INTRODUCTION TO TURBULENCE MODELING

Goodarz Ahmadi
Department of Mechanical and Aeronautical Engineering
Clarkson University
Potsdam, NY 13699-5725

In this section, an introduction to the historical development in turbulence

modeling is provided.

Outline

Viscous Fluid

Turbulence

Classical Phenomenological models (Mixing Length)

One-Equation Models

Two-Equation Models (The k Model)

Stress Transport Models

Material-Frame Indifference

Continuum Approach

Consistency of the k Model

Rate-Dependent Model

ME637 1 G. Ahmadi
VISCOUS FLOW

The conservation laws for a continuous media are:

Mass

+ ( u ) = 0
t

Momentum

du
= f + t
dt

Angular Momentum

tT =t

Energy

e& = t : u + q + h

Entropy Inequality

q h

& ( ) >0
T T

Constitutive Equation

Experimental evidence shows that for a viscous fluid, the stress is a function of
velocity gradient. That is

kl = p kl + G kl (u i , j )

The velocity gradient term may be decomposed as

u i , j = d ij + ij

where d ij is the deformation rate tensor and ij is the spin tensor. These are given as

1 1
d kl = (u k ,l + u l, k ), kl = (u k, l u l, k )
2 2

ME637 2 G. Ahmadi
The principle of Material Frame-Indifference of continuum mechanics implies that the
stress is generated only by the deformation rate of media and the spin has no effect. This
is because both stress and deformation rate tensors are frame-indifferent while spin is not.
Thus, the general form of the constitutive equation is given as

kl = p kl + Fkl (d ij )

For a Newtonian fluid, the constitutive equation is linear and is given as

kl = ( p + u i ,i ) kl + 2d kl

The entropy inequality imposed the following restrictions on the coefficient of viscosity:

3 + 2 > 0 , > 0 ,

Using the constitutive equation in the balance of momentum leads to the celebrated
Navier-Stokes equation. For an incompressible fluid the Navier-Stokes and the
continuity equations are given as

u i u i p 2ui
( + uj )= + ,
t x j x i x j x j

u i
=0
xi

These form four equations for evaluating four unknowns u i , p.

TURBULENT FLOW

In turbulent flows the field properties become random functions of space and
time. Thus

u i = U i + u i u i = Ui , u i = 0

p = P + p p = P , p = 0

Substituting the decomposition into the Navier-Stokes equation and averaging leads to
the Reynolds equation.

ME637 3 G. Ahmadi
Reynolds Equation

U i U i 1 P 2 U i u i u j
+ Uj = +
t x j x i x j x j x j

Here

Tij = u i u j =Turbulent Stress Tensor

First Order Modeling (Classical Phenomenology)

Boussinesq Eddy Viscosity:

dU
T21 = u v = T
dy

Tij U i U j 1
= u i u j = T ( + ) u k u k ij
x j x i 3

Prandtl Mixing Length

U U
T21 = l 2m | |
y y

U T T
T = l 2m | |, T v =
y T y

Kolmogorov-Prandtl Expression

Eddy Viscosity

T c u l , u = velocity scale, l = length scale , c = const.

Kinematic Viscosity

c , c = speed of sound, = mean free path

Let
U U
u ~ lm | |, l = l m T = l 2m | |
y y

For free shear flows

ME637 4 G. Ahmadi
 l m ~ cl 0 ( l 0 = half width)

Close to a wall

l m = y (y = distance from the wall)

Local Equilibrium

For local equilibrium

Production = Dissipation Mixing length Hypothesis

Shortcomings of the Mixing Length Model

U
When = 0 T = 0
y
Lack of transport of scales of turbulence

Estimating the mixing length, l m .

Reattachment Point

U
At the reattachment point = 0 which leads to vanishing eddy diffusivity and
y
thus negligible heat flux. Experiments, however, show that the heat flux is maximum at
the reattachment point.

ME637 5 G. Ahmadi
One-Equation Models

Eddy Viscosity

1
T = c k 1/ 2 l , k= u i u i = Turbulence Kinetic Energy
2

Exact k-equation

d u i u i u u P U i u u i 2 u i u i
= u k ( i i + ) u i u j i +
dt
1
424 23 14 x k4424 2 443 x j x j x j x j x j 2
Convective
1424 3 142 43 142 4 43 4
TurbulenceDiffusion Pr oduction Dissipation Viscous Diffusion
Transport

where
d u i u i d
= convective transport, = +Uj
dt 2 dt t x j
u u P
u k ( i i + ) = turbulence diffusion
x k 2
U i
u i u j = production
x j
u i u i
= , = dissipation
x j x j
2 u i u i
= viscous diffusion
x j x j 2

Modeled k-equation

dk T k U U j U i k 3/ 2
= ( ) + T ( i + ) cD ,
dt x j k x j x j x i x j l

where

T k
( ) = turbulence diffusion, k 1 (turbulence Prandlt number)
x j k x j

U i U j U i
T ( + ) = production,
x j x i x j

ME637 6 G. Ahmadi
 k 3/ 2
cD = = dissipation.
l

Note that the turbulence length scale l is given by an algebraic equation.

Bradshaws Model

Modeled k-equation

dk U k 3/ 2
= (Bk Max ) + ak cD
dt y y l

where

U
ak = production
y
k3/2
cD = dissipation
l

u v = ak (shear stress kinetic energy)

Max y y
B= g( ) , l = f ( )
v 0
2

Shortcomings of One-Equation Models

Use of an algebraic equation for the length scale is too restrictive.

Transport of the length scale is not accounted for.

Transport of Second Scale (Boundary Layer)

The transport of a second turbulence scale, z, is given as

z-Equation

dz T z U k
= ( ) + z[c1 T ( ) 2 c 2 ] + Sz
dt y z y k y T
where

ME637 7 G. Ahmadi
 T z
( ) = diffusion
y z y
U
c1 T ( ) 2 = production
k y
k
c2 = destruction,
T
S z = secondary source

Choices for z

Turbulence Time Scale = l2 / k

Turbulence frequency Scale = k / l 2
Turbulence mean-square vorticity Scale = k / l 2
u u i
Turbulence Dissipation = = i
x j x i
z = kl

-Equation (exact):

d u i u i u k 2 u i 2 u i
= (u j ') 2 2
dt x j x k x l x l x k x l x k x l
1424 3 1442443 144 42444 3
Diffusion Generation by Viscous destruction
vortexstretching

Note that
k 3/ 2 k 3/ 2 E(k) Universal Equilibrium
~ , l~
l
Thus ,
2
k ,
T ~ k l ~

Note also that is also the amount

of energy that paths through the Inertia
entire spectrum of eddies of Subrange
turbulence. k

Schematics of turbulence energy spectrum.

ME637 8 G. Ahmadi
Two-Equation Models

The k Model

c k 2 U i U j 2
T = , u i u j = T ( + ) k ij
x j x i 3
k-equation

dk T k U U j U i
= ( ) + T ( i + )
dt x j k x j x j x i x j 123
144244 3 144424443 dissipation
Diffusion Pr oduction

-equation

d T k U i U j U i 2
= ( ) + c 1 T ( + ) c2 ,
dt x j x j k x j x i x j 123 k
144244 3 14444244443 Distruction
Diffusion Generation

where

c = 0.09 , c 1 = 1.45 , c 2 = 1.9 , k = 1 , = 1.3 , (Jones and Launder, 1973)

Momentum

dU i 1 P
= u iu j
dt x i x j

Mass

U i
=0
x i

Closure: Six Equations for six unknowns, v i , P, k , .

Kolmogorov Model

dk 1 k k
= 2 T S ij Sij k 2 + A ( )
dt 1 424 31 4224 3 x j x j
Pr oduction
Dissipation
1442443
Diffusion

ME637 9 G. Ahmadi
 dw 7 k k
= 2 + 2 A ( )
dt 10 x j x j

Ak 1 U U j
T = , S ij = ( i + )
2 x j x i

Comparison of Model Predictions

In this section comparisons of the predictions of the mixing length and one and
two-equation models with the experimental data for simple turbulent shear flows are
presented.

Development of Plane Mixing Layer (Rodi, 1982)

It is seen that the k model captures the features of the flow more accurately
when compared with the one-equation and mixing length model.

ME637 10 G. Ahmadi
 Turbulent Recirculating Flow (Durst and Rastogi, 1979)

The k model predictions for turbulent flow in a channel with an

obstructing block are compared with the experimental data.

b) Velocity profiles

Flow in a Square Cavity (Gosman and Young)

The k model predictions for a square cavity are shown in this section.

ME637 11 G. Ahmadi
Free-Stream Turbulence

The free stream turbulence affects the skin friction coefficient. The mixing length
model can not predict such effects. The k model does a reasonable job in predicting
the increase of skin friction coefficient with the free stream turbulence.

Turbulent Channel Flow (Rodi, 1980)

Distribution of mean velocity and turbulence quantities in fully developed two-

dimensional channel flows was predicted by Rodi (1980) using an algebraic stress model
(a modified k model)

Closed Channel
Flow

ME637 12 G. Ahmadi
Open Channel
Flow

Jets Issuing in Co-flowing Streams (Rodi, 1982)

1
U E 0
For a jet, = U ( U U E )dy = Excess momentum thickness

Plane Jet Round Jet

ME637 13 G. Ahmadi
Shortcomings of the k Model

U i U j 2
u i u j = T ( + ) k ij
x j x i 3

Limited to an eddy viscosity assumption.

Eddy viscosity and diffusivity are assumed to be isotropic.
Convection and diffusion of the shear stresses are neglected.
Normal turbulent stresses are not considered.
Main assumption is: u i u j ~ k .

Stress Transport Models

Subtracting the Navier-Stokes equation form the Reynolds equation, we find an

evolution equation for the turbulence fluctuation velocity. That is

u i u i 1 p 2 u i U i
+ Uk = + + u i u k (u i u k ) u k (1)
t x k x i x k x k x k x k x k

u j u j
1 p' 2 u j U j
+ Uk = + + u j u k (u j u k ) u k (2)
t x k x j x k x k x k x k x k

Multiplying Equation (1) by u j , and Equation (2) by u i , adding the resulting equations
and averaging leads to the exact stress transport equation:

U j U i u u j p' u i u j
( + Uk )u i u j = [u i u k + u j u k ] 2 i + ( + )
t x x x k x k x k x j x i
144424k 44 3 14444k244443 142 43 144244 3
Convection Pr oduction Dissipation Pr essure strain
,
p'
[u i u j u k + ( u i jk + u j ik ) u i u j ]
x k x k
1444444442444444443
Diffusion

where

U j U i
[ u i u k + u j u k ] = production,
x k x k

ME637 14 G. Ahmadi
 u i u j
2 = dissipation,
x k x k

p ' u i u j
( + ) = pressure strain
x j x i
p'
[ u i u j u k + (u i jk + u j ik ) u i u j ] = diffusion.
x k x k

Modeling Diffusion:

k u u u u u i u j
u i u j u k = c s (u i u l i k + u j u l k i + u k u l )
x l x l x l

Pressure diffusion 0
Viscous diffusion 0

Modeling Dissipation

u i u j 2
2 = ij
x k x k 3

Modeling Pressure-Strain

p' u i u j u u m u i u j
( + ) = dx1 G (x, x1 ){ ( l )1 ( + )
x j x i x x x x
x1
1444444 4m244 l
444 j
44 3i

(ij1 )

U u u u j
+ 2( l )1 ( m )1 ( i + ) }
x m x l x j x i
14444244443
(ij2 ) + (ji2 )

where
2
(ij1) = c 1 ( )(u i u j ij k ) (Return to isotropy)
k 3

2
(ij2) + (ji2) = (Pij P ij ) (Rapid term)
3
Here
U j U i
Pij = ( u i u k + u i u k )
x k x k

ME637 15 G. Ahmadi
Pressure-Strain Correlation

p' u i u j
Modeling the pressure-strain correlation, ( + ) , is critical to the stress
x j x i
transport equations.

Navier-Stokes Equation

u i u i 1 p 2 ui
+ uk = + (1)
t x k x i x j x j

Taking the divergence of (1), we find

2uiuk 1 2p
= , (2)
x i x k x i x i

or

p 2u i u k
2 = . (3)
x i x k

Averaging Equation (3), the result is:

p 2 U i U k 2 u i u k
2
= . (4)
x i x k x i x k

Subtracting (4) from (3), we find

p' U u k u i u k u i u k
2 = 2 i + . (5)
x k x i x k x i x k x i

Introducing the Green function G (x, x 1 ) for the Poisson equation. i.e.,

2 G(x, x 1 ) = (x x 1 ) , (6)

Equation (5) may be restated as

ME637 16 G. Ahmadi
 U i u k u i u k u i u k
p' = G (x, x 1 )[2 + ]dx 1 (7)
x1
x k x i x k x i x k x i

The pressure-strain rate correlation then becomes

p' u i u j u u k u i u j U u u u j
( + ) = G( x, x 1 )[( l )1 ( + ) + 2( l )1 ( k )1 ( i + )]dx 1
x j x i x1
x k x l x j x i x k x l x j x i
(8)
or
p' u i u j
( + ) = ij(1) + ( (ij2 ) + (ji2 ) ) (9)
x j x i
Note that for unbounded regions

1 1
G(x, x 1 ) = . (10)
4 | x x 1 |

Modeled Stress Transport Equation

U i U j 2
( + Uk )u iu j = [u j u k + u i u k ] ij
t x k x k x k 323
1
144 424 44 3 1444 424444 3 Dissipatio n
Convection Pr oduction

2
c1 (u i u j ij k ) + ( (ij2 ) + (ji2 ) ) + ( (ijw ) + (jiw) )
14k 4444 342444444 3 142 4 43 4
Wall effects
Pr essure strain

k u j u k u u u i u j
+ cs { [ u i u l + u j u l k i + u k u l ]}
x k x l x l x l
1444444444 424444444444 3
Diffusion
where

U i U j
[ u j u k + u i u k ] = production
x k x k
2
ij = dissipation,
3

2
c1 (u i u j ij k ) = pressure-strain
k 3

k u j u k u u u i u j
cs { [u i u l + u j u l k i + u k u l ]} = diffusion.
x k x l x l x l

ME637 17 G. Ahmadi
Dissipation Equation

k U i 2
( + Uk ) = c ( u k u i ) c 1 u i u k c 2 ,
t x k x k x i k x k 123 k
1442443 144424443 14 4244 3 Destruction
Convection Diffusion Generation

where

k
c ( u k u l ) = diffusion,
x k x k
U i
c 1 u i u k = generation
k x k

2
c2 = destruction.
k

Reynolds Equation

1 P
( + Uj )U i = u i u j
t x j x i x j

Continuity Equation

U i
=0
xi

Closure

There are eleven equations for eleven unknowns, U i , P , u i u j , .

ME637 18 G. Ahmadi
Comparison of Model Predictions

In this section comparisons of the predictions of the stress transport model with
the experimental data for simple turbulent shear flows are presented.

Curved Mixing Layer (Gibson and Rodi, 1981)

ME637 19 G. Ahmadi
Asymmetric Channel Flow (Launder, Reece and Rodi, 1975)

Mean Velocity and Turbulent Shear Stress

Turbulence Intensities

ME637 20 G. Ahmadi
Algebraic Stress Transport Model (Rodi, ZAMM 56, 1976)

A simplified stress transport model is given as:

d k U j U i
u i u j = c s ( u k u m u u ) u i u k u j u k
dt x k x m i j x k x
14 444244 44 3 1444 24443k
Diffusion Pr oduction
, (1)
2 2 2
c1 (u i u j ij k ) (Pij ij P) ij
14k 44444 32444444 3 3 1323
Pr essure Strain Dissipation

where

k
D ij = ( u k u l u i u j ) = diffusion,
x k x l
U j U i
Pij = u iu k u j u k = production,
x k x k
2 2
c1 (u i u j ij k) ( Pij ij P ) = pressure-strain,
k 3 3

2
ij = dissipation.
3

1
Here, P = Pii is the production rate of turbulent kinetic energy. Contracting
2
Equation (1), we find the transport equation for k:

dk k k U k
= cs ( u k u m ) u k u m
{ , (2)
dt x k x m x m Dissipation
14442444 3 14243
Diffusion Pr oduction

where

k k
D= ( u k u l ) =diffusion
x k x l
U k
P = u k u l = production
x l

Rodi (1976) assumed that

d u iu j dk u i u j
u i u j D ij = ( D) = ( P ) , (3)
dt k dt k

ME637 21 G. Ahmadi
Using (3) in (1) and rearranging, the result is:

Pij 2 P
2 ij
1 3
u i u j = k ij + . (4)
3 c1
1 + ( 1)
1 P
c1

Equation (4) provides an algebraic expression for u i u j .

For simple shear flows, it may be shown that equation (4) reduces to the
Kolmogorov-Prandtl hypothesis with

k2
T = c (5)

and

1 P
[1 (1 )]
2 (1 ) c1
c = with = 0.6 and c1 = 1.8 2.2 . (6)
3 c1 [1 1 ( P 1)]2
+
c1

Conclusions (Existing Models)

Available models can predict the mean flow properties with reasonable accuracy.
Small adjustments of parameters are sometimes necessary!
First-order modeling gives reasonable results only when a single length and velocity
scale characterizes turbulence.
The k model gives relatively accurate results when a scalar eddy viscosity is
sufficient to characterize the flow. That is there is no preferred direction for example
through the action of a body force.
The stress transport models have the potential to most accurately represent the mean
turbulent flow fields.

Deficiencies of Existing Models

Adjustments of coefficients are sometimes needed.

The derivation of the models are somewhat arbitrary.
There is no systematic method for improving a model when it loses its accuracy.
Models for complicated turbulent flows (such as multiphase flows) are not available.
Realizability and other fundamental principles are sometimes violated.

ME637 22 G. Ahmadi
 For example, the transport equations for u ' 2 , v '2 , w '2 must always lead to positive
values of these quantities. In addition, the transport equations for the cross terms
must also lead to cross correlations that satisfy Schwarts inequalities. i.e.,
2
u 2 v' 2 u v ' > 0.

ME637 23 G. Ahmadi
Anisotropic Rate-Dependent Model

Averaged Balance Laws:

Mass

v i ,i = 0

Linear Momentum

v& i = t ji , j + t Tji , j + f i

Thermal Energy

e& = q i ,i + q iT,i + t ij v j ,i + + r

Fluctuation Energy

k& = t Tij v j,i + K i,i

Clausius-Duhem Inequality

& (q i ) ,i R iT,i r + & T SiT,i > 0

Helmholtz free Energy Function

T
=e , T = k T

Heat Flux-Coldness Correlation

R Ti = q Ti

Fluctuation Energy Flux-Turbulence Coldness Correlation

S Ti = K i T E i

Total Heat Flux

Qi = q i + q Ti

ME637 24 G. Ahmadi
Clausius-Duhem Inequality

& 1
& 2 Q i ,i + t ij v j,i +

T 1 T&T
1
+ T
& T 2 T K i T,i + T + t Tij v j,i > 0
( ) E i ,i

Constitutive Equations

Stress

t ij = p ij + 2d ij

2
t Tij = k ij +
T Dd ij
( 1
)
+ T 2 + 2 d kl d kl d ij + d kl d kl ij d ik d kj
3 Dt 3

Jaumann Derivative

Dd ij
= d& ij + d ik kj + d jk ki
Dt

Deformation Rate and Spin Tensors

d ij =
1
(vi, j + v j,i ) , ij =
1
( v i , j v j ,i ) ,
1
= d ij d ij
2 2 2

Heat Flux

T
Q i = + C ,i

Fluctuation Energy flux

T k
K i = + k k ,i ,i

Heat Capacity

2
C =
2

ME637 25 G. Ahmadi
Thermodynamic Constraints

T > 0 , > 0 , > 0 , < 48

2

T = C k

Turbulence free Energy Function

o
= k ln
T
+ C 2 + C 0
k

Turbulence Stress

2 D 1
t Tij = kij + T 2d ij + d ij + 2 d lk d kl d ij + d lk d kl ij d ik d kj
3 Dt 3

Basic Equations

v i ,i = 0

2 Dd ij 1
v& i = p + k + 2( + T )d ij + T [ + ( d lk d kl ij d ik d kj ) + 2 d lk d kl d ij ] + f i
3 ,i Dt 3 , j

T
C & = ( + C ) ,i + 2d ij d ij + + r
,i

T k Dd ij
k& = ( + k )(k ,i ,i ) + P + T d ij
,i Dt

P = T [ 2d ij d ij d ik d kjd ij + 2 (d ij d ji ) 2 ]

ME637 26 G. Ahmadi
Scale Transport Equations

T T k k
& = ( + T ) ,i + C 1 P + C 3 + k 2 [k , i ,i ][k ,i ,i ]
,i k k
T 2C
+ + T [ 2
]( 2 ) ,i ,i C 2 C D
0 + 2C

1 1 1
> C 1 > 0 , C 2 > , > C 3 > 0 , 0 > 0
0m 0 0m

0 m max .( 0 + 2C 2 )

k
= CD

T T k k
& = ( + ) ,i + C 1 P + C 3 + k 2 [k ,i ,i ][k ,i , i ]
,i k k
T 2 C k2 2 2
2
+ + [ ]( ) C C
2
,i ,i
k2 k
0 + 2C 2

k2 k
= C
T
, =

When = 0 , = 0.93 , = 0.54

2
> , = 0.005
48

C = 0.09 , C1 = 1. 45 , C 2 = 1.92 , k = 1 , = 1.3

ME637 27 G. Ahmadi
 Comparison with the Experimental Data for Duct Flow

In this section the rate-dependent model predictions are compared with the
experimental data of Kreplin and Eckelmann, DNS of Kim et al. and the k- model
predictions

Comparison of mean velocity profile with Comparison of axial turbulence

the experimental data of Kreplin and intensity with the experimental data of
Eckelmann for Reynolds numbers of 8200 Kreplin and Eckelmann, DNS of Kim et
and 5600. al. and k- model.

Comparison of vertical turbulence Comparison of lateral turbulence

intensity with the experimental data of intensity with the experimental data of
Kreplin and Eckelmann, DNS of Kim et Kreplin and Eckelmann, DNS of Kim et
al. and k- model. al. and k- model.

ME637 28 G. Ahmadi
 Comparison of turbulence shear stress with the experimental data of
Kreplin and Eckelmann, and DNS of Kim et al.

Comparison with Experimental the Data for Backward Facing Step Flows

In the section the rate-dependent model predictions are compared with the
experimental data of Kim et al. and the algebraic model of Srinivasan et al.

Schematics of the flow over a backward facing step.

ME637 29 G. Ahmadi
 Comparison of the mean velocity profiles with the data of Kim et al. (1978).
(Dashed lines are the model predictions of Srinivasan et al. (1983).

Comparison of the turbulence kinetic energy profiles with the data of Kim et al. (1978).
(Dashed lines are the model predictions of Srinivasan et al. (1983).

ME637 30 G. Ahmadi
 Comparison of the turbulence dissipation profiles.
(Dashed lines are the model predictions of Srinivasan et al. (1983).

Comparison of the axial turbulence intensity profiles with the data of Kim et al. (1978).
(Dashed lines are the model predictions of Srinivasan et al. (1983).

ME637 31 G. Ahmadi
 Comparison of the vertical turbulence intensity profiles with the data of Kim et al.
(1978). (Dashed lines are the model predictions of Srinivasan et al. (1983).

Comparison of the turbulence shear stress profiles with the data of Kim et al. (1978).
(Dashed lines are the model predictions of Srinivasan et al. (1983).

ME637 32 G. Ahmadi
 Schematics of the flow in an axisymmetric pipe expansion.

Comparison of the mean velocity profiles with the data of Junjua et al. (1982) and
Chaturvedi (1963). (Dashed lines are the model predictions of
Srinivasan et al. (1983))

Comparison of the axial turbulence intensity profiles with the data of Junjua et al. (1982)
and Chaturvedi (1963). (Dashed lines are the model predictions of
Srinivasan et al. (1983))

ME637 33 G. Ahmadi
 Comparison of the vertical turbulence intensity profiles with the data of Junjua et al.
(1982) and Chaturvedi (1963). (Dashed lines are the model predictions of
Srinivasan et al. (1983))

Comparison of the turbulence shear stress profiles with the data of Junjua et al. (1982)
and Chaturvedi (1963). (Dashed lines are the model predictions of
Srinivasan et al. (1983))

ME637 34 G. Ahmadi