Lecture Notes in Turbulence

Steve Berg

2nd June 2004

Contents

1 Conservation Equations 2
1.1 Conservation of Mass . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Conservation of Moment . . . . . . . . . . . . . . . . . . . . . . 4

2 Turbulent Theory 9
2.1 What is Turbulence ? . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Turbulence Spectrum . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3 Correlation in Turbulence . . . . . . . . . . . . . . . . . . . . . 16
2.4 Turbulence Equations . . . . . . . . . . . . . . . . . . . . . . . 23

3 Turbulence Models 26
3.1 Zero Equation Model . . . . . . . . . . . . . . . . . . . . . . . . 26
3.2 One Equation Model . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3 Two Equation Model . . . . . . . . . . . . . . . . . . . . . . . . 29
3.4 Reynolds Stress Model (RSM) . . . . . . . . . . . . . . . . . . . 31

4 Summary 36

A Turbulence Pictures 38

B Examination Questions 41

1
Chapter 1

Conservation Equations

Transport equations are the basics for all CFD simulations. Before we discuss
turbulence models we will derive these equations for laminar flow. Turbulence
will be added in chapter 2.4. This chapter is based on lectures notes from
Mathiesen (2000) and as an alternative you can also read chapter 6 in Munson
et al. (2002).

1.1 Conservation of Mass

This equation is also known as equation of continuity. It is derived from the

concept of control volumes, ie the mass entering a volume also has to leave it.
Before we start discussing the details about mass flow, let us first define the
basic equation for mass flow in one direction:

kg
mx = ρux A [ ]
s

This equation is used for all directions for calculation of mass flow. In our
case we will only derive the equation of mass for two direction (x and y). It is
easier to understand the procedures for the derivation of the equations when
there are a few variables as possible. The third flow direction part (z-direction)
can easily be added to the two direction equation.

2
 (ρuy )|y+∆y ∆x∆z
6

y + ∆y

(ρux )|x ∆y∆z - ∆x∆y∆z = ∆V - (ρux )|x+∆x ∆y∆z

y
6

x x + ∆x
(ρuy )|y ∆x∆z

Accumulations of Incoming mass Outgoing mass

mass per = per time unit - per time unit
time unit

Accumulation of mass per time unit:

∂ρ ∂
∆V = (ρux )∆x∆y∆z
∂t ∂t
Mass flow in and out of the control volume is then given by:

(ρux ) |x ∆y∆z +(ρuy ) |y ∆x∆z −(ρux ) |x+∆x ∆y∆z −(ρuy ) |y+∆y ∆x∆z
| {z } | {z } | {z } | {z }
inf low−x inf low−y outf low−x outf low−y

We then put the inflow and outflow for x directed flow in the same equation,
and do the same for the y directed flow:

{(ρux ) |x −(ρux ) |x+∆x } ∆y∆z

{(ρuy ) |y −(ρuy ) |y+∆y } ∆x∆z

We want to express this variables with the mean value:

− {(ρux ) |x+∆x −(ρux ) |x }

∆y∆x∆z
∆x

− {(ρuy ) |y+∆y −(ρuy ) |y }

∆y∆x∆z
∆y

The only thing we have done is multiplying with respectively ∆x and ∆y in

numerator and denominator. If we add all this parts together in one equation,
the final result will be:

3
∂ρ {(ρux ) |x+∆x −(ρux ) |x } {(ρuy ) |y+∆y −(ρuy ) |y }
∆x∆y∆z = − ∆y∆x∆z− ∆y∆x∆
∂t ∆x ∆y

∆z can be removed from all parts of the equations above.

∂ρ {(ρux ) |x+∆x −(ρux ) |x } {(ρuy ) |y+∆y −(ρuy ) |y }

∆x∆y = − ∆y∆x− ∆y∆x
∂t ∆x ∆y

If the control volumes becomes very small (∆V → 0, then ∆x → 0 and

∆y → 0) , then the final equation for mass will be:

∂ρ ∂ ∂
= − (ρux ) − (ρuy )
∂t ∂x ∂y

If we re-arrange this equation and add the last term for the flow in z-
direction, we will end up with the final expression for mass conservation in 3
dimensions.

∂ρ ∂ ∂ ∂
+ (ρux ) + (ρuy ) + (ρuz ) = 0
∂t
|{z} |∂x {z } ∂y |∂z {z }
| {z }
Accumulation convection−x convection−y convection−z

1.2 Conservation of Moment

The equation for conservation of moment is derived from the conservation of

impulse (ρ~v ). The word impulse and moment have the same meaning and
describes a force. In the deduction below we will concentrate on 2 dimensions
only and the flow is x directed (ρux ). The 3rd dimension can easily be added
afterwards.
(ρux )uy |y+∆y ∆x∆z
6

y + ∆y

(ρux )ux |x ∆y∆z - ∆x∆y∆z = ∆V - (ρux )ux |x+∆x ∆y∆z

y
6

x x + ∆x
(ρux )uy |y ∆x∆z

4
 Accumulations of Incoming impuls Outgoing impuls Sum of all forces
impuls per = per time unit - per time unit + on the system
time unit

Accumulation of impulse per time unit can be written as:

∂
(ρu~x )4V
∂t

4V is the size of the control volume 4x4y4z. For the convective trans-
port we can derive equations in the same manner as in section 1.1.

 ρu u |
x y y+∆y − ρux uy |y
 ρu u |
x x x+∆x − ρux ux |x
 
− ∆x∆y∆z − ∆x∆y∆z
| ∆x {z } | ∆y
{z }
changes of x-impuls in x-direction changes of x-impuls in y-direction

Sum of all forces on the system is a function of pressure, viscous forces and
gravity.
Pressure is an isotropic variable, ie it has the same value in all directions.
The conservation equation for pressure can be written as:
P |y+∆y ∆x∆z

?
y + ∆y

P |x ∆y∆z - ∆x∆y∆z = ∆V  P |x+∆x ∆y∆z

y
6

x x + ∆x
P |y ∆x∆z

P|
x+∆x − P |x

(P |x − P |x+∆x )∆x∆z = − ∆V
∆x

Viscous forces are diffusion of impulse caused by molecular transport (we

do not include turbulence - yet). We define these stresses as shear- and normal
stresses. τxy = τyx are the shear stresses, and τxx and τyy are the normal
stresses. Both have the units (N/m2 ). Stresses that works in the y-direction
comes in by the y-directed impulse.

5
 ? τxy |y+∆y ∆x∆z

y + ∆y
6

τxx |x ∆y∆z - ∆x∆y∆z = ∆V  τxx |x+∆x ∆y∆z

?
y -

6 τxy |y ∆x∆z

x x + ∆x

Shear stress is a deformation force that alters the control volume shape
while the normal stresses changes the volume.
The viscous forces are given by:
τ |
xy y+∆y − τxy |y
τ |
xx x+∆x − τxx |x
 
− ∆x∆y∆z − ∆x∆y∆z
∆x ∆x

-
gx

?
gy

Gravity forces are given by ρgx ∆V

Now we summarise all the equations above.

 ρu u |
x y y+∆y − ρux uy |y
 ρu u |
∂ x x x+∆x − ρux ux |x
 
∆V (ρux ) = − ∆V − ∆V
∂t ∆x ∆y
τ |
P |x+∆x − P |x xy y+∆y − τy |y
τ |
xx x+∆x − τx |x
 
− ∆V − ∆V − ∆V + ρgx ∆V
∆x ∆x ∆y

∆V → 0 (∆x → 0, ∆y → 0)
The conservation of impulse in 2 dimensions in x-direction is:

6
 ∂ ∂ ∂ ∂P ∂τxx ∂τxy
(ρux ) + (ρux ux ) + (ρux uy ) = − − − + ρgx
∂t ∂x ∂y ∂x ∂x ∂y

Equally, the conservation of 2 dimensional impulse in y-direction is:

∂ ∂ ∂ ∂P ∂τyy ∂τxy
(ρuy ) + (ρuy ux ) + (ρuy uy ) = − − − + ρgy
∂t ∂x ∂y ∂y ∂y ∂x

For a 3rd dimension, an impulse in z-direction can be written in the same

manner. The final equations for conservation of momentum in 3 dimensions
are:

∂ ∂ ∂ ∂ ∂P ∂τxx ∂τxy ∂τxz

(ρux )+ (ρux ux )+ (ρux uy )+ (ρux uz ) = − − − − +ρgx
∂t ∂x ∂y ∂z ∂x ∂x ∂y ∂z
∂ ∂ ∂ ∂ ∂P ∂τxx ∂τxy ∂τxz
(ρuy )+ (ρuy ux )+ (ρuy uy )+ (ρuy uz ) = − − − − +ρgy
∂t ∂x ∂y ∂z ∂x ∂x ∂y ∂z
∂ ∂ ∂ ∂ ∂P ∂τxx ∂τxy ∂τxz
(ρuz )+ (ρuz ux )+ (ρuz uy )+ (ρuz uz ) = − − − − +ρgz
∂t ∂x ∂y ∂z ∂x ∂x ∂y ∂z

Written in tensor notation:

∂ ∂ ∂P ∂τij
(ρui ) + (ρui uj ) = − − + ρgj
∂t ∂x ∂xj ∂xj

Can also be written as:

Duj ∂P ∂τij
ρ =− − + ρgi
Dt ∂xj ∂xi

The stress tensor, τij , is fluid depended. We have already mentioned that
viscosity is divided in shear- and normal stresses. For a Newtonian fluid we
can write:
 
∂ui ∂uj 2 ∂uk
τij = −µ + − δij (1.1)
∂xj ∂xi 3 ∂x
| {z } | {z k}
shear stresses normal stresses

∂ux ∂ux ∂uy ∂uz

= + + (= 0) (1.2)
∂xk ∂x ∂y ∂z

The term k is the sum of x, y and z, and is given by equation 1.2. δij is the
Kroenicker delta and is equal to one when i = j, and zero when i 6= j. That
means, when Kroenicker delta is equal to zero, normal stresses are absent, and
the flow is incompressible (ρ is constant). The volume will not change, so we
can ignore the last term in equation 1.1. If we include combustion, there will

7
be large temperature gradients, and we can no longer assume that the density
is constant.
The stresses in equation 1.1 can be simplified to:

 ∂u
i ∂uj 
τij = −µ +
∂xj ∂xi

The derived function for τij can be written as:

∂τij ∂  ∂ui ∂uj 

= −µ + (µ = constant)
∂xi ∂xi ∂xj ∂xi

∂τij  ∂2u ∂ 2 ui 
i
= −µ +
∂xi ∂xi xj ∂xi xi

∂τij  ∂ ∂u
i ∂ 2 ui 
= −µ +
∂xi ∂xj ∂xi ∂xi xi

From continuity of mass we know that for an incompressible flow:

∂ui
=0
∂xi

∂τij  ∂2u 
i
= −µ
∂xi ∂xi xj

From this we write the two general equation which are generally used
in CFD: Navier-Stokes equation (1.3) and Euler equation (1.4). For Navier-
Stokes we assume that both ρ and µ are constant. The Euler equation is often
used in aero dynamics, since the viscosity for air is negligible (µ ∼ 0).

∂ ∂ ∂P ∂ 2 ui
(ρui ) + (ρui uj ) = − +µ + ρgi (1.3)
∂t ∂xi ∂xi ∂xi ∂xj

∂ ∂ ∂P
(ρui ) + (ρui uj ) = − + ρgi (1.4)
∂t ∂xi ∂xi

8
Chapter 2

Turbulent Theory

In this chapter we will explain what turbulence really is, how it is created, and
how it is transferred from the mean flow to the smaller whirls. We will describe
the differences between isotropic-, homogeneous- and stationary turbulence.
The turbulent flow can be divided in spectrum, and we will derive equations
for the smallest scales. Correlations are important in turbulence. From these
numbers we can derive energy spectrum for all scales, and derive the Taylor
micro scales and the Reynolds number, Reλ , associated with it. At the end of
this chapter we will derive the basic Reynolds Averaged Navier-Stokes (RANS)
equation which is used in a CFD code for simulation of turbulent flow.

2.1 What is Turbulence ?

In flows which are originally laminar, turbulent arises from instabilities at

large Reynolds number. In a pipe, the flow is laminar when Re ≤ 2100, and
is turbulent when ≥ 4100. Between these two limits, there is an transitional
zone (2100 < Re < 4100). This is illustrated in figure 2.1. For a flat plate
the transition in boundary layers between laminar and turbulent flow occurs
at Reynolds number at approximately 500 000.
The characteristic for a turbulent flow is that the fluid no longer moves
as individual molecules but as macroscopic “balls” of fluid. The flow contains

Random turbulent fluctuations Turbulent

Re = UD/v

4000

Turbulent bursts Transitional

2000

Laminar

t, sec

Figure 2.1: Transition from laminar to turbulent flow in a pipe

9
 U
Moving plate

u
h

Wall

Figure 2.2: Example of Couette flow (Munson et al., 2002)

a wide range of length and time scales of the turbulent motion. The size
of the largest whirls are determined by the geometry, and the size of the
smallest whirls are determined by Kolmogorov scales. The flow is almost
always 3 dimensional. Dissipation by viscous forces dampen turbulent unless
new energy is constantly supplied. Turbulence relies on extracting energy from
the mean flow, eg a shear layer. Vortex stretching is an important mechanism
feeding energy into vortices.
Boussinesq assumed as early as 1877 that transport by turbulence is an
extra stress, or an extra viscosity.

τeff = τ + τturb µeff = µ + µturb (2.1)

Before we go any further, we shall discus what viscosity is. A typical

example is given between two plates. One plate is at rest, and the other is
moving with the flow (see figure 2.2). From experiments we know that the
fluid adheres to both plates. The velocity of the lower plate is zero and the
velocity at the upper plate is moving at the same velocity as the fluid. The
velocity distribution between the two plates is linear. The fluid velocity is
proportional to the distance y from the lower plate. In order to support the
motion it is necessary to apply an tangential force to the upper plate. It is
known from experiments that this force is proportional to the velocity U of
the upper plate, and inversely proportional to the distance h. The frictional
forces per unit area, denoted by τ , is proportional to u/h, for which in general
we substitute with du/dy. The proportional factor between τ and du/dy, is
often denoted by µ. And this factor depends on the nature of the fluid. It is
small for thin fluids, and large for fluids like oil (viscous). The frictional forces
can be written as:

du
τ =µ (2.2)
dy

An other word for frictional forces, τ , is shearing stress. To explain why

there is a viscous force, and a velocity gradient at the boundary layer, we can
imagine that there are molecules diffusing from the higher velocities in the free
flow to lower velocites at the wall. And opposite, low velocity molecules are

10
 y y

A A A A

Figure 2.3: Turbulence balls is assumed to behave in the same way as molecules
in laminar flow

migrating from low velocity layer to a high velocity layer. This momentum
transfer will result in a shear layer, where the layer near the wall has nearly zero
velocity, and the shear layers further up have an gradual increasing velocity.
This idea about momentum transfer, caused by molecular movement between
velocity gradients, is not restricted to boundary layers only. It also takes place
in other types of flow like jets and turbulent mixing layers.
Prandtl used Boussinesq’s idea in equation 2.1 to develop a simple way
to calculate turbulence (algebraic or zero equation model). Prandtl used the
wall mixing model where he assumed that the eddies behaves in the same way
as molecules in laminar flow. Fluid lumps moving toward the low velocity
regions, across line A-A, have some of their excess moment removed by the
lower velocity fluid. Conversely, lumps moving away from the lower velocity
region gain momentum from their new high velocity surrounding fluid. This
is illustrated in figure 2.3. Lower velocity regions acts as a momentum sink.
The eddies transport mass from one place to another. Flux of momentum
is called stress, or said in another way, turbulent stresses is the transport of
momentum (ρu0 ) with the υ fluctuation.

τturb = υ 0 (ρu0 ) (2.3)

This way of representing stress, where we assume that the eddies in tur-
bulent flow behave in the same manner as molecules in laminar flow, is called
the “eddy-viscosity approximation”. We use this approximation in zero-, one-
and two equations model, see section 2.4 and chapter 3. To verify that the
this flux of momentum really is a stress, we can compare the units for both
molecular viscosity in equation 2.2 and turbulent viscosity in equation 2.3.
 
du kg m/s kg N
τ =µ · = = 2
dy m · s m m · s2 m
 
0 0 kg m m kg N
τturb = ρu v 3
· · = 2
= 2
m s s m·s m

11
Turbulent Transport Larger whirls increases the transport across the flow.
They transport fluids at distances equal to largest whirls, le . Smaller whirls
does the same, but over a shorter distance. They are less important for trans-
port, but are important for chemical reactions 1 . Larger whirls are more
important for prediction of flow than smaller whirls.

Energy transfer in turbulence Energy are transferred from the mean

flow to the larger whirls. This is often explained by the verse2 :

Big whirls have little whirls that feed on their velocity, and little
whirls have lesser whirls and so on to viscosity - in the
molecular sense

Figure 2.4: Breakdown of whirls from larger to smaller whirls (Ertesvåg, 2000)

Breakdown of turbulence (from larger to smaller whirls) takes place by

two different mechanisms; larger whirls accelerate surrounding fluid to smaller
whirls, and/or the whirls are stretched to longer and thinner whirls. This is
illustrated in figure 2.4.
The energy is transferred to smaller whirls, until these whirls starts to
disintegrate into a molecular movement. Or said in another way: the molecule
transport changes from convection to diffusion where the kinetic energy is
transferred into heat. This is shown in figure 2.5. This process is called
dissipation. Dissipation takes place at all levels, but for high Reynolds number
most of the dissipation takes place in the smaller whirls.
We notice from figure 2.5 that there can be defined a Reynolds for each
individual scale. This Reynolds numbers are defined in the same way as the
Reynolds number for a pipe:

ρvD vD
Repipe = =
µ υ

v is the velocity for the fluid along the pipe and D is the pipe diameter. µ
is dynamic viscosity and υ is kinematic viscosity. For a vortex, the diameter
is defined as `, and the vortex rotating velocity is u. Notice the difference
between velocity in a pipe and in a vortex; axial velocity for a pipe and radial
velocity for a vortex.
1
We will come back to this in Combustion Technology, FACE 9
2
From the English scientist Lewis Fry Richardson, 1922

12
 Instabilities in flow causes a degredation u00 `00
of larger whirls to smaller ones with
Reu00 = ν
1
0
`
characteristic length `00 , and velocity u00
u0

`00
The larger whirls are defined by the
characteristic length `0 ,
u00
and velocity u0 .
u0 `0
Reu0 = ν
1
η
This continues until the Reynolds number υ
are equal to 1. Characteristic
length is η, and velocity is v.
υη
Reη = ν
=1

Figure 2.5: Energy is transferred from the larger to the smaller whirls
(Ertesvåg, 2000)

Isotropic, Homogeneous and Stationary Turbulence Researchers have

always (specially earlier times) tried to simplify turbulence. It is important to
notice that these simplified turbulence hardly exist in practice. They are, how-
ever, easier to study than “real” turbulence. Some of the constants used in the
turbulence models described in chapter 3, are derived under these conditions.

• In homogeneous turbulence, the fluctuating components of velocity u0 (x, t)

and pressure p0 (x, t) are statistically homogeneous in the entire volume.
It follows that imposed mean velocity gradients (∂ ūi /∂xj ) must also be
uniform, although they can vary with time.
Since there are not any gradients in mean flow velocities the production
of turbulence can be ignored. This will simplify the turbulence equations
and they can be solved analytically.

• Isotropic turbulence is statistical independent of direction, ie the normal

0 0 0
stresses are all equal: u12 = u22 = u32 . For shear stresses we got: u01 u02 =
0 (i 6= j). An mathematical explanation is that the correlations remains
unchanged if the coordinate system is turned. More about correlation in
section 2.3. We often assume that the smallest whirls in high Reynolds
number flow is isotropic. In some special cases the larger whirls can be
isotropic, eg flow behind a grid in a wind tunnel where the measuring
probe is moving with the flow. Isotropic turbulence must in practice be
homogeneous, ie all points in space have the same properties.

• Stationary turbulence is statistical independent of time.

2.2 Turbulence Spectrum

The main flow will transfer energy to the larger whirls. These whirls are
determined by the the outer dimensions of the flow, eg the pipe diameter.
Time scales for the small whirls are short. They are in close equilibrium with

13
 Universal Equilibrium Range
Energy Containing
Range
Dissipation Range Inertial Subrange

η `e

Figure 2.6: Turbulent spectrum with the larger energy containing eddies to
right and dissipation eddies to left

the local properties. They are called universal equilibrium range. The term
universal is used because this assumption goes for all high Reynolds number
flows. The different scales in a turbulence spectrum are illustrated in figure
2.6. The spectrum is divided in 2- or 3 parts. High Reynolds number flow
has 3 parts while low Reynolds number flow has only two parts. The inertial
subrange vanish for lower Reynolds number. The universal equilibrium range
can be divided in two separate ranges, inertial sub-range and dissipation range.

Inertial subrange In the inertial region the turbulence scales are indepen-
dent of both the large scales and the small scales if the Reynolds number is
large. For turbulent flow with low Reynolds number this region is “small”, ie
only dissipation- and energy containing range are present.
This region is characterised by the amount of energy transported through
the spectrum per time (ie ε) and the sizes of the eddy (ie 1/κ). The energy in
the region can be estimated as:

E ≈ ε a κb

Dimensional analysis for ε is L2 /T 3 , for κ it is 1/L, and for E it is L3 /T 2 .

This gives:

a  b
L2 L2a−b

a b 1
E(κ) = (ε κ ) = =
T3 L T 3a

2
which gives: a = 3 and b = − 53

2 5
E ∼ CK · ε 3 κ− 3

This expression is called the Kolmogorov spectrum law or −5/3 power

law and is illustrated in figure 2.7. This function is derived from dimensional
analysis, and measurements have verified its existence. CK is the Kolmogorov
constant and is approximately 1.5.

Dissipation range Viscous forces have a greater impact on smaller whirls

than the larger ones. When viscous forces overcomes the inertia movement,

14
 E(κ)

slope = − 53

κ1

Figure 2.7: - 5/3 low for inertial subrange. Both the wavenumber κ- and the
energy E(κ) axis are given in logarithmic function

the whirls will dissipate into a molecular movement. These whirls are named
after the person who first came up with this ”idea”, Kolmogorov. Kolmogorov
also assumed that:

• small whirls are independent of larger whirls

• flow directions for the smaller whirls are independent of larger whirls -
local isotropy

As illustrated in figure 2.5, Reynolds number for the smallest whirls are
equal to 1, ie inertia forces equals viscous forces (or there is a balance between
momentum forces and diffusion).

ηυ
Reη = =1
ν

The scales of the smallest eddies are determined by viscosity, ν (m2 /s), and
dissipation (increase of thermal energy), ε (energy/time=m2 /s3 ). The length
scale, η, for the smallest eddies can be expressed as:

η = ν a εb

From dimensional analysis we can find that:

m : 1 = 2a + 2b s : 0 = −a − 3b
3
which gives a = 4 and b = − 14 . The Kolmogorov length scale is then given
by:

15
  14
ν3

η=
ε

Writing the velocity scales in the same way gives the Kolmogorov velocity
scale υ and the Kolmogorov time scale τ as:

1
υ = (νε) 4

ν 1
2
τ=
ε

We have seen above that there can be derived expressions for inertial- and
dissipation range. But these expressions, derived by Kolmogorov, are derived
from dimensional analysis, and not from experimental measurements. There
are actually smaller scales than the Kolmogorov scales. We will come back to
this later in this chapter.

2.3 Correlation in Turbulence

Auto correlation and cross correlation are two important statistical factors in
turbulent analysis. These correlations are meant for isotropic turbulence but
we also use them as approximations to real turbulence. First we need a basic
explanation about what correlation really is.
Correlation is a measure of linear relationship, or the study the strength
of relationship between two random variables. One simple kind of association
between the variables x and y produces pairs of values or, graphically, points
that scatters around a straight line. A numerical measure of this relationship
is called the sample correlation coefficient (SCC), and is given in equation
2.4. Small amount of scatter around a line indicates strong association and
large amount of scatter is a result of weak association. Examples are given in
figure 2.8. High values for SCC indicates a strong relationship, as shown in
(a). High negative values in (b) indicates that there is an anti-correlation. If
values for SCC approximates 0.5, as in (c), there is not any clear relationship.
A further description about correlation can be found in any basic book about
basic statistics, such as Bhattacharyya and Johnson (1977).
Pn
i=1 (xi− x̄)(yi − ȳ)
SCC = q P (2.4)
n 2
 Pn
2

i=1 (xi − x̄) i=1 (yi − ȳ) )

Informations about connections in time and space between fluctuations can

be gathered from two-points correlations. Two-points correlations between
two velocity components can be defined by equation 2.5.

Qij = u0i (x)u0j (x + ∆x) (2.5)

16
 (a) SCC=0.915 (b) SCC=-0.912 (c) SCC=0.492

Figure 2.8: Examples for different values of SCC

For standardisation we often define this function with help of the root-
square-mean (σ) values at the same points. The result is:

Qij
Rij =
σi σj

The correlation coefficient is thereby limited to the interval [−1, 1]. In

isotrop turbulence all directions possibilities are equivalent which results inn
Rij = 0 for (i 6= j). Rij can then be defined by two functions, f (r) and g(r),
where r is the distance between the points (r = ∆x).
Consider two discretionary, statistic, stationary functions, u0 and v 0 , which
respective mean values is zero. We introduce the term Ru0 v0 for the specific
two-point correlation between u0 and v 0 (time-space correlation):

u0 (x, t)v 0 (x + ∆x, t + ∆t)

Ru0 v0 (x, ∆x, ∆t) =
σu σv

We call this correlation coefficient cross-correlation; if u0 and v 0 describe

the same variable, either u0 or v 0 , we call it autocorrelation.
Auto correlation is often used in Laser Doppler Anemometry (LDA) to
avoid velocity bias (Dantec, 2000). In LDA we do not measure fluid velocity
directly but instead we measure seeding particles immersed in the fluid. Dur-
ing periods of higher velocities, a larger volume of fluid is swept through the
measuring volume, and consequently a great number of velocity samples will
be recorded. As a direct result, an attempt to evaluate the statistics on the
flow field using arithmetic averaging will bias the result in favour of the higher
velocities. One way to avoid this biasing is to find the time where two mea-
surements are independent of each other. The autocorrelation will provide
us this information. Or said in another simplified way, the autocorrelation
gives us the time for a whirl to pass by the measured volume. In this kind of
measurements ∆x = 0, ie we are measuring in the same volume, over time.
Correlations can also be measured at two different locations. This will
give us some useful information about the length scales of the eddies in turbu-
lence. If we measure along a straight line, we can either measure the velocity
fluctuations along- or by the line (longitudinal and lateral).
The longitudinal correlation is given by:

17
 f (r) g(r)
1 1

0 0 r
r
0 0

Figure 2.9: Typical examples of longitudinal- and lateral correlation

u0` r u0` (r)

- - -

u0` u0` (r)

R`` (r) = u0` · u0` (r) = f (r) · u02 ⇒ f (r) =
u02

The lateral correlation is given by:

6
6
u0n u0n (r)
r
-

u0n u0n (r)

Rnn (r) = u0n · u0n (r) = g(r) · u02 ⇒ g(r) =
u02

If two points are at the same spot, we are “looking” at the same fluctuation.
The correlation is then 1.

f (0) = g(0) = 1

When distances increases, the correlation between them will gradually be-
come smaller. If correlation is equal to zero, they are uncorrelated.

f (∞) = g(∞) = 0

The value for f (r) can be negative for a part of of r. The value for g(r)
must be negative for a part of r. A fluctuation across the line must result in
fluctuation in opposite direction (continuity). This is illustrated in figure 2.9
We can define a length scale Lf where the area f (0) is equal to the area
under the curve f (r) ≥ 0, as indicated in figure 2.10. Since f (0) = 1, we can
write the longitudinal macroscale:

18
 f (r) f (r)
1 1

0 0
r r
0 Lf 0 λf

Figure 2.10: The macroscale is defined as Lf and microscale as λf

Z ∞
Lf = f (r)dr
0

We can define the lateral correlation function, g(r), in the same way:
Z ∞
Lg = g(r)dr
0

It can be shown that Lf = 2Lg . These two length scales tells us the
distance fluctuations will influence each other.
Taylor series are polynomials that can be used to approximate other func-
tions. The general formula for Taylor series are given in equation 2.6.3

f 00 (a) f ka f na
Pn (x) = f (a)+f 0 (a)(x−a)+ (x−a)2 . . .+ (x−a)k +. . .+ (x−a)n
2! k! n!
(2.6)
Taylor series for f (r) around r = 0 gives:

1 00
f (r) = f (0) + f 0 (0)r + · f (0) · r2 + . . .
2

We know that f (0) = 1 and f 0 (0) = 0. If we round off the Taylor series
after the r2 link, the result will be a parable which crosses the r-axis. We
define a length scale λf as a values instead of r where the parable crosses
the r-axis (f (r = λf ) = 0). This length scale is called the longitudinal micro
scales.

2
λ2f = −
f 00 (0)

From the lateral correlation function g(r) we can define a lateral microscale
in the same manner as for the longitudinal micro scales.
3
A more detailed discussion about this topic can be found in Kreyszig (1993)

19
 2
λ2g = −
g 00 (0)
√
It can be shown that (Hinze, 1975) λf = 2λg. Dissipation of turbulence
energy is:

∂u0i ∂u0i
ε=v
∂xj ∂xj

For isotropic turbulence:

2 2 2
∂u01 ∂u02 ∂u03
  
= =
∂x1 ∂x2 ∂x3

2 2 2 2 2 2
∂u01 ∂u01 ∂u02 ∂u02 ∂u03 ∂u03
     
= = = = =
∂x2 ∂x3 ∂x1 ∂x3 ∂x1 ∂x2

  0 2  0 2 
∂u1 ∂u1
ε=v 3 +6
∂x1 ∂x2

According to Hinze (1975):

2
∂u01 2u02

= −u02 · f 00 (0) =
∂x1 λ2f

2  0 2
∂u01 4u02 2u02

00 00 ∂u1
= −u02 02
· 2f (0) = 2 = 2 = −u · g (0) = 2
∂x2 λf λg ∂x1

2
∂u01 u02 u02

ε = 15v = 30v 2 = 15v 2
∂x1 λf λg

It can be practical to define a length scale such as: λ = λg

u02 k 2
ε = 15v 2
= 10v 2 , u02 = k
λ λ 3
We only have one length scalar, as in isotropic turbulence. λ is equal to
λg , and with this length scale we can define a turbulence number:

u0 λ
Reλ =
v
We know that the larger whirls are determined by the outer dimensions,
and the smaller whirls by viscous forces. A turbulence Reynolds number, such
as Reλ , gives a measure for the degree of turbulence. It also gives the difference

20
between the smaller (η) and the larger (`e ) whirls. If Reλ is high, there will
be a large difference between small and large whirls. It can also be shown that
for isotropic turbulence (Hinze, 1975):

`e 3
∼ 0.1Reλ2
η

What are Typical Values for Reλ ? Ertesvåg (2000) discusses different
values for Reλ . The highest Reynolds numbers are found in the atmosphere
and oceans. The outer dimensions for these flows are huge, and typical
Reynolds number are 104 . Typical flows in laboratories are much lower. A
Reynolds number (Reλ ) at 150-200 is considered high. In special cases a
Reynolds number up to 103 is achievable.

Relationship Between Reλ and ReT In turbulence models with an equa-

tion for ε, there is often defined a Reynolds number:

k2
ReT =
vε

The relation between Reλ and ReT is according to Ertesvåg (2000):

 1  1
0 2 2 10vk 2
u = k λ=
3 ε

1
u0 λ 20 k 2

2
Reλ = =
v 3 vε

2
3k · 10vk 20 k 2 20
Re2λ = = = ReT
v2ε 3 vε 3

Energy spectrum The two point correlation given by:

Rij (~r, t) = u0i (~x, t)u0j (~x + ~r, t)

can be transformed by a Fourier transformation:

F
Rij (~r, t) → Eij (~κ, t)

Eij is the energy spectrum tensor and ~κ is the wave number vector (1/l).
The energy spectrum E(κ, t) are independent of directions, and is defined as:
Z ∞
1
E(κ, t)dκ = u0i u0i = k(t)
0 2

21
 E(κ)
D(κ)

`0 κe, `e κd, `d η −1

Figure 2.11: Energy and dissipation spectrum. The x-axis is logarithmic and
y-axis is linear (Ertesvåg, 2000)

This energy spectrum is the area under the E(κ) curve in figure 2.11.
Dissipation spectrum D(κ, t) is the area under the other curve in the same
figure, and is defined as:
Z ∞
D(κ, t)dκ = ε(t)
0

Both curves in figure 2.11 have a maxima at different locations from the
smallest (η) and largest whirls (`0 ). The largest whirls in a flow have the same
diameter as the outer dimensions, which is `0 . For example, the largest whirls
in a pipe have the same length as the pipe diameter. There are not many of
these larger whirls; most of the energy containing whirls are smaller. They
are indicated in figure 2.11 by the letters `e . The larger whirls, `0 , have more
energy than the smaller whirls, `e . But `e are more numerous, so most of
the energy is located around this scale. We learned earlier that the smallest
whirls are the Kolmogorov scales. This scales are derived from dimensional
analysis, and experiments has shown that there are actually smaller scales in
a flow. Most of the dissipation, however, takes place in larger whirls (`d ).
Hinze (1975) is referring to some measurements which indicates that `d is
typically 10 times larger than the Kolmogorov scales for isotropic flow. Pope
(2000) is referring to other measurements for real flows (non-isotropic) where
`d can be up to 60 times larger. Taylor scales are larger than the Kolmogorov
scales, and some people often relate this scales to `d , ie that the Taylor scales
represents the eddies sizes in which most of the dissipation occurs. There are
some disagreement about this statement. According to Tennekes and Lumley
(1972) the Taylor scales does not represent any group of eddy sizes in which
dissipative effects are strong. It is not a dissipation scale, because it is defined
with the assistance of a velocity scale which is not relevant for the dissipating
eddies.

What Happens with Energy Spectrum when Turbulence Decreases?

When turbulence decreases, k, ε, and Reλ has to decrease as well. But the
length scales are increasing. We can imagine that when the whirls are loosing

22
 E(κ)

Figure 2.12: Changes in energy spectrum when turbulence decreases

(Ertesvåg, 2000)

their “strength”, the particles will slip from each other and the whirls becomes
larger. The area under the energy spectrum is the turbulence energy k; this
area is decreasing as well. The length scales `e is increasing, and the top of
the curve will move towards left (lower wave number and higher/larger length
scales). Kolmogorov length scales are also increasing. The relationship `e /η
(is a measure for the extent of the spectrum) is decreasing. All this are shown
in figure 2.12.
The energy spectrum is decreasing and it shifts toward the longer length
scales when turbulence decreases.

2.4 Turbulence Equations

We derived the Navier-Stokes equation for laminar flow in chapter 1. We

repeat the equations for conservation of mass (equation 2.7) and momentum
(2.8) for simplicity. We assume that ρ = constant.

∂ρ ∂
+ =0 (2.7)
∂t ∂xj

∂ ∂ ∂p τij
(ρui ) + (ρui uj ) = − + + ρSi (2.8)
∂t ∂xj ∂xi ∂xj

  ρ = const.
∂ui ∂uj 2 ∂uk 
>
τij = µ + − µ  δij
∂xj ∂xi 3 
∂xk

Reynolds divided the velocities in a mean- and a fluctuating part. This is

illustrated in figure 2.13 where the velocity in a turbulent flow is measured at
a point. The instantaneous velocity is then given by equation 2.9 as the sum
of mean- and fluctuating velocity.

23
 u0
u

ū

Figure 2.13: Velocities fluctuate around the mean value

ui = ui + u0i (2.9)

To derive the Reynolds averaged equation, we need to insert ui from equa-

tion 2.9 in equation 2.8.

∂ ∂ ∂p τij
(ρ(ui + u0i )) + (ρ(ui + u0i )(uj + u0j )) = − + + ρSi
∂t ∂xj ∂xi ∂xj

(ui + u0i )(uj + u0j ) = (ui uj + ui u0j + u0i uj + u0i u0j )

∂ ∂ ∂p τij
(ρ(ui + u0i )) + (ρ(ui uj + ui u0j + u0i uj + u0i u0j )) = − + + ρSi
∂t ∂xj ∂xi ∂xj

We need to average each term in the equation. The mean value for a term
with only one fluctuation is equal to zero. We need three laws for averaging:

φ=φ

φ+ψ =φ+ψ

ψφ0 = ψ · φ0

0 0 0
∂ ∂ ∂p τ ij
0 0 + u0  u0i u0j )) = −
>

(ρ(ui + u )) + (ρ(u u + u u u + + + ρS i


>

i i j i j i j
∂t
∂xj   ∂xi ∂xj

The final result is Reynolds averaged equation (we assume that ρ = con-
stant) are given in equations 2.10 and 2.11.

∂uj ∂u0j
=0 and =0 (2.10)
∂xj ∂xj

24
 ∂ ∂ ∂p ∂
(ρui ) + (ρui uj ) = − + (τ ij − ρu0i u0j ) + ρS i (2.11)
∂t ∂xj ∂xi ∂xj

ρu0i u0j is the Reynold stresses, and is a new unknown variable that needs to
modelled. That is the purpose for a turbulence model. There are two different
approaches to model the Reynold stresses, either the eddy-viscosity approach
introduced in chapter 2.1, or we can model all six equations for turbulence
stresses. The eddy viscosity approach is often modelled by equation 2.12.
The six equation for turbulent stresses, represented by equation 2.13, can be
derived from Navier-Stokes equations in the same way as we derived the ex-
pressions for the turbulent equation of momentum (equation 2.11). Definition
of the different terms in equation 2.13 are discussed in chapter 3.4.

ρ = const.
   
∂ui ∂uj 2 ∂u
l
7
−ρui uj = µt + − ρk + µt  δij (2.12)
∂xj ∂xi 3 ∂xl

∂u0i u0j
       
∂ 0 0 ∂ 0 0 0 0 ∂uj 0 0 ∂ui ∂
ρui uj + ρui uj uk = − ρui uk + ρuj uk + µ
∂t ∂xk ∂xk ∂xk ∂xk ∂x
| {z } | {z } | {z k }
ρCij ρPij ρDij,v

∂u0i ∂u0 ∂u0j

∂u0j
   
∂
− ρu0i u0j u0k − p0 u0i δjk + p0 u0j δik + p0 −2µ i


+ +
∂x ∂xj ∂xi ∂xk ∂xk
| k {z } | {z }| {z }
ρDij,t ρΦij ρεij

(2.13)

Unfortunately, turbulence models often introduce new unknown correla-

tions that are not known and further approximations are required.

25
Chapter 3

Turbulence Models

This chapter is about the different turbulence models. This are zero equation
model, one equation model and the two equation models. Zero- and one
equation models are (there are exceptions) incomplete, ie before we start a
simulation we need to know the length scales (or time scales). Two equation
models was derived in the beginning of the 70’, and was a break through in
turbulence modelling. The most popular two equation model, the standard
k-ε model, is the most (some would probably say - “only”) used turbulence
model in industry.

3.1 Zero Equation Model

Zero equation model are the simplest of all turbulence model. It was developed
by Prandtl in 1925. He visualised a simplified model for turbulent motion in
a boundary layer where the turbulent balls moves around as molecules. He
showed that the two velocity fluctuations can be given by:

du du
u0 = ` v0 = `
dy dy

Turbulent stresses is the transport of momentum (ρu0 ) with the υ fluctu-

ation

 2
0 0 2 du
τturb ∼ υ (ρu ) ∼ ρ`
dy

` is a mixing length and is a function of outer dimensions, eg pipe diameter

or a distance to a wall. Since ` is unknown, it has to be determined before
any calculation starts. The zero equation model is therefore an incomplete
turbulence model. This model is hardly in use any more, except some modified
versions in aerodynamics. Further details are given in Wilcox (2002).

26
3.2 One Equation Model

In one equation models we are modelling the turbulence energy, and solving
dissipation (ε) analytical. They are incomplete models since the length scales
have to be related to some typical flow dimensions. This model was a fore-
runner for the two equation model. There are (lately) developed complete one
equation models, eg Spalart-Allmaras turbulence model.
We need to develop an equation for turbulence energy - k. Kinetic energy
per mass unit in a flow is 12 ui ui . If we subtract the mean values from this,
we end up with the kinetic energy for the turbulent fluctuations 12 u0i u0i . The
mean value of this expression is the “mean kinetic turbulence energy” or just
“turbulence energy”

1 1
k = u0i u0i = (u02 + u02 02
2 + u3 )
2 2 1

We will not derive the equation for k, but we will use a method from
Ertesvåg (2000) and show how it can be done in 5 steps.

1. Based on equation for movement (ρui ), and assume that ρ is constant

2. Reynolds averaging: ui = ui + ui

3. Subtract the last equation from the first. The result is an equation for
the fluctuation u0i = ui − ui .

∂ ∂ ∂
(ρu0i ) + . . . = . . . =




(ρui ) + . . . = . . . − (ρūi ) + . . . = . . .
∂t ∂t ∂t

4. multiply this equation with fluctuation u0i ; the result is an equation for
1 0 0
2 ui ui (fluctuating turbulence energy)

5. Averaging this equation will result is an expression for mean turbulence

energy, k = 12 u0i u0i

∂ ∂ 1 0 0
averaging ∂ 1 0 0
u0i (ρu0i )+. . . = . . . =



(ρ ui ui )+. . . = . . . −→ (ρ ui ui )+. . . = . . .
∂t ∂t 2 ∂t 2

The result, with constant ρ, can be written as:

∂u0 ∂u0i
 
∂ ∂ ∂ ūi ∂ ∂k ∂ 1
0 0 − ρu0i u0i u0j − pu0j − µ i


(ρk) + (ρkūj ) = −ρui uj + µ +
∂t ∂xj ∂xj ∂xj ∂xj ∂xj 2 ∂xj ∂xj
| {z } | {z } | {z } | {z } | {z }
ρCk ρPk ρDk,v ρDk,t ρε

This equation is exact, ie it has been derived from the basic Navier-stokes
equation and we have not made any assumption, simplifications or trying to
model any terms.

ρCk transient- and convective term

27
Pk production term

Dk,t turbulent diffusion or mean convective transport with turbulent move-

ments

Dk,v viscous gradient term

ε dissipation

We cannot solve the exact equation for k. Some of these terms are unknown
and have to modelled (the equations is not exact any more). The triple corre-
lation u0i u0i u0j and pressure-velocity correlation p0 u0j are unknown, and Dk,t has
to be modelled. The viscous gradient term, Dk,v , can be calculated directly,
and does not need any modelling. In high Reynolds number flow, this term is
small compared to the turbulent viscosity, Dk,t . Pk can be calculated directly.
Dissipation term, ε, has to be modelled.
Turbulent diffusion, Dk,t , is modelled as a gradient model:

µt ∂k
(3.1)
σk ∂xj

This can be interpreted in two different ways:

1. As a model for “− 12 ρu0j u0i u0i − ρ0 u0j ”

2. As a model for “− 12 ρu0j u0i u0i ”, and “ρ0 u0j ” is ignored

Turbulent diffusion model is:

 
∂ µt ∂k
ρDk,t =
∂xj σk ∂xj

Turbulent “balls/lumps” have a length scale L and fluctuating velocity u0 .

The “force” against their flow is: F ∼ ρu02 · A ∼ ρu02 · L2 (similar to drag
force). Force per time unit (effect) is: F · u0 ∼ ρu03 · L2 . This is energy loss
for one ball. If we divide by volume, V ∼ L3 :

u03 k 3/2
ρε ∼ ρ → ε ∼ CD
L L

CD is a constant found from experiments. L is a characteristic length scale

for the larger whirls. The expression for ε can also be found by dimensional
analysis.
Molecular viscosity for a gas is given by1 µ ∼ ρ`ῡ, where ` is a length scale
and ῡ is mean velocity for molecules. We assume that the “balls/lumps” in
turbulent flows behave as molecules in gas. Turbulent viscosity can then be
defined as: µt = ρ`0 u0
1
From kinetic theory of gases

28
 √ √
vt ∼ u0 `0 u0 ∼ 2k ∼ k `0 = L

√
vt = CL kL (CL = 0.09)

  
∂ ∂ ∂ µt ∂k
(ρk) + (ρkuj ) = µ+ + ρPk − ρε
∂t ∂xj ∂xj σk ∂xj
 
∂ui ∂uj ∂ui
ρPk = µt +
∂xj ∂xi ∂xj

√
µt = ρvt = ρ kL

k 3/2
ε = CD
L
L is an algebraic expression and must be determined before the simulation
starts.

3.3 Two Equation Model

As opposed to zero (analytical)- and one equation models, two equations mod-
els are complete. They can be used to predict properties of a given turbulent
flow with no prior knowledge of the turbulence structures. We are using the
same k - equation as for the “one equation model”, and in addition we need
to find a transport equation for the length- or time scale. There are several
different approaches to find this extra scales, and the result is a variety of two
equation models. Some of them are well documented, whilst other are not
documented at all. The most popular model is the “Standard k − ε model”.
An exact equation for ε can be derived from the Navier-Stokes equations. The
final result will be:

∂ε ∂ε  ∂ ūi ∂ 2 ūi
= −2v u0i,k u0j,k + u0k,i u0k,j − 2v u0k u0i,j − 2v u0i,k u0i,m u0k,m

+ ūj
∂t ∂xj ∂xj ∂xk ∂xj
 
2 0 0 ∂ ∂ε 0 0 0 v 0 0
−2v ui,km ui,km + v − v uj ui,m ui,m − 2 pm uj,m
∂xj ∂xj ρ

According to Wilcox (2002), this equation is far more complicated than

the turbulence energy (k) equation and involves several new unknown double
and triple correlations of fluctuating velocity, pressure and velocity gradients.
These correlation are hopelessly difficult to measure with any degree of ac-
curacy, so there is little hope of finding reliable guidance from experiments
regarding suitable approximations.
The simplified ε equation (3.3) can be derived in a way that makes it look
similar as the k equation (3.2).

29
   
∂ ∂ ∂ µt ∂k
(ρk) + (ρkūj ) = µ+ + ρPk − ρε (3.2)
∂t ∂xj ∂xj σk ∂xj
  
∂ ∂ ∂ µt ∂k
(ρε) + (ρεūj ) = µ+ + ρPε − ρQε (3.3)
∂t ∂xj ∂xj σk ∂xj

Production and disintegration of ε is assumed to be proportional to pro-

duction and disintegration of k. Each of these terms have to be multiplied
with ε/k for get the right dimension.

ε ε
Pε = Cε Pk Qε = Cε2 ε
k k

The idea behind this relationship is that when turbulence energy increases,
the disintegration has to increase as well. Or said in another way: when Pk
increases, ε should also increase, if it does not, k can get unphysical high
values. The best way to increase ε is by increasing Pε - and that is the reason
why Pε is dependent on Pk . When turbulence energy decreases, disintegration
must also decrease. The complete set of equation and its constants for the
standard k − ε turbulence model are (Ertesvåg, 2000):

k2
µt = ρvt = Cµ ρ
ε
 
∂ ūi ∂ ūj 2
−ρu0i u0j = µt + − ρkδij
∂xj ∂xi 3
  
∂ ∂ ∂ µt ∂k
(ρk) + (ρkūj ) = µ+ + ρPk − ρε
∂t ∂xj ∂xj σk ∂xj

  
∂ ∂ ∂ µt ∂ε ε ε
(ρε) + (ρεūj ) = µ+ + Cε1 ρPk − Cε2 ρε
∂t ∂xj ∂xj σε ∂xj k k

 
∂ ūi ∂ ūj ∂ ūi
ρPk = µt +
∂xj ∂xi ∂xj

σk = 1.0 σε = 1.3 Cε1 = 1.44 Cε2 = 1.92 Cµ = 0.09

Standard k − ε models includes 5 universal constants. Some of these are

found from boundary layer flow, Cµ and Cε1 , others from wind tunnel exper-
iments, C2 . σk and σε are determined from computer optimisation.

30
What is ε ? For high Reynolds number, most of the dissipation takes place
in the smaller whirls. The mechanical energy is transferred from the mean
flow to the larger whirls, and thereafter to the smaller whirls.
The exact equation for ε represents the dissipating eddies, or the length
scale for these. Since this equation is hard to model, we are using a simplified
model where ε is given by the larger whirls. ε in the k − ε model is not the
dissipation, but it represents the energy transfer from larger to smaller whirls.
In stationary flow the energy transfer from the larger whirls are equal to the
dissipation in the smaller whirls.
Since the ε equation we use in k−ε models (and RSM) differs from the exact
equation derived from the Navier -Stokes equation, this equation is often seen
as the weak part in the turbulence model. In standard k − ε model (and other
two equation models) we only have one equation for the Reynolds stresses.
Advantages with the k − ε model are:

• fast (only two extra equations needs to be modelled)

• stable

Disadvantages:

• poor performance in in a variety of flows where the Reynolds stresses

are not equal in all directions

– curvatures in the flow (vortex, swirling flow, bends . . . )

– rotating flows

3.4 Reynolds Stress Model (RSM)

The Reynold stress model (RSM) does not use the eddy viscosity approach
described in chapter 2.1. Instead we are deriving 6 individual equations for the
stresses. This equation is already given in 2.13, but for simplicity we repeat
it:

∂u0i u0j
       
∂ 0 0 ∂ 0 0 0 0 ∂uj 0 0 ∂ui ∂
ρui uj + ρui uj uk = − ρui uk + ρuj uk + µ
∂t ∂xk ∂xk ∂xk ∂xk ∂x
| {z } | {z } | {z k }
ρCij ρPij ρDij,v

∂u0i ∂u0j ∂u0 ∂u0j

   
∂
− ρu0i u0j u0k − p0 u0i δjk + p0 u0j δik + p0 −2µ i


+ +
∂x ∂xj ∂xi ∂xk ∂xk
| k {z } | {z }| {z }
ρDij,t ρΦij ρεij

We need an explanation what the different terms means:

Cij transient term and transport with mean flow

31
Pij production, or energy transferred from the mean flow to the Reynolds
stresses

Dij,v viscous diffusion

Dij,t turbulence diffusion

Φij re-distribution term, exchange of energy between components

εij dissipation, transfer of kinetic energy to thermal energy

The convection term Cij , production term Pij , and viscous diffusion term
Dij,v can be solved directly. The viscous diffusion term is often negligible
compared to the turbulent diffusion term, but it remains important where
there are significant gradients. The other terms like turbulence diffusion Dij,t ,
re-distribution term Φij , and dissipation εij needs to be modelled.
The turbulent diffusion model is modelled as:

k 0 0 ∂u0i u0j
 
∂
Dij,t = Cs uk u` (3.4)
∂xk ε ∂x`

Cs is often set to 0.22. k/ε is used as a time scale. Dij,t can be interpreted
in two different ways (in the same way as in the k in equation 3.1):

1. as a model for the expression Dij,t

2. as a model for the triple correlation gradient, and the pressure term is
ignored

In practice will these two interpretations give the same expression/meaning.

Another approach to model turbulent diffusion is:

∂u0i u0j
 
∂ 0 k
Dij,t = Cµ k (3.5)
∂xk ε ∂xk

This model (equation 3.5) is easier to program than the other model in
equation 3.4 , and in some cases its proved to be more stable. Fluent uses a
similar simplified approach.
Dissipation tensor, εij , is assumed to be isotropic, ie dissipation is equal
for all normal stresses.

2
εij = εδij
3

Dissipation takes place in the smaller whirls. These whirls are indepen-
dent of the larger whirls and the main flow. For the small whirls, the flow
probability for the different directions are all equal. The dissipation must be
equal distributed between the three energy components u02 02 02
1 , u2 and u3 . This
gives: ε11 = ε22 = ε33 = 2/3 ε

32
 The pressure strain term is the most uncertain terms we are modelling in
the Reynolds stress equation. The classical approach to modelling Φij uses
the following decomposition:

Φij = Φij,1 + Φij,2 + Φij,w

Φij,1 slow pressure strain term, also known as return-to-isotropy term

Φij,2 rapid pressure strain term

Φij,w wall reflection term

Pressure strain term is also called the re-distribution term. It takes from
the terms who has a lot, and gives it away the the terms who has less (“Robin
Hood term”). Or said in another way: it shares the energy between the
components, or it makes the turbulence isotropic.
The slow pressure strain, Φij,1 , is modelled as:
 
ε 0 0 2
Φij,1 = −C1 ρ ui uj − δij k
k 3

where C1 = 1.8. The rapid pressure strain, Φij,2 , is modelled as:

 
2
Φij,2 = −C2 (Pij − Cij ) − δij (P + C)
3

Where C2 = 0.60. Φij,2 is called the “rapid term” because C2 is found

by “rapid distortion”, ie isotropic turbulence is distorted . The equation for
turbulence is in these cases are simple and can be solved analytically, which
gives C2 = 0.60. Pij and Cij are taken from the Reynolds stress transport
equation, P = Pkk and C = Ckk
Φij,1 and Φij,2 re-distributes velocity fluctuations in the “free flow”. When
the flow approaches a wall, the velocity components will be influenced by the
solid surface, and the turbulence are not seeking isotropy any more. An extra
term for the re-distribution for these parts of the flow is needed.
The wall reflection term, Φij,w , is responsible for the redistribution of nor-
mal stresses near the wall. It tends to damp the normal stresses perpendicular
to the wall, while enhancing the stresses parallel to the wall. Ertesvåg (1991)
gives some details about near wall turbulence. All velocity fluctuation are
reduced to zero at the wall (y → 0), but the relationship between fluctua-
tions and mean velocity approaches finite values. According to measurements:
u01 /ū1 ∼ 0.25 and u03 /ū1 ∼ 0.065 when y = 0. The fluctuations around mean
values are still intense, even when it approaches the surface. This is illustrated
in figure 3.1 where the fluctuation u01 increases in the mean flow direction ū1 ,
when the flow approaches the wall. The normal fluctuation, u02 , decreases
while the u03 fluctuations remains unchanged.
The wall reflection term, Φij,w , is given by the sum of Φij,1w and Φij,2w :

33
 k 3/2
 
ε 3 3
Φij,1w = C1w u0k u0m nk nm δij − u0i u0k nj nk − u0j u0k ni nk
k 2 2 C` εd

k 3/2
 
3 3
Φij,2w = C2w φkm,2 nk nm δij − φik,2 nj nk − φjk,2 ni nk
2 2 C` εd

where C1w = 0.5, C2w = 0.3. nk is the xk component of the unit normal
3/4
to the wall, d is the normal distance to the wall, and C` = Cµ /κ, where
Cµ = 0.09 and κ is the von Karman constant (= 0.4187).
The wall reflection term dampens the normal fluctuations near the solid
surfaces. But this term is not only a near-wall term, it also influences the
flow outside the boundary layer. For flow in complex geometries, it may
be difficult to define the wall distance. Developers of boundary layers and
numerical techniques makes the necessary fine mesh close to the wall, and the
results from the RSM are often satisfying. In industry, there are often limited
time or resources to use fine grids. According to Ertesvåg (2000), there are
indications that a corse grid near the wall can worsen the result compared
to a k-ε model. A typical simulation which require the solution of Reynolds
stresses due to swirl, rotations, etc could be solved more correctly with a k-ε
model if we cannot afford a fine mesh.
u2
ū
u01 /ū1
u02 /ū1
u03 /ū1

u3

u1

Figure 3.1: Turbulence variations for the 3 velocity components close to the
wall

The advantages with RSM is that it reproduces (or trying to reproduce) the
dynamics of each stress component which enables better modelling of stresses.
This will lead to better capturing of (Hanjalic, 2004):

• streamline curvature (flow separation, recirculation and stagnation re-

gions)

• strong pressure gradients

• swirling flows

34
• secondary motions (pressure induced and stress induced)

• three dimensionality effects

• compressibility and flow discontinuities (eg shock waves)

35
Chapter 4

Summary

When we use the theory for isotropic turbulence, and measure the correlation,
in our case the autocorrelation, we can find the macro scales and the Taylor
scales (or Taylor micro scales). The last is interesting because it gives us Reλ .
This number gives us the difference between the larger and smaller whirls,
and this number must be higher than 150-200 for the existence of the inertial
sub-range.
The theory for the two most interesting turbulence models, standards k −ε
and RSM, are discussed. The former is the most popular turbulence model
and is widely used in industry. There are many other turbulence models that
resemble the standard k − ε model, and they might perform better in some
types of flow.
The problem in turbulence modelling is to know when to use the different
turbulent models. In most cases the two equation models will be adequate for
many tasks. If the flow has sudden changes in flow directions, the Reynolds
stress model should be used. But there are back sides with the RSM. We have
already mentioned that the boundary layer should be better resolved when
we are using RSM. If we cannot resolve the boundary layer, we may have to
stick with two equation models, even if the physics of the flow indicates that
the RSM is preferred. Another problem with the RSM are difficulties with
convergence - it is harder to make a simulation converge when using RSM
compared to two equation models. This may be one of the reasons why many
CFD users prefer the two equations models.

36
Bibliography

Gouri K. Bhattacharyya and Richard A. Johnson. Statistical Concepts and

Methods. John Wiley & sons, 1977.

Dantec. BSA/FVA Flow Software, 4 ed. Publication nr. 9040U5714, 2000.

Ivar S. Ertesvåg. Utvikling av turbulensmodell for låge reynoldstal med likn-

ing for reynoldsspenningane og likning for karakteristisk frekvens. Doktor
ingeniøravhandling, Norges Tekniske Høgskole, 1991.

Ivar S. Ertesvåg. Turbulent strøyming og forbrenning. Tapir akademisk Forlag,

Trondheim, 2000.

K. Hanjalic. Advanced turbulence closures: Second-moment- and related mod-

els. von Karman Institute for Fluid Dynamics - Lecture series in Introduc-
tion to turbulence modelling, 2004.

J. O. Hinze. Turbulence, 2 edition. McGraw-Hill series in mechanical engi-

neering, 1975.

Erwin Kreyszig. Advanced Engineering Mathematics, 7 ed. John Wiley &

Sons, Inc, 1993.

Vidar Mathiesen. Lecture notes in process modelling 1. In norwegian, 2000.

Bruce R. Munson, Donald F. Young, and Theodore H. Okiishi. Fundamentals

of Fluid Mechanics, 4 edition. John Wiley & Sons, Inc, 2002.

Stephen B. Pope. Turbulent Flows. Cambridge University Press, 2000.

H. Tennekes and J. L. Lumley. A First Course In Turbulence. The MiT press,

1972.

David C. Wilcox. Turbulence modelling for CFD, 2 edition. DCW Industries,

Inc, 2002.

37
Appendix A

Turbulence Pictures

A mixture of soap/water is flowing over pin-shaped obstacles. When the

Reynolds number increases, parts of the flows becomes turbulent, and vor-
tices are created.

38
 Figure A.1: Flow at low Reynolds number

Figure A.2: A von Kárman street is created below the bolts at low Reynolds
number

Figure A.3: Flow velocity is increased and some parts in the flow becomes
turbulent

39
Figure A.4: Further increase in Reynolds number will make the flow turbulent,
and there is a “stretching” of vortices

Figure A.5: Close-up of figure A.4

40
Appendix B

Examination Questions

Question 1: Reynolds averaging

• Describe in general terms the main features of turbulence

• Define the Reynolds averaging process

• Perform Reynolds averaging on the Navier-Stokes equations

Question 2: Size and time scales 1

• Describe in general terms the main features of turbulence

• Explain the energy cascade concept

• Define the largest and smallest scales in turbulent flow

Question 3: Size and time scales 2

• Describe in general terms the main features of turbulence

• Explain the relationship between the mean flow Reynolds number and
the range of scales

Question 4: Size and time scales 3

• Describe in general terms the main features of turbulence

• Define and give examples of spatial and time correlations

Question 5: General turbulent flow

• Describe in general terms the main features of turbulence

• Define the regions of confined turbulent flow, and discuss physical aspects
of these

41
Question 6: Reynolds averaging and closure

• Describe in general terms the main features of turbulence

• Derive the Reynolds-Averaged Navier-Stokes (RANS) equations

Question 7: Turbulence models

• Describe in general terms the main features of turbulence

• There are different turbulence models to solve the new terms, −u0i u0j , to
close the RANS equations. What is the major difference between 0-, 1-,
2-equation models and Reynolds stress models?

Question 8: Viscous stress analogy

• Describe in general terms the main features of turbulence

• Dimension analysis show that the new terms −ρu0i u0j are also stresses,
same as the viscous stresses, τij . For incompressible Newtonian fluids,
the viscous stresses can be expressed as:
 
∂ui ∂uj
τij = τji = µ +
∂xj ∂xi

In direct analogy to the viscous stresses, how do you expect to model

the new terms −ρu0i u0j

Question 9: Boussinesq hypothesis (also known as “eddy viscosity approach”)

• Describe in general terms the main features of turbulence

• Boussinesq’s hypothesis reads:

 
∂ui ∂uj 2
−ρu0i u0j = ρνt + − ρkδij
∂xj ∂xi 3

The last term seems beyond the direct analogy to the viscous stresses.
Why does the hypothesis include this term?

Question 10: The closure problem

• Describe in general terms the main features of turbulence

• By introducing Boussinesq’s hypothesis, the closure problem of RANS

equations turns out to be the term νt . What is the common basis of the
mixing length (0-equation), turbulent kinetics model (1-equation) and
k − ε (2-equation) in solving νt ? What are their differences?

42