You are on page 1of 66

Turbulence and its Modeling

Introduction
• Turbulence is a very complex phenomenon
• Dominated by a wide range of length and time
scales
• Turbulence is always 3D and unsteady
• An ideal model should introduce the minimum
amount of complexity while capturing the essence
of the relevant physics
• Often the required result necessitates a very simple
turbulence model
• Alternatively a seemingly simple quantity might
require comprehensive resolution of turbulence
• Knowledge, understanding and experience vital in
making a choice
Physical Description
• For small length scales and low velocities flows are laminar
• Viscous diffusion dominates, damping vorticity and
momentum
• At higher velocities and larger length scales the viscous
stresses are overcome by the fluid’s inertia, rapid fluctuations
are observed and the flow becomes essentially 3D and
unsteady
• Turbulent motion is characterized by swirling motions or
eddies over a wide range of length scales
• Time and length scales are represented by frequencies and
wavelengths obtained from a Fourier analysis of the turbulent-
flow time history
• Length scales range from the mean flow width down to the
smallest length scale of turbulence (much larger than
molecular length scales)
Physical Description -
continued
• Turbulence develops as an instability of laminar flows
• There is strong interaction between fluctuations of differing
wavelengths and directions
• Turbulent motion is spread across wavelengths primarily
through vortex stretching
• Energy is gained if vortex elements are orientated as to allow
the mean velocity gradients to stretch them
• Larger wavelengths interact the most strongly with the mean
flow
• Large-scale turbulence carries the most energy – responsible
for enhanced diffusivity
• Large eddies stretch smaller eddies, transferring or cascading
energy to them
• Finally the energy is dissipated by viscosity in the shortest
length scales
Physical Description -
continued
• Rate of dissipation set by long-wave motion,
shorter length scales simply adjust
• Turbulent flow is always dissipative
• Large eddies contain smaller eddies
• Large eddies move across the flow carrying smaller
eddies with them
• Large eddies have a long lifetime, can persist for
distances of up to 30 times the width of the flow
• As such, turbulence at any location depends on
history of flow, cannot be based on the local strain-
rate tensor as in laminar flow
• Enhances diffusivity of the flow
Physical Description - continued
Physical Description - continued
Physical Description - continued
Physical Description - continued
Physical Description - continued
Physical Description - continued
Physical Description - continued
Smallest Scales
• Cascade process involves transfer of turbulent
kinetic energy (per unit mass), k, from large to
small eddies
• Small scale eddies associated with a short time
scale Smallest scales in
turbulent flow
• Dissipation rate should be nearly equal to rate at
which energy is transferred to smaller scales
η ≡ (ν 3 / ε )
1
• Kolmogorov’s universal equilibrium theory: 4
motion of smallest scales depend on
– Rate of energy supplied by large eddies, ε = -dk/dt
τ ≡ (ν / ε )
1
– Kinematic viscosity, ν 2
• Kolmogorov scales: length, time and velocity
υ ≡ (νε )
1
• Turbulent boundary layer associated with an air
flow of 100 km/h: η ≈ 4.6 micron 4

• Air, mean free path: lmfp ≈ 0.064 micron


• η/lmfp ≈ 72
Spectral Representation
• Spectral distribution of
energy
• Fourier analysis
– Wavenumbers: κ
– Wavelengths: λ = 2π/κ (eddy
size)
• E(κ)dκ - turbulent kinetic ∞
k = ∫ E (κ )dκ
energy contained between
wavenumbers κ and κ+dκ
• E(κ) – Energy spectral
density or Energy Spectrum 0
Function
k = turbulent specific kinetic energy
SR - continued
Know what each part of spectral
representation graph means!

Applies to inertial sub-range only

λ = wavelength
κ = 2π/λ = wavenumber
Large eddies => small κ
Small eddies => large κ
SR - continued
• E(κ) is a function of
– l – length characteristic of larger eddies
– S – mean strain rate
– ν and ε - turbulence is dissipative
– κ - by definition
• Dimensional analysis and measurements confirm
that for high Re (Taylor):

3
2
⇒ k ∝ (εl )
k 2
ε∝ 3
l
• l – integral length scale
SR - continued
• Most turbulence models assume a large separation of scales (high
Re)

l >> η l is integral length scale, wtf is that?

• Substituting estimate of ε into Kolmogorov length scale

1
3 2
l k l
∝ Re , ReT ≡
4
η ν
T

• ReT – Turbulent Reynolds number


• ReT >> 1 – necessary conditions for a large separation of scales
SR - continued
• According to Kolmogorov’s universal equilibrium theory, there exists a
range of eddies between largest and smallest where cascade process is
– Independent of statistics of energy containing eddies (S and l can be ignored)
– Independent of direct effect of molecular viscosity (ν can be ignored)
• In this range energy is transferred through inertial effects only
• E(κ) depends only on ε and κ
• On dimensional grounds

E (κ ) = Cκ ε 2 3κ −5 3 ,
1 1
<< κ <<
l η
• CK – Kolmogorov constant
• Range of wavenumbers – inertial subrange
• Existence of inertial subrange confirmed experimentally
• Kolmogorov -5/3 law, central in DNS and LES
SR - continued

Understand this
image well since it's come
up twice!
Turbulent Boundary Layer
• Primarily an empirical investigation
• At wall: viscous shear dominates
• In outer region: turbulent shear
dominates
• Middle region: both viscous and
turbulent shear are important
TBL - continued
• Law of the Wall – 1930, Ludwig Prandtl deduced that close to the
wall, velocity is unaffected by shear layer thickness
• u* - friction velocity
• y+ - dimensionless distance from wall (Reynolds number based on
friction velocity and normal distance from wall)

τw ρyu ∗
yu ∗

u = y+ = =
ρ µ ν

u = φ (ρ , µ , τ w , y )

u+ = y+
TBL - continued
• Velocity Defect Law – 1933, Theodore Von Kármán,
deviation in velocity in the outer layer is
– Independent of molecular viscosity
– Dependent on boundary layer thickness, δ and other properties
• U – free stream velocity

(U − u ) = φ (ρ , δ , τ w , y )
U −u  y
= G 
δ 

u
TBL - continued
• Logarithmic Overlap Layer – 1937, C.B Millikan
uses functional analysis to demonstrate that only a
logarithmic distribution could span middle region
• The equation holds for smooth surfaces in the
absence of any adverse pressure gradients

+
u =
1
κ
( )
ln y + + C

κ ≈ 0.41
C ≈ 5 .0
TBL - continued
• Composite
turbulent
velocity profile
TBL - continued
25

Wall • Transition close


20
Overlap to the wall

15
Shows that log-layer law
u+

comes into effect near y+ = 30


10
logarithmic overlap layer

5
linear/viscous sublayer (law of the wall)
u+ = y+

0
0.1 1 10 100 1000
y+
logarithmic overlap layer TBL - continued
2.5
Wall
• Water flow in pipe,
Overlap
demonstrating thickness
of laminar sub-layer
2

1.5
u (m/s)

d = 13.6 mm
uave= 2.7586 m/s
1 Re = 24700 Use these values to make calculations!
f Blasius = 0.00532
2
το = 20.24 N/m
0.5

linear/viscous sublayer (law of the wall)


u+ = y+
0
0 0.05 0.1 0.15 0.2
y (mm)
TBL - continued
3.5
Wall • Water flow in pipe,
3 Overlap logarithmic across
diameter
2.5

2
u (m/s)

logarithmic overlap layer


1.5 d = 13.6 mm
uave= 2.7586 m/s
1 Re = 24700
Use these values to make calculations!
f Blasius = 0.00532
2
0.5 το = 20.24 N/m

0
0 2 4 6 8 still looking at fairly small scale

y (mm)
TBL - continued
6
Wall • Laminar vs. turbulent
5
Overlap
Laminar
velocity profile for the
same flow rate
4
laminar velocity profile
u (m/s)

d = 13.6 mm
2
uave= 2.7586 m/s
Re = 24700
1 f Blasius = 0.00532
2
το = 20.24 N/m
turbulent velocity profile (log-overlap region)
0 steep velocity change!
0 2 4 6 8
y (mm)
TBL - continued
• For rough walls the law of the wall still holds
• C is a function of the average roughness height, ks

k ≡ + u * ks
C → 8.5 −
1
κ
( )
ln k s+ , k s+ >> 1
ν
s

• For a completely rough wall

+  y
1
u = ln  + 8.5
κ  ks 
TBL - continued
• Defect layer lies between the log layer and edge of boundary
layer
• Velocity deviates noticeably from logarithmic behavior as free
stream is approached
• Law of the Wake:

2Π π y 
u+ =
1
κ
( )
ln y + + C +
κ
sin 2 
δ

 2 

• Π is Cole’s wake parameter and varies with pressure


gradient
• Π ≈ 0.6 for constant pressure (zero pressure gradient)
Modeling: A Brief History
• Time averaged N-S equations forms the basis of
most models
• Concept originated with Reynolds (1895)
• First attempts were modeled on gradient-diffusion
process
• Boussinesq (1877) introduced the concept of a so-
called eddy viscosity
• Prandtl (1904) discovers boundary layers
• Prandtl (1925) introduced the mixing length to
calculate eddy viscosity, an approach borrowed
from the mean-free path concept used to
characterize viscosity of gasses
• Mixing length models dominate for the next 20
years
A Brief History - continued
• Mixing length models are algebraic or zero equation models
• n-equation models – n refers to number of additional
differential transport equations used to include turbulent
effects
• Prandtl (1945) – introduces a model where eddy viscosity
depends on k, which is approximated with a PDE
– Takes history effects into account – one-equation model
– The model is incomplete as a length scale is required
– More than just boundary and initial conditions are required to
close the set of equations
– Length scales change as the flow changes
• Kolmogorov (1942) – introduces the first complete model k-w
– Introduces ω, “rate of dissipation of energy in unit volume and
time”
– ω satisfies a differential equation similar to that used for k
– Two-equation model
– Model went unnoticed for 25 years due to lack of computing
power
A Brief History - continued
• Rota (1945) and Chow (1951) lays foundation of models that
do away with Boussinesq approximation
– Rota devised a model for the differential equation governing
evolution of the turbulent stress tensor
– Stress-transport models
– Also referred to as second-order or second-moment closure
– Automatically includes effects such as streamline curvature,
rigid-body rotation and body forces
• By 1950 the following categories of turbulence models were
established
– Algebraic (zero-equation) models
– One-equation models
– Two-equation models
– Stress-transport models
• The rapid increase in computing power since the 1960’s has
seen development of all of these differing modeling strategies
Reynolds Averaging
• Three pertinent forms:
– Time averaging – stationary turbulence
(average does not change with time)
– Spatial averaging – homogeneous
turbulence (uniform in all directions)
– Ensemble averaging – decaying
turbulence
RA - continued
• Time averaging
t +T
FT (x ) = lim ∫ f (x, t )dt
1
T →∞ T
t

• Spatial averaging
FV (t ) = lim ∫∫∫ f (x, t )dV
1
V →∞ V
V

• Ensemble averaging
N
FE (x, t ) = lim ∑ f (x, t )
1
n
N →∞ N
n =1
RA - continued
• Stationary turbulence

ui (x, t ) = U i (x ) + u 'i (x, t )

• “Unsteady” stationary
turbulence
• T1 << T2

ui (x, t ) = U i (x, t ) + u 'i (x, t )


RA - continued
φ = Φ + φ ' , ψ = Ψ +ψ '

c1φ + c2ψ = c1Φ + c2 Ψ

∂u i ∂U i
≈ , ui ' << U i
∂t ∂t

φψ = ΦΨ + Φψ '+ Ψφ '+φ 'ψ ' = ΦΨ + φ 'ψ '


RA - continued
• Averaging process leaves continuity equation
unchanged
• Introduces new terms into N-S equations

∂U i
=0
∂xi

ρ
∂U i
∂t


∂x j
( )
U jU i + u j ' ui ' = −
∂P
+

∂xi ∂x j
(2µSij )

• The new terms are moved to the RH-side and


conveniently termed, Reynolds “stresses”

ρ
∂U i
∂t


∂x j ∂xi ∂x j
(
(U jU i ) = − ∂P + ∂ 2µSij − ρ u j ' ui ' )
RA - continued
• Reynolds stress tensor is symmetric and introduces 6 new
variables -u_x u_y = -u_y u_x; -u_x u_z = -u_z u_x; -u_y u_z = -u_z u_y
• Continuity and N-S: 4 equations
• Pressure, velocity and Reynolds stresses: 10 variables
• τij – specific Reynolds stress tensor

 − u x ' u x ' − u x ' u y ' − u x ' u z '


 
τ ij = − u y ' u x ' − u y ' u y ' − u y ' u z '
 
 − u ' u ' − u ' u ' − u ' u '
 z x z y z z 

Know this! Seemed to go back to it a lot!


Turbulence Intensity
• An expression of how intense the turbulent fluctuations
are
• Quantified in terms of the normal Reynolds stresses
• Relative intensity: normalized to free stream mean
velocity

^ u '2 ^ v '2 ^ w'2


u≡ , v≡ , w≡
Ue Ue Ue
TI - continued
• Incompressible flat-plate boundary layer

Isotropic further from


wall?
TI - continued
• Turbulent kinetic energy (per unit mass), k

1 2
( 1
)
k = u ' + v' + w' = ui ' ui ' RHS acts like dot product?
2
2 2

• Under the assumption of isotropic turbulence

2 k
T ' ≡ 100
3 U e2
Algebraic Models
Boussinesq hypothesis used to
• Boussinesq eddy-viscosity approximation: model Reynolds stress term

Reynolds stress tensor is replaced by the


product of an eddy viscosity and the strain-
rate tensor
• Eddy viscosity is computed in terms of the
mixing length
• Mixing length is analogous to the mean free
path in gas
• Molecular viscosity is a function of the gas
whilst the mixing length is a function of the
flow
Algebraic Models- continued
• A shear flow is considered, U=U(y)i
• Flux of momentum across y = 0 is considered
• Upward migrating molecules cause a deficit of
momentum and vice versa
Algebraic Models- continued
• At a molecular level the velocity can be decomposed as

u = U + u' '

• Instantaneous flux across the plane, y = 0, is


proportional to the velocity normal to the plane, v’’
• dpxy – instantaneous x-momentum flux across a
differential area normal to the y-direction

dp xy = ρ (U + u ' ' )v' ' dS )


Algebraic Models- continued
• Performing an ensemble average over all molecules

dPxy = ρu ' ' v' 'dS

• The stress on plane y = 0, σxy = dPxy/dS


• In fluid mechanics we write: σij = pδij – tij
• tij – viscous stress tensor

t xy = ρu ' ' v' '

• Very similar to the Reynolds stress tensor


Algebraic Models- continued
• From the kinetic theory of gasses it can be shown that
molecules moving between points Q and P, one mean
free path, lmfp, apart, set up shear stress at plane y = 0
1 dU
t xy ≈ ρvth lmfp
2 dy

• Where vth is the thermal velocity of the gas molecules


• For plane shear flow: txy = µ dU/dy

1
µ = ρvthlmfp
2
Algebraic Models- continued
• Prandtl (1925) – mixing length hypothesis
where clumps of fluid (eddies) behave in a
manner analogous to gas molecules
• lmix ↔ lmfp, vmix ↔ vth

dU
vmix = constant ⋅ lmix
dU νT = l 2
mix
dy dy

• Mixing length needs to be specified and is related to the


flow
Algebraic Models- continued
• For wall bounded flows the Knudsen number, KN, is
close to unity, casting suspicion on the assumed
linear relationship between turbulence stresses and
strain rate
• It can be shown that lmix ∼ vmix / |dU/dy| and thus the
clumps of fluid undergo significant changes
between points Q and P
• Theoretical basis for mixing length models very
shallow
• Models have however performed well for very
specific problems mostly due to tuning
• Seems to perform better for equilibrium turbulent
flows
Algebraic Models- continued
• Simple and easy to implement
• Numerically stable
• Work only well for flows to which they have been fine tuned
• Cebeci-Smith and Baldwin-Lomax reproduce skin friction and
velocity profiles faithfully for incompressible turbulent
boundary layers for weak pressure gradients
• Baldwin-Lomax superior only in the sense that the set-up is
less arduous for complex flows
• Both models perform badly for separated flows
• Johnson-King model
– Performs better for separated flows
– Many ad hoc closure coefficients spoil simplicity
– Requires an iterative procedure to function
– Specifically formulated for wall-bounded flows
– Performs well for transonic flows
1 & 2 Equations Models
• Both make use of a differential equation to
calculate k
• 1 equation models are incomplete and thus
a length scale based on the flow needs to
be defined
• 2 equation models solve an additional
differential expression from which a length
scale can be computed and is thus said to
be complete
• 1 equation models have never been very
popular or widely used
1 & 2 EM - continued
• The trace of the Reynolds stress tensor, τii = 2k
• Differential expressions for the terms in the Reynolds
stress tensor can be derived by averaging the product of
the N-S equations with a fluctuating property
• N(ui) – Navier-Stokes operator

∂ui ∂ui ∂p ∂ 2 ui
N (ui ) = ρ + ρu k + −µ
∂t ∂xk ∂xi ∂xk ∂xk

• N-S equation is thus N(ui) = 0

ui ' N (u j ) + u j ' N (ui ) = 0


1 & 2 EM - continued
?
∂τ ij ∂τ ij ∂U j ∂U i ∂  ∂τ ij 
+Uk = −τ ik − τ jk + ε ij − Π ij + ν + Cijk 
∂t ∂xk ∂xk ∂xk ∂xk  ∂xk 

p '  ∂ui ' ∂u j ' 


Π ij = +

p  ∂x j ∂xi 

∂ui ' ∂u j '


ε ij = 2ν
∂x j ∂xi

ρCijk = ρ ui ' u j ' uk ' + p' ui 'δ jk + p' u j 'δ ik


?

1 and 2 EM - continued
∂k ∂k ∂U i ∂  ∂k 1 1 
+U j = τ ij −ε + ν − ui ' ui ' u j ' − p ' u j '
∂t ∂x j ∂x j ∂x j  ∂x j 2 ρ 
Transport equation?
Derive!

∂ui ' ∂ui '


ε =ν
∂xk ∂xk

• LH – side: unsteady and convection terms


• RH-side
– 1st term: Production – rate of energy transfer from mean flow to turbulence
– 2nd term: Dissipation – rate at which k is converted into thermal energy
– ν∂k/∂xj: Molecular Diffusion – diffusion of turbulence energy due to fluid’s
natural diffusion transport processes
– Triple velocity correlation: Turbulent Transport – rate of turbulence energy
transport through fluctuations
– Pressure term: Pressure Diffusion – turbulent ransport due to the pressure-
velocity correlation
1 and 2 EM - continued
• Specific Reynolds stress tensor
– Boussinesq approximation is assumed
2
τ ij = 2ν T Sij − kδ ij
3
• Second term ensures that τii = -2k
• Turbulent Transport and Pressure Diffusion
– Turbulent Transport - Modeled as a diffusion process
– Pressure Diffusion - In the absence of any additional data it is
grouped with the preceding term (σk – closure coefficient)

1 1 ν T ∂k
ui ' ui ' u j ' + p ' u j ' = −
2 ρ σ k ∂x j
1 and 2 EM - continued

∂k ∂k  2  ∂U i ∂  ν T  ∂k 
+U j =  2ν T Sij − kδ ij  −ε + ν +  
∂t ∂x j  3  ∂x j ∂x j  σ k  ∂x j 

Derive! Transport equation?


1 and 2 EM - continued
• Dissipation
– Manner in which it is determined not unique amongst turbulence
models
– Along with a length scale, dissipation is needed to close the
turbulence model
– On dimensional grounds Taylor stated that

3
2
k
ε∝
l

• Specification of the length scale differs between


turbulence models
1 Equation Models
• Do not present a significant improvement over
algebraic models
• As such they will not be discussed in any detail
• Taylor’s argument is used to relate dissipation
rate to the length scale
• The length scale is then based on practices
similar to that employed for algebraic models

3
2
k
ε = CD
l
2 Equations Models
• The choice of the 2nd turbulence transport equation
is nearly unlimited
• Length and velocity scales are constructed based
on dimensional grounds
• No guarantee that dimensional analysis would
correctly mirror the physics of turbulence
• There is little reason to believe that 2-equation
models would be more accurate or physically
realistic than algebraic or 1-equation models
• A given is that they are numerically more difficult to
implement
2 EM: k-ε Model
• Most popular model
• Earliest work: Chou (1945), Davidov (1961) and Harlow &
Nakayama (1968)
• Jones and Launder (1972) – central paper Standard k-εε model
• Launder and Sharma (1974) – closure coefficients
• The exact equation for dissipation rate is formed by

∂ui ' ∂
2ν [N (ui )] = 0
∂x j ∂x j
• Results in a host of unknown terms that even defy
measurement
• Drastic surgery is performed to approximate these terms
based on dubious concepts and ideas
2 EM: k-ε Model -continued
ν T= Cµ k 2 / ε

∂k ∂k  2  ∂U ∂  ν T  ∂k 
+U j =  2ν T S ij − kδ ij  i − ε + ν +  
∂t ∂x j  3  ∂x j ∂x j  σ k  ∂x j 

∂ε ∂ε ε 2  ∂U i ε2 ∂  ν T  ∂ε 
+U j = Cε 1  2ν T Sij − kδ ij  − Cε 2 + ν +  
∂t ∂x j k 3  ∂x j k ∂x j  σ ε  ∂x j 

Cε 1 = 1.44, Cε 2 = 1.92, C µ = 0.09, σ k = 1.0, σ ε = 1.3


2 EM: k-ω Model

• Developed by Kolmogorov (1942) – 1st 2-equation


model
• 2nd parameter, ω, dissipation per unit turbulent kinetic
energy
• Saffman (1970), without prior knowledge of the
Kolmogorov-model developed an even better version
• Since then a host of improvements have been made
• As opposed to the ε-equation, the equation for ω has
been developed by a combination of physical
observation and dimensional analysis
2 EM: k- ω Model - continued
2 EM: Wall Functions
• Perturbation analysis show that most 2-equation models fail
to predict the turbulent velocity profile accurately
• It is possible to integrate the model equations through the
boundary layer, but an extremely fine grid is required
• One way of resolving this is to not solve for the boundary
layer but replace this with empirically determined wall
functions
• The relevant turbulent properties at the first node from the
wall is calculated from the wall functions and specified
• Wall functions do not perform well where flow separation
occurs
• Results are sensitive to the location of the first node from the
wall surface
1&2 EM: General Comments
• 1-Equation models holds only a slight
advantage over algebraic models
• 1-Equation models are numerically more
well-behaved than 2-equation models
• Recent 1-equation models make use of an
approximation for eddy viscosity and are
thus complete
– Baldwin-Barth (1990)
• Inaccurate for attached boundary layers
• Very inaccurate for detached flows
– Spalart-Allmaras (1992)
• Good prediction of skin friction
• Superior solutions than algebraic models for separated
flow
1&2 EM: General Comments
• k-ε most widely used
• k-ε even less accurate than Baldwin-Barth for flow with
adverse pressure gradients
• k-ε extremely difficult to integrate through viscous sub-layer,
requires viscous corrections
• k-ε requires a fair amount of fine-tuning to predict accurate
results
• k- ω more accurate for 2D boundary layers under favorable
and adverse pressure gradients
• Can be easily integrated through viscous sub-layer without
introduction of viscous-correction
• Performs well in recirculating flows
• With viscous correction the k- ω model reproduces subtle
features of turbulence kinetic energy close to solid walls and
even describes boundary-layer transition reasonably well

You might also like