Professional Documents
Culture Documents
DOI 10.1007/s10494-010-9264-5
F. R. Menter · Y. Egorov
Received: 21 August 2009 / Accepted: 14 May 2010 / Published online: 10 June 2010
© Springer Science+Business Media B.V. 2010
Abbreviations
EARSM Explicit Algebraic RSM;
DES Detached Eddy Simulation;
DIT Decaying Isotropic Turbulence;
HWN High Wave Number (damping);
KSKL K Square-root-K L (model);
LES Large Eddy Simulation;
MILES Monotonically Integrated LES;
RANS Reynolds-Averaged Navier–Stokes (equations);
RSM Reynolds Stress Model;
SAS Scale-Adaptive Simulation;
SRS Scale Resolving Simulation;
SKL Square-root-K L (model);
SST Shear-Stress Transport (model);
1 Introduction
calibrated to comply with the logarithmic law without additional terms depending on
wall distance. This deficiency has eventually led to the advent of the k − ε model
as the major industrial two-equation model. The following quote from Rodi [29]
stresses that this was largely due to the k–kL model’s inability of handling the log-
layer: “In the late 60’s and early 70’s, some investigations have been carried out at
Imperial College, London, with Rotta’s kL equation, but this requires an extra term
near the wall to conform with the log law. Hence in the early 70’s there was soon
a switch over to the k − ε model . . . By choosing the constants properly, the (k–ε)
model can be made consistent with the log law.”
In recent years, steps of modernizing the kL-equation have been taken (Menter
and Egorov [18–23]). It was argued that Rotta’s assumptions, leading to the term
with the third derivative of the velocity in his kL-equation, is not consistent with the
nature of the underlying term in the exact equation. It was shown that it is more
consistent in inhomogeneous flows to keep the second derivative of the velocity
instead of the third as in Rotta’s model. With the second derivative, the k–kL model
can satisfy the logarithmic layer equations without additional terms. Furthermore,
the inclusion of the second derivative allows the model to adjust to resolved turbulent
structures without dissipating them as classical RANS models would. The ability of
the new model to operate in Scale-Resolving Simulation (SRS) mode is termed Scale-
Adaptive Simulation (SAS). It essentially means that, under certain conditions, the
model automatically balances the contributions of modelled and resolved parts of the
turbulent stresses. One of the more interesting features of the SAS approach is that
for unstable flows, the model changes smoothly from an LES model through various
stages of eddy-resolution back to a steady RANS model, based on the specified time
step, as will be demonstrated for a 2D periodic hill flow. The √ new model has been
formulated as a one- and a two-equation model using 8 = kL as the new scaling
variable in Menter et al. [22]. In the current article, only the two-equation model will
be discussed. It is also important to note that the model as presented uses an eddy
viscosity to compute the turbulent stresses, but this is not inherent to the formulation.
Like any other two-equation model, it can also serve as a basis for EARSMs and
RSMs.
SAS models behave in many situations similar to Detached Eddy Simulation
(DES) models (e.g. Spalart et al. [37], Spalart [36], Strelets et al. [39]), but without an
explicit influence of the grid spacing on the RANS mode of the model. This allows for
a safer passage from RANS to SRS, especially for complex applications, where high
quality ‘LES’-meshes cannot easily be generated for the detached flow regions. Due
to space limitations, the few comparisons of the current approach to hybrid models
will be restricted to the most widely used DES method. A more detailed account of
hybrid turbulence modelling can be found in Sagaut et al. [34] and Fröhlich and von
Terzi [9].
There is a potential for confusion in terminology for simulations based on
unsteady RANS (URANS) models. There are two schools of thought. The first
is the “classical” concept of ‘separation of scales’ where URANS averages out all
turbulent fluctuations and resolves only frequencies far lower than those of the
turbulent fluctuations (typically resulting from variations in geometry or unsteady
boundary conditions). This concept is also often (mentally) applied to unsteady
vortex shedding behind bluff bodies, where classical URANS models (like k–ε or
k–ω) typically only reproduce the single vortex shedding frequency. The second
116 Flow Turbulence Combust (2010) 85:113–138
2 Rotta’s k − kL Model
Rotta [32, 33] formulated an exact transport equation for the variable kL based on
the definition of an integral length scale, L, and used it as a starting point of a term-
by-term closure. This paragraph will give a short recapitulation of Rotta’s derivation.
Assuming flows with a dominant shear strain in the y-direction (shear flows), the
following length scale, L, can be defined (3/16 is an arbitrary scaling factor [33]):
Z∞
3 ¡ ¢
kL = Rii xE, r y dr y (1)
16
−∞
where k is the turbulent kinetic energy. In this equation, Rii (Ex, r y ) is the sum of
the diagonal of the two-point correlation tensor, Rij, measured at a location xE with
two probes at distance r y measuring the three fluctuating velocity components, ui , as
shown in Fig. 1. The correlation tensor Rij is then defined as:
¡ ¢
Rij = ui (Ex) u j xE + r y (2)
For Rii the summation convention is applied. The overbar represents a time (or
ensemble) average. It is intuitively clear that the correlation function has a maximum
at r y = 0 and decays to zero for large r y , as sketched in Fig. 2:
From Eq. 1 it can be seen that the integral length scale, L, is proportional to the
surface area under the correlation curve shown in Fig. 2 divided by the turbulent
kinetic energy, k. The two-point correlation allows therefore an exact definition of an
integral length scale, L. Based on this definition, Rotta derived a transport equation
for the quantity 9 = kL for shear flows. The derivation is essentially achieved by
multiplying the transport equation for the fluctuating velocity at the fixed probe
Flow Turbulence Combust (2010) 85:113–138 117
location with the fluctuating velocity at the traversing probe location and vice versa.
After adding, integration over r y and (ensemble) averaging the equation reads:
Z∞ " ¡ ¢ #
∂9 ∂9 3 ∂U xE + r y ∂U (Ex)
+ Uj + − Rii dr y =
∂t ∂xj 16 ∂x ∂x
−∞
| {z }
Convection
Z∞ Z∞ ¡ ¢
3 ∂U (Ex) 3 ∂U xE + r y
− R21 dr y − R12 dr y
16 ∂ y 16 ∂y
−∞ −∞
| {z }
Production
Z∞ Z∞ 2
3 ∂ ¡ ¢ 3 ∂ Rii
+ R(ik)i − Ri(ik) dr y + ν dr y
16 ∂rk 8 ∂rk ∂rk
−∞ −∞
| {z }
Destruction
Z∞ · ¸
∂ 3 1¡ ′ ′ ¢ ∂9
− R(i2)i + p v + v ′ p′ dr y − ν (3)
∂ y 16 ρ ∂y
−∞
| {z }
Di f f usion
The first index in the correlations refers to the fixed probe location (grouping as
indicated by parenthesis for three indices). The mean shear direction in this equation
is aligned with the y-coordinate. The terms in the equation can be interpreted as
convection, production, destruction and diffusion in the grouping as they appear
above. U refers to the (time or ensemble) averaged velocity in x-direction.
Using this equation as a starting point, Rotta models the unknown correlations on
a term-by-term basis. The first integral term on the left hand side can be neglected
as it involves derivatives in x-direction, which are small compared to y-derivatives in
shear layers. Therefore, this term is small compared to the integrals on the right hand
side. The most interesting term in Eq. 3 is the second production term on the right
hand side:
Z∞ ¡ ¢
3 ∂U xE + r y
− R12 dr y (4)
16 ∂y
−∞
containing the mean velocity gradient at the location of the second probe. For
modelling this term, Rotta expands the velocity gradient into a Taylor series:
¡ ¢
∂U xE + r y ∂U (Ex) ∂ 2 U (Ex) 1 ∂ 3 U (Ex) 2
= + ry + r + ... (5)
∂y ∂y ∂y 2 2 ∂ y3 y
Z∞ ¡ ¢
∂U xE + r y
R12 dr y →
∂y
−∞
Z∞ Z∞ Z∞
∂U (Ex) ∂ 2 U (Ex) 1 ∂ 3 U (Ex)
R12 dr y + R12 r y dr y + R12 r2y dr y + ... (6)
∂y ∂ y2 2 ∂ y3
−∞ −∞ −∞
The first term on the right hand side of this equation can be added to the first
production term in Eq. 3. For the remaining terms, Rotta makes the assumption
that the term with the second derivative of the mean velocity is negligible, leaving
the term with the third derivative as the main additional contribution to the integral
(higher order terms neglected). This estimate is based on the observation that in
homogenous turbulence the function R12 is symmetric with respect to r y (see Fig. 2).
The product of R12 r y is therefore asymmetric and the integral becomes zero as the
contributions from −r y balance the contributions from +r y .
Finally, Rotta introduces the following definitions:
Z∞
3
L12,1 = (R12 + R21 ) dr y
16u′1 u′2
−∞
1/n
Z∞
3
L12,n = R12 rn−1
y dr y
, (n > 1) (7)
16 (n − 1)!u′1 u′2
−∞
In Eq. 8 the underlying assumption is that all lengths-scales based on the integrals in
Eq. 7 are proportional to one another—a logical consequence of using a single scale
equation.
The main destruction term is modelled using dimensional arguments as:
Z∞
3 ∂ ¡ ¢
− R(ik)i − Ri(ik) dr y = ζ̃3 · k3/2 (9)
16 ∂rk
−∞
Using a gradient diffusion model [33], the final two-equation model can be written in
the following boundary-layer form (y is a coordinate across a shear layer):
3/2 µ ¶
∂k ∂k 3/4 k ∂ νt ∂k
+ Uj = Pk − cµ +
∂t ∂xj L ∂ y σk ∂ y
µ ¶ · ¸
∂9 ∂9 ∂U ∂ 3U ∂ νt ∂9
+ Uj = −u′ v ′ ζ̂1 L + ζ̂2 L3 3 − ζ̂3 · k3/2 +
∂t ∂xj ∂y ∂y ∂ y σ9 ∂ y
9 ∂U i
9 = kL; νt = cµ1/4 √ ; Pk = −ui′ u′j ; (10)
k ∂xj
1/4
A factor of cµ in the definition of the eddy viscosity corresponds to the selected
definition of the length scale L, which returns L = κ y in the logarithmic part of a
near-wall boundary layer.
The constants in the Rotta [33] model are not fully specified but can be estimated
as follows: ζ̂1 ≈ 1.2 based on correlation measurements from Rose (see Rotta, [33])
in a homogenous shear flow. Rotta further estimates that ζ̂3 ≈ 0.11−0.13 which
covers the range of plausible Loitsianskii coefficients σ ∼ 2−4 for decaying isotropic
turbulence. Assuming further a diffusion coefficient of σ9 ≈ 1.0 the value of ζ̂2
can be calculated from the logarithmic layer requirements (dU/dy = uτ /κ y; k =
√
cµ u2τ ; L = κ y; νt = uτ κ y):
à !
1 1 1 2 1
ζ̂2 = − 2 ζ̂1 − ζ̂3 3/4 + κ (11)
2κ cµ σ9 cµ1/2
Obviously the k–ε model does not have a term corresponding to ζ̂2 —the logarithmic
layer can still be satisfied through the inclusion of the ε-diffusion term.
Table 1 Model constants for k–kL model resulting from Rotta’s and from the k–ε model
ζ̂1 ζ̂2 ζ̂3
Rotta 1.20 (−3.24)–(−2.88) 0.11–0.13
From k–ε 1.06 – 0.095
120 Flow Turbulence Combust (2010) 85:113–138
The principle difference of the kL-equation to other scale equations is the ap-
pearance of the third derivative of the velocity. All other terms are equivalent to
corresponding terms in the ω- or the ε-equation. This additional term is a result of
the integral given by Eq. 4 and the Taylor series expansion given in Eqs. 5–6. The
main assumption made by Rotta is:
Z∞
∂ 2 U (Ex)
R12 r y dr y ≈ 0 (13)
∂ y2
−∞
is positive in a logarithmic layer (R12 and u′ v ′ being mainly negative and the third
derivative being positive). It therefore acts as a source term, while a sink term would
be required in order to satisfy the logarithmic law of the wall. Based on the derivation
of the term, it is not easily physically justified to introduce a negative coefficient ζ̂2 .
The final reason why the third derivative term should be avoided is that it is difficult
to compute in a general purpose CFD code and that it is most likely erratic in a
three-dimensional flow field.
However, without any higher derivative term, the model does not allow satisfying
the logarithmic law and would in fact be inferior to models using the standard length
scale equations.
On closer inspection, the argument that the integral in Eq. 13 should be assumed
zero is inconsistent. While it is true that the function R12 is symmetric in homoge-
nous turbulence, the entire term (13) would be zero under those conditions, as
homogenous
± turbulence can only exist in a zero or constant shear environment where
∂ 2 U ∂ y2 = 0. In other words, the term is an inhomogeneous term by its very nature.
Menter and Egorov [18] argue therefore that the second derivative term should
be kept as the leading order term in the equations, instead of the third derivative
term. Interestingly, this term also provides the correct sign in the logarithmic layer.
The heuristic correctness of this argument can best be explained for the flow in
the logarithmic region of the law-of-the-wall. In this situation, sketched in Fig. 3,
Probe 1 is fixed and Probe 2 is shifted by ±r y . As the size of the large turbulent
Flow Turbulence Combust (2010) 85:113–138 121
Fig. 3 Two-point correlation in inhomogeneous flow in a logarithmic layer with a linear variation
of L
eddies increases linearly with distance from the wall like L = κ y, it is clear that
the configuration in Fig. 3 will result in a smaller correlation for −r y and a larger
correlation for +r y . Therefore, the correlation measured in such flows is asymmetric
with respect to ±r y . As a result, both the integral and the second derivative of Eq. 13
are non-zero.
In early papers on the new model (e.g. Menter and Egorov [18]), a linear depen-
dency of the term in Eq. 13 on the second velocity derivative was proposed, whereas
in the latest version a quadratic formulation is used (Menter et al. [22]). This is based
R∞
on the observation that the integral itself R12 r y dr y should be zero for homogenous
−∞
turbulence due to the symmetry of the two-point correlations in homogenous flows
(Rotta’s argument). In other words, this term should be proportional to a quantity
which is zero under homogenous conditions. Here we assume that this term itself
is also proportional to the second derivative of the velocity (used as an indicator of
inhomogeneity).
Z∞ Ã ± !
2 1 ∂ 2 U ∂ y2
R12 r y dr y = −const · u′ v ′ L ± ·L (15)
κ ∂U ∂ y
−∞
with a positive constant as indicated by the logarithmic layer. The entire term reads
therefore:
Z∞ µ ¶2
3 ∂ 2 U (Ex) 9 L
− R12 r y dr y = −const · Pk · (16)
16 ∂ y2 k LvK
−∞
122 Flow Turbulence Combust (2010) 85:113–138
using −u′ v′ ∂U/∂ y = Pk and with the von Karman length scale, LvK ,
¯ ¯
¯ ∂U ±∂ y ¯
¯ ¯
LvK = κ ¯ 2 ± 2¯ (17)
¯∂ U ∂y ¯
The absolute value in the definition of Lvk is not required in the derivation of the
model, as the square of Lvk , is finally used, but is introduced to ensure a consistent
positive definition of Lvk .
All other terms in the 9-equation are modelled like in Rotta’s model, yielding the
following equation:
à µ ¶ ! · ¸
∂9 ∂9 9 L 2 ∂ νt ∂9
+ Uj = Pk ζ̃1 − ζ̃2 − ζ̃3 · k3/2 + (18)
∂t ∂xj k LvK ∂ y σ9 ∂ y
In order to distinguish this new 9 equation from the original Rotta (Eq. 10), the
model parameters are written here with a tilde overbar ζ̃i rather than with a caret
overbar ζ̂i .
While this equation represents a suitable scale equation, which could serve as
a basis for calibration, for practical reasons it was decided in Menter et al. [22]
to introduce a further step√ in the derivation and transform the equation to one
with a new variable 8 = kL. This variable has the advantage of being directly
proportional to the eddy viscosity and therefore allows formulating a one-equation
model in addition to the proposed two-equation model (see Menter et al. [22]). A
simple transformation of variables leads to the final two-equation model, written in
the full three-dimensional form (omitting overbars from now on):
¡ ¢ µ ¶
∂ (ρk) ∂ ρU jk 3/4 k2 ∂ µt ∂k
+ = Pk − cµ · ρ +
∂t ∂xj 8 ∂ x j σk ∂ x j
¡ ¢ Ã µ ¶ ! µ ¶
∂ (ρ8) ∂ ρU j8 8 L 2 ∂ µt ∂8
+ = Pk ζ1 − ζ2 − ζ3 · ρk +
∂t ∂xj k LvK ∂ x j σ8 ∂ x j
µt = cµ1/4 ρ8;
¯ ′¯ s
¯U ¯ ′′ ∂ 2Ui ∂ 2Ui
LvK = κ ¯¯ ′′ ¯¯ ; U = ;
U ∂ x2k ∂ x2j
q µ ¶
1 ∂U i ∂U j
U′ = S = 2 · Sij Sij; Sij = + (19)
2 ∂xj ∂ xi
Table 2 Model constants for k–8 model (current version and transformation from Rotta)
ζ1 ζ2 ζ3 σk σ8 cµ κ
From k–9 0.7 0.0278–0.0478 1 1 0.09 0.41
k–8 [22] 0.8 1.47 0.0288 2/3 2/3 0.09 0.41
The relationship between the constants of the new model and Rotta’s model is:
1 κ2 1
ζ1 = ζ̂1 − ; ζ2 = ζ1 − c−3/4
µ ζ3 + c−1/2
µ ; ζ3 = ζ̂3 − cµ3/4 ; σ8 = σ9 (20)
2 σ8 2
where ζ 2 results from the logarithmic-layer relationship.
Parameter values, resulting from a direct transformation of the constants from
Rotta’s model, are presented in the first row of Table 2 (note that ζ 2 cannot be
obtained from the transformation, due to the difference between the second and
third derivative). The second row of the constants are those proposed in Menter
et al. [22] after some optimization for a range of boundary layer and free shear flows.
It should be noted that the calibration of the constants is not necessarily final, but is
also not critical for the purpose of the present article concentrating on the unsteady
behaviour of the model. In Menter et al. [22] there are more details concerning the
integration of the model through the viscous sublayer and the relationship between
a one-
√ and¡ two-equation formulation.
¢ The terminology “KSKL model” results from
k − kL K-Square-root K L .
The emphasis of the current article is on the unsteady characteristics of the
KSKL model. However, in order to demonstrate that the model is a suitable
RANS formulation, Fig. 4 shows a comparison of velocity profiles for steady state
computations around the NACA 4412 airfoil of Coles and Wadcock [2] at an angle of
attack α = 13.9◦ and Re = 1.5·106 . The figure shows a comparison of the KSKL (two-
equation model), SKL (Square-root-K L) (one-equation model) as given in Menter
et al. [22], the SST model [14] and the Spalart and Allmaras [35] model. It can be
seen that the optimized version of the KSKL (and the related SKL) model as given
by Menter et al [22] are competitive against more established one- and two-equation
turbulence models. In Menter et al. [22] other steady cases have been computed
successfully (e.g. flat plate zero pressure gradient boundary layer, axi-symmetric
diffuser flow, backward facing step and impinging jet with heat transfer).
The exact transport equation underlying the current model indicates that higher
derivatives of the velocity should appear in the length-scale equation. It is interesting
to go back to Rotta’s interpretation of the influence of such terms. In his book [33]
Rotta states (translated from German): “Neglecting the convective and dif fusive terms
and replacing c·k3/2 with the help of the equation for the turbulent kinetic energy, one
obtains the following interesting simplif ication:
∂U/∂ y
L2 = 2κ 2 (21)
∂ 3 U/∂ y3
This relation has a remarkable similarity to the formula:
¯ ¯
¯ ∂U/∂ y ¯
L=κ¯ 2 ¯ ¯ (22)
∂ U/∂ y2 ¯
which von Karman derived based on similarity arguments”.
This statement indicates that Rotta viewed it as one of the main characteristics
of his model that it provided a length scale in the source terms, which is missing in
other models. In other words, forming a source term equilibrium using k–ε or k–ω
type models does not allow the determination of a length L (or related) scale. This
can best be seen from the standard k–ω model:
¡ ¢ µ ¶
∂ (ρk) ∂ ρU jk k¡ 2 ¢ ∂ µt ∂k
+ =ρ S − cµ ω 2 +
∂t ∂xj ω ∂ x j σk ∂ x j
¡ ¢ µ ¶
∂ (ρω) ∂ ρU jω ¡ ¢ ∂ µt ∂k
+ = ρ cω1 S2 − cω2 ω2 + (23)
∂t ∂xj ∂ x j σω ∂ x j
When considering the turbulence model source terms of classical two-equation
models as a black box, the only variable entering from the outside is the shear strain
rate S ∼ 1/T. The source terms can therefore only help determining one turbulence
scale—in this case the turbulent frequency ω ∼ (1/T) ∼ S. The second scale is not
defined from the source terms alone.
Despite the widespread usage of two-equation models, their mechanism of de-
termining the turbulent length scale is often not appreciated. It results from the
inclusion of the diffusion terms into the estimate as follows (for a generic variable
2):
µ ¶
∂ νt ∂2 2
∝ νt 2 (24)
∂y σ ∂y δ
where δ is the thickness of the turbulent layer (shear layer thickness). Introducing
this estimate into standard two-equation models gives:
L∼δ (25)
In other words, the maximum of the turbulent length scale, L, in a shear layer
calculated by a standard two-equation model will always be proportional to the
thickness of the turbulent layer due to the action of the diffusion terms. This is
the appropriate scaling for steady shear flows, but for unsteady flows it damps out the
formation of resolved scales, resulting in the classical URANS behaviour as discussed
Flow Turbulence Combust (2010) 85:113–138 125
above. Models using higher derivatives in the source terms behave differently—they
allow the detection of resolved turbulent scales from the source terms of the model.
In the case of the KSKL model, the source term equilibrium results in a turbulent
length scale, L, independent of the diffusion terms:
v .
u
u ζ1 − ζ3 c3/4
µ
t
L= LvK (26)
ζ2
What is the physical meaning of the additional source term, which contains LvK ?
Obviously any spatial variation in the strain rate (meaning a non-zero second
derivative) reduces the effectiveness of the production term in the Φ equation. Under
constant shear (homogenous conditions) the turnover frequencies ω of two small
eddies are the same independent of their location, as they are driven by the same
constant strain rate, S(ω ∼ S). They can therefore merge into one larger vortex with
the same turbulent frequency. This corresponds to the actual situation in constant
shear flows—the turnover frequency, ω, of turbulence is proportional to the strain
rate, S, while the length scale grows to infinity. Under non-homogenous conditions
(strain rate not spatially constant), the individual vortices have locally different
turnover frequencies, ω, proportional to the local strain rate, S(y). From a certain
size on, they can therefore not merge into a larger vortex, as one vortex cannot have
two different frequencies. This results in a finite vortex size depending on the local
strain rate and its spatial variation. The spatial variation is given to first order by the
von Karman length scale. This is the physical rationale why the von Karman length
scale should appear in the length scale equation of RANS models.
Another interesting situation arises in the following generic experiment where it
is assumed that a 1-D mean flow profile of the form U(y) = U 0 × sin(2π y/Lm ) could
be established by external (e.g. magnetic) forces as shown in Fig. 5. This mean flow
has a characteristic length Lm . Under such conditions, one would assume that the
turbulence model would provide a modelled scale which is also L ∼ Lm as only scales
Fig. 5 Generic 1D test. Left: Frozen 1D mean flow U(y) = U0 × sin(2π y/Lm) used to compute
steady turbulence models solutions. Right: Turbulent length scale returned by different models. Walls
applied at y/d = ±1 (thick lines) and y/d = ±2 (thin lines). Comparison of SST and SST–SAS model
in 1D code
126 Flow Turbulence Combust (2010) 85:113–138
Fig. 6 Circular cylinder in a cross flow at Re = 3.6·106 , Left: SST–URANS, Right: KSKL model.
Iso-surface of Q = S2 − Ä2 , coloured according to the eddy viscosity ratio µt /µ (smaller by factor 14
in right f igure)
Flow Turbulence Combust (2010) 85:113–138 127
in the SST or other classical URANS models. The effect by which this is achieved
is also visible from Fig. 6 where the colour on the iso-surface shows that the eddy
viscosity for the SST model is around an order of magnitude larger than for the SAS
model, as LvK adjusts to the smallest scales of the spectrum.
A standard test for the unsteady characteristics of the model is Decaying Isotropic
Turbulence (DIT). The interest is not in running DIT in RANS mode, meaning
simply specifying values for k and Φ and following their decay as given by their corre-
sponding transport equations. Instead, the DIT case is run in a mode typically applied
in LES, meaning a synthetically generated velocity field is initially specified and its
decay is computed from the three-dimensional unsteady momentum equations (the
initial field is computed by a widely used computer program from the group of Prof.
Strelets in St. Petersburg). In order to obtain initial conditions for k and Φ, the KSKL
model equations are solved on the frozen initial velocity field (it is again interesting to
point out that standard two-equation models would not provide converged solutions
under such conditions, as the flow is computed with periodic boundary conditions,
not providing a finite layer thickness). Using these initial conditions, the KSKL
model is then run coupled with the momentum equations. In addition, solutions with
the LES–WALE model and the standard k − ε model are also computed (for k − ε
the transformed initial conditions of the KSKL model are used). Simulations are
computed on a 323 grid using a second order central scheme for the convective fluxes
and a time step corresponding to CF L ∼ 0.5.
Figure 7 shows the turbulent spectrum for the Comte-Bellot experiment (Comte-
Bellot and Corrsin [3]) in comparison with the experimental data after t = 2 non-
dimensional times. The most interesting result is that the KSKL model allows the
formation of a turbulent spectrum, but does not provide sufficient damping at the
high-wave number limit to dissipate the energy at the smallest scales (KSKL no
HWN damping in Fig. 7). This is in contradiction to the expectations for classical
URANS models and the central aspect of the new model. The k − ε model behaves
as expected and damps out the small scales quickly, while the WALE model returns
its calibrated behaviour in agreement with the experimental spectrum.
Considering the above arguments, the behaviour of the KSKL model is actually
not all that surprising, as the von Karman scale adjusts to the smallest scales and
thereby produces an eddy viscosity small enough to allow the formation of even
smaller eddies until the grid limit is reached. At that point, no smaller eddies can
form. However, consistently, the partial differential equations of the SAS model,
having no information on the cut-off limit, again provide an eddy viscosity small
enough to allow further cascading to smaller scales. As this is not possible due to
the resolution limit, the energy accumulates at the high wave number limit. The
behaviour of insufficient damping can easily be remedied, as will be shown later.
This is a good point for reconsidering the situation. First a new turbulence model
based on RANS arguments has been derived. It was found that this model can be
calibrated to produce steady RANS solutions for numerous flows like the airfoil
flow shown above, or other flows computed by Menter et al. [22]. However, under
certain conditions, the model behaves in contradiction to the expectations concerning
RANS models and allows the formation of a turbulent spectrum down to the grid
limit. As RANS is based on averaging out random fluctuations, this points to a
conceptual problem in the definition of RANS (and more so classical URANS)
models. While classical RANS models like k − ε do not show this discrepancy, they
do however instead produce a disturbing and unphysical behaviour for the generic
flows discussed above. Especially for the generic 1-D mean flow U(y) = U 0 × sin(λy)
it is unphysical that standard models return L → ∞ for t → ∞. In other words,
both types of models produce results which are intuitively in contradiction with
expectations. The choice is between models which either do not recognize resolved
scales at all and thereby allow length scales larger than the driving flow scale, or
models which do adjust to resolved structures and as a result allow the formation
of a turbulent spectrum as long as the numerical method and the grid and time step
allows it.
Leaving aside the theoretical considerations and turning to more practical aspects,
the new model offers an interesting passage into scale-resolving simulations. Given
that a sufficiently strong instability is present, the model allows the formation of a
turbulent spectrum, similar to a LES model. At the same time, stable flow regions will
still be covered in RANS mode. This indicates a behaviour similar to the Detached
Eddy Simulation (DES) concept proposed by Spalart et al. [35, 36], but on an entirely
different basis and using different modelling mechanisms. Some practical results for
unsteady simulations using the KSKL model will be shown below and in Egorov
et al. [7].
It can only be speculated which impact Rotta’s assumption concerning the second
term in the Taylor series expansion had on turbulence modelling. Clearly, the model
with the second derivative satisfies the logarithmic layer without additional terms.
Furthermore, the KSKL model avoids one of the main problems of the k–ε models,
namely the difficulties in formulating extensions to integrate the equations through
the viscous sublayer (‘low-Re’ extensions) as shown in Menter et al. [22]. It is
therefore not unlikely that the k–kL model would have been the RANS model
of choice for engineering flows (see Rodi [29]). This would also have significantly
reduced the difficulties in computing adverse pressure gradient and separating flows
as shown in Menter et al. [22]. Even more interestingly, the model, using the second
derivative, exhibits an entirely different behaviour for unsteady flow simulations
compared to classical RANS models, which is advantageous to many engineering
Flow Turbulence Combust (2010) 85:113–138 129
flows. This behaviour could have benefited unsteady simulations for several decades,
offering an attractive alternative to Large Eddy Simulation (LES) early on. Finally,
the availability of SAS at an earlier stage would naturally have had a strong impact
on hybrid RANS/LES model formulations. As will be shown below, SAS models
can cover a significant portion of flows for which DES models have initially been
developed.
The tests for the DIT case have shown that the KSKL model (and other SAS models)
does not provide sufficient damping of the smallest scales at the grid limit and for a
small time step (CFL < 1). This is not a difficult practical problem, as classical LES
technologies can be applied for achieving the required dissipation of energy. The
simplest and most pragmatic way of achieving this goal has turned out to be the
enforcement of a lower limit on the eddy viscosity coming from the SAS model. The
lower limit should however not impact steady state RANS solutions. This can be
achieved by using the WALE–LES model [26] for this purpose:
¡ ¢
νt = cµ1/4 8 → νt = max cµ1/4 8, νtW ALE (27)
where 1 is the grid spacing based on the third root of the control volume, Cw ∼ 0.5
and Gija is the traceless part of Gij = 1/2(gik gkj + g jk gki ) and gij = ∂U i /∂ x j.
The WALE model was selected as a HWN damping limit, as it gives νtW ALE = 0
for steady shear flows, ensuring that the RANS model remains unaffected even for
relatively coarse grids. This would not be the case with a simpler Smagorinsky type
model. Figure 7 shows the turbulent spectrum for the DIT case computed without
and with the HWN limiter. As the HWN limiter is reached by the SAS model, these
results are identical to the WALE simulation.
During the development of SAS models, different approaches have been utilized
for achieving HWN damping. In Menter and Egorov [21] a numerical discretization
of LvK was presented which ensured a sufficient level of damping at the HWN limit.
While this approach works well, it was linked to a specific numerical formulation,
not easily generalized. In Menter and Egorov [23] an alternative HWN damping
was therefore given for the SST–SAS model. It ensures that Lvk > c1 and allows
the calibration of the HWM limit through a constant c. This limiter is efficient and
currently employed in ANSYS–CFX and ANSYS–Fluent for the SST–SAS model.
However, in its proposed form it works only for the SST–SAS model, but would
impact the RANS–KSKL formulation for very coarse grids. Equation 27 is therefore
a more generally applicable formulation. It should also be noted that one could also
achieve HWN damping through numerical dissipation using the discretization of the
convective scheme, as in implicit LES methods.
130 Flow Turbulence Combust (2010) 85:113–138
The KSKL model can be transformed to other variables, which can help introducing
the unsteady characteristics of this model into existing two-equation models. The
high Reynolds number equations for k–ε and k–ω are (assuming σk = σ8 ):
à µ ¶ ! à !
∂ρε ∂ρU jε ε L 2 ε2 ζ3
+ = Pk 2 − ζ1 + ζ2 −ρ 2 − 3/4
∂t ∂xj k LvK k cµ
µ ¶ µ 4 ¶
∂ µt ∂ε 2cµ ρ k ∂ ³ ´
ε ∂ ³ ε ´
+ − 2
d (29)
∂ x j σ8 ∂ x j σ8 ε ∂xj k ∂xj k
à µ ¶ !
∂ρω ∂ρU jω ω L 2 ¡ ¢
+ = Pk 1 − ζ1 + ζ2 − ρω2 cµ − cµ1/4 ζ3
∂t ∂xj k LvK
µ ¶ µ ¶
∂ µt ∂ω 2ρ 1 ∂k ∂ω k ∂ω ∂ω
+ + − 2 (30)
∂ x j σ8 ∂ x j σ8 ω ∂ x j ∂ x j ω ∂xj ∂xj
To complete the model derivation, the additional terms have also been introduced
into the widely used SST model (Menter [14]). The modification of the SST model is
based on the transformation given by Eq. 30. The two last terms of this equation are
introduced into the ω-equation of the SST model. For details see Egorov and Menter
[6]. The formulation results in an additional term in the ω-equation of the SST model
(Q SAS added to the right hand side of the SST ω-equation):
" µ ¶ µ ¶ #
2 L 2 2ρk 1 ∂k ∂k 1 ∂ω ∂ω
Q SAS = max ρζ2 S − C SAS max , , 0 (31)
LvK σ8 k2 ∂ x j ∂ x j ω 2 ∂ x j ∂ x j
with C SAS = 2. The max function and the k-derivative term have been introduced
to avoid any change to the RANS performance of the SST model for boundary
layer flows. In unsteady situations, the term including the von Karman length scale is
dominating the other terms, resulting in the full activation of the SAS functionality.
The resulting model is termed SST–SAS model.
Figure 8 shows a visual example of scale-adaptivity for the periodic hill flow testcase
as described by e.g. Fröhlich et al. [8]. The Reynolds number is Re = U B h/ν = 104
(U B —bulk velocity, h—height of hill). The three pictures show a visualization of the
Q-criterion for three simulations of this flow. The left picture was computed with a
time step of ∆t = 0.045U B / h which corresponds to an LES time resolution (CF L ∼
1). The middle and the right pictures where obtained using a factor 2 and 4 larger ∆t
on the same numerical mesh (2.5 · 106 nodes). The colour of the figures depicts the
ratio of eddy viscosity to molecular viscosity. The larger time steps do not allow the
same spatial resolution as the small ones, resulting in significantly larger turbulent
structures. The important aspect is that the eddy viscosity adjusts accordingly and
increases from left to right. It thereby compensates for the non-resolved portion of
the spectrum. Further increasing the time step will result in a steady RANS solution.
It is an essential feature of the SAS concept that a continuous variation of solutions
from LES-type to RANS type can be achieved based on the time step selected.
Figure 9 shows the time-averaged velocity profiles for the different simulations
compared to a steady-state RANS solution using the SST model. For SAS, only the
smallest and largest ∆t results are shown for clarity. Even for the largest ∆t, there is
a substantial improvement compared to RANS.
It is interesting to consider the behaviour of a classical Smagorinsky LES model
vt = (c1)2 S and other LES models for this situation. As the grid is the same in all
three simulations, and as the strain rate is lower for large scales than for small scales,
one would obtain actually the opposite behaviour to SAS. The Smagorinsky model
would predict a lower eddy viscosity for the larger structures and is therefore not
scale-adaptive. This behaviour would also carry over to DES methods, as they scale
like the Smagorinsky model when running in LES mode.
Fig. 8 Turbulent structures for periodic hill flow. Iso-surface for Q = S2 − Ä2 (S—Strain Rate, Ä—
Vorticity). Colour gives the eddy viscosity ratio µt /µ (it increases from left to right by factor ∼10).
Left: 1t = 0.045U B /h, middle: 1t = 0.09U B /h, right: 1t = 0.18U B /h
132 Flow Turbulence Combust (2010) 85:113–138
It is this robust behaviour with respect to space and time resolution which makes
the SAS model an attractive concept for engineering simulations. In many technical
flows, a LES grid and a LES time step (CFL ∼ 1) cannot be maintained in the
entire unsteady flow domain. The SAS model will however always have a fall-
back URANS or RANS solution if the resolution is not sufficient for resolving the
turbulent scales. In contrast, LES and DES models can return incorrect results and
potentially numerical instabilities if the numerical grid is too coarse (insufficient for
LES) or the time-step is too large (substantially larger than CF L ∼ 1). An example
of such a SST–DES simulation is shown in Fig. 10, where a backward facing step
flow is computed on a grid with insufficient resolution in the spanwise direction (12
nodes over a spanwise extent of 2h − h being the step height). The DES switch then
already affects the RANS model, but the flow instability is too weak to generate a
sufficient number of unsteady structures on the under-resolved grid. This results in a
mean reattachment point of xr ∼ 15h, contrary to an experimental value of xr ∼ 6h.
The SST–SAS model produces a steady RANS solution under such conditions with
the correct reattachment length (using the same small time step of CFL ∼ 1 as in
Fig. 10 SST–DES solution for backward facing step on spanwise under-resolved grid (contour
of instantaneous velocity). Time averaged reattachment length xr ∼ 15 h. (Picture courtesy BMW
AG/Dr. Sohm)
Flow Turbulence Combust (2010) 85:113–138 133
Fig. 11 Numerical grid for x–y plane of flow around triangular cylinder (mean flow in x-direction)
SST–DES). It has to be noted that the unsteady DES simulation does recover the
correct mean value of xr under grid refinement, (∼25 nodes required in spanwise
direction). However, such a consistent grid refinement is non-trivial for complex
industrial flows.
The limitation of SAS is that it will not switch into scale-resolving mode if the
flow is not sufficiently unstable, as discussed for the backward facing step flow.
In this case, unsteady behaviour cannot be achieved with the SAS model. In most
cases, this poses no accuracy problem, as the RANS mode of the model typically
produces reasonably accurate solutions under such conditions. However, if unsteady
information is required, e.g. for acoustics simulations, such unsteady characteristics
cannot be obtained without additional provisions. Examples of such cases have
been reported by Davidson [5]. The cases studied there are all dominated by the
turbulence coming from inside the upstream boundary layers and are therefore not
suitable for unsteady SAS model simulations without an explicit conversion of energy
to the resolved part (e.g. forcing). One possibility for such energy transfer is the
usage of forcing terms as described in Menter et al. [24], or the definition of an
interface, where modelled energy is explicitly transferred to resolved energy through
the generation of synthetic turbulence.
The flow around a triangular cylinder (experiment of Sjunesson et al., [38])
constitutes an example of a suitable case for SAS model simulations. Figure 11 shows
the 2-D section of the grid used. It is extended in the third direction to cover six
Fig. 12 Flow structures for flow around triangular cylinder. Left: SST–SAS model, right: SST–DES
model
134 Flow Turbulence Combust (2010) 85:113–138
times the edge length of the triangle with 81 cells in that direction. The overall grid
has 1.8·106 hexahedral cells. The Reynolds number based on freestream velocity and
edge length is 45,500 with an inlet velocity of 17.3 m/s. Periodic boundary conditions
are applied in spanwise direction. The simulations where run with ANSYS-Fluent
using the BCD (bounded central difference) advection scheme and a time step of
∆t = 10−5 s (CFL ∼ 1 behind cylinder).
This test case was computed with the SST–SAS and the SST–DES model.
Figure 12 shows the turbulent structures for both models using Q = 106 1/s2 . There
is clearly very little difference between the two pictures indicating that both models
are operating in LES mode. The grid shown in Fig. 11 does not resolve the boundary
layers on the walls of the cylinder, as they are not relevant for this flow. It is important
to point out though that both models would automatically cover the boundary layers
in RANS mode if the grid was refined in wall normal direction. It should also be
noted that the flow was computed with steady boundary conditions for both models.
The global instability of this flow is sufficient for triggering the unsteady structures.
Figure 13 shows the time-averaged axial velocity along the centreline behind the
cylinder in comparison with the experimental data. The agreement of both models
with the data is very close. The same is true for the comparison of the velocity profiles
and turbulence profiles at three different stations shown in Fig. 14. The agreement
between the SST–SAS model, the SST–DES model and the experiment is basically
within the experimental uncertainties (considering that the averaged flowfield should
be symmetric).
In engineering flows there are frequently flow regions of the nature represented by
the above test case—e.g. large separation zones past bluff bodies, however, typically
embedded within larger stable flow zones. These separated areas are not predicted
Fig. 14 Velocity profiles and turbulence RMS profiles for three different stations downstream of the◮
triangular cylinder (x/a = 0.375, x/a = 1.53, x/a = 3.75). Comparison of SST–SAS, SST–DES models
and experiment. (a U—velocity, b urms , c vrms , d u’v’)
Flow Turbulence Combust (2010) 85:113–138 135
136 Flow Turbulence Combust (2010) 85:113–138
well by steady RANS methods. With SAS models, these areas are automatically
“detected” and resolved down to the available grid and time step resolution limit. If
the resolution is low, the model remains in RANS (or URANS) mode, making SAS
a safe pathway into scale-resolving engineering flow simulations. In a companion
article [7], numerous more complex and industrially relevant test cases will be
shown to demonstrate the SAS models potential for the simulation of engineering
flows.
5 Summary
Acknowledgements The current work was partially supported by the EU within the research
projects DESIDER (Detached Eddy Simulation for Industrial Aerodynamics) under contract
No. AST3-CT-200-502842 (http://cfd.mace.manchester.ac.uk/desider/). The authors also want to
thank the Editor Prof. W. Rodi for fruitful discussions and suggestions which have helped improving
the article.
Flow Turbulence Combust (2010) 85:113–138 137
References
1. Baldwin, B.S., Barth, T.J.: A one-equation turbulence transport model for aerodynamic flows.
AIAA Paper 92-0439 (1992)
2. Coles, D., Wadcock, A.J.: Flying hot-wire study of flow past an NACA 4412 airfoil at maximum
lift. AIAA J. 17(4), 312–329 (1979)
3. Comte-Bellot, G., Corrsin, S.: Simple Eulerian time correlation of full- and narrow-band velocity
signals in grid-generated, ‘isotropic’ turbulence. J. Fluid Mech. 48(Part 2), 273–337 (1971)
4. Craft, T.J., Launder, B.E., Suga, K.: Development and application of a cubic eddy-viscosity
model of turbulence. Int. J. Heat Fluid Flow 17(2), 108–115 (1996)
5. Davidson, L.: Evaluation of the SST–SAS model: channel flow, asymmetric diffuser and axi-
symmetric hill. In: Proceedings European Conference on Comp. Fluid Dyn. ECCOMAS CFD (2006)
6. Egorov, Y., Menter, F.R.: Development and application of SST–SAS model in the DESIDER
project. In: Advances in Hybrid RANS–LES Modelling. Notes on Num. Fluid Mech. Multidiscip.
Des., vol. 97, Springer (2008)
7. Egorov, Y., Menter, F.R., Cokljat, D.: The scale-adaptive simulation method for unsteady tur-
bulent flow predictions. Part 2: application to complex flows. Flow Turbul. Combust. (2010).
doi:10.1007/s10494-010-9265-4
8. Fröhlich, J., Mellen, C.P., Rodi, W., Temmerman, L., Leschziner, M.: Highly resolved large-eddy
simulation of separated flow in a channel with streamwise periodic constrictions. J. Fluid Mech.
526, 19–66 (2005)
9. Fröhlich, J., von Terzi, D.: Hybrid LES/RANS methods for simulation of turbulent flows. Prog.
Aerosp. Sci. 44(5), 349–377 (2008)
10. Gatski, T.B., Speziale, C.G.: On explicit algebraic stress models for complex turbulent flows. J.
Fluid Mech. 254, 59–78 (1993)
11. Kim, S.E.: Large Eddy simulation using unstructured meshes and dynamic subgrid-scale turbu-
lence models. AIAA Paper no. 2004-2548 (2004)
12. Launder, B.E., Spalding, D.B.: The numerical computation of turbulent flows. Comput. Methods.
Appl. Mech. Eng. 3, 269–289 (1974)
13. Launder, B.E., Reece, G.J., Rodi, W.: Progress in the development of a Reynolds-stress turbu-
lence closure. J. Fluid Mech. 68, 537 (1975)
14. Menter, F.R.: Two-equation eddy-viscosity turbulence models for engineering applications.
AIAA J. 32(8), 1598–1605 (1994)
15. Menter, F.R.: Eddy viscosity transport equations and their relation to the k–ε model. NASA-
TM-108854 (1994)
16. Menter, F.R.: Eddy viscosity transport equations and their relation to the k–ε model. J. Fluids
Eng. 119, 876–884 (1997)
17. Menter, F.R, Kuntz, M., Bender R.: A scale-adaptive simulation model for turbulent flow pre-
dictions. AIAA Paper 2003-0767 (2003)
18. Menter, F.R., Egorov, Y.: Re-visiting the turbulent scale equation. In: Proc. IUTAM Symp. One
Hundred Years of Boundary Layer Research. Springer, Göttingen (2004)
19. Menter, F.R., Egorov, Y.: A scale-adaptive simulation model using two-equation models. AIAA
Paper 2005-1095, Reno/NV (2005)
20. Menter, F.R., Egorov, Y.: Turbulence models based on the length-scale equation. In: Fourth
International Symposium on Turbulent Shear Flow Phenomena, Williamsburg, 2005—Paper
TSFP4-268 (2005)
21. Menter, F.R., Egorov, Y.: SAS turbulence modelling of technical flows. In: DLES 6—6th ER-
COFTAC Workshop on Direct and Large Eddy Simulation September, Poitiers (2005) √
22. Menter, F.R., Egorov, Y., Rusch D.: Steady and unsteady flow modelling using the k − kL
model. In: Hanjalic, K., Nagano, Y., Jakirlic, S. (eds.) Proc. Turbulence, Heat and Mass Transfer,
vol. 5 (2006)
23. Menter, F.R., Egorov, Y.: Formulation of the Scale-Adaptive Simulation (SAS) model during
the DESIDER Project. In: Haase, W., Braza, M., Revell, A. (eds.) Notes on Num. Fluid Mech.
and Multidisc. Design, vol. 103, Springer (2009)
24. Menter, F.R., Garbaruk A., Smirnov P.: Scale adaptive simulation with artificial forcing. In: Proc.
3rd Symposium on Hybrid RANS–LES Methods (2009)
25. Moffatt, H.K.: Turbulence and stochastic processes: Kolmogorov’s ideas 50 years on. In: Hunt,
J.C.R., Phillips, O.M., Williams, D. (eds.) Proceedings of the Royal Society, London, A, vol. 434,
1991, pp. 1–240 (1991)
138 Flow Turbulence Combust (2010) 85:113–138
26. Nicoud, F., Ducros, F.: Subgrid-scale stress modelling based on the square of the velocity gradient
tensor. Flow Turbul. Combust. 62, 183–200 (1999)
27. Pope, S.B.: A more general effective-viscosity hypothesis. J. Fluid Mech. 72, 331–340 (1975)
28. Rodi, W.: A new algebraic relation for calculating the Reynolds stresses. Z. Angew. Math. Mech.
56, 219–221 (1976)
29. Rodi, W.: Turbulence modelling for boundary layer calculations. In: Proc. IUTAM Symp. One
Hundred Years of Boundary Layer Research, Göttingen, Springer (2004)
30. Rodi, W., Mansour, N.N.: Low Reynolds number modelling with the aid of direct numerical
simulation data. J. Fluid Mech. 250, 509–529 (1993)
31. Rotta, J.C.: Statistische theorie nicht-homogener turbulenz I und II. Z. Phys. 129, 547–572; 131,
51–77 (1951)
32. Rotta, J.C.: Über eine methode zur Berechnung turbulenter Scherströmungen, aerodynamische
Versuchsanstalt Göttingen. Rep. 69 A14 (1968)
33. Rotta, J.C.: Turbulente Strömumgen. BG Teubner Stuttgart (1972)
34. Sagaut, P., Deck, S., Terracol, M.: Multiscale and Multiresolution Approaches in Turbulence.
Imperial College Press, London (2006)
35. Spalart, P.R., Allmaras, S.R.: A one-equation turbulence model for aerodynamic flows. La
Recherche Aerospatiale n 1, 5–21 (1994)
36. Spalart, P.R.: Strategies for turbulence modelling and simulations. Int. J. Heat Fluid Flow. 21, 2
(2000)
37. Spalart, P.R., Jou, W., Strelets, M., Allmaras, S.: Comments on the feasibility of LES for wings,
and on a hybrid RANS/LES approach. In: Advances in DNS/LES, 1st AFOSR Int. Conf. on
DNS/LES (1997)
38. Sjunnesson, A., Henriksson, R., Lofstrom C.: CARS measurements and visualization of reacting
flows in bluff body stabilized flame. AIAA Paper. 92-3650 (1992)
39. Strelets, M.: Detached Eddy simulation of massively separated flows. AIAA Paper 2001-879
(2001)
40. Temmerman, L., Leschziner, M.A.: Large eddy simulation of separated flow in a streamwise
periodic channel constriction. In: Proceedings, 2nd Symp. on Turbulence and Shear-Flow Phe-
nomena, Stockholm (2001)
41. Tennekes, H., Lumley, J.L.: A First Course in Turbulence. MIT Press, London (1992)
42. Travin, A., Shur, M., Spalart, P.R., Strelets, M.: On URANS solutions with LES-like behaviour.
In: Proc. ECCOMAS 2004, Jyväskylä (2004)
43. Wallin, S., Johansson, A.V.: An explicit algebraic Reynolds stress model for incompressible and
compressible turbulent flows. J. Fluid Mech. 403, 89–132 (2000)
44. Wilcox, D.C.: Turbulence Modeling for CFD. DCW Industries Inc., 2. Edition (1998)