See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/310452765

Turbulence theory and modelling - Lecture notes

Technical Report · July 2014

DOI: 10.13140/RG.2.2.34953.24169

CITATION READS

1 7,459

1 author:

Bjørn H Hjertager
University of Stavanger (UiS)
187 PUBLICATIONS 6,359 CITATIONS

SEE PROFILE

All content following this page was uploaded by Bjørn H Hjertager on 17 November 2016.

The user has requested enhancement of the downloaded file.

Turbulence Theory and Modelling

by

Professor Bjørn H. Hjertager

Dept. of Mechanical and Structural Engineering and Material Science
University of Stavanger
N-4036 Stavanger, Norway

Lecture notes*

July, 2001
Revised July, 2014

*
With inputs from publications by: Davidson (2014), Ferziger and Peric (1995)
and Ranade (2002)
 2

Table of contents page

TURBULENT FLOWS 3
Introduction 3
Characteristics of turbulence 5
Turbulent scales 7
Energy spectrum 10
The cascade process created by vorticity 18
Classification of approaches to predicting turbulent flow 23

REYNOLDS AVERAGED NAVIER-STOKES EQ. (RANS) 25

Introduction 25
Averaging the Navier-Stokes equations 26
‘Eddy’ viscosity 30
Prandtl’s mixing length theory 31
Scalar variables 32
General turbulence models 33
The k- model 35
The  equation 37
Turbulent viscosity 38
Reynolds stresses 39
Influence of walls 41
Low-Reynolds number modelling 43
Models for flows with non-isotropic transport coefficients/
Reynolds stress models 44

LARGE EDDY SIMULATION (LES) 47

Introduction 47
Large eddy model equations 49
Smagorinski SGS model 52
Wall modifications 53
Other SGS models 54
RANS versus LES 56

DIRECT NUMERICAL SIMULATION (DNS) 59

LITERATURE 61
 3

TURBULENT FLOWS

Introduction

Almost all fluid flow which we encounter in daily life is turbulent. Typical
examples are flow 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 flow 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 fluid flow it will most likely be
turbulent. They are characterised by the following properties (Examples of
turbulent flows are given in Figures 1-3):

- Turbulent flows are highly unsteady

- They are three-dimensional

- They contain a great deal of vorticity

- Turbulence increases the rate at which conserved quantities are stirred

- By the processes mentioned above, turbulence brings fluids of

differing momentum into contact

- Turbulent flows fluctuate on a broad range of length and time scales

All the above properties are important.

The effects produced by turbulence may or may not be desirable, depending on

application:

- intense mixing is useful when chemical reaction or heat transfer is

needed

- on the other hand, increased mixing of momentum results in increased

frictional forces thus increasing the pumping power etc

Engineers need to be able to understand and predict these effects in order to

obtain a good design.
 4

Figure 1. Flow past cylinder (from Hinze, 1975)

Figure 2. Vortices in a mixing layer (from Banerjee, 1992)

Figure 3. Instantaneous PIV image (from http://pivnet.dm.go.dlr.de/PivNet)

 5

Characteristics of turbulence

In turbulent flow we usually divide the velocities in one time-averaged part ui ,

which is independent of time (when the mean flow is steady), and one
fluctuating part ui so that ui  ui  ui

There is no definition on turbulent flow, but it has a number of characteristic

features (see Pope (200) and Tennekes & Lumley (1972)) such as:

I-Irregularity. Turbulent flow is irregular and chaotic (they may seem random,
but they are governed by Navier-Stokes equation). The flow consists of a
spectrum of different scales (eddy sizes). We do not have any exact definition
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 flow 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 flow the diffusivity increases. 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) and heat transfer
in internal flows such as in channels and pipes.

III-Large Reynolds Numbers. Turbulent flow occurs at high Reynolds

number. For example, the transition to turbulent flow in pipes occurs that ReD ≃
2.300, and in boundary layers at Rex ≃ 500.000.

IV-Three-Dimensional. Turbulent flow is always three-dimensional and

unsteady. However, when the equations are time averaged, we can treat the flow
as two-dimensional (if the geometry is two-dimensional).

V-Dissipation. Turbulent flow is dissipative, which means that kinetic energy in

the small (dissipative) eddies are transformed into thermal energy. The small
 6

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 flow. 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 flow they are
much larger than the molecular scale and we can treat the flow as a continuum.

All the fluctuating properties of a turbulent flow contain energy across a wide
range of frequencies or wave numbers (κ =2f/U, where f is the frequency).

This is demonstrated in Figure 4, which gives the energy spectrum of

turbulence.

Log E(κ)
Universal equilibrium range

Viscous subrange
Energy
containing
range

Inertial subrange

κd log κ

Figure 4. Energy spectrum (cascade) of turbulence

The smallest eddies of motion in turbulent flow is dictated by viscosity.

In typical engineering flows these scales have length of the order 0.1 to 0.01 mm
and frequencies around 10 kHz.
 7

Turbulent scales

The largest scales are of the order of the flow geometry (the boundary layer
thickness, for example), with length scale ℓ0 and velocity scale u0. These scales
extract kinetic energy from the mean flow which has a time scale comparable to
the large scales (see Figure 5), i.e.

u1 u
 t01  0 (1)
x2 0

Part of the kinetic energy of the large scales is lost to slightly smaller scales with
which the large scales interact. Through the cascade process, 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 kinetic energy transferred
from eddy-to-eddy (from an eddy to a slightly smaller eddy) is the same per unit
time for each eddy size.

The dissipation is denoted by ε which is energy per unit time and unit mass (ε =
[m2/s3]). The dissipation is proportional to the kinematic viscosity, ν, times the
fluctuating velocity gradient up to the power of two. The friction forces exist of
course at all scales, but they are largest at the smallest eddies. In reality a small
fraction is dissipated at all scales. However, it is assumed that most of the
energy that goes into the large scales per unit time (say 90%) is finally
dissipated at the smallest (dissipative) scales.

The smallest scales where dissipation occurs are called the Kolmogorov scales
whose velocity scale is denoted by uη , length scale by ℓη and time scale by τη .
We assume that these scales are determined by viscosity, ν, and dissipation, ε.
The argument is as follows.

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.
 8

Figure 5: Cascade process with a spectrum of eddies. The energy-containing

eddies are denoted by u0; ℓ1 and ℓ2 denotes the size of the eddies in the inertial
sub range such that ℓ2 < ℓ1 < ℓ0; ℓη is the size of the dissipative eddies.(From
Davidson (2014))

dissipation: The amount of energy per unit time that is to be dissipated is ε. 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 uη , ℓη and τη in ν and ε using dimensional analysis.
We write

uη = νa εb

[m/s] = [m2/s] [m2/s3] (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, (3)
 9

and one for seconds [s]

−1 = −a − 3b, (4)

which give a = b = 1/4. In the same way we obtain the expressions for ℓη and τη
so that
1 1
1  3  4
  2
u    4
        (5)
   

The above analyses, called the Kolmogorov’s first hypothesis, states that for
large Reynolds numbers there is a range of high wave numbers where turbulence
is determined by the energy dissipation rate () and the kinematics viscosity ()
only. Hence, it is statistically independent of the large scales.

- This range is usually referred to as the viscous subrange. See Figure 4.

- The inertial subrange together with the viscous subrange forms the

- Universal equilibrium range.

 10

Energy spectrum

As mentioned above, the turbulence fluctuations are composed of a wide range

of scales. We can think of them as eddies, see Fig. 5. It turns out that it is often
convenient to use Fourier series to analyze turbulence. In general, any periodic
function, g, with a period of 2L (i.e. g(x) = g(x + 2L)), can be expressed as a
Fourier series, i.e.

1
g ( x)  a0    an cos  n x   bn sin  n x   (6)
2 n 1

where x is a spatial coordinate and

n 2
n  or  = (7)
L L
Variable κn is called the wavenumber. The Fourier coefficients are given by
L
1
an   g ( x) cos  n x  dx
L L
L
1
bn   g ( x) sin  n x  dx
L L

Parseval’s formula states that

L 
L 2
 g 2  x  dx  a0  L  an2  bn2  (8)
L
2 n 1

Let now g be a fluctuating velocity component, say u1' . The left side of Eq. 8
expresses u1' 2 in physical space (vs. x) and the right side u1' 2 in wavenumber space
(vs. κn). The reader, who is not familiar to the term “wavenumber”, is probably
more familiar to “frequency”. In that case, express g in Eq. 6 as a series in time
rather than in space. In this case the left side of Eq. 8 expresses u1' 2 as a function
of time and the right side expresses u1' 2 as a function of frequency.
 11

The turbulent scales are distributed over a range of scales which extends from
the largest scales which interact with the mean flow to the smallest scales where
dissipation occurs, see Fig. 5. Now let us think about how the kinetic energy of
the eddies varies with eddy size. Intuitively we assume that large eddies have
large fluctuating velocities which implies large kinetic energy, u1' u1' 2 . It is now
convenient to study the kinetic energy of each eddy size in wavenumber space.
In wavenumber space the energy of eddies can be expressed as

E(κ)dκ (9)

where Eq. 9 expresses the contribution from the scales with wavenumber
between κ and κ + dκ to the turbulent kinetic energy k. The energy, E(κ),
corresponds to g2(κ) in Eq. 8. The dimension of wavenumber is one over length;
thus we can think of wavenumber as proportional to the inverse of an eddy’s
diameter, i.e κ ∝ 1/d. The total turbulent kinetic energy is obtained by
integrating over the whole wavenumber space i.e.

k   E ( ) d   L  g 2 ( n ) (10)
0

Think of this equation as a way to compute the kinetic energy by first sorting all
eddies by size (i.e. wavenumber), then computing the kinetic energy of each
eddy size (i.e. E(κ)dκ), and finally summing the kinetic energy of all eddy sizes
(i.e. carrying out the integration). Note that the physical meaning of E is kinetic
energy per unit wavenumber of eddies of size ℓκ ∝ κ-1. Hence the dimension of
E is u1' 2 /κ, see Eq. 10.

The kinetic energy is the sum of the kinetic energy of the three fluctuating
velocity components, i.e.

k
1 '2
2
  1
u1  u2'2  u3'2  ui' ui'
2
(11)
 12

Figure 6: Spectrum for turbulent kinetic energy, k. I: Range for the large,
energy containing eddies. II: the inertial subrange. III: Range for small,
isotropic scales. The wavenumber, κ, is proportional to the inverse of the length
scale of a turbulent eddy, ℓκ, i.e. κ ∝ ℓk -1. (From Davidson (2014))

The spectrum of E is shown in Fig. 6. We find region I, II and III which

correspond to:

I. Large energy containing eddies. In this region we have the large eddies which
carry most of the energy. These eddies interact with the mean flow and extract
energy from the mean flow. This energy transfer takes places via the production
term, Pk , in the transport equation for turbulent kinetic energy, see Eq. 20. The
energy extracted per unit time by the largest eddies is transferred (per unit time)
to slightly smaller scales. The eddies’ velocity and length scales are u0 and ℓ0,
respectively.

III. DissipationIviscous 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 ε in the transport
equation for turbulent kinetic energy, see Eq. 20. The scales of the eddies are
described by the Kolmogorov scales (see Eq. 5)
 13

II. Inertial subrange. The existence of this region requires that the Reynolds
number is high (fully turbulent flow). The eddies in this region represent the
mid-region. The turbulence is also in this region isotropic. This region is a
“transport region” (i.e. in wavenumber space) in the cascade process. The
“transport” in wavenumber space is called spectral transfer. Energy per time
unit, Pk = ε, is coming from the large eddies at the lower part of this range and is
transferred per unit time to the dissipation range at the higher part. Note that the
relation Pk = {dissipation at small scales}, see Fig. 6, is given by the assumption
of the cascade process, i.e. that the energy transfer per unit time from eddy-size–
to–eddy-size is the same for all eddy sizes.

The kinetic energy, k  u' ,i u' ,i / 2 , of an eddy of size (length scale), 1/κ,
represents the kinetic energy of all eddies of this size. The kinetic energy of all
eddies (of all size) is computed by Eq. 11. 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 per unit time (ε) and the size of
the eddies, 1/κ. Dimensional analysis gives

E = κa εb

[m3/s2] = [1/m] [m2/s3] (12)

We get two equations, one for meters [m]

3 = ≗ a + 2b,

and one for seconds [s]

≗ 2 = ≗ 3b,

so that b = 2/3 and a = ≗ 5/3. Inserted in Eq. 12 we get

E(κ) = CK ε 2/3 κ-5/3 (13)

where the Kolmogorov constant CK ≃ 1.5. This is a very important law

(Kolmogorov spectrum law or the ≗ 5/3 law) which states that, if the flow is
fully turbulent (high Reynolds number), the energy spectra should exhibit a
≗ 5/3 - decay in the inertial region (region II, Fig. 6).
 14

Above we state that the eddies in Region II and III are isotropic. This means that
– in average – the eddies have no preferred direction, i.e. the fluctuations in all
directions are the same so that u1  u2  u3 . Note that is not true
'2 '2 '2

instantaneously, i.e. in general u1  u2  u3 . Furthermore, isotropic turbulence

' ' '

implies that if a coordinate direction is switched, nothing should be changed.

' '
For example if the x1 coordinate direction is rotated 180o the u1 u2 should
remain the same, i.e. u1' u2'  u1' u2' . This is possible only if u1' u2'  0 . Hence, all
shear stresses are zero in isotropic turbulence.

Using our knowledge in tensor notation, we know that an isotropic tensor can be
written as const x δij . Hence, the Reynolds stress tensor for small scales can be
written as u1' u2'  const x δij which, again, shows us that the shear stresses are
zero in isotropic turbulence.

As discussed on p. 9, the concept of the cascade process assumes that the energy
extracted per unit time by the large turbulent eddies is transferred (per unit time)
by non-linear interactions through the inertial range to the dissipative range
where the kinetic energy is transformed (per unit time) to thermal energy
(increased temperature). The spectral transfer rate of kinetic energy from eddies
of size 1/κ to slightly smaller eddies can be estimated as follows. An eddy loses
(part of) its kinetic energy during one revolution. The kinetic energy of the eddy
is proportional to uκ2 and the time for one revolution is proportional to ℓκ/uκ.
Hence, the energy spectral transfer rate, εκ, for an eddy of length scale 1/κ can be
estimated as (see Fig. 6)

u2 u3
  
(14)
k k
uk

Kinetic energy is transferred per unit time to smaller and smaller eddies until the
transfer takes place by dissipation (i.e. increased temperature) at the
Kolmogorov scales. In the inertial subrange, the cascade process assumes that εκ
= ε. Applying Eq. 14 for the large energy-containing eddies gives
 15

u02 u03
0    k  
0 0 (15)
u0

The dissipation at small scales (large wavenumbers) is determined by how much

energy per unit time enters the cascade process at the large scales (small
wavenumbers). We can now estimate the ratio between the large eddies (with u0
and ℓ0) to the Kolmogorov eddies (uη and ℓη). Equations 5 and 15 give

u0
   u0   u03  0  u0   u03  0u0 4    u0  0    Re 0 4
1 4 1 4 1 4 14 1

u
1 1 1 1
 0  3  4
 3  4
 3  4
 3  4
3
  0   3  0   3  0 4   3 3   Re 0 4
     u0  0   u0  0   u0  0  (16)
1 1 1 1
 0   2    2
u3 2
 0  u0 3  0 2  2
1
 0   3  0   0     Re 0 2
      u0  0    0  u0    0 u0 2 

where Re  u0 0  . We find that the ratio of the velocity, length and time scales
0

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. Hence, 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.
 16

To quantify the turbulence further scales can be defined. Consider the spatial
auto-correlation coefficient in the ‘1’ direction, plotted in Figure 7.

f()
1.0


0
 l0

Figure 7. Auto-correlation coefficient

u'1 ( x1 )  u '1 ( x1  1 )
f ( ) 
u'1 ( x1 )  u '1 ( x1 )
(17)

Here (-) denotes time averaging and u'1  u1  u1 is the velocity fluctuation
around the mean value. By integration of the auto-correlation one obtains the
integral scale l, which is a measure of the largest energy containing eddies
present in the flow. Hence:

l0   f ( )d 
0

(18)
Furthermore one may define a characteristic eddy size for the turbulence called
the Taylor micro scale. This scale is obtained by a series expansion of the auto-
correlation coefficient around =0:

2
1  
d2 f
d 2
(19)
If we consider isotropic turbulence we may estimate the size of 1 from the
 17

turbulent energy budget. For equilibrium turbulent flow we may assume that
production of turbulence (P) equals dissipation of turbulence ():

 u u   u ' u' j   u'i u ' j 

P  u'i u' j  i  j     i  
    

 x j xi   x j xi   x j xi 
(20)
which states that production of energy by the turbulent stresses is balanced by
the rate of viscous dissipation. The dissipation may in isotropic turbulence be
written by   15 u '1 / x1  . By defining the Taylor microscale (  ) as:
2

 u'1  u '1 
2 2

   2
 x1  
(21)
we will obtain:

u2
  15
2
(22)
where u  u'1  . Since the production term P is associated with the integral
2 2

scale, l0, i.e. P=A u3/l0. From this we may deduce:

1 1
0  A 2  1
   Re 2 and  Re 2
  15  0 0 0
(23)
where Rel=u l0/ is the turbulent Reynolds number.

From Eq. 16 we have also deduced the relation:

0 3  3
 Re 0 4 and   Re 0 4 (24)
 0

These relations show the difference between the large scales and the microscales
increases when the Reynolds number is increased. This will be used in
determining the resolution and simulation times needed in turbulence modeling
to be discussed later.
 18

The cascade process created by vorticity

The interaction between vorticity and velocity gradients is an essential

ingredient to create and maintain turbulence. Disturbances are amplified by
interaction between the vorticity vector and the velocity gradients; the
disturbances are turned into chaotic, three-dimensional fluctuations, i.e. into
turbulence. Two idealized phenomena in this interaction process can be
identified: vortex stretching and vortex tilting.

The equation for the instantaneous vorticity ( i  i  i ' ) may be deduced by

taking the curl of the Navier-Stokes equations and reads

i i ui  2i

uj  j 
 t x j x j x j x j
unstady       
convection stretching / tilting viscous / diffusion

 u3 u2
 
 2 x3
x
u  u u
i   ijk k   1  3
x j  x3 x1
 u u
 2 1
 x1 x2
 u1 u u
 1  2 1  3 1
 x1 x2 x3
ui  u2 u u
j   1  2 2  3 2
x  x1 x2 x3
j
stretching / tilting  u u u (25)
 1 3  2 3  3 3
 x1 x2 x3
 19

This equation is not an ordinary convection-diffusion equation: it has an

additional term on the right side which represents amplification and
rotation/tilting of the vorticity lines (the first term on the right side).

The i = j components of this term represent (vortex stretching. A positive ∂u1/∂x1

will stretch a cylinder and from the requirement that the volume must not change
(incompressible continuity equation) we find that 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. Thus we can assume that there are no viscous
stresses acting on the cylindrical fluid element surface which means that the
angular momentum

r2ω1 = const. (26)

remains constant as the radius of the fluid element decreases. Equation 26 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 fluid element and an in- crease of the vorticity
component (i.e. the tangential velocity component) aligned with the element. For
example, an extension of a fluid element in one direction (x1 direction) decreases
the length scales in the x2 direction and increases ωi′, see Fig. 8.

u1'
Figure 8: A fluid element is stretched by  0 . Its radius decreases (from
x1
dashed line to solid line). (From Davidson (2014))
 20

Figure 9: The rotation rate of the fluid element (black circles) in Fig. 8
u2'
increases and its radius decreases. This creates a positive  0 which
x2
stretches the small red fluid element aligned in the x2 direction and increases
ω2′. The radius of the red fluid element decreases. (From Davidson (2014))

At the same time, vortex tilting creates small-scale vorticity in the x2 and x3
direction, ω2′ and ω3′. The increased ω1′ means that the velocity scale in the x2
direction is increased, see Fig. 9. The increased u2′ velocity component will
stretch smaller fluid elements aligned in the x2 direction, see Fig. 9. This will
increase their vorticity ω2′ and decrease its radius, r2. In the same way will the
increased ω1′ also stretch a fluid element aligned in the x3 direction and increase
ω3′ and decrease r3. At each stage, the length scale of the eddies – whose
velocity scale are increased – decreases. Figure 10 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. 6); 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 “don’t 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. 10 and Table 1 The large original eddy (1st generation) is
 21

aligned in the x1 direction. It creates 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.

Figure 10: 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. (From Davidson (2014))

Table 1: Number of eddies at each generation with their axis aligned in the x1,
x2 or x3 direction, see Fig. 10. (From Davidson (2014))
 22

The i ≠ j components in the first term on the right side in Eq. 25 represent vortex
tilting. Again, take a slender fluid element, now with its axis aligned with the x2
axis. The velocity gradient ∂u1/∂x2 will tilt the fluid element so that it rotates in
the clock-wise direction. As a result, the second term ω2 ∂u1/∂x2 in line one in
Eq. 25 gives a contribution to ω1. This shows how vorticity in one direction is
transferred to the other two directions through vortex tilting.

Vortex stretching and vortex tilting qualitatively explain how interaction

between vorticity and velocity gradient create vorticity in all three coordinate
directions from a disturbance which initially was well defined in one coordinate
direction. Once this process has started it continues, because vorticity generated
by vortex stretching and vortex tilting interacts with the velocity field and
creates further vorticity and so on. The vorticity and velocity field becomes
chaotic and three-dimensional: turbulence has been created. The turbulence is
also maintained by these processes.

From the discussion above we can now understand why turbulence always must
be three-dimensional (Item IV on p. 5). If the instantaneous flow is two-
dimensional (x1 − x2 plane) we find that the vortex-stretching/tilting term on the
right side of Eq. 25 vanishes because the vorticity vector and the velocity vector
are orthogonal. The only non-zero component of vorticity vector is ω3 because

u3 u2
1   0
x2 x3
u1 u3
2   0
x3 x1

Since u3 =0, we get  j ui x j  0 .

 23

Classification of approaches to predicting turbulent flows

Use of correlations

- Friction factors as function of Reynolds number or

- Nusselt number of heat transfer as function of Reynolds number and
Prandtl number
- Sherwood number of mass transfer as function of Reynolds number
and Schmidt number
- Only applicable to simple flows
- No large computers needed

Averaging of the equations of motion over time (RANS)

- The so-called one-point closure leads to a set of partial differential

equations called the Reynolds averaged Navier-Stokes ( RANS)
equations
- These equations does not lead to a closed set of equations
- Require introduction of approximations referred to as turbulence
models

Large eddy simulation (LES)

- Solves for the largest scale of the fluid motion

- Models are only needed for the small scale motion
- Compromise between one-point methods and direct numerical
simulation (see below)

Direct numerical simulation (DNS)

- Solves the Navier-Stokes equations for all motions in the turbulent

flow
- Needs very fine numerical resolution

Going down the list more and more of the turbulent motion are computed and
fewer are approximated.
The methods at the bottom are more exact but computation time is increased
considerably.
Figure 11 gives an overview of the various modelling concepts.
 24

Zero-Equation Models

One-Equation Models

Include Two-Equation Models Increased

More Computational
Standard k- Cost
Physic RNG k- Per Iteration
Reynolds-Stress Model
Algebraic
Differential RANS-based
models

Large-Eddy Simulation

Direct Numerical Simulation

Figure 11: Modelling Approaches for Turbulent Flows (from V.Ranade, 2002)
 25

REYNOLDS AVERAGED NAVIER-STOKES EQUATIONS (RANS)

Introduction

Transient flow in a point:

u’
u

Time, t
T
Figure 12. Schematic description of velocity variation as function of time.

The following may be defined:

u = u + u

T
t+
1 2
u=
T
 u  dt
T
t-
2

u = 0 and u 2  0
(26)
 26

Averaging of the Navier-Stokes equations

The influence of turbulence on the mass and momentum equations (Navier-

Stokes) may be found if we look at the x-direction, steady state form:

Mass balance.

u +  v +  w = 0
x y z
(27)

Momentum balance; x-direction:

   p  2 u 2 u 2 u 
 u  u  +  v  u  +  w  u  = - +    2 + 2 + 2 
x y z x x  y  z 
(28)

Define:
u = u + u and v = v + v 

w = w + w and p = p + p

and  is assumed constant

(29)
 27

Put this into (27):


 ( u + u)+   ( v + v)+   ( w + w) = 0
x y z
(30)

Time average the equation and get:


  u  +   v  +   w  = 0
x y z
(31)
As we may observe this is the same form as (27), where the instantaneous
velocity is exchanged by the time averaged velocity.

If we insert (29) into equation (28) we get:


 ( u + u)  ( u + u)+   ( v + v)  ( u + u)+   ( w + w)  ( u + u)=
x y z

( p + p)  2 ( u + u)  2 ( u + u)  2 ( u + u) 

- +    + +
x   z 2 
2 2
 x y
(32)
Multiply left side and get:

 ( u2 + 2uu + u2 )+   ( vu + vu + vu + vu)+   ( wu + wu + wu + wu)=
x y z

( p + p)   2 ( u + u)  2 ( u + u)  2 ( u + u) 

- +  + +
x   z 2 
2 2
 x y
(33)
 28

Time average the equation and obtain:


 u2 +  u2 +   v  u +  vu+   w  u +  wu =
x y z

p  2u u u
+    2 +  2 +  2 
2 2

-
x  x  y z 
(34)
Rearrange the equation so that terms with fluctuations are on the right hand side:

 u 2  +    v  u  +   w  u  =
x y z

p     2u u u
+ (-  u2 ) + (-  v u ) + (-  wu ) +     2 +  2 +  2 
2 2

-
x x y z  x  y z 
(35)
The new terms on the right hand side are often named apparent stresses or
Reynolds stresses.

Generally for all the three momentum equation the stresses may be defined as:

 ij = -  ui u j
(36)
also denoted as the stress tensor.
 29

On matrix form it may be written as:

  u2  v u  wu

 
 ij = -   uv   v   wv 
2

 
  uw  v w  w2 

(37)
Prior to solving equation (35) the stress tensor must be modelled.
Relating the Reynolds stresses to the time averaged velocity may do this.
These are done with the so-called turbulence models.
 30

“Eddy” viscosity

Boussinesq suggested:

du
 u v  = -  t
dy
(38)

Generally for 3D flow:

 u u  2
  ui' u 'j   t  i  j    ij    k
 x j xi  3
where  ij  1 when i  j and  ij  0 otherwise

(39)
Here the turbulent kinetic energy is introduced, defined as:

1 2
k=  u + v2 + w2 
2
(40)
 31

Prandtl’s mixing length theory

Modelling of the expression (38) above are done by assuming the following:

u
du
u = v  = l 
dy du
dy
(41)

Insert in (38) and get: l l

du du
 uv  = -  l 2 y
dy dy
(42) Figure 13. Velocity profile.

“Eddy” viscosity becomes:

du
t =  l2
dy
(43)

Generally we can see that the “eddy” viscosity or turbulent viscosity may be
written as:

 t =   ut  l
(44)
Which means that turbulent viscosity the product of density, typical turbulence
velocity and typical length scale for the turbulence.
 32

Scalar variables

For a scalar variables, , (enthalpy, concentration, etc) will we obtain additional

terms in conservation equations of the form (      ' ):

-  u'i 
(45)
These terms are modelled analogous as for the Reynolds stresses:


-  u'i  =  ,t
xi

(46)
The turbulent transport coefficient, ,t, is related to the turbulent viscosity, t:

t
 ,t =

(47)
Here the effective turbulent Prandtl/Schmidt number ,, is introduced.
If we assume a simple mixing length theory we may write:

du
 l2
dy
 ,t =

(48)
The Prandtl mixing length theory may be denoted a zero-equation turbulence
model.
 33

General turbulence models

More general turbulence models starts with the expression for turbulent
viscosity:

 t =   ut  l
(49)
As we observe there are two variables, in addition to the density that determines
the turbulent viscosity, namely ut and l. This has resulted in development of the
so-called "two-variable or two-equation" models. This means that differential
equations for the two variables may be written that can be related to ut and l.

The turbulence velocity ut may be related to the turbulent kinetic energy as:

2 1 2
ut = k= ( u  + v 2 + w2 )
3 3
(50)
A differential equation may be put up for turbulent kinetic energy, k. This
equation may be deduced from the Navier-Stokes equations.
 34

Equations suitable for determining the length scale, l, are proposed by several.
The suggested variables are:

3
k k k2
(kl)= k  l , W = 2 ,   and  =
l l l

(51)
Where W is the vorticity fluctuations squared, ω is the vorticity fluctuations
(frequency) and  is the dissipation of turbulent kinetic energy.

Experience with these different variables has shown that it is the formulation
based on the variable, , that is the most robust. I.e. gives the best result over a
large range of slow cases.
 35

The k -  model

The variables are:

1) turbulent kinetic energy:

1 1
k = ui  ui = ( u2 + v 2 + w2 )
2 2
(52)
and 2) dissipation of turbulent kinetic energy:

 u'i u ' j
=  
 xk xk
(53)
The differential equations for both k and  may be deduced from the Navier-
Stokes equations. This is a comprehensive task and we will only refer to the
result for the k-equation. We now introduce notation Ui = ui and find:

  '  ui u j 
' '
k  ' ' U i ui' ui'
+  U i k  =  ui   
+ p  -  ui u j

-
t  xi  xi   2   x  xj  xj
 j

I II III IV V
(54)
Here:
I - transient term
II - convective term
III- diffusive transport (MUST BE MODELLED)
IV - production of turbulent kinetic energy from the mean flow
V - viscous dissipation of turbulent kinetic energy (MUST BE
MODELLED)
 36

Modelled form of the k - equation:

III is replaced with:

   t k 
 xi  k  xi 
(55)
Which means similar type of assumption as in the Boussinesq approximation.
I.e. gradient diffusion.

V is replaced with:

 
(56)

Final form for the k-equation:

k  
+

 U i k  =    t k  -  ui'u 'j  U i -   
t xi xi  k xi  x j

(57)
where k is a constant turbulent Schmidt number for k.
 37

he  - equation

 ui' u j
'

=  
 xk xk
(58)
An exact transport equation for  may be deduced from the Navier-Stokes
equations. It will be outside our scope to deduce this and to refer the exact form.
We will only give the modelled form:

    t    U i  
+  U i   =   x  + C1 k -  u'i u' j x  - C 2  k 
t xi xi   i   j 

I II III IV V
(59)

Here:
I - transient term
II - convective term
III- diffusive transport
IV - production of 
V - destruction of 
, C1 and C2 are constants
 38

Turbulent viscosity

From previously we have:

 t =   ut  l

2
ut = k
3

k
2
from dimensional considerations
l
(60)

This means:

l = k and
2

3
2

k = k
1 2
t    k 2 further
 

t = C k

(61)

C is a constant.
 39

Reynolds stresses

 U i
U j  2  U k 
 ijt    ui'u 'j  t     ij   t
   k
 x j xi  3  xk 
where  ij  1 when i  j and  ij  0 otherwise

(62)
Newton’s law of friction for laminar flow:

 U i U j  2 U k
 ij =  lam   -  ij   lam
lam
+
 x j xi  3 xk

(63)

If these are combined we obtain:

 U i U j  2  U k 
 ij =  ij +  ij =  eff   -  ij    eff + k 
eff t lam
+
 x j xi  3  xk 
(64)

Where eff = t + lam

A popular alternative to the standard k -  given above is the so-called RNG k -
 model based on the Re-Normalisation Group theory. The main difference is
the determination of the various constants (C, C1, C2, k,  and ). Details of
this model may be found in the book of e.g. Versteeg & Malalasekera.
 40

Compilation of the constants for both k -  and RNG k -  models that gives "the
best" agreement with experimental data:

Constant/ Standard k- RNG k-

Parameter
C 0.09 0.0845
C1 1.44 1.42
C2 1.92   G

C 3 1  k
 4.38 t
1.68    
1 0.0123 

k 1.0 ~ 0.7179 (high Re limit)*

 1.3 ~ 0.7179 (high Re limit) *

G is the production of turbulent kinetic energy.

* General expression for estimating effective Prandtl numbers for k and  is:

0.6321 0.3679
1 1
 1.3929  2.3929
  

0.3929 3.3929 T

Table 2. Parameters of k- Model (from Ranade, 2002)

 41

Influence of walls

Close to the wall it is assumed that "the universal velocity-profile” applies.

P
yP
w

qw

Figure 14. Conditions close to a wall.

This has the form as shown in the figure below:

u+
linear
scale

u+=1/ ln y+ +B (fully turbulent)

u+= y+ (laminar)

y+ (log. scale)
Figure 15. The universal velocity profile.
The following definitions apply for the boundary layer close to the wall:

where u =  w
u
u 
+

u 
(65)
  u  y
y 
+


 42

Based on the assumptions about the boundary layers the shear stress w and the
heat flux qw at the wall may be related to the variables for point close to the wall
(Figure 14 and 15):

  C  2  k P   uP
1 2

 
w= 1

 ln E  y P

+

(66)
and

  C  2  k P   c p  T P - T w 
1 2

qw =  
1

  h  ln E  y P + P

+

(67)
Here are:

  h,lam 0,75 
P = 9.24    - 1.0 
  h  
(68)
Where , h,lam, h and E are constants; cp is the specific heat and T is the
temperature.
For the k and  equations the wall law will also give relevant boundary
conditions at the wall.
 43

Low-Reynolds number modelling

The k- model is mainly applicable for large Reynolds numbers. If the Reynolds
number is small the k- model may be modified with following terms on the
right hand side of the equations:

2
  k 12 
k: - 2.0   lam   
  xj 
 

 2 U i 
: - 2.0   lam   t   
  x j  x j 
(69)

Further the constants C and C2 are functions of the local Reynolds number:

 
 2.5 
C  = C  ,  e
- 
 1.0+ Ret 
 50 


C 2 = C 2,  1.0 - 0.3  e Re
- t
2

where

t   k2
Ret = =
 lam  lam  
(70)
C, and C2, are the constants for large Reynolds numbers, given before.
 44

Models for flow with non-isotropic transport coefficients/

Reynolds stress models

In some flow cases we can not assume isotropic viscosity:

 U i U j  2
 ijt    ui' u 'j   t      ij    k
 x j xi  3

(71)
These flows may be:
1) flow in a channel with square sectioned cross-section,
2) flow with strong rotation, etc.

In these situations transport equations for the unknown correlations in equation

(58) may be derived (Reynolds stress model (RSM)):

   U j U i 
t
 
  u 'iu ' j +
 xi
 
  U i  u 'iu ' j = D ij -   u 'iu 'k 
  xk
+ u ' j u 'k 
  +  ij   ij
xk 
(72)
Here the terms on the left-hand side are: transient and convection of the
correlations, whereas the terms on the right hand side are:

1) diffusion,
2) production,
3) redistribution of the correlations and
4) dissipation .
 45

A simplification of the RSM model is the so-called algebraic shear stress model
(ASM).
The Algebraic Reynolds Stress Model is a simplified Reynolds Stress Model.
The RSM and k − ε models are written in symbolic form as:

RSM : Cij − Dij = Pij + Φij − εij

k − ε : C k − Dk = P k – ε (73)

In ASM we assume that the transport (convective and diffusive) of u 'iu ' j is
related to that of k, i.e.

Cij – Dij = u 'iu ' j /k (P k – ε) (74)

Inserting equation based on the assumption in Eq. 60:

Ck – Dk = u 'iu ' j /k (P k – ε) (75)

Hence we may determine the shear stress as:

  
 u'i u' j = f  u' p u 'q , k,  , U l 
  xm 
(76)

The table on the following page gives an overview over advantages and
disadvantages for Two-equation models and Reynolds stress models.
 46

Advantages Disadvantages
Standard k-  Simplest model to  More expensive than zero
represent variation of equation models
turbulence length and  Assumes isotropic eddy
velocity scales viscosity
 Robust & economical  Performs poorly for:
 Excellent performance + some unconfined flows
for many industrial flows + rotating flows
 The most widely + non-circular ducts
validated model + curved boundary layers
RNG k-  Performs better than  Assumes isotropic eddy
standard model for viscosity
some:  Not sufficiently validated
+ separated flows so far
+ swirling flows
Algebraic Stress  Accounts for anisotropy  Restricted to flows where
Models (ASM)  Combines generality of convection and diffusion
approach with the terms are negligible
economy of the k-  Performs as poorly as k-
model in some flows due to
 Good performance for problems with  equation
isothermal and buoyant  Not widely validated
thin shear layers
Reynolds Stress  Most general model of  Computationally
Models (RSM) all classical turbulence expensive (seven extra
models PDEs)
 Performs well for many  Performs as poorly as k-
complex flows including in some flows due to
non-circular ducts and problems with  equation
curved flows  Not widely validated

Table 3: Summary of Two-equation and Reynolds Stress Models

(from Ranade, 2002)
 47

LARGE EDDY SIMULATION (LES)

Introduction

As we have seen turbulent flows contain a wide range of length and time scales.
The range of eddy sizes that might be found in turbulent flow is shown
schematically on the left-hand side if Figure 16 below.

The figure shows a typical velocity component in a point in the flow. The range
of scales on which fluctuations occur is obvious.

RANS

Figure 16. Schematic Representation of Scales in Turbulent Flows and Their

Relationship with Modelling Approaches (Adapted from Fergizer and Peric,
1995 by Ranade, 2001)

Figure 16 also gives the relationships between the three modeling approaches
RANS, LES and DNS models.
 48

The large-scale motion are generally more energetic than the small scales.
- their size and strength make them by far the most effective transporters
of the conserved properties.

The small scales are usually much weaker and provide little transport of these
properties.

A simulation that treats the large eddies more exactly then the small ones make
sense  LES is such an approach.

Large eddy simulations are:

- three-dimensional,
- time-dependent
- expensive
- but less expensive than Direct Numerical Simulations (DNS)

Generally, DNS is, because of accuracy, the preferred method if feasible.

LES is the preferred method for flows in which Reynolds number is too high or
the geometry is to complex to allow application of DNS.
 49

Large eddy simulation model equations

It is essential to define the quantities to be computed precisely. We need a

velocity field that contains only the large-scale components of the total field.
Filtering does this:
- the large or resolved scale field is the local average of the complete
field

We use one-dimensional notation and generalisation to 3-D is straightforward.

The filtered velocity is defined by:
ui ( x )   G ( x, x ' )ui ( x ' )dx '

(77)
where G(x,x’) is a localised filter. Filters used in LES include:
1/ 2 2

6 1  6 ( x  x ´)
i i

- Gaussian: G ( xi  xi ´)    e ,
2

  
- box filter (simple local average):
1 
G ( xi  x´i )  if | xi  xi ´|  and 0 otherwise
 2
- and a cut-off (filter that eliminates all Fourier coefficients belonging to
wave-numbers above a cut-off)

Every filter has a length scale , associated with it.

In a rough sense it may be stated that eddies with a size larger than  are large-
scale eddies whereas those smaller than  is the small eddies that need to be
modelled.
 50

When Navier-Stokes equations for constant density are filtered, we obtain a set
of equations very similar to the RANS equations given above:

 (  ui )  (  u j ui ) p     ui  u j 
       

t x j xi x j   j x x i 
(78)
Since the continuity equation is linear filtering does not change its form:

 (  ui )
0
xi
(79)
It is important to note that since

ui u j  ui u j

(80)
and the quantity on the left side is not easy to compute, a modelling
approximation for the difference between the two sides of this inequality:

 ijs    (ui u j  ui u j )
(81)
must be introduced.
 51

When (81) is introduced into (78) we obtain the model momentum equation for
the resolved scale flow:

 (  u )  (  u j ui )  S    u  u 
i   p ij     i  j 
  
t x x x x   x x 
j i j j   j i 

(82)

In the context of LES, s is called the sub-grid scale (SGS) Reynolds stress.
ij
This stress is in fact the large-scale momentum flux caused by the action of the
small or unresolved scales.

The models used to approximate the SGS Reynolds stress (Eq. 81) are called
sub-grid scale (SGS) models. Although the name SGS indicates that the filter
width  is related to the grid size, this is not generally true. However, it has in
practise become usual to fix the filter width to the grid size.
 52

Smagorinski SGS model

The most commonly used sub-grid scale model is the one proposed by
Smogarinski (1963):
- it is an eddy viscosity model
- the effect of the SGS Reynolds stress are increased transport and
dissipation
- as these processes are due to viscosity in laminar flows it seem
reasonable to assume the following model:

1   ui u j 
 ijs   kks  ij   t     2  t Sij
3  x j xi 

(83)
where t is the SGS eddy viscosity, ij is the Kroenecker delta defined before
and Sij is the strain rate of the large scale or resolved field.

The form of the sub-grid scale eddy viscosity may be derived by dimensional
arguments as:
 t  C S2 2 | S |
(84)
where CS is a model parameter to be determined,  is the filter length scale and
| S | ( Sij Sij )1 / 2 . This form of the equation may be derived in a number of ways.

Theories provide estimates for the constant CS and a value for CS = 0.2 is found.
Unfortunately, CS is not constant and may be a function of Reynolds number and
other non-dimensional parameters.
The filter width  in the Smogarinski model is put equal to the grid size.
 53

Wall modifications

The Smogarinski model, although relatively successful, is not without

problems. For example, to simulate channel flows several modifications are
required:
- CS in the bulk flow has to be lowered from 0.2 to 0.065
- This changes the viscosity by almost an order of magnitude
- Changes of this magnitude is also needed for all shear flows
- In regions close to the walls of the channel the CS value has to be reduced
further
- One successful recipe is the van Driest damping that has been used to reduce
the near-wall eddy viscosity in RANS models:


Cs  Cs 0 1  e  y

/ A

2

(85)
Here y+ is the distance from the wall in viscous units defined as:
y  u
y 

(86)
where u is the shear velocity defined as:
1/ 2
 
u   w 

(87)
Here w is the shear stress at the wall and A+ is a constant usually taken to be
approx. 25.
 54

Other LES SGS models

If we wish to simulate more complex and/or higher Reynolds number flows it

may be important to have a more accurate model.

The scale similarity model

The small scales resolved in a simulation are in many ways similar to the still
smaller scales that are treated via the SGS model. This idea leads to an
alternative sub-grid scale model denoted the scale similarity model.
The main argument is that the important interactions between the resolved and
unresolved scales involve the smallest eddies of the former and the largest of
the latter, i.e. eddies that are a little larger or a little smaller than the length scale
 associated with the filter. Arguments based on this concept lead to the
following model:
 ijs    (ui u j  ui u j )
(88)
where the double overline indicates a quantity that has been filtered twice. This
model correlates well with the actual SGS Reynolds stress, but hardly dissipates
any energy. It transfers energy from the smallest resolved scales to the larger
scales. To correct for lack of dissipation it is necessary to combine the
Smagorinski and scale similarity models to produce a mixed model with the
following expression for the SGS shear stress:
1
 ijs   kk ij   (ui u j  ui u j )  2CS2 2 | S | Sij
3
(89)
 55
Dynamic models
One major drawback of the above scale similarity model is the need for ad hoc
setting of the model parameter CS. To avoid this one may use the dynamic
approach in which the magnitudes of the model parameters are based on local
conditions and are recalculated during the simulation. Dynamic models are
based on the assumption that SGS stresses exhibit asymptotic behaviour as the
filter size is reduced. Therefore by using the smallest resolved scales one may
calculate the model parameters. A second filtering which is broader than  is
used and the Sub Test Filter Scale (STS) stress tensor may be calculated. The
resolved SGS stresses may also be calculated from (89). By comparing these
two stresses we may test the quality of the model in a direct way. Or more
importantly may compute the model parameters. This can be done at every
point in space and at every time step. The value can then be applied to the SGS
model of the large-scale simulation. In this way some kind of self-consistent
SGS model is produced.

The dynamic procedure removes many of the difficulties described earlier:

- In shear flows the Smogarinski model parameter needs to be smaller than the
isotropic one: The dynamic model produces this change automatically.
- The model parameters have to be reduced further near walls: The dynamic
model automatically decreases the parameter in the correct manner near
walls.

Although the dynamic model improves the Smagorinski model there are still
problems. The model parameter it produces is a rapidly varying function of
space and time so the eddy viscosity takes on large values of both signs.
 56
One-equation ksgs models

A one-equation model can be used to model the SGS turbulent kinetic energy.
The equation can be written on the same form as the RANS k-ε equation, i.e.

ksgs    ksgs 
t

xi
 ui ksgs   
xi 
   sgs  xi 
  Pksgs  
3/2
ksgs (90)
 sgs  ck ksgs , Pksgs  2 sgs Sij Sij ,   C
1/2

The production term, Pksgs , is calculated from the resolved velocity field. Typical

values for the contants are: ck  0.07 and C  1.07

RANS versus LES

Any numerical procedure for RANS can also be used for LES; for example
pressure correction methods such as SIMPLE are often used for LES.
What are the specific requirements to carry out LES with a finite volume code?
If you have a RANS finite volume code, it is very simple to transform that into
an LES code. A LES code is actually simpler than a RANS code. Both the
discretization scheme and the turbulence model are simpler in LES than in
RANS, see Table 4.
 57

Table 4: Differences between a finite volume RANS and LES code. (From
Davidson (2014)).

It is important to use a non-dissipative discretization scheme which does not

introduce any additional numerical dissipation, hence a second-order (or
higher) central differencing scheme should be employed.
The time discretization should also be non-dissipative. The Crank-Nicolson
scheme is suitable.
As mentioned above, turbulence models in LES are simple. There are two
reasons: first, only the small-scale turbulence is modeled and, second, no
equation for the turbulent length scale is required since the turbulent length
scale can be taken as the filter width, ∆.
In LES we are doing unsteady simulations. The question then arises, when can
we start to time average and for how long? This is exactly the same question we
must ask ourselves whenever doing an experiment in, for example, a wind
tunnel. We start the wind tunnel: when has the flow (and turbulence) reached
fully developed conditions so that we can start the measure the flow? Next
question: for how long should we carry out the measurements?
Both in LES and the wind tunnel, the recorded time history of the u1 velocity at
a point may look like in Fig. 17.
 58
u1

Figure 17: Time averaging in LES. (From Davidson (2014))

Time averaging can start at time t1 when the flow seems to have reached fully
developed conditions. It is difficult to judge for how long one should carry out
time averaging. Usually it is a good idea to form a characteristic time scale from
a velocity, V (free-stream or bulk velocity), and a length scale, L (width of a
wake or a body, 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 for the flow around a bluff body; a value of 10 may be sufficient
if L is the length of a recirculation region.
 59

DIRECT NUMERICAL SIMULATION (DNS)

The most accurate approach to turbulence simulation is to solve the Navier-

Stokes equations without averaging or approximation other than discretizations
whose errors can be estimated and controlled. In such simulations all of the
motions contained in the flow are resolved.

In a direct numerical simulation (DNS) the domain on which the computation is

performed must be at least as large as the largest turbulent eddy. A useful
measure of this scale is the integral scale (l0) of the turbulence. Each linear
dimension of the domain must be at least a few times the integral scale. A valid
simulation must also capture all of the kinetic energy dissipation. This occurs at
the smallest scales where the viscosity is active. This means that the size of the
grid must be no larger than the viscously determined scale called Kolmogorov
scale l (see figure 4 and text related to it). Based on this the number of grid
points must be at least l0/ l. As shown in (24) on page 16 this ratio is
proportional to Rel3/4. Here Rel is a Reynolds number based on the magnitude of
the velocity fluctuations and the integral scale. This parameter is about 0.01
times the macroscopic Reynolds number that engineers use to describe a flow.
Since this number of points must be employed in each of the three co-ordinate
directions and that the time step is related to grid size, the cost of simulation

scales as Re3 / 4  Re3 / 4  Re3 / 4  Re1/ 2  Re11/ 4  Re2.75 . For LES the same
 l  
 l  
 l  
 l
 l l
x  direction y  direction z  direction time

arguments regarding resolving the Taylor microscale (eq. 23) indicate that the

number of grid points scales with Re1/ 2  Re1/ 2  Re1/ 2  Re1/ 2  Re4 / 2  Re2 .
l
 l
 l
  l
 l l
x  direction y  direction z  direction time
 60

If we for example increase the Reynolds number by a factor of two we need to

increase the number of grid points by 22.75 = 6.72 and 22=4 for DNS and LES
computations, respectively. Since the number of grid points that may be used in
a computation is limited by processing speed and memory of the computer,
DNS is possible only at relatively low Reynolds number. On present computers
it is possible to make DNS of homogeneous flows at Reynolds numbers up to
about 200. As noted above, this corresponds to overall Reynolds number of
about 20.000. This allows DNS to reach the low end of Reynolds numbers of
engineering interest.

We may conclude that the major role that DNS can fill is as a research tool.
Some examples of this kind of uses are:

- Understanding of mechanisms of turbulence production, energy transfer and

dissipation in turbulent flows
- Simulation of the production of aerodynamic noise
- Understanding the effects of compressibility on turbulence
- Understanding the interaction of chemical reaction and turbulence

Other applications of DNS have already been made and many others will be
proposed in the future.
 61

LITERATURE

Davidson, L (2014), ‘Fluid mechanics, turbulent flow and turbulence modeling’,

Lecture Notes, Chalmers Institute of Technology, Gothenburg, Downloaded from:
http://www.tfd.chalmers.se/~lada/postscript_files/solids-and-fluids_turbulent-
flow_turbulence-modelling.pdf on July 15th

Ferziger, J.H. and Peric, M. (1995), ‘Computational Methods for Fluid Dynamics’,
Springer Verlag, Berlin.

Hinze, J.O. (1975), ‘ Turbulence’, 2nd edition, McGraw- Hill Publishing Co., New
York.

Pope, S.B. (2000), ‘Turbulent Flows’, Cambridge University Press, Cambridge.

Tennekes, H. and Lumley, J.L. (1972) ‘A first course in turbulence’, MIT press,
Cambridge, Massachusetts

Ranade, V. (2002), ‘Computational Flow Modeling for Chemical Reactor

Engineering’, Academic Press, San Diego.

Vesteeg, H.K. and Malalasekara, W. (1995), ‘An Introduction to Computational

Fluid Dynamics’, Longman Scientific & Technical, Essex, England.

View publication stats