Professional Documents
Culture Documents
net/publication/255608559
CITATIONS READS
49 777
1 author:
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Jonas Per Bredberg on 14 April 2016.
JONAS BREDBERG
JONAS BREDBERG
Department of Thermo and Fluid Dynamics
Chalmers University of Technology
ABSTRACT
This report explains and discuss two main boundary condi-
tions for turbulence models: the integrating and the wall fun-
ction approach. The report thoroughly derives the law-of-the-
wall for both momentum and thermal field. The deviation of
these laws from DNS-data is discussed. The Wilcox tur-
bulence model constants are derived on the basis of the law-
of-the-wall and experiments. Using a simplistic treatment,
the low-Reynolds number modifications (damping functions)
to the model are explained.
Included in the report is also a guideline on the implemen-
tation of wall functions, both the standard and the Chieng-
Launder two-layer variant.
A new boundary condition for turbulence models combining
wall function and integration approach is presented.
ii
Contents
Nomenclature iv
1 Introduction 1
2 Near-wall Physics 1
2.1 Buffer Layer . . . . . . . . . . . . . . . . . . . . . 1
2.2 Viscous sub-layer . . . . . . . . . . . . . . . . . . 1
2.3 Inertial Sub-layer . . . . . . . . . . . . . . . . . . 2
3 Law-of-the-Wall 3
3.1 Momentum . . . . . . . . . . . . . . . . . . . . . . 3
3.2 Energy/Temperature . . . . . . . . . . . . . . . . 4
3.3 Turbulence . . . . . . . . . . . . . . . . . . . . . . 6
3.4 Turbulence Model Constants . . . . . . . . . . . 7
3.4.1 Coefficient, . . . . . . . . . . . . . . . . 7
3.4.2 Coefficient, . . . . . . . . . . . . . . . . . 7
3.4.3 Coefficients, . . . . . . . . . . . . . . 7
3.4.4 Coefficient, . . . . . . . . . . . . . . . . . 7
3.5 HRN Turbulence Model . . . . . . . . . . . . . . 8
iii
Nomenclature
Greek Symbols
Latin Symbols Thermal diffusivity +,-./
Turbulence model constant Turbulence model coefficients
Specific heat
Turbulence model coefficients
Various constants k Turbulence model coefficient
Length-scale constant Characteristic length scale "
8 >@ /4.
Velocity North face value
~
8d>@ 8dB> Fluctuating velocity /4. ` Node value
G45 Turbulent quantity
8 X Reynolds stresses _ Thermal quantity
8 X Friction velocity, e .
]gfWfhfji 8 X W W4. w Viscous sub-layer value
Normalized value:
D Streamwise coordinate " Wall value
U Wall normal coordinate " Quantity based on the
friction velocity
iv
1 Introduction 2 Near-wall Physics
One of the most common engineering problems is com- The near-wall region may be sub-divided into three dif-
puting turbulent flows that are influenced by an adja- ferent areas, Tennekes and Lumley [31]:
cent wall. Examples of this are flows in turbomachinery,
around vehicles, and in pipes. The main two effects of a viscous sub-layer
f U s
wall are:
buffer layer
U s hf
Damping the wall normal components, making the
turbulent flow anisotropic.
inertial sub-layer
hf U s Wf f
Increasing the production of turbulence through the
shearing mechanism in the flow. The turbulence is negligible in the viscous sub-layer,
while the viscous effects are small in the inertial sub-
The wall gives rise to a boundary layer, where the ve- layer. In the buffer layer, however, both turbulent and
locity changes from the no-slip condition at the wall to viscous effects are of importance, see Sahay and Sreeni-
its free stream value. The variation is usually largest in vasan [29] and Fig. 1.
the near-wall region, and hence the strongest gradients
are found here. Similarly, for heat transfer applications, 1.5
there exists a thermal boundary layer with equally large Total shear
gradients. Because both heat transfer and friction are Turbulent shear
computed using gradients of the dependent variable, it Laminar shear
is very important to accurately capture this near-wall
variation. 1
The standard method is to apply a very fine mesh close
to the wall, to resolve the flow. This method is called the
integration method, which necessitates an LRN (low-
Reynolds-number) type of turbulence model. At higher
Reynolds numbers, the region under the wall influence 0.5
diminishes. However, because it is equally
PSfragimportant to
replacements
accurately capture the near-wall gradients, a large num-
ber of nodes are then necessary. From an engineering
point of view, this becomes inconceivable and a function
that bridge the near-wall region is instead introduced, 0
thereby significantly reducing the computational requi- 0 10 20 30 40 50
rements. The anticipation is that this can be done with
only a small deterioration in the results. This latter ap- Figure 1: Laminar and turbulent shear in the near wall
proach is denoted the wall function method, which uses region. Channel flow, DNS-data, H !X t [25].
an HRN (high-Reynolds-number) type of turbulence mo-
del.
1
0.25 10
, DNS
0.2 Production 9
Model
Dissipation ,, DNS
0.15 8 , Model
, DNS
0.1 7 , Model
, DNS
0.05 PSfrag replacements 6 , Model
0
5
−0.05 4
−0.1 3
acements −0.15 2
−0.2 1
−0.25 0
0 10 20 30 40 50
0 2 4 6 8 10
0.25
20
0 8
acements
6
−0.05
4
−0.1 2
PSfrag replacements 0
0 10 20 30 40 50
Us
10 20 30 40 50 60 70 80 90 100
s t
8 XWM o U
lues. In light of this, a plausible turbulence model would
be to assume a variation in both the velocity profile and R (8)
2
where o
t f ] ,
t f f and t are justified
3 Law-of-the-Wall
by DNS-data. For air, Kays and Crawford [20] give the
following values: o
t 9E N o t f and t in The variation in the different parameters in the loga-
the relation for the temperature. The inertial sub-layer rithmic region, postulated in the previous section, is ba-
is also denoted the logarithmic (log) region, due to the sed on the law-of-the-wall and its consequences. Be-
above characteristic profiles, which feature a logarith- low, the law-of-the-wall for both momentum and tem-
mic behaviour. perature are derived. The latter is an extension of the
The models assumed for the inertial sub-layer are former, based on the modified Reynolds analogy. The as-
compared with DNS-data in Fig. 5. There is excellent sumed variation in the turbulence quantities can only
agreement for s
tions beyond U
both
Wf the velocity and temperature rela-
. The turbulent quantities are ho-
be proven, see section 3.3 and 3.4, in conjunction with a
specific turbulence model. In this paper the Wilcox
wever less accurate, and the assumed constant values turbulence model is used. The constants included are
for the turbulence kinetic energy are not correct. At derived in section 3.4.
U s t ]gfhf there is a descrepancy of around 25% when
using either Eq. 7 or 8 as compared with DNS-data.
3.1 Momentum
Assuming that a Couette-like flow prevails, the
following simplifications can be made to the momentum
equations:
1D flow, with
rGt $ variation only in the wall-normal
direction: PAUQ .
Fully developed flow with zero gradients in the
streamwise direction; C C?D
t f , apart from a
pressure gradient: ~ P D )UQ .
Negligible convection.
f t C 9 ^ w > w >
C U U
(10)
f t 9 U ^ q 8 > w >
D U U U
9
f t
U ^ q
8 > w >
z
D U .
(12)
3
Note that
8 >w >
is negative and hence the left-handside 25
(LHS) of Eq. 14 – i.e. the total shear stress – linearly
DNS
decreases from the wall to the centre of the channel.
Linear
This is of course in excellent agreement with DNS-data 20 Logarithmic
for fully developed flow, see Fig. 1.
8
If Eq. 14 is normalised with the friction velocity, X ,
and a length-scale based on wall quantities: 15
s U 8 X s
U (15)
M
10
The relation becomes:
s s
s
8 w s t ]
U
H !X PSfrag replacements
(16)
U 5
s
10 10 10
total shear is constant. Hence the velocity
s
stress can be written as a function of U only:
and shear
AP U Q
s s
t $ PAU
Figure 6: U-velocity profile with linear and log-law ap-
Q (17) proximation. Channel flow, DNS-data, H ! X t [25].
s
8 w s t d
PAU Q (18)
These relations are called the ’law-of-the-wall’ expres-
sed in inner variables, as proposed by Prandtl [27]. lue of
where is an integration constant, which is given a va-
based on experimental data. The logarith-
If the inertial sub-layer is instead approached from mic law or log-law for the velocity is thus:
the core of the flow, and the equation is normalised with
s t ]
f ] PAU Q ^
the mean flow length scale: s
(25)
n *U (19)
The velocity
s t profile using this relation and the linear re-
and the friction velocity, the following relation is given: lation, U s , is compared with DNS-data in s Fig.t ]h6.] ,
] s 8?w s t ] The intercept between the two laws is found at U
H ! X n n (20) located in the buffer layer, which is, as is obvious from
the figure, not accurately modelled by either of the equa-
For large
H
! X the first term on the LHS can be neglected tions.
and thus the relation reduces to:
8 w s t ( P n Q
0.01 1
(21)
Note however that this does not give a relation for Prt=νt/αt →
itself. On the basis of purely dimensional grounds, it
is possible to establish a relation for the velocity pro-
file, similar to Eq. 17, although as a function of the wall
distance in outer variables only, van Karman [33]: 0.005 ← αt 0.5
8 t & P n Q
, ← νt
X (22)
s s
where is the centreline velocity and is the de-
fect velocity. This relation is called the ’velocity-defect
law’, Tennekes and Lumley [31]
Somewhere in the flow it is plausible to assume
PSfrag that
replacements 0 0
0 100 200 300 400
both relations for the -velocity hold. Differentiating sU
Eqs. 17 and 22 and adding a proportional constant as
done by Clauser [9] gives:
$ ]
Figure 7: Diffusivities and turbulent
air. Channel flow, DNS-data !X
H t Prandtl number for
& t
n n U ss t [25], [18].
o U (23)
4
9E s
mechanisms would be similar only if the Prandtl num- U @ N
bers (both the molecular and turbulent) are unity. The Air: 0.7 13.2
definition of the turbulent Prandtl number is: Water: 5.9 7.55
9E N 8 > w > _ w > ` > (26)
U U Table 1: Critical Us values, air and water
The turbulent Prandtl number and the turbulent mo-
mentum diffusivity, M N , and turbulent thermal diffusi-
vity, N , are plotted in Fig. 7. The definitions of M N and N Re-arranging and integrating once more gives:
are:
8 >w > F t q q N _
^
9E 9E N
MN
U
u U (27)
F (33)
w >` > _ t
U 9 E
N _ U (28)
q
9 E ^ q N N
Reynolds, in an attempt to describe the heat exchange On introducing normalised values:
process, proposed a model which is based on similarities
between the momentum and energy transfer. Reynolds’ U
s U8 X (34)
analogy presupposes that the two diffusivities, M N and
M
N , are equal and hence, from _ s P_ _ Q 8 X
9E N t Eq. 26, that the turbulent
] . For F (35)
Prandtl number is unity, a fluid with a unit
9E t ] , this
molecular Prandtl number, would result in and re-arranging, the equation yields:
identical temperature and velocity profiles. The validity
of the Reynolds analogy can be observed in Fig. 7. As _ s t
U s
9E N can
noted the assumption of a constant only be ac-
] 9E ^ P M N MOQ)
9 EN (36)
ceptable in the inertial sub-layer, and with a value less
than unity,
9E N f . _ s s
This gives the variation of as a function of U , alt-
Although the validity of the Reynolds analogy is limi- hough in an integrated form, which is not preferable
ted, there still exists a strong similarity between the ve- when making comparisons with the law-of-the-wall for
locity field and the temperature field, which implies that the momentum equation. The formulation could ho-
there could exist a corresponding law-of-the-wall for the wever be simplified if it were possible to divide the near-
thermal field. wall region into a laminar and a turbulent layer. In
With the same assumptions as for the momentum the case of the momentum equation, the velocity profile
equation, i.e. 1D fully developed flow field and ther- follows approximate a linear relation in the viscous sub-
mal field, the energy (temperature) equation in the wall- layer (laminar) region and a logarithmic relation in the
normal direction can be simplified to: inertial sub-layer (turbulent) region. If the thermal fi-
f t ] F t _ w >` >
eld behaved similarly, the above integral can be divided
U U U (29) and integrated successfully. The location of this parti-
tion depends on the molecular Prandtl number, which
where is the specific heat of the fluid and is the for a value of unity would
(viscous) thermal diffusivity. s tbe ]Widentical to that of the mo-
] . For a lower
mentum equation, i.e. U value of the
9E t fc] ), this critical
Note that the pressure gradient, contrary to the momen- Prandtl number, e.g. air ( value
tum equation, does not appear in this equation. If a non-
negligible pressure gradient were present, much of the
will be higher, with the opposite being
Prandtl number fluid, e.g. water (
true for a
9E t ). Examples high
similarity between the momentum and energy equation s
of these cross-over locations, or critical U values, are gi-
would be lost.
Assuming that the turbulent heat flux,
w > ` > , can be ven in table 1, see also Kays and Crawford [20]. Taking
the value for air, the two integrals would be:
described with the turbulent thermal diffusivity and the
turbulent Prandtl number, the relation can be re-written _ s t [ = 9E s ^
U s
as: U
N 9 EN
(37)
f t
N _ [ = 3P M MQ'
9q E ^ 9q E N
U (30)
U where it was assumed that the turbulent part could be
neglected in the first integral and the laminar part in the
When integrating this equation from the wall to the
centre of the channel, the heat flux is found as: second integral. The first integral is easily integrated to
t
N _ yield the linear-law for the temperature as:
9q E ^ 9q E N (31)
U
_ s 9 EUs
t (38)
[%=
where the integration constant, , is a function of the
However the second part can only be integrated if the
heat flux applied at the wall:
variation in M N M and
9E N is known. In the inertial sub-
t F (32) layer the turbulent Prandtl number can with acceptable
accuracy be assumed constant. The ratio M N M , or the
5
turbulent Reynolds number, can be approximated using model (see section 3.5) reduces in the inertial sub-layer
the mixing-length theory, which gives: to:
MN t 8 XoU
MN t oUs (39)
M f t M N
(42)
Adopting this and evaluating the integral yields:
U U
t
_ s t
9 EN
U s t
9 EN
U] s
f M N U ^ U M N U
(43)
o s o (40)
[%=
[%= U f t
MN
^ (44)
Thus the law-of-the-wall for the thermal field, or the U U U
temperature log-law, is, adding Eq. 38 and 40 MN t (45)
s
_ s t ] 9E ^ 9E N U t f s ^ SP U Q
o ]
where the laminar viscosity in the diffusion terms is
(41)
neglected. Note that the neglected pressure gradient in
the momentum equation (Eq. 42) is a severe limitation
where the constants are found by using properties for air
9E t f and 9E N t f ) and the van Karman constant,
( t in the following derivation. This makes the wall func-
o f ] . tion approach even more questionable for the turbulent
The two formulations, the linear-law and the log-law, quantities than for the velocity field.
are compared with DNS-data in Fig. 8. Similar to the The above equations can be satisfied in the inertial
momentum (velocity) model, the agreement is less ac- sub-layer with the following set of wall functions, see
curate in the buffer layer where neither equation gives Wilcox [36]:
acceptable performance. For
the two laws is found at U
s t air,] the cross-over between
, and hence the lami-
8 X U8 X
t ^
nar thermal layer extends further out than the laminar o M (46)
momentum layer.
t 8 X
V (47)
20
t 8 X
V o (48)
8 X o U U
18 DNS
MN t
Linear
16 Logarithmic (49)
14
12
These relations can be proven by re-inserting them into
_ s the above transport equations. Using Eqs. 49 and 46 in
10 Eq. 42 yields:
8 X o U 8 X U 8 X ^ t
8
f t
U o M
6
acements U
4 8
t 8 XWo U X ] t 8 X t f (50)
2 U o U U
0 0 1 2
Thus these wall functions satisfy the momentum equa-
Us
10 10 10
tion. Eqs. 46, 47, 48 and 49 are substituted into Eq. 43
which gives:
Figure 8: Temperature profile with linear
H and
approximation. Channel flow, DNS-data, !WX
t log-law
[18].
f t 8 WX o U
8 X U8 X 8 X 8 X
o V ^
o U V
U M
3.3 Turbulence 8 8 X t
^ WX o U
There is a number of different approaches to the turbu-
U U V
8 X ] 8 X= f t 8 X= 8 X= t f
t 8 XoU
o U oU ^
lent law-of-the-wall, or log-laws, where the derivations
below mirror that of Wilcox [36]. If 1D fully developed oU oU
flow is assumed, and the pressure influence is neglected (51)
(even in the streamwise direction) the Wilcox 1
HRN
The Wilcox
1 Note that the log-law may be derived for any turbulence model.
6
using the wall functions, Eqs. 46, 47, 48 and 49, gives: from Eq. 47 and the measured relation of above, Eq. 53,
f t 8 X U8 X 8 X
we can write the coefficient as:
^ 8 X
8 X t f t f V
o U V o M U
8 X 8 X ] (55)
^ 8 XWo U V
t o t f f
o U
U U U
8 X 8 X ]
t
o U ^ U
8 hX o U V o U 3.4.2 Coefficient,
t 8 X 8 X ] t
^ By studying how grid-turbulence dies out, it is possible
oU V
U to determine the coefficient, . In decaying turbulence,
t ^ o t f the turbulence equations simplify to:
V
V t
o t (56)
t (57)
(52)
To satisfy the last relation, it is necessary to establish Solving this equation system gives:
an identity between coefficients o , , , and . I I
Z ' x (58)
3.4 Turbulence Model Constants Measurements [32] indicate that the turbulence kinetic
energy decays as:
0.3
} t ] ,f f
Z (59)
0.25
and hence the ratio g
t ] f f
. With
t f f
from above, the correct value for would be:
t f f
f f
0.2
0.15
The turbulence model by Wilcox [34] gives this
coefficient the value of
t f t f f .
0.1
0.05
3.4.3 Coefficients,
acements There is no consensus within the turbulence modelling
0 community about the values for these coefficients. Diffe-
0 20 40 sU 60 80 100
rent Schmidt numbers are arrived at depending on the
set of experimental data used. In recent years there has
also been a greater tendency to introduce modifications
Figure 9: Shear stress to turbulent kinetic energy.
Channel flow, DNS-data,
H -
! X t [25]. for these
coefficients, and the values used in turbulence
models thus differs. The two Schmidt numbers for the
model [34] are optimised to:
3.4.1 Coefficient,
t ]
Experiments have shown that the shear stress and the
]
t
turbulent kinetic energy is related as
f (53)
3.4.4 Coefficient,
in the logarithmic region. The validity of this approx-
imation is shown in Fig. 9. If the shear stress can be The final coefficient, , is set using the deduced relation
assumed to be constant and the laminar part is neglec- from the specific dissipation rate equation, Eq. 52. With
ted, then:
o t f ] and the above constants, is calculated as:
t
^ N t . 8 X (54) o t f f f ] f t f
t
f f
V f f
V
(60)
where the friction velocity is defined as X e . . 8 In the Wilcox model is given the value of
t
Now, using the relation for the turbulent kinetic energy
f
.
7
3.5 HRN Turbulence Model 4 Modelling Near-wall Turbu-
On the basis of the above derived turbulence constants lence
it is possible to write the Wilcox
model as follows:
MN t
With respect to wall treatment, there are two different
(61) categories of turbulence models. The first are the LRN-
models, which use a refined mesh close to wall in order
C C
to resolve all the important physics. The second met-
t 9 ; ^ C?D B P3M ^ M N Q ?C D B (62) hod employs the HRN-models, which bridge the near-
wall region using wall functions. The latter approach is
of course less demanding of computer resources, but a
t 9 ; C ^ N C significant amount of information is lost.
^
?C D B P M M Q ?C D B (63) The wall function method is based on the previously
mentioned ’law-of-the-wall’ and is strictly valid only in
with the constants as derived earlier: the inertial sub-layer (log-layer), 1D fully developed flow,
t t ]gfhf t f
t ] W t ]
where the pressure gradient can be neglected. The situ-
ations in which all these simplifications are fulfilled are
relatively few, and the predictions made using wall fun-
ction or HRN turbulence models are thus generally less
accurate than those that apply the integrated method of
LRN turbulence models.
The reason why there is still an interest in wall func-
tions is the need to reduce computational requirements,
as well as the better numerical stability and convergence
speed. If it were possible to improve the predictions
using wall functions, more problems of engineering in-
terest could be solved with academic quality, resulting
in fewer safety margins and less need of experimental
validation when designing new products.
The previous section derived the commonly used wall
functions based on the law-of-the-wall, while this section
will focus on alternatives to these wall functions.
However, before embarking on strategies in con-
structing wall functions, it is beneficial to study the
near-wall physics from a modelling point of view. Atten-
tion is first directed towards the LRN-models and their
damping functions.
8
section and the LRN model [35] is: part of the buffer layer, while the HRN and LRN
formulations become similar in the inertial sub-layer.
MN t (64) In the limit of !N
H t f , i.e. at the wall, the coefficients
H H ;
t ^ ! N for the LRN and HRN formulation give:
] ^ H !NH ; (65)
LRN HRN LRN/HRN
0.025 1 0.025
t 9<; C C
0.025 0.09 0.28
^ C?D B P3M ^ M N Q ?C D B (66)
2.22 0.55 4
t ] ^ H !gN HuI v
0.075 0.075 1
]gfhf ] ^ P H P !N H ! I Q v Q (67) P
Q 0.055 0.55 0.1
t ] g] f t < t f t ] W t ] W
DNS
0.09
the damping functions, Eqs. 65, 67 and 69, are set equal 0.06
]
to ,
g
] W
f f and , respectively.
0.05
4 0.04
0.03
3.5 DNS:
PSfrag replacements
0.02
DNS:
3
0.01
2.5
DNS:
0
Us
0 20 40 60 80 100
’s
2
acements
Figure 11:
-coefficient, LRN and DNS
H !-X t
1.5
[25].
1
0.5 4.1.2
modification,
The most important damping function is the damping
of the turbulence kinetic viscosity, M N . This is done by
0
Us
0 5 10 15 20 25 30
9
, while the turbulent viscosity is defined as in Eq. 2
64. The effective
thus becomes: DNS
t
1.8 LRN
HRN
(72) 1.6
f f . The
is the value of in the logarithmic region, simplest correction is to reduce the destruction term in
Eqs. 46 and 48, were used in the the dissipation rate equation. This is usually done in a
log-laws,
relations for u U and , respectively. The approximate R model by introducing a damping function $ , see
Patel et al. [26].
identity is correct only in the inertial sub-layer for which
the log-laws are valid. In the near-wall region, U
s There is more to the damping functions than balan-
]f , the above relation may only be used qualitatively. cing the equations, however. As previously discussed, it
The a priori computed production-to-dissipation ratios is also important to predict an accurate production-to-
for the HRN and LRN formulation are compared with dissipation ratio, which puts an t additional demand on
DNS-data in Fig. 12. the actual values of and R . In a mo-
The LRN formulation, similar to the DNS-data, incre- del, where the dissipation term is a function of both
ases the production in the buffer layer, which results in and , it is unrealistic to believe that 9merely ; R , asmodifying
a
f higher peak of for the LRN-model as compared the
-equation would yield a correct well as a
with the HRN-model. Slightly dependent on the Rey- correct level of . Hence, apart from modifying the speci-
fic dissipation rate equation, the dissipation term in the
nolds number, DNS-datas give
netic9 ; energy at around U
t ] the peak of turbulent ki-
, close to the maxima of -equation is also modified. In the LRN model
these
the R -ratio given in Fig. 12. two aspects are accomplished through both an increase
of production of (increasing ), and a reduction of the
dissipation term in the -equation (decreasing ). Note
4.1.4 Rationale behind the and coefficients
however that, contrary to the JR model, the destruction
As was previously shown, the damping function, , term in the secondary equation P R Q is unchanged.
was introduced to reduce the turbulence viscosity in the Considering Patel et al. [26], it seems likely that the
near-wall region. This also effectively reduces the pro- minimum number of damping functions required for an
duction terms in the - and -equations. In order to pre- Eddy-Viscosity Model (EVM) to correctly predict the tur-
dict the same level of turbulence, the dissipation term bulent kinetic energy, , in the near-wall region is three.
in the -equation must also be reduced. This can be ar- In that paper the only model that accurately reproduces
ranged in three different ways: the profile of is the Lam-Bremhorst [21] ,R model,
10
which is one of only two models tested employing three The turbulent kinetic energy is set in the first node by
damping functions. iteratively computing the friction velocity from the law-
of-the-wall. This iterative process is needed because X
8
appears implicitly
8 X
4.2 HRN-models: Wall Functions
Eq. 47, as X
8 t :[)inx v the log-law. Initially is set from
a new X
8
e . Using this relation,
The rationale behind wall function is the reduced com- is found from the log-law:
putational requirement and the increase numerical sta-
8 X t os s 8 X
bility and convergence speed. By adopting a mesh, # where U t U M (78)
where the first interior node is located in the inertial P U Q
sub-layer, it is possible to use the law-of-the-wall to spe-
The new value is then used to compute U and re-fed
s
cify the boundary condition for the dependent variables
8 , and . 8
into the X expression. This process is repeated until
convergence, with the turbulent kinetic energy finally
set as:
4.2.1 Standard wall functions
t 8 X
In its simplest form, the logarithmic laws, Eqs. 46, 47, V (79)
48 are directly applied to the first interior node. These
wall functions are here referred to as the standard wall
For the specific dissipation rate, , the friction velocity,
8 X , is substituted by the relation for in the log-law as:
function method. The three steps involved in a CFD-
code would then be:
t 8 X t :[)x v e t e -
Solve the momentum equation with a modified wall V o -U V o -U : [)x v o U
(80)
friction, either through an added source term or via
a modified effective viscosity.
4.2.2 Launder-Spalding methodology
Set at the first node iteratively with the use of the
law-of-the-wall. The standard wall function method is very limited in its
usage. This is especially true for re-circulating flows,
Set with .
where the turbulent kinetic energy becomes zero in se-
In a turbulent boundary layer, the strongest velocity parating and re-attachment points, where, by definition,
8 X is zero. This singular behaviour severely deteriora-
gradient is found close to the wall. With a wall-
function based turbulence model, which utilises a relati- tes the predictions of, for instance, heat transfer in rib-
vely coarse mesh, it is impossible to resolve these near- roughened channels. Launder-Spalding [24] proposed a
wall gradients. The predicted wall friction would thus modification to the standard wall function method that
be largely in error if a modification is not introduced: involves the following steps:
l
C
. q l U t q U
q C
U
(74)
Solve the momentum equation with a modified wall
viscosity.
where the subscript ~ is used for the first interior node. Solve the turbulent kinetic energy, with modified in-
The necessary modification could either be made tegrated production and dissipation terms.
through
Set R using the predicted .
(i) an added source term simulating the correct wall
friction or The same reason and method of specifying the wall
viscosity as for the standard wall functions
8 X in theapply in this
(ii) a modified viscosity, an effective viscosity, q , that case. However, instead of using log-law, the
8 t
[)x v V is utilised:
ensures the correct friction even though the velocity identity X
gradient is erroneous.
v
Through the law-of-the-wall: t
%[ x e U o
] q # (81)
s P U Q
8 X t o P # U Q (75)
where U is defined with as:
[)x v V
the wall friction is computed as
. t 8 #X o U t U
(82)
s (76) M
P U Q
with 8 X . In the -equation, the integrated production term is de-
fined from (1D):
For a wall function based model
directly using a source term,
Y t .this can either be set
, or via a modified 9 ;U NC U
effective wall viscosity [10]: C U (83)
t t 8 X o t 8 XUo
q U . # s q # s (77) In the inertial sub-layer the laminar shear stress,
P U Q P U Q q C J C U , is negligible and, by assuming that the shear
11
stress is constant, the turbulent shear stress, N , is equal
to the wall shear stress: north
PSfrag replacements
t ^ N t g
t f
N t .
(84)
P
The accuracy of this assumption is shown in Fig. 1. The
vs
production can be set using the law-of-the-wall:
9 ;U t C U t l t q l t wall
C U U .
v V o (85)
t Eq. 81 t
[)x # l
Figure 13: Assumed variation of variables in the first
P U Q computational node, Chieng-Launder model. vs=edge of
l
The integration of P C u C UQ U is equal to , viscous sub-layer.
since is constant.
To set the value for the dissipation rate, production is
Comparing the two formulations, Eqs. 88 and 91, gives:
assumed to be equal to dissipation:
9 ; t R l f t
#
(86) P U Q P Q P U Q (92)
Eq. 88 Eq. 91
The production can be re-written as:
9 ; t 8 > w > C t 8 X 8 X The latter formulation is used in the Launder-Spalding
C U oU (87) model, with the integrated dissipation rate in the -
equation set as:
where the constant shear stress approximation, Eq. 54, v
8 t 8 > w > t 8 X , and the law-of-the-wall, C u C U t
R U
t
=:x ='x' #
X
P o U
Q , were used. Integrating the dissipation rate o P U Q (93)
l
between the wall and the edge of the first cell, U , yi- 8 v
elds: where X= in Eq. 91 has been substituted by
='x
='x' .
v
8
X= t 8 X= Note that in [24],
is used rather than
=:x , although
t l f
R U
o
U U o P UQ P Q
(88) this is most probably a typographical error.
The dissipation rate equation is not solved for the first
The mathematical singularity at the wall of this equa- interior node, and R is instead fixed according to:
f
tion ( P Q ) necessitates re-consideration. The short-
t 8 X= t
=:x v =:x'
coming of the above derivation is the assumption that R oU oU (94)
the law-of-the-wall is valid all the way from the wall,
which is erroneous.
However, by choosing a different relation for the in- which is a direct consequence of substituting the friction
tegrated R , other approximations could be found. The velocity with the log-law for turbulent kinetic energy,
8 X t
[%x v V , into Eq. 86.
integrated velocity gradient in the production term can,
according to the mean value theorem, be re-written as:
4.2.3 Improvements of the near-wall representa-
C t l C tion, Chieng-Launder model
C
U U U C U (89)
It is evident from the above derivation that there is room
for a certain value of C J C U
. Approximating
l l the velocity for improvement in the Launder-Spalding model, espe-
gradient with discrete values, u U , and, assuming cially in its representation of the near-wall region. Chi-
that this value could be established in the inertial sub- eng and Launder [7] enhanced the model by introdu-
layer, the velocity gradient could be re-written with the cing a more complete model of the near-wall turbulence.
help of the law-of-the-wall: They argued that there exists two distinct regions, the
l 8
C l t l X P # U Q viscous
different
sub-layer and the inertial sub-layer, which have
turbulent structure. The partition line between
C U U o U (90)
these regions is defined by a viscous sub-layer Reynolds
Hence the integrated dissipation rate can also be written number:
as:
H !T t UTV T (95)
t 9<; U t Eq. 87 t
8 > w > C U t M
R U
C U The near-wall variation in the variables assumed in the
t t 8 X l U l 8 X # 8 X
t = #
Eq. 90 o U P U Q o P U Q
Chieng and Launder model is shown in Fig. 13.
The turbulent shear stress is zero for U U T and is
(91) linearly interpolated between the wall friction and the
12
shear stress at the edge of the cell for U
UT
where the new constants, o : and , are
t f # t f
N t f U U T related to the standard values, through the identities:
o t o
[)x v and # t #
[%x v . For local equilibrium flow,
t U t 8 X , and with the above identities, the modified
N ^ PS
z Q U
U UT
(96)
e
R o U U
UT In the second part, the velocity gradient is re-written
using the law-of-the-wall:
The turbulent kinetic energy gradient in the first rela- C
PS
Q U U
C U U t
tion is approximated using the previously assumed vari-
13
H !T t fis a good partition value, since the two formula- viscous sub-layer in order to accurately predict the wall
tions become a laminar and a turbulent approximation, friction and heat transfer. To be able to significantly im-
respectively. prove the results, it is necessary to mimic the LRN met-
However, in a subsequently paper by Johnson and hod of resolving the viscous sub-layer, especially in non-
Launder [14], it was found that the heat transfer predic- equilibrium regions. Thus the wall function approach
tion could be improved by letting the viscous sub-layer of bridging the viscous sub-layer seems unattainable,
thickness, i.e. the sub-layer Reynolds number, vary ac- apart from regions in which the flow is in equilibrium.
cording to the turbulence level: Hence the only possible path to reducing the computa-
H !T t H !T H !T T W
tional demands and maintain accuracy would be a tur-
T (107) bulence model that adopts itself to the flow, employing
a LRN type model where necessary and switching to a
where
t f
, and h is the linear extrapolated HRN type model otherwise. While such artificial intel-
turbulent kinetic energy at the wall. ligence may sound incredible it is possible to achieve it
Ciofalo and Collins [8] used an approach however through a smartly devised model.
where they proposed to let the sub-layer thickness vary, The second problem is associated with the underlying
dependent on the turbulence intensity. physics of the wall functions, i.e. the law-of-the-wall.
When employing wall function, it is necessary to apply
4.2.5 Extensions to the Chieng-Launder model the first computational node in the inertial sub-layer.
This fact is less of a problem in equilibrium flow but it
Amano et al. ([2], [3]) devised two different models ba- is accentuated in regions of flow separation 2 , where the
sed on the same zonal principle as the Chieng-Launder wall shear stress and hence U s decreases. For these re-
model. In both models, the dissipation rate equation is gions, the size of the first cell must be enlarged to ensure
solved using the layered approach rather than allowing that the cell node is kept within the logarithmic region.
it to be set as in the Chieng-Launder model. The results Hence the mesh generation becomes a delicate business
were only marginally improved however, and hence the using wall functions for re-circulating flows.
two-layered approach was extended in the second model
Both of the above explained problems are solved with
to a three-layered version. Here the parameters, R
the two-layer Chieng-Launder model. The CL-model,
were allowed to vary in accordance with measured data
however gives rise to additional complexity. Through
for the viscous sub-layer, the buffer layer and the iner-
prolonged mathematical operations the viscous sub-
tial sub-layer.
layer and the inertial sub-layer are modelled simultane-
ously, although the author encountered numerical dif-
4.3 New Model: Blending Integrating ficulties using this approach. If the two regionscould
and Wall Function Boundary Condi- be separate mathematically, a numerically simpler and
tions more robust wall function could be constructed.
Such a turbulence model will be presented here. A less
The physical treatment improved, w hen progressing complete derivation is given in Bredberg et al. [6], where
from the Launder-Spalding model to the more advanced predictions of heat transfer in rib-roughened channels
models, although at the expense of increased numerical are included. The reader is referred to the paper for a
complexity. The modifications by Johnson and Launder discussion of the accuracy and for comparisons between
[14] and Ciofalo and Collins [8] showed that the impro- the model and other turbulence models.
vements with the zonal approach were not particularly
Dependent on some flow quantity, the model will eit-
great if a variable sub-layer thickness was not included.
her employ a wall function or use the integrating method
The further refinement by Amano et al. enhanced the
as a wall boundary condition, and hence greatly simp-
predictions only marginally, with a significantly increa-
lify the numerical scheme as compared to the Chieng-
sed complexity.
Launder model. The new model will, like the wall fun-
Thus, here, instead of further optimising the layered
ction, approach reducing the computational demand –
models, a new strategy will be adopted.
by coarsening the mesh, when possible, and still be able
Two main problems with wall-functions must be add-
to produce accurate results through the use of the LRN
ressed to make such an effort worthwhile:
part of the model, when necessary. The physical conse-
The negligence of the physics in the viscous sub- quence of separating and re-circulating regions can ac-
layer, tually be utilised with such an adopting turbulence mo-
del. Ins non-equilibrium nodes where the physics reduces
The necessity of locating the first computational
the U value, the hybrid model automatically switches to
node in the inertial sub-layer
the LRN mode, producing accurate results even for a re-
The first problem has been discussed in numerous pa- lative coarse mesh. Thus even a uniform mesh could be
pers, see e.g. Launder [22]. The conclusion made in sufficient for complex re-circulating flow, and, as a re-
these studies is that the standard wall function formu- sult, reduces the workload of mesh-generation substan-
lation (HRN-models) are indisputably inferior to the in- tially.
tegrating formulation (LRN-models). This is particu-
larly true for non-equilibrium flows, where it is of cri- 2 Although the log-law in itself is rather questionable in such cir-
14
4.3.1 Blending function improves predictions for non-equilibrium flows. This is
utilised by the Launder-Spalding model and subsequent
The critical part of a hybrid turbulence model is the con- models. In addition the HRN part of the new model as-
struction of the function that effectively changes the mo- sumes a variation of the near-wall turbulence similar to
del into either LRN or HRN mode. It is numerically what is done in the Chieng and Launder model. Ho-
undesirable to adopt a function that abruptly switches wever, since the viscous sub-layer is treated separately
between the two different modes. It is also physically from the HRN part in the new model, a simplified model
questionable, because the flow is only strictly laminar
in the immediate near-wall region – enabling an LRN
is sufficient here. The model is also adopted to a -
type turbulence model.
model to be used – while only in the inertial sub-layer The assumed variation of the dependent variables for
can the flow be treated as fully turbulent. The latter is this model are shown in Fig. 14. Both the turbulent
of course a requirement for using a wall function-based kinetic energy and the shear stress have lost their li-
HRN-model, see Fig. 1. In the buffer layer it is thus most near variation in the inertial sub-layer, as compared to
appropriate to employ a smoothing function that blends the Chieng-Launder model. This of course simplifies the
the two formulation together. One of the simplest mat- treatment immensely. The specific dissipation is assu-
hematical functions that accomplishes this is the expo- ]
med to vary either according to the log-law, U , or the
nential function, which can be designed to operate from ]
linear-law, U .
f to ] dependent on some flow quantity. On the grounds
As was stated above, only the region above the viscous
of the above discussion, it would be s most beneficial to sub-layer is of interest in this section. The integrated
base the exponential function on U . However owing to
numerical issues, it is disadvantageous to include any gion, i.e. from U T
production of turbulent kinetic energy in the HRN re-
U U
, is found as:
wall distance, and hence the blending function is ins-
C
9 ; U t
8 > w > C U U
tead constructed using the turbulent Reynolds number,
H !N t M N M as:
P
TQ (110)
PSfrag replacements
4.3.2 HRN part: Simplified Chieng-Launder mo- 0.2
del applied to a
-type turbulence model
8
The logarithmic equations based on X , Eqs. 25 and 41 0
0 10 20 30 40 50
are unsuitable in separated flows. This applies especi-
ally for heat transfer since the predicted Nusselt num-
ber becomes zero in separation and re-attachment points Figure 15: The accuracy of the assumption, N.
using these equations. In contrast, experiments indicate Channel flow.
a maximum level of heat transfer at or in the vicinity of
8
these points (Launder [23]). Substituting X with
V the turbulent production is negligible in the near-wall
15
region; hence any discrepancy comparing to the Chieng- 10
Launder model is a consequence of the simplistically as-
9
DNS
LRN
sumed variation in the turbulent shear stress. At this New, sum
moment it may perhaps be of interest to compare the 8 New, LRN part
turbulent shear stress given by DNS-data, Fig. 1, with New, HRN part
7
the assumed variations of the two models. The assump-
T
tion of a constant turbulent shear stress beyond U (new
model) is perhaps not any worse than the linear varia-
6
tion from the wall shear to edge of the cell in the Chieng-
5
Launder model. Note especially in Fig. PSfrag
1 thatreplacements
the (lami-
4
nar) wall shear is higher than any given turbulent shear,
3
and hence the linear increase in the turbulent shear
from the wall shear is not physically founded. 2
The integrated dissipation rate of the turbulent kine- 1
tic energy is based on the logarithmic law, Eq. 48:
t V 0
U t
U t
Us
0 20 40 60 80 100
R U
: [%x v o U
t
)=:x v =:x'
U
Figure 16: Integrated Hproduction
t term,
-equation.
o UT Channel flow, DNS-data, ! X [25].
(112)
1
where was assumed constant. 0
Contrary to the Chieng-Launder model, the dissipa-
−1
tion rate is not set using the log-law. Rather, the mean
integrated value from the log-law is used: −2 DNS
t ]
e U t −3 LRN
TZ
U
UT
: [%x v o U
−4
New, sum
New, LRN part
New, HRN part
T (113)
t e PSU
U Q PSfrag replacements
−5
'[)x
v o PSU
U T Q −6
−7
4.3.3 LRN part: Standard −8
l $
C ^ ] $ .
T ^ 9 >
The asymptotic boundary conditions applicable for the
turbulent quantities are:
U M C U
N
P Q U
f
(114) t
M
U (117)
U (115)
v =:x'
l$ M ^ ] $ ) :
= x U
U U - P Q o PSU
U T Q U T
The latter
within U
s relation
is valid only for computational nodes
(Wilcox [36]). To establish a grid inde-
pendent solution, 7-10 nodes are required in this range. where
9 > is the contribution of the non-primary shear
Both requirements can however be eased, especially the stress components (
t u U ). The production term and
latter, with negligible deterioration of the predictions, the dissipation term are shown in Fig. 16 and Fig. 17,
see e.g. Bredberg et al. [5], [4]. respectively.
16
2 3
DNS
LRN
2.5
DNS
LRN
1.5 New New
2
1
1.5
0.5 1
0.5
0
acements PSfrag replacements 0
−0.5
−0.5
−1 −1
Us
0 20 40 60 80 100
Us
0 20 40 60 80 100
0.5
The wall viscosity is set as:
0 o mU V T
M t $ M ^ P] $ Q M ^
acements −1.5 $
To close the model, the blending function, (Eq. 108),
−2 and the viscous sub-layer values, U T and T , must be de-
−2.5
fined. The viscous sub-layer thickness, U T , is calculated
using the definition of the viscous sub-layer Reynolds
−3 number and the assumed variation of the turbulent ki-
Us
0 20 40 60 80 100
netic energy:
t $
M ^ P ] $ Q e PAU
U T Q T and U T .
: [%x v o AP U
zU T Q
(118) and the above routines for the specification of
U
17
1 5 Modelling Near-wall Heat
0.9 Transfer
0.8
In contrast to previous sections, this one is written in
0.7
a brief and and abbreviated style. This is not to imply
0.6 that the thermal field is any easier to predict or that the
$ modelling is any simpler, but is merely a consequence
0.5
of the lesser efforts made to improve the thermal mo-
0.4 dels. To exemplify this, it can be noted that, in some
0.3 predictions, an RSM turbulence model – which involves
seven turbulent transport equations – is combined with
0.2
acements a constant turbulent Prandtl number heat transfer mo-
0.1 del, i.e. a zero-equation isotropic model. The turbulent
0
Prandtl number heat transfer model using the simple
gradient diffusion hypothesis, or SGDH, relies on the si-
Us
0 5 10 15 20
milarity of the turbulence and heat transfer fields. The
$ SGDH assumes that they are identical apart from a pro-
Figure 21: Blending function, , using LRN data
portionality factor, the turbulent Prandtl number. The
validity of this approximation in fully developed chan-
8
nel flow can be seen in Fig. 7. Using the SGDH ap-
7 proximation, it is also implied that the thermal field is
isotropic. With an RSM the individual Reynolds stres-
6 ses are known however, and can be used to predict the
turbulent anisotropic level in the flow. If it is assumed
5
that the anisotropicity of the heat transfer is identical to
r >4 that of the turbulence, a GGDH, or Generalized Gradi-
4
ent Diffusion Hypothesis, model can be employed for the
heat transfer prediction.
3
Although a GGDH would be more appropriate than
acements 2 a SGDH, the fundamental approximation of an identity
between the flow field and thermal field is still used, and
1 should be questioned. Similar to the turbulence field,
it is also possible to solve a set of equations for ther-
mal quantities. Examples of this approach are the two-
` R N , of the Nagano group, Abe et al.
0
Us
0 10 20 30 40 50
equation model,
[1] and EAHF (Explicit Algebraic Heat Flux) models of
Figure 22: Comparison of the new models variation of So and Sommer [30] and Dol et al. [11]. Here, however,
H t
quantities, with those given at !WT
viscous sub-layer
]h] sU
(constant lines).
only the simple model based on the turbulent Prandtl
number is used.
18
0.5 5.3 New Blending Model
0.45 During the numerical optimisation of the new model it
0.4 was found that the most appropriate blending parame-
ter for s the thermal field was the normalised tempera-
0.35 ture,
_ . The non-dimensionalised wall distance in Eq.
0.3 DNS 125, is however based on U U e :[)x v M s instead of
o New model the commonly used although less satisfying U :
_ t ] PAU Q ^
0.25
flow. The above relation for the_ wall lower value, o
t f
s temperature is not
commonly used, however, and s t is instead expressed
.
where
9E t 9 E N , and 9E N t o o , [16]. The Acknowledgments
constant,
9 E , depends on the Prandtl number, which
Funding for the present work has been provided by
according to Jayatillaka [13] is: STEM, the Volvo Aero Corporation and ALSTOM Power
9 E ='x v ] via the Swedish Gas Turbine Center.
9E t
9E N
] ^ f
9 EN
(127)
f fWf 9 E
19
References [13] C.L.V. Jayatillaka. The influence of Prandtl num-
ber and surface roughness on the resistance of the
[1] K. Abe, T. Kondoh, and Y. Nagano. A new turbu- laminar sublayer to momentum and heat transfer.
lence model for predicting fluid flow and heat trans- Progress in Heat and Mass Transfer, 1:193, 1969.
fer in separating and reattaching flows - I. flow fi-
[14] R.W. Johnson and B.E. Launder. Discussion of: On
eld calculations. Int. J. Heat and Mass Transfer,
the calculation of turbulent heat transfer transport
37:139–151, 1994.
downstream an abrupt pipe expansion. Numerical
[2] R.S. Amano. Development of a turbulence near- Heat Transfer, 5:493–496, 1982.
wall model and its application to separated and re- [15] W.P. Jones and B.E. Launder. The prediction of la-
attached flows. Numerical Heat Transfer, 7:59–75, minarization with a two-equation model of turbu-
1984. lence. Int. J. Heat and Mass Transfer, 15:301–314,
1972.
[3] R.S. Amano, M.K. Jensen, and P. Goel. A nume-
rical and experimental investigation of turbulent [16] B.A. Kader and A.M. Yaglom. Heat and mass trans-
heat transport downstream from an abrupt pipe ex- fer laws for fully turbulent wall flows. Int. J. Heat
pansion. J. Heat Transfer, 105:862–869, 1983. and Mass Transfer, 15:2329–2351, 1972.
[4] J. Bredberg. Prediction of flow and heat trans- [17] N. Kasagi, Y. Tomita, and A. Kuroda. Direct nu-
fer inside turbine blades using EARSM, yR merical simulation of passive scalar in a turbulent
and turbulence models. Thesis for the De- channel flow. J. Heat Transfer, 114:598–606, 1992.
gree of Licentiate of Engineering, Dept. of Thermo
[18] H. Kawamura, H. Abe, and Y. Matsuo. DNS of tur-
and Fluid Dynamics, Chalmers University of Te-
bulent heat transfer in channel flow with respect to
chnology, Gothenburg, 1999. Also available at
Reynolds and Prandtl number effects. Int. J. Heat
www.tfd.chalmers.se/˜bredberg.
and Fluid Flow, 20:196–207, 1999.
[5] J. Bredberg and L. Davidson. Prediction of flow and [19] W.M. Kays. Turbulent Prandtl number - where are
heat transfer in a stationary 2-D rib roughened pas- we? J. Heat Transfer, 116:284–295, 1994.
sage using low-Re turbulent models. In 3:rd Eu-
ropean Conference on Turbomachinery, pages 963– [20] W.M. Kays and M.E. Crawford. Convective Heat
972, London, 1999. IMechE. and Mass Transfer. McGraw-Hill, Inc, 1993.
[6] J. Bredberg, S-H. Peng, and L. Davidson. On the [21] C.K.G. Lam and K.A. Bremhorst. Modified form of
wall boundary condition for computing heat trans- the R -model for predicting wall turbulence. J.
fer with models. In J.H. Kim, editor, HTD- Fluids Engineering, 103:456–460, 1981.
Vol. 366-5, ASME Heat Transfer Divison - 2000, vo- [22] B.E. Launder. Numerical computation of convective
lume 5, pages 243–250, Orlando, 2000. The Ameri- heat transfer in complex turbulent flows: Time to
can Society of Mechanical Engineers. abandon wall functions? Int. J. Heat and Mass
Transfer, 27:1485–1491, 1984.
[7] C.C. Chieng and B.E. Launder. On the calculation
of turbulent heat transfer transport downstream [23] B.E. Launder. On the computation of convective
an abrupt pipe expansion. Numerical Heat Trans- heat transfer in complex turbulent flows. J. Heat
fer, 3:189–207, 1980. Transfer, 110:1112–1128, 1988.
20
[29] A. Sahay and K.R. Sreenivasan. The wall-normal
position in pipe and channel flows at which viscous
and turbulent shear stresses are equal. Physics of
Fluids, 11:3186–3188, 1999.
[30] R.M.C So and T.P. Sommer. An explicit algebraic
heat-flux model for the temperature field. Int. J.
Heat and Mass Transfer, 39:455–465, 1996.
[31] H. Tennekes and J.L. Lumley. A First Course in
Turbulence. Massachusetts Institute of Technology,
Cambridge, 1972.
[32] A.A. Townsend. The structure of turbulent shear
flow. Cambridge University Press, Cambridge,
1976.
[33] T. von Karman. Mechanishe ähnlichkeit un turbu-
lenz. Nachr. Ges. Wiss., Göttingen, pp 68, 1930.
[34] D.C. Wilcox. Reassessment of the scale-
determining equation for advanced turbulence
models. AIAA Journal, 26:1299–1310, 1988.
[35] D.C. Wilcox. Comparison of two-equation turbu-
lence models for boundary layers with pressure gra-
dient. AIAA Journal, 31:1414–1421, 1993.
[36] D.C. Wilcox. Turbulence Modeling for CFD. DCW
Industries, Inc., 1993.
21