Professional Documents
Culture Documents
modeling
Lars Davidson
Division of Fluid Dynamics
Department of Applied Mechanics
Chalmers University of Technology
SE-412 96 G teborg, Sweden
o
http://www.tfd.chalmers.se/lada, lada@chalmers.se
June 12, 2010
Abstract
This course material is used in three courses in the International Masters programme Solid and uid mechanics at Chalmers. The three courses are TME075
Mechanics of solids and uids (Part II: Fluid mechanics), MTF256 Turbulent Flow
and MTF270 Turbulence Modeling. MSc students who follow these courses are
supposed to have taken at least one course in uid mechanics.
This document can be downloaded here
http://www.tfd.chalmers.se/lada/turbulent flow/lecture notes.html
The Fluid courses in the MSc programme are presented here
http://www.tfd.chalmers.se/lada/msc/msc-programme.html
The MSc programme is presented here
http://www.chalmers.se/en/sections/education/masterprogrammes
Contents
1
Motion, ow
1.1
Eulerian, Lagrangian, material derivative
1.2
Viscous stress, pressure . . . . . . . . .
1.3
Strain rate tensor, vorticity . . . . . . . .
1.4
Deformation, rotation . . . . . . . . . .
1.5
Irrotational and rotational ow . . . . .
1.5.1 Ideal vortex line . . . . . . . . .
1.5.2 Shear ow . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
9
9
10
11
13
15
15
16
Governing ow equations
2.1
The Navier-Stokes equation . . . .
2.1.1 The continuity equation . .
2.1.2 The momentum equation .
2.2
The energy equation . . . . . . . .
2.3
Transformation of energy . . . . .
2.4
Left side of the transport equations
2.5
Material particle vs. control volume
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
17
17
17
17
18
19
20
21
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
21
21
25
25
25
28
29
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
30
30
32
35
35
Turbulence
5.1
Introduction . . . . . . . . . . . . . . .
5.2
Turbulent scales . . . . . . . . . . . . .
5.3
Energy spectrum . . . . . . . . . . . . .
5.4
The cascade process created by vorticity
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
38
38
39
40
43
Turbulent mean ow
6.1
Time averaged Navier-Stokes . . . . . . . . . . .
6.1.1 Boundary-layer approximation . . . . . .
6.2
Wall region in fully developed channel ow . . .
6.3
Reynolds stresses in fully developed channel ow
6.4
Boundary layer . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
46
46
47
47
52
53
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
54
57
57
8.2
8.3
8.4
60
61
62
64
69
10 Correlations
10.1 Two-point correlations . . . . . . . . . . . . . . . . . . . . . . . . .
10.2 Auto correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . .
70
70
72
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
75
75
75
76
76
77
78
79
80
80
81
81
83
83
86
91
94
94
95
96
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
97
97
99
101
102
13 Realizability
13.1 Two-component limit . . . . . . . . . . . . . . . . . . . . . . . . .
103
104
105
108
111
112
112
112
117
.
.
.
.
17.1
17.2
17.3
17.4
17.5
17.6
17.7
17.8
17.9
17.10
17.11
17.12
17.13
17.14
17.15
17.16
17.17
17.18
17.19
17.20
17.21
17.22
17.23
17.24
17.25
17.26
117
118
119
120
121
121
122
122
123
123
124
125
126
127
128
129
129
130
131
132
133
134
134
135
136
136
136
18 Unsteady RANS
18.1 Turbulence Modeling . . . . . . . . . . . . . . . . . . . . . . . . .
18.2 Discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
138
141
141
19 DES
19.1 DES based on two-equation models . . . . . . . . . . . . . . . . . .
19.2 DES based on the k SST model . . . . . . . . . . . . . . . . .
142
143
145
20 Hybrid LES-RANS
20.1 Momentum equations in hybrid LES-RANS . . . . . . . . . . . . .
20.2 The equation for turbulent kinetic energy in hybrid LES-RANS . . .
20.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
145
150
150
151
.
.
.
.
152
152
152
154
154
156
156
157
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
22.3
22.4
22.5
22.6
22.7
22.8
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
A TME075: identity
157
157
157
157
158
158
161
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
162
163
163
163
163
164
164
164
164
164
165
165
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
166
166
166
167
167
167
167
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
168
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
170
170
171
173
173
175
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
176
176
177
177
178
178
179
179
179
179
180
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
F.11
F.12
F.13
Dissipation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Two-point correlations . . . . . . . . . . . . . . . . . . . . . . . . .
Do something fun! . . . . . . . . . . . . . . . . . . . . . . . . . . .
180
181
182
183
188
189
191
191
192
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
193
193
193
194
195
195
197
198
L MTF270, Assignment 2:
L.1 Task 2.1 . . . . .
L.2 Task 2.2 . . . . .
L.3 Task 2.3 . . . . .
L.4 Task 2.4 . . . . .
L.5 Task 2.5 . . . . .
L.6 Task 2.6 . . . . .
L.7 Task 2.7 . . . . .
L.8 Task 2.9 . . . . .
L.9 Task 2.10 . . . . .
L.10 Task 2.11 . . . . .
L.11 Task 2.12 . . . . .
LES
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
199
199
200
200
201
202
202
203
204
204
204
204
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
205
205
205
208
209
N Transformation of a tensor
N.1 Rotation to principal directions . . . . . . . . . . . . . . . . . . . .
N.2 Transformation of a velocity gradient . . . . . . . . . . . . . . . . .
211
212
213
214
P Greens formulas
P.1
Greens rst formula . . . . . . . . . . . . . . . . . . . . . . . . . .
P.2
Greens second formula . . . . . . . . . . . . . . . . . . . . . . . .
223
223
223
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
P.3
P.4
Q References
223
226
227
1. Motion, ow
9
T (Xi , t1 )
Xi
T (x1 , t1 )
i
T (Xi , t2 )
T (x2 , t2 )
i
Figure 1.1: The temperature of a uid particle described in Lagrangian, T (Xi , t), or
Eulerian, T (xi , t), approach.
1 Motion, ow
Recommended exercises: Mase [2] 4.10, 4.12, 4.14, 4.15, 4.17, 4.18, 4.19, 4.26, 4.27,
4.30, 4.36, 4.41, 4.42, 4.44, 4.45, 5.22, 5.23 1
(1.1)
10
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
x2
12
11
13
x1
Figure 1.2: Denition of stress components on a surface.
Note that we have to use partial derivative on T since it is a function of more than one
(independent) variable. The rst term on the right side is the local rate of change; by
this we mean that it describes the variation of T in time at position xi . The second term
on the right side is called the convective rate of change, which means that it describes
the variation of T in space when is passes the point xi . The left side in Eq. 1.1 is called
the material derivative and is in this text denoted by dT /dt.
Equation 1.1 can be illustrated as follows. Put your nger out in the blowing wind.
The temperature gradient youre nger experiences is T /t. Imagine that youre a
uid particle and that you ride on a bike. The temperature gradient you experience is
the material derivative, dT /dx.
Exercise 1 Write out Eq. 1.1, term-by-term.
(1.2)
dvi
ji
+ fi
=
dt
xj
(1.3)
where the rst and the second term on the right side represents, respectively, the net
force due to surface and volume forces (ij denotes the stress tensor). Stress is force
per unit area. The rst term includes the viscous stress tensor, ij . As you have learnt
earlier, the rst index relates to the surface at which the stress acts and the second
index is related to the stress component. For example, on a surface whose normal is
ni = (1, 0, 0) act the three stress components 11 , 12 and 13 , see Fig. 1.2.
The diagonal components of ij represent normal stresses and the off-diagonal
components of ij represent the shear stresses. In Part I you learnt that the pressure is
dened as minus the sum of the normal stress, i.e. p = kk /3. The pressure, p, acts
local rate
of change
Conv. rate
of change
Material
derivative
11
(1.4)
ij is the deviator of ij . The expression for the viscous stress tensor is found in Eq. 2.4
at p. 17. The minus-sign in front of p appears because the pressure acts into the surface.
When theres no movement, the viscous stresses are zero and then of course the normal
stresses are the same as the pressure. In general, however, the normal stresses are the
sum of the pressure and the viscous stresses, i.e.
11 = p + 11 ,
22 = p + 22 ,
33 = p + 33 ,
(1.5)
Exercise 2 Consider Fig. 1.2. Show how 21 , 22 , 23 act on a surface with normal
vector ni = (0, 1, 0). Show also how 31 , 32 , 33 act on a surface with normal vector
ni = (0, 0, 1).
Exercise 3 Write out Eq. 1.4 on matrix form.
1 vi
vi vj
vj
vi
=
+
xj
2 xj
xj xi
xi
=
1
2
(1.6)
=0
2vi /xj
vi
vj
+
xj
xi
1
2
vi
vj
xj
xi
= Dij + Vij
where
Dij is a symmetric tensor called the strain-rate tensor
Vij is a anti-symmetric tensor called the vorticity tensor
The vorticity tensor is related to the familiar vorticity vector which is the curl of
the velocity vector, i.e. q = v, or in tensor notation2
qi = ijk
vk
xj
(1.7)
(1.8)
2 In this course we use Mases notation for the vorticity vector, q . Note, however, that in the literature
i
the common notation is i .
Strain-rate
tensor
vorticity tensor
12
The vorticity represents rotation of a uid particle. Inserting Eq. 1.6 into Eq. 1.7
gives
qi = ijk (Dkj + Vkj ) = ijk Vkj
(1.9)
since ijk Dkj = 0 because the product of a symmetric tensor (Dkj ) and a antisymmetric tensor (ijk ) is zero (see Solved Problem no. 1.37 in Mase). Let us show
this for i = 1 by writing out the full equation. Recall that Dij = Dji (i.e. D12 = D21 ,
D13 = D31 , D23 = D32 ) and ijk = ikj = jki etc (i.e. 123 = 132 = 231 . . . ,
113 = 221 = . . . 331 = 0)
1jk Dkj = 111 D11 + 112 D21 + 113 D31
+ 121 D12 + 122 D22 + 123 D32
+ 131 D13 + 132 D23 + 133 D33
= 0 D11 + 0 D21 + 0 D31
+ 0 D12 + 0 D22 + 1 D32
(1.10)
(1.11)
(1.12)
This can easily be proven by writing all the componenents, see Table A.1 at p. A.1.
Hence we get with Eq. 1.7
1
Vm = im qi
(1.13)
2
or, in a more common form
1
Vij = ijk qk
(1.14)
2
A much easier way to go from Eq. 1.9 to Eq. 1.14 is to write out the components of
Eq. 1.9. Here we do it for i = 1
q1 = 123 V32 + 132 V23 = V32 V23 = 2V23
(1.15)
and we get
1
V23 = q1
2
which indeed is identical to Eq. 1.14.
(1.16)
Exercise 4 Write out the second and third component of the vorticity vector given in
Eq. 1.7 (i.e. q2 and q3 ).
Exercise 5 Complete the proof of Eq. 1.10 for i = 2 and i = 3.
Exercise 6 Write out Eq. 1.15 also for i = 2 and i = 3 and nd an expression for V12
and V13 (cf. Eq. 1.16). Show that you get the same result as in Eq. 1.14.
13
x1
v1
x2 t
x2
x2
v2
x1 t
x1
x2
x1
Figure 1.3: Rotation of a uid particle during time t. Here v1 /x2 = v2 /x1
so that V12 = q3 /2 = v2 /x1 > 0.
Exercise 7 In Eq. 1.16 we proved the relation between Vij and qi for the off-diagonal
components. What about the diagonal components of Vij ? What do you get from
Eq. 1.6?
Exercise 8 From you course in linear algebra, you should remember how to compute
a vector product using Sarrus rule. Use it to compute the vector product
1
e
2
e
3
e
q = v = x1 x2 x3
v1
v2
v3
Verify that this agrees with the expression in tensor notation in Eq. 1.7.
rotation
14
x1
v1
x2 t
x2
x2
v2
x1 t
x1
x2
x1
Figure 1.4: Deformation of a uid particle by shear during time t. Here v1 /x2 =
v2 /x1 so that D12 = v1 /x2 > 0.
x1
v1
x1 t
x1
x2
x2
x1
Figure 1.5: Deformation of a uid particle by elongation during time t.
In general, a uid particle experiences a combination of rotation, deformation and
elongation as indeed is given by Eq. 1.6.
Exercise 9 Consider Fig. 1.3. Show and formulate the rotation by q1 .
15
Exercise 10 Consider Fig. 1.4. Show and formulate the deformation by D23 .
Exercise 11 Consider Fig. 1.5. Show and formulate the elongation by D22 .
vi =
(1.17)
xi
1.5.1 Ideal vortex line
The ideal vortex line is an irrotational (potential) ow where the uid moves along
circular paths, see Fig. 1.6. The velocity eld in polar coordinates reads
v =
,
2r
vr = 0
(1.18)
(1.19)
1
r
x2
,
2(x2 + x2 )
1
2
v2 =
x1
.
2(x2 + x2 )
1
2
(1.20)
To verify that this ow is a potential ow, we need to show that the vorticity, qi =
ijk vk /xj is zero. Since it is a two-dimensional ow (v3 = /x3 = 0), q1 = q2 =
0, we only need to compute q3 = v2 /x1 v1 /x2 . The velocity derivatives are
obtained as
x2 x2
x2 x2
v1
v2
1
2
2
1
=
=
(1.21)
2,
x2
2 (x2 + x2 )
x1
2 (x2 + x2 )2
1
2
1
2
and we get
q3 =
(x2 x2 + x2 x2 ) = 0
1
1
2
2 (x2 + x2 )2 2
1
2
(1.22)
which shows that the ow is indeed a potential ow, i.e. irrotational (qi 0).
It may be little confusing that the ow path forms a vortex but the ow itself has no
vorticity. Thus one must be very careful when using the words vortex and vorticity.
By vortex we usually mean a recirculation region of the mean ow. That the ow has
no vorticity (i.e. no rotation) means that a uid particle moves as illustrated in Fig. 1.6.
vortex vs.
vorticity
16
Figure 1.6: Ideal vortex. The uid particle does not rotate.
v1
v1
x2
a
x3
x1
v2 = 0
(1.23)
with c > 0, see Fig. 1.7. The vorticity vector for this ow reads
q1 = q2 = 0,
q3 =
v1
v2
= c
x1
x2
(1.24)
When the uid particle is moving from position a, via b to position c it is indeed
rotating. It is rotating in clockwise direction. Note that the positive rotating direction
is dened as the counter-clockwise direction, indicated by in Fig. 1.7. This is why
the vorticity, q3 , is negative (= c).
2. Governing ow equations
17
2 Governing ow equations
Recommended exercises: Mase 5.1, 5.25, 5.26, 5.37, 7.1, 7.2, 7.14, 7.27, 7.28, 7.29,
7.46 3
(2.2)
vi
=0
xi
(2.3)
(2.4)
Inserting Eq. 2.4 into the balance equations, Eq. 1.3, we get
ji
p
p
dvi
+
+ fi =
+
=
dt
xi xj
xi xj
2 vk
2 Dij
ij
3 xk
+ fi (2.5)
where denotes the dynamic viscosity.4 This is the Navier-Stokes equations (sometimes the continuity equation is also included in the name Navier-Stokes). It is also
called the transport equation for momentum. If the viscosity, , is constant it can be
moved outside the derivative. Furthermore, if the ow is incompressible the second
term in the parenthesis on the right side is zero because of the continuity equation. If
these two requirements are satised we can also re-write the rst term in the parenthesis
as
2 vi
vj
vi
=
(2 Dij ) =
+
(2.6)
xj
xj xj
xi
xj xj
because of the continuity equation. Equation 2.5 can now for constant and incompressible ow be written
3 priority
4 In
2 vi
p
dvi
+
+ fi
=
dt
xi
xj xj
(2.7)
18
Exercise 12 Formulate the Navier-Stokes equation for incompressible ow but nonconstant viscosity.
(2.8)
where u denotes inner energy. ci denotes the conductive heat ux5 and z the net radiative heat source. The latter can also be seen as a vector, zi,rad ; for simplicity, we
neglect the radiation from here on. Change of notation gives
vi
ci
du
= ji
dt
xj
xi
(2.9)
In Part I we formulated the constitutive law for the heat ux vector (Fouriers law)
ci = k
T
xi
(2.10)
Inserting the constitutive laws, Eqs. 2.4 and 2.10, into Eq. 2.9 gives
du
2
vi
T
xi
(2.11)
where we have used Dij vi /xj = Dij (Dij + Vij ) = Dij Dij because the product
of a symmetric tensor, Dij , and an anti-symmetric tensor, Vij , is zero (Solved problem
no. 1.37 in Mase). Two of the viscous terms (denoted by ) represent irreversible viscous heating (i.e. transformation of kinetic energy into thermal energy); these terms are
important at high-speed ow6 (for example re-entry from outer space) and for highly
viscous ows (lubricants). The rst term on the right side represents reversible heating and cooling due to compression and expansion of the uid. Equation 2.11 is the
transport equation for (inner) energy, u.
Now we assume that the ow is incompressible for which
du = cp dT
(2.12)
where cp is the heat capacity (see Part I) so that Eq. 2.11 gives
cp
5 In this
dT
=+
dt
xi
T
xi
(2.13)
course we use Mases notation, ci , for the heat ux vector. In the literature, the common notation
is qi .
6 High-speed ows relevant for aeronautics will be treated in detail in the course Compressible ow in
the MSc programme.
19
(2.14)
where = k/(cp ) is the thermal diffusivity. From the thermal diffusivity, the Prandtl
number
Pr =
(2.15)
1 d(vi vi )
dk
dvi
=
=
dt
2
dt
dt
(2.17)
(vi vi /2 = v 2 /2 = k) so that
dk
ji
+ vi fi
= vi
dt
xj
(2.18)
vi
vi ji
dk
ji
+ vi fi
=
dt
xj
xj
(2.19)
This is the transport equation for kinetic energy, k. Adding Eq. 2.19 to Eq. 2.9 gives
d(u + k)
ci
ji vi
+ vi fi
=
dt
xj
xi
(2.20)
This is an equation for the sum of inner and kinetic energy, u + k. This is the transport
equation for total energy, u + k.
Let us take a closer look at Eqs. 2.9, 2.19 and 2.20. First we separate the term
ji vi /xj in Eqs. 2.9 and 2.19 into work related to the pressure and viscous stresses
respectively (see Eq. 1.4), i.e.
ji
vi
vi
vi
= p
+ ji
xj
xi
xj
a
b=
(2.21)
thermal
diffusivity
20
The physical meaning of the a-term in Eq. 2.21 which include the pressure, p
is heating/cooling by compression/expansion. This is a reversible process, i.e.
no loss of energy but only transformation of energy.
The physical meaning of the b-term in Eq. 2.21 which include the viscous stress
tensor, ij is a dissipation, which means that kinetic energy is transformed to
thermal energy. It is denoted , see Eq. 2.11, and is called viscous dissipation.
It is always positive and represents irreversible heating.
The dissipation, , appears as a sink term in the equation for the kinetic energy,
k (Eq. 2.19) and it appears a source term in the equation for the inner energy,
u (Eq. 2.9). The transformation of kinetic energy into inner energy takes place
through this source term.
, does not appear in the equation for the total energy u + k (Eq. 2.20); this
makes sense since only affects u and k, not its sum, u + k.
Dissipation is very important in turbulence where transfer of energy takes place at
several levels. First energy is transferred from the mean ow to the turbulent uctuations. The physical process is called production of turbulent kinetic energy. Then we
have transformation of kinetic energy from turbulence kinetic energy to thermal energy; this is turbulence dissipation (or heating). At the same time we have the usual
viscous dissipation from the mean ow to thermal energy, but this is much smaller than
that from the turbulence kinetic energy to thermal energy. For more detail, see section
2.4 in [3]7 .
=
+ vj
dt
t
xj
(2.22)
This is often called the non-conservative form. Using the continuity equation, Eq. 2.2,
it can be re-written as
d
+
=
+ vj
dt
t
xj
vj
d
+
dt
xj
=0
+ vj
+
t
xj
=
(2.23)
vj
+ vj
+
t
xj
xj
The two underlined terms will form a time derivative term, and the other three terms
can be collected into a convective term, i.e.
d
vj
=
+
dt
t
xj
(2.24)
nonconservative
21
x2
V0
x1
Figure 3.1: The plate moves to the right with speed V0 for t > 0.
This is called the conservative form. With = 1 we get the continuity equation. When
solving the Navier-Stokes equations numerically using nite volume methods, the left
side in the transport equation is always written in the form of Eq. 2.24; in this way we
ensure that the transported quantity is conserved. The results may be inaccurate due
to too coarse a numerical grid, but no mass, momentum, energy etc is lost (provided a
transport equation for the quantity is solved): what comes in goes out.
dV =
V
d
vi
+
dt
xi
dV =
V
vi
+
t
xi
dV
(2.25)
was formulated in order to move the time derivative inside the integral. in the equation above can be (mass), vi (momentum) or u (energy). This equation applies to
any volume at every instant and the restriction to a material particle is no longer necessary. Hence, in uid mechanics the transport equations (Eqs. 2.2, 2.5, 2.9, . . . ) are
valid both for a material particle as well as for a volume; the latter is usually xed (this
is not necessary).
7 can
2 v1
v1
=
t
x2
2
(3.1)
conservative
22
x2
t3
t2
t1
0
0.2
0.4
0.6
0.8
v1 /V0
Figure 3.2: The v1 velocity at thee different times. t3 > t2 > t1 .
We will nd that the diffusion process depends on the kinematic viscosity, = /,
rather than the dynamic one, . The boundary conditions for Eq. 3.1 are
v1 (x2 , t = 0) = 0,
v1 (x2 = 0, t) = V0 ,
v1 (x2 , t) = 0
(3.2)
The solution to Eq. 3.1 is shown in Fig. 3.2. For increasing time (t3 > t2 > t1 ), the
moving plate affects the uid further and further away from the plate.
It turns out that the solution to Eq. 3.1 is a similarity solution; this means that the
number of independent variables is reduced by one, in this case from two (x2 and t) to
one (). The similarity variable, , is related to x2 and t as
x2
=
2 t
(3.3)
If the solution of Eq. 3.1 depends only on , it means that the solution for a given uid
will the same (similar) for many (innite) values of x2 and t as long as the ratio
be
x2 / t is constant. Now we need to transform the derivatives in Eq. 3.1 from /t and
/x2 to d/d so that it becomes a function of only. We get
dv1
x2 t3/2 dv1
1 dv1
v1
=
=
=
t
d t
2 t d
4 d
dv1
1 dv1
v1
=
=
x2
d x2
2 t d
1
2 v1
v1
1 dv1
=
=
=
x2
x2 x2
x2 2 t d
2 t x2
2
dv1
d
1 d2 v1
4t d 2
(3.4)
v1
V0
(3.5)
(3.6)
similarity
solution
23
3
2.5
2
1.5
1
0.5
0
0
0.2
0.4
0.6
0.8
f
Figure 3.3: The velocity, f = v1 /V0 , given by Eq. 3.11.
We have now successfully transformed Eq. 3.1 and reduced the number of independent
variables from two to one. Now let us nd out if the boundary conditions, Eq. 3.2, also
can be transformed in a physically meaningful way; we get
v1 (x2 , t = 0) = 0 f ( ) = 0
v1 (x2 = 0, t) = V0 f ( = 0) = 1
(3.7)
v1 (x2 , t) = 0 f ( ) = 0
Since we managed to transform both the equation (Eq. 3.1) and the boundary conditions
(Eq. 3.7) we conclude that the transformation is suitable.
Now let us solve Eq. 3.6. Integration once gives
df
= C1 exp( 2 )
d
(3.8)
exp( 2 )d + C2
f = C1
(3.9)
exp( 2 )
(3.10)
At the limits, the error function takes the values 0 and 1, i.e. erf(0) = 0 and erf(
) = 1. Taking into account the boundary conditions, Eq. 3.7, the nal solution to
(3.11)
The solution is presented in Fig. 3.3. Compare this gure with Fig. 3.2 at p. 22; all
graphs in that gure collapse into one graph in Fig. 3.3. To compute the velocity, v1 ,
we pick a time t and insert x2 and t in Eq. 3.3. Then f is obtained from Eq. 3.11 and
the velocity, v1 , is computed from Eq. 3.5. This is how the graphs in Fig. 3.2 were
obtained.
24
3
2.5
2
1.5
1
0.5
0
6
21 /(V0 )
Figure 3.4: The shear stress for water ( = 106 ) obtained from Eq. 3.12 at time
t = 100 000.
From the velocity prole we can get the shear stress as
21 =
V0 df
V0
v1
=
=
exp 2
x2
2 t d
t
(3.12)
where we used = /. Figure 3.4 below presents the shear stress, 21 . The solid
line is obtained from Eq. 3.12 and circles are obtained by evaluating the derivative,
df /d, numerically using central differences (fj+1 fj1 )/(j+1 j1 ).
As can be seen from Fig. 3.4, the magnitude of the shear stress increases for de
creasing and it is largest at the wall, w = V0 / t
The vorticity, q3 , across the boundary layer is computed from its denition (Eq. 1.24)
q3 =
V0 df
V0
v1
=
=
exp( 2 )
x2
2 t d
t
(3.13)
From Fig. 3.2 at p. 22 it is seen that for large times, the moving plate is felt further
and further out in the ow, i.e. the thickness of the boundary layer, , increases. Often
the boundary layer thickness is dened by the position where the local velocity, v1 (x2 ),
reaches 99% of the freestream velocity. In our case, this corresponds to the point where
v1 = 0.01V0 . From Fig. 3.3 and Eq. 3.11 we nd that this occurs at
= 1.8 =
2 t
= 3.6 t
(3.14)
It can be seen that the boundary layer thickness increases with t1/2 . Equation 3.14 can
also be used to estimate the diffusion length. After, say, 10 minutes the diffusion length
for air and water, respectively, are
air = 10.8cm
water = 2.8cm
(3.15)
As mentioned in the beginning of this section, note that the diffusion length is determined by the kinematic viscosity, = / rather than by dynamic one, .
Exercise 14 Consider the graphs in Fig. 3.3. Create this graph with Matlab.
diffusion
length
25
V
P1
h
P2
P1
V
x2
x1
Figure 3.5: Flow in a horizontal channel. The inlet part of the channel is shown.
Exercise 15 Consider the graphs in Fig. 3.2. Note that no scale is used on the x2 axis
and that no numbers are given for t1 , t2 and t3 . Create this graph with Matlab for both
air and engine oil. Choose suitable values on t1 , t2 and t3 .
Exercise 16 Repeat the exercise above for the shear stress, 21 , see Fig. 3.4.
26
P2
P1
x2
x1
(3.16)
where V denotes the bulk (i.e. the mean) velocity, and Dh = 4A/Sp where Dh ,
A and Sp denote the hydraulic diameter, the cross-sectional area and the perimeter,
respectively. For ow between two plates we get Dh = 2h.
Let us nd the governing equations for the fully developed ow region; in this
region the ow does not change with respect to the streamwise coordinate, x1 (i.e.
v1 /x1 = v2 /x1 = 0). Since the ow is two-dimensional, it does not depend
on the third coordinate direction, x3 (i.e. /x3 ), and the velocity in this direction is
zero, i.e. v3 = 0. Taking these restrictions into account the continuity equation can be
simplied as (see Eq. 2.3)
v2
=0
(3.17)
x2
Integration gives v2 = C1 and since v2 = 0 at the walls, it means that
v2 = 0
(3.18)
across the entire channel (recall that we are dealing with the part of the channel where
the ow is fully developed; in the inlet section v2 = 0, see Fig. 3.5).
Now let us turn our attention to the momentum equation for v2 . This is the vertical
direction (x2 is positive upwards, see Fig. 3.5). The gravity acts in the negative x2
direction, i.e. fi = (0, g, 0). The momentum equation can be written (see Eq. 2.7
at p. 17)
v2
v2
p
2 v2
dv2
+ v2
=
+
g
(3.19)
v1
dt
x1
x2
x2
x2
2
Since v2 = 0 we get
p
= g
x2
(3.20)
p = gx2 + C1 (x1 )
(3.21)
Integration gives
27
(3.22)
Hence the pressure, p, decreases with vertical height. This agrees with our experience
that the pressure decreases at high altitudes in the atmosphere and increases the deeper
we dive into the sea. Usually the hydrostatic pressure, P , is used in incompressible
ow. This pressure is zero when the ow is static, i.e. when the velocity eld is zero.
However, when you want the physical pressure, the gx2 as well as the surrounding
atmospheric pressure must be added.
We can now formulate the momentum equation in the streamwise direction
dv1
v1
v1
dP
2 v1
+ v2
=
+
v1
dt
x1
x2
dx1
x2
2
(3.23)
where p was replaced by P using Eq. 3.22. Since v2 = v1 /x1 = 0 the left side is
zero so
dP
2 v1
=
(3.24)
x2
dx1
2
Since the left side is a function of x2 and the right side is a function of x1 , we conclude
that they both are equal to a constant. The velocity, v1 , is zero at the walls, i.e.
v1 (0) = v1 (h) = 0
(3.25)
x2
h dP
x2 1
dx
2
h
1
(3.26)
The minus sign on the right side appears because the pressure gradient is decreasing
for increasing x1 ; the pressure is driving the ow. The negative pressure gradient is
constant and can be written as dP/dx1 = P/L.
The velocity takes its maximum in the center, i.e. for x2 = h/2, and reads
v1,max =
h P h
2 L 2
1
2
h2 P
8 L
(3.27)
(3.28)
The mean velocity (often called the bulk velocity) is obtained by integrating Eq. 3.28
across the channel, i.e.
v1,mean =
v1,max
h
h
0
4x2 1
2
x2
dx2 = v1,max
h
3
(3.29)
hydrostatic
pressure
28
0.8
0.6
x2 /h
0.4
0.2
0
0
0.2
0.4
0.6
0.8
v1 /v1,max
Figure 3.7: The velocity prole in fully developed channel ow, Eq. 3.28.
3.2.3 Force balance
To formulate a force balance, we start with Eq. 1.3 which reads for i = 1
dv1
j1
=
dt
xj
(3.31)
The left hand side is zero since the ow is fully developed. Forces act on a volume and
its bounding surface. Hence we integrate Eq. 3.31 over the volume of a slice (length
L), see Fig. 3.8
j1
0=
dV
(3.32)
V xj
Recall that this is the form on which we originally derived the the momentum balance
(Newtons second law) in Part I. Now use Gauss divergence theorem
0=
V
j1
dV =
xj
j1 nj dS
(3.33)
The bounding surface consists in our case of four surfaces (lower, upper, left and right)
so that
Slef t
Supper
Slower
Sright
j1 nj dS (3.34)
j1 nj dS +
j1 nj dS +
j1 nj dS +
0=
The normal vector on the lower, upper, left and right are ni,lower = (0, 1, 0), ni,upper =
(0, 1, 0), ni,lef t = (1, 0, 0), ni,right = (1, 0, 0). Inserting the normal vectors and using Eq. 1.4 give
0=
(p + 11 )dS +
Slef t
Sright
(p + 11 )dS
21 dS
21 dS +
Slower
Supper
(3.35)
11 = 0 because v1 /x1 = 0 (fully developed ow). The shear stress at the upper and
lower surfaces have opposite sign because w = (v1 /x2 )lower = (v1 /x2 )upper .
Using this and Eq. 3.22 give (the gravitation term on the left and right surface cancels
and P and w are constants and can thus be taken out in front of the integration)
0 = P1 W h P2 W h 2w LW
(3.36)
29
L
w,U
P1
P2
walls
x2
w,L
x1
Figure 3.8: Force balance of the ow between two plates.
where W is the width (in x3 direction) of the two plates (for convenience we set W =
1). With P = P1 P2 we get Eq. 3.30.
3.2.4 Balance equation for the kinetic energy
In this subsection we will use the equation for kinetic energy, Eq. 2.19. Let us integrate
this equation in the same way as we did for the force balance. The left side of Eq. 2.19
is zero because we assume that the ow is fully developed; using Eq. 1.4 gives
0=
vi
vi ji
ji
+ vi fi
xj
xj
=0
vi
vi
vj p vi ji
+
+ pij
ji
=
xj
xj
xj
xj
(3.37)
pvj
ji vi
+
dV
xj
xj
(3.38)
(pvj + ji vi )nj dS
dV
(3.39)
where S is the surface bounding the volume. The unit normal vector is denoted by nj
which points out from the volume. For example, on the right surface in Fig. 3.8 it is
nj = (1, 0, 0) and on the lower surface it is nj = (0, 1, 0). Now we apply Eq. 3.39
to the uid enclosed by the at plates in Fig. 3.8. The second term is zero on all
four surfaces and the rst term is zero on the lower and upper surfaces (see Exercises
below). We replace the pressure p with P using Eq. 3.22 so that
Slef t &Sright
v1 n1 dS
Slef t &Sright
= P v1,mean W h
30
21 (x + 0.5x2 )
2
v1 = cx2
p(x 0.5x1 )
1
x2
(x , x )
1
2
p(x + 0.5x1 )
1
21 (x 0.5x2 )
2
x1
Figure 4.1: Surface forces in the x1 direction acting on a uid particle (assuming 11 =
0).
because gx2 n1 v1 on the left and right surfaces cancels; P can be taken out of the
integral as it does not depend on x2 . Finally we get
P =
1
W hv1,mean
dV
(3.40)
Exercise 17 For the fully developed ow, compute the vorticity, qi , using the exact
solution (Eq. 3.28).
Exercise 18 Show that the rst and second terms in Eq. 3.39 are zero on the upper and
the lower surfaces in Fig. 3.8.
Exercise 19 Show that the second term in Eq. 3.39 is zero also on the left and right
surfaces Fig. 3.8.
Exercise 20 Using the exact solution, compute the dissipation, , for the fully developed ow.
Exercise 21 From the dissipation, compute the pressure drop. Is it the same as that
obtained from the force balance (if not, nd the error; it should be!).
31
x1 momentum equation acting on a uid particle in a shear ow. Looking at Fig. 4.1 it
is obvious that only the shear stresses are able to rotate the uid particle; the pressure
acts through the center of the uid particle and is thus not able to affect rotation of the
uid particle.
Let us have a look at the momentum equations in order to show that the viscous
terms indeed can be formulated with the vorticity vector, qi . In incompressible ow
the viscous terms read (see Eqs. 2.4, 2.5 and 2.6)
2 vi
ji
=
xj
xj xj
(4.1)
xj xj
xj xi
2 vi
2 vj
xj xi
xj xj
2 vj
2 vk
inm mjk
xj xi
xj xn
(4.2)
The rst on the right side is zero because of continuity. Lets verify that
2 vj
2 vi
xj xi
xj xj
= inm mjk
2 vk
xj xn
(4.3)
2 vk
2 vk
2 vi
2 vk
= = (ij nk ik nj )
=
xj xn
xj xn
xi xk
xj xj
(4.4)
The rst term on the right side is zero because of continuity and hence we nd that
Eq. 4.2 can indeed be written as
2 vk
2 vi
= inm mjk
xj xj
xj xn
(4.5)
(4.6)
(4.7)
(4.8)
Thus, there is a one-to-one relation between the viscous term and vorticity: no viscous
terms means no vorticity and vice versa. Hence, inviscid ow (i.e. friction-less ow)
has no rotation. (The exception is when vorticity is transported into an inviscid region,
but also in that case no vorticity is created or destroyed: it stays constant, unaffected.)
Inviscid ow is often called irrotational ow (i.e. no rotation) or potential ow.
The main points that we have learnt in this section are:
potential
32
1. The viscous terms are responsible for creating vorticity; this means that the vorticity cant be created or destroyed in inviscid (friction-less) ow
2. The viscous terms in the momentum equations can be expressed in qi ; considering Item 1 this was to be expected.
Exercise 22 Prove the rst equality of Eq. 4.6 using the --identity.
Exercise 23 Write out Eq. 4.8 for i = 1 and verify that it is satised.
1
vi
= vj (Dij + Vij ) = vj Dij ijk qk
xj
2
(4.9)
where Eq. 1.14 on p. 12 was used. Inserting Dij = (vi /xj + vj /xi )/2 and
multiplying by two gives
2vj
vi
= vj
xj
vi
vj
+
xj
xi
ijk vj qk
(4.10)
vj
1 (vj vj )
=
xi
2 xi
(4.11)
( 1 v 2 )
vi
= 2
ijk vj qk
xj
xi
no rotation
(4.12)
rotation
where v 2 = vj vj . The last term on the right side is the vector product of v and q, i.e.
v q.
The trick we have achieved is to split the convective term into one term without
rotation (rst term on the right side of Eq. 4.12) and one term including rotation (second
term on the right side). Inserting Eq. 4.12 into the incompressible momentum equation
(Eq. 2.7) yields
1 v2
1 p
vi
2 vi
ijk vj qk =
+
+ fi
+ 2
t
xi
xi
xj xj
no rotation
rotation
(4.13)
frictionless
33
The volume source is in most engineering ows represented by the gravity which is
conservative meaning that it is uniquely determined by the position (in this case the
vertical position). Hence it can be expressed as a potential, h; we write
fi = gi =
h
xi
(4.14)
The negative sign appears because height is dened positive upwards and the direction
of gravity is downwards.
Since the vorticity vector is dened by the cross product pqi vi /xq ( v in
vector notation, see Exercise 8), we start by applying the operator pqi /xq to the
Navier-Stokes equation (Eq. 4.13) so that
1
2 2 v2
vj qk
2 vi
+ pqi
pqi ijk
txq
xi xq
xq
1 2p
3 vi
2h
= pqi
+ pqi
pqi
xi xq
xj xj xq
xq xi
pqi
(4.15)
where the body force in Eq. 4.15 was re-written using 4.14. We know that ijk is antisymmetric in all indices, and hence the second term on line 1 and the rst and the last
term on line 2 are all zero (product of a symmetric and an anti-symmetric tensor). The
last term on line 1 is re-written using the - identity (see Table A.1 at p. A.1)
pqi ijk
vj qk
vp qk
vq qp
vj qk
= (pj qk pk qj )
=
xq
xq
xk
xq
(4.16)
qi
=
xi
xi
ijk
vk
xj
= ijk
2 vk
=0
xj xi
(4.17)
vj qk
vp
qp
= qk
vk
xq
xk
xk
(4.18)
2
3 vi
=
xj xj xq
xj xj
pqi
vi
xq
2 qp
xj xj
(4.19)
Inserting Eqs. 4.18 and 4.19 into Eq. 4.15 gives nally
vp
qp
2 qp
qp
dqi
= qk
+
+ vk
dt
t
xk
xk
xj xj
(4.20)
We recognize the usual unsteady term, the convective term and the diffusive term. Furthermore, we have got rid of the pressure gradient term. That makes sense, because as
mentioned in connection to Fig. 4.1, the pressure cannot affect the rotation (i.e. the vorticity) of a uid particle since the pressure acts through the center of the uid particle.
34
q1
x2
x1
Figure 4.2: Vortex stretching.
q2
v1 (x2 )
x2
q2
x1
q v1 +q v1 + q v1
1
2
3
x1
x2
x3
vp
v2
v2
v2
qk
(4.21)
=
+q2
+ q3
q
1 x1
xk
x2
x3
v3
v3
v3
q
1
+q2
+ q3
x1
x2
x3
The diagonal terms in this matrix represent vortex stretching. Imagine a slender,
cylindrical uid particle with vorticity qi . We introduce a cylindrical coordinate system
with the x1 -axis as the cylinder axis and r2 as the radial coordinate (see Fig. 4.2) so
that qi = (q1 , 0, 0). We assume that a positive v1 /x1 is acting on the uid cylinder;
it will stretch the cylinder. From the continuity equation in polar coordinates
v1
1 (rvr )
+
=0
x1
r r
Vortex
stretching
(4.22)
we nd that since v1 /x1 > 0 the radial derivative of the radial velocity, vr , must
be negative, i.e. the radius of the cylinder will decrease. Hence vortex stretching will
either make a uid element longer and thinner (as in the example above) or shorter and
thicker (when v1 /x1 < 0).
The off-diagonal terms in Eq. 4.21 represent vortex tilting. Again, take a slender
uid particle, but this time with its axis aligned with the x2 axis, see Fig. 4.3. The
velocity gradient v1 /x2 will tilt the uid particle so that it rotates in clock-wise
direction. The second term q2 v1 /x2 in line one in Eq. 4.21 gives a contribution to
q1 . This means that vorticity in the x2 direction creates vorticity in the x1 direction..
Vortex stretching and tilting are physical phenomena which act in three dimensions:
uid which initially is two dimensional becomes quickly three dimensional through
Vortex
tilting
35
these phenomena. Vorticity is useful when explaining why turbulence must be threedimensional. For more detail, see section 2.4 in [3]9 . You will learn more about this in
other MSc courses on turbulent ow.
Exercise 24 The transport equation for q is derived above (Eq. 4.20). It is also derived
in Mase, Exercises 5.30. The two equations do not agree. Why? Is there a mistake in
the derivation above? By Mase? Or . . .
(4.23)
(Greek indices are used to indicate that they take values 1 or 2). This equation is
exactly the same as the transport equation for temperature in incompressible ow, see
Eq. 2.14. This means that vorticity diffuses in the same way as temperature does. In
fully developed channel ow, for example, the vorticity and the temperature equations
reduce to
2 q3
x2
2
2T
0=k 2
x2
0=
(4.24)
For the temperature equation the heat ux is given by c2 = T /x2 ; with a hot lower
wall and a cold upper wall (constant wall temperatures) the heat ux is constant and
goes from the lower wall to the upper wall. We have the same situation for the vorticity.
Its gradient, i.e. the vorticity ux, q3 /x2 , is constant across the channel. You have
plotted this quantity in Assignment 2.
4.3.1 Diffusion length from the Rayleigh problem
In Section 3.1 we studied the Rayleigh problem (unsteady diffusion). The diffusion
time, t, or the diffusion length, , in Eq. 3.14 can now be used to estimate the thickness
of the boundary layer. The vorticity is generated at the wall and it diffuses away from
the wall. There is vorticity in a boundary layer and outside the boundary layer it is
zero.
Consider the boundary layer in Fig. 4.4. At the end of the plate the boundary
thickness is (L). The time it takes for a uid particle to travel along the plate is L/V0 .
During this time vorticity will diffuse the length according Eq. 3.14. If we assume
that the uid is air with the speed V0 = 3m/s and that the length of the plate L = 2m
we get from Eq. 3.14 that (L) = 1.2cm.
9 can
36
V0
x2
111111111
000000000
x1
111111
000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
L
Figure 4.4: Boundary layer. The boundary layer thickness, , increases for increasing
streamwise distance from leading edge (x1 = 0).
Exercise 25 Note that the estimate above is not quite accurate because in the Rayleigh
problem we assumed that the convective terms are zero, but in a developing boundary
layer they are not (v2 = 0 and v1 /x1 = 0). The proper way to solve the problem is
to use Blasius solution (you have probably learnt about this in your rst uid mechanics
course; if not, you should go and nd out). Blasius solution gives
5
=
1/2
x
Rex
(4.25)
37
5. Turbulence
38
5 Turbulence
5.1 Introduction
Almost all uid ow which we encounter in daily life is turbulent. Typical examples
are ow around (as well as in) cars, aeroplanes and buildings. The boundary layers
and the wakes around and after bluff bodies such as cars, aeroplanes and buildings are
turbulent. Also the ow and combustion in engines, both in piston engines and gas
turbines and combustors, are highly turbulent. Air movements in rooms are turbulent,
at least along the walls where wall-jets are formed. Hence, when we compute uid
ow it will most likely be turbulent.
In turbulent ow we usually divide the velocities in one time-averaged part vi ,
which is independent of time (when the mean ow is steady), and one uctuating part
vi so that vi = vi + vi .
There is no denition on turbulent ow, but it has a number of characteristic features (see Pope [5] and Tennekes & Lumley [6]) such as:
I. Irregularity. Turbulent ow is irregular, random and chaotic. The ow consists
of a spectrum of different scales (eddy sizes). We do not have any exact denition of
an turbulent eddy, but we suppose that it exists in a certain region in space for a certain
time and that it is subsequently destroyed (by the cascade process or by dissipation, see
below). It has a characteristic velocity and length (called a velocity and length scale).
The region covered by a large eddy may well enclose also smaller eddies. The largest
eddies are of the order of the ow geometry (i.e. boundary layer thickness, jet width,
etc). At the other end of the spectra we have the smallest eddies which are dissipated by
viscous forces (stresses) into thermal energy resulting in a temperature increase. Even
though turbulence is chaotic it is deterministic and is described by the Navier-Stokes
equations.
II. Diffusivity. In turbulent ow the diffusivity increases. This means that the
spreading rate of boundary layers, jets, etc. increases as the ow becomes turbulent.
The turbulence increases the exchange of momentum in e.g. boundary layers, and
reduces or delays thereby separation at bluff bodies such as cylinders, airfoils and cars.
The increased diffusivity also increases the resistance (wall friction) in internal ows
such as in channels and pipes.
III. Large Reynolds Numbers. Turbulent ow occurs at high Reynolds number.
For example, the transition to turbulent ow in pipes occurs that ReD 2300, and in
boundary layers at Rex 500 000.
IV. Three-Dimensional. Turbulent ow is always three-dimensional and unsteady.
However, when the equations are time averaged, we can treat the ow as two-dimensional
(if the geometry is two-dimensional).
V. Dissipation. Turbulent ow is dissipative, which means that kinetic energy in
the small (dissipative) eddies are transformed into thermal energy. The small eddies
receive the kinetic energy from slightly larger eddies. The slightly larger eddies receive
their energy from even larger eddies and so on. The largest eddies extract their energy
from the mean ow. This process of transferring energy from the largest turbulent
scales (eddies) to the smallest is called the cascade process.
VI. Continuum. Even though we have small turbulent scales in the ow they are
much larger than the molecular scale and we can treat the ow as a continuum.
turbulent
eddy
cascade
process
39
ow of kinetic energy
large scales
intermediate scales
....
dissipative scales
Figure 5.1: Cascade process with a spectrum of eddies. The energy-containing eddies
are denoted by v0 ; denotes the size of the eddies in the inertial subrange are denoted
2 < 1 < 0 ; is the dissipation at the dissipative eddies.
(5.1)
The kinetic energy of the large scales is lost to slightly smaller scales with which the
large scales interact. Through the cascade process the kinetic energy is in this way
transferred from the largest scale to the smallest scales. At the smallest scales the
frictional forces (viscous stresses) become large and the kinetic energy is transformed
(dissipated) into thermal energy.
The dissipation is denoted by which is energy per unit time and unit mass ( =
[m2 /s3 ]). The dissipation is proportional to the kinematic viscosity, , times the uctuating velocity gradient up to the power of two (see Section 8.1). The friction forces
exist of course at all scales, but they are largest at the smallest eddies. Thus it is not
quite true that eddies, which receive their kinetic energy from slightly larger scales,
give away all of that to the slightly smaller scales. This is an idealized picture; in reality a small fraction is dissipated. However it is assumed that most of the energy (say
90%) that goes into the large scales is nally dissipated at the smallest (dissipative)
scales.
The smallest scales where dissipation occurs are called the Kolmogorov scales
whose velocity scale is denoted by v , length scale by and time scale by . We
assume that these scales are determined by viscosity, , and dissipation, . The argument is as follows.
40
viscosity: Since the kinetic energy is destroyed by viscous forces it is natural to assume
that viscosity plays a part in determining these scales; the larger viscosity, the
larger scales.
dissipation: The amount of energy that is to be dissipated is . The more energy that
is to be transformed from kinetic energy to thermal energy, the larger the velocity
gradients must be.
Having assumed that the dissipative scales are determined by viscosity and dissipation,
we can express v , and in and using dimensional analysis. We write
v
=
a
b
2
2 3
[m/s] = [m /s] [m /s ]
(5.2)
where below each variable its dimensions are given. The dimensions of the left and the
right side must be the same. We get two equations, one for meters [m]
1 = 2a + 2b,
(5.3)
1 = a 3b,
(5.4)
which give a = b = 1/4. In the same way we obtain the expressions for and so
that
1/4
v = ()
, =
1/4
, =
1/2
(5.5)
f (x) =
1
a0 +
(an cos(n x) + bn sin(n x))
2
n=1
(5.6)
1
L
1
L
f (x)2 dx =
L 2
a +L
(a2 + b2 )
n
n
2 0
n=1
(5.7)
41
Pk
E()
II=
in
e
rti
al
ran
ge
E
)
(
3
5/
III=dissipating range
Figure 5.2: Spectrum for turbulent kinetic energy, k. I: Range for the large, energy
containing eddies. II: the inertial subrange. III: Range for small, isotropic scales.
For readers not familiar to Fourier series, a brief introduc tion is given in Section E. Let
now f be a uctuating velocity component, say v1 . The left side of Eq. 5.7 expresses
2
2
v1 in physical space (vs. x) and the right side v1 in wavenumber space (vs. n ).
The reader who is not familiar to the term wavenumber, is probably more familiar to
frequency. In that case, express f in Eq. 5.6 as a series in time rather than in space.
2
In this case the left side of Eq. 5.7 expresses v1 in the time domain and the right side
2
expresses v1 in the frequency domain.
The turbulent scales are distributed over a range of scales which extends from the
largest scales which interact with the mean ow to the smallest scales where dissipation
occurs, see Fig. 5.1. Now let us think about how the kinetic energy of the eddies varies
with eddy size. Intuitively we assume that large eddies have large uctuating velocities
which implies large kinetic energy, vi vi /2. It is now convenient to study the kinetic
energy of each eddy in wavenumber space. In wavenumber space the energy of eddies
from to + d can be expressed as
E()d
(5.8)
where Eq. 5.8 expresses the contribution from the scales with wavenumber between
and + d to the turbulent kinetic energy k. The energy, E(), corresponds to f 2 ()
in Eq. 5.7. The dimension of wavenumber is one over length; thus we can think of
wavenumber as proportional to the inverse of an eddys diameter, i.e 1/d. The
total turbulent kinetic energy is obtained by integrating over the whole wavenumber
space i.e.
k=
E()d = L
f 2 (n )
(5.9)
The kinetic energy is the sum of the kinetic energy of the three uctuating velocity
components, i.e.
k=
1
1 2
2
2
v + v2 + v3 = vi vi
2 1
2
(5.10)
42
The spectrum of E is shown in Fig. 5.2. We nd region I, II and III which correspond to:
I. In this region we have the large eddies which carry most of the energy. These
eddies interact with the mean ow and extract energy from the mean ow. This
energy transfer takes places via the production term, P k , in the transport equation
for turbulent kinetic energy, see Eq. 8.13. The energy of the largest eddies is
transferred to slightly smaller scales. The eddies velocity and length scales are
v0 and 0 , respectively.
III. Dissipation range. The eddies are small and isotropic and it is here that the
dissipation occurs. The energy transfer from turbulent kinetic energy to thermal
energy (increased temperature) is governed by in the transport equation for
turbulent kinetic energy, see Eq. 8.13. The scales of the eddies are described by
the Kolmogorov scales (see Eq. 5.5)
II. Inertial subrange. The existence of this region requires that the Reynolds number
is high (fully turbulent ow). The eddies in this region represent the mid-region.
This region is a transport region (in wavenumber space i.e.) in the cascade
process. The transport in wavenumber space is called spectral transfer. Energy
per time unit, P k = , is coming from the large eddies at the lower part of this
range and is given off to the dissipation range at the higher part (note that the
relation P k = is given by the concept of the cascade process). Since the
cascade concept assumes that all turbulent kinetic energy is transferred from
large to small eddies, it means than the number of small eddies must be much
larger than that of large eddies, i.e. N k is constant (N and k = v,i v,i /2
denote the number of eddies and kinetic energy eddies of size proportional to
1/, respectively). The eddies in this region are independent of both the large,
energy-containing eddies and the eddies in the dissipation range. One can argue
that the eddies in this region should be characterized by the spectral transfer of
energy () and the size of the eddies 1/. Dimensional analysis gives
E
=
[m /s2 ] =
3
a
b
2 3
[1/m] [m /s ]
(5.11)
E() = const. 3 3
This is a very important law (Kolmogorov spectrum law or the 5/3 law) which
states that, if the ow is fully turbulent (high Reynolds number), the energy
spectra should exhibit a 5/3-decay in the inertial region.
Above we state that the small eddies are isotropic. This means that in average
the eddies have no preferred direction, i.e. the uctuations in all directions are the
spectral
transfer
43
2
2
2
same so that v1 = v2 = v3 . Note that is not true instantaneously, i.e. in general
The kinetic energy is transferred to smaller and smaller eddies until it is dissipated at
the dissipative scales. In the inertial subrange the cascade process assumes that = .
Applying Eq. 5.12 for the large energy-containing eddies gives
0 = O
2
v0
0 v0
=O
3
v0
0
= =
(5.13)
o
=
1/4
0 =
1/2
0 =
1/4
3 0
3
v0
0
3
v0
1/4
v0 = (v0 0 /)
1/4
0 =
1/2
0 =
3
3
v0 3
0
3
v0
0
1/2
= Re1/4
1/4
= Re3/4
0
=
v0
v0 0
1/2
= Re1/2
(5.14)
where Re = v0 0 /. We nd that the ratio of the velocity, length and time scales
of the energy-containing eddies to the Kolmogorov eddies increases with increasing
Reynolds number. This means that the eddy range (wavenumber range) of the intermediate region, (region II, the inertial region), increases with increasing Reynolds number.
The larger the Reynolds number, the larger the wavenumber range of the intermediate
range where the eddies are independent of both the large scales and the viscosity.
isotropic
turbulence
44
Figure 5.3: Family tree of turbulent eddies (see also Table 5.1). Five generations. The
large original eddy, with axis aligned in the x1 direction, is 1st generation. Adapted
from [7]
generation
1st
2nd
3rd
4th
5th
6th
7th
x1
1
0
2
2
6
10
22
x2
0
1
1
3
5
11
21
x3
0
1
1
3
5
11
21
Table 5.1: Number of eddies at each generation with their axis aligned in the x1 , x2 or
x3 direction, see Fig. 5.3.
45
The equation for the instantaneous vorticity (qi = qi + qi ) reads (see Eq. 4.20)
vj
qi
vi
2 qi
= qj
+
xj
xj
xj xj
vk
qi = ijk
xj
(5.15)
As we learnt in Section 4.2 this equation is not an ordinary convection-diffusion equation; it has an additional term on the right side which represents amplication and
rotation/tilting of the vorticity lines (the rst term on right side). The i = j components of this term represent (see Eq. 4.21) vortex stretching. A positive v1 /x1 will
stretch the cylinder, see Fig. 4.2 and from the continuity equation we nd that the radial
derivative of the radial velocity v2 must be negative (see Eq. 4.22), i.e. the radius of
the cylinder will decrease. We have neglected the viscosity since viscous diffusion at
high Reynolds number is much smaller than the turbulent one and since viscous dissipation occurs at small scales (see p. 39). Thus we can assume that there are no viscous
stresses acting on the cylindrical uid element surface which means that the rotation
momentum
r2 q1 = const.
Vortex
stretching
(5.16)
remains constant as the radius of the uid element decreases. Note that also the circulation, which is the integral of the tangential velocity round the perimeter is
constant. Equation 5.16 shows that the vorticity increases if the radius decreases (and
vice versa). As was mentioned above, the continuity equation shows that stretching
results in a decrease of the radius of a slender uid element and an increase of the
vorticity component (i.e. the tangential velocity component) aligned with the element.
Hence an extension of a uid element in one direction (x1 direction) decreases the
length scales and increases the velocity scales in the other two coordinate directions
(x2 and x3 ). The increased v2 and v3 velocity components will in next stage stretch
smaller uid elements aligned in these two directions and so on. At each stage, the
length scale of the eddies whose velocity scale are increased decreases. Figure 5.3
illustrates how a large eddy whose axis is oriented in the x1 axis in a few generations
creates through vortex stretching smaller and smaller eddies with larger and larger
velocity gradients. Here a generation is related to a wavenumber in the energy spectrum (Fig. 5.2); young generations correspond to high wavenumbers. The smaller the
eddies, the less the original orientation of the large eddy is recalled. In other words, the
small eddies dont remember the characteristics of their original ancestor. The small
eddies have no preferred direction. They are isotropic. The creation of multiple eddies
by vortex stretching from one original eddies is illustrated in Fig. 5.3 and Table 5.1 The
large original eddy (1st generation) is aligned in the x1 direction. It creates two eddies
in the x2 and x3 direction (2nd generation), which in turn each create new eddies in
the x1 and x3 (3rd generation) and so on. For each generation the eddies become more
and more isotropic as they get smaller.
The i = j components in the rst term on the right side in Eq. 4.21 represent vortex
tilting. Again, take a slender uid element, now with its axis aligned with the x2 axis,
Fig. 4.2. The velocity gradient v1 /x2 will tilt the uid element so that it rotates in
the clock-wise direction. As a result, the second term q2 v1 /x2 in line one in Eq. 4.21
Vortex
tilting
6. Turbulent mean ow
46
q1
q2
v3
v2
0
x2
x3
v1
v3
0.
x3
x1
=
=
6 Turbulent mean ow
6.1 Time averaged Navier-Stokes
When the ow is turbulent it is preferable to decompose the instantaneous variables (for
example the velocity components and the pressure) into a mean value and a uctuating
value, i.e.
vi = vi + vi
p = p + p
(6.1)
where the bar, , denotes the time averaged value. One reason why we decompose the
+
t
xj
= 0
=
(6.2)
p
2 vi
+
xi
xj xj
(6.3)
The gravitation term, gi , has been omitted which means that the p is the hydrostatic pressure (i.e. when vi 0, then p 0, see p. 27). Inserting Eq. 6.1 into
47
the continuity equation (6.2) and the Navier-Stokes equation (6.3) we obtain the time
averaged continuity equation and Navier-Stokes equation
i
v
xi
i vj
v
xj
(6.4)
p
+
xi
xj
i
v
vi vj
xj
(6.5)
velocity, vi , i.e.
vi
=0
(6.6)
xi
Reynolds
equations
closure
problem
1
v
1
v
,
x1
x2
(6.7)
Eq. 6.5 reads (we write the left side on non-conservative form, see Section 2.4 at p. 20)
1
v
1
v
1
v
+ 2
v
x1
x2
1
v
p
+
v1 v2
x1
x2
x2
(6.8)
tot
x1 and x2 denote the streamwise and wall-normal coordinate, respectively, see Fig. 6.1.
Note that the two terms on the left side are of the same order, because they both include
the product of one large (1 or /x2 ) and one small (2 or /x1 ) part.
v
v
In addition to the viscous shear stress, 1 /x2 , an additional turbulent one a
v
Reynolds shear stress appears on the right side of Eq. 6.8. The total shear stress is
thus
tot =
1
v
v1 v2
x2
shear
stress
48
x2
v1 (x2 )
2
xA
2
x1
x3
Figure 6.1: Flow between two parallel plates. The width (i.e. length in the x3 direction)
of the plates, Zmax , is much larger that the separation between the plates, i.e. Zmax
.
quantity is the wall shear stress which is dened as
w =
1
v
x2
(6.9)
w
From the wall shear stress, we can dene a wall friction velocity, u , as
w = u2 u =
1/2
(6.10)
In order to take a closer look at the near-wall region, let us, again, consider fully
developed channel ow, see Fig. 6.1. In fully developed channel ow, the streamwise
derivative of the streamwise velocity component is zero (this is the denition of fully
developed ow), i.e. 1 /x1 = 0. The continuity equation gives now v2 = 0, see
v
Eq. 3.18 at p. 26. The rst term on the left side of Eq. 6.8 is zero because we have
fully developed ow (1 /x1 = 0) and the last term is zero because v2 0. The
v
+
x1
x2
1
v
v1 v2
x2
(6.11)
v 2
2
x2
x2
(6.12)
Integration gives
0=
v 2
p
+ 2
x2
x2
2
2
dx2 = p(x1 , x2 ) pw (x1 ) + v2 (x1 , x2 ) v2 |x2 =0
(6.13)
2
Using the wall boundary condition v2 |x2 =0 = 0 gives
2
v2 + p pw (x1 ) = 0
(6.14)
pw (x1 ) = p(x1 , 0) is the integration constant determined from the pressure boundary
condition at the wall. Since it is an integration constant stemming from integration over
wall
friction
velocity
103
49
102
101
x2 /
outer region
overlap region
Wall
inner region
Centerline
log-law region
buffer region
viscous region
x+ = y +
2
1
5 10
30
100
1000
10000
Figure 6.2: The wall region (adapted from Ch.7 in [5]) for Re = 10 000. denotes
half width of the channel, see Fig. 6.1 and x+ = x2 u / denotes the normalized wall
2
distance.
x2 (see Eq. 6.13) it is not a function of x2 . Recall that we have assumed that the ow
2
is fully developed which means that v2 is independent of x1 . Derivation of Eq. 6.14
with respect to x1 gives now
p
dw (x1 )
p
=
(6.15)
x1
dx1
which shows that p/x1 is a function only of x1 (i.e. it is independent of x2 ).
Now we can nally write the ow equations for fully developed turbulent ow
between two parallel plates (see Eq. 6.11 and 6.15)
0=
+
x1
x2
1
v
v1 v2
x2
(6.16)
The wall region can be divided into one outer and one inner region, see Fig. 6.2.
The inner region includes the viscous region (dominated by the viscous diffusion) and
the logarithmic region (dominated by turbulent diffusion); the logarithmic region is
sometimes called the inertial region, because the turbulent stresses stem from the inertial (i.e. the non-linear convection) term. The buffer region acts as a transition region between these two regions where viscous diffusion of streamwise momentum is
gradually replaced by turbulent diffusion. In the inner region, the total shear stress is
approximately constant and equal to the wall shear stress w , see Fig. 6.3. Note that the
total shear stress is constant only close to the wall (Fig. 6.3b); further away from the
wall it decreases (in fully developed channel ow it decreases linearly by the distance
form the wall, see Fig. 6.3a). The Reynolds shear stress vanishes at the wall because
we go away from the wall the viscous stress decreases and turbulent one increases and
at x+ 11 they are approximately equal. In the logarithmic layer the viscous stress is
2
negligible compared to the Reynolds stress.
At the wall, the velocity gradient is directly related to the wall shear stress, i.e. (see
Eq. 6.9 and 6.10)
1
1
v
= u2 = u2
(6.17)
x2 w
50
2000
1500
x+
2
200
150
x+
2
1000
100
500
a)
0
0
50
0.2
0.4
0.6
0.8
b)
0.2
0.4
0.6
0.8
Figure 6.3: Reynolds shear stress. Re = 2000. a) lower half of the channel; b) zoom
: v1 v2 /w ;
: (1 /x2 )/w .
v
near the wall. DNS data [8, 9].
2000
24
20
v1 /u
1500
x+ 1000
2
16
12
8
500
4
a)
0
0
10
15
20
25
v2 /u
b)
10
100
1000
x+
2
: v1 /u = x+ .
2
2
: DNS
v1 =
1 2
u x2 + C1
Since the velocity, v1 , is zero at the wall, the integration constant C1 = 0 so that
v1
u x2
=
u
(6.18)
Equation 6.18 is expressed in inner scaling (or wall scaling) which means that v1 and
x2 are normalized with quantities related to the wall, i.e. the friction velocity stemming
from the wall shear stress and the viscosity (here we regard viscosity as a quantity
related to the wall, since the ow is dominated by viscosity). Often the plus-sign ( + )
is used to denote inner scaling and equation Eq. 6.18 can then be written
v1 = x+
+
2
(6.19)
51
used in the entire region. What about the length scale? Near the wall, an eddy cannot
be larger than the distance to the wall and it is the distance to the wall that sets an
upper limit on the eddy-size. Hence it seems reasonable to take the wall distance as the
characteristic length scale; a constant is added so that
= x2
(6.20)
(6.21)
based on the velocity scale, u , and the length scale x2 . Another way of deriving
the expression in Eq. 6.21 is to use the Boussinesq assumption in which a turbulent
Reynolds stress is assumed to be equal to the product between the turbulent viscosity
and the velocity gradient as
1
v
v1 v2 = t
(6.22)
x2
The turbulent viscosity, t , represents the turbulence and has the same dimension as ,
i.e. [m2 /s]. Hence t can be expressed as a product of a turbulent velocity scale and a
turbulent length scale, and in the log-law region that gives
t = u x2
(6.23)
so that Eq. 6.22 gives (inserting v1 v2 = u2 ) read
u2 = u x2
1
v
1
v
u
=
x2
x2
x2
(6.24)
(6.25)
1
ln x+ + B or
2
1
v1
x2 u
+B
= ln
u
(6.26)
where B is an integration constant. From Eq. 6.26 we can dene the viscous length
scale, , as
x+ = x2 / =
(6.27)
2
u
a a
Equation 6.26 is the logarithmic law due to von K rm n [10]. The constant, , is called
the von K rm n constant. The constants in the log-law are usually set to = 0.41 and
a a
B = 5.2.
3000 (x2 /
0.3).
As can be seen in Fig. 6.2 the log-law applies for x+
2
Figures 6.4 and 6.8 where the Reynolds number is lower than in Fig. 6.2 show that
the log-law t the DNS up to x+ 500 (x2 / 0.25).
2
In the outer region of the boundary layer, the relevant length scale is the boundary
layer thickness. The resulting velocity law is the defect law
v1,c v1
x2
= FD
u
(6.28)
log-law
52
2000
1800
1600
150
1400
x+
2
x+
2
100
1200
1000
800
50
600
400
0
150
a)
100
50
50
100
150
b)
200
2
Figure 6.5: Fully developed channel ow. Re = 2000. Forces in the v1 equation,
see Eq. 6.16. a) near the lower wall of the channel; b) lower half of the channel
excluding the near-wall region. DNS data [8, 9].
: (v1 v2 /x2 )/w ;
:
: ( p/x1 )/w .
( 2 v1 /x2 )/w ;
32,tot
x2
x3
Figure 6.6: Symmetry plane of channel ow.
3
v
2
v
+
x2
x3
v2 v3 = 0
The viscous part is zero because v3 = 2 /x3 = 0. Hence, also the turbulent part
v
must be zero, i.e. v2 v3 = 0. With the same argument, v1 v3 = 0. However note that
2 = v 2 = 0. The reason is that although the time-averaged ow is two-dimensional
v3
3
(i.e. v3 = 0), the instantaneous turbulent ow is always three-dimensional and un
2
steady. Hence v3 = 0 and v3 = 0. Hence v3 = 0. Consider, for example, the time
but
2
2
v3 = v3 = (0.25)2 + 0.1252 + 0.1252 + (0.2)2 + 0.22 /5 = 0.03475 = 0.
Figure 6.3 presents the Reynolds and viscous shear stresses for fully developed
ow. As can be seen, the viscous shear stress is negligible except very near the wall. It
53
is equal to one near the wall and decreases rapidly for increasing wall distance. On the
other hand, the Reynolds shear stress is zero at the wall (because the uctuating velocities are zero at the wall) and increases for increasing wall distance. The intersection
of the two shear stresses takes place at x+ 11.
2
Looking at Eq. 6.16 we nd that it is not really the shear stress that is interesting,
but its gradient. The gradient of the shear stress, (v1 v2 )/x2 and 2 v1 /x2
2
represent, together with the pressure gradient, p/x1 , the forces acting on the uid.
Figure 6.5 presents the forces. Start by looking at Fig. 6.5b which shows the forces in
the region away from the wall. The pressure gradient is constant and equal to one: this
is the force driving the ow. This agrees fortunately with our intuition. We can
imagine that the uid (air, for example) is driven by a fan. Another way to describe
the behaviour of the pressure is to say that there is a pressure drop. The pressure must
decrease in the streamwise direction so that the pressure gradient term, p/x1 , in
Eq. 6.16 takes a positive value which pushes the ow in the x1 direction. The force
that balances the pressure gradient is the gradient of the Reynolds shear stress. This is
the force opposing the movement of the uid. This opposing force has its origin at the
walls due to the viscous wall force (viscous shear stress multiplied by area).
Now lets have a look at the forces in the wall region, see Fig. 6.5a. Here the forces
are two orders of magnitude larger than in Fig. 6.5b but they act over a very thin region
(x+ 40 or x2 / < 0.02). In this region the shear stress gradient term is driving
2
the ow and the opposing force is the viscous force. We can of course make a force
balance for a section of the channel, as we did for laminar ow, see Eq. 3.36 at p. 28
and Fig. 3.8 at p. 29 which reads
0 = p1 Zmax 2 p2 Zmax 2 2w LZmax
(6.29)
p
w
p
=
=
L
x1
(6.30)
As can be seen the pressure drop is directly related to the wall shear stress. In turbulent
ow the velocity prole in the center region is much atter than in laminar ow (cf.
Fig. 6.4 and Fig. 3.7 at p. 28). This makes the velocity gradient near the wall (and
the wall shear stress, w ) much larger in turbulent ow than in laminar ow: Eq. 6.30
shows why the pressure drop is larger in the former case compared to the latter; or
in other words why a larger fan is required to push the ow in turbulent ow than in
laminar ow.
2
2
2
Figure 6.7 presents the normal Reynolds stresses, v1 , v2 and v3 . As can
be seen, the streamwise stress is largest and the wall-normal stress is smallest. The
former is largest because the mean ow is in this direction; the latter is smallest because
the turbulent uctuations are dampened by the wall. The turbulent kinetic energy,
2
k = vi vi /2, is also included. Note that this is smaller than v1 .
54
2000
100
80
1500
60
x+
2
x+
2
1000
40
500
a)
0
0
20
b)
0
0
Figure 6.7: Normal Reynolds stresses and turbulent kinetic energy. Re = 2000. DNS
2
2
2
: v1 /w ;
: v2 /w ;
: v3 /w ; : k/u2 .
data [8, 9].
2000
24
20
v2 /u
1500
x+ 1000
2
16
12
8
500
4
a)
0
0
10
15
v2 /u
20
25
b)
10
100
1000
x+
2
: v2 /u = x+ .
2
2
Second, in a boundary layer ow the wall shear stress is not determined by the pressure
drop; the convective terms must also be taken into account. Third, the outer part of the
boundary layer is highly intermittent, consisting of turbulent/non-turbulent motion.
However, the inner region of a boundary layer (x2 / < 0.1) is principally the same
as for the fully developed channel ow, see Fig. 6.8: the linear and the log-law regions
are very similar for the two ows. However, in boundary layer ow the log-law is
valid only up to approximately x2 / 0.1 (compared to approximately x2 / 0.3 in
channel ow)
1/2
root-meansquare
RMS
(7.1)
The RMS is the same as the standard deviation which is equal to the square of the
variance. In order to extract more information, probability density function is a useful
statistical tool to analyze turbulence. From the velocity signals we can compute the
standard
deviation
variance
55
20
1.5
15
0.5
10
0.5
5
0
5
10
12
14
16
18
20
1.5
10
12
14
16
18
20
6
10
12
14
16
18
20
0.25
v 0.15
0.2
1.5
0.2
0.25
0.15
0.1
0.1
0.05
0.5
0.05
0
5
10
15
20
0
1.5
0.5
0.5
1.5
0
6
(a) Point 1. S = 1.51 and F = (b) Point 2. S = 0.27 and F = (c) Point 3. S = 0.09 and F =
7.68.
5.81.
2.77.
Figure 7.2: Probability density functions av time histories in Fig. 7.1. Vertical red
lines show vrms . The skewness, S, and the atness, F , are given for the three time
histories.
probability densities (sometimes called histograms). With a probability density, fv , of
the v velocity, the mean velocity is computed as
v=
vfv (v)dv
(7.2)
fv (v)dv = 1
(7.3)
Here we integrate over v. The mean velocity can of course also be computed by
integrating over time, as we do when we dene a time average, , (see Eq. 6.1 at p. 46),
i.e.
T
1
vdt
(7.4)
v=
2T T
where T is sufciently large.
Consider the probability density functions of the uctuations. The second moment
corresponds to the variance of the uctuations (or the square of the RMS, see Eq. 7.1),
i.e.
v 2 fv (v )dv
v 2 =
1
2T
T
T
v 2 (t)dt
56
skewness
v 3 fv (v )dv
v 3 =
3
and is commonly normalized by vrms , so that the skewness, Sv , of v is dened as
Sv =
1
3
vrms
v 3 fv (v)dv =
1
3
2vrms T
v 3 (t)dt =
T
1
4
vrms
v 4 fv (v)dv
The uctuations at Point 1 (see Fig. 7.1a) includes some samples which are very large
and hence its atness is large (see caption in Fig. 7.2a), whereas the uctuation for
Point 3 all mostly clustered within 2vrms giving a small atness, see Fig. 7.1c and
the caption in Fig. 7.2c.. For a Gaussian distribution
f (v ) =
1
vrms
exp
v rrms
2
2vrms
= 3
atness
57
(8.1)
Bj
Ai
Ai Bj
=
Bj
xk
xk
xk
(8.2)
Ai
Ai
Ai
+ Ai
xj
xj
= Ai
Ai
xj
(8.3)
Ai
1 Ai Ai
=
xj
2 xj
(8.4)
vi
[vi vj vi vj ] =
xj
vi vj
2
1
vi
[p p] + vi
[vi vi ] +
v
xi
xj xj
xj i
(8.5)
vi
(i + vi )(j + vj ) vi vj = vi
v
vi vj + vi vj + vi vj .
xj
xj
(8.6)
Using the continuity equation vj /xj = 0 (see Eq. 6.6), the rst term is rewritten as
vi
i .
vi vj = vi vj
xj
xj
(8.7)
v
We obtain the second term in Eq. 8.6 (using j /xj = 0) using Trick 2
1
(j k) = vj
v
vv
xj
xj 2 i i
1
= vj
2
vi
vi
v
+ vi i
xj
xj
vi
(j vi )
v
xj
(8.8)
58
The third term in Eq. 8.6 can be written as (replace vj by vj and use the same technique
as in Eq. 8.8)
1
v v v .
2 xj j i i
(8.9)
The rst term on the right side of Eq. 8.5 is re-written using the continuity equation
as
1 p
1 p vi
vi
=
xi
xi
(8.10)
The second term on the right side of Eq. 8.5 can be written
vi
vi
=
xj xj
xj
vi
v
xj i
vi vi
xj xj
(8.11)
applying Trick 1 (if we apply the chain-rule on the rst term on the right side of
Eq. 8.11 we get the left side and the second term on the right side). For the rst term
in Eq. 8.11 we use the same trick as in Eq. 8.8 so that
vi
v
xj i
xj
xj
1
2
xj
vi vi
xj
1
2
vi
v
vi + i vi
xj
xj
=
(8.12)
1 2 vi vi
2k
=
=
2 xj xj
xj xj
The last term on the right side of Eq. 8.5 is zero because it is time averaging of a
uctuation, i.e. ab = ab = 0. Now we can assemble the transport equation for the
turbulent kinetic energy. Equations 8.7, 8.8, 8.10, 8.11 and 8.12 give
j k
v
v
i
= vi vj
xj
xj xj
I
II
1 1
v vi
k
i
vj p + vj vi vi
2
xj
xj xj
III
(8.13)
IV
59
0
0 are determined by the large eddies and are constant for a given spectrum, we nd
3
3
from 2 / = 2 /0 that for decreasing size of eddies (decreasing ), the time scale,
0
also decreases, i.e.
2/3
0
(8.15)
=
0
2/3
1
Hence we nd that the velocity gradient of an eddy, proportional to , increases for decreasing eddy size, i.e. the dissipation is indeed largest at high wavenumbers.
In 2D boundary-layer ow, for which /x2 /x1 and v2 v1 , the exact k
equation reads
v
v
1 k 2 k
v
1
+
= v1 v2
x1
x2
x2
v vi
k
1
1
p v2 + v2 vi vi
i
x2
2
x2
xj xj
(8.16)
Note that the dissipation includes all derivatives. This is because the dissipation term
is at its largest for small, isotropic scales for which all derivatives are of the same order
and hence the usual boundary-layer approximation /x1 /x2 does not apply
for these scales.
Figure 8.1 presents the terms in Eq. 8.16 for fully developed channel ow. The left
side is since the ow is fully developed zero. In the outer region (Fig. 8.1b) all
terms are negligible except the production term and the dissipation term which balance
each other. Closer to the wall (Fig. 8.1a) the other terms do also play a role. Note that
the production and the dissipation terms close to the wall are two orders of magnitude
60
0.3
0.04
0.03
0.2
0.02
0.1
0.01
0
0.01
0.1
0.02
0.2
a)
0.03
10
20
30
40
x+
2
0.04
b)
100
200
300
400
500
x+
2
: P k;
:
Re = 2000. a) Zoom near the wall; b) Outer region. DNS data [8, 9].
; : v p /x2 ; +: v2 vi vi /2/x2 ; : 2 k/x2 .
2
40
35
30
25
x+ 20
2
15
10
5
0
0
0.2
0.4
0.6
0.8
1 /x+
v+
2
Figure 8.2: Channel ow at Re = 2000. DNS data [8, 9].Velocity gradient vs. x+ .
2
larger than in the logarithmic region (Fig. 8.1b). At the wall the turbulent uctuations
are zero which means that the production term is zero. Since the region near the wall
is dominated by viscosity the turbulent diffusion terms due to pressure and velocity
are also small. The dissipation term and the viscous diffusion term attain their largest
value at the wall and they much be equal to each other since all other terms are zero or
negligible.
The turbulence kinetic energy is produced by its main source term, the production
v
term, P k = v1 v2 1 /x2 . The velocity gradient is largest at the wall (see Fig. 8.2)
where the shear stress is zero (see Fig. 6.3); the former decreases and the magnitude of
the latter increases with wall distance and their product takes its maximum at x+ 11.
2
Since P k is largest here so is also k, see Fig. 6.7. k is transported in the x2 direction
by viscous and turbulent diffusion and it is destroyed (i.e. dissipated) by .
homogeneous
turbulence
61
2
zero, for example v1 /xi = 0, k/xi = 0 etc. (Note that the derivatives of the
instantaneous turbulent uctuations are non-zero even in homogeneous turbulence, i.e.
The instantaneous turbulent kinetic energy, k = v,i v,i /2, of this eddy may either be
transferred in wavenumber space or transported in physical (spatial) space, or both. It
may rst be transported in physical space towards the center, and there lose its kinetic
energy to smaller eddies. This should be kept in mind when thinking in terms of the
cascade process. Large eddies which extract their energy from the mean ow may not
give their energy to the slightly smaller eddies as assumed in Figs. 5.2 and 5.1, but k
may rst be transported in physical space and then transferred in spectral space.
In the inertial range (Region II), however, the cascade process is still a good approximation. The reason is that the transfer of turbulent kinetic energy, k , from eddy-toeddy occurs at a much faster rate than the spatial transport by convection and diffusion.
In other words, the time scale of the cascade process is much smaller than that of convection and diffusion which have no time to transport k in space before it is passed on
to a smaller eddy by the cascade process. We say that the turbulence at these scales are
in local equilibrium.
In summary, care should be taken in non-homogeneous turbulence, regarding the
validity of the cascade process for the large scales (Region I).
(8.17)
Similarly, the net effect of the dissipation term is a negative contribution, i.e.
dV < 0
(8.18)
local
equilibrium
62
0.25
0.2
0.15
0.1
0.05
0
0
10
15
20
x+
2
Figure 8.3: Channel ow at Re = 2000. Comparison of mean and uctuating dissipa: (1 /x2 )2 ;
v
tion terms. Both terms are normalized by u4 /. DNS data [8, 9].
: .
What about the transport terms, i.e. convection and diffusion? Integration of the convection term over V gives, using Gauss divergence law
V
j k
v
dV =
xj
vj knj dS
(8.19)
where S is the bounding surface of V . This shows that the net effect of the convection
term occurs only at the boundaries. Inside the domain, the convection merely transports
k with out adding or subtracting anything to k; the convection act as a source term in
part of the domain, but in the remaining part of the domain it acts as an equally large
sink term. Similarly for the diffusion term, we get
1
k
V
vj vk vk + p vj
xj
V xj
1
1
k
nj dS
=
v v v + p v
2 j k k j
xj
S
(8.20)
The only net contribution occurs at the boundaries. Hence, Eqs. 8.19 and 8.20 show
that the transport terms only as the word implies transports k without giving any
net effect except at the boundaries. Mathematically these terms are called divergence
terms, i.e. they can both be written as the divergence of a vector Aj ,
Aj
xj
where Aj for the convection and the diffusion term reads
vj k
convection term
1
k
1
Aj =
v v v + pv
diffusion term
2 j k k j
xj
(8.21)
(8.22)
The equation for K = vi vi /2 is derived in the same way as that for vi vi /2. Multiply
vi
vi vj
2 vi
1 p
i vj
v
= vi
+ i
v
vi
xj
xi
xj xj
xj
(8.23)
divergence
terms
63
= vj
vi vj
xj
xj
xj
xj
i vi
v
1 i vi
v
1 i vi
v
j K
v
= vj
vj
= vj
=
xj
2 xj
2 xj
xj
(8.24)
where Trick 1 was used twice: rst to obtain the term on the left side on the rst line
and then to get the second term on the left side of the second line. The continuity
equation was used to get the rst term on the right side of the rst line.
Using the continuity equation the rst term on the right side of Eq. 8.23 can be
written as
i p
v
p
=
.
(8.25)
i
v
xi
xi
The viscous term in Eq. 8.23 is rewritten in the same way as the viscous term in Section 8.1, see Eqs. 8.11 and 8.12, i.e.
i
v
i
v
K
i i
v v
=
.
xj xj
xj xj
xj xj
(8.26)
vi vj
j K
v
v
2K
1 i p
i i
v v
=
vi
.
xj
xj xj
xi
xj xj
xj
(8.27)
i vi vj
v
vi vj
v
i .
=
+ vi vj
xj
xj
xj
(8.28)
j K
v
v
2K
1 i p
i i
v v
=
xj
xj xj
xi
xj xj
i vi vj
v
v
i .
+ vi vj
xj
xj
(8.29)
On the left side we have the usual convective term. On the right side we nd: transport
by viscous diffusion, transport by pressure-velocity interaction, viscous dissipation,
transport by velocity-stress interaction and loss of energy to the uctuating velocity
eld, i.e. to k. Note that the last term in Eq. 8.29 is the same as the last term in
Eq. 8.13 but with opposite sign: here we clearly can see that the main source term in
the k equation (the production term) appears as a sink term in the K equation.
equation the dissipation term and the negative production term (representIn the K
ing loss of kinetic energy to the k eld) read
v
i i
v v
i ,
+ vi vj
xj xj
xj
(8.30)
and in the k equation the production and the dissipation terms read
vi vj
v vi
i
v
i
xj
xj xj
(8.31)
64
The gradient of the time-averaged velocity eld, vi , is much smoother than that of
the uctuating velocity eld, vi . Hence the dissipation by the uctuations, , is much
larger than the dissipation by the mean ow (left side of Eq. 8.30). This is clearly seen
in Fig. 8.3. The mean dissipation, (1 /x2 )2 , is largest at the wall where it takes the
v
value = 1/2000 (normalized by u4 /).
vj
[vi vk vi vk ] =
xk
2
v v
1
vi
p + vj
+ i k vj
vj
xi
xk xk
xk
(9.1)
Equation 6.5 is written with the index i as free index, i.e. i = 1, 2 or 3 so that the
equation is an equation for v1 , v2 or v3 . Now write Eq. 6.5 as an equation for vj and
vi
[vj vk vj vk ] =
xk
vj vk
2 vj
1
vi
p + vi
+
v
xj
xk xk
xk i
(9.2)
It may be noted that Eq. 9.2 is conveniently obtained from Eq. 9.1 by simply switching
65
vj
[v v v v ] =
[vi vk vi vk ] + vi
j k
j k
xk
xk
1 p
1 p
vi
vj
xj
xi
+vi
2
2 vj
vi
+ vj
xk xk
xk xk
(9.3)
vj vk vi vk
vi +
v
xk
xk j
Note that each line in the equation is symmetric: if you switch indices i and j in any
of the terms nothing changes. This is important since the tensor vi vj is symmetric.
Furthermore, you can check that the equation is correct according to the tensor notation
rules. Indices i and j appear once in each term (not more and not less) and index k (the
dummy index) appears exactly twice in each term (implying summation). Note that it
is correct to use any other index than k in some terms (but you must not use i and j).
You could, for example, replace k with m in the rst term and with q in the second
term; it is permissible, but usually we use the same dummy index in every term.
vj
[i vk + vi vk + vi vk ] + vi
v
vj vk + vj vk + vj vk
xk
xk
(9.4)
Using the continuity equation the rst terms in the two group are rewritten as
vj vk
i
v
v
j
+ vi vk
xk
xk
(9.5)
vj
vj vk
vj
vi vk
v
+ vi
= vk vj i + vk vi
xk
xk
xk
xk
vi vj vk
vi vj
=
= vk
xk
xk
(9.6)
The continuity equation was used twice (to get the right side on the rst line and to get
the nal expression). Re-writing also the third terms in the two groups in Eq. 9.4 in the
same way, the second and the third terms in Eq. 9.4 can be written
(k vi vj ) +
v
vv v
xk
xk i j k
(9.7)
1 1 1 vi
1 vj
vi p
vj p + p
+ p
xj
xi
xj
xi
(9.8)
It will later turn out that it is convenient to express all derivatives as /xk . Therefor
we re-write the derivative in the two rst terms as
= jk
and
= ik
xj
xk
xi
xk
(9.9)
so that
jk
66
1
1 vj
1 1 vi
+ p
vi p ik
vj p + p
xk
xk
xj
xi
(9.10)
vj
vi
xk
xk
v
vj i
xk
xk
vi vj
xk xk
Trick 1 is used again to merge the two rst terms so that the third line in Eq. 9.3
reads
=
xx
vi vj
xk
xk
vi
vj
v
+ vj i
xk
xk
vi vj
xk xk
2 vi vj
v vj
v vj
2 i
=
2 i
xk xk
xk xk
xk xk
(9.11)
We can now put everything together. Put the rst term in Eq. 9.7 on the left side and
the second term on the right side together with Eqs. 9.5, 9.10 and 9.11 so that
v
v
i v v j
(k vi vj ) = vj vk
v
i k
xk
xk
xk
I
xk
II
vi vj
1
1
vi vj vk + jk vi p + ik vj p
xk
(9.12)
III
1
+ p
vj
vi
+
xj
xi
vi vj
xk xk
IV
Note that the manipulation in Eq. 9.9 allows the diffusion (term III) to be written on a
more compact form. After a derivation, it is always useful to check that the equation is
correct according to the tensor notation rules.
Every term or group of terms should include the free indices i and j (only
once);
Every term or group of terms should be symmetric in i and j;
A dummy index (in this case index k) must appear exactly twice (=summation)
in every term
Equation 9.12 can also be written in a simplied easy-to-read symbolic form as
Cij = Pij + Dij + ij ij
(9.13)
vj
vi
+
xj
xi
(9.14)
67
0.1
0.05
0.05
a)
0.5
0
10
20
30
40
x+
2
0.1
b)
100
200
300
400
500
x+
2
2
Figure 9.1: Channel ow at Re = 2000. Terms in the v1 equation scaled by u4 /.
a) Zoom near the wall; b) Outer region. DNS data [8, 9].
: P11 ;
: 11 ;
:
v 2 )/x ; : 2 v 2 /x2 .
11 ; +: (v2 1
2
2
1
Equation 9.12 is the (exact) transport equation of the Reynolds stress, vi vj . It is called
the the Reynolds stress equations. Since it is an equation for a second-order tensor,
it consists of nine equations, but since it is symmetric we only need to consider six
of them. Compare Eq. 9.12 with the equation for turbulent kinetic energy, Eq. 8.13.
An alternative and maybe easier way to derive Eq. 8.13 is to rst derive Eq. 9.12
and then take the trace (setting i = j) and dividing by two. In both the k and the
vi vj equations there is a convection term (I), a production term (II), a diffusion term
(III) and a dissipation term (IV). In the vi vj equation there is a fth term (V), see
Eq. 9.14, which is called the pressure strain term. The physical meaning of this term is
to redistribute energy between the normal stress components (if we transform Eq. 9.12
to the principal coordinates of vi vj there are no shear stresses, only normal stresses).
2
The average of the normal stresses is vav = vi vi /3. For a normal stress that is larger
2
than vav , the pressure-strain term is negative and vice-versa. It is often called the Robin
Hood term because it as Robin Hood takes from the rich and gives to the poor.
Note that the trace of the pressure-strain term is zero, i.e.
ii =
v
vi
+ i
xi
xi
1
p
=0
(9.15)
because of the continuity equation and this is the reason why this term does not appear
in the k equation.
For 2D boundary layer ow, Eq. 9.12 reads
v
v
i v v j
(1 vi vj ) +
v
(2 vi vj ) = vj v2
v
i 2
x1
x2
x2
x2
v
vi j
1
vi vj v2 + j2 vi p + i2 vj p
x2
x2
1
+ p
vj
vi
+
xj
xi
(9.16)
vi vj
xk xk
Now lets look at this equation for fully developed channel ow for which
v2 = v3 = 0
()
()
=
=0
x1
x3
(9.17)
Reynolds
stress
equations
pressure
strain
Robin Hood
68
The second line shows that it is the streamwise and spanwise derivative that operate on
time-averaged quantities that are zero, not those that operate on instantaneous quantities such as in ij and ij .
The production term in Eq. 9.16 reads
Pij = vj v2
i
v
v
j
vi v2
x2
x2
(9.18)
2
2
2
For the v1 (i = j = 1), v2 (i = j = 2), v3 (i = j = 3) and v1 v2 (i = 1, j = 2)
equations we get
1
v
x2
v
2 = 0
= 2v2 v2
x2
v
3 = 0
= 2v3 v2
x2
1
v
v
v v v 2 = v 2 1
= v2 v2
1 2
x2
x2
x2
P11 = 2v1 v2
(9.19a)
P22
(9.19b)
P33
P12
(9.19c)
(9.19d)
v1 v2
12 =
=0
(9.20)
xk xk
However, very close to the wall, x+ 10, 12 = 0 because here the wall affects
2
the dissipative scales making them non-isotropic; 12 is positive since v1 v2 < 0, see
Fig. 9.3.
If you want to learn more how to derive transport equations of turbulent quantities,
see [12] which can be downloaded here
http://www.tfd.chalmers.se/lada/allpapers.html
69
0.02
0.015
0.01
0.005
0
0.005
0.01
0.015
a)
0.05
0
10
20
30
40
x+
2
0.02
b)
100
200
300
400
500
x+
2
a) Zoom near the wall; b) Outer region. DNS data [8, 9].
: P22 ;
: 22 ;
2
2
: 22 ; +: (v2 v2 )/x2 ; : 2 v2 /x2 .
: 2v p /x2 ;
2
0.15
0.05
0.1
0.05
0
0.05
0.1
a)
10
20
x+
2
30
40
0.05
b)
100
200
300
400
500
x+
2
Figure 9.3: Channel ow at Re = 2000. Terms in the v1 v2 equation scaled by u4 /.
a) Zoom near the wall; b) Outer region. DNS data [8, 9].
: P12 ;
: 12 ;
2
: 12 ; +: (v1 v2 )/x2 ; : 2 v1 v2 /x2 .
: v p /x2 ;
2
(the turbulent uctuating) velocity v2 . At its new location the v1 velocity is in average
smaller than at its old, i.e. v1 (x2,A ) < v1 (x2,B ). This means that when the particle
at x2,B (which has streamwise velocity v1 (x2,B )) comes down to x2,A (where the
streamwise velocity is v1 (x2,A )) it has an excess of streamwise velocity compared to
its new environment at x2.A . Thus the streamwise uctuation is positive, i.e. v1 > 0
v < 0).
and the correlation between v1 and v2 is in average negative (v1 2
If we look at the other particle (dashed line in Fig. 9.4) we reach the same con
clusion. The particle is moving upwards (v2 > 0), and it is bringing a decit in v1
so that v1 < 0. Thus, again, v1 v2 < 0. If we study this ow for a long time and
average over time we get v1 v2 < 0. If we change the sign of the velocity gradient so
that 1 /x2 < 0 we will nd that the sign of v1 v2 also changes.
v
10. Correlations
70
x2,B
v2 > 0
v2 < 0
v1 (x2 )
x2,A
x2
x1
Figure 9.4: Sign of the Reynolds shear stress v1 v2 in a boundary layer.
|xA xC |
1
1
In cases where the shear stress and the velocity gradient have the same sign (for
example, in a wall jet) the reason is that the transport terms are more important than
the production term.
10 Correlations
10.1 Two-point correlations
Two-point correlations are useful when describing some characteristics of the turbulence. Pick two points along the x1 axis, say xA and xC , and sample the uctuating
1
1
velocity, in, for example, the x1 direction. We can then form the correlation of v1 at
these two points as
(10.1)
B11 (xA , xC ) = v1 (xA )v1 (xC )
1
1
1
1
Often, it is expressed as
1
1
1
1
1
(10.2)
71
Lint
xA
1
x1
(a) Small integral length scale
Lint
xA
1
x1
(b) Large integral length scale
Figure 10.2: Schematic relation between the two-point correlation, the largest eddies
(thick lines) and the integral length scale, Lint .
72
It is obvious that if we move point A and C closer to each other, B11 increases;
when the two points are moved so close that they merge, then B11 = v 2 (xA ). If,
1
on the other hand, we move point C further and further away from point A, then B11
will go to zero. Furthermore, we expect that the two-point correlation function will be
related to the largest eddies. It is convenient to normalize B11 so that it varies between
1 and +1. The normalized two-point correlation reads
norm
B11 (xA , x1 ) =
1
1
v1,rms (xA )v1,rms (xA
1
1
+ x1 )
1
1
(10.3)
where subscript rms denotes root-mean-square, which for v1 , for example, is dened
as
2
v1,rms = v1
1/2
(10.4)
RMS is the same as standard deviation (Matlab command std) which is the squareroot of the variance (Matlab command var).
Consider a ow where the largest eddies have an eddy scale (length scale) of Lint ,
see Fig. 10.2. We expect that the two point correlation, B11 , approaches zero for separation distance, |xA xC | > Lint because for separation distances larger than |xA xB |
1
1
1
1
there is no correlation between v1 (xA ) and v1 (xC ). Hence, ows with large eddies will
1
1
have a two-point correlation function which decreases slowly with separation distance.
For ows with small eddies, the two-point correlation, B11 , decreases rapidly with x1 .
From the two-point correlation, B11 , an integral length scale, Lint which is dened
as the integral of B11 over the separation distance, i.e.
Lint (x1 ) =
0
B11 (x1 , x1 )
d1
x
A
C
v1,rms v1,rms
integral
length scale
(10.5)
uctuation with a separation in time. If we again choose the v1 uctuation, the auto
correlation reads
(10.6)
B11 (tA , t ) = v1 (tA )v1 (tA + t )
where t = tC tA , is the time separation distance between time A and C. If the mean
ow is steady, the time direction is homogeneous and then B is independent on tA ;
B11 (t ) = v1 (t)v1 (t + t )
(10.7)
v (t)v1 (t
2
v1,rms 1
+ t)
(10.8)
In analogy to the integral length scale, Lint , the integral time scale, Tint , is dened
integral
time scale
73
Tint =
0
norm
B11 (t)dt
(10.9)
74
75
a mean part vi and a uctuating part vi so that vi = vi + vi . Time average and we get
(j ) = 0
v
+
t
xj
p
2 vi
ij
i
v
(i vj ) =
v
+
0 ( 0 )gi
+
t
xj
xi
xj xj
xj
(11.1)
(11.2)
(note that denotes temperature) where the turbulent stress tensor (also called Reynolds
stress tensor) is written as:
ij = vi vj
The pressure, p, denotes the hydro-static pressure, see Eq. 3.22, which means that when
the ow is still (i.e. vi 0), then the pressure is zero (i.e. p 0).
The body force fi which was omitted for convenience in Eq. 6.4 has here been
re-introduced. The body force in Eq. 11.2 is due to buoyancy, i.e. density differences.
The basic form of the buoyancy force is fi = gi where gi denotes gravitational acceleration. Since the pressure, p, is dened as the hydro-static pressure we must rewrite
(11.3)
where h denotes change in height). If we let density depend on pressure and temperature, differentiation gives
d =
d +
p
dp
(11.4)
Our ow is incompressible, which means that the density does not depend on pressure,
i.e. /p = 0; it may, however, depend on temperature and mixture composition.
Hence the last term in Eq. 11.4 is zero and we introduce the volumetric thermal expansion, , so that
(11.5)
d = d, =
p
Now we can re-write the buoyancy source as
fi = 0 ( 0 )gi = 0 ( 0 )gi
(11.6)
which is the last term in Eq. 11.2. Consider the case where x3 is vertically upwards.
Then gi = (0, 0, g) and a large temperature in Eq. 11.6 results in a force vertically
upwards, which agrees well with our intuition.
Reynolds
stress
tensor
11.2. The exact vi vj equation
76
component as = + . The transport equation for reads (see Eq. 2.14 where
temperature was denoted by T )
2
vi
=
+
t
xi
xi xi
(11.7)
i
v
2
v
=
i
xi
xi xi
xi
(11.8)
Introducing = + we get
The last term on the right side is an additional term whose physical meaning is turbulent
heat ux vector. This is similar to the Reynolds stress tensor on the right side of the
time-averaged momentum equation, Eq. 11.2. The total heat ux vector viscous plus
turbulent in Eq. 11.8 reads (cf. Eq. 2.10)
ci,tot = ci + ci,turb =
vi
xi
(11.9)
11.2 The exact vi vj equation
Now we want to solve the time-averaged continuity equation (Eq. 11.1) and the three
momentum equations (Eq. 11.2). Unfortunately there are 10 unknowns; the four usual
ones (i , p) plus six turbulent stresses, vi vj . We must close this equation system; it is
v
called the closure problem. We must nd some new equations for the turbulent stresses.
We need a turbulence model.
The most comprehensive turbulence model is to derive exact transport equations
for the turbulent stresses. An exact equation for the Reynolds stresses can be derived
from the Navies-Stokes equation. It is emphasized that this equation is exact; or, rather,
as exact as the Navier-Stokes equations. The derivation follows the steps below.
Eq. (A)
Subtract Eq. (B) from Eq. (A) equation for vi , Eq. (C)
Multiply Eq. (C) with vj and Eq. (D) with vi , time average and add them together
v
equation for vi j
In Section 9 at p. 64 these steps are given in some detail. More details can also be
a
found in [12] (set the SGS tensor to zero, i.e. ij = 0).
closure
problem
11.3. The exact vi equation
77
The nal vi vj -equation (Reynolds Stress equation) reads (see Eq. 9.12)
vi vj
vi vj
v
v
j v v i + p
= vi vk
+ vk
j k
t
xk
xk
xk
Cij
vj
vi
+
xj
xi
Pij
ij
2 vi vj
p vj
p vi
vi vj vk +
ik +
jk +
xk
xk xk
Dij,
Dij,t
gi vj gj vi 2
Gij
(11.10)
vi vj
xk xk
ij
where Dij,t and Dij, denote turbulent and viscous diffusion, respectively. The total
diffusion reads Dij = Dij,t + Dij, . This is analogous to the momentum equation
where we have gradients of viscous and turbulent stresses which correspond to viscous
and turbulent diffusion. Equation 11.10 can symbolically be written
Cij = Pij + ij + Dij + Gij ij
where
Cij Convection
Pij Production
ij Pressure-strain
Dij Diffusion
Gij Buoyancy production
ij Dissipation
If temperature variations occurs we must solve for the mean temperature eld, see
Eq. 11.8. To obtain the equation for the uctuating temperature, subtract Eq. 11.8 from
Eq. 11.7
v
(vk + vk + vk ) =
+ k
+
t
xk
xk xk
xk
(11.11)
To get the equation for the uctuating velocity, vi , subtract the equation for the mean
velocity vi (Eq. 11.2) from the equation for the instantaneous velocity, vi (Eq. 6.3) so
that
v v
vi
1 p
2 vi
(vk vi + vk vi + vk vi ) =
+
+ i k gi (11.12)
+
t
xk
xi
xk xk
xk
78
Multiply Eq. 11.11 with vi and multiply Eq. 11.12 with , add them together and
time average
vi
(v + v + v ) + ( v + v v + v v )
k
vi k k i
+ vi
i k
k
t
xk k
xk
2
2 vi
p
=
+ vi
+
gi
xi
xk xk
xk xk
(11.13)
The Reynolds stress term in Eq. 11.12 multiplied by and time averaged is zero, i.e.
vi vj
vi vj
= 0
=
xk
xk
The rst term in the two parantheses on line 1 in Eq. 11.13 are combined into two
production terms
v
+ v
vi vk
(11.14)
k
xk
xk
The second term in the two parenthesis on the rst line of Eq. 11.13 can be re-written
as a convection term
vi vk
(11.15)
xk
Finally, the last two terms are re-cast into turbulent diffusion terms
vi vk
xk
(11.16)
Inserting Eqs. 11.14, 11.15 and 11.16 into Eq. 11.13 gives the transport equation for
v
vi
v i p v v
vk vi = vi vk
+
k
t
xk
xk
xk xi xk k i
Pi
+
xk
vi
xk
+
xk
Di,
v
i
xk
Di,t
( + )
vi
xk xk
gi
(11.17)
Gi
where Pi , i and Di,t denote the production, scramble and turbulent diffusion term,
respectively. The production term include one term with the mean velocity gradient
and one term with the mean temperature gradient. On the last line, Di, , i and Gi
denote viscous diffusion, dissipation and buoyancy term, respectively. The unknown
terms i , Di , i , Gi have to be modelled as usual; this is out of the scope of
the present course but the interested reader is referred to [13].
1 2
1
2
2
v1 + v2 + v3 vi vi
2
2
79
By taking the trace (setting indices i = j) of the equation for vi vj and dividing by two
we get the equation for the turbulent kinetic energy:
k
k
v
i vi vi
+ vj
= vi vj
t
xj
xj
xj xj
Ck
Pk
(11.18)
xj
vj
1
+ vi vi
2
gi vi
xj xj
k
D
k
Dt
Gk
k
k
where as in the vi vj equation Dt and D denotes turbulent and viscous diffusion,
k
k
k
respectively. The total diffusion reads D = Dt + D . Equation 11.18 can symbolically be written:
C k = P k + Dk + Gk
(11.19)
where C = 0.09. An exact equation for the transport equation for the dissipation
=
vi vi
xj xj
can be derived (see, e.g., [14]), but it is very complicated and in the end many terms
are found negligible. It is much easier to look at the k equation, Eq. 11.19, and to setup
a similar equation for . The transport equation should include a convective term, C ,
a diffusion term, D , a production term, P , a production term due to buoyancy, G ,
and a destruction term, , i.e.
C = P + D + G
(11.20)
The production and destruction terms term in the k equation are used to formulate the
corresponding terms in the equation. The terms in the k equation have the dimension
k/t = [m2 /s3 ] whereas the terms in the equation have the dimension /t =
[m2 /s4 ]. Hence, we must multiply P k and by a quantity which dimension [1/s]. One
quantity with this dimension is the mean velocity gradient which might be relevant for
the production term, but not for the destruction. A better choice should be /k = [1/s].
Hence, we get
(11.21)
P + G = (c1 P k + c3 Gk c2 )
k
The turbulent diffusion term is expressed in the same way as that in the k equation but
with its own turbulent Prandtl number, (see Eq. 11.27), i.e.
D =
xj
xj
(11.22)
80
= (c1 P + c3 Gk c2 ) +
+ vj
t
xj
k
xj
xj
(11.23)
j
v
i
v
+
xj
xi
2
2
+ ij k = 2t sij + ij k
3
3
(11.24)
The last term is added to make the equation valid also when it is contracted (i.e taking
the trace); after contraction both left and right side are equal (as they must be) and
equal to vi vi = 2k. When Eq. 11.24 is included in Eq. 11.2 we replace 6 turbulent
stresses with one new unknown (the turbulent viscosity, t ). This is of course a drastic
simplication.
If the mean temperature equation is solved for we need an equation for the heat
ux vector, vi . One option is to solve its transport equation, Eq. 11.17. If an eddyviscosity model (i.e. Eq. 11.24) is used for the Reynolds stresses, an eddy-viscosity
model is commonly used also for the heat ux vector. The Boussinesq assumption
reads
(11.25)
vi = t
xi
where t denotes the turbulent thermal diffusivity. It is usually obtained from the
turbulent viscosity as
t
t =
(11.26)
t
t
(11.27)
It is important to recognize that the viscosity (), the Prandtl number (P r), the
thermal diffusivity () are physical parameters which depend on the uid (e.g. water
or air) and its conditions (e.g. temperature). However, the turbulent viscosity (t ), the
turbulent thermal diffusivity (t ) and the turbulent Prandtl number ( ) depend on the
ow (e.g. mean ow gradients and turbulence).
vi , k and equations and formulate modeling assumptions for the remaining terms in
the Reynolds stress equation. This will give us the Reynolds Stress Model [RSM] (also
called the Reynolds Stress underline Transport Model [RSTM]) where a (modelled)
transport equation is solved for each stress. Later on, we will introduce a simplied
81
algebraic model, which is called the Algebraic Stress Model [ASM] (this model is also
called Algebraic Reynolds Stress Model, ARSM)
Summary of physical meaning:
v
v i
= vi vk
k
xk
xk
Pij = vi vk
Pi
Pk =
1
v
i
Pii = vi vj
2
xj
(11.28)
The k is usually not solved for in RSM but a length-scale equation (i.e. or ) is
always part of an RSM and that equation includes P k .
In the k model, the Reynolds stresses in the production term are computed
using the Boussinesq assumption, which gives
vi vj = t
P k = t
sij =
1
2
j
v
i
v
+
xj
xi
i
v
j
v
+
xj
xi
i
v
j
v
+
xj
xi
2
ij k
3
i
v
= 2t sij sij
xj
(11.29)
xj
D =
xj
Dk =
t
k
t
+
k
xj
xj
(11.30)
The gradient hypothesis simply assumes that turbulent diffusion acts as to even out
k
all inhomogeneities. In other words, it assumes that the turbulent diffusion term, Dt ,
82
transports k from regions where k is large to regions where k is small. The turbulent
diffusion ux of k is expressed as
dk =
j,t
1
t k
vj vi vi =
2
k xj
(11.31)
Only the triple correlations are included since the pressure diffusion usually is negligible (see Fig. 8.1 at p. 60). Taking the divergence of Eq. 11.31 (including the minus
sign in Eq. 11.18) gives the turbulent diffusion term in Eq. 11.30.
Solving the equations for the Reynolds stresses, vi vj , opens possibilities for a more
advanced model of the turbulent diffusion terms. Equation 11.31 assumes that if the
gradient is zero in xi direction, then there is no diffusion ux in that direction. A more
general gradient hypothesis can be formulated without this limitation, e.g.
dk
j,t,G vj vk
k
xk
(11.32)
which is called the general gradient diffusion hypothesis (GGDH). It was derived in
[15] from the transport equation of the triple correlation vj vi vi . In GGDH the turbulent
ux dk , for example, is computed as
1,t,G
dk
1,t,G v1 v1
k
k + v v k
+ v1 v2
1 3
x1
x2
x3
(11.33)
xk
k
The diffusion term, Dt , in the k equation is obtained by taking the divergence of this
equation
dk
k k
j,t,G
k
ck vj vk
(11.35)
Dt =
=
xj
xj
xk
This diffusion model may be used when the k equation is solved in an RSM or an ASM.
The corresponding diffusion terms for the and vi vj equations read
Dt =
Dij,t
xj
=
xk
c vj vk
k
xk
ck vk vm
k vi vj
xm
(11.36)
Equation 11.36 often causes numerical problems. A more stable alternative is to model
the diffusion terms as in 11.30 which for vi vj reads
Dij,t =
xm
t vi vj
k xm
(11.37)
83
2. All shear stresses are zero because the uctuations in two different coordinate
directions are not correlated.
What applies for the small-scale uctuations (Items 1 and 2, above) must also apply
to the gradients of the uctuations, i.e.
v1 v1
v2 v2
v3 v3
=
=
xk xk
xk xk
xk xk
v
vi
j
= 0 if i = j
xk xk
(11.38)
2
ij
3
(11.39)
1
where the factor 2/3 is included so that = 2 ii is satised, see Eqs. 11.10 and 11.18.
so that v1 /x1 < 0, see Fig. 11.1. As a result the uctuating pressure p increases at
O so that
v
p 1 < 0
x1
The uid in the x1 direction is performing work, moving uid particles against the
pressure gradient. The kinetic energy lost in the x1 direction is transferred to the x2
and x3 directions and we assume that the collision makes uid particles move in the
other two directions, i.e.
v3
v2
> 0,
>0
(11.40)
x2
x3
Indeed, if v1 /x1 < 0, the continuity equation gives v2 /x2 + v3 /x3 > 0.
However, in Eq. 11.40 we assume that not only their sum is positive but also that they
84
v1
v1
x2
x1
Figure 11.1: Physical illustration of the pressure-strain term.
both are positive. If this is to happen the kinetic energy in the x1 direction must be
2
2
2
2
larger than that in the x2 and x3 direction, i.e. v1 > v2 and v1 > v3 . If this were not
true, the uctuation v1 would not be able to create an acceleration of both v2 and v3 .
The amount of kinetic energy transferred from the x1 direction to the x2 and x3
directions, should be proportional to the difference of their energy, i.e.
1
1 v1
p
x1
2
2
2
2
2
v1 v2 + v1 v3
3 2 1 2
2
2
= v1
v + v2 + v3
2
2 1
2
= v1
1 2
2
v + v3
2 2
3 2
v k
2 1
(11.41)
The expression in Eq. 11.41 applies only to the normal stresses, i.e. the principal axis
of vi vj . By transforming to a coordinate system which is rotated /4 it is shown
that the sign of p (vi /xj + vj /xi ) and vi vj are opposite. Assume that we express Eq. 11.41 in principal coordinates, (x1 , x2 ), and then transform the equation to
(x1 , x2 ) by rotating it angle = /4, see Appendix N. Replacing u12 in Eq. N.6b by
v1 v2 we get
2
2
v1 v2 = 0.5 v1 v2
(11.42)
since v1 v2 = v2 v1 . Now we have transformed the right side of Eq. 11.41. Next
step is to transform the left side, i.e. the velocity gradients. We use Eqs. N.6b and N.6c:
replacing u12 and u21 by v1 /x2 and v2 /x1 , respectively, and adding them gives
v
v1
v
v2
+ 1 =
2
x1
x2
x1
x2
(11.43)
v2
v
+ 1
x1
x2
= p
v1
v
2
x1
x2
(11.44)
Now we apply Eq. 11.41 using the right side of Eq. 11.44
p
v1
v
2
x1
x2
2
2
v1 v2
(11.45)
85
v1
x2
x2
a)
b)
x1
x3
Figure 11.2: Decaying grid turbulence. The circles (a) and the thin rectangles (b)
illustrates past of the grid which consists of a mesh of circular cylinders.
Inserting Eqs. 11.42 and 11.44 into Eq. 11.45 gives nally
p
v2
v
+ 1
x1
x2
= v1 v2
(11.46)
This shows that the pressure-strain term acts as a sink term in the shear stress equation.
Thus, Eqs. 11.41 and 11.46 lead as to write
ij,1 p
vj
vi
+
xj
xi
= c1
2
vi vj ij k
3
(11.47)
where denotes the modelled pressure-strain term and subscript 1 means the slow part;
the concept slow and rapid is discussed at p. 86. We have introduced the turbulent
time scale k/. This pressure-strain model for the slow part was proposed by Rotta in
1951 [16].
Let us investigate how Eq. 11.47 behaves for decaying grid turbulence, see Fig. 11.2.
Flow from left with velocity v1 passes through a grid. The grid creates velocity gradi
ents behind the grid which generate turbulence. Further downstream the velocity gradients are smoothed out and the mean ow becomes constant. From this point and further
downstream the ow represents homogeneous turbulence which is slowly approaching
isotropic turbulence; furthermore the turbulence is slowly dying (i.e. decaying) due
to dissipation. The exact vi vj equation for this ow reads (no production or diffusion
because of homogeneity)
v1
vi vj
p
=
x1
vj
vi
+
xj
xi
ij
(11.48)
Rottas pressure-strain model is supposed to reduce anisotropy. Thus it should be interesting to re-write Eq. 11.48 expressed in the normalized anisotropy Reynolds stress
86
vi vj
1
ij
(11.49)
2k
3
We introduce bij (Eq. 11.49), Rottas model (Eq. 11.47) and the model for the dissipation tensor (11.39) into Eq. 11.48 so that
bij =
21
v
(kbij )
1 k
+ ij
x1
3 x1
2
= 2c1 bij ij
3
(11.50)
k
=
x1
(11.51)
bij
(11.52)
Provided that c1 > 1 Rottas model does indeed reduce non-isotropy as it should.
The model of the slow pressure-strain term in Eq. 11.47 can be extended by in
cluding terms which are non-linear in vi vj . To make it general it is enough to include
terms which are quadratic in vi vj , since according to the Cayley-Hamilton theorem, a
second-order tensor satises its own characteristic equation (see Section 1.20 in [2]);
this means that terms cubic in vi vj can be expressed in terms linear and quadratic in
v . The most general form of
vi j
ij,1 can be formulated as [17]
1
ij,1 = c1 aij + c aik akj ij ak ak
1
3
vi vj
2
aij =
ij
k
3
(11.53)
aij is the anisotropy tensor whose trace is zero. In isotropic ow all its components
are zero. Note that the right side is trace-less (i.e. the trace is zero). This should be so
87
y
V
x2 , y2
x1 , y1
Figure 11.3: The exact solution to Eq. 11.56. The integral is carried out for all points,
y, in volume V .
Now we want to derive an exact equation for the pressure-strain term, ij . Since
it includes the uctuating pressure, p , we start by deriving an exact equation for p
starting from Navier-Stokes equations.
1. Take the divergence of incompressible Navier-Stokes equation assuming con2
stant viscosity (see Eq. 6.2) i.e. x xi (vi vj ) = . . . Equation A.
j
2. Take the divergence of incompressible time-averaged Navier-Stokes equation as2
v
suming constant viscosity (see Eq. 6.4) i.e. x xi (i vj ) = . . . Equation B.
j
Subtracting of Equation B from Equation A gives a Poisson equation for the uctuating pressure p
1 2 p
2
i vj
v
v v vi vj
= 2
xj xj
xj xi xi xj i j
rapid term
(11.55)
slow term
(11.56)
there exists an exact analytical solution given by Greens formula (it is derived from
Gauss divergence law)
(x) =
1
4
(11.57)
where the integrals at the boundaries vanish because it is assumed that f 0 at the
88
boundaries, see Fig. 11.3. Applying Eq. 11.57 on Eq. 11.55 gives
p (x) =
4
(y) v (y)
2
vi
j
+
v (y)vj (y) vi (y)vj (y)
2
yi
yi yj i
yj
rapid term
slow term
dy3
|y x|
(11.58)
where dy = dy1 dy2 dy3 . Now make two assumptions in Eq. 11.58:
Assumption ii) means that the mean velocity gradient can be taken outside the integral.
Now multiply Eq. 11.58 with vi /xj + vj /xi . Since this term is not a function of
y it can be moved in under the integral. We obtain after time averaging
vi (x) vj (x)
+
xj
xi
1
p (x)
k (x) 1
v
x 2
vi (x) vj (x)
+
xj
xi
v (y) dy3
yk |y x|
Mijk
1
4
vi (x) vj (x)
+
xj
xi
(11.59)
dy3
2
vj
vi
+
xj
xi
= Aij + Mijk
k
v
= ij,1 + ij,2
x
(11.60)
where the rst term represents the slow term, ij,1 (see Eq. 11.47), and second term
the rapid term, ij,2 (index 2 denotes the rapid part).
Now we will take a closer look at rapid part (i.e. the second term) of Mijk . The
second term of Mijk in the integral in Eq. 11.59 can be rewritten as
vj (x) v (y)
=
xi
yk
yk
v (y)
vj (x)
xi
v (y)
v (y)vj (x)
yk xi
yk
v (y)vj (x)
yk xi
2 vj (x)
yk xi
vj (x)
v (y)
xi
(11.61)
89
2 vj (x)/yk xi on line 1 is zero because vj (x) is not a function of y. For the same
reason the last term on line 2 is zero.
Note that the terms above as well as in Eq. 11.59 are two-point correlations, the
two points being x and y. Introduce the distance vector between the two points
ri = yi xi
(11.62)
ri
yi
xi
(11.63)
=
xi
ri
=
yi
ri
(11.64)
We can now write Mijk in Eq. 11.59, using Eqs. 11.61 and 11.64, as
Mijk =
1
2
2
2
v vj +
v v
rk ri
rk rj i
= aijk + ajik
dr3
|r|
(11.65)
It can be shown that aijk is symmetric with respect to index j and (recall that v and
aijk = aikj
(11.66)
(11.67)
Now let us formulate a general expression of aijk which is linear in vi vj and
symmetric in (j, ) and (i, k). We get
aijk = c1 ik vj v
+ c2 j vi vk
+ c3 (ij vk v + kj vi v + i vk vj + k vi vj )
+ c4 j ik k
+ c5 (ij k + jk i )k
(11.68)
90
Each line is symmetric in (j, ) and (i, k). For example, on line 3, term 1 & term 3 and
term 2 & term 4 are symmetric with respect to j and and term 1 & term 2 and term 3
& term 4 are symmetric with respect to i and k.
Consider Eq. 11.59. Here it is seen that if i = j then Mijk = 0 due to the
continuity equation; looking at Eq. 11.65 we get
aiik = 0
(11.69)
+ c4 i ik k + c5 (3k + ik i )k
= c1 vk v + c2 v vk + c3 (3vk v + vk v + vk v + 2k k)
(11.70)
+ c4 k k + c5 (3k + k )k
= vk v (c1 + c2 + 5c3 ) + kk (c4 + 2c3 + 4c5 )
(11.71)
(11.72)
= 3c1 vj v + 2c2 j k + 4c3 vj v + (3c4 + 2c5 )j )k
= vj v (3c1 + 4c3 ) + j k(2c2 + 3c4 + 2c5 )
c4 + 2c3 + 4c5 = 0
3c1 + 4c3 2 = 0,
(11.73)
for the ve unknown constants. Let us express all constants in c2 which gives
c1 =
4c2 + 10
,
11
c3 =
3c2 + 2
,
11
c4 =
50c2 + 4
,
55
c5 =
20c2 + 6
55
(11.74)
+c3
(11.75)
We nd that the c1 term and the second and third part of the c3 term can be merged.
Furthermore, the c2 term and the third and fourth part of the c3 term can be merged as
c1 (Eq. 11.47
c2 (Eq. 11.77)
c2 (Eq. 11.78)
91
LRR model
1.5
0.4
LRR-IP model
1.5
0.6
ij,2 =
(11.76)
11
3
55
ij,2 =
(11.77)
(11.78)
It can be noted that there is a close similarity between the Rotta model and Eq. 11.78:
both models represent return-to-isotropy, the rst expressed in vi vj and the second
in Pij . The model in Eq. 11.78 is commonly called the IP model (IP=Isotropization
by Production) . Since two terms are omitted we should expect that the best value of
should be different than (c2 + 8)/11; a value of = 0.6 was (c2 = 1.4) found to
give good agreement with experimental data. Since Eq. 11.78 is a truncated form of
Eq. 11.77 it does not satisfy all requirements that Eq. 11.77 do. Equation 11.78 does
satisfy symmetry condition and continuity but it does not satisfy the integral condition
in Eq. 11.71. Although Eq. 11.78 is a simpler, truncated version of Eq. 11.77 is often
gives more accurate results [20]. Since the IP model is both simpler and seems to
be more accurate than Eq. 11.77 this is one of the most popular models of the rapid
pressure-strain term. The coefcients for the slow and rapid terms in the LRR and
LRR-IP models are summarized in Table 11.1
11.7.6 Wall model of the pressure-strain term
When we derived the rapid pressure-strain model using Greens function in Eq. 11.58
we neglected the inuence of any boundaries. In wall-bounded domains it turns out
that the effect of the walls must be taken into account. Both the rapid term in the LRR
model and the IP model must be modied to include wall modeling.
92
x2
x1
(11.79)
(11.80)
where xi xi,w and ni,w denotes the distance vector to the wall and the unit wallnormal vector, respectively. As explained above, this damping is inviscid (due to pressure) and affects the turbulent uctuations well into the log-region. It has nothing to do
with viscous damping. Away from the wall, in the fully turbulent region, the damping
function goes to zero since the the distance to the wall (|xi xi,w |) increases faster
than the turbulence length scale. In the viscous region this term is not relevant since it
should account for inviscid damping. Function f should not exceed one.
93
(11.81)
The requirement that the sum of the pressure strain term should be zero. i.e. ii,1w =
0, is now satised since 11,1w + 22,1w + 33,1w = 0.
The wall model for the shear stress is set as
3
12,1w = c1w v1 v2 f
2
k
(11.82)
The factor 3/2 is needed to ensure that ii,1w = 0 is satised when the coordinate system is rotated. You can easily prove this by rotating the matrix [11,1w , 12,1w ; 21,1w , 22,1w ]
and taking the trace of in the principal coordinates system (i.e. taking the sum of the
eigenvalues).
The general formula for a wall that is not aligned with a Cartesian coordinate axis
reads [21]
3
3
vk vm nk,w nm,w ij vk vi nk,w nj,w vk vj ni,w nk,w f
2
2
(11.83)
An analogous wall model is used for the rapid part which reads
ij,1w = c1w
3
3
ij,2w = c2w km,2 nk,w nm,w ij ki,2 nk,w nj,w kj,2 ni,w nk,w f
2
2
(11.84)
11.8. The modeled vi vj equation with IP model
94
11.8 The modeled vi vj equation with IP model
vi vj ij k
c1
k
3
2
c2 Pij ij P k
3
3
+c1w [ vk vm nk nm ij vi vk nk nj
k
2
t
3
vj vk nk ni ]f
2
xn
3
+c2w [ km,2 nk nm ij ik,2 nk nj
2
t
3
jk,2 nk ni ]f
2
xn
vk
2 vi vj
xk xk
k vi vj
xm
+
+
ck vk vm
xk
(convection)
(production)
(slow part)
(rapid part)
gi vj gj vi (buoyancy production)
2
ij (dissipation)
3
(11.86)
In ASM we assume that the transport (convective and diffusive) of vi vj is related to
that of k, i.e.
vi vj
Cij Dij =
C k Dk
k
Inserting Eq. 11.86 into the equation above gives
Pij + ij ij =
vi vj
Pk
k
(11.87)
Thus the transport equation (PDE) for vi vj has been transformed into an algebraic
equation based on the assumption in Eq. 11.86.
95
Now we want to re-write this equation as an equation for vi vj . Insert the IP models
for ij,1 (Eq. 11.47) and ij,2 (Eq. 11.78) and the isotropic model for ij (Eq. 11.39)
in Eq. 11.87 and multiply by k/ so that
k
2
k
Pij c1 vi vj ij k c2
2
Pij ij P k
3
+ij,1w + ij,2w =
2
ij k
3
vi vj
Pk
Collect all vi vj terms so that
vi vj
Pk
1 + c1 =
2
+ ij k(1 + c1 )
3
k
2
Pij c2 Pij ij P k + ij,1w + ij,2w
3
2
2 k
2
k
Pij ij P c2 Pij ij P k + ij,1w + ij,2w + ij k(P k 1 + c1 )
=
3
3
3
Finally, we can write
vi vj =
2
2
k (1 c2 ) Pij 3 ij P k + ij,1w + ij,2w
ij k +
3
c1 + P k / 1
(11.88)
v
c1 1 + c2 P k / k 2
2
(1 c2 )
3
(c1 1 + P k /) y
c
s
(k 2 /2 )ij ji , (k 3 /3 )ij sjk ski
4 4
3 3
(k 2 /2 )ij sji ,
s
(11.89)
2
k2
k3
s
kij + G(1) sij + +G(2) 2 (ik kj ik skj )
(11.90)
96
x2
n
v1 (x2 )
Xx x1
s
Xx
= sij ,
2
Tij
n
G(n) Tij
n=1
1
3
Tij = sik skj ij sik ski
3
5
Tij = ik skm smj sim smk kj
= sik kj sjk ki ,
1
4
Tij = ik kj ij ik ki ,
3
2
6
Tij = im mk skj + sik km mj ij pm mk skp
3
8
7
Tij = im smk kn nj im mk skn nj , Tij = sim mk skn snj sim smk kn snj
2
9
Tij = im mk skn snj sim smk kn nj ij pm mk skn snp
3
T 10 = im smk skn np pj im mk skn snp pj
ij
(11.91)
n
where Tij may depend on the ve invariants in Eq. 11.89. Equation 11.91 is a general
form of a non-linear eddy-viscosity model. Any ASM may be written on the form of
Eq. 11.91.
It may be noted that Eq. 11.91 includes only linear and quadratic terms of sij and
ij . That is because of Cayley-Hamilton theorem which states that a second-order tensor satises its own characteristic equation (see Section 1.20 in [2]); hence cubic terms
or higher can recursively be expressed in linear ( sij ) and quadratic tensors (ik skj ).
s
Furthermore, note that all terms in Eq. 11.91 are symmetric and traceless as required
by the left side, vi vj 2ij k/3.
j
v
v
i
vj vk
xk
xk
97
2
P12 = v2
1
v
x2
1
v
x2
P22 = 0
2
Is v2 zero because its production term P22 is zero? No! The sympathetic term ij
2
2
which takes from the rich (i.e. v1 ) and gives to the poor (i.e. v2 ) saves the unfair
situation! The IP model for ij,1 and ij,2 gives
2
2
k v2 > 0
k 3
2 1
1
v
>0
= c2 P11 = c2 v1 v2
3
3
x2
22,1 = c1
22,2
2
Note also that the dissipation term for the v1 v2 is zero, but it takes the value 3 for the
2
2
v1 and v2 equations (see p. 83). Since the modelled v1 v2 does not have any dissipation
term, the question arises: what is the main sink term in the v1 v2 equation? The answer
is, again, the pressure strain term ij,1 and ij,2 .
upwards (i.e. /x3 > 0), then the ow is stably stratied. This is illustrated in
Fig. 12.1. Consider /x3 > 0. This means that the density decreases with increasing vertical height, i.e. /x3 < 0. If a uid particle is displaced form its equilibirum
level 0 upto level 2, see Fig. 12.1, it is heavier then the surrounding at this new level
(0 > 2 ). Hence, the buoyancy forces the particle back to its original positition 0. In
this way the vertical turbulent uctuations are dampened. Similarily if a particle originating at level 0, is moved down to level 1. Here it is lighter than its new environment,
and hence buoyance makes it to move make to its original level 0.
For the case of unstable stratication, the situation is reversed. Cold uid is
located on top of hot uid, i.e. /x3 < 0 and /x3 > 0. In Fig. 12.1 we would
then have 2 > 0 ). If a uid particle at level 0 is displaced upward to level 2, it is at
this location lighter than its new enviroment; hence it continues to move upwards. If it
is moved down to level 1 it is heavier than its new enviroments and it will then continue
98
2 > 0
2 < 0
/x3 > 0
/x3 < 0
x3
1 < 0
1 > 0
Figure 12.1: Stable stratication due to positive temperature gradient /x3 > 0.
downwards. Hence, turbulent uctuations are enhanced. This ow situation is called
unstable stratication.
The production term due to buoyancy reads (see Eq. 11.10)
G33 = 2gv3
(12.1)
since gi = (0, 0, g). From the equation for the turbulent heat ux, v3 (i.e. Eq. 11.17
with i = 3), we nd the production term for v3
P3 = v3 vk
3
v
vk
xk
xk
(12.2)
In the case illustrated in Fig. 12.1, the production term due to temperature gradient
2
reads P3 = v3 /x3 < 0 (recall that we assume that buoyancy dominates so that
the rst term in Eq. 12.2 is much larger than the second one). Since the main source
term in the v3 equation, P3 , is negative, it makes v3 < 0 so that G33 < 0 (see
Eq. 12.1). Thus, for the case illustrated in Fig. 12.1, we nd that the production term,
G33 , due to buoyancy yields a damping of the vertical uctuations as it should.
Note that the horizontal turbulent uctuations are not affected by the buoyancy
term, Gij , since G11 = G22 = 0 becaue the gravity is in the x3 direction (i.e. g1 =
g2 = 0).
If the situation in Fig. 12.1 is reversed so that /x3 < 0 the vertical uctuations
are instead augmented. This is called unstably stratied conditions.
When eddy-viscosity models are used, transport equations are usually not solved
Gk = 0.5Gii = gi vi
(12.3)
For gi = (0, 0, g), it reads Gk = gv3 which with Eq. 11.25 gives
Gk = g
t
xi
(12.4)
99
A
0
0
B
r
B
V (r)
Figure 12.2: Flow in a polar coordinate system illustrating streamline curvature. The
streamline is aligned with the axis.
r
v1
x2
streamline
x1
Figure 12.3: Streamline curvature occurring when the ow approaches, for example, a
separation region or an obstacle.
Hence it is seen that in stably stratied conditions, Gk < 0 as required. The difference between an eddy-viscosity model and a Reynolds stress model, is that the former
reduces k whereas the latter reduces only the vertical uctuations.
100
V /r > 0
convex curvature
concave curvature
V /r < 0
stabilizing
destabilizing
destabilizing
stabilizing
=0
r
r
(12.5)
Here the variables are instantaneous or laminar. The centrifugal force exerts a force in
the normal direction (outward) on a uid following the streamline, which is balanced
by the pressure gradient. Since we have assumed that V /r > 0, Eq. 12.5 shows
that the pressure gradient increases with r. If the uid is displaced by some disturbance
(e.g. turbulent uctuation) outwards to level A, it encounters a pressure gradient larger
than that to which it was accustomed at r = r0 , as (V )A > (V )0 , which from Eq. 12.5
gives (p/r)A > (p/r)0 . Hence the uid is forced back to r = r0 . Similarly, if
the uid is displaced inwards to level B, the pressure gradient is smaller here than at
r = r0 and cannot keep the uid at level B. Instead the centrifugal force drives it back
to its original level.
It is clear from the model problem above that convex curvature, when V /r > 0,
has a stabilizing effect on (turbulent) uctuations, at least in the radial direction. It is
discussed below how the Reynolds stress model responds to streamline curvature.
Assume that there is a at-plate boundary layer ow, see Fig. 12.3. The ratio
2
2
of the normal stresses v1 and v2 is typically 5. At one x1 station, the ow is
deected upwards. How will this affect the turbulence? Let us study the effect of
concave streamline curvature. The production terms Pij owing to rotational strains
(1 /x2 , 2 /x1 ) can be written as (see Eq. 11.10):
v
v
2
RSM, v1 eq. : P11 = 2v1 v2
2
RSM, v1 v2 eq. : P12 = v1
1
v
x2
2
v
v
2 1
v2
x1
x2
2
RSM, v2 eq. : P22 = 2v1 v2
P = t
(12.6a)
2
v
x1
1
v
2
v
+
x2
x1
(12.6b)
(12.6c)
(12.6d)
12.3. Stagnation ow
101
v1 (x2 )
x2
x1
12.3 Stagnation ow
The k model does not model the normal stresses properly, whereas ASM/RSM do.
The production term in the k equations for RSM/ASM and k model in stagnation
ow (see Fig. 12.5) due to 1 /x1 and 2 /x2 is:
v
v
2
RSM : 0.5 (P11 + P22 ) = v1
k : P k = 2t
1
v
v
v
2 2 = 1 (v 2 v 2 )
v2
2
x1
x2
x1 1
1
v
x1
2
v
x2
(12.7)
(12.8)
102
x2
x2
x1
x1
x2
x1
Disadvantages:
2 2 2
i) isotropic, and thus not good in predicting normal stresses (v1 , v2 , v3 )
13. Realizability
103
13 Realizability
There are a number of realizability constraints. The usual two ones are that all normal
stresses should stay positive and that the correlation coefcient for the shear stress
should not exceed one, i.e.
vi vj
2
vi
2
vj
2
vi 0 for all i
1/2
(13.1)
These criteria are seldom used in RSMs. However, satisfying the rst criteria is actually
of importance for eddy-viscosity models in stagnation ow [28]. Assume that the ow
is in the x1 direction and that it approaches the wall (see Fig. 12.5). The Boussinesq
2
assumption for the normal stress v1 reads (cf. Eq. 12.7)
2
v1 =
2
1
v
2
= k 2t s11
k 2t
3
x1
3
(13.2)
2
It is seen that if s11 gets too large then v1 < 0 which is unphysical, i.e. non-realizable.
Lets now briey repeat the concept invariants. This means something that is
independent of the coordinate system. Here we mean independent of rotation of the
coordinate system. If a tensor is symmetric, then we know that it has real eigenvalues
which means that we can rotate the coordinate system so that the off-diagonal components vanish (see, e.g., [2]). For the strain tensor this means that the off-diagonal
components of sij vanish and this is the coordinate system where the diagonal com
ponents become largest (e.g. s11 in Eq. 13.2). Thus this is the coordinate system in
2
which the danger of negative v1 from Eq. 13.2 is largest. The equation for nding the
eigenvalues of a tensor Cij is (see e.g. [2] or [29])
|Cij ij | = 0
(13.3)
which gives in 2D
C11
C21
C12
C22
=0
(13.4)
2D
I1 = Cii
1
2D
I2 = (Cii Cjj Cij Cij ) = det(Cij )
2
(13.5)
Since the above equation is the same irrespectively of how the coordinate system is
2D
2D
rotated, it follows that its coefcients I1 and I2 are invariants.
In 3D Eq. 13.3 gives
C11
C12
C21
C22
C31
C32
C13
C23
C33
=0
(13.6)
104
which gives
3D
3D
3D
3 I1 2 + I2 I3 = 0
3D
I1 = Cii
1
3D
I2 = (Cii Cjj Cij Cij )
2
1
3D
I3 = (2Cij Cjk Cki 3Cij Cji Ckk + Cii Cjj Ckk ) = det(Cij )
6
(13.7)
3D
3D
3D
The invariants are I1 , I2 and I2 .
Lets go back to Eq. 13.2 and assume incompressible 2D ow. The rst invariant
reads (cf. Eq. 13.5)
2D
I1 = sii = s11 + s22 = 1 + 2 = 0
(13.8)
It is zero due to the continuity equation. The second invariant of sij reads
2D
I2 = ij sij /2,
s
(13.9)
(see Eq. 13.5) which is the same in all coordinate systems (hence the name invariant).
The solution to Eq. 13.5, using Eq. 13.8, is
2D
1,2 = I2
1/2
sij sij
2
1/2
(13.10)
The eigenvalues of sij correspond to the strains in the principal axis. As discussed
above, we apply Eq. 13.2 in principal coordinate directions of sij . Hence, s11 in
Eq. 13.2 is replaced by the largest eigenvalue so that
2
v1 =
2
k 2t 1
3
(13.11)
2
The requirement v1 0 gives now together with Eq. 13.11
k
k
=
3|1 |
3
2
sij sij
1/2
(13.12)
In 3D, Eq. 13.7 instead of Eq. 13.5 is used, and Eq. 13.10 is replaced by [28]
|k | = k
2ij sij
s
3
1/2
(13.13)
2
Dvi
0
Dt
(13.14)
where D/Dt denotes the material derivative (think of Eq. 13.14 in Lagrangian coordinates, i.e. we follow a uid particle as it approaches the wall). Eq. 13.14 requires that
105
2
when vi approaches zero, the left side (and thus also the right side) of the transport
2
equation of vi should also do so. Since we are here concerned about the pressure2
strain term, well take a look at how it behaves near walls when vi 0. This is
of some relevance in near-wall turbulence where the wall-normal stress goes to zero
faster than the wall-parallel ones: this state of turbulence is called the two-component
limit [30]. Neither the form of ij,2 in Eq. 11.78 nor Eq. 11.77 satisfy the requirement
2
that 22,2 = 0 when v2 = 0 [17]. In Eq. 11.78, for example,
2
22,2 ij P k
3
(13.15)
Very complex forms of ij,2 have been proposed [31] [CL96] which include terms cu
bic in vi vj . The CL96 model does satisfy the two-component limit. Another advantage
of the CL96 model is that it does not need any wall distances, which is valuable in
complex geometries.
The models of the slow pressure-strain in Eq. 11.47 (linear model) and Eq. 11.53
(non-linear model) do also not satisfy the two-component limit. The Rotta model, for
example, gives
2
22,1 c1
(13.16)
3
The only way to ensure this is to make c1 0 when the wall is approached. A
convenient parameter proposed by [30] is A which is an expression of A2 and A3 (the
second and third invariant of aij , respectively), i.e.
9
A2 = aij aji , A3 = aij ajk aki , A = 1 (A2 A3 )
8
(13.17)
sij
k
i
v
j
v
+
xj
xi
aij = 2t
sij =
1
2
(14.1)
vi vj
2
ij
k
3
(14.2)
The relation between the stress vi vj and the velocity gradient in Eq. 14.1 is, as can
be seen, linear. One way to make eddy-viscosity models more general, is to include
106
non-linear terms [22]. A subset of the most general form reads [32]
aij = 2c sij
1
+ c1 2 sik skj sk sk ij + c2 2 ik skj sik kj
3
1
+ c3 2 ik jk k k ij + c4 3 sik sk j i sk skj
3
2
+ c5 3 i m smj + si m mj mn n sm ij
3
+ c6 3 sk sk sij + c7 3 k k sij
1
ij =
2
(14.3)
i
v
j
v
xj
xi
where is a turbulent time scale; for a non-linear k model = k/, and for a nonlinear k model = 1/. The tensor groups correspond to a subset of Eq. 11.91:
1
Line 1: Tij ,
3
2
Line 2: Tij and Tij
4
5
Line 3: Tij and Tij
6
Line 4: Tij
1
Line 5: Tij multiplied by the invariants sk sk and k k
v
The expression in Eq. 14.3 is cubic in i /xj . However, note that it is only
quadratic in sij and ij . This is due to Cayley-Hamilton theorem which states that a
tensor is only linearly independent up to quadratic terms, see p. 86; this means that, for
example, s3 = sik sk sj can be expressed as a linear combination of s2 = sik skj and
ij
ij
sij .
aij is symmetric and its trace is zero; it is easily veried that the right side of
Eq. 14.3 also has these properties. Examples of non-linear models (sometimes also
called explicit algebraic Reynolds stress models, EARSM) in the literature are the models presented in [33, 34, 32, 35]. EARSMs are very popular especially the model in
[35] in the aeronautical community where explicit time-marching solver are used.
They are computationally cheap, more accurate than linear eddy-viscosity models and
they dont give rise to any numerical instabilities as in implicit solvers (like SIMPLE).
In implicit solvers a large turbulent viscosity in the diffusion term of the momentum
equations is needed to stabilize the solution procedure.
107
12
1
v
x2
a11 =
1 2
12
1
v
x2
a22 =
1
= 2
6
1
v
x2
a33
a12 = c
(c1 + 6c2 + c3 )
(c1 6c2 + c3 )
(14.4)
(c1 + c3 )
1
1
v
+ 3
x2
4
1
v
x2
(c5 + c6 + c7 )
0.82 2
12
a22 =
0.5 2
12
1
v
x2
u1
x2
a33
a12
2
0.82 2
k+
k
3
12
2
v1 =
2
0.5 2
k
k
3
12
1
v
x2
2
2
v2 =
v
0.16 2 1
=
12
x2
k 1
v
= c
x2
1
v
x2
2
0.16 2
2
v3 = k
k
3
12
1
v
x2
(14.5)
We nd that indeed the non-linear model gives anisotropic normal Reynolds stresses.
v
In Eqs. 14.4 and 14.5 we have assumed that the only strain is 1 /x2 . When
we discussed streamline curvature effects at p. 101 we found that it is important to
investigate the effect of secondary strains such as 2 /x1 . Lets write down Eq. 14.3
v
for the strain 2 /x1
v
1 2
12
2
v
x1
a11 =
1 2
12
2
v
x1
a22 =
1
= 2
6
2
v
x1
1
a12 = 3
4
2
v
x1
a33
(c1 6c2 + c3 )
(c1 + 6c2 + c3 )
(14.6)
(c1 + c3 )
(c5 + c6 + c7 )
Inserting with values on the constants from [32] (see above) we obtain
0.5 2
12
2
v
x1
0.82 2
12
2
v
x1
a11 =
a22 =
a33 =
0.16 2
12
2
v
x1
(14.7)
2
, a12 = 0
108
As can be seen the coefcient for a22 is larger than that in Eq. 14.5, and hence the
model is slightly more sensitive to the secondary strain 2 /x1 than to the primary
v
one 1 /x2 . Thus, the non-linear models are able to account for streamline curvature,
v
but due to the choice of constants so that c5 + c6 + c7 = 0 this effect is weak.
1 v2
v 2
( + t ) 2 2v2 p /x2 22
+
=
x1
x2
x2
x2
(15.1)
in which the diffusion term has been modelled with an eddy-viscosity assumption, see
Eq. 11.37 at p. 82. Note that the production term P22 = 0 because in boundary-layer
approximation v2 v1 and /x1 /x2 . The model for the dissipation 22 is
taken as
v 2
model = 2
22
k
Adding model on both sides of Eq. 15.1 yields
22
2 v2
v 2
v 2
1 v2
v 2
+
+ 2 =
x1
x2
k
v 2
v 2
( + t ) 2 2v2 p /x2 22 + 2
x2
x2
k
(15.2)
(15.3)
v 2
( + t ) 2 + P 2
+
=
x1
x2
x2
x2
k
(15.4)
109
2
P is the source term in the v2 -equation above, and it includes the velocity-pressure
gradient term and the difference between the exact and the modelled dissipation. Note
that this term is commonly split into the pressure-strain term and a diffusion term as
v2 p /x2 =
v2 p
p v2 /x2
x2
Physically, the main agent for generating wall-normal stress is indeed the pressurestrain term via re-distribution, see example in Section 11.11.
A new variable f = P/k is dened and a relaxation equation is formulated for f
as
L2
22
1
2f
f =
x2
k
T
2
T = max
22
C1
=
k
T
k
, CT
2
2 v2
3
k
2
v2
2
k
3
1/2
t
+ C2
k
k 3/2
, C
L = CL max
1
v
x2
(15.5)
1/4
where 22 is the IP model of the pressure-strain term, see Eqs. 11.47 and 11.78, the
rst term being the slow term, and the second the rapid term. The constants are given
the following values: c = 0.23, CT = 6, ce1 = 1.44, c2 = 1.9, k = 0.9, =
1.3, C1 = 1.3, C2 = 0.3, CL = 0.2, C = 90.
2
2
The boundary condition for f is obtained from v2 equation. Near the wall, the v2
equation reads
2
2
v2
2 v2
0=
(15.6)
2 + fk k
x2
Near the wall, Taylor analysis gives = 2k/x2 [3]; using this expression to replace
2
k in Eq. 15.6 gives
2
2
2 v2
f x2
2v2
2
0=
+
2
(15.7)
2
2
x2
2
x2
Assuming that f and are constant very close to the wall, this equation turns into an
ordinary second-order differential equation with the solution
2
v2 = Ax2 +
2
B
x4
f 2 2
x2
20
2
2
Since v2 = O(x4 ) as x2 0, both constants must be zero, i.e. A = B = 0, so we get
2
f =
2
20 2 v2
x4
2
(15.8)
j v 2
v
( + t )
+ f k
=
xj
xj
xj
k
(15.9)
110
1.5
L=0.04,S=1
L=0.08,S=1
L=0.12,S=1
L=0.2,S=1
L=0.2, S=2
L=0.2, S=0.5
0.5
0
0
0.2
0.4
x2
0.6
0.8
(15.10)
( + t ) 2 + 22
+
=
x1
x2
x2
x2
3
(15.11)
2
which is the usual form of the modelled v2 -equation with isotropic dissipation. Thus
the f equation acts so as to let f go from the value of its source term to its (negative)
wall value over lengthscale L. This is how the reduction of the source term P in
Eq. 15.4 is achieved as the wall is approached. The behavior of the equation for f
(Eq. 15.5) for different right sides is illustrated in the Fig. 15.1 where the equation
L2
2f
f +S =0
x2
2
(15.12)
has been solved with f = 0 at the wall and with different L and S.
As can be seen, f is, as required, reduced as the wall is approached. Furthermore,
f approaches the value of the source term as x2 > L. The inuence of the lengthscale
L is nicely illustrated: the larger L, the further away from the wall does f go to its
far-eld value.
111
(15.13)
The k and -equations are also solved (without damping functions). For convenience, the boundary conditions are given again
2
k = 0, v2 = 0
= 2k/x2
2
(15.14)
2
20 2 v2
f =
x4
2
The boundary condition for f makes the equation system numerically unstable.
2
One way to get around that problem is to solve both the k, and v2 , f equations
coupled [38].
2
xj xj
v 2
k2
2
The resulting v2 and f -equation read [39]
v 2
v2
j v 2
v
=
( + t )
+ kf 6
xj
xj
xj
k
L2
2f
v2
1
Pk
2
(C1 6) (C1 1) + C2
+ f =
xj xj
T
k
3
k
j i
v
i
v
v
+
P k = t
xj
xi xj
1/2
k
,6
T = max
L = CL max
k 3/2
, C
(15.15)
(15.16)
1/4
112
k
0.6k
,
2 ( s )1/2
6C v sij ij
k 3/2
k 3/2
L = min
,
1/2
6C v 2 (2ij sij )
s
(15.17)
These realizable conditions have been further investigated by Sveningsson [38, 40, 41,
42, 43], and it was found that the limitation on T is indeed important, whereas that for
L is not. Furthermore, it was found that it is important to impose the limitation on T
in a consistent manner. For instance, if the limit is used in the f equation, it must for
consistency also be used for /k in Eq. 15.15.
In the V2F model, v 2 denotes the generic wall-normal stress. Thus it should be the
smallest one. This is not ensured in the V2F models presented above. Below the
simple modication proposed by [1] is presented.
The source term kf in the v 2 -equation (Eq. 15.15) includes the modeled velocitypressure gradient term which is dampened near walls as f goes to zero. Since v 2
represents the wall-normal normal stress, it should be the smallest normal stress, i.e.
2
2
2
2
2
2
v2 v1 and v2 v3 , and thus v2 should be smaller than or equal to 3 k. In
the homogeneous region far away from the wall, the Laplace term is assumed to be
negligible i.e. 2 f /xj xj 0. Then Eq. 15.16 reduces to f = right side.
It turns out that in the region far away from the wall, the Laplace term is not negli2
gible, and as a consequence v 2 gets too large so that v 2 > 3 k. A simple modication
is to use the right side of Eq. 15.16 as an upper bound on the source term kf in the
v 2 -equation, i.e.
2
vsource = min kf,
2k
(C1 6)v 2
(C1 1) + C2 P k
k
3
(15.18)
This modication ensures that v 2 2k/3. For more details, see [1].
113
Figure 16.1: Flow around an airfoil. Pressure contours. Red: high pressure; blue: low
pressure
One example of adverse pressure gradient is the ow along the surface of an airfoil,
see Fig. 16.1. Consider the upper surface (suction side). Starting from the leading edge,
the pressure decreases because the velocity increases. At the crest (at x/c 0.15)
the pressure reaches its minimum and increases further downstream as the velocity
decreases. This region is called the adverse pressure gradient (APG) region.
The k model is better than the k model at predicting adverse pressure
gradient ow and the standard model of [45] does not use any damping functions.
However, the disadvantage of the standard k model is that it is dependent on the
free-stream value of [46]
In order to improve both the k and the k model, it was suggested in [44]
to combine the two models. Before doing this, it is convenient to transform the k
model into a k model using the relation = /( k), where = c . The lefthand side of the equation will consist of the convection term, D/Dt, which denotes
the material derivative assuming steady ow. Let us express the left-hand side of the
equation as a combination of the left-hand sides of the and the k equations by using
the chain rule, i.e.
d/( k)
1 d
d(1/k)
d
=
=
+
dt
dt
k dt
dt
dk
1 d dk
1 d
2
=
=
k dt
k dt
k dt
k dt
(16.1)
Now we have transformed the left side of the equation. The right side should be
transformed in the same manner. For example, the production of the equation will
consist of two terms, one term from the equation
1
P
k
(16.2)
Pk
k
(16.3)
114
(cf. with Eq. 16.1). In the same way we transform the entire right side inserting the
modelled equations for k and so that
D
1
1
= P P k k +
Dt
k
k
k
k
Production, P
1
k
T
D
Destruction,
T
D
k k
T
Turbulent diffusion, D
2
2 k
k x2
k x2
j
j
(16.4)
Viscous diffusion, D
Production term
1
P P k = C1 2 P k P k
k
k
k
k
k
= (C1 1) P
k
(16.5)
k = C2
k
k
k
k
= (C2 1) 2
(16.6)
P =
Destruction term
=
D =
2 k
2 k 2 k
2
2 = k x2 k x2
k x2
k xj
j
j
j
k xj
=
k
=
+k
xj
xj
2k
k x2
j
k
2k
k
2
2k
+ 2 +
+k 2
xj xj
xj
xj xj
xj
k x2
j
2 k
+
k xj xj
xj
(16.7)
xj
The turbulent diffusion term is obtained as (the derivation is found in [47] which
can be downloaded from www.tfd.chalmers.se/lada)
T
D =
2t k
+
k xj xj
k
+
k
1
1
t
t
t k
+
xj xj
xj
2k
+
x2
j
t
xj
(16.8)
2t k
+
k xj xj
xj
t
xj
(16.9)
115
+
(j ) =
v
xj
xj
t
2
+
+
k
+ P k 2
xj
k
k
xi xi
(16.10)
k 500
42 k
4
,
(16.11)
F1 = tanh( ), = min max
,
y y 2
CD y 2
are used. F1 = 1 in the near-wall region and F1 = 0 in the outer region. The coefcient, for example, is computed as
SST = F1 k + (1 F1 )k
(16.12)
where k = 0.075 and k = 0.0828. The last term in the equation should only
be active in the k region; hence it is multiplied by (1 F1 ).
At p. 112 it was mentioned that the k model is better than the k model
in predicting adverse pressure-gradient ows because it predicts a smaller shear stress.
Still, the predicted shear stress is too large. This brings us to the second modication
(see p. 112). When introducing this second modication, [44] noted that a model (the
Johnson - King model [JK]) which is based on transport of the main shear stress v1 v2 ,
predicts adverse pressure gradient ows much better than the k model. In the JK
model, the v1 v2 transport equation is built on Bradshaws assumption [48]
v1 v2 = a1 k
1/2
where a1 = c
written as
v1 v2
(16.13)
k 1
c k 2
v
v
c k 2 1
=
=
= c1/2 k
x2
x2
2
1
v
x2
2 1/2
= a1 k
Pk
1/2
(16.14)
It is found from experiments that in the boundary layer of adverse pressure gradient
ow the production is much larger than the dissipation (P k ) and v1 v2 a1 k,
which explains why Eq. 16.14 over-predicts the shear stress and works poorly in this
type of ow. To reduce |v1 v2 | in Eq. 16.14 in adverse pressure gradient ow, [44]
proposed to re-dene the turbulent eddy viscosity including the expression in Eq. 16.13.
We have two expressions for the turbulent viscosity
a1 k
v1 v2
=
k
a1 k
t = =
a1
t =
(16.15a)
(16.15b)
116
where is the absolute vorticity (in boundary layer ow = 1 /x2 ); in (a) the
v
Boussinesq assumption together with Eq. 16.13 were used and (b) is taken from the
k model. We want (a) to apply only in the boundary layer and hence we multiply
it with a function F2 (similar to F1 ) which is 1 near walls and zero elsewhere. Then we
take the minimum of (a) and (b) so that
t =
a1 k
max(a1 , F2 )
(16.16)
When the production is large (i.e. when is large), Eq. 16.16 reduces t . It is important
to ensure that this limitation is not active in usual boundary layer ows where P k .
It can be seen that t is reduced only in regions where P k > , because if P k < then
< a1 since
= c 2 = a 2 2
2 = t 2 = P k <
1
t
k
k
(16.17)
been introduced [49]. The rst modication is that the absolute vorticity in Eq. 16.16
has been replaced by || = (2ij sij )1/2 which is the norm (i.e. the length) of the
s
s
tensor i /xj + j /xi in the Boussinesq assumption (see Eq. 11.24), i.e.
v
v
i
v
j i
v
v
+
= 2ij (ij + ij ) = 2ij sij
s s
s
xj
xi xj
1 i
v
j
v
ij =
2 xj
xi
|| =
s
(16.18)
117
0.5
0.45
v1 , v1
0.4
0.35
0.3
0.25
0.2
no filter
one filter
0.15
0.1
2
2.5
x1
3.5
1
2T
(t)dt, = +
(17.1)
(note that we use the notation . for time averaging). In LES we lter (volume average)
the equations. In 1D we get (see Fig. 17.1)
x+0.5x
1
(, t)d
(x, t) =
x x0.5x
= +
Since in LES we do not average in time, the ltered variables are functions of space
and time.
The equations for the ltered variables have the same form as Navier-Stokes, i.e.
1 p
2 vi
ij
i
v
(i vj ) =
v
+
+
t
xj
xi
xj xj
xj
vi
=0
xi
(17.2)
ij = vi vj vi vj
vi = 0
v i = vi
(17.3)
118
This is true for box lters. Note that for the spectral cut-off lter vi = v i , see
p. 120. However, in nite volume methods, box lters are always used. In this course
we use box lters, if not otherwise stated.
Lets look at the ltering of Eq. 17.2 in more detail. The pressure gradient term, for
example, reads
1
p
p
=
dV
xi
V V xi
Now we want to move the derivative out of the integral. When is that allowed? The
answer is if the integration region is not a function of xi , i.e. if V is constant. In nite
volume methods, the ltering volume, V , is (almost always) identical to the control
volume. In general, the size of the control volume varies in space. Fortunately, it can
be shown that if V is a function of xi , the error we do when moving the derivative out
of the integral is proportional to V 2 , i.e. it is an error of second order. Since this is the
order of accuracy of our nite volume method anyway, we can accept this error. Now
lets move the derivative out of the integral, i.e.
p
=
xi
xi
1
V
pdV
V
+O V2 =
+O V2
xi
=
xj
xj
1
V
vi vj dV
V
+O V2 =
(vi vj ) + O V 2
xj
There is still a problem with the formulation of this term: it includes an integral of a
1
2T
v dt =
T
1
v 2T = v
2T
Lets lter vI once more (lter size x, see Fig. 17.2. For simplicity we do it in
1
x
1
x
x/2
v ()d =
x/2
1
x
x
x
vA +
vB .
2
2
x/2
v ()d
v ()d +
x/2
119
x
x
A
I 1
x
I
I +1
x
Figure 17.2: Box lter illustrated for a control volume.
cut-off
E()
I
II
III
1
1
3
3
1
vI1 + vI +
vI + vI+1
2
4
4
4
4
1
v
v
= (I1 + 6I + vI+1 ) = vI
8
vI =
(17.4)
120
GB (r)v(x r)dr
v (x) =
1/, if r /2
0, if r >
GB (r) =
(17.5)
GB (r)dr = 1
It is often convenient to study the ltering process in the spectral space. The lter in
spectral space is particular simple: we simply set the contribution from wavenumbers
larger than cut-off to zero. Hence the cut-off lter lters out all scales with wavenumber
larger than the cut-off wavenumber c = /. It is dened as
1/ if c
0
otherwise
GC () =
(17.6)
1
2
v(r) exp(r)dr
(17.7)
v(r) =
v () exp(r)d
(17.8)
where denotes the wavenumber and = 1. Note that it is physically meaningful to use Fourier transforms only in a homogeneous coordinate direction; in nonhomogeneous directions the Fourier coefcients which are not a function of space
have no meaning. Using the convolution theorem (saying that the integrated product of two functions is equal to the product of their Fourier transforms) the ltering in
Eq. 17.5 is conveniently written
v () = v () =
=
0
=
0
v () exp()d
(17.9)
If we lter twice with the cut-off lter we get (see Eq. 17.9)
v
v = GC G = GC v = v
(17.10)
121
0.4
0.4
v1 0.2
v1 0.2
node 1
0.2
0.2
0
0.5
1.5
0.5
1.5
x1 /L
x1 /L
Figure 17.4: Physical and wavenumber space. Sinus curves with different wavenumbers illustrated in physical space.
using Eqs. 17.9 and 17.6. Thus, contrary to the box-lter (see Eq. 17.4), nothing happens when we lter twice in spectral space. The box lter is sharp in physical space
but not in wavenumber space; for the cut-off lter it is vice versa.
In nite volume methods box ltering is always used. Furthermore implicit ltering
is employed. This means that the ltering is the same as the discretization (=integration
over the control volume which is equal to the lter volume, see Eq. 17.14).
1 = 2/L
(17.11)
and it covers two cells (x1 /L = 0.5). If we dene this as the cut-off wavenumber we
get 1,c 2x1 = 2 so that
1,c = 2/(2x1 ) = /x1
(17.12)
It is of course questionable if v1 in Fig. 17.4a really is resolved since the sinus wave
covered only by two cells. However this is the usual denition of the cut-off wavenumber.
If we require that the highest resolved wavenumber should be covered by four cells
(x1 /L = 0.25), as in Fig. 17.4b, then the cut-off wavenumber is given by 1,c =
2/(4x1 ) = /(2x1 ).
1
ij ij kk = sgs
3
2
sgs = (CS )
= 2sgs sij
(17.13)
122
= (VIJK )
(17.14)
The scalar || is the norm (i.e. the length) of i /xj + j /xi in the Boussinesq
s
v
v
assumption, see Eq. 16.18.
Near the wall, the SGS viscosity becomes quite large since the velocity gradient is
very large at the wall. However, because the SGS turbulent uctuations near a wall go
to zero, so must the SGS viscosity. A damping function f is added to ensure this
f = 1 exp(x+ /26)
2
(17.15)
A more convenient way to dampen the SGS viscosity near the wall is simply to use
the RANS length scale as an upper limit, i.e.
1/3
= min (VIJK )
, n
(17.16)
i
v
j
v
+
xj
xi
i
v
xj
1/2
1/2
= 2 (2ij sij )
s
2 ||.
s
(17.17)
from the resolved turbulence. This energy is transferred to the SGS scales and act as
production term (Pksgs ) in the ksgs equation. The SGS kinetic energy is then transferred to higher wave-numbers via the cascade effect and the kinetic energy is nally
dissipated (=physical dissipation) in the dissipation range. It should be mentioned
that this process is an idealized one. We assume that ALL dissipation takes place in the
dissipation range. This is a good approximation, but in reality dissipation (i.e. transfer
of energy from kinetic energy to internal energy, i.e. increase in temperature) takes
place at all wave numbers, and the dissipation increases for increasing wave number.
E()
123
PK
in
e
rti
al
gs
E()
ran
ge
ks
5/3
gs
dissipating range
c
k=
E()d
0
Changing the lower integration limit to wavenumbers larger than cut-off (i.e. c ) gives
the SGS kinetic energy
E()d
ksgs =
(17.18)
ksgs =
C5/3 2/3 d
(Note that for these high wavenumbers, the Kolmogorov spectrum ought to be replaced
by the Kolmogorov-Pau spectrum in which an exponential decaying function is added
for high wavenumbers [53, Chapter 3]). Carrying out the integration and replacing c
with / we get
2/3
3
(17.19)
ksgs = C
2
In the same way as ksgs can be computed from Eq. 17.18, the resolved turbulent kinetic
energy, kres , is obtained from
c
E()d
kres =
0
124
cut-off, grid lter
test lter
E()
III
c = /
= /
+
t
xj
vivj
1 p
2 v i
Tij
+
xi
xj xj
xj
(17.20)
where the subgrid stresses on the test level now are given by
Tij = vi vj v i v j
(17.21)
125
test lter
grid lter
vi
+
t
xj
vivj
1 p
2 v i
ij
+
xi
xj xj
xj
vi vj v i v j
xj
=
(17.22)
(17.23)
(17.24)
1
ij ij kk
3
(17.25)
In the energy spectrum, the test lter is located at lower wave number than the grid
lter, see Fig. 17.6.
1
2x
v dx =
1
2x
P
W
1
1
(w x + ve x) =
v
2x
2
1
= (W + 2P + vE )
v
v
4
=
v dx +
v dx
vW + vP
vP + vE
+
2
2
(17.26)
126
I, J + 1, K
I, J, K
I 1, J, K
I + 1, J, K
x2
x1
I, J 1, K
1
(I1/2,J1/2,K1/2 + vI+1/2,J1/2,K1/2
v
8
+I1/2,J+1/2,K1/2 + vI+1/2,J+1/2,K1/2
v
+I1/2,J1/2,K+1/2 + vI+1/2,J1/2,K+1/2
v
(17.27)
+I1/2,J+1/2,K+1/2 + vI+1/2,J+1/2,K+1/2 )
v
where
s ij
1 v i
vj
=
+
, | s |=
2 xj
xi
(17.29)
1/2
2 s ij s ij
127
Note that C in Eq. 17.28 is not squared (cf. the Smagorinsky model, Eq. 17.13 at
2
p.121). Hence, C should be compared with CS . Applying the test lter to Eq. 17.28
(assuming that C varies slowly), substituting this equation and Eq. 17.29 into Eq. 17.25
gives
2
1
(17.30)
ss
Lij ij Lkk = 2C | s | s ij 2 ||ij
3
Note that the constant C really is a function of both space and time, i.e. C =
C(xi , t).
Equation 17.30 is a tensor equation, and we have ve (ij is symmetric and traces
less) equations for C. [57] suggested to satisfy Eq. 17.30 in a least-square sense. Let
us dene the error as the difference between the left-hand side and the right-hand side
of Eq. 17.30 raised to the power of two, i.e.
Q=
1
Lij ij Lkk + 2CMij
3
(17.31a)
Mij =
ss
| s | s ij 2 ||ij
(17.31b)
The error, Q, has a minimum (or maximum) when Q/C = 0. Carrying out the
derivation of 17.31a gives
Q
1
= 4Mij Lij ij Lkk + 2CMij
C
3
=0
(17.32)
128
vI
I 1
I +1
v
x
= vI
I,exact
vI vI1
+ O (x)
x
(17.34)
vI1 = vI x
v
x
1
+ (x)2
2
I,num
vI vI1
=
x
v
x
1
x
2
I,num
2v
x2
I,num
+ O (x)3
so that
2v
x2
I,num
+ O (x)2
v
x
= vI
I,exact
v
x
1
x
2
I,num
2v
x2
I,num
+ O (x)2
When this expression is inserted into the LES momentum equations, the second term
on the right-hand side will act as an additional (numerical) diffusion term. The total
diffusion term will have the form
diusion term =
( + sgs + num )
v
x
(17.35)
where the additional numerical viscosity, num 0.5|I |x. This means that the total
v
dissipation due to SGS viscosity and numerical viscosity is (cf. Eq. 17.17)
sgs+num = 2(sgs + num )ij sij
s
For more details on derivation of equations transport equations of turbulent kinetic
energies, see [12].
129
v
ij = vi vj vi vj = (i + vi )(j + vj ) vi vj
= vi vj + vi vj + vj vi + vi vj vi vj
= (i vj vi vj ) + vi vj + vj vi + vi vj
where the term in brackets is denoted the Leonard stresses, the term in square brackets
is denoted cross terms, and the last term is denoted the Reynolds SGS stress. Thus
ij = Lij + Cij + Rij
Lij = vi vj vi vj
Cij = vi vj + vj vi
(17.36)
Rij = vi vj .
Note that the Leonard stresses Lij are computable, i.e. they are exact and dont need
to be modelled.
In scale-similarity models the main idea is that the turbulent scales just above cutoff wavenumber, c , (smaller than ) are similar to the ones just below c (larger than
); hence the word scale-similar. Looking at Eq. 17.36 it seems natural to assume
that the cross term is responsible for the interaction between resolved scales (i ) and
v
Cij = cr (i vj v i v j )
v
(17.37)
M
and Rij = 0 (superscript M denotes Modelled). It was found that this model was not
sufciently dissipative, and thus a Smagorinsky model was added
M
Cij = cr (i vj v i v j )
v
M
2
Rij = 2CS 2 ||ij
ss
(17.38)
Lm = cr (i vj v i v j )
v
ij
m
Cij = 0
(17.39)
m
Rij = Rij = vi vj
Note that the modied Leonard stresses is the same as the unmodied one plus
the modelled cross term Cij in the Bardina model with cr = 1 (right-hand side of
Eq. 17.37), i.e.
M
Lm = Lij + Cij
ij
130
E()
10
+
SGS
SGS
10
20
30
0
50
100
150
x+
2
Figure 17.10: Dissipation terms and production term from DNS data. 963 mesh data
: + ;
: ; +: SGS .
ltered onto a 483 mesh. Re = 500.
SGS
SGS
In order to make the model sufciently dissipative a Smagorinsky model is added,
and the total SGS stress ij is modelled as
ij = vi vj v i v j 2(CS )2 ||ij
ss
(17.40)
= vi v j v i v j = Lm
cr ij
(17.41)
2 vi
ik
1 p
di
v
=
+
dt
xi
xk xk
xk
(17.42)
where d/dt and ik denote the material derivative and the SGS stress tensor, respectively. In the scale-similarity model, the SGS stress tensor is given by
ik = vi vk vi vk
(17.43)
This model is not sufciently dissipative. Let us take a closer look at the equation for
17.19. Forcing
131
the resolved, turbulent kinetic energy, K = vi vi /2, which reads
dK
p vi
1 vk vi vi
vi
2 vi
+
+
=
v
+ vk vi
dt
xk
xi
2 xk
xk xk i
2 vi
ik
ik
ik
vi =
v =
xk
xk
xk xk i
xk i
2K
i i
v v
ik
xk xk
xk xk
xk i
(17.44)
SGS
The rst term on the last line is the viscous diffusion term and the second term, , is
the viscous dissipation term which is always positive. The last term, SGS , is a source
term arising from the SGS stress tensor, which can be positive or negative. When it is
positive, forward scattering takes place (i.e. it acts as a dissipation term); when it is
negative, back scattering occurs.
Figure 17.10 presents SGS dissipation, SGS in Eq. 17.44, computed from ltered
DNS data. The forward scatter, + , and back scatter, , SGS dissipation are
SGS
SGS
dened as the sum of all instants when SGS is positive and negative, respectively. As
can be seen, the scale-similarity model is slightly dissipative (i.e. SGS > 0) , but the
forward and back scatter dissipation are both much larger than SGS .
One way to make the SGS stress tensor strictly dissipative is to set the back scatter
to zero, i.e. max(SGS , 0). This could be achieved by setting ik /xk = 0 when its
sign is different from that of vi (see the last term in Eq. 17.44). This would work if we
were solving for K. Usually we do not, and the equations that we do solve (the ltered
Navier-Stokes equations) are not directly affected by the dissipation term, SGS .
Instead we have to modify the SGS stress tensor as it appears in the ltered NavierStokes equations, Eq. 17.42. The second derivative on the right side is usually called a
diffusion term because it acts like a diffusion transport term. When analyzing the stability properties of discretized equations to an imposed disturbance, v , using Neumann
analysis (see, for example, Chapter 8 in [63]), this term is referred to as a dissipation
term. In stability analysis the concern is to dampen numerical oscillations; in connection with SGS models, the aim is to dampen turbulent resolved uctuations. It is shown
in Neumann analysis that the diffusion term in the Navier-Stokes equations is dissipative, i.e. it dampens numerical oscillations. However, since it is the resolved turbulent
uctuations, i.e. K in Eq. 17.44, that we want to dissipate, we must consider the ltered Navier-Stokes equations for the uctuating velocity, vi . It is the diffusion term in
this equation which appears in the rst term on the right side (rst line) in Eq. 17.44.
To ensure that SGS > 0, we set ik /xk to zero when its sign is different from that
of the viscous diffusion term (cf. the two last terms on the second line in Eq. 17.44).
This is achieved by dening a sign function; for details, see [62].
17.19 Forcing
An alternative way to modify the scale-similarity model is to omit the forward scatter,
i.e. to include instants when the subgrid stresses act as counter-gradient diffusion. In
hybrid LES-RANS, the stresses can then be used as forcing at the interface between
URANS and LES. This new approach is the focus of [64].
132
vi
n+1/2
= vi + tH v n , vi
n
(17.45)
pn+1/2
pn
(1 )t
xi
xi
where H includes convective, viscous and SGS terms. In SIMPLE notation this equation reads
pn+1/2
n+1/2
V
=
aP vi
anb v n+1/2 + SU t
xi
nb
where SU includes all source terms except the implicit pressure. The face velocities
n+1/2
n+1/2
n+1/2
+ vi,j1 ) (note that j denotes node number and i is a tensor
vf,i
= 0.5(i,j
v
index) do not satisfy continuity. Create an intermediate velocity eld by subtracting
the implicit pressure gradient from Eq. 17.45, i.e.
n+1/2
vi = vi + tH v n , vi
n
n+1/2
vi = vi
+ t
(1 )t
pn
xi
pn+1/2
xi
(17.46a)
(17.46b)
n+1/2
/xi = 0 so that
v
1 f,i
2 pn+1
=
xi xi
t xi
(17.47)
The Poisson equation for pn+1 is solved with an efcient multigrid method [65]. In the
pn+1
xi
(17.48)
A few iterations (typically two) solving the momentum equations and the Poisson pressure equation are required each time step to obtain convergence. More details can be
found [66]
1. Solve the discretized ltered Navier-Stokes equation for v1 , v2 and v3 .
3. The Poisson equation (Eq. 17.47) is solved with an efcient multigrid method [65].
4. Compute the face velocities (which satisfy continuity) from the pressure and the
intermediate face velocity from Eq. 17.48
5. Step 1 to 4 is performed till convergence (normally two or three iterations) is
reached.
6. The turbulent viscosity is computed.
133
RANS
2D or 3D
steady or unsteady
2nd order upwind
1st order
more than two-equations
Domain
Time domain
Space discretization
Time discretization
Turbulence model
LES
always 3D
always unsteady
central differencing
2nd order (e.g. C-N)
zero- or one-equation
Table 17.1: Differences between a nite volume RANS and LES code.
v1
t
t1 : start
t2 : end
E()
134
I
II
III
Figure 17.12: Spectrum for k. I: Range for the large, energy containing eddies; II:
the inertial subrange for isotropic scales, independent of the large scales () and the
dissipative scales (); III: Range for small, isotropic, dissipative scales.
ask ourself whenever doing an experiment in, for example, a windtunnel. We start the
windtunnel: when has the ow (and turbulence) reached fully developed conditions so
that we can start the measure the ow. Next question: for how long should we carry
out the measurements.
Both in LES and the windtunnel, the recorded time history of the v1 velocity at
a point may look like in Fig. 17.11. Time averaging can start at time t1 when the
ow seems to have reached fully developed conditions. It is difcult to judge for how
long one should carry out time averaging. Usually it is a good idea to form a nondimensional time scale from a velocity, V (free-stream or bulk velocity), and a length
scale, L (width of a wake, width or length of a recirculation region), and use this to
estimate the required averaging time; 100 time units, i.e. 100L/V , may be a suitable
averaging time.
( + sgs )
+ Pksgs
(j ksgs ) =
v
+
t
xj
xj
xj
3/2
sgs =
1/2
ck ksgs ,
Pksgs = 2sgs sij sij ,
(17.49)
ksgs
= C
Note that the production term, Pksgs , is equivalent to the SGS dissipation in the equation for the resolved turbulent kinetic energy (look at the ow of kinetic energy discussed at the end of [69]).
135
dissipation and the wavenumber II 1/. Thus we can write the SGS viscosity
as
sgs = a (CS )b
(17.50)
Dimensional analysis yields a = 1/3, b = 4/3 so that
sgs = (CS )4/3 1/3 .
(17.51)
The scales in the inertial subrange are isotropic, and thus they are in local equilibrium,
i.e. in the ksgs equation we have that production balances dissipation (see Eq. 17.49)
2sgs sij sij = .
(17.52)
sgs = (CS )2 ||
s
(17.53)
1/2
|| = (2ij sij )
s
s
ksgs
(j ksgs ) = Pksgs +
v
+
t
xj
xj
ef f
3/2
ksgs
xj
ksgs
(17.54)
1
2
a
a
1
2
with ef f = + 2Chom ksgs . The C in the production term Pksgs is computed
dynamically (cf. Eq. 17.33). To ensure numerical stability, a constant value (in space)
of C (Chom ) is used in the diffusion term in Eq. 17.54 and in the momentum equations.
Chom is computed by requiring that Chom should yield the same total production of
ksgs as C, i.e.
1
2
2Cksgs sij sij
xyz
2
= 2Chom ksgs sij sij
xyz
ksgs
.
(17.55)
Now we want to nd a dynamic equation for C . The equations for ksgs and K read in
symbolic form
ksgs Tf (vi,j , vi,j ) = C
3/2
ksgs
K 3/2
(17.56)
136
Since the turbulence on both the grid level and the test level should be in local equilibrium (in the inertial 5/3 region), the left-hand side of the two equations in Eq. 17.56
should be close to zero. An even better approximation should be to assume T (ksgs ) =
T (K), i.e.
3/2
K 3/2
1
= P K C
,
P ksgs C ksgs
so that
n+1
C
=
P K P ksgs +
1 n 3/2
C k
sgs
K2
(17.57)
The idea is to put the local dynamic coefcients in the source terms, i.e. in the production and the dissipation terms of the ksgs equation (Eq. 17.54). In this way the dynamic
coefcients C and C dont need to be clipped or averaged in any way. This is a big
advantage compared to the standard dynamic model of Germano (see discussion on
p. 127).
137
E33 (k3 )
10
10
10
10
10
10
10
3 = 2k3 /x3,max
Figure 17.13: Energy spectra in fully developed channel ow [97]. denotes half
channel width. Number of cells expressed as (/x1 , /x3 ).
: (10, 20);
:
(20, 20);
: (10, 40); : (5, 20); +: (10, 10).
L1
V1,in
H
y
h
x
L2
L
138
E33 (3 )
10
10
10
10
100
200
3 = 2k3 /x3,max
Figure 17.15: Energy spectra E33 (3 ) in the recirculation region and the shear layer
downstream the bump (x1 /H = 1.2). Thick dashed line shows 5/3 slope.
:
: x2 /H = 0.13;
: x2 /H = 0.34 (in the shear
x2 /H = 0.0035 (near the wall);
layer).
18 Unsteady RANS
To perform an accurate LES, a very ne mesh must be used. This causes problems, for
example, near walls. LES is very good for wake ow, where the ow is governed by
large, turbulent structures, which can be captured by a fairly coarse mesh.
However, if attached boundary layers are important, LES will probably give poor
predictions in these regions, unless ne grids are used.
An alternative to LES for industrial ows can then be unsteady RANS (ReynoldsAveraged Navier-Stokes), often denoted URANS (Unsteady RANS) or TRANS (Transient
RANS).
In URANS the usual Reynolds decomposition is employed, i.e.
v (t) =
1
2T
t+T
v(t)dt, v = v + v
(18.1)
tT
The URANS equations are the usual RANS equations, but with the transient (unsteady)
term retained, i.e. (on incompressible form)
vi vj
1 p
2 vi
i
v
(i vj ) =
v
+
+
t
xj
xi
xj xj
xj
vi
=0
xi
(18.2)
Note that the dependent variables are now not only function of the space coordinates,
but also function of time, i.e. vi = vi (x1 , x2 , x3 , t), p = p(x1 , x2 , x3 , t) and vi vj =
v (x , x , x , t).
vi j 1 2 3
Even if the results from URANS are unsteady, one is often interested only in the
time-averaged ow. We denote here the time-averaged velocity as v , which means
v = v + v = v + v + v
(18.3)
139
see Fig. 18.1. The modelled turbulent uctuation, v is not shown in the gure; if this
is added to v + v we obtain v.
What type of turbulence model should be used in URANS? That depends on type
of ow. If the ow has strong vortex shedding, the standard high-Re number k
model can be used, i.e.
k j k
v
=
+
t
xj
xj
j
v
=
+
t
xj
xj
t
k
k
+ P k
xj
(18.4)
+
c1 P k c2
xj
k
(18.5)
k2
(18.6)
i
v
1 p
i vk
v
i
v
( + t )
=
+
+
t
xk
xi
xk
xk
(18.7)
So we are doing unsteady simulations, but still we time average the equations. How
is this possible? The theoretical answer is that the time, T , in Eq. 18.1 should be much
smaller than the resolved time scale, i.e. the modelled turbulent uctuations, v , should
have a much smaller time scale than the resolved ones, v . This is called scale separa
tion. In practice this requirement is often not satised [66]. On the other hand, how do
the momentum equation, Eq. 18.7, know how they were time averaged? Or if they were
volume ltered? The answer is that they dont. The URANS momentum equation and
the LES momentum equation are exactly the same, except that we denote the turbulent
viscosity in the former case by t and in the latter case by sgs . In URANS, much
more of the turbulence is modelled than in LES, and, hence, the turbulent viscosity, t ,
is much larger than the SGS viscosity, sgs .
The common denition of URANS is that the turbulent length scale is not determined by the grid, whereas in LES it is. In URANS we do usually not care about scale
v, v
v
t
Figure 18.1: Decomposition of velocities in URANS.
140
Figure 18.2: Conguration of the ow past a triangular ameholder. Flow from left to
right
The standard k model (Eq. 18.4 and 18.5) was used by [102] for URANS simulations computing the ow around a triangular ame-holder in a channel, see Fig. 18.2.
This ow has a very regular vortex shedding. and the ow actually has a scale
separation. In the gure below the v2 velocity in a point above the ame-holder is
shown. Figure 18.3 shows that the velocity varies with time in a sinusoidal manner.
When were doing URANS, the question arises how the results should be time
averaged, i.e. when should we start to average and for how long. This issue is the same
when doing LES, and this was discussed in the LES-lecture.
141
18.2 Discretization
In LES it is well-known that non-dissipative discretization schemes should be used.
The reason is that we dont want to dampen out resolved, turbulent uctuations. The
same is to some extent true also for URANS. In the predictions on the ame-holder
presented above, the hybrid discretization scheme for the convective terms was used
together with fully implicit rst-order discretization in time; this gives rst-order accuracy in both space and time. The turbulence model that was used was the standard
k model. Thus, both the discretization and the turbulence model have high dissipation. The reason why the unsteadiness in these computations was not dampened out
is that the vortex shedding in this ow is very strong.
In general a discretization scheme which has little numerical dissipation should be
19. DES
142
19 DES
DES (Detached Eddy Simulation) is a mix of LES and URANS. The aim is to treat
the boundary layer with RANS and capture the outer detached eddies with LES. The
model was originally developed for wings at very high angles of attack.
The RANS model that was originally used was the one-equation model by [106].
It can be written [106, 100, Sect. 4.6]
t
j t
v
=
+
t
xj
xj
t = t f1
+ t t
t xj
Cb2 t t
+P
t xj xj
(19.1)
143
f
2 d2 2
1/2
s = (2ij sij )
= Cw1 fw
(19.2)
d = min(d, Cdes )
(19.3)
x , x and x denote the cell length in the three grid directions , and . The
constant Cdes is usually set to 0.65.
In the boundary layer d < Cdes and thus the model operates in RANS mode.
Outside the turbulent boundary layer d > Cdes so that the model operates in LES
mode. The modelled length scale is reduced and the consequence is that the destruction
term Y increases, which gives a reduction in the turbulent viscosity t . A reduced t
gives a smaller production term P so that the turbulent viscosity is further reduced.
At rst sight it may seem that as the model switches from RANS mode to LES
mode thus reducing d, this would give rise to an increased production term P through
the second term (see Eq. 19.2). However, this second term is a viscous term and is
active only close to the wall. This term is sometimes neglected [108]
144
scale. Recently, DES models based on two-equation models have been formulated
by [109, 110, 111]. In these models the turbulent length scale is either obtained from
the two turbulent quantities (e.g. k 3/2 / or k 1/2 /) or the lter width . A model
based on the k model can read
k
(j k) =
v
+
t
xj
(j ) =
v
+
t
xj
xj
xj
t
k
t
+
P k = 2t sij sij ,
k
+ P k T
xj
+ (C1 P k C2 )
xj
k
t = k 1/2 t
The turbulent length scale t and the turbulent dissipation, T , are computed as [111,
112]
t = min C
k 3/2
, Ck
T = max , C
k 3/2
In other models [109, 49] only the dissipation term, T is modied. When the
grid is sufciently ne, the length scale is taken as . The result is that the dissipation in the k equation increases, k decreases which gives a reduced t . A third
alternative is to modify only the turbulent length scale appearing in the turbulent viscosity [112]. A rather new approach is to reduce the destruction term in the equation
as in PANS [113, 114] (Partially Averaged Navier-Stokes) and PITM [115] (Partially
Integrated Transport Modeling). In these models increases because of its reduced
destruction term which decreases both k and t .
The values of the constants can be
(C , , Ck , C , C1 , C2 ) = (0.09, 1.31, 0.07, 1.09, 1.44, 1.92)
(19.4)
Note that a low-Re k model should be used. The AKN model [116] could be a
suitable one.
In regions where the turbulent length scales are taken from (LES mode) the equation is still solved, but is not used. However, it is needed as soon as the model
switches to RANS model again.
In the RANS mode the major part of the turbulence is modelled. When the model
switches to LES mode, the turbulence is supposed to be represented by resolved turbulence. This poses a major problem with this type of models. If the switch occurs at
location x1 , say, it will take some distance L before the momentum equations start to
resolve any turbulence. This is exactly what happens at an inlet in an LES simulation
if no real turbulence is given as inlet boundary conditions. One way to get around this
is to impose turbulence uctuations as forcing conditions [117, 118, 69, 119, 120, ] at
the location where the model switches from RANS mode to LES mode. The forcing is
added in the form of a source term (per unit volume) in the momentum equations.
145
(j k) =
v
+
t
xj
(j ) =
v
+
t
xj
t
k
+ Pk k
k xj
t
Pk
+
+
2
xj
t
1 k
+ 2(1 F1 )2
xi xi
k 500
4
F1 = tanh( ), = min max
,
d d2
t =
xj
xj
a1 k
max(a1 , ||F2 )
s
F2 = tanh( 2 ),
= max
42 k
CD d2
(19.5)
2k 1/2 500
,
d d2
where d is the distance to the closest wall node. The SST model behaves as a k
model near the wall where F1 = 1 and a k model far from walls (F1 = 0). All
coefcients are blended between the k and the k model using the function F1 .
In DES the dissipation term in the k equation is modied as [49]
k kFDES ,
FDES = max
= max {x1 , x2 , x3 } ,
Lt =
Lt
,1
CDES
k 1/2
Again, the DES modication is meant to switch the turbulent length scale from a
RANS length scale ( k 1/2 /) to a LES length scale ( ) when the grid is sufciently ne. When FDES is larger than one, the dissipation term in the k equation
increases which in turn decreases k and thereby also the turbulent viscosity. With a
smaller turbulent viscosity in the momentum equations, the modelled dissipation (i.e
the damping) is reduced and the ow is induced to go unsteady. The result is, hopefully,
that a large part of the turbulence is resolved rather than being modelled.
In some ows it may occur that the FDES term switches to DES in the boundary
layer because z is too small (smaller than the boundary layer thickness, ). Different
proposals have been made [121, 122] to protect the boundary layer from the LES mode
FDES = max
Lt
(1 FS ), 1
CDES
20 Hybrid LES-RANS
When simulating bluff body ows, LES (Large Eddy Simulation) is the ideal method.
Bluff body ows are dominated by large turbulent scales that can be resolved by LES
146
without too ne a resolution and accurate results can thus be obtained at an affordable
cost. On the other hand, it is a challenging task to make accurate predictions of wallbounded ows with LES. The near-wall grid spacing should be about one wall unit in
the wall-normal direction. This is similar to the requirement in RANS using low-Re
number models. The resolution requirements in wall-parallel planes for a well-resolved
LES in the near-wall region expressed in wall units are approximately 100 (streamwise)
and 30 (spanwise). This enables resolution of the near-wall turbulent structures in the
viscous sub-layer and the buffer layer consisting of high-speed in-rushes and low-speed
ejections [96], often called the streak process.
An event of a high-speed in-rush is illustrated in Fig. 20.1. In the lower part of
the gure the spanwise vortex line is shown. Initially it is a straight line, but due to a
disturbance e.g. a turbulent uctuation the mid-part of the vortex line is somewhat
lifted up away from the wall. The mid-part of the vortex line experiences now a higher
v1 velocity (denoted by U in the gure) than the remaining part of the vortex line. As
a result the mid-part is lifted up even more and a tip of a hairpin vortex is formed.
The vorticity of the legs lift each other through self-induction which helps lifting the
tip even more. In the x1 x2 plane (upper part of Fig.. 20.1) the instantaneous and
mean velocity proles (denoted by U and U in the gure, respectively) are shown
as the hairpin vortex is created. It can be seen that an inexion point is created in the
instantaneous velocity prole, U , and the momentum decit in the inner layer increases
for increasing x1 . Eventually the momentum decit becomes too large and the highspeed uid rushes in compensating for the momentum decit. The in-rush event is also
called a sweep. There are also events which occurs in the other direction, i.e. lowspeed uid is ejected away from the wall. These events are called bursts or ejections.
The spanwise separation between sweeps and bursts is very small (approximately 100
viscous units, see Fig. 20.1). This is the main reason why the grid must be very ne
in the spanwise direction. The streamwise distance between the events is related to
the boundary layer thickness (4, see Fig. 20.1). The process by which the events are
formed is similar to the later stage in the transition process from laminar to turbulent
ow. Figure 20.2 presents the instantaneous eld of the streamwise velocity uctuation,
v1 in the viscous wall region. As can be seen, the turbulent structures very elongated
in the streamwise direction.
At low to medium Reynolds numbers the streak process is responsible for the major
part of the turbulence production. These structures must be resolved in an LES in order
to achieve accurate results. Thus, for wall-bounded ows at high Reynolds numbers
of engineering interest, the computational resource requirement of accurate LES is
prohibitively large. Indeed, the requirement of near-wall grid resolution is the main
reason why LES is too expensive for engineering ows, which was one of the lessons
learned in the LESFOIL project [100, 101].
The object of hybrid LES-RANS (and of DES) is to eliminate the requirement of
high near-wall resolution in wall-parallel planes. In the near-wall region (the URANS
region), a low-Re number RANS turbulence model (usually an eddy-viscosity model)
is used. In the outer region (the LES region), the usual LES is used, see Fig. 20.3.
The idea is that the effect of the near-wall turbulent structures should be modelled by
the RANS turbulence model rather than being resolved. In the LES region, coarser
grid spacing in wall-parallel planes can be used. The grid resolution in this region is
presumably dictated by the requirement of resolving the largest turbulent scales in the
ow (which are related to the outer length scales, e.g. the boundary layer thickness)
rather than the near-wall turbulent processes. The unsteady momentum equations are
solved throughout the computational domain. The turbulent RANS viscosity is used in
147
x3
x1
Figure 20.2: Fluctuating streamwise velocity in a wall-parallel plane at x+ = 5. DNS
2
of channel ow [69].
wall
URANS region
LES region
URANS region
x2
wall
x1
Figure 20.3: The LES and URANS region.
x+
2,ml
148
1
30
2
B11 (1 )/v1,rms
x
25
20
v1 15
+
10
5
0
10
x+
2
10
0.8
0.6
0.4
0.2
0
0
10
0.5
x1
1.5
LES region
URANS region
x+
2,ml
x2
wall
x1
Figure 20.5: Using forcing at the interface between the LES and URANS region.
the URANS region, and the turbulent SGS viscosity is used in the LES region.
Much work on hybrid LES-RANS has been carried out. In [123, 66, 124] twoequation models were used in the URANS region and a one-equation SGS model
was employed in the LES region. One-equation models were used in both regions
in [125, 126]. The locations of the matching planes were determined in different ways.
In some work [66, 124] it was chosen along a pre-selected grid plane. In [125] it was
determined by comparing the URANS and the LES turbulent length scales or was computed from turbulence/physics requirements. In [123] they used a two-equation model
in the URANS region and blended it into a one-equation model in the LES region. Different partial differential equations for automatically nding the matching plane were
investigated in [126]. A one-equation model was used in both regions in [127], and the
c coefcient at the interface was computed dynamically to yield a smoother transition between the URANS and LES regions. In [128] they proposed a k turbulence
model, later also used by [129], in which the c2 is made into a function of the ratio of
the RANS and LES length scales. On a ne mesh the model switches smoothly to LES
149
j + 1/2
v2,f
v1,f
Control volume j
LES region
Interface
j 1/2
URANS region
An
Figure 20.6: Added uctuations, vf,1 , vf,2 , vf,3 , in a control volume (j = jml + 1)
in the LES region adjacent to the interface. The uctuations are either synthesized
(subscript f = S) or taken from channel DNS (subscript f = DN S).
150
URANS region
T
C
3/4
c n[1 exp(0.2k 1/2 n/)]
1/4
c k 1/2 n[1 exp(0.014k 1/2n/)]
1.0
LES region
=
0.07k 1/2
1.05
Table 20.1: Turbulent viscosity and turbulent length scales in the URANS and LES
regions. n and denote the distance to the nearest wall and von K rm n constant
a a
(= 0.41), respectively. = (V )1/3
.
ations at the interface. In [118] backscatter was introduced in the interface region with
the object of generating resolved uctuations.
One way to improve hybrid LES-RANS is to add uctuations to the momentum
equations at the interface [119, 69], see Figs. 20.5 and 20.6. The object is to trigger the
equations to resolve turbulence. Adding uctuations in order to trigger the equations
to resolve turbulence is actually very similar to prescribing uctuating turbulent inlet
boundary conditions for DNS or LES (or hybrid LES-RANS). If no triggering inlet
boundary conditions are prescribed in DNS or LES, the resolved turbulence near the
inlet will be too small and a large streamwise distance is required before the equations
trigger themselves into describing turbulent ow. This is also the case in hybrid LESRANS: if no triggering (forcing) is applied at the interface between the LES region and
the URANS region, the resolved turbulence in the LES region near the URANS region
will be too small.
i
v
1 p
( + T )
(20.1)
(i vj ) =
v
+
+
t
xj
xi
xj
xj
where T = t (t denotes the turbulent RANS viscosity) for x2 x2,ml (see Fig. 20.3)
and, for x2 > x2,ml , T = sgs . The turbulent viscosity, T , is computed from an algebraic turbulent length scale (see Table 20.1) and kT ; the latter is obtained by solving
its transport equation, see Eq. 20.2.
kT
( + T )
(j kT ) =
v
+
t
xj
xj
xj
PkT = ij sij ,
3/2
+ PkT C
kT
(20.2)
ij = 2T sij
In the inner region (x2 x2,ml ) kT corresponds to the RANS turbulent kinetic energy,
k; in the outer region (x2 > x2,ml ) it corresponds to the subgrid-scale kinetic turbulent
energy (ksgs ). No special treatment is used in the equations at the matching plane except that the form of the turbulent viscosity and the turbulent length scale are different
in the two regions, see Table 20.1. At the walls, kT = 0.
20.3. Results
151
30
25
25
20
20
15
15
10
10
v /u,w
30
10
x+
2
10
10
(a) Solid lines: isotropic forcing, MS = 0.25; dashdotted lines: isotropic forcing, MS = 1; dashed lines:
no forcing; +: present 963 DNS.
x+
2
10
kT /u2
,w
0.8
0.5vi vi /u2 ,
,w
0.6
v1 v2 /u2 ,
,w
+ 12 /u2
,w
0.4
0.2
0
0
0.2
0.4
x2
0.6
0.8
6
5
3
2
1
0
0
0.2
0.4
x2
0.6
0.8
Figure 20.8: Shear stress and turbulent kinetic energy [69]. Solid lines: no forcing;
dashed lines: forcing with isotropic uctuations with MS = 0.25; : present 963 DNS.
Thick lines: resolved; thin lines: modelled.
20.3 Results
Fully developed channel ow at Re = u / = 2000 ( denotes the channel half
width) is used as a test case to evaluate the effect of different forcing conditions. This
ow may seem to be an easy test case, but it is not. In attempts to improve the performance of LES in wall-bounded ows, the Achilles heel is the near-wall ow region.
The bulk velocity in fully developed channel ow with periodic boundary conditions
(see Eq. 20.1) is entirely determined by the wall shear stress; consequently the ow is
extremely sensitive to the turbulence in the near-wall region.
The streamwise velocity proles obtained with and without forcing are compared
in Fig. 20.7 with the present DNS and the log-law. It can be seen that the centerline velocity is strongly over-predicted when no forcing is used, whereas forcing with
MS = MDN S = 0.25 gives excellent agreement with the log-law (MS and MDN S
152
denote forcing with synthetic and DNS uctuations, respectively). The reason for the
overly large velocities without forcing is that the resolved shear is too small. It can be
seen in Fig. 20.8a that it is the resolved shear stress that increases when forcing is introduced, indicating that the resolved shear stress without forcing is too small. This was
also observed by [118]: when forcing is introduced, the resolved shear stress increases,
which reduces the bulk and centerline velocity.
Recently a novel way for generating uctuations to be used as forcing at the interface was presented [97]. In this work backscatter obtained from a scale-similarity
model was used.
v /x2
2 v /x2
2
(21.1)
which includes the second velocity gradient is a suitable length scale for detecting
unsteadiness. The von K rm n length scale is smaller for an instantaneous velocity
a a
prole than for a time averaged velocity, see Fig. 21.1. This is interesting because, as
noted in [133], the von K rm n length scale decreases when the momentum equations
a a
resolve (part of) the turbulence spectrum.
The rst and second derivatives in Eq. 21.1 are given in boundary layer form. We
want to extend this expression to a general one, applicable in three dimensions. In the
same way as in, for example, the Smagorinsky model, we take the rst derivative as
|| = (2ij sij )1/2 . The second derivative can be generalized in a number of ways. In
s
s
the SAS model it is taken as
U =
2 vi
2 vi
xj xj xk xk
0.5
(21.2)
Hence, the general three-dimensional expression for the von K rm n length scale reads
a a
LvK,3D =
||
s
|U |
(21.3)
153
T1 = 2 S 2
LvK,3D
1 1 k k
2k
max
,
T2 =
2 xj xj k 2 xj xj
L=
k 1/2
1/4
(21.5)
154
0.8
x2
0.6
Lvk,3D
Lvk,1D
0.4
0.2
0
0
10
v1
15
20
Figure 21.1: Velocity proles from a DNS of channel ow. Solid line: time-averaged
velocity with length scale Lvk,1D , Eq. 21.1 ; dashed line: instantaneous velocity with
length scale Lvk,3D , Eq. 21.3.
vj+1 vj
,
x2
=
j+1/2
2v
x2
2
=
j
v
x2
=
j1/2
vj vj1
x2
(x2 )2 4 v
vj+1 2vj + vj1
+
(x2 )2
12 x4
2
j+1
2v
x2
2
and then
vj+2 vj
,
2x2
=
j
v
x2
=
j1
vj vj2
2x2
100
200
x+
2
300
400
500
0
0.1
155
x+
2
10
0.01
0.02
15
20
25
0.03
0.04
0.05
0.3
0.08
0.25
0.06
0.2
0.15
0.04
0.1
0.02
0.05
0
0
0.2
0.4
0.6
0.8
0
0
x2 /
x2 /
Figure 21.2: Turbulent length scales in fully developed channel ow. Left: global
view; right: zoom. DNS. 963 mesh. Re = 500. x1 / = 0.065, x3 / = 0.016,
x2 -stretching of 9%.
: LvK,3D ;
: LvK,1D ;
: (x1 x2 x3 )1/3 ; : x2 .
0
x+
2
500
1500
2000
0.25
0
0.1
x+
2
20
40
0.01
0.02
60
80
100
0.03
0.04
0.05
0.08
0.2
0.06
0.15
0.04
0.1
0.02
0.05
0
0
0.2
0.4
0.6
x2 /
0.8
0
0
x2 /
Figure 21.3: Turbulent length scales in fully developed channel ow. Hybrid LESRANS. Left: global view; right: zoom. 32 64 32 mesh. Re = 2000. x1 / =
0.39, x3 / = 0.19, x2 -stretching of 17%.
: LvK,3D ;
: LvK,1D ;
:
1/4
(x1 x2 x3 )1/3 ; : x2 ; +: k = k 0.5 /(c ).
The von K rm n length scale, LvK,1D , based on the averaged velocity prole
a a
v1 = v1 (x2 ) is also included in Fig. 21.2, and as can be seen it is much larger than
LvK,3D . Near the wall LvK,1D increases because the time-average second derivative,
2 v1 /x2 , goes to zero as the wall is approached. No such behavior is seen for the
2
three-dimensional formulation, LvK,3D .
In Fig. 21.3, data from hybrid LES-RANS are used (taken from [69]). When using
hybrid LES-RANS, part of the turbulence is resolved and part of the turbulence is modelled. The resolved turbulence is dissipated by a modelled dissipation, 2 T sij sij
(T denotes SGS or RANS turbulent viscosity), and T . As a result, the length
scale of the smallest resolved turbulence is larger in hybrid LES-RANS than in DNS.
Close to the wall in the URANS region (x2 < 0.031), the resolved turbulence is
156
dampened by the high turbulent viscosity, and as a results LvK,3D follows closely
LvK,1D .
The RANS turbulent length scale, k , from a 1D RANS simulation at Re =
2000 with the k SST model is also included in Fig. 21.3. In the inner region
(x2 < 0.5), its behavior is close to that of the von K rm n length scale, LvK,1D . In the
a a
center region the RANS turbulent length scale continues to increase which is physically
correct. However, the von K rm n length scale, LvK,1D , goes to zero because the
a a
velocity derivative goes to zero.
Two lter scales are included in Figs. 21.2 and 21.3. In the DNS-simulations,
x2 < (x1 x2 x3 )1/3 near the wall, whereas far from the wall x2 > (x1 x2 x3 )1/3
because of the stretching in the x2 direction and because of small x1 and x3 . In the
hybrid simulations, it can be noted that the three-dimensional lter width is more that
twice as large as the three-dimensional formulation of the von K rm n length scale,
a a
i.e. (x1 x3 x3 )1/3 > 2 LvK,3D .
In [137], the SST-SAS model has been evaluated in channel ow, ow in an asymmetric diffuser and ow over an axi-symmetric hill.
v (x) = 2
n=1
un cos(n x + n ) n
(22.1)
157
n
where un , n and i are the amplitude, phase and direction of Fourier mode n. The
un = (E())1/2
E() = cE
2
u2
(/e )4
rms
e[2(/ ) ]
e [1 + (/e )2 ]17/6
= (i i )1/2 ,
(22.2)
= 1/4 3/4
The coefcient cE is obtained by integrating the energy spectrum over all wavenumbers
to get the turbulent kinetic energy, i.e.
k=
E()d
(22.3)
(22.4)
158
n
2
n
i
x3
n
1
n
3
n
3
n
1
n n
i
x2
x1
n
Figure 22.1: The wave-number vector, n , and the velocity unit vector, i , are orthogi
onal (in physical space) for each wave number n.
where
(z) =
ez xz1 dz
(22.5)
j
N
v1 = 2
un cos( n )1
n=1
N
un cos( n )2
v2 = 2
n=1
(22.6)
v3 = 2
n =
un cos( n )3
n=1
n
k1 x1 +
n
n
k2 x2 + k3 x3 + n
In this way inlet uctuating velocity elds (v1 , v2 , v3 ) are created at the inlet x2 x3
plane.
E()
159
E() 5/3
E(n )
log()
n
Figure 22.2: Modied von K rm n spectrum
a a
(V3 )m
a(V3 )m1
(22.7)
b(v3 )m
(22.8)
and thus Eq. 22.7 is a convenient way to prescribe the turbulent time scale of the uctuations. The inlet boundary conditions are prescribed as (we assume that the inlet is
located at x1 = 0 and that the mean velocity is constant in the spanwise direction, x3 )
v1 (0, x2 , x3 , t) = V1,in (x2 ) + u (x2 , x3 , t)
1,in
(22.9)
v3,in (x2 , x3 , t)
where v1,in = (V1 )m , v2,in = (V2 )m and v3,in = (V3 )m (see Eq. 22.7). The mean
inlet proles, V1,in , V2,in , V3,in , are either taken from experimental data, a RANS
solution or from the law of the wall; for example, if V2,in = V3,in = 0 we can estimate
V1,in as [145]
+
x+ 5
x2
2
+
+
(22.10)
V1,in =
3.05 + 5 ln(x2 ) 5 < x+ < 30
2
1
ln(x+ ) + B
x+ 30
2
2
160
1
0.8
0.6
B( )
0.4
0.2
0
0
0.2
0.4
0.6
0.8
Eq. 22.8;
: computed from synthetic data, (V1 )m , see Eq. 22.7.
The method to prescribed uctuating inlet boundary conditions have been used for
channel ow [144], for diffuser ow [138] as well as for the ow over a bump and an
axisymmetric hill [146].
A. TME075: identity
161
TME075: identity
n
2
1
2
j
1
1
2
k
2
2
1
inm mjk
12m m12 = 123 312 = 1 1 = 1
21m m12 = 213 312 = 1 1 = 1
12m m21 = 123 321 = 1 1 = 1
ij nk ik nj
10=1
0 1 = 1
0 1 = 1
1
3
1
3
1
3
1
1
3
3
3
1
10=1
0 1 = 1
0 1 = 1
2
3
2
3
2
3
2
2
3
3
3
2
10=1
0 1 = 1
0 1 = 1
162
V =1
h
x2
x1
L
Figure B.1: Flow between two plates (not to scale).
B.1
163
Fully developed conditions mean that the ow does not change in the streamwise direction, i.e. v1 /x1 = 0. If we dene fully developed as the location where the
velocity gradient in the center becomes smaller than 0.01, i.e. |v1 /x1 | < 0.01, how
long distance from the inlet does the ow become fully developed?
Another way to dene fully developed conditions can be the x1 position where the
centerline velocity has reached, for example, 99% of its nal value. What x1 value do
you get?
In Section 3.2.2, a distance taken from the literature is given. How well does this
agree with your values?
In the fully developed region, compare the velocity prole with the analytical prole (see Section 3.2.2).
Look at the vertical velocity component, v2 . What value should it take in the fully
developed region (see Section 3.2.2)? What value does it take (at x2 = h/4, for
example)?
B.2
On the lower wall, the wall shear stress, w,L (index L denotes Lower), is computed as
w,L 21,w,L =
v1
x2
(B.1)
L
ti
= ji nj
(B.2)
which is a general way to compute the stress vector on a surface whose normal vector
is n. The expression for ij can be found in Eqs. 1.4 and 2.4; recall that the ow
in incompressible. On the top wall, the normal vector pointing into the surface (i.e.
nj = (0, 1, 0)). Use Eq. B.2 to compute the wall shear stress at the upper wall. Plot
the two wall shear stresses in the same gure. How do they compare? In the fully
developed region, compare with the analytical value (see Eq. 3.30).
B.3
Inlet region
In the inlet region the ow is developing from its inlet prole (v1 = V = 1) to the
fully developed prole somewhere downstream. The v1 velocity is decelerated in the
near-wall regions, and hence the velocity in the center must increase due to continuity.
In Section B.1 you looked at the streamwise velocity in the center. Now plot v1 in the
center and near the wall as a function of x1 . Plot also v1 /x1 . If you, for a xed x1 ,
sum v1 /x1 , what do you get. Why?
B.4
In Section B.3 we found that in the developing region v1 near the walls decreases for
increasing x1 . What about v2 ? How do you explain the behaviour of v2 ?
B.5. Vorticity
B.5
164
Vorticity
B.6
Deformation
In Section 1.4, we divided the velocity gradient into a strain-rate tensor, Dij , and a
vorticity tensor, Vij . Since the ow is two-dimensional, we have only two off-diagonal
terms (which ones?). Plot and compare one of the off-diagonal term of Dij and Vij .
Where are they largest in the domain? Why? What is the physical meaning of Dij and
Vij , respectively? Compare Vij with the vorticity you plotted in Section B.5. Are they
similar? Any comment?
B.7
Choose a slice in the fully developed region. Apply a force balance as in Eq. 3.33 and in
Fig. 3.8. Use this to compute the pressure drop. Compare with the drop from your pressure data, i.e. mean(p 2d(i1,:)) - mean(p 2d(i2,:)) where x1 2d(i1,1)
and x1 2d(i2,1) are the left and right x1 coordinate of the slice, respectively.
Apply the momentum balance to the entire domain. Since the left side of the
momentum equation is not zero in the inlet region, Eq. 3.33 is not valid. You must
use the full equation, Eq. 3.31. Re-write the left-hand side on conservative form, see
Eq. 2.24. Then apply the Gauss divergence theorem as in Eq. 3.33. Compute the
pressure drop and compare it with the actual pressure drop, mean(p 2d(ni,:)) mean(p 2d(1,:)).
B.8
Dissipation
Compute and plot the dissipation, = ji vi /xj . What is the physical meaning
of the dissipation? Where do you expect it to be largest? Where is it largest? Any
difference it its behaviour in the inlet region compared to in the fully developed region?
The dissipation appears as a source term in the equation for internal energy, see
Eq. 2.8. This is discussed in some detail at p. 19. Integrate (sum) the dissipation for the
entire computational domain. Estimate how large temperature increase that is created
by the ow. Assume that the upper and lower walls are adiabatic; furthermore we can
neglect the heat ux by conduction, c1 , (see Eq. 2.10) at the inlet and outlet. Re-write
the left side on conservative form and then apply the Gauss divergence theorem as in
Section B.7.
Hint: see Eqs. 2.12 and 2.13
B.9
Eigenvalues
Compute and plot the eigenvalues of the viscous stress tensor, ij . Use the Matlab
command eig. If you have computed the four elements of the ij matrix you can use
the following commands:
tau=[tau_11 tau 12; tau_21 tau_22];
[n,lambda]=eig(tau)
B.10. Eigenvectors
165
B.10
Eigenvectors
Compute and plot the eigenvectors of ij . Recall that at each point you will get two
eigenvectors, perpendicular to each other. It is enough to plot one of them. An eigenvector is, of course, a vector. Use the Matlab command quiver to plot the eld of the
eigenvectors. Try to analyze why the eigenvectors behave as they do.
B.11
Now you will make a new simulation with Comsol. A tutorial on how to use Comsol
can be found on the course www-page.
Add a bend to the channel (either at the end the channel in Section B or somewhere
closer to the inlet). Try to make the height, h, of the bend equal to that of the channel;
try also to make the length of the channel and the bend equal (a small h and/or large
length, L, increases the pressure loss which makes it difcult to compare the pressure
loss in two channels). If you make the bend sharp, a recirculation region will probably
appear. In that case you should make sure that the channel is long enough after the
bend so that the recirculation region does not extend to the outlet. Feel free to make a
more complex geometry if you want.
The following steps are suggested:
construct the bend so that no recirculation region occurs; try different ways to
add the bend so that the pressure drop (i.e. the losses) gets as small as possible.
construct the bend so that a recirculation region does appear.
Analyze the results of the two congurations.
How large is the pressure drop?
Plot the dissipation in order to nd out where the losses occur. The dissipation variable is not pre-dened in Comsol. You must compute it yourself. Open
Postprocessing/Doman plot parameters. If you now want to plot
dissipation as a surface, choose tab surface. Choose a surface under subdomain selection (usually you have only one). In Expression ll in, for
example, 1.8e-5*(ux*vy) (I assume that = 1.8 105 ). and click on
Apply. Youre now plotting (v1 /x1 )(v2 /x2 ). Now repeat the process
and type the entire expression for the dissipation, . The range of the dissipation
is huge which makes readability poor. To enhance readability, try, for example,
to plot0.1 (in the example above, we would type (1.8e-5*(ux*vy))0.1.
Try to analyze the ow in other ways.
166
To start Comsol Multiphysics, choose Comsol from the Start/Programme menu. Then
select Comsol Multiphysics. In the window that pops up select
Fluid Dynamics/Incompressible Navier-Stokes/
Steady-state analysis
and click
OK
C.1 Geometry
Change the geometry by opening
Options/Axes/Grid Settings
De-activate Axis equal and ll, for example, in
x min: 0
x max: 0.4
y min: 0
y max: 0.1
In the same panel, select the tab
Grid
and de-activate Auto and ll in
x spacing: 0.01
y spacing: 0.001
Press Apply and OK
Now lets draw the geometry. Make the height of the channel approximately 0.01.
From the shortcuts to the left, pick Line. Start by drawing the lower wall. Switch
to 3rd Degree Bezier by left-clicking on the button 3rd Degree Bezier to the left and
draw the bend. By switching between Line and 3rd Degree Bezier draw the complete geometry. When youre done, complete the geometry by right-clicking. Probably
you need to re-draw the geometry a couple of times to get it right.
As an alternative, you can create the grid as a combination of two rectangles (one
horizontal near the inlet and one vertical near the outlet) and a bend drawn with 3rd
Degree Bezier. Use the button Union and Delete interior boundaries to join the
three domains into one single domain.
C.2 Mesh
Generate the mesh by clicking on the button Initialize Mesh on the top. You probably
need to rene the mesh: click on Rene mesh once. To rene the mesh locally in a
region, click on the button Rene Selection and select a region (N.B. a region, not
a point) of the mesh with the left mouse. Note that the ner you make the mesh, the
longer time the computations will take.
167
C.6 Post-processing
Below we show a few examples of what you can plot.
When the solution has converged, a surface plot of the velocity eld is shown. In
order to make a surface plot of another variable, open
Postprocessing/Domain Plot Parameters
and click on tab Surface. Pick a variable in Prened quantities (e.g. Pressure).
Try to understand why the pressure is low in some regions and high in others.
Plot also the vorticity. Does it look as you expect?
Now lets plot the x component of the wall-shear stress, w,x along the lower wall.
Open
Postprocessing/Domain Plot Parameters
and click on tab Line/Extrusion. Select in Predened quantities the viscous force
per area, x comp. Use boundary selection to the left, and click on the boundaries (press
Ctrl and use left-click to pick the boundaries elements that dene the lower wall).
Make sure that the button Line/Extrusion plot is active. Then click on Apply.
Now we want to plot the vorticity along a line. Open
Postprocessing/Cross Section Plot Parameters
and click the tab Line/Extrusion. In Predened quantity choose Vorticity (this
means v2 /x1 v1 /x2 which is the denition of q3 ). Plot it along a vertical line
in the rst part of the channel by lling start (x0, y0) and end coordinate (x1, y1). How
does it compare with your Matlab plot in Part B.
Now its up to you to plot and analyze a number of interesting quantities!
Tutorials on Comsol Multiphysics can be found at here
www.comsol.com/products/multiphysics/research/tutorials/
168
169
170
V2
(V2 |T)
V1
(V1 |T)
Figure E.1: Scalar product.
E.1
Orthogonal functions
(E.1)
see Fig. E.1. Now dene the scalar product of two vectors, a and b, as a b = (a|b).
The coordinates, ci , can be determined by making a scalar product of Eq. E.1 and Vi
which gives
(T|Vi ) = (c1 V1 |Vi ) + (c2 V2 |Vi ) + (c3 V3 |Vi )
= (c1 V1 |V1 ) + (c2 V2 |V2 ) + (c3 V3 |V3 )
2
(E.2)
where |Vi | denotes the length of Vi ; the second line follows because of the orthogonality of Vi . Hence the coordinates, ci , are determined by
ci = (T|Vi )/|Vi |2
(E.3)
171
(f |gn ) =
f (x)gn (x)dx
(E.4)
Then, in a similar way to Eq. E.1, any function can, over the interval [a, b], be expressed
as
f=
cn g n
(E.5)
n=1
As above, we must now nd the coordinates (cf. the coordinates, ci , in Eq. E.1).
Multiply, as in Eq. E.2, f with the basis functions, gi , i.e.
(f |gi ) =
n=1
cn (gn |gi )
(E.6)
Since we know that all gn are orthogonal, Eq. E.6 is non-zero only if i = n, i.e.
(f |gi ) = (c1 g1 |gi ) + (c2 g2 |gi ) . . . ci (gi |gi ) . . . ci+1 (gi+1 |gi ) . . . =
= ci (gi |gi ) = ci ||gi ||2
(E.7)
Similar to Eq. E.3, the coordinates can be found from (switch from index i to n)
cn = (f |gn )/||gn ||2
(E.8)
The coordinates, cn , are called the Fourier coefcients to f in system {g} and
1
||gn || is the length of gn (cf. |Vi | which is the length of Vi in Eq. E.3), i.e.
1/2
1/2
gn (x)gn (x)dx
(E.9)
Let us now summarize and compare the basis functions in physical space and the
basis functions in functional space
1. Any vector in R3 can be expressed in
the orthogonal basis vectors Vi
2. The length of the basis vector, Vi , is
|Vi |
3. The coordinates of Vi are computed
as ci = (T|Vi )/|Vi |2
E.2
Trigonometric functions
(E.10)
172
gn (x) =
for n = 2k = 2, 4, . . .
for n = 2k + 1 = 1, 3, . . .
(E.11)
(n |k ) =
=
cos(nx) cos(kx)dx =
1
2
1
1
1
sin((n + k)x) +
sin((n k)x)
2 n+k
nk
= 0 for k = n
(E.12)
Length of k
(k |k ) = ||k ||2 =
(0 |0 ) = ||0 ||2 =
x 1
+ sin(2x)
2 4
cos2 (kx)dx =
= for k > 0
1 dx = 2
(E.13)
Orthogonality of n and k
(n |k ) =
sin(nx) cos(kx)dx =
1
2
1
1
1
cos((n + k)x) +
cos((n k)x)
2 n+k
nk
=0
(E.14)
Orthogonality of n and k
(n |k ) =
=
sin(nx) sin(kx)dx =
1
2
1
1
1
sin((n k)x)
sin((n + k)x)
2 nk
n+k
= 0 for k = n
(E.15)
Length of k
(k |k ) = ||k ||2 =
sin2 (kx)dx =
x 1
sin(2x)
2 4
= for k 1
(E.16)
E.3
173
Now that we have proved that {g} in Eq. E.11 forms an orthogonal system of func1
tions, we know that we can express any periodic function, f (with a period of 2) in
{g} as
1
f (x) = c +
(E.17)
n=1
1
an = (f |n )/||n ||2 =
1
c = (f |0 )/||0 ||2 =
2
bn = (f |n )/||n ||2 =
f (x) sin(nx)dx
(E.18a)
(E.18b)
f (x)dx
(E.18c)
f (x) =
a0
+
(an cos(nx) + bn sin(nx))
2
n=1
1
an = (f |n )/||n ||2 =
bn = (f |n )/||n ||2 =
E.4
(E.19a)
f (x) sin(nx)dx
(E.19b)
f (x) cos(nx)dx
(E.19c)
(f (x))2 dx =
2
a +
(a2 + b2 )
n
n
2 0
n=1
We will try to prove this formula. Assume that we want to approximate the function
f as well as possible with an orthogonal series
an g n
(E.20)
n=1
Now we want to prove that the Fourier coefcients are the best choice to minimize the
difference
N
||f
n=1
an gn ||
(E.21)
Later we will let N . Using the denition of the norm and the laws of scalar
174
||f
n=1
an gn ||2 =
= (f |f )
n=1
n=1
an (f |gn )
k=1
= (f |f ) 2
an g n f
ak (f |gk ) +
ak g k
k=1
n=1 k=1
an ak (gn |gk ) =
(E.22)
n=1
an (f |gn ) +
n=1
a2 (gn |gn )
n
because of the orthogonality of the function system, {g}N . Express f in the second
1
term using the Fourier coefcients cn (see Eqs. E.5 and E.8) gives
N
(f |f ) 2
n=1
an cn (gn |gn ) +
n=1
a2 (gn |gn )
n
= ||f ||2 +
n=1
= ||f ||2 +
n=1
(E.23)
n=1
||gn ||2 c2
n
The left side of Eq. E.22 is thus minimized if the coefcients an are chosen as the
Fourier coefcients, cn so that
N
||f
n=1
n=1
||gn ||2 c2
n
(E.24)
(f (x))2 dx
for all N
(E.25)
As N is made larger, the magnitude of the left side increases, and its magnitude gets
closer and closer to that of the right side, but it will always stay smaller than ||f ||2 .
This means the series on the left side is convergent. Using the Fourier coefcients in
Eq. E.19 and letting N it can be shown that we get equality of the left and right
side, which gives Parsevals formula,
||f ||2
(f (x))2 dx =
2
(a2 + b2 )
a +
n
n
2 0
n=1
Note that /2 and on the right side correspond to the length of ||gn ||, i.e. ||0 ||,
||n || and ||n ||, respectively.
If you are interested how to make a spectrum from a two-point correlation, read the
short report [149] which can be downloaded here
http://www.tfd.chalmers.se/lada/allpapers.html
E.5
175
Equation E.19 gives the Fourier series of a real function. It is more convenient to
express a Fourier series in complex variables even if the function f itself is real. On
complex form it reads
f (x) =
cn exp(nx))
(E.26a)
n=
cn =
1
2
f (x) exp(nx)dx
(E.26b)
where the Fourier coefcients, cn , are complex. Below we verify that if f is real, then
Eq. E.26 is equaivalent to Eq. E.19. The Fourier coefcients, cn , read assuming that
f is real according to Eq. E.26
cn =
1
2
f (x)(cos(nx) sin(nx))dx =
1
(an bn ),
2
n>0
(E.27)
where an and bn are given by Eq. E.19. For negative n in Eq. E.26 we get
cn = c =
n
1
2
f (x)(cos(nx) + sin(nx))dx =
1
(an + bn ),
2
n > 0 (E.28)
1
2
f (x)dx =
1
a0
2
(E.29)
see Eq. E.19. Inserting Eqs. E.27, E.28 and E.29 into Eq. E.26 gives
f (x) =
1
1
a0 +
(an bn ) exp(nx) + (an + bn ) exp(nx)
2
2 n=1
1
1
a0 +
(an bn )(cos(nx) + sin(nx)) + (an + bn )(cos(nx) sin(nx))
2
2 n=1
1
1
a0 +
an cos(nx) 2 bn sin(nx) = a0 +
an cos(nx) + bn sin(nx)
2
2
n=1
n=1
(E.30)
which veries that the complex Fourier series for a real function f is indeed indentical
to the usual formulation in Eq. E.19 although the Fourier cofcients, cn , are complex.
One advantage of Eq. E.26 over the formulation in Eq. E.19 is that we dont need any
special denition for the rst Fourier coefcient, a0 . The trick in the formulation in
Eq. E.26 is that the imaginary coefcients for negative and positive n cancel whereas
the real coefcients add. This means that the real coefcients are multiplied by a factor
two except the rst coefcient, a0 , which makes up for the factor 1 in front of a0 in
2
Eq. E.19.
176
In this exercise you will use data from a Direct Numerical Simulation (DNS) for fully
developed channel ow. In DNS the unsteady, three-dimensional Navier-Stokes equations are solved numerically. The Re number based on the friction velocity and the
half channel width is Re = u h/ = 500 (h = = u = 1 so that = 1/Re ).
A 96 96 96 mesh has been used. The streamwise, wall-normal and spanwise
directions are denoted by x (x1 ), y (x2 ) and z (x3 ) respectively. The cell size in x and
z directions are x = 0.0654 and z = 0.0164. Periodic boundary conditions were
applied in the x and z direction (homogeneous directions). All data have been made
non-dimensional by u and .
You can do the assignment on your own or in a group of two. You should write a
report where you analyze the results following the heading H1H13. It is recommended
A
A
(but the not required) that you use LTEX(an example of how to write in LTEXis available
on the course www page). It is available on Linux. On Windows you can use, for
example, Lyx (www.lyx.org) or MikTex (www.miktex.org) which are both free
to donwload.
177
Compute and plot also the instantaneous uctuations; v1 at node 1, for example, is
computed as
u1_mean=mean(u1);
u1_fluct=u1-u1_mean;
hist (note that the probability density function of v1 is different from that of v1 ).
Normalize the probability functions, so that (here shown for v1 )
fv1 (v)dv = 1
(F.1)
To carry out the integration you can, for example, use the Matlab command trapz
(type help trapz in the Matlab command window). Investigate how a probability
density changes as the number of bins changes (to nd out what a bin is, type help
hist in the Matlab command window)
Consider the probability density functions of the uctuations. The second moment
corresponds to the variance of the uctuations (or the square of the RMS, see Eq. F.5),
i.e.
v 2 f (v )dv
v 2 =
v 2
As in Eq. 7.4,
is usually computed by integrating in time.
A probability density function is symmetric if positive values are as frequent and
large as the negative values. What about fu 1 . . . fv 4 , do they seem to be symmetric?
Instead of looking at the probability density functions, we should use a denition of
the degree of symmetry, which is the skewness. It is dened as
v 3 f (v )dv
v 3 =
3
and is commonly normalized by vrms , so that the skewness, Sv , of v is dened as
Sv1 =
1
3
v1,rms
3
v1 fv 1 (v)dv
F.4. Mean ow
178
3
3
Again, the integration can also be carried out in time, i.e. Sv1 = v1 /v1,rms . Compute
S for fu 1 . . . fv 4 . Note that you must use a normalized f (see Eq, F.1).
There is yet another statistical quantity which sometimes is used for describing
turbulent uctuations, namely the atness. The variance (the square of RMS) tells us
how large the uctuations are in average, but it does not tell us if there are few very
large uctuations or frequent very small ones. The atness gives this information, and
4
it is dened computed from v 4 and normalized by vrms , i.e.
Fv1 =
1
4
vrms
4
v1 fv 1 (v)dv
4
4
fg (v ) =
1
vrms
exp
v rrms
2
2vrms
= 3
F.4 Mean ow
All data in the data les below have been stored every 10th time step.
Download the le uvw inst small.dat, y.dat and the Matlab le pl vel.m
which reads the data les. The data le includes v1 , v2 and v3 from the same DNS as
above, but now you are given the time history of all x2 nodes at one chosen x1 and x3
node. There are nj = 98 nodes in the x2 direction; node 1 and nj are located at the
lower and upper wall, respectively.
Your data are instantaneous. Compute the mean velocity. Plot it both as linearlinear plot and a log-linear plot (cf. Fig. 6.4).
+
In the log-linear plot, use x+ for the wall distance. Include the linear law, v1 = x+ ,
2
2
+
+
1
and the log law, v1 = ln x2 + B ( = 0.41 is the von K rm n constant and
a a
B = 5.2). How far out from the wall does the velocity prole follow the linear law?
At what x+ does it start to follow the log-law?
2
Compute the bulk velocity
V1,b =
1
2h
2h
v1 dx2
(F.2)
(recall that h denote half the channel width) What is the Reynolds number based on
V1,b and centerline velocity, V1,c , respectively?
2 v1
v1 v2
1 p
+
2 x
x1
x2
2
(F.3)
This equation is very similar to fully developed laminar ow which you studied in Assignment 2 in TME075 Mechanics of solids and uids, see Eq. 3.24; the difference
is that we now have an additional Reynolds stress term. Recall that all terms in the
equation above represent forces (per unit volume). Let us investigate how these forces
179
(the pressure gradient, the viscous term and the Reynolds stress term) affect uid particles located at different x2 locations. Compute and plot the three terms. (the le
uvw inst small.dat does not include p; set p/x = 1.)
If a term is positive it means that it pushes the uid particle in the positive x1
direction. What about the viscous term? Is it always negative? Where is it largest? At
that point, which term balances it? How large is the third term? Why? The pressure
term should be a driving force. The Reynolds stress term, where is it positive and where
is it negative?
vi = vi vi
(F.4)
Now you can easily compute all stresses vi vj . Plot the normal stresses in one gure
and the shear stresses in one gure (plot the stresses over the entire channel, i.e. from
x2 = 0 to x2 = 2h). Which shear stresses are zero?
1/2
(F.5)
Compute the RMS of the wall shear stress. This is a measure of the uctuating tangential force on the wall due to turbulence. If heat transfer is involved, the uctuating
temperature at the wall inducing uctuating heat transfer may be damaging to the material of the walls causing material fatigue. This is probably the most common form of
uid-solid interaction.
180
vj
vi
+
xj
xi
(F.6)
Our data are obtained from incompressible simulations, in which the pressure may
vary unphysically in time (p/t does not appear in the equations). Hence, we prefer
to compute the velocity-pressure gradient term
p =
ij
vj p
vi p
,
xj
xi
(F.7)
2
see the second line in Eq. 9.3. The pressure diffusion term in the v2 equation which
is the difference between Eqs. F.6 and F.7 (the two rst terms in Eq. 9.8) is small
except very close to the wall (see Figs. 9.2 and 9.3). Hence, the difference between p
ij
and ij is small.
Download the data le p inst small.mat and the Matlab le pl press strain.m
which reads the data le. The time histories of the pressure along ve x2 lines [(x1 , x2 , x3 ),
(x1 x1 , x2 , x3 ) and (x1 , x2 , x3 x3 )] are stored in this le. This allows you to
compute all the three spatial derivatives of p . Using the velocities stored in uvw inst small.dat
(see Section F.4), you can compute all the terms in Eq. F.7.
Plot the pressure strain, p , for the three normal stresses and the shear stress across
ij
the channel. For which stresses is it negative and positive? Why?
Which term do you think is the largest source and sink term, respectively, in Eq. 9.12?
F.11 Dissipation
The dissipation of turbulent uctuations (i.e. transformation of turbulent kinetic energy
into internal energy, i.e. increased temperature) reads (see Eq. 9.12)
ij =
vi vj
xk xk
(F.8)
Download the le diss inst.mat and the Matlab le pl diss.m which reads it.
The data le includes the time history of the velocities along ve x2 lines [(x1 , x2 , x3 ),
(x1 x1 , x2 , x3 ) and (x1 , x2 , x3 x3 )] so that you can compute all spatial derivatives. The data cover only the lower half of the channel. Compute and plot all components of ij . Which is largest and which are close to zero (they would be identically
zero had I had provided more samples, i.e. longer time histories)? Plot also = 0.5ii ,
i.e.
v vi
= i
(F.9)
xk xk
Where is it largest and why? In which equation does this quantity appear?
Let us now consider the equations for the mean kinetic energy, K = vi vi /2 (Eq. 8.29)
and turbulent kinetic energy, k = vi vi /2 (Eq. 8.13). The dissipation in the K equation
reads
i i
v v
mean =
(F.10)
xk xk
181
n
ea
Pk
Figure F.1: Transfer of energy between mean kinetic energy (K), turbulent kinetic
energy (k) and internal energy (denoted as an increase in temperature, T ). K =
1
1
2 vi vi and k = 2 vi vi . The dissipations, and mean , are dened in Eqs. F.9 and F.10,
respectively.
The ow of kinetic energy between K, k and T is illustrated in Fig. F.1. Compute
and plot also mean and Pk . Which is large and which is small? How is the major part
of the kinetic energy transformed from K to T ? Is it transformed via k or directly
from K to T ?
correlations of v3 in the x3 direction in which the ow (and turbulence) is homogeneous (this means that the derivative with respect to this spatial coordinate of all time
averaged quantities is zero, i.e. 1 /x3 = 0, vi vj /x3 = 0, /x3 = 0 etc).
v
Download the data le w inst k.mat and y half every forth.dat, and
the Matlab le pl twocorr.m which reads the data les. The data le includes the
time history of v3 in the lower half of the channel and only for every fourth grid point
in the x2 direction. For every x2 grid point the values along the x3 are included (only
every second point).
Now we will compute the two-point correlation between the x3 = 0 node and all
other x3 nodes. For this case the normalized two-point correlation of v3 reads
norm
Bv3 v3 (x2 , x3 ) =
1
v (x2 , x3
2
v3,rms 3
1
2
v3,rms
= 0)v3 (x2 , x3 )
(F.11)
norm
and plot Bv v (x2 , x3 ) for different x2 locations. Does it increase or decrease for
3 3
increasing x2 ? Explain why.
From the two-point correlation, the integral length scale can be computed, see
Eq. 10.5. Compute the integral length scale and plot it versus x2 . The upper limit
in this integral is innity (). You can either carry out the integration using all available data in the x3 direction; in general this may not be optimal and another option
182
is to terminate the integration at the point where the two-point correlation becomes
negative. Compare these two options.
When plotting two-point correlations, it is no point showing both symmetric parts;
show only half of it (cf. the two-point correlations in Section 10.1).
The integral length scale is as the word implies the integral (i.e. an average) of
the size (length) of all eddies. It is mathematically equal to the area below the B norm
function. How large is it compared to half the channel width (which corresponds to the
boundary layer thickness)? How does the integral length compare to the wall distance?
Can it be larger than x2 ? Can an eddy be larger than the wall distance?
183
Week 1
1. What characterizes turbulence? Explain the characteristics. What is a turbulent
eddy?
2. Explain the cascade process. How large are the largest scales? What is dissipation? What dimensions does it have? Which eddies extract energy from the
mean ow? Why are these these eddies best at extracting energy from the
mean ow?
3. What are the Kolmogorov scales? Use dimensional analysis to derive the expression for the velocity scale, v , the length scale, and the time scale, .
4. Make a gure of the energy spectrum. Show the relation between the Fourier
coefents of a uctuating velocity and the energy spectrum. The energy spectrum consists of thee subregions: which? describe their characteristics. Show
the ow of turbulent kinetic energy. Given the energy spectrum, E(), how is
the turbulent kinetic energy, k, computed? Use dimensional analysis to derive
the 5/3 Kolmogorov law.
5. What does isotropic turbulence mean?
6. How is the energy transfer from eddy-to-eddy, , for the large eddies and the
eddies in the inertial range estimated? Show how the ratio of the large eddies to
the dissipative eddies depend on the Reynolds number.
7. Describe the cascade process created by vorticity. Write the vortex stretching/tilting term in tensor notation. What is its physical meaning? Describe the
physical process of vortex stretching which creates smaller and smaller eddies.
Show and discuss the family tree of turbulence eddies and their vorticity. Show
that in 2D ow the vortex stretching/tilting term vanishes.
8. How is the mean velocity computed with a PDF function? How should the function be normalized? Show how the second moment of a uctuating velocity
is computed with a PDF function. How are the skewness and the atness of a
uctuating velocity dened? What is the physical meaning of a small and large
skewness? What is the physical meaning of a small and large atness?
184
185
1. Derive the exact transport equation for turbulent kinetic energy, k. Discuss the
physical meaning of the different terms in the k equation. Which is the main
source and sink term, respectively? Which terms do only transport k?
2. In the cascade process, we assume that the dissipation is largest at the smallest
2/3
1
0 , see Eq. 8.15 at p. 59. Show this. For which
scales, i.e. =
0
eddies is the production largest? Why?
3. Given the exact k equation, give the equation for boundary-layer ow. All spatial
derivatives are kept in the dissipation term: why? In the turbulent region of
the boundary layer, the k equation is dominated by two terms. Which ones?
Which terms are non-zero at the wall? Why? Where is the viscous and turbulent
diffusion terms large? Why? Where is the production term largest?
4. Discuss the difference of spatial transport of k and spectral transfer of k. Give
an example of how they are combined in non-homogeneous turbulence. What is
homogeneous turbulence?
5. Derive the exact transport equation for mean kinetic energy, K. Discuss the
physical meaning of the different terms. One term appears in both the k and
the K equations: which one? Consider the dissipation terms in the k and the
equations: which is largest? Why? Show where they appear in the energy
K
spectrum.
186
1. Derive the exact transport equation for turbulent Reynolds stress, vi vj . Take the
trace of the vi vj equation to obtain the k equation.
2. Discuss the physical meaning of the different terms. The mean normal stress can
2
be dened as vrms,av = vi vi /3; what sign will the pressure-strain term have for
2
normal stresses, respectively, larger and smaller than vrms,av ? What role does
12 has? What sign? Why do we call the pressure-strain term the Robin Hood
term?
3. What about the dissipation term, 12 , for the shear stress?
4. What is the role of the convection and diffusion terms? (use the Gauss divergence
theorem)
5. Consider fully developed channel ow: how are the expressions for the production terms simplied? Which production terms are zero and non-zero, respec
tively? What about the production term for v1 v2 ? Which sign does it have in
the lower and upper part of the channel, respectively? Why is there no pressurestrain term in the k equation?
6. Consider the fully turbulent region in fully developed channel ow: which are
2
2
2
the main source and sink terms in the v1 , v2 , v3 and v1 v2 equations? Which
terms are the largest terms at the wall? Which terms are zero at the wall?
7. Consider channel ow and use physical reasoning to show that v1 v2 must be
negative and positive in the lower and upper half of the channel, respectively. Is
this consistent with the sign of P12 ?
187
1. Being able to use the basic equations learned during the rst part on simple real
ows, for example, to be able to explain why a ne mesh screen acts to reduce
the turbulence level in a wind tunnel, or to explain why a contraction reduces
turbulence.
2. Understanding and being able to estimate uncertainties in measurements of mean
velocities. Understanding the concept of integral time (or lenght) scale.
3. Understand a power (energy) spectrum and being able to estimate how different
lter settings and sample rates affect the power spectrum.
4. Being able to explain Taylors hypothesis and its use.
5. Understanding and being able to estimate the Kolmogorov scales for varying
turbulent Reynolds numbers.
6. Being able to discuss the relation between large and small scales at different
Reynolds numbers.
7. Having acquired a basic knowledge about visualization methods.
8. Being able to describe in general terms how it is possible to nd similarity scales
from the basic equations.
188
In this Appendix we will investigate some properties of aijk in Eq. 11.65 at p. 89.
Introduce the two-point correlation function
Bj (r) = vj (x)v (x + r)
We get
Bj (r)
2 Bj (r)
2 Bj (r)
Bj (r)
=
=
ri
ri
rk ri
rk ri
(H.1)
Since Eq. H.1 in the denition of aijk in Eq. 11.65 is integrated over r3 covering both
r and r (recall that v and vj are separated by r), aijk is symmetric with respect to
index j and , i.e.
aijk = aikj
(H.2)
Greens third formula (it is derived from Gauss divergence law) reads
(x) =
1
4
2
dy3
|y x|
(H.3)
where the boundary integrals have been omitted. Setting = v vj in Eq. H.3 gives
vj v =
1
4
2 v vj dy3
1
= aiji
xi xi |y x|
2
(H.4)
189
In [60] he found that the Leonard term Lij and the cross term Cij are not Galilean
invariant by themselves, but only the sum Lij + Cij is. As a consequence, if the cross
term is neglected, the Leonard stresses must not be computed explicitly, because then
the modelled momentum equations do not satisfy Galilean invariance.
Below we repeat some of the details of the derivation given in [60]. Galilean invariance means that the equations do not change if the coordinate system is moving with a
constant speed Vk . Lets denote the moving coordinate system by , i.e.
x = xk + Vk t, t = t, vk = vk + Vk
k
(I.1)
(xi , t)
j
=
+
=
xk
xk xj
xk t
x
k
x
t
(xi , t)
k
=
+ t t = Vk x + t .
t
t xk
k
(I.2)
From Eq. I.2 is it easy to show that the Navier-Stokes (both with and without lter)
is Galilean invariant [60, 150]. Transforming the material derivative from the (t, xi )coordinate system to the (t , x )-coordinate system gives
i
= + Vk + (vk Vk )
+ vk
t
xk
t
xk
xk
= + vk ,
t
xk
It shows that the left hand side does not depend on whether the coordinate system
moves or not, i.e. it is Galilean invariant.
Now, lets look at the Leonard term and the cross term. Since the ltering operation
the Leonard and the cross term we get (note that since Vi is constant Vi = Vi = Vi )
v
v
v
v
L = vi vj vi vj = (i + Vi )(j + Vj ) (i + Vi )(j + Vj )
ij
= vi vj + vi Vj + vj Vi vi vj vi Vj Vi vj
= vi vj vi vj + Vj (i vi ) + Vi (j vj )
= Lij Vj vi Vi vj
(I.3)
v
v
Cij = vi vj + vj vi = (i + Vi )vj + (j + Vj )vi =
= vi vj + vj Vi + vj vi + vi Vj = Cij + vj Vi + vi Vj
From Eq. I.3 we nd that the Leonard term and the cross term are different in the two
coordinate systems, and thus the terms are not Galilean invariant. However, note that
the sum is, i.e.
190
M
Cij (Eq. 17.36) and the modelled one, Cij (i.e. Eq. 17.37). The exact form of Cij
transforms as in Eq. I.3. The Bardina term transforms as
M
Cij = cr (i vj v i v j )
v
v
v
= cr (i + Vi )(j + Vj ) (i + Vi )(j + Vj )
v
v
(I.5)
= cr [i vj vi vj (i vi )Vj (j vj )Vi ]
v
M
= Cij + cr v i Vj + v j Vi .
M
M
As is seen, Cij = Cij , but here this does not matter, because provided cr = 1 the
M
modelled stress, Cij , transforms in the same way as the exact one, Cij . Thus, as for
M
M
the exact stress, Cij (see Eq. I.4), we have Cij + L = Cij + Lij . Note that in order
ij
to make the Bardina model Galilean invariant the Leonard stress must be computed
explicitly.
191
x3
dAi
n
n
i
x2
x1
0 n 2
0 n 2
0 n
0 n 2
(J.1)
n
J.2. Unit vector i
192
n
i
(1, 0, 0)
(1, 0, 0)
n
i
(0, 0, 1)
(0, 1, 0)
n
0
90
(0, 1, 0)
(0, 1, 0)
(0, 0, 1)
(1, 0, 0)
0
90
(0, 0, 1)
(0, 0, 1)
(0, 1, 0)
(1, 0, 0)
0
90
n
Table J.2: Examples of value of n , i and n from Eqs. J.1 and J.3.
i
n
J.2 Unit vector i
n
Continuity requires that the unit vector, i , and n are orthogonal. This can be seen
j
by taking the divergence of Eq. 22.1 which gives
N
v = 2
n=1
un cos(n x + n ) n n
(J.2)
n
n
i.e. i n = 0 (superscript n denotes Fourier mode n). Hence, i will lie in a plane
i
n
normal to the vector i , see Fig. 22.1. This gives
n
1 = cos(n ) cos(n ) cos(n ) sin(n ) sin(n )
n
2 = sin(n ) cos(n ) cos(n ) + cos(n ) sin(n )
n
3
(J.3)
= sin( ) cos( )
n
n
n
The direction of i in this plane (the 1 2 plane) is randomly chosen through
n
. Table J.2 gives the direction of the two vectors in the case that i is along one
coordinate direction and = 0 and = 90o .
193
K.1 Two-dimensional ow
You can do the assignment on your own or in a group of two. It is recommended (but
A
A
the not required) that you use LTEX(an example of how to write in LTEXis available
on the course www page). It is available on Linux. On Windows you can use, for
example, Lyx (www.lyx.org) or MikTex (www.miktex.org) which are both free
to download.
Youll use data from a coarse DNS. Although some of the data are probably not
fully accurate, in this exercise we consider the data to be exact. You will use Matlab to
2
2
2
read data les of the mean ow (1 , v2 , p) and turbulent quantities. (v1 , v2 , v3 , v1 v2 ,
v
and ). You will analyze one of the following four ows:
Case 1: Flow in an asymmetric diffuser of angle = 10o . Re = 18 000 ( = 5.56105,
= 1) based on inlet bulk velocity and inlet channel height. Hybrid LES-RANS
simulations on a coarse mesh are presented in [119]
Case 2: Flow over a bump (the inlet is located at the middle of the bump). Re =
65 000 ( = 1.00 105 , = 1) based on inlet bulk velocity and inlet channel height. Hybrid LES-RANS simulations at a higher Reynolds number are
presented in [99]
Case 3: Flow over a 2D hill [151, ]. Re = 18 000 ( = 5.56 105 , = 1) based on the
bulk velocity in the channel and the hill height. Hybrid LES-RANS of the ow
over the 3D hill are presented in [152] (the 2D hill corresponds to the centerplane
of the 3D hill).
Case 4: Flow over two small hills Re = 18 000 ( = 5.56 105 , = 1) based on the
bulk velocity in the channel and the height of the hill at the lower wall.
The work should be carried out in groups of two (if you want to work on you own
that is also possible) . Contact the teacher to get a Case No. Download the data from
http://www.tfd.chalmers.se/lada/comp turb model. At the wwwpage you can download a M-le (pl vect.m) which reads the data and plots the vector eld and the pressure contours. You must also download the function dphidx dy.m
which computes the gradients. Make sure you put this function in the directory where
you execute pl vect.m.
The report, along with the Matlab les(s), should be submitted electronically at
the Student Portal www.student.portal.se; the deadline can be found at the
Student Portal.
K.2 Analysis
Study the ow. In which regions do you expect the turbulence to be important?
Now lets nd out. The two-dimensional time-averaged Navier-Stokes for the x1
momentum reads (the density is set to one, i.e. = 1)
1 v1
v
1 v2
v
p
2 v1
v 2
2 v1
v v
+
=
+
1 +
1 2
x1
x2
x1
x2
x1
x2
x2
1
2
(K.1)
K.2. Analysis
194
Recall that all the terms on the right-hand side represent x components of forces per
unit volume.
K.2.1 The momentum equations
2
The le pl vect.m loads the data le and plots the proles of v1 at some x stations,
the velocity vector eld and a contour plot of velocity gradient 1 /x2 . Compute
v
all terms in Eq. K.1. You will need to compute the derivatives of e.g. v1 and p. In
v
v
pl vect.m the function dphidx dy.m is used to compute 1 /x1 and 1 /x2 .
Use this function to compute all derivatives that you need. Find two (or more) x1
2
locations (vertical grid lines) where the v1 stress is large and small, respectively. One
way to nd these locations is to use the Matlab surf command. Do not pick grid
lines too far downstream, since the statistics may be poor in this region giving shaky,
oscillating proles (especially for gradients).
Assignment 1.1. Plot the stresses along vertical grid lines at these two locations using the Matlab
command plot(x,y). Please make sure that in your report the numbering on
the axis and the text in the legend is large enough; you can use the command
h1=gca;
set(h1,fontsize,[20]) %the number 20 gives the fontsize
The size of the labels and the title is similarly controlled by
xlabel(x/H,fontsize,[20])
ylabel(y/H,fontsize,[20])
title(velocity,fontsize,[20])
Assignment 1.2. Plot also all terms in Eq. K.1. To enhance readability you may omit the small
terms or use two plots per vertical grid line. Make also a zoom near the walls.
For example, for a x y plot
plot(u,y,linew,2) % linewidth=2
you may want to zoom in on y=[0 0.01] and u=[-0.1 0.4]; this is achieved
by
axis([-0.1 0.4 0 0.01])
The axis command can be used together with any plot, e.g. with surf and
quiver.
Which terms are negligible? Can you explain why they are negligible?
What about the viscous terms: where do they play an important role? Which terms
are non-zero at the wall? (you can show that on paper).
So far we have looked at the v1 -momentum equation. The database corresponds to
a two-dimensional ow. Now lets think of the forces as vectors. The normal stresses
in the x1 x2 plane represent the vector
N =
2
v 2
v1
, 2
x1
x2
(K.2)
K.2. Analysis
195
v1 v2
v v
, 1 2
x2
x1
(K.3)
This is the force per unit volume ([N/m3 ]) by which the normal stresses affect the ow.
Assignment 1.3. Plot the vector eld N to nd out some features. Zoom-in on interesting regions.
Assignment 1.4. Plot also vector elds of the shear stress, S (see Eq. K.3), the pressure gradient
and the viscous terms. Zoom up in interesting regions. Anything interesting?
2
When v2 reaches a maximum or a minimum along a grid line normal to the
wall, what happens with the vector eld? Zoom-in on interesting regions.
K.2.2 The turbulent kinetic energy equation
The exact transport equation for for the turbulent kinetic energy, k, reads
2k
(j k) =
v
+ Pk + Dk
xj
xj xj
v
i
Pk = vi vj
xj
(K.4)
vk vi vj =
+ Pij + ij + Dij ij
xk
xk xk
v
v
j v v i
Pij = vi vk
j k
xk
xk
(K.5)
The pressure-strain term, ij , and the diffusion term, Dij , need to be modelled. Here
we use the models in Eqs. 11.79, 11.47, 11.78, 11.83 and 11.84.
1. In the damping function, f (see Eq. 11.80), xn denotes the distance to the nearest
wall. If, for example, the lower wall is the closest wall to node (I, J), then
xn = (x(I, J) x(I, 1))2 + (y(I, J) y(I, 1))2
1/2
(K.6)
K.2. Analysis
196
2. ni,w denotes the unit normal vector of the wall to which the distance xn is computed. If we assume, again, that the lower wall is the closest wall to cell (I, J)
and that the lower wall is horizontal, then ni,w = (0, 1). The compute ni,w for
the general case, compute rst the vector which is parallel to the wall, si,w , and
compute then ni,w from si,w (see Eq. K.11)
3. The diffusion terms Dij and D can be modelled using the Generalized Gradient
Diffusion Hypothesis GGDH of [153]
Dij =
xm
cuk um
k vi vj
xk
(K.7)
This diffusion model can cause numerical problems, and the GGDH is then replaced by a simple eddy viscosity model
Dij =
xm
t vi vj
k xm
, t = C k 2 /
(K.8)
2
(2k/3)ij where t = c k /. Compare the eddy-viscosity stresses with two of
the Reynolds stresses from the database. Make also a zoom-in near walls.
When using the Boussinesq assumption the production of turbulent kinetic energy
P k = 2t sij sij
(K.9)
is always positive. The exact production of turbulent kinetic energy (see Eq. K.4) is
usually positive. It can however become negative.
Assignment 1.9. Compute the exact production in Eq. K.4 in the entire domain to investigate if
the production is negative anywhere. If so, explain why.
The reason why the eddy-viscosity production in Eq. K.9 must be positive is of
course that neither t nor sij sij can go negative. Another way to explain this fact is
v
that the modelled Reynolds stress, vi vj , and the strain rate tensor, i /xj are parallel.
To nd out to what degree the exact Reynolds stress and the strain rate are parallel, one
can compute the eigenvectors.
197
Assignment 1.10. Compute the eigenvalues and eigenvectors of the strain tensor, sij . The eigen
values correspond to the normal strain in the direction of the eigenvectors (see
Section 13). If the shear strains (i.e. the off-diagonal components) dominate,
you will get eigenvectors in the direction /4 /2 and if the normal strains
(i.e. the diagonal components) dominate the direction of the eigenvectors will
be along the x1 and x2 axes (explain why!). Plot the eigenvectors as a vector
eld. Our ow is 2D; thus we get two eigenvectors and two eigenvalues. Since
the two eigenvectors are perpendicular to each other it is sufcient to plot one of
them ( for example, the eigenvectors (/4, /4), (/4, /4), (/4, /4)
and (/4, /4), all represent the same principal coordinate system). Zoom in
on interesting regions.
Assignment 1.11. Compute the eigenvalues and eigenvectors of the Reynolds stresses, vi vj . The
eigenvalues correspond to the normal stresses in the direction of the eigenvectors.
Zoom in on interesting regions. In which regions are the eigenvectors of the
Reynolds stress tensor and those of the strain tensor not parallel? This should
indicate regions in which an eddy-viscosity model would perform poorly. Zoom
in on interesting regions.
1
(v1,sw + v1,se + v1,nw + v1,ne )
4
(K.10)
Lets compute v1 /x1 . In order to do that we use Gauss law over a control
volume centered at face e (dashed control volume in Fig. K.1). To compute v1 /x1
we use the x1 -component of the normal vector n = (n1 , n2 ), i.e.
v1
dV =
x1
v1 n1 dA
A
v1 n1 dA
A
1
V
(v1 n1 A)i =
i=e,n,w,w
1
{(v1 An1 )e + (v1 An1 )n + (v1 An1 )w + (v1 An1 )s }
V
198
N
ne
nw
se = (s1e , s2e )
P
ne = (n1e , n2e )
E
W
sw
se
S
Figure K.1: Control volume. The velocity v1 is stored at the corners (ne, nw, . . . ).
Coordinates x1 , x2 are given at the corners (ne, nw, ...).
K.3.1 Geometrical quantities
It is useful to rst compute the unit vectors s along the control volume. For the east
face, for example, we get
s1e =
x1,ne x1,se
de
s2e =
x2,ne x2,se
de
de =
(note that the area of the east face Ae is equal to de since z = 1). The relation
between the normal vector n, s and the unit vector in the z-direction
sn =0
s z = n,
(K.11)
199
You can do the assignment on your own or in a group of two. You will receive data
from a DNS of fully developed ow in a channel. It is recommended (but the not
A
A
required) that you use LTEX(an example of how to write in LTEXis available on the
course www page). It is available on Linux. On Windows you can use, for example, Lyx (www.lyx.org) or MikTex (www.miktex.org) which are both free to
download.
The equations that have been solved are
vi
=0
xi
(L.1)
vi
p
1
2 vi
(vi vj ) = i1
+
+
t
xj
xi
Re xj xj
The Re number based on the friction velocity and the half channel width is Re =
u h/ = 500 (h = = u = 1 so that = 1/Re ).
A 96 96 96 mesh has been used. The streamwise, wall-normal and spanwise
directions are denoted by x (x1 ), y (x2 ) and z (x3 ) respectively. The cell size in x and
z directions are x = 0.0654 and z = 0.0164. Periodic boundary conditions were
applied in the x and z direction (homogeneous directions).
At the www-page (http://www.tfd.chalmers.se/lada/comp turb model)
you nd data les with three instantaneous ow elds (statistically independent). The
data les include the instantaneous variables u (v1 ), v (v2 ), w (v3 )and p (made nondimensional by u and ). Use Matlab to analyze the data. You nd a Matlab program
at the www-page which reads the data and computes the mean velocity. The data les
are Matlab binary les. Since the data les are rather large, it is recommended that you
do all tasks using only data les 1. When everything works, then use also data les
2 and 3 averaging by use of the three les.
L.1
Task 2.1
vi = vi + vi , v = p + p
The symbol . denotes averaging in the homogeneous directions x and z. Note that
in reality . always denote time averaging. It is only in this special academic test case
where we have three homogeneous directions (x, z, t) where we can in addition to
time averaging also can use x and z averaging. Compute the six stresses of the stress
tensor, vi vj . Use the denition to compute the stresses, for example
2
v1 = (v1 v1 )
2
= v1 2 v1
2
2
2
= v1 2 v1 v1
+ v1
2
= v1 v1 2 .
Wait with analysis of the results till you have done next part.
v1
L.2
200
Task 2.2
v
= 2 p 1
x
v
v
= p 1 + p 2
y
x
v
= 2 p 2
y
Pk = v1 v2
P11
P12
11
12
22
Do the production terms and the pressure-strain term appear as you had expected? (see
the previous course MTF256)
Now analyze the uctuations in the previous subsection. Which stresses do you
think are symmetric with respect to the centerline? or anti-symmetric? Whats the
reason?
When averaging, we use only three time steps (three les). If we would use many
more time steps or, in general, if we let T when time avaraging, e.g.
1
T 2T
+T
= lim
dt
T
L.3
Task 2.3
lter width = 2x1 (this corresponds to a test lter, see Eq. 17.26. Compare
v1 and v2 with v1 and v2 .
2. Do the same thing again but with a lter width of = 4x1 (now you must
derive the expression on your own!). Discuss the differences between no lter,
= 2x1 and = 4x1 .
In LES we almost always assume that the lter width is equal to the control volume
(i.e. we use an implicit lter). Above, in Item 1 and 2 you have just carried out explicit
ltering.
Repeat Item 1, but now for a 2D lter (x1 and x3 direction); the formula for a 3D
lter is given in Eq. 17.27. Compare the results along the same lines as in Item 1 and
2.
201
cut-off
E()
I
II
III
L.4
Task 2.4
Compute the SGS stress 12 from the Smagorinsky model, which reads
ij = 2sgs sij , sgs = (Cs f )
sij =
1
2
i
v
j
v
+
xj
xi
2ij sij
s
(L.2)
f = 1 exp(x+ /26)
2
The ltered velocities, vi , are taken from Task 2.3 using the 2D lter (in x1 and x3 );
we should really have used a 3D lter, but in order to keep it simple, we use the 2D
lter. Before doing the 2D lter, look in the Lecture Notes how a 3D lter is done. The
constant Cs = 0.1.
Compare the SGS stress 12 with the resolved stress u v and compare the SGS
viscosity with the physical one. Plot them across the channel. Any thoughts?
As an alternative to the damping function, f , compute the lter length as
= min{n, }
(L.3)
where n is the distance to the nearest wall and = 0.4 (von K` rm` n constant). In this
a a
case you should set f = 1.
L.5
202
Task 2.5
Repeat the Task 2.4, but now for the WALE model by [154], which reads
i 2
v
, g = gik gkj
xj ij
1
1 2
2
2
g + gji ij gkk
sd =
ij
2 ij
3
gij =
sgs =
(L.4)
3/2
sd sd
ij ij
2
(Cm )
5/4
5/2
(ij sij ) + sd sd
s
ij ij
with Cm = 0.5.
L.6
Task 2.6
vi vi
xj xj
ij = vi vj vi vj
(L.5)
i
v
xj
(L.6)
Now, what is the relation between sgs and ? Considering the cascade process, what
did you expect?
Recall that when we do traditional Reynolds decomposition, the production term
in the equation for turbulent kinetic energy appears as a sink term in the equation for
the mean kinetic energy, see Eq. 8.29. This is the case also in LES, but now we have
vi = vi + vi = vi + vi + vi
(L.7)
resolved uctuations, vi , is made. Hence the SGS dissipation in Eq. L.6 appears as an
instantaneous production term in the equation for ksgs [70, 71, ] (can be downloaded
from www.tfd.chalmers.se/lada).
Plot (along x2 ), compare and discuss the four dissipations (see Fig. L.2)
203
ksgs
s
sg
s ij
s ij
gs
vi vj
vi
xj
x i v
j
x i
j
vi
v i xj
xj
T
1
1
v1
v1 v2
i i
v v
i
v v
v
ij xi + ij xj , cf. Eq. L.6
j
=
sgs
sgs
i i
v v
L.7
1
v
x2
v1
sgs
x2
v1
x2
Task 2.7
Above the ltered velocities were computed using the lter width = 2x1 . In
dynamic models, we often dene the test lter as twice the usual lter, i.e. = 2.
Use this denition (1D lter, i.e. = 4x1 ) to compute the dynamic Leonard stress
L12 from the denition
Lij = vi vj v i v j
(L.8)
and compare it (across the channel) with the resolved stress v1 v2 and the SGS stress
12 dened in Eq. L.5. Do you expect the magnitude of stresses to be similar?
L.8
204
Task 2.9
2
2
What is the near-wall behavior of v1 , v1 and v2 (i.e., for v1 , what is m in v1 =
m
O(x2 )). In order to estimate m, plot the quantities in log-log coordinates. Do the
quantities exhibit the near-wall behaviour that you expected?
L.9
Task 2.10
1
96 96
96
96
I=1 K=1
(L.9)
where xK and m are the spanwise locations of the two points. Take advantage of the
3
fact that the ow is periodic, but be careful when integrating the correlation above in
the x3 direction. We have 96 cells in the x3 direction. If, for example, m = 2x3 ,
and one of the points (x1 ) is at K = 1 then the other (x1 2) is at K = 95.
3
3
Plot the two-point correlation at a couple of x2 positions. When plotting two-point
correlations, it is no point showing both symmetric parts; show only half of it (cf. the
two-point correlations in Section 10.1 and Fig. M.1).
Compute and plot the integral length scale, L1 , which is dened by
L1 (x2 ) =
1
2
v1,rms
B11 (x2 , )d
(L.10)
L.10
Task 2.11
The energy spectrum of any second moment can be obtained by taking the FFT of the
corresponding two-point correlation. The energy spectrum of any second moment can
be obtained by taking the FFT of the corresponding two-point correlation. You can nd
some details on how to use Matlabs FFT in Appendix M.
If you have computed the Fourier coefcients properly, the sum of all coefcients
should give the energy. The reason is that the Fourier coefcients correspond to the
energy spectrum, and if we integrate the energy spectrum over all wave numbers we
get the total energy. When we take the FFT of Eq. L.9, for example, we get
B11 (z ) = F F T (B11 )
and summation gives
N
B11 /N
2
v1,rms =
(L.11)
see Appendix M
Plot the energy spectra at a couple of x2 locations. Conrm that Eq. L.11 is satised. When plotting two energy spectrum, it is no point showing both symmetric parts;
show only half of it (cf. the energy spectrum in Fig. M.5 b).
L.11
Task 2.12
205
(M.1)
where x3 is the separation between the two points. Here we assume that x3 is an
1
M
M
m=1
(M.2)
where m denotes summation in homogeneous directions (i.e. time plus spatial homogeneous directions).
In the following section we give a simple example how to use Matlab to Fourier
transform a signal where we know the answer. Then we show how to derive the energy
spectrum from a spatial two-point correlation. Finally, some comments are given on
how to create an energy spectrum versus frequency from an autocorrelation (i.e. from
a two-point correlation in time).
(M.3)
where L is the length of the domain and N = 16 is the number of discrete points,
see Fig. M.2. Lets use this function as input vector for the discrete Fourier transform
(DFT) using Matlab. The function u is symmetric, so we expect the Fourier coefcients
to be real. In Matlab the DFT of u is dened as (type help fft at the Matlab prompt)
N
U (k) =
un exp
n=1
1kN
2(k 1)(n 1)
N
(M.4)
206
1
L
u(x) exp(x)dx,
= 2/L
(M.5)
Note that the discrete Fourier U (k) coefcients in Eq. M.4 must be divided by N , i.e.
U (k)/N , in order to correspond to the Fourier coefcients U c (N corresponds to L
in Eq. M.5). Furthermore, it can be noted that in Eq. M.4 the period [0, 2] is used
whereas the formulation in Eq. M.5 is based on the interval [, ].
In Matlab, we generate the function u in Eq. M.3 using the commands
N=16;
n=1:1:N;
u=1+cos(2*pi*(n-1)/N);
The u function is shown in Fig. M.2. 16 nodes are used; node 1 is located at x = 0
and node 16 is located at 15L/16.
Now we take the discrete Fourier transform of u. Type
U=fft(u);
Instead of using the built-in fft command in Matlab we can program Eq. M.4
directly in Matlab as
U=zeros(1,N);
for k=1:N
for n=1:N
arg1=2*pi*(k-1)*(n-1)/N;
U(k)=U(k)+u(n)*cos(-arg1);
1.4
1.2
1
B33 (3 )
x
0.8
0.6
0.4
0.2
0
0.2
0
0.1
0.2
0.3
0.4
0.5
0.6
x3
Figure M.1: Two-point correlation, B(3 ) = v3 (x3 x3 )v3 (x3 ) , of DNS data in
x
channel ow taken from [69].
207
2
2
16
1.5
0.5
9
0
0
0.5
1.5
2.5
(M.6)
which corresponds to cos(2x/L) in Eq. M.3. It can be noted that the interval [k =
N/2 + 1, N = 9, 16] corresponds to the negative, symmetric part of the wavenumbers
in the physical formulation (cf. Eqs. M.4 and M.5). The rst Fourier coefcient corresponds as always to the mean of u, i.e. U (1)/N = u . This is easily veried from
Eq. M.4 by inserting k = 1. The remaining coefcients are zero.
In Fig. M.3, U/N is plotted versus non-dimensional wavenumber, k, and versus
wavenumber = 2(n 1)/L.
The energy, u2 , of the signal in Fig. M.2 can be computed as
u2 =
1
L
u2 (x)dx =
u2 /N = 1.5
n
(M.7)
n=1
In wavenumber space the energy is according to Parsevals formula, see Eq. E.4
equal to the integral of the square of the Fourier coefcients, i.e.
u2 =
1
L
U 2 ()d =
1
N
N
2
Un /N = 1.5
n=1
(M.8)
1.2
208
1
0.8
0.8
U/N
0.6
0.6
U/N
0.4
0.4
0.2
0.2
0
0.2
0
10
15
0
0
20
10
k
(a) Versus non-dimensional wavenumber, k.
15
20
96
1.2
1
B33 (3 )
x
0.8
0.6
0.4
0.2
0
0.2
0
0.5
1.5
x3
Figure M.4: Periodic two-point correlation, B33 (3 ) = v3 (x3 )v3 (x3 + x3 ) , of DNS
x
B33 (k) =
2(k 1)(n 1)
N
(M.9)
The simulations have been carried out with periodic boundary conditions in x3 direction (and x1 ), and hence B33 (3 ) is symmetric, see Fig. M.4. Thus, it is sufcient to
x
use the cosine part of Eq. M.9, i.e.
N
B33 (k) =
2(k 1)(n 1)
N
(M.10)
In Fig. M.5a the Fourier coefcients B33 3 are presented versus wavenumber 3 =
2(n 1)/x3,max , where x3,max 1.55, see Fig. M.4. Figure M.5b shows the same
energy spectra in log-log scale (only half of the spectrum is included), which is the
209
0.25
1
10
B33 (k)/N
0.2
0.15
10
0.1
10
0.05
0
0
100
200
300
400
10
10
10
2
Figure M.5: The energy spectrum of v3 versus wavenumber, 3 . Dashed line in b)
show 5/3 slope.
common way to present energy spectra. The dashed line shows the 5/3 slope which
indicates that the energy spectra from the DNS follows the Kolmogorov 5/3 decay.
As usual, the Fourier coefcient for the rst non-dimensional wavenumber, i.e.
B33 =
N
n=1
B33 (n)
1
B33 (1)
N
(M.11)
compare with Eq. M.10. Note that this is almost the same expression as that for the
integral length scale which reads (see Eq. 10.5)
Lint (3 ) =
x
1
2
v3
B33
2
v3
(M.12)
Hence the integral length scale is related to the rst Fourier mode as
Lint =
B33 (1)
(M.13)
2
v3
2
2
The two-point correlation for zero separation is equal to v3 , i.e. B33 (0) = v3 =
2
1.51. Another way to obtain v3 is to integrate the energy spectrum in Fig. M.5, i.e.
2
v3 =
1
x3,max
B33 (3 )d3 =
N
(M.14)
n=1
When computing the energy spectra of the v3 velocity, say, versus frequency, the time
series of v3 (t) is commonly Fourier transformed and the energy spectrum is obtained
by plotting the square of the Fourier coefcients versus frequency, f . We can also split
the time signal into a number subsets, Fourier transform each subset and then average.
In Matlab, the command pwelch is a convenient command which does all this.
210
the same way as in Section M.3. The only difference is that B33 ( ) is a function of
N. Transformation of a tensor
211
Transformation of a tensor
The rotation of a vector from the xi coordinate system to xi reads (see, e.g., Chapter
1 in [2])
ui = bij uj
(N.1)
where bij denotes the cosine between the axis
bij = cos (xi , xj )
(N.2)
(N.3)
The relations bik bjk = bki bkj = ij are fullled as they should.
For a second-order tensor, the transformation reads
uij = bik bjm ukm
(N.4)
b22 = 1/ 2
(N.5)
(N.6a)
(N.6b)
(N.6c)
(N.6d)
x2
x2
x1
x1
Figure N.1: Transformation between the coordinate systems (x1 , x2 ) and (x1 , x2 ).
212
(N.8)
where
c11 = cos ,
(N.9)
c22 = cos
see Fig. N.1. It can be seen that the relation cji = bij is satised as it should. The
eigenvectors for Eq. N.7 are any two orthogonal vectors with angles /4, 3/4.
Let us choose /4 and 3/4 for which the transformation in Eq. N.8 reads ( = /4)
s11 = c11 c11 s11 + c12 c11 s21 + c11 c12 s12 + c12 c12 s22
1
= (s11 + s21 + s12 + s22 )
2
s12 = c11 c21 s11 + c12 c21 s21 + c11 c22 s12 + c12 c22 s22
1
= (s11 s21 + s12 + s22 )
2
s21 = c21 c11 s11 + c22 c11 s21 + c21 c12 s12 + c22 c12 s22
1
= (s11 + s21 s12 + s22 )
2
s22 = c21 c21 s11 + c22 c21 s21 + c21 c22 s12 + c22 c22 s22
1
= (s11 s21 s12 + s22 )
2
(N.10a)
(N.10b)
(N.10c)
(N.10d)
s12 = 0,
s21 = 0,
s22 = s21
(N.11)
Since the diagonal elements are zero it conrms that the coordinate system x1 x2
with = /4 is indeed a principal coordinate system. The eigenvalues, (k) , of sij
correspond to the diagonal elements in Eq. N.11, i.e.
(1) s11 = s12 =
1 v1
,
2 x2
1 v1
2 x2
(N.12)
213
A11 =
A12
A21
A22
(N.13)
Insert Eq. N.9 with = /4 and replace Aij by the velocity gradient
v1
v1
1 v1
=
=
,
x1
x2
2 x2
v2
v2
1 v1
=
=
,
x1
x2
2 x2
(N.14)
214
Week 1
1. How is the buoyancy term, gi , re-written in incompressible ow?
2. Which terms in the vi vj equation need to be modelled? Explain the physical
meaning of the different terms in the vi vj equation. Discuss and show how the
dissipation term, ij , is modelled.
3. How are the Reynolds stress, vi vj , and the turbulent heat ux, vi , modelled in
the Boussinesq approach?
4. How is the turbulent diffusion ux for k modeled in the k model (dk ) and
i,t
in the RSM (dk ), respectively? Explain the advantage of the GGDH model
i,t,G
over the usual Boussinesq model. What is the nal expressions of the modelled
diffusion terms in the k, and vi vj equation.
5. How is the production term in the k equation modelled (Boussinesq)?
6. Given the transport equation for the temperature, , derive the transport equation
for . Derive the transport equation for vi (Eq. 11.17). Discuss the physical
meaning of the different terms. Which terms need to be modelled?
7. What is the expression for the total heat ux which appear in the equation?
8. Derive the transport equation for (start from the k equation)
9. Use physical reasoning to derive a model for the diagonal components of the
pressure-strain term (slow part).
215
Week 2
2
1. The slow pressure-strain model reads ij,1 = c1 k vi vj 3 ij k . The
v v
i
anisotropy tensor is dened as bij = 2kj 1 ij . Show that for decaying grid
3
turbulence, the model for the slow pressure-strain model indeed acts as to make
the turbulence more isotropic if c1 > 1.
2. Derive the exact equation for pressure uctuation, Eq. 11.55. For a Poisson
equation
2
=f
xj xj
there exists an exact analytical solution
(x) =
1
4
(O.1)
Using Eq. O.1, give the exact analytical solution for the pressure-strain term.
What are the slow and rapid terms? Which two assumptions are used to
simplify this equation? Give the simplied analytical solution for p .
3. Using the exact solution of the p equation, the rapid pressure-strain term is written
k (x) 1
v
x 2
vi (x) vj (x)
+
xj
xi
v (y) dy3
k
v
= Mijk
yk |y x|
x
Derive the expression for Mijk (i.e. Eq. 11.65) assuming homogeneity.
4. The pressure-strain term is split into two terms, which ones? How are they modelled?
5. The rapid pressure-strain term is expressed in aijk (Eq. 11.65). aijk is symmetric with respect to i, k and j, . If we assume that aijk is linear in the
Reynolds stresses, give the most general expression for aijk (i.e. Eq. 11.68)
which satises the symmetry conditions. Equation 11.68 must satisfy continuity
requirement (i.e. aiik = 0) and the requirement given by Greens formula (i.e.
aiji = 2vj v ). Use these two requirements to derive the four equations for the
ve coefcients (i.e. Eq. 11.73)
6. Derive the algebraic stress model (ASM). What main assumption is made?
7. Describe the physical effect of the pressure-strain term in the near-wall region.
What sign must hence 22,1w have?
2
8. The modelled slow and rapid pressure strain term read ij,1 = c1 k vi vj 3 ij k
2
and ij,2 = c2 Pij 3 ij P k , respectively. Give the expression for the production terms, modelled pressure-strain terms and modelled dissipation terms
for a simple shear ow. In some stress equations there is either no production
terms or no dissipation term. How come? Which is their main source term (or
sink term)?
216
Week 3
1. Consider buoyancy-dominated ow with x3 vertically upwards. The production
Gij = gi vj gj vi ,
Pi = vi vk
xk
respectively. Show that the Reynolds stress model dampens and increases the
vertical uctuation in stable and unstable stratication, respectively. Show also
that k in the k model is affected in the same way.
2. Consider streamline curvature for a streamline formed as a circular arc (convex
curvature). Show that the turbulence is dampened if V /r > 0 and that it
is enhanced if the sign of V /r is negative. Now consider a boundary layer
where the streamlines are curved away from the wall (concave curvature). Show
that this streamline curvature gives an enhanced positive feedback between the
production terms. Show that the effect of streamline curvature in the k model
is much smaller.
3. Consider stagnation ow. Show that in the Reynolds stress model, there is only a
small production of turbulence whereas eddy-viscosity models (such as the k
model) give a large production of turbulence.
4. What is a realizability constraint? There are two main realizability constraints
on the normal and the shear stresses: which ones? Show that the Boussinesq
assumption may give negative normal stresses. In which coordinate system is
the risk largest for negative normal stresses? Derive an expression how to avoid
negative normal stresses by reducing the turbulent viscosity.
5. What is the two-component limit? What requirement does it put on the pressurestrain models? Show that the standard IP model and the Rotta model do not
satisfy this requirement.
6. What is a non-linear eddy-viscosity model? When formulating a non-linear
model, the anisotropy tensor aij = 2t sij /k is often used. Show the three
rst terms (S, S 2 , 2 , S) in the non-linear model in the lecture notes. Show
that each term is traceless and symmetric (as aij ).
7. What is the modeling idea of the V2F model? Which equations are solved?
2
8. The transport equation for v2 reads (the turbulent diffusion terms are modelled)
1 v2
v 2
v 2
v2
v 2
( + t ) 2 2v2 p /x2 22
+
=
x1
x2
x2
x2
22
22
1
2f
f =
x2
k
T
2
2
2
v2
k
3
k
,
k 3/2
217
Show how the magnitude of the right side and L affect f . How does f enter into
2
the v2 equation? What it the physical meaning of f ? Show that far from the
2
2
walls, the V2F model (i.e. the f and the v2 equation) returns to the v2 equation
in the Reynolds stress model.
218
Week 4
1. Consider the boundary condition for the f equation. Very close to the wall, how
2
is the v2 equation simplied? Use this equation to derive a wall boundary for f .
2. What does the acronym SST mean? The SST model is a combination of the
k and the k model. In which region is each model being used and why?
How is expressed in k and ?
3. Derive a transport equation for from the k and transport equations. In the
SST model, a blending function F1 is used; what does this function do? In which
region is each model being used and why?
219
Week 5
1. Describe the shear stress limiter. Show that the eddy-viscosity assumption gives
too high shear stress in APG since P k / 1.
2. Show the difference between volume averaging (ltering) in LES and timeaveraging in RANS.
3. Consider the spatial derivative of the pressure in the ltered Navier-Stokes: show
that the derivative can be moved outside the ltering integral (it gives an additional second-order term). The ltered non-linear term has the form
vi vj
xj
Show that it can be re-written as
i vj
v
xj
giving an additional term
ij
(vi vj ) +
(i vj ) =
v
xj
xj
xj
show that v I = vI .
5. Consider the energy spectrum. Show the three different regions (the large energycontaining scales, the 5/3 range and the dissipating scales). Where should the
cut-off be located? What does cut-off mean? Show where the SGS scales, grid
(i.e resolved) scales and the cut-off, c are located in the spectrum.
6. Show how a sinus wave sin(c x) corresponding to cut-off is represented on a
grid with two and four nodes, respectively. How is c is related to the grid size
x for these cases?
7. Show the Smagorinsky model.
220
Week 6
1. Discuss the energy path in connection to the source and sink terms in the k, K
and ksgs computed from the energy spectrum?
and the ksgs equations. How are k
2. What is a test lter? Grid and test lter Naiver-Stokes equation and derive the
relation
(O.2)
vi vj v i v j + ij = Tij
Draw an energy spectrum and show which wavenumber range k, ksgs , ksgs,test
cover.
3. Formulate the Smagorinsky model for the grid lter SGS stress, ij , and the test
lter SGS stress, Tij . Use Eq. O.2 and derive the relation
1
Lij ij Lkk = 2C
3
ss
| s | s ij 2 ||ij
This equation is a tensor equation for C. Use this relation and derive the nal
expression for the dynamic coefcient, C, Eq. 17.33.
4. Show that when a rst-order upwind schemes is used for the convection term,
an additional diffusion term and dissipation terms appear because of a numerical
SGS viscosity.
5. What is a scale-similarity model? How are the cross terms, Cij , modelled? What
is the exact form of the the cross term, Cij ?
221
Week 7
1. Show that the Bardina scale-similarity model is not Galilean invariant. How is
the the model modied to make it Galilean invariant?
2. What are the ve main differences between a RANS nite volume and a LES
code? What do you need to consider in LES when you want to compute timeaveraged quantities?
3. How ne does the mesh need to be in the wall region be?
2
4. What is DES? The length scale in the RANS S-A model reads t ; how is it
d
computed in DES? How is the length scale computed in a k two-equation
DES model?
222
Week 8
1. The modied (reduced) length scale in two-equation DES models can be introduced in different ways in the equations. Discuss two different ways. What is
the effect on the modelled, turbulent quantities?
2. Describe hybrid LES-RANS based on a one-equation model. What is the difference between DES and hybrid LES-RANS? Standard hybrid LES-RANS is
found to give too large velocity in the outer region. What is the reason? How
can it be xed?
3. Describe URANS. How is the instantaneous velocity decomposed? What turbulence models are used? What is scale separation?
4. Describe the SAS model. How is the von K rm n length scale dened? An
a a
additional source term is introduced in the equation: what is the form of this
term? What is the object of this term? When is it large and small, respectively?
P. Greens formulas
223
Greens formulas
In this appendix we will derive Greens three formulas from Gauss divergence law.
In the last section we will derive the analytical solution to the Poisson equation. The
derivations below are partly taken from [155].
Fi
dV =
xi
Fi ni dS
(P.1)
where S is the bounding surface of the volume, V , and ni is the normal vector of S
xi
xi
ni dS
xi
(P.2)
2
+
xi xi
xi xi
(P.3)
dV =
S
xi
xi
2
dV +
xi xi
dV =
xi xi
dV =
xi xi
ni dS
xi
(P.4)
ni dS
xi
(P.5)
2
dV +
xi xi
2
2
xi xi
xi xi
dV =
xi
xi
ni dS
(P.6)
1
|r rP |
(P.7)
As usual we are considering a volume V with bounding surface S and normal vector
ni . Since function (r) is singular for r = rP , consider a small sphere in V , see
224
ni
S
n
i
r
S
rP
x2
x1
Figure P.1: Greens third formula. A volume V with bounding surface S with normal
vector ni . In V there is a small sphere S located at rP with radius and normal vector
n .
i
Fig. P.1. In Eq. P.6 we need the rst and the second derivative of . The rst derivative
of 1/ri is computed as
1
r
xi
r/xi
ri
= 3
2
r
r
(P.8)
since the derivative of a distance X is a vector along the increment of the distance, i.e.
X/xi = Xi /X where X = |Xi |. The second derivative is obtained as
2
xi xi
1
r
xi
1
r3
= 3
ri
r3
=
ri
+
r
ri
xi
3ri
r4
r 3ri
xi r4
3
r2 3
= 3 +
r
r r4
1
r3
(P.9)
=0
To get the right side on the second line we used the fact that ri ri = r2 . Now we replace
ri = r by r rP = ri rP,i in Eqs. P.8 and P.9 which gives
xi
2
xi xi
1
|r rP |
1
|r rP |
ri rP,i
|r rP |3
(P.10)
=0
P
for ri = ri , i.e. for V excluding the sphere S , see Fig. P.1. Apply Greens second
formula for this volume which has the bounding surfaces S and S with normal vectors
ni (outwards) and n (inwards), respectively. We get
i
V S
2
1
dV =
|r rP | xi xi
+
S
P
ri ri
|r rP |3
|r rP | xi
P
ri ri
3
|r rP |
|r rP | xi
ni dS
(n )dS
i
(P.11)
225
where the volume integral is taken over the volume V but excluding the sphere S , i.e.
V S . Note the minus sign in front of the normal vector in the S integral; this is
because the normal vector must point out of the volume V S , i.e. into the sphere,
S . In the sphere the normal vector, n , is the direction from point rP , i.e.
i
n =
i
r rP
ri rP,i
=
|r rP |
|ri rP,i |
(P.12)
where we have normalized the vector ri rP,i in order to make its length equal to one.
The length of the vector ri rP,i is the radius of sphere S , i.e.
|r rP | =
(P.13)
The surface area, dS, for sphere S can be expressed in spherical coordinates as
dS = 2 = 2 sin dd
(P.14)
where is the solid angle. Inserting Eqs. P.12, P.13 and P.14 in the last integral in
Eq. P.11 gives
IS =
S
ri rP,i
+
2
2
xi
2 d =
S
+ (ri rP,i )
xi
(P.15)
To re-write this integral we will use the mean value theorem for integrals. In one
dimension this theorem simply states that for the integral of a function, g(x), over the
interval [a, b], there exists (at least) on point for which the the relation
b
a
g(x)dx = (a b)g(xQ )
(P.16)
holds, where xQ denotes a point on [a, b]. Applying this theorem to the integral in
Eq. P.15 gives
IS = (rQ )
S
d + rQ,i rP,i )
(rQ )
xi
(P.17)
since
(P.18)
1
ri ri
1
1
ni dS +
ni dS
+
4 S |r rP |3
4 S |r rP | xi
(rP ) =
1
4
(P.19)
(P.20)
Hence it is seen that the volume element dV goes to zero faster than the singularity
1/|r rP |.
226
(P.21)
where we assume that goes to zero at innity and that the right side is limited.
Greens third formula, Eq. P.19, gives
1
4
f (r)
dV
|r rP |
V
P
1
1
ri ri
+
ni dS +
4 S |r rP |3
4
(rP ) =
1
ni dS
|r rP | xi
(P.22)
We choose the volume as a large sphere with radius R. Using Eqs. P.12, P.13 and P.14,
the rst surface integral can be written as
1
4
P
1
ri ri
ni dS =
|r rP |3
4R2
ni ni dS =
S
1
4
(P.23)
1
1
ni dS =
|r rP | xi
4R
S
1
2
1
dV =
=
4R S xi xi
4R
ni dS
xi
(P.24)
f dV
V
This integral also goes to zero for large R since we have assumed that f is limited.
Hence the nal form of Eq. P.22 reads
(rP ) =
1
4
f (r)
dV
|r rP |
(P.25)
Q. References
227
References
Q. References
228
[18] B.E. Launder, G.J. Reece, and W. Rodi. Progress in the development of a
Reynolds-stress turbulence closure. Journal of Fluid Mechanics, 68(3):537
566, 1975.
[19] D. Naot, A. Shavit, and M. Wolfshtein. Interactions between components of the
turbulent velocity correlation tensor. Israel J. Techn., 8:259, 1970.
[20] B. E. Launder. An introduction to single-point closure methodology. In T.B.
Gatski, M.Y. Hussaini, and J.L. Lumpley, editors, Simulation and Modeling of
Turbulent Flows, New York, 1996. Oxford University Press.
[21] M.M. Gibson and B.E. Launder. Ground effects on pressure uctuations in the
atmospheric boundary layer. Journal of Fluid Mechanics, 86:491511, 1978.
[22] S.B. Pope. A more general effective-viscosity hypothesis. Journal of Fluid
Mechanics, 472:331340, 1975.
[23] P. Bradshaw. Effects of streamline curvature on turbulent ow. Agardograph
progress no. 169, AGARD, 1973.
[24] W. Rodi and G. Scheuerer. Calculation of curved shear layers with two-equation
turbulence models. Physics of Fluids, 26:14221435, 1983.
[25] B.E. Thompson and J.H. Whitelaw. Characteristics of a trailing-edge ow with
turbulent boundary-layer separation. Journal of Fluid Mechanics, 157:305326,
1985.
[26] S. Wallin and A.V. Johansson. Modelling streamline curvature effects in explicit
algebraic reynolds stress turbulence models. International Journal of Heat and
Fluid Flow, 23(5):721730, 2002.
[27] L. Davidson. Prediction of the ow around an airfoil using a Reynolds stress
transport model. Journal of Fluids Engineering, 117:5057, 1995.
[28] P.A. Durbin. On the k 3 stagnation point anomaly. International Journal of
Heat and Fluid Flow, 17:8990, 1996.
[29] J. Hult. Cartesiska tensorer: Ett praktiskt konspr k. Almqvist & Wiksell,
a
1972.
[30] J.L. Lumley. Computational modeling of turbulent ows. Advances in Applied
Mechanics, 18:123176, 1978.
[31] T.J. Craft and B.E. Launder. A Reynolds stress closure designated for complex geometries. International Journal of Heat and Fluid Flow, 17(3):245254,
1996.
[32] T.J. Craft, B.E. Launder, and K. Suga. Prediction of turbulent transitional phenomena with a nonlinear eddy-viscosity model. International Journal of Heat
and Fluid Flow, 18:1528, 1997.
[33] T.B. Gatski and C.G. Speziale. On explicit algebraic stress models for complex
turbulent ows. Journal of Fluid Mechanics, 154:5978, 1993.
[34] T.-H. Shih, J. Zhu, and J.L. Lumley. A new Reynolds stress algebraic equation
model. Comput. Methods Appl. Engng., 125:287302, 1995.
Q. References
229
[35] S. Wallin and A.V. Johansson. A new explicit algebraic Reynolds stress model
for incompressible and compressible turbulent ows. Journal of Fluid Mechanics, 403:89132, 2000.
[36] P.A. Durbin. Near-wall turbulence closure modeling without damping functions.
Theoretical and Computational Fluid Dynamics, 3:113, 1991.
[37] P.A. Durbin. Application of a near-wall turbulence model to boundary layers
and heat transfer. International Journal of Heat and Fluid Flow, 14(4):316
323, 1993.
[38] A. Sveningsson. Analysis of the performance of different v 2 f turbulence
models in a stator vane passage ow. Thesis for licentiate of engineering, Dept.
of Thermo and Fluid Dynamics, Chalmers University of Technology, G teborg,
o
Sweden, 2003.
[39] F-S Lien and G. Kalitzin. Computations of transonic ow with the v2f turbulence model. International Journal of Heat and Fluid Flow, 22(1):5361, 2001.
[40] A. Sveningsson and L. Davidson. Assessment of realizability constraints in v 2
f turbulence models. International Journal of Heat and Fluid Flow, 25(5):785
794, 2004.
[41] A. Sveningsson and L. Davidson. Computations of ow eld and heat transfer in
a stator vane passage using the v 2 f turbulence model. Paper GT2004-53586,
ASME TURBO EXPO 2004, Vienna, 2004.
[42] A. Sveningsson and L. Davidson. Computations of ow eld and heat transfer
in a stator vane passage using the v 2 f turbulence model. Journal of Turbomachinery, 127(3):627634, 2005.
[43] A. Sveningsson. Turbulence Transport Modelling in Gas Turbine Related Applications. PhD thesis, Div. of Fluid Dynamics, Dept. of Applied Mechanics,
Chalmers University of Technology, G teborg, Sweden, 2006.
o
[44] F.R. Menter. Two-equation eddy-viscosity turbulence models for engineering
applications. AIAA Journal, 32:15981605, 1994.
[45] D.C. Wilcox. Reassessment of the scale-determining equation. AIAA Journal,
26(11):12991310, 1988.
[46] F.R. Menter. Inuence of freestream values on k turbulence model predictions. AIAA Journal, 30(6):16571659, 1992.
[47] J. Bredberg. On two-equation eddy-viscosity models. Report 01/8, can be
downloaded from www.tfd.chalmers.se/lada/allpaper.html, Dept. of Thermo
and Fluid Dynamics, Chalmers University of Technology, 2001.
[48] P. Bradshaw, D.H. Ferriss, and N.P. Atwell. Calculation of boundary-layer development using the turbulent energy equation. Journal of Fluid Mechanics,
28:593616, 1967.
[49] F.R. Menter, M. Kuntz, and R. Langtry. Ten years of industrial experience of the
SST turbulence model. In K. Hanjali , Y. Nagano, and M.J. Tummers, editors,
c
Turbulence Heat and Mass Transfer 4, pages 624632, New York, Wallingford
(UK), 2003. begell house, inc.
Q. References
230
Q. References
231
[66] L. Davidson and S.-H. Peng. Hybrid LES-RANS: A one-equation SGS model
combined with a k model for predicting recirculating ows. International
Journal for Numerical Methods in Fluids, 43:10031018, 2003.
[67] S.V. Patankar. Numerical Heat Transfer and Fluid Flow. McGraw-Hill, New
York, 1980.
[68] H.K. Versteegh and W. Malalasekera. An Introduction to Computational Fluid
Dynamics - The Finite Volume Method. Longman Scientic & Technical, Harlow, England, 1995.
[69] L. Davidson and M. Billson. Hybrid LES/RANS using synthesized turbulence
for forcing at the interface. International Journal of Heat and Fluid Flow,
27(6):10281042, 2006.
[70] L. Davidson. Large eddy simulation: A dynamic one-equation subgrid model
for three-dimensional recirculating ow. In 11th Int. Symp. on Turbulent Shear
Flow, volume 3, pages 26.126.6, Grenoble, 1997.
[71] L. Davidson. Large eddy simulations: A note on derivation of the equations for
the subgrid turbulent kinetic energies. Technical Report 97/11, Dept. of Thermo
and Fluid Dynamics, Chalmers University of Technology, Gothenburg, 1997.
[72] A. Sohankar, L. Davidson, and C. Norberg. Large eddy simulation of ow past
a square cylinder: Comparison of different subgrid scale models. Journal of
Fluids Engineering, 122(1):3947, 2000.
[73] A. Sohankar, L. Davidson, and C. Norberg. Erratum. Journal of Fluids Engineering, 122(3):643, 2000.
[74] S. Krajnovi and L. Davidson. A mixed one-equation subgrid model for largec
eddy simulation. International Journal of Heat and Fluid Flow, 23(4):413425,
2002.
[75] S. Krajnovi and L. Davidson. Large eddy simulation of the ow around a bluff
c
body. AIAA Journal, 40(5):927936, 2002.
[76] S. Krajnovi . Large Eddy Simulations for Computing the Flow Around Vehic
cles. PhD thesis, Dept. of Thermo and Fluid Dynamics, Chalmers University of
Technology, G teborg, Sweden, 2002.
o
[77] J. Pallares and L. Davidson. Large-eddy simulations of turbulent ow in a rotating square duct. Physics of Fluids, 12(11):28782894, 2000.
[78] J. Pallares and L. Davidson. Large-eddy simulations of turbulent heat transfer in
stationary and rotating square duct. Physics of Fluids, 14(8):28042816, 2002.
[79] S. Krajnovi and L. Davidson. Numerical study of the ow around the busc
shaped body. Journal of Fluids Engineering, 125:500509, 2003.
[80] S. Krajnovi and L. Davidson. Flow around a simplied car. part I: Large eddy
c
simulations. Journal of Fluids Engineering, 127(5):907918, 2005.
[81] S. Krajnovi and L. Davidson. Flow around a simplied car. part II: Understandc
ing the ow. Journal of Fluids Engineering, 127(5):919928, 2005.
Q. References
232
Q. References
233
[96] S.R. Robinson. Coherent motions in the turbulent boundary layer. Annual Review of Fluid Mechanics, 23:601639, 1991.
[97] L. Davidson. Large eddy simulations: how to evaluate resolution. International
Journal of Heat and Fluid Flow, 30(5):10161025, 2009.
[98] L. Davidson. How to estimate the resolution of an les of recirculating ow (to
appear). In ERCOFTAC, Quality and Reliability of Large-Eddy Simulations II.
Springer, 2010.
[99] L. Davidson. Inlet boundary conditions for embedded LES. In First CEAS
European Air and Space Conference, 10-13 September, Berlin, 2007.
[100] L. Davidson, D. Cokljat, J. Fr hlich, M.A. Leschziner, C. Mellen, and W. Rodi,
o
editors. LESFOIL: Large Eddy Simulation of Flow Around a High Lift Airfoil,
volume 83 of Notes on Numerical Fluid Mechanics. Springer Verlag, 2003.
[101] C.P. Mellen, J. Fr hlich, and W. Rodi. Lessons from LESFOIL project on large
o
eddy simulation of ow around an airfoil. AIAA Journal, 41(4):573581, 2003.
[102] S. Johansson, L. Davidson, and E. Olsson. Numerical simulation of vortex shedding past triangular cylinders at high reynolds numbers using a k turbulence
model. International Journal for Numerical Methods in Fluids, 16:859878,
1993.
[103] S. Perzon and L. Davidson. On CFD and transient ow in vehicle aerodynamics.
In SAE Technical Paper 2000-01-0873, Detroit, 2000.
[104] S. Perzon and L. Davidson. On transient modeling of the ow around vehicles
using the Reynolds equation. In J.-H. Wu, Z.-J. Zhu, F.-P. Jia, X.-B. Wen, and
W. Hu, editors, ACFD 2000, Bejing, 2000.
[105] T.H. Shih, J. Zhu, and J. L. Lumley. A realizable Reynolds stress algebraic
equation model. Technical memo 105993, NASA, 1993.
[106] P. Spalart and S. Allmaras. A one equation turbulence model for aerodynamic
ows. AIAA paper 92-0439, Reno, NV, 1992.
[107] P.R. Spalart, W.-H. Jou, M. Strelets, and S.R. Allmaras. Comments on the
feasability of LES for wings and on a hybrid RANS/LES approach. In C. Liu and
Z. Liu, editors, Advances in LES/DNS, First Int. conf. on DNS/LES, Louisiana
Tech University, 1997. Greyden Press.
[108] F.R. Menter, M. Kuntz, and R. Bender. A scale-adaptive simulation model for
turbulent ow prediction. AIAA paper 20030767, Reno, NV, 2003.
[109] A.K. Travin, M. Shur, , M. Strelets, and P.R. Spalart. Physical and numerical upgrades in the detached-eddy simulations of complex turbulent ows. In
R.Friederich and W.Rodi, editors, Fluid Mechanics and its Applications. Advances in LES of Complex Flows, volume 65 of Euromech Colloquium 412.
Fluid Mechanics and its Applications. Advances in LES of Complex Flows,
pages 239254. Academic Publishers, Dordrecht, 2002.
Q. References
234
[110] C. De Langhe, B. Merci, K. Lodeer, and E. Dick. Hybrid RANS-LES modelling with the renormalization group. In K. Hanjali , Y. Nagano, and M.J.
c
Tummers, editors, Turbulence Heat and Mass Transfer 4, pages 697704, New
York, Wallingford (UK), 2003. begell house, inc.
[111] J.C. Kok, H.S. Dol, B. Oskam, and H. van der Ven. Extra-large eddy simulation
of massively separated ows. AIAA paper 2004-264, Reno, NV, 2004.
[112] J. Yan, C. Mocket, and F. Thiele. Investigation of alternative length scale substitutions in detached-eddy simulation. Flow, Turbulence and Combustion, 74:85
102, 2005.
[113] S.S. Girimaji. Partially-Averaged Navier-Stokes model for turbulence: A
Reynolds-averaged Navier-Stokes to direct numerical simulation bridging
method. Journal of Fluids Engineering, 73(2):413421, 2006.
[114] J. Ma, S.-H. Peng, L. Davidson, and F. Wang. A low Reynolds number partiallyaveraged Navier-Stokes model for turbulence. In 8th International ERCOFTAC
Symposium on Engineering Turbulence, Modelling and Measurements, Marseille, France, 9-11 June, 2010.
[115] R. Schiestel and A. Dejoan. Towards a new partially integrated transport model
for coarse grid and unsteady turbulent ow simulations. Theoretical and Computational Fluid Dynamics, 18:443468, 2005.
[116] K. Abe, T. Kondoh, and Y. Nagano. A new turbulence model for predicting
uid ow and heat transfer in separating and reattaching ows - 1. Flow eld
calculations. Int. J. Heat Mass Transfer, 37:139151, 1994.
[117] P. Batten, U. Goldberg, and S. Chakravarthy. Interfacing statistical turbulence
closures with large-eddy simulation. AIAA Journal, 42(3):485492, 2004.
[118] U. Piomelli, E. Balaras, H. Pasinato, K.D. Squire, and P.R. Spalart. The innerouter layer interface in large-eddy simulations with wall-layer models. International Journal of Heat and Fluid Flow, 24:538550, 2003.
[119] L. Davidson and S. Dahlstr m. Hybrid LES-RANS: An approach to make LES
o
applicable at high Reynolds number. International Journal of Computational
Fluid Dynamics, 19(6):415427, 2005.
[120] J. Larsson, F.S. Lien, and E. Yee. The articial buffer layer and the effects of
forcing in hybrid LES/RANS. International Journal of Heat and Fluid Flow,
28(6):14431459, 2007.
[121] F.R. Menter and M. Kuntz. Adaption of eddy-viscosity turbulence models to unsteady separated ows behind vehicles. In Rose McCallen, Fred Browand, and
James Ross, editors, The Aerodynamics of Heavy Vehicles: Trucks, Buses, and
Trains, volume 19 of Lecture Notes in Applied and Computational Mechanics.
Springer Verlag, 2004.
[122] M. Strelets. Detached eddy simulation of massively separated ows. AIAA
paper 20010879, Reno, NV, 2001.
Q. References
235
[123] X. Xiao, J.R. Edwards, and H.A. Hassan. Inow boundary conditions for
LES/RANS simulations with applications to shock wave boundary layer interactions. AIAA paper 20030079, Reno, NV, 2003.
[124] F. Hamba. An approach to hybrid RANS/LES calculation of channel ow. In
W. Rodi and N. Fueyo, editors, Engineering Turbulence Modelling and Experiments 5, pages 297305. Elsevier, 2003.
[125] P. Tucker and L. Davidson. Zonal k-l based large eddy simulation. Computers
& Fluids, 33(2):267287, 2004.
[126] P. Tucker. Differential equation based length scales to improve DES and RANS
simulations. AIAA paper 2003-3968, 16th AIAA CFD Conference, 2003.
[127] L. Temmerman, M. Had iadbic, M.A. Leschziner, and K. Hanjali . A hybrid
z
c
two-layer URANS-LES approach for large eddy simulation at high Reynolds
numbers. International Journal of Heat and Fluid Flow, 26:173190, 2005.
[128] A. Dejoan and R. Schiestel. Large-eddy simulation of non-equilibrium pulsed
turbulent ow using transport equations subgrid model. In E. Lindborg, A. Johansson, J. Eaton, J. Humphrey, N. Kasagi, M. Leschziner, and M. Sommerfeld,
editors, The Second International Symp. on Turbulence and Shear Flow Phenomena, volume 2, pages 341346, Stockholm, 2001.
[129] S. Kenjere and K. Hanjali . LES, T-RANS and hybrid simulations of thermal
s
c
convection at high RA numbers. In W. Rodi and M. Mulas, editors, Engineering
Turbulence Modelling and Measurements 6, pages 369378. Elsevier, 2005.
[130] P.R. Spalart. Strategies for turbulence modelling and simulations. International
Journal of Heat and Fluid Flow, 21:252263, 2000.
[131] N.V. Nikitin, F. Nicoud, B. Wasistho, K.D. Squires, and P. Spalart. An approach
to wall modeling in large-eddy simulations. Physics of Fluids, 12(7):16291632,
2000.
[132] L. Temmerman, M.A. Leschziner, and K. Hanjali . A-priori studies of near-wall
c
RANS model within a hybrid LES/RANS scheme. In W. Rodi and N. Fueyo,
editors, Engineering Turbulence Modelling and Experiments 5, pages 317326.
Elsevier, 2002.
[133] F.R. Menter and Y. Egorov. Revisiting the turbulent length scale equation. In IUTAM Symposium: One Hundred Years of Boundary Layer Research, G ttingen,
o
2004.
[134] F.R. Menter, M. Kuntz, and R. Bender. A scale-adaptive simulation model for
turbulent ow predictions. AIAA-2003-0767, Reno, 2003.
[135] J.C. Rotta. Turbulente Str mungen. Teubner Verlag, Stuttgart, 1972.
o
[136] F.R. Menter and Y. Egorov. A scale-adaptive simulation model using twoequation models. AIAA paper 20051095, Reno, NV, 2005.
[137] L. Davidson. Evaluation of the SST-SAS model: Channel ow, asymmetric
diffuser and axi-symmetric hill. In ECCOMAS CFD 2006, September 5-8, 2006,
Egmond aan Zee, The Netherlands, 2006.
Q. References
236
[138] L. Davidson. Hybrid LES-RANS: Inlet boundary conditions for ows including
recirculation. In 5th International Symposium on Turbulence and Shear Flow
Phenomena, volume 2, pages 689694, 27-29 August, Munich, Germany, 2007.
[139] N. Jarrin, S. Benhamadouche, D. Laurence, and R. Prosser. A synthetic-eddymethod for generating inow conditions for large-eddy simulations. International Journal of Heat and Fluid Flow, 27(4):585593, 2006.
[140] M. Billson. Computational Techniques for Turbulence Generated Noise. PhD
thesis, Dept. of Thermo and Fluid Dynamics, Chalmers University of Technology, G teborg, Sweden, 2004.
o
[141] M. Billson, L.-E. Eriksson, and L. Davidson. Jet noise prediction using stochastic turbulence modeling. AIAA paper 2003-3282, 9th AIAA/CEAS Aeroacoustics Conference, 2003.
[142] L. Davidson. Hybrid LES-RANS: Inlet boundary conditions. In B. Skallerud
and H.I. Andersson, editors, 3rd National Conference on Computational Mechanics MekIT05 (invited paper), pages 722, Trondheim, Norway, 2005.
[143] L. Davidson. Hybrid LES-RANS: Inlet boundary conditions for ows with recirculation. In Second Symposium on Hybrid RANS-LES Methods, Corfu island,
Greece, 2007.
[144] L. Davidson. Using isotropic synthetic uctuations as inlet boundary conditions for unsteady simulations. Advances and Applications in Fluid Mechanics,
1(1):135, 2007.
[145] J.R. Welty, C.E.Wicks, and R.E. Wilson. Fundamentals of Momentum, Heat,
and Mass Transfer. John Wiley & Sons, New York, 3 edition, 1984.
[146] L. Davidson. HYBRID LES-RANS: Inlet boundary conditions for ows with
recirculation. In Advances in Hybrid RANS-LES Modelling, volume 97 of Notes
on Numerical Fluid Mechanics and Multidisciplinary Design, pages 5566.
Springer Verlag, 2008.
[147] L. Davidson and B. Farhanieh. CALC-BFC: A nite-volume code employing
collocated variable arrangement and cartesian velocity components for computation of uid ow and heat transfer in complex three-dimensional geometries. Rept. 95/11, Dept. of Thermo and Fluid Dynamics, Chalmers University
of Technology, Gothenburg, 1995.
[148] F. Eriksson. Flerdimensionell analys (in Swedish). Studentlitteratur, Lund, Sweden, 4 edition, 1977.
[149] L. Davidson.
How to make energy spectra with Matlab (available at
www.tfd.chalmers/lada). Report 2006/05, Div. of Fluid Dynamics, Dept. of
Applied Mechanics, Chalmers University of Technology, G teborg, Sweden,
o
2006.
[150] R.L. Panton. Incompressible Flow. John Wiley & Sons, New York, 1984.
[151] L. Davidson. private communication. Division of Fluid Dynamics, Dept. of
Applied Mechanics, Chalmers University of Technology, 2007.
Q. References
237