You are on page 1of 13

This article was downloaded by: [Hong Kong Polytechnic University]

On: 7 September 2009


Access details: Access Details: [subscription number 912320008]
Publisher Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK

International Journal of Computational Fluid Dynamics


Publication details, including instructions for authors and subscription information:
http://www.informaworld.com/smpp/title~content=t713455064

Review of the shear-stress transport turbulence model experience from an


industrial perspective
Florian R. Menter a
a
ANSYS Germany GMBH, Otterfing, Germany
Online Publication Date: 01 April 2009

To cite this Article Menter, Florian R.(2009)'Review of the shear-stress transport turbulence model experience from an industrial

perspective',International Journal of Computational Fluid Dynamics,23:4,305 316


To link to this Article: DOI: 10.1080/10618560902773387
URL: http://dx.doi.org/10.1080/10618560902773387

PLEASE SCROLL DOWN FOR ARTICLE


Full terms and conditions of use: http://www.informaworld.com/terms-and-conditions-of-access.pdf
This article may be used for research, teaching and private study purposes. Any substantial or
systematic reproduction, re-distribution, re-selling, loan or sub-licensing, systematic supply or
distribution in any form to anyone is expressly forbidden.
The publisher does not give any warranty express or implied or make any representation that the contents
will be complete or accurate or up to date. The accuracy of any instructions, formulae and drug doses
should be independently verified with primary sources. The publisher shall not be liable for any loss,
actions, claims, proceedings, demand or costs or damages whatsoever or howsoever caused arising directly
or indirectly in connection with or arising out of the use of this material.

International Journal of Computational Fluid Dynamics


Vol. 23, No. 4, AprilMay 2009, 305316

Review of the shear-stress transport turbulence model experience from an industrial perspective
Florian R. Menter*
ANSYS Germany GMBH, Staudenfeldweg 12, Otterng, 83624, Germany

Downloaded By: [Hong Kong Polytechnic University] At: 06:19 7 September 2009

(Received 7 January 2009; nal version received 22 January 2009)


The present author was asked to provide an update on the status and the more recent developments around the
shear-stress transport (SST) turbulence model for this special issue of the journal. The article is therefore not
intended as a comprehensive overview of the status of engineering turbulence modelling in general, nor on the
overall turbulence modelling strategy for ANSYS computational uid dynamics (CFD) in particular. It is clear from
many decades of turbulence modelling that no single model nor even a single modelling approach can solve all
engineering ows. Any successful CFD code will therefore have to oer a wide range of models from simple Eddyviscosity models through second moment closures all the way to the variety of unsteady modelling concepts
currently under development. This article is solely intended to outline the role of the concepts behind the SST model
in current and future CFD simulations of engineering ows.
Keywords: SST turbulence model; engineering ows; SAS; unsteady ows; scale-adaptive simulation; laminarturbulent transition

Introduction
Some 15 years ago, the author proposed a new
turbulence model for aerodynamic simulations, termed
the shear stress transport (SST) model (Menter 1994).
The need for this model arose from the situation of
CFD in external aerodynamics in the early 1990s. At
that time, the increase in computing power allowed for
the rst time the systematic application of CFD to
three-dimensional aerodynamic congurations. The
standard turbulence model used in aeronautics codes
was the Baldwin-Lomax (BL) algebraic model (see
Wilcox 1998). The BL model had already been
generalised from the CebeciSmith model (see Wilcox
1998) for easier use outside boundary layer codes, but
was clearly imposing severe limitations on the geometric complexity as well as on the mesh topologies
that could be handled. In attention, it required nonlocal search algorithms which turned out to be code
specic, making a consistent implementation between
dierent CFD codes problematic. Lastly, new technologies like CFD on unstructured grids and parallel
processing, based on domain decomposition, required
models which avoided non-local operations.
This need could only be served by models based on
transport equations. Although such models had been
developed for many decades, the aeronautics community was reluctant in adopting them for various
reasons. The rst was that the community had early
on followed the philosophy of integrating the

*Email: orian.menter@ansys.com
ISSN 1061-8562 print/ISSN 1029-0257 online
2009 Taylor & Francis
DOI: 10.1080/10618560902773387
http://www.informaworld.com

equations through the viscous sublayer, while most


other industries using CFD adopted less accurate wall
functions. This was a stumbling stone mainly for the
k-e model, which had and has proven notoriously
dicult to integrate to the wall, despite the many lowRe number extensions proposed. The second issue was
the inability of standard models to accurately predict
ow separation and stall characteristics. Signicant
progress in this aspect had been made by Johnson and
King (JK) (Wilcox 1998), in the framework of
algebraic models, and the community was reluctant
in accepting transport models falling behind the high
accuracy that the JK model already provided.
The most widely used alternative to the k-e model
was the k-o model in the formulation developed by
Wilcox (1998). This model oered a much more robust
and accurate viscous sublayer formulation and also
proved more accurate for boundary layers in adverse
pressure gradients. However, the Wilcox model also
had its weaknesses. The rst was a severe sensitivity of
the results to the freestream values specied for o
outside boundary and shear layers (Menter 1992,
Wilcox 1998). This freestream sensitivity introduced a
strong dependency of the solution on relatively
arbitrary values specied for k and o at the inlet, as
well as on their decay upstream of the aerodynamic
device. Furthermore, the model lacked (like all
standard two-equation models) the eect of SST,
which had been demonstrated in its importance for

Downloaded By: [Hong Kong Polytechnic University] At: 06:19 7 September 2009

306

F.R. Menter

separation predictions by Johnson and King (JK) (see


Wilcox 1998). The eects of freestream dependency
have also not been eliminated by the new version of the
k-o model (Wilcox 2007), see Appendix.
The principle idea behind the SST models was to
combine the best elements of the k-e, the k-o and the
JK models. This was achieved by introducing functions
which gradually blended the dierent elements of these
models into a single formulation. Obviously, this is a
pragmatic engineering approach, justied only by an
improved model performance. It is to be explicitly
stressed that the SST model owes much of its success to
the robust and accurate near wall formulation of the
1988 Wilcox k-o model.
The SST model is implemented into most aeronautics, as well as into most general-purpose commercial CFD codes. In other words, the models application
range has expanded signicantly beyond the original
aerodynamic target. It has proven successful, at least in
comparison with other models, for a wide range of ows
dominated by boundary layer behaviour, including
applications where wall heat transfer is of relevance
(Esch and Menter 2003). Furthermore, the concept of
blending the o- and the e-equations has found its way
into more complex closures, like Explicit Algebraic
Reynolds Stress Models (EARSM) (Hellsten and Laine
2000) or full Dierential Reynolds Stress Models
(DRSM) (Eisfeld 2006).
On the RANS side, a relatively recent extension was
the introduction of a formulation for the prediction of
laminar-turbulent boundary layer transition (Langrty
et al. 2006, Menter et al. 2006b,c). Although the
transition phenomena themselves are handled by two
separate transport equations, this extension still benets
from some of the characteristics of the underlying SST
model. For reliable transition predictions, it is essential
that the viscous sublayer formulation of the model does
not mimic transitional behaviour itself. The k-o and
therefore the SST model avoid low-Re extensions and
can therefore be combined with specic transition
models in a robust and reliable way.
Also in recent years, the blending concept has
proven useful in the development of scale-resolving
turbulence models. The SST model (together with the
SpalartAllmaras (SA) model (Spalart and Allmaras
1994) has served as one of the main platforms for
Detached Eddy Simulation (DES) models as proposed
by Spalart (2000) and Strelets (2001). It is also one of
the options for the concept of Scale-Adaptive Simulation (SAS) developed by the authors group (Menter
et al. 2003, Menter and Egorov 2004, 2005, 2006,
Menter et al. 2006a). It is interesting to note that the
blending functions have become an essential element of
the DES approach to prevent undesired eects of the
Large Eddy Simulation (LES) limiter on boundary

layers (Menter and Kuntz 2003), resulting in the


Delayed Detached Eddy Simulation (DDES) methodology (Menter and Kuntz 2003, Spalart et al. 2006).
The current article will review the SST model
formulation and discuss some of its specics. Extensions will cover transition prediction and model
enhancements for unsteady ows. Results obtained
with the model for a limited number of applications
will be presented. Finally, model limitations and
potential advancements will be discussed briey.
The SST model
In this section, the complete formulation of the SST
model is given, with the limited number of modications
from the original version highlighted (Menter 1994).
@ rk @ rUi k

P~k  b rko
@t
@xi


@
@k

m sk m t
@xi
@xi
@ r o @ rUi o
1

a P~k  bro2
@t
@xi
nt


@
@o

m so m t
@xi
@xi
1 @k @o
21  F1 rsw2
o @xi @xi
p
a1 k
; S 2Sij Sij
nt
max a1 o ; S F2


@Ui @Ui @Uj

Pk mt
@xj @xj @xi
~
1
! Pk min Pk ; 10  b rko
8(
<

# )4 9
!
p
=
k 500n 4rso2 k
F1 tanh
;
min max  ; 2
;
:
;
b oy y o CDko y2
2"
!#2 3
p
2
k
500n
5
F2 tanh4 max  ; 2
b oy y o


1 @k @o 10
CDkw max 2rso2
;10
o @xi @xi
"

where, k is the turbulence kinetic energy, o is the


turbulence frequency, y is the distance to the nearest
wall, S is the invariant measure of the strain rate, r is
the density and Ui is the ow velocity. F1 and F2 are
blending functions which are equal to zero away from
the surface (k-e model), and switches over to one inside
the boundary layer (k-o model). It should be noted
that a production limiter is used in the SST model to
prevent the build-up of turbulence in stagnation
regions as an essential part of the SST model.

International Journal of Computational Fluid Dynamics

307

Downloaded By: [Hong Kong Polytechnic University] At: 06:19 7 September 2009

All constants are computed by a blend from the


corresponding constants of the k-e and the k-o model via
a a1F1 a2 (17F1), etc. The constants for this model
are: b* 0.09, a1 5/9, b1 3/40, sk1 0.85, so1 0.5,
a2 0.44, b2 0.0828, sk2 1, so2 0.856.
The only modications from the original formulation (Menter 1994) are the use of the strain rate, S, in
the denition of the Eddy-viscosity instead of the
vorticity and the use of the factor 10 in the production
limiter, instead of 20 as proposed by Menter (1994).
Another recent model enhancement is the inclusion
of eects of streamline curvature and system rotation
(Smirnov and Menter 2008). It is not repeated here
because of space limitations.
Near wall treatment
The near wall treatment is of equal importance in
practical industrial CFD simulations as the formulation of the turbulence model itself. Integration
through the viscous sublayer is typically referred to
as low-Re approach. The terminology low-Reynolds number refers to the turbulent Reynolds
2
number Ret kev  vvt , which is low near the wall.
This is one of the most confusing terminologies in
turbulence modelling as many engineers relate it to
the device Reynolds number. A more appropriate
terminology is Viscous Sublayer Model (VSM). In
aeronautics, VSMs are typically employed and the
equations are integrated to the wall, with a grid
resolution of the order of y*1 near the wall. Such
demands on grid resolution cannot be enforced in
general purpose codes/applications. This has been
recognised by many groups over the last decade and
numerous y-insensitive wall treatments have been
proposed (Kim et al. 2002, Esch and Menter 2003,
Craft et al. 2004, Kalitzin et al. 2005, Popovac and
Hanjalic 2007). The goal is to avoid the stringent grid
resolution requirements of VSMs and to ensure that
the predicted wall shear-stress and heat transfer is
largely independent of the near wall (y) resolution.
This is of course only reasonable if there are
otherwise enough grid points in the boundary layer.
There are many dierent names for such methods.
The author would like to propose the terminology
y-insensitive wall treatment.
Figure 1 shows velocity proles for Couette ow
simulations on three vastly dierent grids (y*0.2;
y*9; y*100). Despite the large dierences in the
near wall spacing, the computed wall shear-stress
varies by less than 5% and all solutions follow the
logarithmic prole. As a result, such a wall formulation can signicantly improve the predictive accuracy
for general industrial applications, as the user inuence
via the grid generation is drastically reduced.

Figure 1. Velocity proles for three dierent grids using the


automatic wall treatment.

It should be noted that y-insensitive wall formulations depend on the numerical method used in the
code. There is a signicant dierence, e.g. if the solver
uses a cell-centred or a vertex based formulation. It is
therefore not possible to judge the quality of formulations entirely on theoretical arguments (Popovac and
Hanjalic 2007). The best quality characterisation
would be: how much does the wall shear stress vary
for a Couette ow of suciently high Reynolds
number as the grid is coarsened? Typically, an entirely
constant value for tw cannot be achieved but maximum
variations of 510% should be achievable.
In recent years (Durbin 1995), the concept of
elliptic relaxation was introduced to avoid the need for
damping functions in the viscous sublayer region. It
could be argued that the k-o model has a built-in
elliptic relaxation. Near the wall the o-equation
reduces to (Wilcox 1998) (for constant viscosity):
m

@2o
bro2
@x2j

This equation in combination with the values


specied for o at the wall eciently communicates the
wall presence to the ow in the near wall region, without
a need for additional transport equations or damping
functions. This makes the o-equation a highly attractive
and robust modelling framework. The SST model has
proven good accuracy for a wide range of heat transfer
testcases (Esch and Menter 2003), largely because of this
ecient near wall formulation.
Aerodynamic simulations
The main motivation for the development of the SST
model was the desire to accurately predict the

Downloaded By: [Hong Kong Polytechnic University] At: 06:19 7 September 2009

308

F.R. Menter

behaviour of boundary layers under adverse pressure


gradients up to separation and possibly with the
inclusion of moderate separation bubbles. Figure 2
shows a typical example of such a ow. It is the NACA
4412 airfoil as investigated experimentally in work
done by Coles and Wadock (1979). This ow was
already part of the model validation in the work by
Menter (1994) and more details on the set up can be
found there. The current grid with 5136129 nodes
provides mesh-independent results. Figure 2 shows a
model comparison with some more recent models
(Menter 1994, Cokljat et al. 2003, Spalart and
Allmaras 2004, Wilcox 2007) in the rear part of the
suction surface of the airfoil. Visibly, the SST model
provides the best agreement against the experimental
data. Similar relations between models are observed
for other ows of the same nature (Menter 1994).
In a more severe test, the SST model was applied to
the test cases of the 2003 AIAA Drag Prediction
Workshop. The goal of the workshop was to evaluate
if current turbulence models and CFD methods allow
the prediction of the increase in drag caused by the
installation of an engine nacelle. The two congurations tested are shown in Figure 3. The models have
been designed and tested experimentally by DLR
(Brodersen and Sturmer 2001). Figure 4 shows the

Figure 2. Comparison of velocity proles for NACA 4412


airfoil (a 13.878, Re 1.5 6 106) using dierent turbulence
models.

Figure 3.

drag polars predicted with ANSYS-CFX for both


congurations using the SST model. The agreement
with the data is generally quite close and clearly
displays the o-set in drag caused by the addition of
the pylon-nacelle. The small but systematic
dierence for the wing-body-nacelle-pylon (WBNP)
case is most likely a result of the relatively low mesh
resolution of around 12 million cells. Langtry et al.
(2004) modied the a1 coecient slightly, demonstrating the sensitivity of the current ows to small
modelling details. Changing the a1 constant from
a1 0.31 to a1 0.35 resulted in a solution change
corresponding to an angle of attack variation of
Da*0.38, which is a signicant value near cruise
conditions. This corresponded to the dierence between the SA and the SST model for simple airfoil ow
like that seen in Figure 2.
Transition model formulation
Modelling of laminar-turbulent transition in boundary
layers has proven one of the most challenging tasks in
CFD for many decades. Although many industrial
ows are in the range of 104 5Re 5 106, meaning in
regimes where signicant portions of the boundary
layers can be laminar, there was simply no reliable way

Figure 4.

Drag polars for WB and WBNP testcase.

Geometries of DLR-F6 models for AIAA drag prediction workshop.

Downloaded By: [Hong Kong Polytechnic University] At: 06:19 7 September 2009

International Journal of Computational Fluid Dynamics


of including these eects even to rst order in generalpurpose CFD codes. A model extension to the SST
model which allows the inclusion of such eects into
general CFD simulations has recently been developed
by the authors group (Langrty et al. 2006, Menter
et al. 2006b,c). The model solves two additional
transport equations and incorporates experimental
correlations to trigger the transition onset. The model
formulation is strictly local and therefore fully
compatible with modern general purpose CFD codes.
Discussion of the model equations is beyond the scope
of the current article and is therefore not presented
here.
The transition model has been validated against a
wide range of experimental data (Langrty et al. 2006,
Menter et al. 2006b,c) and is used today successfully in
many industrial CFD simulations. The application
range covers turbomachinery blades, wind turbines,
racing cars all the way to the design of sailing
yachts for the Americas cup (Parolini and Quarteroni
2005).
Figure 5 shows results for the transition location
and the drag coecient for a 2D GE wind turbine
airfoil. The experiment was conducted at dierent
angles of attack. Because of the resulting change in
pressure gradients, the transition location on both
sides of the airfoil changes signicantly. Figure 5 (left)
shows the location of the transition point computed by
the model in comparison with the data and a simplied
en model from the 2D XFOIL code. The agreement
with the data is quite good, considering the sensitivity
of the transition process to experimental details. Figure
5 (right) shows the drag coecient versus angle of
attack. Scales are omitted because of data condentiality, but the relative change in the drag coecient is
apparent and of signicance in wind turbine design.
More details can be found in work done by Langrty
et al. (2006).

309

Other approaches to transition modelling have


recently become available. One of the more promising
is from Walters and Cokljat (in press). It would
however be desirable to combine this formulation with
the SST model to avoid the freestream sensitivity of the
standard k-o model underlying its current
formulation.
It is important to note that transition models can
only be combined with turbulence models which do
not mimic transitional behaviour because of their VSM
formulation. Otherwise, the interference of both
formulations would result in an unpredictable
behaviour.
Unsteady ows and scale-adaptive simulation
The long-term trend in industrial CFD simulations
goes to unsteady simulations. This is partly driven by
the increasing number of inherently unsteady applications like internal combustion engines and other
geometries with moving parts, etc. The main driver
is however the need for increased accuracy which goes
beyond the capabilities of steady state simulations.
Such simulations typically require the resolution of at
least a portion of the turbulent spectrum in at least a
portion of the ow domain. Although LES has been
around for many decades, its impact on industrial
ow simulations has been relatively moderate and
mostly restricted to free shear ows. However, most
engineering applications feature wall boundary layers,
and the high cost of LES in such regions has limited
its application in engineering ow simulations.
Spalart (2000) and Strelets (2001) were the rst to
recognise that RANS and LES models can be
combined in a very eective and simple way,
exploiting the accuracy of RANS models in attached
boundary layers and the ability of LES to more
accurately predict strongly separated ows. The

Figure 5. Right: transition location on pressure and suction side for GE wind turbine airfoil. Right: drag for transitional and
fully turbulent predictions. Comparison with experiments and XFOIL results.

Downloaded By: [Hong Kong Polytechnic University] At: 06:19 7 September 2009

310

F.R. Menter

resulting DES model (and many variants of it) has


been applied successfully to numerous industrial
ows, and is currently also been investigated as an
approach for avoiding the excessive Reynolds number
scaling of LES for wall (Peng and Haase 2008). As
stated in the introduction, the SST model has been
one of the main platforms for DES model formulations and the concept of using blending functions has
developed into an essential element of such model
formulations (Menter and Kuntz 2003, Spalart et al.
2006).
One of the motivations for DES was the inability
of standard RANS models to resolve unsteady
turbulent structures in separated and free shear
ows, even if the grid and the time step resolution
would allow that. The behaviour of RANS models in
unsteady mode (URANS) is depicted in the left part
of Figure 6. It can be seen that URANS models
typically produce single mode large scale unsteady
structures without resolving any of the details of
turbulence. It was long believed that this is a result
of the Reynolds averaging applied to the RANS
equations, which is consistent with damping small
scale turbulence. Although this argument is convincing at rst, it was recently shown in a series of
articles (Menter et al. 2003, Menter and Egorov
2004, 2005, 2006, Menter et al. 2006a) that this
behaviour is not directly related to the Reynolds
averaging, but to the specic way such models have
been derived in the past. The weak link in any
historic RANS formulation has been the scale
equation (typically the e- or o-equation). Although
an exact transport equation is available as basis
for the modelled turbulent kinetic energy equation,
no such equation can be used as basis for e or
o. The reason is that the exact e-equation describes
the smallest (dissipative) scales of turbulence, while
the modeled e-equation describes the large scales
(which are relevant for the momentum exchange).
The e-equation (and subsequently the o-equation) is
therefore modelled in pure analogy to the k-

Figure 6.

equation. This is a very weak concept and an exact


scale equation would signicantly improve the basis
of RANS models. This was recognised by Rotta
(1972), who derived an exact equation for the
turbulent length scale, L, times the turbulent kinetic
energy, kL. This formulation was re-visited recently
(Menter and Egorov 2004) and it was argued that
Rottas modelling assumptions were not entirely
consistent with the inherent nature of the terms in
the equations. In particular, a leading order term
was omitted in the Rotta model based on arguments
of homogenous turbulence, although the term proved
to be of in-homogenous nature. As a result, the
Rotta model was re-formulated into an equation
containing the second derivative of the mean ow.
The model is documented in detail by Menter et al.
(2006a) and only the basic high Re form is given
here:




@ rk @ rUj k
k3=2
@ mt @k

Pk  c3=4
r
m
@t
@xj sk @xj
L
@xj



 !
@ rF @ rUj F
F
L 2

Pk z1  z2
@t
k
LvK
@xj


@ mt @F
 z3 rk
3
@xj sF @xj
mt
nt c1=4
Pk mt S2 ;
m F;
r


1 @Ui @Uj

Sij
2 @xj @xi

LvK

0
U
k 00 ;
U

jU j S;

p
2Sij Sij ;

s
@ 2 Ui @ 2 Ui
jU j
@x2k @x2j
00

(z1 0.8, z2 1.47, z3 0.0288, sk 2/3, sF 2/3,


cm 0.09, k 0.41). The model can be run in a oneor a two-equation mode. Model constants as well as

Turbulent structures computed for cylinder in crossow. Left: SSTRANS, Right: SSTSAS model.

International Journal of Computational Fluid Dynamics

Downloaded By: [Hong Kong Polytechnic University] At: 06:19 7 September 2009

extensions for VSM modelling are given in Menter


et al. (2006a) and are not repeated here. It should be
noted that these models are also an attractive framework for steady RANS formulations (see Menter et al.
2006a) which is worthwhile further exploration.
In order to avoid the development of an entirely
new modelling framework, the above model has also
been transformed into a k-o basis. The transformed oequation reads:

!
@ro @rUj o o
L 2

Pk 1  z1 z2
@t
@xj
k
LvK




@ mt @o
 ro2 cm  c1=4
z

3
m
@xj sF @xj


2r 1 @k @o
k @o @o
 2

4
sF o @xj @xj o @xj @xj
This equation could be used within a k-o formulation. However, to preserve the characteristics of the SST
model for boundary layers ows, the SAS terms where
further modied as additional terms to the o-equation
in the SST model, in a way which does not aect the
near wall SSTRANS model (for more details, see
Menter and Egorov 2006, Menter et al. 2006a):
"


L 2
2
QSAS r  max z2 kS2
Cc
LvK
sF

 
1 @o @o 1 @k @k
k  max
;
;0
5
o2 @xj @xj k2 @xj @xj
(Cc 2). The resulting model is termed SSTSAS model.
The most relevant aspect of SAS models for the
current discussion is the appearance of the von
Karman lengths scale LvK. It is a direct result of the

311

term-by-term modelling of the exact Rotta equation. It


was found that the inclusion of this term alters the
model behaviour for unsteady ow simulations entirely
compared to conventional URANS models. The
reason is that LvK allows the model to adjust to
already resolved scales in the ow eld, thereby
avoiding the excessive damping introduced by conventional URANS. This can be seen in the right portion of
Figure 6. It should be emphasised that a numerical
scheme with low diusion has to be used in the
resolved ow regime.
Figure 7 shows the ow structures computed by
this model for a periodic hill ow. The two pictures
represent simulations on the same mesh (*2.5 million
nodes) using two dierent time steps. The time step in
the left part of the gure corresponds to a typical LES
time step (CFL 5 1), in the right part the time step is
increased by a factor of four. Further increasing the
time step would result in a steady state RANS
solution. Figure 7 illustrates the terminology ScaleAdaptive, which allows the model to adjust to the
mesh and time step resolution provided, resulting in a
continuous variation of the simulation from LES to
steady-state RANS. The colour in Figure 7 displays
the ratio of Eddy-viscosity to molecular viscosity. In
the left part of Figure 7 this ratio is of order 510 and
in the right part of order 3050. The ability of the
model to adjust its Eddy-viscosity to the resolved scales
is unique and cannot be achieved with standard LES
models. For Smagorinsky type models (vt (cD)2S),
the length scale is xed by the grid spacing, D. For
large scales the strain rate, S, is lower than for small
scales. Such a model would therefore produce a lower
Eddy-viscosity for large structures than for small ones.
Figure 8 shows the velocity proles computed with
the SAS model using the two dierent time steps in
comparison with the reference LES (Frohlich et al.

Figure 7. Turbulent structures for ow over periodic hill. Left: LES time step (CFL *1), Right 6 4 larger time step. Colour:
ratio of Eddy-viscosity to molecular viscosity.

Downloaded By: [Hong Kong Polytechnic University] At: 06:19 7 September 2009

312

F.R. Menter

Figure 8. Velocity proles for the SAS simulations with dierent time steps for periodic hill ow in comparison with reference
LES and SSTRANS solution.

2005) and the SSTRANS solution. It can be seen that


even the large time step leading to the structures seen
in the right side of Figure 7 gives a signicant
improvement in the velocity proles compared to the
steady state solution (this case is known to be
challenging for RANS models).
Considering that the current model is derived
entirely on RANS arguments makes it clear that the
URANS behaviour of classical models as shown in
Figure 6 (left) is not a result of RANS averaging, but
of the specic details of the model formulation.
It is important to emphasise that the current model
is still a (U)RANS model and that it will not go
unsteady for all ows. The model gives steady results
for attached and mildly separated wall bounded ows
and unsteady solutions mostly for ows with large
separation and mixing zones. Davidson (2006) performed interesting tests with the SSTSAS model for
attached and moderately separated ows. As expected,
the model reverted back to RANS mode in channel
ow with prescribed LES inlet turbulence, albeit
slower than classical URANS. This is of principal
interest, but it should be emphasised that it does not
represent the potential of the model to predict
unsteady industrial owelds, where the unsteady
regimes are often characterised by strong mixing in
largely separated regions.
A signicant number of test cases has been
presented in the publications from the authors group
(Menter et al. 2003, Menter and Egorov 2004, 2005,
2006, Menter et al. 2006a). In addition, numerous
industrial simulations have been carried out successfully for ows behind blu bodies (stalled airfoils,
mixers, ame holders, buildings, etc.) or for strongly
swirling ows as observed in gas turbine combustion
chambers. One of the most interesting studies was

performed for the ow in an Internal Combustion


Engine (blow-down set-up) (Grahs and Othmer 2006).
The unsteady ow behind the valve was captured very
well, resulting in proper tumble and massow rates.
Another interesting example is the ow in a chemical
mixer (Honkanen et al. 2007).
Figure 9 shows SAS simulations over a generic
airplane geometry. The simulation is in good agreement with the experimental data (right part of gure).
The simulation (Re 2.8 6 106, a 158) has been
carried out on an unstructured mesh with 11 6 106
control volumes. The left part shows the geometry
and the turbulent structures produced by the
simulation. The right part shows a comparison
between the experimental data and the time averaged
simulation. Although the comparison is not perfect, it
showed still one of the better agreements with the data
during the EU project DESIDER (Peng and Haase
2008).
It should be noted that there is a wide variety of
unsteady industrial ows, and SAS is just one of the
approaches of tackling them. It has the advantage that
it is relatively safe as it has a RANS fall-back position
in case of overly coarse grids/time steps or insucient
ow instability. This is not the case with the DES
model, which introduces a grid dependency into the
RANS portion of the model to ensure a LES mode
under certain conditions. On the other hand, DES
models can more easily be forced into unsteady mode
which can be of signicant advantage for otherwise
stable ows. It is expected that the formulation of
hybrid RANS/LES models will dominate turbulence
modelling for the next decade. SAS models should be
an interesting element for such formulations, as they
have RANS capabilities, while at the same time
understanding LES structures. They should therefore

Downloaded By: [Hong Kong Polytechnic University] At: 06:19 7 September 2009

International Journal of Computational Fluid Dynamics

313

Figure 9. Flow over generic airplane conguration FA-5. Left: ow structures. Right: comparison of experiments and SAS
axial ow component. (Geometry and data are courtesy of EADS Deutschland) Printed with permission from EADS Military
Aircraft, Germany.

be able to provide a soft interface between RANS and


LES regions.
Future directions
Within the RANS modelling approach, the inclusion
of anisotropic eects should oer some potential for
increased accuracy. Examples are secondary ows near
corners (rectangular duct ow etc.). There are strong
indications, that corner ows under adverse pressure
gradients separate later than predicted by linear EVM.
The main eect seems to be the transfer of momentum
into the corner by secondary ows driven by the
anisotropy of turbulence. Such eects should most
eciently be modelled within the framework of
EARSM or non-linear EVM. One of the primary
requirements for such models will however be a similar
robustness as existing EVM for complex ow.
A second and long standing open problem of all
existing RANS models is their inability to predict the
ow-recovery downstream of moderate to large
separation zones. This was discussed many years ago
(Johnson et al. 1994), but is still an unsolved issue,
although it should be solvable within the RANS
modelling framework.
Transition modelling based on locally formulated
transport equations is a relatively new area in
industrial CFD. It is anticipated that such models
can and will be further enhanced and simplied over
the next years. Additional eects like cross-ow
instabilities should be taken into account.
The main thrust of turbulence modelling will be in
the area of hybrid RANSLES methods. There is
good potential of achieving formulations which avoid
the excessive grid resolution requirements of current

LES models near walls. Although some in the LES


community follow such developments with skepticism,
the potential of such methods is much too high to
ignore them. After all, LES is not an exact science, but
a pragmatic attempt of solving problems which are
beyond Direct Numerical Simulation on todays
computers. Like in all areas of physics modelling,
those formulations will be successful which oer the
best compromise between accuracy, robustness and
computational eciency. Combinations of (U)RANS
and LES models in various forms have a good chance
of becoming the foundation of the next generation of
engineering turbulence models.
Acknowledgements
The author thanks his colleagues Yuri Egorov, Robin Langtry
(now at The Boeing Company), Richard Lechner and Martin
Kuntz, who have performed the computations shown in this
review article. Some of the results have been produced during
the European Research Projects Flomania and DESIDER
and have been funded by the European Union.

References
Brodersen, O. and Sturmer, A., 2001. Drag prediction of
engine airframe interference eects using unstructured
NavierStokes calculations. In: 15th AIAA Computational Fluid Dynamics Conference. June 1114, 2001,
Anaheim, CA. AIAA Paper 20012414.
Cokljat, S.E., et al., 2003. A comparative assessment of the
V2F model for recirculating ows. In: Paper 20030765,
Reno, USA.
Coles, D. and Wadock, A.J., 1979. A ying-hot-wire study of
two-dimensional mean ow past an NACA 4412 airfoil
at maximum lift. AIAA Journal, 17 (4), 321328.
Craft, T.J., et al., 2004. A new wall function strategy for
complex turbulent ows. Numerical Heat Transfer, Part
B 45, 301317.

Downloaded By: [Hong Kong Polytechnic University] At: 06:19 7 September 2009

314

F.R. Menter

Davidson, L., 2006. Evaluation of the SSTSAS model:


channel ow, asymmetric diuser and axi-symmetric hill.
In: Proceedings ECCOMAS CFD 2006, Egmond aan
Zee, The Netherlands.
Durbin, P.A., 1995. Separated ow computations with the ke-v2 model. AIAA Journal, 33, 659664.
Eisfeld, B., 2006. Numerical simulation of aerodynamic
problems with a Reynolds stress turbulence model, Notes
on numerical uid dynamics and multidisciplinary design. Berlin/Heidelberg: Springer.
Esch, T. and Menter, F.R., 2003. Heat transfer predictions
based on two-equation turbulence models with advanced
wall treatment. Turbulence Heat Mass Transfer, 4, 6333
6640.
Frohlich, J., et al., 2005. Highly-resolved large Eddy
simulations of separated ow in a channel with streamwise periodic constrictions. Journal of Fluid Mechanics,
526, 1966.
Grahs, T. and Othmer, C., 2006. Evaluation of aerodynamic
noise generation: parameter study of a generic side
mirror evaluating the aeroacoustic source strength. In:
Proceedings of ECCOMAS CFD.
Hellsten, A. and Laine, S., 2000. Explicit algebraic Reynolds
stress modelling in decelerating and separating ows. In:
AIAA Paper 20002313, Denver.
Honkanen, M., et al., 2007. Time-resolved stereoscopic PIV
experiments for the validation of transient CFD simulations. In: 7th international symposium on particle image
velocimetry, Roma.
Johnson, D.A., Menter, F.R., and Rumsey, C.L., 1994. The
status of turbulence modeling for external aerodynamics.
In: 32nd AIAA Aerospace Sciences Meeting, Reno, NV,
January 2004. AIAA Paper 19942226.
Kalitzin, G., et al., 2005. Near-wall behaviour of RANS
turbulence models and implications for wall functions.
Journal of Computational Physics, 204, 265291.
Kim, S.E., Cokljat, D., and Choudhury, D., 2002. A
pragmatic near-wall treatment for turbulent ows and
heat transfer. In: Paper presented at poster session, 5th
ASME-ISHMT Heat and Mass Transfer Conference, 35
January, Calcutta, India.
Langtry, R., Kuntz, M., and Menter, F.R., 2004. Drag
prediction of engine-airframe interference eects with
CFX-5. In: AIAA Paper 20040392, Reno.
Langtry, R.B., Menter, F.R., Likki, S.R., and Suzen, Y.B.,
2006. A correlation-based transition model using local
variables Part II: Testcases and industrial applications.
Journal of Turbomachinery, 128123. 423.
Menter, F.R., 1992. Inuence of freestream values on k-o
turbulence model predictions. AIAA Journal, 30 (6),
16571659.
Menter, F.R., 1994. Two-equation Eddy-viscosity turbulence
models for engineering applications. AIAA Journal, 32
(8), 269289.
Menter, F.R. and Egorov, Y., 2004. Revisiting the turbulent
length scale equation. In: IUTAM Symposium: one
hundred years of boundary layer research, Gottingen.
Menter, F.R. and Egorov, Y., 2005. SAS turbulence modelling
of technical ows. In: Proceedings of direct and large Eddy
simulation, Netherlands: Springer, 687694.
Menter, F.R. and Egorov, Y., 2006. A scale adaptive
simulation model using two-equation models. In: AIAA
Paper 20051095.

Menter, F.R. and Kuntz, M., 2003. Adaptation of Eddyviscosity turbulence models to unsteady separated ow
behind vehicles. In: Proceeding conference on the aerodynamics of heavy vehicles: trucks, busses and trains, ,
Asilomar, Springer.
Menter, F.R., Egorov, Y., and Rusch, D., p
2006a.
Steady and
unsteady ow modelling using the k  kL model. In: K.
Hanjalic, Y. Nagano, and S. Jakirlic, eds., Turbulence,
Heat and Mass Transfer, 5.
Menter, F.R., Kuntz, M., and Bender, R., 2003. A scaleadaptive simulation model for turbulent ow predictions.
In: 41st AIAA Aerospace Sciences Meetings, Reno, NV,
January 2003. AIAA Paper 20030767.
Menter, F.R., Langtry, R.B., Likki, S.R., and Suzen, Y.B.,
2006b. A correlation-based transition model using local
variables Part I: model formulation. Journal of
Turbomachinery, 128, 413.
Menter, F.R., Langtry, R.B., and Volker, S., 2006c.
Transition modelling for general purpose CFD codes.
Journal of Flow Turbulence and Combustion, 77 (14),
277303.
Parolini, N. and Quarteroni, A., 2005. Mathematical models
and numerical simulations for the Americas Cup.
Computer Methods in Applied Mechanics and Engineering, 194, 10011026.
Peng, S.-H. and Haase, W, 2008., Advances in hybrid
RANS-LES modelling. In: S.-H. Peng and H. Werner,
eds. Notes on Numerical Fluid Mechanics and Multidisciplinary Design, Vol. 97. Berlin: Springer.
Popovac, M. and Hanjalic, K., 2007. Compound wall
treatment for RANS computation of complex turbulent
ows and heat transfer. Journal of Flow Turbulence and
Combustion, 78 (2), 177202.
Rotta, J.C., 1972. Turbulente Stromungen. 1. Au., Stuttgart:
B.G. Teubner, ISBN 3-519-02316-4.
Smirnov, P. and Menter, F.R., 2008. Sensitizing of the SST
turbulence model to rotation and curvature by applying
the Spalart-Shur correction term. In: Proceedings of
ASME Turbo Expo, Berlin.
Spalart, P.R., 2000. Strategies for turbulence modelling and
simulations. International Journal of Heat Fluid Flow, 21,
252263.
Spalart, P.R. and Allmaras, S.R., 1994. A one-equation
turbulence model for aerodynamic ows. La Recherche
Aerospatiale, 1, 521.
Spalart, P., et al., 2006. A new version of detached Eddy
simulation, resistant to ambiguous grid densities. Journal
of Theoretical and Computational Fluid Dynamics, 20,
181195.
Strelets, M., 2001. Detached Eddy simulation of massively
separated ows. In: 39th AIAA Aerospace Sciences
Meeting. AIAA Paper, 20010879.
Walters, D.K. and Cokljat, D., in press. A three-equation
Eddy-viscosity model for Reynolds averaged Navier
Stokes simulations of transitional ow. Journal of Fluids
Engineering.
Wilcox, D.C., 1998. Turbulence modeling for CFD. 1st ed. La
Canada, CA: DCW Industries Inc.
Wilcox, D.C., 2007a. Formulation of the ko turbulence
model revisited. In: 45th AIAA Aerospace Sciences
Meeting. Reno, NV, USA. AIAA Paper 20071408.

315

International Journal of Computational Fluid Dynamics

Downloaded By: [Hong Kong Polytechnic University] At: 06:19 7 September 2009

Appendix. Comments on the 2006 Wilcox k-w model


In an eort to improve the original k-o model, Wilcox has
recently proposed a modied k-o model formulation (Wilcox
2007). The main goals were to avoid/reduce the freestream
sensitivity of the 1988 model and to improve its predictive
accuracy for ows separating under adverse pressure
gradients. It is interesting that Wilcox adopted two of the
elements of the SST model, namely the cross-diusion term
and the SST limiter.
There are some noteworthy aspects in comparison
with the SST model. Firstly, instead of using a blending
function to activate the cross diusion term, the activation
is determined by the sign of this term. Secondly, the constant
sd (equal to 0.125 if (@k/@xj  @o/@xj) 4 0) of the
cross diusion term in the Wilcox model is an order of
magnitude lower than in the SST model (where it
corresponds to 1.7). Thirdly, the SST limiter in the
denition of the Eddy viscosity translates into a value of
a1 0.34, and is therefore less stringent than in the SST
model. The result of this higher a1 value can be seen in
Figure 2 of the main paper, where the model shows a
reduced tendency for predicting ow separation under
adverse pressure gradients.
However, particularly the low value of the constant sd in
the positive part of the cross-diusion term is of concern, as
this term was introduced into the SST model for avoiding the
freestream sensitivity of the original k-o model. Applying
only a fraction of this term bears the danger of maintaining a
signicant portion of this sensitivity in the model. This is
actually the case as will be shown below.
Wilcox (2007b) argues that specied freestream values
for o should be small compared to o values inside the
corresponding shear layer. However, specifying low values
for o is not as easy as it appears on rst inspection. There are
several reasons why this approach is not suitable for general
industrial CFD simulations. Firstly, it has to be considered
that a high freestream value for one shear layer is a low value
for another, as o scales with the strain rate inside the layer.
In most technical applications, there are numerous boundary
layers and free shear layers in a single simulation domain and
a proper freestream value for all sub-regions is hardly
achievable. Secondly, low values for o typically result in unphysically low values for the turbulent kinetic energy, k (due
to the demand of keeping mt/m low at the inlet). Thirdly,
modelling the eect of laminar-turbulent boundary layer
transition is of increasing importance in modern industrial
CFD simulations (Langtry et al. 2006, Menter et al. 2006b).
Transition depends crucially on the freestream values for
turbulence outside the boundary layer. The simulations have
to mimic the level and the decay of the freestream turbulence
in agreement with the physical conditions. This requires the
specication of both k and o at the inlet to match the level
and the decay of k in the domain. In such simulations there is
no liberty in selecting the inlet values of o independent of the
ow physics. In addition, one should consider that very low
freestream values for o (compared to strain rates in all shear
layers in the domain) increase the danger of instabilities
during the simulations. For very low o, any small production
of k can lead to excessive levels of the Eddy-viscosity during
the iterative process because of the appearance of o in the
denominator of the production term of the Eddy-viscosity
Pmt  1=oPk :
More importantly however, arguing for low freestream
values for o (relative to its values inside shear layers) means

arguing against the equations and their natural behaviour.


This can be understood by examining the self-similar ow
equations. The full equations are given by Wilcox (1998) and
only the relevant part is reproduced here. The equation for
the non-dimensional o-equation for a mixing layer reads
(stress-limiter and round/plane jet terms omitted for
simplicity of argument):



 
dW d
dW
W dU 2

sN
UW aN
dZ dZ
dZ
K dZ
s
d dK dW
 bW2
W dZ dZ

kx; y U21 KZ; ox; y

Ux; y U1 UZ;
N

K
;
W

U1
WZ;
x

y
x

where, U1 is the freestream velocity in the upper stream


(for the current equations the lower stream has zero
velocity), x is the streamwise and y the normal co-ordinate.
The interesting aspect of this equation is that it allows for
two self-similar freestream values for W for the U1
freestream (U 1):
W1d 0;

W2d

1
b

Zero freestream values are obviously not possible in


practice, but low values closer to W1d than to W2d could be
specied. However, specication of a value for W much
lower than W2d would not result in a self-similar layer
development. This can be seen by evaluating the freestream
development of o directly from the o-equation for uniform
parallel ow:
U1

dod
bo2d !
dx
1
o@
1=o0 b=U1  x

where, o0 is the value specied at x 0. For x !? one


gets:
od
for

U1
! W W2d
bx
x!1

In other words, even if a low but nite value is specied


for o0 at the inlet (x 0), the non-dimensional freestream
value will nevertheless approach W2d for long running
lengths, x. As the spreading rates for W1d and W2d are
dierent, a self-similar solution will only be achieved once W
approaches W2d. The nal self-similar solution will
correspond to the high freestream value W2d (which can
be around 2530% of the maximum value in the layer).
Similar arguments can be made for all self-similar ows with
non-zero freestream velocity (e.g. mixing layers, boundary
layers).
Figure A1 shows the spreading rate computed for a
mixing layer of two streams with velocities U1 9.18 m/s,

316

F.R. Menter

Downloaded By: [Hong Kong Polytechnic University] At: 06:19 7 September 2009

U2 15.18 m/s (Brodersen and Sturmer 2001) using the full


NavierStokes equations and the full new k-o model. Three
dierent freestream values were specied at the inlet. The

Figure A1. Spreading rates for mixing layer computed with


SST and new k-o model using dierent freestream values.

SST model was run with the same settings. For the SST
model only one curve is shown as its results do not depend on
the o inlet value. As expected from the above argument, the
spreading rate of the new k-o model is lowest for the high
freestream value and highest for the lowest one. It can also be
seen that the low and particularly the medium freestream
values do not produce a self-similar solution. The spreading
rate changes from a high value near the inlet to lower values
downstream. Eventually all curves converge to the spreading
rate corresponding to the W W2d solution. The W2d
spreading rate of the new k-o model is about 30% lower
than the experimental value. Figure A2 shows the nondimensional velocity and shear-stress proles for the same
simulations. Consistent with Figure A1 the high value for o
results in a signicant underprediction of the turbulent stress
level and of the spreading rate.
The above discussion demonstrates that freestream
values can have a large eect on CFD solutions for certain
models. In the above simulation, the Eddy-viscosity changes
by a factor of two (!) depending which freestream value is
used for the model. Considering the many uncertainties
already present in industrial CFD simulations, it is desirable
to avoid such additional ambiguities resulting from the
formulation of the turbulence model.

Figure A2. Self-similar solution mixing layer for SST model and new k-o model. Left: velocity proles. Right: principal
turbulent shear stress u0 v0 (both non-dimensionalised by the velocity dierence U2 7 U1;Ymix y/x).

You might also like