University of Ljubljana

Faculty for mathematics and physics

Department of physics

Seminar
Turbulence models in CFD

Jurij SODJA
Mentor: prof. dr. Rudolf PODGORNIK

Ljubljana, March 2007

ABSTRACT

The seminar discusses basic concepts of turbulence modeling in computational fluid

dynamics (CFD). There are two mainstreams present in the field of interest. The first
one are so called Reynolds averaged Navier-Stokes (RANS) models. RANS models
are most widely used nowadays. Therefore theoretical aspects of RANS models are
presented in more detail. The second mainstream presents direct numerical
simulations and large eddy simulations. These models are important since they have
the potential to model turbulence in a much more general and better way than RANS
models. Basic concepts regarding large-eddy simulations are also presented in this
work. There is also a comparison between RANS models and large-eddy models
presented at the end of the seminar.
INDEX

1 INTRODUCTION .................................................................................................3

2 GENERAL REMARKS.........................................................................................3

2.1 Ideal turbulence model...................................................................................3

2.2 Complexity of the turbulence model..............................................................4

2.3 Classification of turbulent models .................................................................4

3 REYNOLDS-AVERAGED NAVIER-STOKES MODELS .................................6

3.1 Reynolds’s decomposition .............................................................................6

3.1.1 Equations describing instantaneous fluid motion ..................................6

3.1.2 Reynolds averaging................................................................................7

3.2 The closure problem ......................................................................................9

3.2.1 Laminar flow, infinitesimal fluctuations and superposition ..................9

3.3 Reynolds stress models ................................................................................10

4 COMPUTATION OF FLUCTUATING QUANTITIES.....................................12

4.1 Direct numerical simulation.........................................................................12

4.2 Large-Eddy simulation.................................................................................13

5 RANS versus LES................................................................................................14

6 CONCLUSION....................................................................................................17

7 REFERENCES ....................................................................................................18

2
1 INTRODUCTION

The abbreviation CFD stands for computational fluid dynamics. It represents a vast
area of numerical analysis in the field of fluid’s flow phenomena. Headway in the
field of CFD simulations is strongly dependent on the development of computer-
related technologies and on the advancement of our understanding and solving
ordinary and partial differential equations (ODE and PDE). However CFD is much
more than “just” computer and numerical science. Since direct numerical solving of
complex flows in real-like conditions requires an overwhelming amount of
computational power success in solving such problems is very much dependent on the
physical models applied. These can only be derived by having a comprehensive
understanding of physical phenomena that are dominant in certain conditions.[1], [2]

Why turbulence?

Whenever turbulence is present in a certain flow it appears to be the dominant over all
other flow phenomena. That is why successful modeling of turbulence greatly
increases the quality of numerical simulations.

All analytical and semi-analytical solutions to simple flow cases were already known
by the end of 1940s. On the other hand there are still many open questions on
modeling turbulence and properties of turbulence it-self. No universal turbulence
model exists yet.

Further more the price tag for our ignorance is immense. That makes the area of CFD
modeling also extremely economically attractive.

2 GENERAL REMARKS

2.1 Ideal turbulence model

Solving CFD problem usually consists of four main components: geometry and grid
generation, setting-up a physical model, solving it and post-processing the computed
data. The way geometry and grid are generated, the set problem is computed and the
way acquired data is presented is very well known. Precise theory is available.
Unfortunately, that is not true for setting-up a physical model for turbulence flows.

3
The problem is that one tries to model very complex phenomena with a model as
simple as possible. [3]

Therefore an ideal model should introduce the minimum amount of complexity into
the modeling equations, while capturing the essence of the relevant physics.

2.2 Complexity of the turbulence model

Complexity of different turbulence models may vary strongly depends on the details
one wants to observe and investigate by carrying out such numerical simulations.
Complexity is due to the nature of Navier-Stokes equation (N-S equation). N-S
equation is inherently nonlinear, time-dependent, three-dimensional PDE.

Turbulence could be thought of as instability of laminar flow that occurs at high

Reynolds numbers ( Re ). Such instabilities origin form interactions between non-
linear inertial terms and viscous terms in N-S equation. These interactions are
rotational, fully time-dependent and fully three-dimensional. Rotational and three-
dimensional interactions are mutually connected via vortex stretching. Vortex
stretching is not possible in two dimensional space. That is also why no satisfactory
two-dimensional approximations for turbulent phenomena are available.

Furthermore turbulence is thought of as random process in time. Therefore no

deterministic approach is possible. Certain properties could be learned about
turbulence using statistical methods. These introduce certain correlation functions
among flow variables. However it is impossible to determine these correlations in
advance.

Another important feature of a turbulent flow is that vortex structures move along the
flow. Their lifetime is usually very long. Hence certain turbulent quantities can not be
specified as local. This simply means that upstream history of the flow is also
important of great importance.

2.3 Classification of turbulent models

Nowadays turbulent flows may be computed using several different approaches.
Either by solving the Reynolds-averaged Navier-Stokes equations with suitable
models for turbulent quantities or by computing them directly. The main approaches
are summarized below.

4
Reynolds-Averaged Navier-Stokes (RANS) Models

• Eddy-viscosity models (EVM)

One assumes that the turbulent stress is proportional to the mean rate of
strain. Further more eddy viscosity is derived from turbulent transport
equations (usually k + one other quantity).

• Non-linear eddy-viscosity models (NLEVM)

Turbulent stress is modelled as a non-linear function of mean velocity

gradients. Turbulent scales are determined by solving transport
equations (usually k + one other quantity). Model is set to mimic
response of turbulence to certain important types of strain.

• Differential stress models (DSM)

This category consists of Reynolds-stress transport models (RSTM) or

second-order closure models (SOC). One is required to solve transport
equations for all turbulent stresses.

Computation of fluctuating quantities

• Large-eddy simulation (LES)

One computes time-varying flow, but models sub-grid-scale motions.

• Direct numerical simulation (DNS)

No modelling what so ever is applied. One is required to resolve the

smallest scales of the flow as well.

Extend of modelling for certain CFD approach is illustrated in the following figure
Figure 2.1. It is clearly seen, that models computing fluctuation quantities resolve
shorter length scales than models solving RANS equations. Hence they have the
ability to provide better results. However they have a demand of much greater
computer power than those models applying RANS methods. [4], [5]

5
 Figure 2.1 Extend of modelling for certain types of turbulent models

3 REYNOLDS-AVERAGED NAVIER-STOKES MODELS

The following chapter deals with the concept of Reynolds’s decomposition or

Reynolds’s averaging. The term Reynolds’s stress is introduced and explained briefly.
Further on methods how to include these ideas into certain numerical models are
presented. [1], [2], [6]

3.1 Reynolds’s decomposition

3.1.1 Equations describing instantaneous fluid motion

For easier understanding of certain mathematical ideas it is convenient to briefly

revise N-S equations describing instantaneous fluid motion at the beginning. All
variables describing instantaneous flow are marked with a tilde. These variables are
fluid’s density ( ρ
), velocity components ( u
i ), pressure ( p
) and components of

viscous stress tensor ( T
ij( v ) ). At this point it is also suitable to point out that these

variables are al time and space dependent.

General N-S equations for both turbulent and non-turbulent flow run:

(v)
 ∂u
i ∂ui  ∂p
∂T
ij
ρ
 + u
j  = − + and (3.1)
 ∂t ∂x ∂x ∂x
 j  i j

6
  ∂ρ
∂ρ
 ∂u

 + u
j  + ρ
i = 0 (3.2)
 ∂t ∂x j  ∂xi

The firs equation (3.1) is called momentum equation (second Newtonian law for
fluids). The second equation (3.2) is known as continuity equation. At this point I
would also like to define viscous stress tensor T
ij( ) as follows:
v

 1 
T
ij( ) = 2 µ  s
ij − s
kk δ ij  , (3.3)
v

 3 

where s
ij means:

1  ∂u
∂u
j 
s
ij =  i +  (3.4)
2  ∂x j ∂xi 

Should one assume incompressible flow the previous equations simplify immensely.
The continuity equation (3.2) is reduced to ∂ui xi = 0 . Having this result in mind the
momentum equation (3.1) can be rewritten as:

 ∂u
i ∂u
i  1 ∂p
µ ∂ 2u
i

 + u  = − + (3.5)
ρ ∂x 2j
j
 ∂t ∂x j  ρ ∂x i

The factor µ ρ is often regarded to as kinematic viscosityν . Viscous stress tensor

simplifies as well:

T
ij( v ) = 2µ s
ij (3.6)

3.1.2 Reynolds averaging

The concept of Reynolds averaging was introduced by Reynolds in 1895. One may
consider Reynolds averaging in many different ways. There are three most common
perceptions of this term: time averaging, space averaging or ensemble averaging.

Time averaging is appropriate when considering a stationary turbulence. That is when

the flow does not vary on the average in time. In such cases time average is defined
by:

  1 t +T  
F ( r ) = lim  ∫ f ( r , t ) dt  (3.7)
T →∞ T
 t 

7
Space average is appropriate for homogenous turbulence. That is a turbulent flow that
on the average does not vary in any direction. Space average is defined by:

1  
F ( t ) = lim 

V →∞ V ∫∫∫ f ( r , t ) dV  (3.8)

Ensemble average is the most general aspect of Reynolds average. It should be

understood as an average of N identical experiments. Ensemble average is both time-
and space-dependent. It is defined by:

 1 N

F ( r , t ) = lim
N →∞ N
∑ f (r,t )
n =1
n (3.9)

The main idea of Reynolds averaging is to decompose the flow to averaged and
fluctuating component:

u
i = U i + ui
p
= P + p (3.10)
T
ij( ) = Tij( ) + τ ij( )
v v v

This process is called Reynolds decomposition. The upper case letters represent the
mean values; the lower case letters represent the fluctuating values on the right hand
side in expressions (3.10). By inserting relations (3.10) into N-S equation (3.1) one
obtains the following expression:

 ∂ (U i + ui )
ρ + (U j + u j )
∂ (U i + ui ) 
 = − +
( (v ) (v)
∂ ( P + p ) ∂ Tij + τ ij ) (3.11)
 ∂ t ∂ x ∂x ∂x j
 j  i

This equation can now be averaged to yield an equation expressing momentum

conservation for the averaged motion. At this point it is important to stress that the
operations of averaging and differentiation commute. It is also assumed that the
average of fluctuating quantities is zero. Therefore the averaged momentum equation
reduces to:

(v)
 ∂U i ∂U i  ∂P ∂Tij ∂u
ρ +U j  =− + − ρ uj i (3.12)
 ∂t ∂x j  ∂xi ∂x j ∂x j


In similar manner continuity equation for incompressible flow can be decomposed.

Such a continuity equation is linear therefore the original form for the instantaneous
motion is preserved:

8
 ∂U i
=0
∂xi
(3.13)
∂ui
=0
∂xi

Using the second relation in equation (3.13) one can rework the last term on the right
hand side of the equation(3.12). The result runs:

(v)
 ∂U i ∂U i  ∂P ∂Tij ∂
ρ
 ∂t
+U j  = −
∂x j  ∂xi
+
∂x j
−
∂x j
(ρ ui u j ) (3.14)


(
Term ρ ui u j ) has the same structure and dimension as the viscous stress tensor.

However this term is not a stress at all. It is just a re-worked contribution of the
fluctuating velocities to the change of the averaged ones. On the other hand as far as
the motion of the fluid is concerned it acts as a stress. Hence its name, Reynolds
stress.

3.2 The closure problem

The problem with the above concept of Reynolds decomposition and averaging is that
it introduces additional variables ( ui2=1,2,3 , u1u2 , u1u2 , u2u3 ), for which there

are no available relations. Not in a general sense at least. [1], [2]

One could pretend that Reynolds stress is indeed a stress and try to write constitutive
relations similar to those for viscous stress. However there is an important difference
among these two stresses. Viscous stress is a property of a fluid. That is why separate
experiments can be carried out in order to determine corresponding constitutive
relations. These relations are valid then whenever a flow in that particular fluid is
observed. On the other hand Reynolds stress is a property of the flow. Hence it is
dependent on the flow variables them-selves. That is the reason why it changes from
flow to flow and no general constitutive relations are available.

3.2.1 Laminar flow, infinitesimal fluctuations and superposition

One solution to the closure problem is to treat the flow as a laminar flow with
fluctuations superimposed. One subtracts the averaged momentum equation from
equation describing instantaneous motion. The result for fluctuating motion reads:

9
  ∂ui ∂u 
(v)
∂p ∂τ ij  ∂U i   ∂u ∂u 
ρ +U j i  = − + − ρ u j  − ρ uj i − uj i  (3.15)
 ∂t 
∂x j  ∂xi ∂x j  ∂x j   
  ∂x j ∂x j 

The equation (3.15) has a similar structure than the averaged N-S equation(3.14). The
only difference is the last to terms on the right hand side. The first of them represents
the production term. It describes the way fluctuating motion extracts momentum from
the averaged motion. The second one is similar to Reynolds stress term in equation
(3.14) except that its mean is zero. Should equation(3.15) be averaged its average is
zero.

This approach requires fluctuations to be small. In the limit of infinitesimal

fluctuations Reynolds stress terms are negligible. Therefore averaged N-S equation
(3.14) yields a laminar flow. Furthermore the equation (3.15) reduces to linear PDE.
As a result of this process one obtains a well-defined – closed, set of equations
describing the observed flow.

3.3 Reynolds stress models

There were many different concepts and attempts to solve the turbulence closure
problem in a general form in the past. Nowadays there are two concepts that underlie
most of the Reynolds stress models. [1], [7], [8]

One and the most obvious attempt was to describe Reynolds stress in a similar way
viscous stress is described: the fluid is simply prescribed another property – turbulent
viscosity. This model had been introduced by Boussinesq back in 1877 even earlier
then Reynolds proposed his decomposition and averaging approach in 1895. There are
many difficulties regarding this model. Probably the major problem is how to obtain
this property without carrying out an actual experiment involving that particular flow.

Major breakthrough was done by Prandtl in 1925. He introduced the mixing length
concept analogous to mean free path of the molecules in gas. He also prescribed an
algebraic expression relating turbulent viscosity to the mixing length. That is why
Prandtl is known as the founder of so called algebraic or zero-equation models. Zero-
equation refers to the fact, that no additional transport equations besides to energy,
mass and momentum equations are needed.

10
Another important breakthrough was done by Prandtl in 1945, by introducing a
concept of turbulent viscosity as a function of turbulent kinetic energy. Major
advantage of this concept over the previous one is that it already takes into account
flows history. Hence it is a physically more realistic model. Prandtl used one
additional transport equation to model turbulent kinetic energy. Models based on this
concept are usually called one-equation models.

Still there is a need to specify a turbulence length scale, which is also a flow
dependent property. Hence one still needs to have certain knowledge about the
studied flow in advance. Therefore such models are called incomplete. Both zero- and
one-equation models are incomplete.

On the other hand complete model would be characterized by the fact that no
knowledge of the flow except the initial and boundary conditions is needed in
advance.

First complete model was introduced by Kolmogorov in 1942. The basic idea of his
model was to model turbulent kinetic energy ( k ) and the rate of energy dissipation
( ω ) and then relate the missing information of length and time scales to these
quantities. Since two additional equations are used to model k and ω these kind of
models are called two-equation models. They are also referred to as k − ω models.
Variations of this concept are so called k − ε models ( ε = k nω m ). Instead of ω ε is
modelled.

Another conceptually different attempt was to model Reynolds stress tensor directly.
At first one tried to derive actual Reynolds stress equations. The idea was to re-work
fluctuating momentum equation (3.15) in such a manner that it would describe
Reynolds stress. Major problem with this attempt is that it introduces even more new
unknown variables for which no constitutive relations are known

In 1951 Rotta managed to successfully model Reynolds stress tensor by using PDE.
This model is concept is more realistic than the Boussinesq’s turbulent viscosity
model. However it introduces six additional equations describing Reynolds stress and
one additional equation describing turbulence length scale.

In the field of RANS models no major conceptual break through was done ever since.
There were many improvements mainly in a sense of adjusting certain models to
particular flow cases.

11
4 COMPUTATION OF FLUCTUATING QUANTITIES

In the following section basic properties regarding direct numerical simulation (DNS)
and large-eddy simulation (LES) are briefly summarized. [1]

4.1 Direct numerical simulation

DNS simply means numerical solving of N-S and continuity equation. When dealing
with turbulent flow one tries to resolve all turbulent phenomena at all length and time
scales simply by numerical solving of N-S and continuity equation. For a successful
simulation one typically needs to know what the smallest length, time and velocity
scales are. This information is crucial in order to set space grid and time steps of
adequate scales. This data can easily be acquired by applying Kolmogorov turbulence
theory in advance. What ones want to extract form these data typically is the number
of grid point and time steps necessary.

Number of uniformly distributed grid points reads:

94 uT L
N uni ≈ (110 ReT ) , ReT = (4.1)
ν

ReT represents turbulent Reynolds number, uT represents frictional velocity, L is

typical length scale, ε = µ ρ is kinematic viscosity of the fluid. All quantities are
defined at the integral turbulence scale. All can be derived solely by applying
Kolmogorov turbulence theory.

Number of time steps is defined by:

∆ttotal 0.003 L
N time = , ∆t ≈ (4.2)
∆t ReT uT

The following table 4.1 lists numerical parameters for a certain flow. Figures handed
under N time represents the number of time steps required in order to reach statistically
steady flow. The figure handed under CPU is the amount of time (in hours) required
to obtain the solution using a standard Intel Core 2 Duo E6700 (12.53 gigaflops).
Time step required to finish one time step is approximately 3.2s. [9], [10]

12
As one can see the biggest problem regarding DNS is their overwhelming requirement
for computer power in a sense of both processor’s speed and a size of the memory for
storing intermediate results.

Table 4.1: Numerical requirements solving turbulent flow characterized by ReL and ReT

Re L Re T N DNS N time CPU [h]

6
12300 360 6.7*10 32000 28
7
30800 800 4.0*10 47000 42
8
61600 1450 1.5*10 63000 56
9
230000 4650 2.1*10 114000 101

DNS is of great importance. As computers develop one gains the capability to

simulate flows at ever higher and higher Reynolds number. Nowadays results
acquired by DNS are so good that one may consider them equivalent to data gained
experimentally.

4.2 Large-Eddy simulation

LES is a computation where large vortexes (eddies) are computed directly, while
small scale eddies are modeled. That is why space grid and time steps may be much
longer than in DNS. Hence LES is much more economical in term of computational
power required than DNS:

 ¨0.4 
N LES ≈  1 4  N DNS (4.3)
 ReT 

The following table 4.2 list numerical parameters regarding LES for the same flow
that is discussed in paragraph 4.1. It seems that LES takes roughly 10% of the DNS
CPU time to compute the solution. [9], [10]

Table 4.2: Numerical requirements solving turbulent flow characterized by ReL and ReT

Re L Re T N DNS N LES N time CPU [h]

6 5
12300 360 6.7*10 6.1*10 2913 2.5
7 6
30800 800 4.0*10 3.0*10 3525 3.15
8 7
61600 1450 1.5*10 1.0*10 4200 3.73
9 8
230000 4650 2.1*10 1.0*10 54285 4.87

13
The idea underlying LES is so called convergent evolution. Behavior of the large-
scale eddies depends strongly on the forces acting on the flow and on initial and
boundary conditions. They are flow-dependent On the other hand small-scale eddies
are generally independent from what is happening on the larger scales. They are flow-
independent. Hence large eddies are directly resolved while small eddies are modeled.
One tries to find a universal model for small eddies.

Another important concept regarding LES is filtering. One applies filtering functions
in order to remove sub-grid fluctuations from resolving. Sub-grid fluctuations are
modeled. This is achieved by averaging. One of the simplest filtering functions is
central-difference approximation it-self:

u ( x + h ) − u ( x − h ) d  1 x+h 
2h
=  ∫
dx  2h x − h
u (ξ ) d ξ 

(4.4)

Length scales of order h are still resolved, while length scales smaller than h are
modeled. They are called sub-grid scales (SGS).

5 RANS versus LES

Turbulent flow may be composed of many different features. Therefore it is very

important for a CFD model to be able to predict as many of them as possible.
Turbulent models are usually tested by simulating a flow past a bluff-body. In
particular example flow past a square block is analyzed. [6], [11]

An example of such a flow is shown in Figure 5.1.

Figure 5.1: Flow past a square block in 2D

14
Streamlines predicted LES, EASM (a sophisticated RANS model) and RANS are
shown in Figure 5.2. One can clearly see a strong influence of the model used in the
RANS calculation. The EASM reveals the similar topological features as the LES, but
differs in the extent of the recirculation zones. RANS predicts considerably larger
recirculation zone and the wake region is much more stretched.

Figure 5.2: Streamlines predicted by LES, EASM and RANS

More quantitative difference among observed models is observed when comparing

predicted information to experimental data. The comparison is shown in the following
set of pictures shown in Figure 5.3 and Figure 5.4. One clearly sees that the simple
RANS model of Wilcox fails terribly at predicting turbulent kinetic energy in the
wake region. However more sophisticated RANS model of EASM is much more
successful. Its results are of comparable quality to the results of LES model.

15
 Figure 5.3: Turbulent kinetic energy in the wake region

Figure 5.4: velocity in the wake region

There are certain examples when even simple RANS models outperform the
sophisticated LES model. One example is shown in Figure 5.5, where flow past an
airfoil (NACA 4412, alpha = 12°, Re = 1.6*106) is observed.

Figure 5.5: Pressure distribution

16
6 CONCLUSION

In the last decade CFD has become a major tool in engineering. Due to the progress in
computer technology CFD seems now able to deal with industrial applications at
moderate costs and turnaround times. The future relevance of CFD will therefore
depend on how accurate complex flows can be calculated. Since many flows of
engineering interest are turbulent, the appropriate treatment of turbulence will be
crucial to the success of CFD.

The flow field of a Newtonian fluid is fully described by the Navier-Stokes equation.
However, turbulent flows contain small fluctuations. The resolution of such small
motions requires fine grids and time steps, such that a direct simulation becomes
unfeasible for high Reynolds numbers.

Using RANS, the computational costs can be reduced by solving the statistically
averaged equation system, which requires closure assumptions for the higher
moments.

LES aims to reduce the dependence on the turbulence model. Hence the major portion
of the flow is simulated without any models, and must be resolved by the grid. Only
scales smaller than the resolution of the grid need a model. Consequently LES
approach is computationally more demanding than RANS. RANS models have a
computing time of only about 5% of the LES.

Sophisticated RANS models like EASM are able to capture important flow features
correctly. At low computational costs that makes them already a useful tool in
industrial design.

17
7 REFERENCES

[1] WILCOX, D.C.. Turbulence Modelling for CFD, DCW Industries, California,
USA, 1994

[2] http://www.cfd-online.com/Wiki/Turbulence_modeling, March 2007

[3] DAVIDSON L. An introduction to turbulence models, Chalmers university of

technology, Getebörg, Sweden, 2003

[4] APSLEY, D., CFD, Turbulence modelling in CFD, 2004

[5] BELL, B. Turbulent flow cases, Fluent Inc. 2003

[6] CELIĆ, A. Performance of Modern Eddy-Viscosity Turbulence Models, Institut

für Aerodynamik un Gasdynamik, Germany, 2004

[7] MENTER F. R. Two-equation eddy-viscosity turbulence models for engineering

applications, 1994

[8] RAMŠAK M Večobmočna metoda robnih elementov za dvoenačbne

turbulentne modele, Fakulteta za strojništvo, Univerza v Mariboru, Maribor,
SLO, 2004

[9] http://en.wikipedia.org/wiki/Flops, March 2007

[10] http://en.wikipedia.org/wiki/Cray, March 2007

[11] http://www.cfd.tu-berlin.de, March 2007

18