Professional Documents
Culture Documents
C _
2004 Kluwer Academic Publishers. Printed in the Netherlands.
187
A General One-Equation Turbulence Model for Free
Shear and Wall-Bounded Flows
EHAB FARES and WOLFGANG SCHR
ODER
Aerodynamisches Institut, RWTH Aachen W ullnerstrasse 5-7, Aachen, Germany;
E-mail: ehab@aia.rwth-aachen.de
Received 17 June 2003; accepted in revised form 6 April 2004
Abstract. The purpose of this workis tointroduce a complete andgeneral one-equationmodel capable
of correctly predicting a wide class of fundamental turbulent ows like boundary layer, wake, jet, and
vortical ows. The starting point is the mature and validated two-equation k- turbulence model of
Wilcox. The newly derived one-equation model has several advantages and yields better predictions
than the SpalartAllmaras model for jet and vortical ows while retaining the same efciency and
quality of the results for near-wall turbulent ows without using a wall distance. The derivation and
validation of the new model using ndings computed by the SpalartAllmaras and the k- models
are presented and discussed for several free shear and wall-bounded ows.
Key words: one-equation, turbulence, model
1. Introduction
Many turbulence models were introduced in the past ranging from simple and in-
complete algebraic models like the mixing length hypothesis by Prandtl [38] to
complete stress-transport models like the LaunderReeceRodi model [20]. How-
ever, the increasing complexity of the turbulence models and the larger computa-
tional effort, when they are applied to realistic technical problems, are not always
justied by a qualitative improvement in the solutions. An overview on some of the
standard and widely used turbulence models is given in [62].
One-equation turbulence models enjoyed a wide popularity in the last decade.
The growing interest in this type of models is explained primarily by the numerical
ease of use compared to standard two-equation models like the k- [18] and k-
models [61]. Algebraic models like the customary BaldwinLomax model [1] are
efcient from a numerical point of view but lack generality and have decits like
the missing transport and diffusion effects. Comparatively, one-equation models
represent the simplest form of turbulence models that include transport effects
and are considered to be a good compromise between the algebraic and classical
two-equation models.
The rst one-equation model for the turbulent kinetic energy k was postulated
by Prandtl [39] and independently introduced by Emmons [9]. Both formulations
were incomplete since they needed the denition of a problem dependent length
188 E. FARES AND W. SCHR ODER
scale. Later Bradshaw, Ferris, and Atwell [6] proposed another model for k based
on the assumption that the Reynolds shear stress in a 2D ow
t
xy
is proportional
to the turbulent kinetic energy k
t
xy
=
b
k with
b
= 0.3 (1)
which is known as the Bradshaw hypothesis and was validated by Townsend [57]
through numerous measurements in boundary layers, wakes and mixing layers.
The rst formulation for a one-equation model directly for
t
was postulated
by Nee and Kovasznay [33] based on phenomenological arguments. A complete
model for
t
was also proposed by Secundov [50] as an extension of the former
Nee andKovasznaymodel withconsiderable success. Its current formulation, which
was unknown in the English speaking community until its publication in 1995 is
called the
t
92
version [51]. In 1991 Baldwin and Barth [2] derived an elaborate
transport model from the classical k- model, which had limited success due to
its inconsistent numerical behavior at turbulent/nonturbulent interfaces [28, 62].
The most successful one-equation model up to date was introduced in 1992 by
Spalart and Allmaras [52]. It was developed generically for the turbulent eddy vis-
cosity
t
with strong emphasis on the numerical behavior. Later Menter [28] pro-
posed a general methodology for deriving one-equation models from two-equation
models and suggested a new model based on the standard k- equations, which
shows superior results when compared to the BaldwinBarth model. Nagano, Pei,
and Hattori [32] also derived another one-equation model from the low-Reynolds-
number k- model with many closure and damping functions and showed good
results compared with DNS and experiments for wall-bounded ows like bound-
ary layers and plane wall jets but did not discuss any aerodynamic or free shear
ows.
The performance and the predictive capabilities of many models including the
BaldwinLomax, BaldwinBarth, SpalartAllmaras, standard k-, k-, and the
SST[27] models are documented in many publications [3, 13, 42, 43, 46, 51, 59, 62]
and corroborate the quality of the one-equation models, especially the Spalart
Allmaras model. There are, however, some limitations in the generality of almost
all one-equation models, since they lack a correct prediction of both wall-bounded
and free shear ows simultaneously. This drawback is eliminated in the proposed
turbulence model, while maintaining the same efciency of other one-equation
models.
First, the model is derived mathematically from the two-equation k- model.
Closure functions and coefcients are stipulated with the help of relevant turbulent
ows. In the second part the nal version of the model is given and compared
with the SpalartAllmaras model. The third part constitutes the detailed valida-
tion and discussion of the model based on wake, jet, vortical, wall-bounded, and
aerodynamic ows. Finally, a short summary is given.
A GENERAL ONE-EQUATION TURBULENCE MODEL 189
2. Derivation
Many technically relevant complex ows are composed of fundamental ows like
boundary layer, wake, jet, and vortical ows. The quality of the overall solution is
determined by the capability to capture the detailed physics of these ows which
is why the new model focuses on the prediction of such basic phenomena.
The starting point will be the mature and validated two-equation k- turbulence
model of Wilcox [62] given in tensor notation
Dk
Dt
=
t
i j
u
i
x
j
k
x
j
_
(
t
)
k
x
j
_
(2)
D
Dt
=
k
t
i j
u
i
x
j
x
j
_
(
t
)
x
j
_
(3)
with the following closure coefcients
= = 0.5 , = 0.52 ,
c
= 0.09 ,
c
= 0.072 (4)
and the closure functions
c
f
=
1 680
2
k
1 400
2
k
k
= max
_
0.,
1
3
k
x
j
x
j
_
(5)
=
c
f
=
1 70
1 80
i j
j k
S
ki
(
c
)
3
(6)
using the denitions
i j
=
1
2
_
u
i
x
j
u
j
x
i
_
S
i j
=
1
2
_
u
i
x
j
u
j
x
i
_
(7)
t
=
k
w
t
i j
= 2
t
S
i j
. (8)
The functions
k
and
wall
produces high gradients near the wall and requires a higher resolution
in the boundary layer. Further criticism [26, 27] concerning the freestream value
init
and
inflow
sensitivity are partly remedied in the new formulation of the k-
model [62]. Such criticismneglects the fact that the boundary values of a turbulence
model like in any other boundary value problem should affect the solution, e.g., in
the order of the free stream value, a fact that is rarely mentioned in the debate.
190 E. FARES AND W. SCHR ODER
An equation for
t
can be derived using the denition of the eddy viscosity
k =
t
and the total derivatives Dk/Dt and D/Dt [2, 28]:
D
t
Dt
=
1
_
Dk
Dt
t
D
Dt
_
= (1 )
t
i j
u
i
x
j
k
t
x
j
_
(
t
)
k
x
j
_
x
j
_
(
t
)
x
j
_
. (9)
The terms on the right-hand side are rewritten
1
by inserting k =
t
and using
the equivalence
t
S
i j
u
i
x
j
(
)
t
x
j
_
(
t
)
t
x
j
_
2
(
t
)
t
x
j
x
j
(10)
where
k
is redened
k
= max
_
0.,
1
3
_
t
_
x
j
_
2
t
x
j
x
j
__
. (11)
Note that since so far only a reformulation without any simplication has been
carried out there is no loss in generality compared to the k- formulation in Equa-
tions (2, 3). The division by in two terms on the right-hand side is no problem if
0 since this is physically only possible in laminar ows
t
= 0, which means
that the numerator also tends to zero. Usually a max(, ) with a small numerical
value , e.g., the machine accuracy, is used to numerically prevent division by zero.
The value should be reconstructed explicitly based on ow quantities such as
velocity gradients. By dening this reconstruction we close our proposed equation.
It is clear that this is the most essential step, since the generality of the reconstructed
must be guaranteed throughout a wide range of ows.
The most appropriate assumption concerning such a reconstruction is the
Bradshaw hypothesis given in Equation (1). This assumption is directly imple-
mented into many turbulence models like in [6, 28] and is indirectly included in
several other turbulence models like k- [18]
2
and k- [62]. For instance, neglect-
ing the convective and dissipative terms of the turbulence energy Equation (2) the
aforementioned hypothesis can be easily derived
k =
c
k
2
/
t
=
t
i j
u
i
x
j
k =
1
_
c
_
t
i j
u
i
x
j
(12)
and for 2D shear layers
k =
1
_
t
xy
=
t
_
du
dy
. (13)
A GENERAL ONE-EQUATION TURBULENCE MODEL 191
DNS and experimental data indicate that this hypothesis is neither exactly valid
in the viscous sublayer of the turbulent boundary layer [32] nor in free shear lay-
ers [48]. Several attempts based on dimensional analysis and on simplications
of the k- equations were studied to achieve other reconstructions. However, they
usually lead to ill numerical behavior, since they include higher-order derivatives
and cross derivatives of u and
t
and very strong gradients near the wall similar
to the underlying k- model. On the one hand, an exact reconstruction means an
innite value of
wall
at the wall, which on the other hand should be avoided
for the sake of efciency and stability of the numerical scheme. Consequently, a
slight modication of the coefcients of Equation (10) and its closure functions
is preferred in the formulation of the new one-equation turbulence model. The
quantity is reconstructed using Bradshaws hypothesis for 2D shear layers
=
k
t
=
1
_
du
dy
(14)
and generalized using the norm of the strain tensor [32]
=
1
_
c
_
2S
i j
S
i j
. (15)
In the next sections several tuning and closure coefcients and functions are pre-
sented, which were evaluated using computations of free shear and wall boundary
layer ows.
2.1. SHEAR EDGE
The value of = 0.5 may lead to a non-consistent behavior of the model at
the turbulent/nonturbulent edge like the one observed with the BaldwinBarth
model [28, 54, 62] and a higher value is suggested. A recalibration of the
constants and of the model especially for wall bounded ows led to the
values
= 1.2 and = 0.29. (16)
2.2. FREE SHEAR FLOWS
The calibration of the closure function f
=
1 650
k
1 100
k
. (17)
A recalibration was necessary to achieve the same good agreement between cal-
culations using the k- model and experiments. Note, that Wilcox introduced this
function and calibrated it to improve the original k- model for the same free shear
ows. This function does not play a role in the vicinity of the wall [62] since
k
approaches zero near solid surfaces.
192 E. FARES AND W. SCHR ODER
2.3. NEAR-WALL BEHAVIOR
The Bradshawhypothesis is exactly fullled in the logarithmic region of the bound-
ary layer. This probably explains the success of this assumption in the turbulence
modeling community. However, it is not valid in the viscous sublayer, where the
laminar viscosity is dominating. Almost all turbulence models include damping
terms near the wall to correct this behavior. Many of these terms are functions
of the wall normal distance d. To avoid such a dependence the viscous damping
function f
v1
of Mellor and Herring [25] with a calibrated value of c
v1
similar to the
approach applied in the SpalartAllmaras model was chosen
t
= f
v1
with f
v1
=
3
3
c
3
v1
=
c
v1
= 9.1. (18)
This approach allows to be equal to u
c
k and
d
dt
=
c
2
(20)
and solved analytically for large t
k(t ) t
c
/
c
= t
5/4
and (t )
1
c
t
(21)
which is in good agreement with experiments. Similar to other one-equation models
the proposed one-equation model in Equation (10) does not predict a decay at all,
A GENERAL ONE-EQUATION TURBULENCE MODEL 193
since all terms include spatial derivatives that vanish in a homogeneous owleading
to d
t
/dt 0. The correct decay of
t
can be derived from the k- equation
t
(t ) = k(t )/(t )
t
(t ) t
1
c
/
c
= t
1/4
(22)
To achieve this behavior a generic term is added to Equation (10)
d
t
dt
= k
init
_
c
/c
1
c
/c
= k
init
_
init
_
5
(23)
where k
init
represents the initial turbulent kinetic energy that can be calculated from
the initially, i.e. prescribed turbulence intensity Tu
Tu =
_
2k
3u
2
. (24)
Furthermore, the following relation can be derived using Equations (20) for t = t
init
d
t
dt
init
=
1
_
dk
dt
k
d
dt
_
init
=(
c
c
) k
init
= k
init
_
init
_
5
(25)
This leads to the value of
= (
c
c
)
1/5
0.4478. (26)
The small value of < 1 taken to the power of 5 does not interfere with the
former calibration of the model. Furthermore it should be noted that this additional
term is not essential in the calculation of turbulent steady ows.
A simulation of isotropic turbulence using the original k- model and the new
model with the additional term given in Equation (23) is presented in Figure 1. It
shows the excellent agreement between the proposed term and the predicted decay
of the original model.
3. Final Version of the Model
Using
t
= f
v1
with f
v1
=
3
3
c
3
v1
and =
(27)
the new one-equation turbulence model reads for the dependent variable
D
Dt
=2(1 )
S
i j
u
i
x
j
(
) k
init
_
init
_
5
x
j
_
( )
x
j
_
2
( )
x
j
x
j
(28)
194 E. FARES AND W. SCHR ODER
Figure 1. Decay of isotropic turbulence.
with the closure coefcients
= 1.2 = 0.29
c
= 0.09
c
= 0.072 c
v1
= 9.1 =
_
c
c
_
1/5
(29)
and the closure functions
c
f
=
1 650
k
1 100
k
k
=max
_
0.,
1
3
_
_
x
j
_
2
x
j
x
j
__
(30)
=
c
f
=
1 70
1 80
i j
j k
S
ki
(
0
)
3
=
1
_
c
_
2S
i j
S
i j
. (31)
The initial and boundary conditions for
init
0.1
wall
= 0
inflow
=
init
n
outflow
= 0 (32)
with the initial turbulent kinetic energy k
init
= 3/2 u
2
Tu
2
[
init
.
There are many similarities between the SpalartAllmaras (SA) model and the
new one-equation model like the same transformation from
t
to and the same
A GENERAL ONE-EQUATION TURBULENCE MODEL 195
Table I. Comparison between the right-hand-side terms of the SpalartAllmaras and
the new model.
Term SA model New model
Production c
b1
_
_
2
i j
i j
2
d
2
_
2(1 )
_
S
i j
u
i
x
j
1
3
i j
u
i
x
j
_
Destruction c
w1
f
w
_
d
_
2
(
)
Diffusion 1
1
x
j
_
( )
x
j
_
x
j
_
( )
x
j
_
Diff./Dest.
c
b2
x
j
x
j
2
( )
x
j
x
j
Decay k
init
_
init
_
5
convective term. In Table I the right-hand side terms of both turbulence models
are juxtaposed. Turbulence is produced in the SA model mainly through vorticity,
whereas in the new model it is produced through shear stress. For a 2D free shear
layer the production term in both models reduces to u/y but with different
proportionality factors. The destruction terms are completely different, since there
is no dependence on the wall distance d in the newmodel. The newmodel possesses
a more difcult formulation of the destruction term determined by the closure
functions f
, f
,
k
,
x
j
x
j
shows that this termis almost always
negative, i.e., this term acts as a destruction term. Unlike in the SA model the
quantity reconstructed through shear stress plays a major role in the production,
destruction, and diffusion terms. Finally, the new model is extended to account for
the decay of turbulence, a feature that is missing in the Spalart-Allmaras model.
4. Validation and Discussion
The Favre-averaged NavierStokes equation, and the equation of the turbulence
model are solved numerically for several ow problems. Boundary layer solu-
tions are achieved using similarity transformations as described in [62]. 2D ows
are calculated on the basis of a nite-volume discretization on block-structured
grids. The viscous stresses were discretized using standard second-order accurate
central schemes and the spatial discretization of the convective terms follows the
AUSM
_
1
r
2
(r
2
)
r
1
r
2
r
r
2
v
/
t
u
/
r
_
(33)
with =r
r
_
v
t
r
_
. (34)
The Reynolds stress is modeled according to the Boussinesq hypothesis
v
/
t
u
/
r
=
t
r
r
_
v
t
r
_
, (35)
which leads to the following tangential momentum equation
v
t
t
=(
t
)
_
2
v
t
r
2
1
r
v
t
r
v
t
r
2
_
_
v
t
r
v
t
r
_
t
r
. (36)
Using the circulation = 2rv
t
Equation (36) yields
t
=(
t
)
_
r
2
1
r
r
_
r
2
r
_
t
r
. (37)
Experimental [17, 36], theoretical [14, 35, 45], LES [56], and DNS [34, 40] data
show that self-similarity exists with the similarity variable = r/r
1
and r
1
t
1/2
,
where the subscript 1denotes the positionof maximumtangential velocity. Hoffman
and Joubert [17] proposed that a triple layer structure of the turbulent vortex core
exists obeying different laws for the ratio of circulation /
1
Visc. layer: 0 < 1
1
= C
v
2
(38)
200 E. FARES AND W. SCHR ODER
Log. layer: 1
1
= 1 ln (39)
Outer layer: 1
1
=
1
= const. > 1 (40)
with the constant C
v
= 1.83 determined experimentally.
3
Furthermore, it was
shown [14, 44] that the integral value of circulation decit yields
_
0
_
1
_
d
2
0 for
0 (41)
at high Reynolds numbers. Given the similarity laws of the turbulent vortex in
Equations (3840) and the ratio
=
init
/ = 1000 at the dimensionless time T =
t v
1
init
/r
1
init
= 1000 give good agreement in the viscous and logarithmic region,
whereas the circulation overshoot presented in Figure 7 deviates strongly depending
on the turbulence model used. The new model gives an overshoot of 2% which is
in good agreement with DNS simulations. The SpalartAllmaras model gives a
much stronger overshoot of more than 20%, a value which is almost unaffected by
the rotation correction term. The temporal development of the maximumtangential
velocity v
1
presented in Figure 8 shows a much stronger decay of the vortex for both
SpalartAllmaras model versions than for the new turbulence model. The slope
of the decay t
1/2
for the self-similar turbulent vortex discussed in [15, 45, 64]
is approximately predicted by the new model and the SpalartAllmaras model
using the rotation and curvature corrections. However, the SpalartAllmaras model
approaches the slope at much smaller values of the maximum tangential velocity
than the newly developed model.
A GENERAL ONE-EQUATION TURBULENCE MODEL 201
Figure 6. Comparison of numerical and analytical solutions of the self-similar turbulent vortex
for Re
for Re
= 1000 using different turbulence models compared with the self-similar decay which is
proportional to t
1/2
.
It can be concluded that the newmodel gives the best prediction of the circulation
overshoot and the decay of the turbulent vortex. The ndings of the newmodel are in
excellent agreement with calculations using complex Reynolds stress models [64]
and DNS [34, 40].
4.3. TURBULENT FLAT PLATE BOUNDARY LAYER
A detailed investigation of the plane incompressible turbulent Couette ow using
dimensional analysis, estimating the order of molecular and turbulent momentum
transfer and the subdivision of the ow into two layers led to the universal law of
the wall [12, 48]. Relations u
(y
c
f
_
1
1
2
Ma
2
1
2
= log(Re
x
c
f
)
1
2
log
_
1
1
2
Ma
2
_
. (42)
Results of the turbulent boundary layer at Ma
= 0.3 and Re
L
= 10
6
are
shown in Figure 9 for the new and the SpalartAllmaras model in comparison
A GENERAL ONE-EQUATION TURBULENCE MODEL 203
with the log law distribution. Note that the beginning of the logarithmic region is
predicted at about y
min
1. The test cases were dened at the
AFSORIFPStanford conference on the computation of Turbulent Boundary Lay-
ers [19, 62]. Table III gives an overview on the calculated ows and the related
experiments.
The gures prove the capability of the novel turbulence model to com-
pete with the other models. The new model behaves either like the k- or the
SpalartAllmaras model. Although the skin-friction coefcient distributions is
somewhat overpredicted for the adverse pressure gradient ows (Figures 13,
15, 16) it is fair to conclude that it yields in all cases satisfactory or good
results.
204 E. FARES AND W. SCHR ODER
Figure 9. Comparison of numerical ndings and the lawof the wall of a at plate at Ma
= 0.3
andRe
L
= 10
6
at different gridresolutions Top: (12965), Center: (6533), Bottom: (3317).
A GENERAL ONE-EQUATION TURBULENCE MODEL 205
Figure 10. Comparison of numerical ndings of a at plate ow at Ma
= 0.3 and Re
L
=
10
6
and theoretical laminar and turbulent (Equation (42)) c
f
distributions at different grid
resolutions Top: (129 65), Center: (65 33), Bottom: (33 17).
206 E. FARES AND W. SCHR ODER
Figure 11. Computed and measured velocity (left) and skin-friction (right) for boundary layer
ow at a favorable pressure gradient (Flow 1300 [24, 62]).
Figure 12. Computed and measured velocity (left) and skin-friction (right) for boundary layer
ow at a favorable pressure gradient (Flow 6300 [62]).
4.5. RAE 2822 AIRFOIL
The transonic turbulent ow over the RAE 2822 airfoil, which was experimen-
tally investigated in [7], represents a standard aerodynamic test case for turbulence
models. The ndings of the simulation of the owat Ma
= 0.73, Re
c
= 6.510
6
and an angle of attack = 2.79
= 0.73, Re
c
= 6.5 10
6
,
and = 2.79
.
Figure 18. Computed Mach number contours of the RAE 2822 airfoil at Ma
= 0.73, Re
c
=
6.5 10
6
, and = 2.79
.
and turbulence modeling play a signicant role in the correct simulation of the
ow. The airfoil geometry is the BAC3 11 in high lift conguration (L1/T2)
as described in [29]. The geometry and the multiblock grid used for the numerical
simulation is shown in Figure 19.
Figure 20 depicts the pressure coefcient distribution on the surface compared
to experimental data from Moir [29] at Ma
= 0.198, Re
c
= 3.52 10
6
,
and = 4
= 0.198, Re
c
= 3.52 10
6
, and = 4
determined
by the new model and the SA model.
4.7. SENSITIVITY ANALYSIS OF THE MODEL
An analysis of the sensitivity of the results to the freestream turbulence and to grid
resolution was done for the newmodel following the methodology in [3]. The value
of the dimensionless freestream eddy viscosity can be dened as
t
/Ux = 10
n
,
where the exponent n is varied within the range of 13 n 0. Furthermore,
the percentage of deviation of the spreading rate value S compared to that obtained
on the nest grid resolution 100(S S
1000
)/S
1000
can be evaluated for different
resolutions ranging from 1001000 grid points. The investigations are shown ex-
emplarly for the mixing layer ow in Figure 22. As can be easily seen the new
model shows a consistent behavior almost independent of the freestream value and
the grid resolution when compared to the k- or the SA models. Investigations of
other types of ows conrm the same behavoir of the new model.
A GENERAL ONE-EQUATION TURBULENCE MODEL 211
Figure 21. Computed Mach number contours for the BAC3-11 airfoil in high lift conguration
at Ma
= 0.198, Re
c
= 3.52 10
6
, and = 4
.
4.8. NUMERICAL ROBUSTNESS AND EFFICIENCY
The numerical implementation of the proposed model is very similar to that of the
SA model. From a computational point of view the new model possesses slightly
more complicated terms that require additional computational time. However, the
new model doesnt require the wall distance. The startup distance calculation is
therefore not necessary, which can be expensive in the case of complicated 3D
geometries [11]. Using the same stability CFL limit and achieving almost similar
convergence rates during the calculation the difference in the CPU time between
simulations using the SAand the proposed model was within 4%. Comparisons with
the k- model conrm that this two equation model has more restrictive stability
limitations and considerably higher demands on the computational resources and
grid resolution especially near the wall.
212 E. FARES AND W. SCHR ODER
Figure 22. Sensitivity of spreading rate to freestream turbulence (left) and grid resolution
(right) for mixing layer ow.
5. Conclusion
Anewone-equation model based on the k-model is derived. The turbulence model
is validated via numerous analytic and experimentally investigated ows to ensure
the generality of the numerical results. The ndings show the novel one-equation
turbulence model to predict a wide range of ows especially jets and vortical ows
more accurately than the SpalartAllmaras model while retaining the same quality
of results for near-wall ows and to be more efcient than the k- two-equation
model.
Acknowledgments
This work has been performed within the collaborative research program401 at the
Aachen University funded by the Deutsche Forschungsgemeinschaft (DFG) under
grant SFB401/TP A5.
Notes
1. Here we chose to eliminate k from the equations. Other choices are also possible but not described
here.
2. Note the relation C
c
=
2
b
= 0.09 with C