You are on page 1of 20

European Congress on Computational Methods in Applied Sciences and Engineering

ECCOMAS 2000
Barcelona, 11-14 September 2000
c ECCOMAS
REPRODUCING THE BLOCKING EFFECT OF THE WALL
IN ONE-POINT TURBULENCE MODELS
Remi Manceau
Thermal and Fluids Sciences Section, Department of Applied Physics
Delft University of Technology
Lorentzweg 1, PO Box 5046, 2600 GA, Delft, The Netherlands
Email: manceau@ws.tn.tudelft.nl
Key words: Near-wall models, Low-Reynolds-number Models, Non-local Eects, Asymp-
totic Behaviours, Elliptic Relaxation, Industrial Applications.
Abstract. The presence of a wall induces dierent eects on turbulence, which are
dicult to account for in one-point closures, in particular the non-local blocking eect.
The balance of the dominant terms of the budget of the Reynolds stresses in the vicinity of
the wall must be reproduced, which requires a careful modelling of the redistribution term.
The elliptic relaxation method enables the reproduction of this balance and, conse-
quently, of the anisotropy in the near-wall region. However, a side eect of this method is
the amplication of the redistribution in the log layer, which can be corrected by improving
the basis assumptions of the model, leading to a modied elliptic relaxation model.
Two simplied models, derived in order to reduce the number of equations, are pre-
sented: the v
2
f model, whose interest for industrial applications is shown through results
in the backstep and the ribbed-channel ow, and a new Reynolds-stress model, the elliptic
blending model, which gives predictions comparable to the elliptic relaxation model, but
with 8 equations instead of 13.
1
Remi Manceau
1 INTRODUCTION
Accounting for the wall-induced eects is one of the most dicult problems turbulence
modellers faced during 20
th
century. The few existing theories are generally valid only in
homogeneous turbulence, e.g., [10], or not to close to solid boundaries [3]. All existing
models derive from these theories, and making them comply with the near-wall behaviour
of turbulence is still nowadays one of the most active elds of research.
This issue is particularly important for real life applications, since ows are always
somehow inuenced by the presence of walls, either as boundaries (internal ows) or as
obstacles (external ows). From an industrial point of view, an accurate prediction of
near-wall turbulence is indispensable, since the quantities of primary interest are usually
those evaluated at the wall, such as the friction and heat transfer coecients C
f
and Nu.
Moreover, very intense energy interactions that inuence the rest of the ow take
place very close to the wall : the production of turbulent kinetic energy that reaches its
maximum in the buer layer, and its dissipation rate into internal energy that reaches its
maximum at the wall itself. The main diculty is that this region is generally very small.
For instance, in external ows, at usual Reynolds numbers, the boundary layer is several
orders of magnitude smaller than the characteristic size of the ow; the viscous sublayer
is also very small compared to the boundary layer, the size ratio behaving as Re
7/8

at
moderate Reynolds number, and as Re
9/10

at high Reynolds number, Re

being based
on the boundary layer thickness and the free-ow velocity.
Therefore, measurements are very dicult in the boundary layer and particularly in
the viscous sublayer. Our knowledge of the phenomena inside the latter remained partial
until the increasing power of super-computers made direct numerical simulation possi-
ble in the middle of the 80s. However, the available data are still limited to very low
Reynolds numbers and very simple geometries (Re 11, 000 for the most recent DNS in
a channel [23]).
The aim of this paper is to present the main properties of near-wall turbulence to be
accounted for in turbulence models and in particular to describe how the blocking eect
can be reproduced in one-point models. First, the inuence of the wall on turbulence, the
asymptotic behaviour of the redistribution term and the weaknesses of standard models
are presented. Secondly, the elliptic relaxation approach is described in the frame of
second moment closure. Attention is then focused on simplied models for industrial
applications: the v
2
f model and the elliptic blending model.
2 MODELLING WALL EFFECTS
2.1 Inuence of the wall on turbulence
The wall exerts a strong inuence on the instantaneous velocity and pressure elds, by
causing the appearance of streaks, hairpin vortices, ejections... However, our concern in
RANS is ensemble means: the diculty is thus to represent the mean eects of all these
instantaneous phenomena on the averaged quantities. In this section, the importance of
distinguishing among dynamic and kinematic eects is emphasized.
2
Remi Manceau
2.1.1 Dynamic eects
Mean shear The viscosity of the uid imposes a no-slip boundary condition on mean
velocities. This induces strong gradients from which the turbulence production
originates.
Viscous damping The no-slip boundary condition at instantaneous level implies that
the uctuating velocity components tangent to the wall u and w behave as y. Con-
sequently, the Reynolds stresses u
2
, w
2
and uw behave as y
2
. The stronger damping
of components involving the wall-normal velocity is due to another phenomenon,
the blocking eect.
Overlapping of scales When the turbulent Reynolds number Re
T
= k
2
/ diminishes,
the scale separation between energetic eddies and dissipative eddies progressively
disappears. Energetic scales are then inuenced by molecular diusion and inuence
in turn the dissipative eddies, which can no longer be considered as isotropic.
2.1.2 Kinematic eects
Wall echo eect The wall echo eect originates from the fact that an image term ap-
pears in the Green function of the domain when it is bounded by a wall. This is
linked to the pressure reection on the wall. Contrary to what is usually claimed,
this eect increases the pressure, and, accordingly, the redistribution term [16, 20],
but is small compared to the blocking eect.
Blocking eect Conjugate eects of the impermeability of the wall and the incompress-
ibility of the uid lead to the damping of the wall-normal instantaneous velocity
component v, which consequently behaves as y
2
in the vicinity of the wall. This
eect is instantaneously felt far from the wall, through the pressure that adjusts to
ensure the incompressibility condition: a velocity directed toward the wall instan-
taneously generates an increase of the pressure that contradicts it.
After using Reynolds decomposition, this eect is inherited by both mean velocity
V and uctuating velocity v. In computations, the blocking of the mean velocity
is well reproduced, since the incompressibility condition is explicitly imposed. This
leads, for instance, to the fact that the mean velocity eld feels obstacles at some
distance upstream. But the continuity equation is not resolved at uctuating level,
and the blocking eect must be accounted for by other means, through the only
term in Reynolds-stress transport equations that contains the pressure, namely the
redistribution term. This issue is the main concern of this paper.
2.1.3 Consequences for modelling
The dynamic and kinematic eects are very dierent by nature and can appear sepa-
rately. In particular, they do not act the same way when the ow reaches the leading edge
of a at plate or when a wall is suddenly introduced into the ow. Dynamic eects act
3
Remi Manceau
progressively, the layer aected by viscosity developing as a function of the streamwise
location [1] or of time [9, 24]. On the contrary, the kinematic eects instantaneously act
far from the wall [9].
Moreover, if all the eects previously described play a role close to a wall, they can
appear separately in other situations. For instance, at the interface between two uids of
dierent density, only kinematic eects are signicant. On the other hand, in a free ow,
when the Reynolds number diminishes, only the overlapping of scales appears. Therefore,
it is impossible, and even dangerous, to model these eects using functions of Re
T
only,
as some models do. They must better be accounted for by dierent mechanisms, and the
blocking eect, which is non-local, needs a very careful modelling, as presented in the
following of this paper.
2.2 Asymptotic behaviours
The aim of this section is to emphasize the importance of the balance between dominant
terms in the Reynolds-stress transport equations in the vicinity of the wall.
In a channel ow, the viscous damping and the blocking eect presented above imply
the following expansions of the mean velocity, the uctuating velocities, the uctuating
pressure, the Reynolds stresses, the turbulent kinetic energy and its dissipation rate:
_

_
U = /
1
y + /
2
y
2
+ O(y
3
)
u = a
1
y + a
2
y
2
+ O(y
3
)
v = b
2
y
2
+ O(y
3
)
w = c
1
y + c
2
y
2
+ O(y
3
)
p = p
0
+ p
1
y + p
2
y
2
+ O(y
3
)
(1)
_

_
u
2
= a
2
1
y
2
+ 2a
1
a
2
y
3
+ O(y
4
)
v
2
= b
2
2
y
4
+ 2b
2
b
3
y
5
+ O(y
6
)
w
2
= c
2
1
y
2
+ 2c
1
c
2
y
3
+ O(y
4
)
uv = a
1
b
2
y
3
+ (a
2
b
2
+ a
1
b
3
) y
4
+ O(y
4
)
(2)
k =
1
2
u
i
u
i
=
1
2
(a
2
1
+ c
2
1
) y
2
+ O(y
3
) (3) = (a
1
2
+c
1
2
) +O(y) (4)
The transport equations of the Reynolds stresses are
u
i
u
j
t
+ U
k
u
i
u
j
x
k
. .
c
ij
=

2
u
i
u
j
x
k
x
k
. .
D

ij

u
i
u
j
u
k
x
k
. .
D
T
ij

u
i
p
x
j

u
j
p
x
i
. .

ij
u
i
u
k
U
j
x
k
u
j
u
k
U
i
x
k
. .
P
ij
2
u
i
x
k
u
j
x
k
. .

ij
(5)
where c
ij
, D

ij
, D
T
ij
,

ij
, P
ij
et
ij
are convection, molecular diusion, turbulent diu-
4
Remi Manceau
D

ij
D
T
ij

ij
P
ij

ij
u
2
2a
2
1
+ 12a
1
a
2
y 4a
2
1
b
2
y
3
4a
1
a
2
y 2/
1
a
1
b
2
y
3
2a
2
1
8a
1
a
2
y
v
2
12b
2
2
y
2
6b
3
2
y
5
4b
2
2
y
2
0 8b
2
2
y
2
w
2
2c
2
1
+ 12c
1
c
2
y 4b
2
c
2
1
y
3
4c
1
c
2
y 0 2c
2
1
8c
1
c
2
y
uv 6a
1
b
2
y 5a
1
b
2
2
y
4
2a
1
b
2
y /
1
b
2
2
y
4
4a
1
b
2
y
Table 1: Asymptotic behaviours of the dierent terms of the budgets of the Reynolds stresses.
sion, velocitypressure gradient correlation (or redistribution), production and dissipation
terms, respectively. Note that

ij
is called herein redistribution term because it mainly
induces an energy redistribution among components of the Reynolds stress, but is not
only redistributive, since it is not traceless.
The asymptotic behaviours of these dierent terms in a channel ow are given in ta-
ble 1. It can be seen that, for all the components, production and turbulent diusion
are negligible. The signicant terms are D

ij
,
ij
and

ij
. Reproducing the balance be-
tween these terms as accurately as possible is thus necessary to predict the anisotropy: in
simulations, the asymptotic behaviours of the Reynolds stresses are related to the terms
involving their second derivatives, in particular to molecular diusion. Therefore, atten-
tion must be focused on the behaviour of
ij
, which is of the same order as D

ij
, but also
on

ij
, which balances the dierence between D

ij
and
ij
:

ij
=
ij
D

ij
+ O(y
n
) (6)

ij
is at a higher order than
ij
and D

ij
in u
2
and w
2
budgets, but at the dominant order
for v
2
and uv. In v
2
budgets,

ij
balances
ij
D

ij
up to 4
th
order (n = 5).
Following Lai & So [11], it must be emphasized that a turbulence model must reproduce
this balance in order to be valid down to the wall. In many cases, the correct behaviour of

ij
is not well reproduced. It is then preferable to built a model for

ij
that compensates
for this shortcoming, by balancing the dierence
ij
D

ij
, even if this does not correspond
to the true behaviour of

ij
.
The redistribution term

ij
is generally split into a traceless part, the pressurestrain
term
ij
, and the pressure diusion D
p
ij
:

ij
=
1

p
_
u
i
x
j
+
u
j
x
i
_
. .

ij

x
k
(u
i
p
jk
+ u
j
p
ik
)
. .
D
p
ij
(7)
in order to distinguish the purely redistributive eect and the diusion eect. Beyond
the non-unicity problem it induces [15], it leads to articial asymptotic behaviours that
are dicult to reproduce. For instance, in v
2
transport equation, the behaviours of the
5
Remi Manceau
0 25 50 75 100
0.05
0
0.05
y
+
(a)
0.01 0.1 1 10 100 1000
1e10
1e09
1e08
1e07
1e06
1e05
1e04
1e03
1e02
1e01
1e+00
1e+01
y
+
(b)
Figure 1: (a) Budget of v
2
with or without the decomposition of

ij
. Channel ow DNS (Moser et al. [23])
at Re

= 590. D

22
; D
T
22
;
22
;

22
;
22
; D
p
22
. (b) Reynolds stresses predicted by
the elliptic relaxation model in a channel at Re

= 590. Symbols: DNS [23]; lines: model. u


2
; v
2
;
w
2
; uv.
two terms of the decomposition
ij
+ D
p
ij
are

22
=
4

p
0
b
2
y +
_
6

p
0
b
3
+
4

p
1
b
2
_
y
2
+ O(y
3
) (8)
D
p
22
=
4

p
0
b
2
y
6

_
p
0
b
3
+ p
1
b
2
_
y
2
+ O(y
3
) (9)
It can be seen that each term behaves as y in the vicinity of the wall, whereas their sum
behaves as y
2
. Thus, this decomposition introduces terms that are dominant in the budget
of v
2
: they must be accurately reproduced such that their sum behaves as y
2
and balances
the dierence between
ij
and D

ij
, as shown in Fig. 1(a). Therefore, it appears preferable
to model directly

ij
, even if its physical meaning is less clear: far from the wall, it can
be considered as purely redistributive, since the pressure diusion is generally negligible,
and near the wall, its most important property is that it balances the dierence between
molecular diusion and dissipation.
2.3 Shortcomings of standard models for the redistribution term
Standard models for the redistribution term, called herein quasi-homogeneous models,
are based on the pioneering works of Chou [3] and Rotta [25]. They all use two assumptions
that lead to problems in predicting near wall ows, i.e., quasi-homogeneity and locality.
First, the quasi-homogeneity assumption consists in considering that the variations of
the mean velocity gradient U
l
/x
m
can be neglected in the integral equation of the rapid
part of the pressurestrain

2
ij
(x) =
_

U
l
x
m
(x

)
u
m
x
l
(x

)
_
u
i
x
j
(x) +
u
j
x
i
(x)
_
dV (x

)
2|x

x|
(10)
6
Remi Manceau
thus allowing to take it outside the integral, which leads to the general form of standard
models for the rapid term

2
ij
=
U
l
x
m
(a
ijml
+ a
jiml
) (11)
Now, Bradshaw et al. [2] have shown, by using a DNS database of a channel ow [22], that
the quasi-homogeneous approximation is valid beyond y
+
= 40, but totally wrong below.
Secondly, the locality assumption consists in forgetting that the pressure-strain is
written as an integral depending on all the points of the domain, and in modelling it by
algebraic expressions, such as (11), depending only on the point x. This is not questionable
from a mathematical point of view (the value of the integral is only a function of x), but
rather from a physical point of view: this leads to the loss of the non-local character of
the Reynolds-stress transport equations and thus makes very dicult the reproduction of
kinematic eects, whose importance has been emphasized in 2.1.
These two assumptions are then only applicable in free ows, and the presence of a wall
requires the introduction of strong corrections to the models: wall echo terms, damping
functions, complex non-linear terms. The wall echo terms [8] have the advantage of
reintroducing non-locality into the equations, since they explicitely depend on the distance
to the wall, but they lead to many problems in complex ows, and, above all, they are
based on an erroneous rationalization (cf. [16, 20]). Damping functions are mainly derived
in order to t particular experimental or DNS data, and have no reason to be valid in
general situations. Moreover, many of them are functions of Re
T
only, which contradicts
the fact, emphasized in 2.1, that a dependence on Re
T
cannot account for both dynamic
and kinematic eects, and they are not sucient for reproducing the correct anisotropy.
Complex non-linear terms also allow to impose elaborate physical constraints, such as
realizability in extreme conditions, but, following Speziale [26], it must be noted that
it should be preferable to avoid the use of quasi-homogeneous and locality assumptions
rather than correcting their consequences by introducing complex terms. Therefore, in the
following of this paper, attention is focused on the elliptic relaxation method, introduced
by Durbin [6, 7], which is based on a totally dierent approach.
3 THE ELLIPTIC RELAXATION MODEL
3.1 Derivation of the model
From the solution of the Poisson equation for the uctuating pressure, the following
integral equation of the redistribution term

ij
can be derived:

ij
(x) =
_

_
u
i
(x)
2
p
x
j
(x

) u
j
(x)
2
p
x
i
(x

)
_
. .

ij
(x, x

)
G

(x, x

)dV (x

) (12)
7
Remi Manceau
where G

(x, x

) is the Green function of the domain. Durbin [6] proposed to model the
two-point correlation
ij
by

ij
(x, x

) =
ij
(x

, x

) exp
_

r
L
_
(13)
where r denotes the distance |x

x|. However, using (13) in (12) leads to the loss of


the correct asymptotic behaviour of

ij
. Indeed, it can be seen in (12) that

ij
goes to
zero when x approaches the wall because of u
i
and u
j
, which are expressed in x. Using
the model (13), u
i
and u
j
are turned into functions of x

, and

ij
has no reason to go to
zero anymore. A way of overcoming this problem is to use the model [16]

ij
(x, x

) = k(x)

ij
(x

, x

)
k(x

)
exp
_

r
L
_
(14)
since the factor k(x) forces

ij
(x) to go to zero at the wall.
Introducing (14) into (12), approximating G

by its free-space value G


IR
3 = 1/4r
and using the fact that (
IR
3(x, x

) = exp(r/L)/4r is the Green function associated to the


operator 1/L
2

2
, the following elliptic relaxation equation is obtained:

ij
k
L
2

ij
k
=

h
ij
k
(15)
In this equation, a quasi-homogeneous model
h
ij
(Rotta+IP, QI, SSG, ...) is used in the
right hand side, noting that in homogenous situations, the second term of the left hand
side vanishes.
The main interest of this approach is that the redistribution term is given by a dier-
ential equation, rather than by an algebraic expression as in the standard approach. The
elliptic character of (15) enables the reproduction of the non-locality of

ij
, which derives
from the model (14) for the correlation function. Moreover, the quasi-homogeneous as-
sumption is only used for the modelling of the source term of (15), which is the form the
model tends to far from solid boundaries, at least suciently far for being in a region of
the ow where this assumption is valid. In the vicinity of the wall, the solution of (15) is
constrained by its boundary conditions, as shown in the following section.
3.2 Near-wall behaviour
It has been noticed in 2.2 that the main point a near-wall turbulence model must
satisfy is the balance between molecular diusion, dissipation and redistribution. First, it
is necessary to pay attention to the behaviour of
ij
, which must take an anisotropic form

ij
in the vicinity of the wall and tend to
2
3

ij
far from it. A possible way of imposing this
is to decompose
ij
into

ij
and (
ij

ij
), and to solve for the latter an elliptic relaxation
equation similar to (15):

ij

ij
k
L
2

ij

ij
k
=
2
3

ij

ij
k
(16)
8
Remi Manceau
By imposing the boundary condition (
ij

ij
)/k = 0 at solid boundaries, the total dissi-
pation

ij
+ (
ij

ij
) approaches the value

ij
in the near-wall region. Far from the wall,
the second term in the left hand side of (16) vanishes and the total dissipation tends to
2
3

ij
.
However, since the operator 1 L
2

2
is linear, and since only the dierence

ij

ij
appears in the Reynolds-stress transport equations, it is not necessary to solve an elliptic
relaxation equation for

ij
/k and another for (
ij

ij
)/k, but only one for their dierence
f
ij
= (

ij

ij
+

ij
)/k, which yields
f
ij
L
2

2
f
ij
=
1
k
_

h
ij

2
3

ij
+

ij
_
(17)
In this equation, the near-wall model

ij
= u
i
u
j
/k is used. Thus, the modelled Reynolds-
stress transport equations become
u
i
u
j
t
+c
ij
= P
ij
+ D

ij
+ D
T
ij
+ kf
ij

u
i
u
j
k
(18)
in which D
T
ij
is given by the Daly & Harlow [4] model.
Thus, according to table 1, in the vicinity of the wall, the budgets of the Reynolds-stress
transport equations reduce to

2
u
i
u
j
y
2

u
i
u
j
k
= kf
ij
(19)
Using /k = 2/y
2
(cf. Eq. 3 and 4), the solution of this dierential equation is
u
i
u
j
= Ay
2
+
B
y


20
2
f
ij
y
4
(20)
Since u
i
u
j
= 0 is imposed at the wall, B = 0, and the asymptotic behaviour of u
i
u
j
de-
pends on the boundary condition used for f
ij
.
For v
2
and uv, the boundary conditions must be chosen in order to cancel the term
Ay
2
in (20). Therefore, for v
2
, the condition f
w
22
= 20
2
v
2
/y
4
is used. Note that f
w
22
is
not singular, since this boundary condition implies that v
2
behaves as y
4
.
For Reynolds stresses behaving as y
2
(cf. Eq. 2), any boundary condition can be applied,
since the leading term in (20) is in y
2
. Therefore, f
w
11
= f
w
33
=
1
2
f
w
22
are used, in order
to ensure that f
w
kk
= 0. The reason for imposing f
w
kk
= 0, whereas

ij
is not traceless,
is that it ensures that the Reynolds-stress transport equations contract to the standard
k equation, thus avoiding the need to modify the coecients of the quasi-homogeneous
model used as the source term.
The case of the Reynolds shear stress uv is more problematic. Indeed, (2) shows that
uv behaves as y
3
, but there is no term containing y
3
in (20). Therefore, the same boundary
condition as for f
22
is used for f
12
, i.e, f
w
12
= 20
2
uv/y
4
, which induces a behaviour of
uv in y
4
.
9
Remi Manceau
f
w
11
f
w
22
f
w
33
f
w
12
f
w
13
f
w
23

1
2
f
w
22

20
2

v
2
y
4

1
2
f
w
22

20
2

uv
y
4
0
20
2

vw
y
4
Table 2: Boundary conditions for the components of the tensor f
ij
.
In general cases, the latter boundary condition is easily extended to the component f
w
23
:
f
w
23
= 20
2
vw/y
4
. For f
w
13
, as for f
w
11
and f
w
33
, any boundary condition can be applied:
f
w
13
= 0 is simply used. Table 2 summarizes the boundary conditions used for the dierent
components of f
ij
.
3.3 Advantages and shortcomings
The main quality of this model is that it reproduces the non-local blocking eect. In-
deed, the model for

ij

ij
+

ij
approaches progressively its asymptotic form kf
w
ij
in the
near-wall region, the transition being ensured by the elliptic operator. In particular, this
enables the stronger damping of v
2
, which leads to the two-component state of turbulence
very close to the wall, as shown in Fig. 1(b).
An easy way of understanding the role played by the model is to compare the equations
of v
2
and w
2
. Suppose that the Rotta+IP model is used for
h
ij
. In a channel, v
2
and w
2
have exactly the same transport equation. Moreover, the elliptic relaxation equations for
f
22
and f
33
are also exactly identical. Thus, if the boundary conditions f
w
22
and f
w
33
were
the same, v
2
and w
2
would be exactly equal across the channel. Fortunately, dierent
boundary conditions are used, which distinguish v
2
and w
2
and lead to the prediction of
the correct anisotropy near the wall, shown in Fig. 1(b). Thus, the reproduction of the
blocking eect is due to the fact that a dierential equation is solved for f
ij
, which enables
the introduction of dierences between the components through the boundary conditions,
without violating the frame indierence principle. In algebraic models for

ij
, the only
way of distinguishing between components is to identify the wall-normal direction, e.g.,
by using the invariants of the Reynolds stress tensor, thus introducing non-linearities,
which complicate the equations and lead to numerical diculties.
Moreover, the elliptic relaxation model only uses the quasi-homogeneous assumption
in the source term
h
ij
, which becomes active suciently far from the wall. This feature
can be illustrated by an a priori test: Eq. (15) is resolved for

22
, taking k, L,
h
ij
and
the boundary conditions from the channel ow DNS [23] (for a priori tests, it is more
convenient to handle separately

ij
and
ij

ij
, the interpretation of the result being
much easier). Fig. 2(a) shows the results obtained using two dierent models for
h
ij
.
It can be seen that the elliptic relaxation equation enables the correction of the near-
wall behaviour of the redistribution, and that the results obtained below y
+
= 40 do not
depend on the quasi-homogeneous model used in the source term: the solution is totally
determined by the boundary conditions. Thus, the quasi-homogeneous model only plays a
10
Remi Manceau

2
2
0 100 200 300 400 500 600
0
0.02
0.04
0.06
y
+
(a)

2
2
0 100 200 300 400 500 600
0
0.02
0.04
0.06
y
+
(b)
Figure 2: A priori tests of the eect of the elliptic relaxation equation. Channel ow at Re

= 590.
(a) Original formulation (17); (b) Modied formulation (22).

22
from the DNS [23];
h
22
given
by the Rotta+IP model (without elliptic relaxation);

22
given by the elliptic relaxation equation
with the Rotta+IP model as the source term;
h
22
given by the SSG model (without elliptic
relaxation);

22
given by the elliptic relaxation equation with the SSG model as the source term.
signicant role further from the wall, where the quasi-homogeneous assumption is valid [2].
On the other hand, some shortcomings of the model must be noted. First, it is not,
contrary to what is sometimes believed, a low-Reynolds-number model, but rather a
near-wall model. In the near-wall region, all the eects described in 2.1 are reproduced,
because the correct asymptotic behaviours are imposed, but low-Reynolds-number eects
far from a wall, in particular the eect of the overlapping of scales, are not accounted for,
since there is no dependence on Re
T
in the model.
Another point to be brought out is the amplication of the redistribution in the log
layer. It can be seen in Fig. 2(a) that the main eect of the elliptic relaxation equation
is to correct the redistribution in the viscous and buer layers. However, a side eect
can be noted: in the logarithmic layer, the redistribution

ij
given by the model is higher
than the one predicted by the quasi-homogeneous model
h
ij
. This is actually due to
the operator 1 L
2

2
, which has an amplication eect. Indeed, in the log layer, the
following behaviours can be assumed:
h
ij
= A/y, k = u

2
/C
1/2

and = u
3

/y, where
A is some constant value. Thus, the source term behaves as 1/y, and if a solution of
the form

ij
=
h
ij
is sought, an amplication factor 1.51 is obtained. With the
Rotta+IP model, which already overestimates the redistribution in the log layer, it would
be preferable to have an amplication factor smaller than 1. With the SSG model, which
reproduces correctly

ij
in this region, a neutral operator, i.e. corresponding to = 1,
would lead to better predictions. This issue is examined in 3.4, where a modied elliptic
relaxation equation [20, 21] is presented.
Finally, it must be emphasized that the elliptic relaxation approach leads to 6 additional
dierential equations in the system, for the 6 independent components of the tensor
f
ij
. This makes the model rather unpopular among industry. Therefore, Durbin [6]
11
Remi Manceau
proposed an eddy-viscosity model that uses an elliptic relaxation equation to reproduce
the behaviour of the eddy viscosity in the vicinity of the wall, the so-called v
2
f model,
which is presented in 4.1. In 4.2, a new approach for second moment closures is proposed,
based on the same general ideas than the elliptic relaxation, but using only 1 additional
equation.
3.4 Reformulation of the elliptic relaxation equation
Manceau et al. [20, 21] sought the reason for the amplication of the redistribution
given by the elliptic relaxation operator in the log layer by investigating the validity of
the basis assumptions of the model using a DNS database. They showed that this problem
is due to the approximation (13) of the correlation function, which does not account for its
anisotropy. In particular, by using a model for the correlation function that is asymmetric
in the main direction of inhomogeneity, identied by the gradient of the length scale
f(x, x

) = exp
_
r
L + (x

x) L
_
(21)
a new formulation of the elliptic relaxation equation can be derived:
(1 + 16(L)
2
)f
ij
L
2

2
f
ij
8LL f
ij
= f
h
ij
(22)
This formulation is characterized by the amplication factor
=
1
1 + 2(12 1)C
2
L
C
3/2


2
(23)
Dierent values of the coecient can be chosen, depending on the eect required in
the log layer. Fig. 2(a) shows that, when the Rotta+IP model is used for
h
ij
, which
overestimates the redistribution in the log layer, a coecient that leads to a damping
of the redistribution is needed. In Fig. 2(b), the

22
prole obtained by an a priori test
with = 0.25 is shown. It can be seen that the use of (22) enables the correction of the
behaviour in the log layer, without degrading the predictions in the viscous and buer
layers. With the SSG model, whose results are degraded in the log layer by the original
elliptic relaxation equation (cf. Fig. 2a), it can be seen in Fig. 2(b) that the formulation
(22), with = 0.17, gives more satisfactory results than the original one.
Results obtained by the full integration of the Reynolds stress model equations (17)
and (18) are plotted in Fig. 3. The SSG model is used for
h
ij
. As noted above, the
original formulation leads to diculties in predicting the redistribution in the log layer,
and the coecients of the model (cf. [28]) have been chosen in order to compensate for
this problem. This leads to the underestimation of the peak of u
2
and of the mean velocity
in the buer layer that can be observed in Fig. 3. Replacing (17) by the formulation (22),
with = 0.083, allows the use of a wider range of coecients, which enables a better
prediction of the peak of u
2
and of the mean velocity in the buer layer. More details
about these simulations can be found in [17].
12
Remi Manceau
U
+
1 10 100 1000
0
5
10
15
20
1 100
0
5
10
15
20
y
+
(a)
u
i
u
j
+
0 100 200 300 400 500 600
0
2
4
6
8
y
+
(b)
Figure 3: Comparison of the results given by the original and modied formulations for the channel ow
at Re

= 590. The SSG model is used for


h
ij
. (a) Mean velocity proles. DNS [23]; original
formulation (22); modied formulation (22). (b) Reynolds stresses. Symbols: DNS ( u
2
, v
2
,
w
2
, _ uv); original formulation (22); modied formulation (22).
4 SIMPLIFIED APPROACHES FOR INDUSTRIAL APPLICATIONS
As emphasized in 3.3, since it involves 6 transport equations for the Reynolds stresses
and 6 elliptic relaxation equations, the full Reynolds-stress model proposed by Durbin is
very seldom used in industry. In the present section, two simplied models suitable for
industrial applications are presented: the v
2
f model and the elliptic blending model.
4.1 The v
2
f model
In industrial applications, k type models are widely used. However, high-Reynolds-
number models also need corrections for reproducing the damping of the eddy viscosity in
the near-wall region. Generally, these corrections are based on the use of damping func-
tions, which introduce strong non-linearities and thus numerical instabilities. In order to
avoid such corrections, Durbin [6] proposed to use the fact, initially noted by Launder [12],
that the eddy-viscosity can be well reproduced by the model
T
= C

v
2
k/ in a channel
ow. Thus, in a channel, by using the same v
2
and f
22
equations as in the full Reynolds
stress model presented in 3, which gives an accurate prediction of the prole of v
2
in the
near-wall region (cf. Fig. 1b), the correct damping of
T
is obtained. This leads to a model
consisting of 3 transport equations, for k, and v
2
, and 1 elliptic relaxation equation.
This model can easily be generalised to general geometries, without changing the equa-
tions, but only by changing the meaning of the variables: v
2
is then no longer a component
of a tensor, but a scalar variable which is given by a transport equation formally identical
to the transport equation of the Reynolds stress component v
2
in a channel (changing
the name of the variable would have made this distinction more clear); in a similar way,
a new scalar variable called f is introduced, which satises the same elliptic relaxation
equation as f
22
in a channel.
More details about this model can be found in [6] or in [19]. The purpose of this section
13
Remi Manceau
is only to present some results that illustrate the interest of the model for wall-bounded
ows.
In Fig. 4, mean velocity, k and v
2
predictions in the case of a channel ow are plotted.
Two dierent sets of results are compared to DNS data. The rst one is given by the
original v
2
f model, using the elliptic operator [1 L
2

2
] in the f equation, the sec-
ond one by the reformulated model, using [(1 + 16(L)
2
) L
2

2
8LL ], with
= 0.083. It can be seen that the k and v
2
proles are correctly reproduced down to the
wall by both formulation. On the other hand, the mean velocity prole is improved by the
modied model, which is due to the suppression of the amplication of the redistribution
in the log layer noted in 3.3, which enables the use of a wider range of coecients. More
details can be found in [17].
Fig. 5(a) shows results obtained with the original v
2
f model in a backstep ow at
Re = 5, 100, compared with the DNS results obtained by Le et al. [13]. It can be seen
that all the characteristics of the ow that models generally fail to predict are correctly
reproduced by the v
2
f model: rst, the recirculation length, which is predicted with
only 3 % error (6.59h instead of 6.39h); secondly, the intensity of the backow, and
consequently the friction coecient (not shown here, cf. [18, 19]); nally, the recovery of
the boundary layer downstream of the recirculation bubble. These results show the ability
of the model to reproduce separated ows as well as the near-wall turbulence, which
illustrates the interest of this model for industrial complex ows: the most important
quantities are often those which are evaluated at solid boundaries and which are strongly
inuenced by separation and reattachment, namely the friction coecient and the Nusselt
number.
Heat transfer predictions are indeed very sensitive to the level of turbulence and the
quality of a turbulence model can be partly evaluated through heat transfer cases. In
1998, a comparison between dierent turbulence models, including low-Reynolds-number
Reynolds stress models, in the case of a periodic ribbed-channel ow, has been proposed
for the 7
th
ERCOFTAC/IAHR workshop on rened turbulence modelling [27]. Fig. 5(bc)
show the velocity proles and the streamlines obtained at Re = 37, 200. It can be ob-
served in Fig. 5(b) that the backow is correctly reproduced in the recirculation zone.
No reattachment point is obtained, the two main recirculation bubbles being actually
connected, as can be seen in Fig. 5(c). Two secondary recirculation zones are also pre-
dicted. The agreement between model and experiments is far from perfect in the region
above the ribs (1 y/h 5), but no conclusion can be drawn because 3D eects are
suspected in the experiments (cf. [19]). The Nusselt number distribution obtained be-
tween two consecutive ribs at Re = 12, 600 is plotted in Fig. 5(d). The heat uxes are
simply modelled by a turbulent diusivity hypothesis. It can be seen that, despite this
very simple heat ux model, the Nusselt number distribution is well predicted, especially
between x/h = 0 and x/h = 3, which shows that, in forced convection, using a turbulence
model that reproduces correctly the near-wall characteristics is more important than us-
ing an elaborate heat ux model. It must be emphasized that the v
2
f model gave the
14
Remi Manceau
U
+
1 10 100 1000
0
5
10
15
20
y
+
(a)
k
,
v
2
0 100 200 300 400 500 600
0
1
2
3
4
5
y
+
(b)
Figure 4: Results given by the v
2
f model in a channel ow at Re

= 590. (a) Mean velocity. DNS [23];


Original formulation (17); Modied formulation (22). (b) Turbulent energy and v
2
. Sym-
bols: DNS ( k, v
2
); Original formulation (17); Modied formulation (22).
y
/
h
0 10 20
0
1
2
U/U
b
(a)
y
/
h
0 2 4 6 8
0
1
2
3
4
5
U/U
b
(b)
y
/
h
x/h
(c)
N
u
/
N
u
s
0 1 2 3 4 5 6
0
1
2
x/h
(d)
Figure 5: Results given by the v
2
f model (original formulation) in two separated ows. (a) Backward-
facing step at Re = 5, 100. Mean velocity proles. DNS [13]. v
2
f. (b) Ribbed-channel ow
at Re = 37, 200. Mean velocity proles. Experiments [5]; v
2
f. (c) Ribbed-channel ow at
Re = 37, 200. Streamlines. (d) Ribbed-channel ow at Re = 12, 600. Heat transfer enhancement Nu/Nu
s
,
Nu
s
being the Nusselt number for a turbulent ow in a smooth circular pipe, given by the Dittus-Boelter
correlation: Nu
s
= 0.023Re
0.8
Pr
0.4
. Experiments [14]; v
2
f.
15
Remi Manceau
best results among all the models used by dierent teams participating to the workshop
(cf. [27]). The present computations with the v
2
f model have been detailed in [18, 19].
4.2 The elliptic blending model
The limitations of the turbulent viscosity approach make necessary the use of full
Reynolds-stress models in number of industrial cases. It has been shown in 3 that
Durbins Reynolds stress model is a very interesting approach for wall-bounded ows, but
remains quite unpopular because of the 6 additional equations it implies.
The main quality of this model, i.e., the reproduction of the blocking eect, is due
to the fact that the elliptic relaxation equations ensure a smooth transition between the
far-from-the-wall form
h
ij
of the redistribution term

ij
, and its near-wall asymptotic
values. Thus, it can be noted that each elliptic relaxation equation provides a transition
depending only on the geometry and of the length scale L, which is the same for all the
components. Thus, it appears that these 6 equations are somewhat redundant. The same
eect can be reproduced by a blending function , which ensures the transition between
far-from-the-wall and near-wall forms in the following manner:

ij
= (1 k)
w
ij
+ k
h
ij
(24)
where k must be zero at the wall and must tend to 1 far from it. The geometrical eect
can be reproduced by using an elliptic relaxation equation for :
L
2

2
=
1
k
(25)
with the boundary condition = 0 at the wall. The reason for using 1/k in the source
term of (25) and then multiplying by k in (24) is that it ensures that the factor k
behaves as y
3
in the vicinity of the wall, which makes the second term in the right hand
side of (24) negligible in this region.
To obtain the same eect as with Durbins model,
w
ij
must be chosen such that the
balance between viscous diusion, dissipation and redistribution in the near-wall region
is respected, as shown in 3.2. Thus,
w
ij
/k must tend to the values f
w
ij
given in table 2.
This can be achieved in a channel by

w
11
=
1
2

w
22
;
w
22
= 5

k
v
2
;
w
33
=
1
2

w
22
;
w
12
= 5

k
uv (26)
since /k tends to 2/y
2
(cf. equations 3 and 4). In a similar way, for the dissipation
tensor, the transition between u
i
u
j
/k and
2
3

ij
is ensured by the same blending method:

ij
= (1 k)
u
i
u
j
k
+ k
2
3

ij
(27)
Fig. 6 shows the results obtained with this model in a channel ow, the Rotta+IP model
being used for
h
ij
. The results given by the full Reynolds stress elliptic relaxation model
16
Remi Manceau
U
+
1 10 100
0
5
10
15
20
y
+
(a)
u
i
u
j
+
0 100 200 300 400 500 600
0
2
4
6
8
y
+
(b)
Figure 6: Comparison in a channel ow at Re

= 590 between the results given by the elliptic blending


model and by the elliptic relaxation model. (a) mean velocity. DNS [23]; elliptic relaxation
model; elliptic blending model. (b) Reynolds stresses. Symbols: DNS [23] ( u
2
, v
2
, w
2
, _
uv); elliptic relaxation model; elliptic blending model.
(original formulation), already shown in Fig. 3, is also plotted for comparison. It can be
seen that the results of the present model are almost as good as those given by the full
elliptic relaxation model. Indeed, even if the prediction of the Reynolds stresses is slightly
less accurate, the anisotropy is globally correctly reproduced, and it must be kept in mind
that this model contains only 8 equations (taking equation into account), instead of
13 for Durbins model. This reduction of the number of equations is very interesting for
industrial applications.
The model needs to be generalised to complex geometries, since (26) is not frame
indierent. Therefore, some directional information must be introduced in a general
tensorial form of (26). In order to avoid the use of geometry-related quantities, such as
the wall-normal vector, which is ill-behaved in complex geometries, it can be noted that
the gradient of is generally normal to the wall, and it is still dened inside the domain.
Thus, a vector n can be dened by n = /||. With this denition, (26) can be
generalized by

w
ij
= 5

k
_
u
i
u
k
n
j
n
k
+ u
j
u
k
n
i
n
k

1
2
u
k
u
l
n
k
n
l
n
i
n
j

1
2
u
k
u
l
n
k
n
l

ij
_
(28)
which is applicable to any geometry.
5 CONCLUSION
The importance of modelling the dierent eects of the wall on turbulence has been
emphasized. In particular, the blocking eect, which is non-local and which is felt by
the Reynolds stresses through the redistribution term, requires a particular attention.
In the vicinity of a wall, this term balances the dierence between molecular diusion
and dissipation, and respecting this balance is indispensable for predicting the correct
17
Remi Manceau
near-wall anisotropy.
Standard models for the redistribution term, based on quasi-homogenous and local-
ity assumptions need strong corrections for being integrable down to solid boundaries,
and the reproduction of the two-component limit of turbulence requires highly nonlinear
terms. On the contrary, Durbins elliptic relaxation model, which is not based on the
same assumptions, does not need nonlinear corrections and predicts accurately the strong
anisotropy in the near-wall region.
Its main shortcoming, the amplication of the redistribution in the log layer by the
elliptic operator, has been corrected by accounting for the anisotropy of the two-point
correlations involved in the integral equation of the redistribution term. The modied
elliptic relaxation equation which is then derived gives better predictions of the mean
velocity prole and of the peak of u
2
in a channel ow.
However, Durbins full Reynolds-stress model is very seldom used in industrial com-
putations, since it involves 6 transport equations for the Reynolds stresses and 6 elliptic
relaxation equations. Two simplied models, suitable to industrial applications, have
been presented.
First, the v
2
f model, which is based on the turbulent viscosity concept, but reproduces
the damping of the latter by elliptic relaxation. Results presented in the cases of ows in
a channels, over a backward-facing step and in a periodic ribbed-channel with one heated
wall, show the ability of the model to reproduce accurately the near-wall turbulence
characteristics, and in particular the friction coecient and the Nusselt number, which
are of primary interest for industry.
Secondly, the elliptic blending model has been proposed. This new model is a full
Reynolds-stress model derived in order to reproduce the main features of the elliptic
relaxation model, but with only 1 additional equation, instead of 6. The results obtained
in a channel ow show the ability of this model to reproduce the anisotropy in the near-
wall region.
Acknowledgments
A part of the work presented in this paper has been supported by

Electricite de France,
during the Ph.D. thesis of the author. In particular, the simulations of the backstep and
ribbed-channel ows have been performed with EDF nite element code N3S.
18
Remi Manceau
REFERENCES
[1] D. Aronson, A. V. Johansson, and L. L ofdahl. Shear-free turbulence near a wall. J.
Fluid Mech., 338:363385, 1997.
[2] P. Bradshaw, N. N. Mansour, and U. Piomelli. On local approximations of the
pressurestrain term in turbulence models. In Proc. of the Summer Program, pages
159164. Center for Turbulence Research, Stanford University, 1987.
[3] P. Y. Chou. On velocity correlations and the solutions of the equations of turbulent
uctuation. Quart. of Appl. Math., 3:3854, 1945.
[4] B. J. Daly and F. H. Harlow. Transport equations in turbulence. Phys. Fluids,
13:26342649, 1970.
[5] L. E. Drain and S. Martin. Two-component velocity measurements of turbulent
ow in a ribbed-wall ow channel. In Intl Conf. Laser Anemometry-Advances and
Applications, pages 99112, 1985.
[6] P. A. Durbin. Near-wall turbulence closure modeling without damping functions.
Theoret. Comput. Fluid Dynamics, 3:113, 1991.
[7] P. A. Durbin. A Reynolds stress model for near-wall turbulence. J. Fluid Mech.,
249:465498, 1993.
[8] M. M. Gibson and B. E. Launder. Ground eects on pressure uctuations in the
atmospheric boundary layer. J. Fluid Mech., 86(3):491511, 1978.
[9] J. C. R. Hunt and J. M. R. Graham. Free-stream turbulence near plane boundaries.
J. Fluid Mech., 84(2):209235, 1978.
[10] A. N. Kolmogorov. The local structure of turbulence in incompressible viscous uid
for very large Reynolds numbers. Dokl. Aked. Nauk., URSS, 30:299303, 1941.
[11] Y. G. Lai and R. M. C. So. On near-wall turbulent ow modelling. J. Fluid Mech.,
221:641673, 1990.
[12] B. E. Launder. On the computation of convective heat transfer in complex turbulent
ows. J. Heat Transfer, 110:11121128, 1988.
[13] H. Le, P. Moin, and J. Kim. Direct numerical simulation of turbulent ow over a
backward-facing step. In Proc. Ninth Symp. Turb. Shear Flows, 132, pages 16.
Kyoto, 1993.
[14] T.-M. Liou, J.-J. Hwang, and S.-H. Chen. Simulation and measurement of enhanced
turbulent heat transfer in a channel with periodic ribs on one principal wall. Intl J.
Heat Mass Transfer, 36:507517, 1993.
19
Remi Manceau
[15] J. L. Lumley. Pressure-strain correlation. Phys. Fluids, 18(6):750, 1975.
[16] R. Manceau. Modelisation de la turbulence. Prise en compte de linuence des parois
par relaxation elliptique. PhD thesis, Universite de Nantes, 1999.
[17] R. Manceau and K. Hanjalic. A new form of the elliptic relaxation equation to
account for wall eects in RANS modelling. To be published in Phys. Fluids.
[18] R. Manceau and S. Parneix. Computations of turbulent ows using the v
2
f model
in a nite element code. In Proc. Fourth Intl Symp. on Engng Turbulence Modelling
and Measurements, pages 319328, 1999.
[19] R. Manceau, S. Parneix, and D. Laurence. Turbulent heat transfer predictions using
the v
2
f model in a nite element code. Intl J. Heat and Fluid Flow, in press.
[20] R. Manceau, M. Wang, and D. Laurence. Inhomogeneity and anisotropy eects on
the redistribution term in RANS modelling. To be published in J. Fluid Mech.
[21] R. Manceau, M. Wang, and D. Laurence. Assessment of inhomogeneity eects on
the pressure term using DNS database: implication for RANS models. In Proc. First
Intl Symp. Turb. Shear Flow Phenomena, 8, pages 239244. Santa Barbara, 1999.
[22] N. N. Mansour, J. Kim, and P. Moin. Reynolds-stress and dissipation-rate budgets
in a turbulent channel ow. J. Fluid Mech., 194:1544, 1988.
[23] R. D. Moser, J. Kim, and N. N. Mansour. Direct numerical simulation of turbulent
channel ow up to Re

= 590. Phys. Fluids, 11(4):943945, 1999.


[24] J. B. Perot and P. Moin. Shear-free turbulent boundary layers: physics and modeling.
Technical report, Dept Mech. Engng, Stanford University, 1993.
[25] J. C. Rotta. Statistische Theorie nichthomogener Turbulenz. Zeitschrift f ur Physik,
129:547572, 1951.
[26] C. G. Speziale. A review of Reynolds stress models for turbulent shear ows. Tech-
nical Report 95-15, ICASE, NASA, 1995.
[27] UMIST, Manchester. Proc. Seventh ERCOFTAC/ IAHR workshop on rened turbulence
modelling, 1998.
[28] V. Wizman, D. Laurence, M. Kanniche, P. Durbin, and A. Demuren. Modeling near-
wall eects in second-moment closures by elliptic relaxation. Intl J. Heat and Fluid
Flow, 17(3):255266, 1996.
20

You might also like