nontrivial turbulent observables rather than on symmetry-compatible mean trends alone.