Turbulence Models February 1–3, 2010

Outline
Turbulence Models • Review last lecture
• Nature of turbulence
• Reynolds-average Navier-Stokes (RNS)
Larry Caretto
Mechanical Engineering 692 • Mixing length theory
• Models using one differential equation
Computational Fluid Dynamics
• Two-equation models, especially k–ε
• Reynolds stress models
February 1–3, 2010
• Large-eddy simulation (LES)
• Direct numerical simulation (DNS)
2

Review Basic Equations Review Solution Approaches

• Have general equation to use in • Solve momentum and continuity for
numerical analysis approaches velocity and…
⎡ ∂ρφ ∂ρ ui ϕ ⎤ ∂ (ϕ ) ∂ϕ
c⎢ + = Γ + S (ϕ ) – Pressure at low Mach numbers where
⎣ ∂t ∂xi ⎥⎦ ∂xi ∂xi density does not depend on pressure
Transient Convective Diffusive " Source" – Density at high Mach numbers where
• Momentum equations separate pressure pressure is found from equation of state
gradient from other source terms ⎡ ρ ⎤
• Density solution approach ⎢ ⎥
∂ρu j ∂ρ ui u j ∂P ∂ ∂u j ⎢ ρu ⎥
Sj** = 0 for + =− + μ + S *j* ∂U ∂E ∂F ∂G ⎢ ρv ⎥
constant μ, ρ ∂t ∂xi ∂x j ∂xi ∂xi + + + =H U=⎢ ⎥
and Bj = 0 ∂ ∂ui ∂
∂t ∂x ∂y ∂z ⎢
ρw
⎥
⎡ 2 ⎤
S *j* = μ + ⎢⎣( κ − 3 μ) Δ ⎥⎦ + ρB j ⎢ρ( e + V 2
/ 2) ⎥
∂xi ∂x j ∂x j 3
⎢ ρW ( K 4)
⎥
⎣ ⎦

Review Fluent Exercise Nature of Turbulence

• Note there are two files case and data • Characterized by fluctuations in the flow
– Case file has grid information • Imperfect ability to model turbulence is
– Data file has results major problem in practical applications
– Can open and save both at same time of CFD
• Will learn about turbulence models this • Turbulent eddies are structures that
week and numerical algorithms next exist in the flow
week – Largest scale structures get energy from
• Use tutorial files as examples of how to main flow and transfer energy to smaller
use Fluent – Smallest scale structures get energy from
larger and dissipate energy to viscosity
5 6

ME 692 – Computational Fluid Dynamics 1

Turbulence Models February 1–3, 2010

Nature of Turbulence II landsat.gsfc.nasa.gov/

No smoking in CSUN • Turbulent
classrooms, but the flows in
smoke from a clouds
cigarette shows how
laminar flows can
from earth
transition into satellite
turbulent flows and shows
the eddy nature of turbulent
the turbulent flow structures
structures

Turbulent pipe flow video

7
Photo credit: Humphrey_Bogart_by_Karsh_(Library_and_Archives_Canada).jpg 8

Fluctuation Quantities Turbulent Kinetic Energy

• Model flow variables in terms of main • The kinetic energy per unit mass due to
flow properties and fluctuations fluctuations in velocity ui’
– Instantaneous value, φ ui ' ui ' u ' u '+v' v'+ w' w'
– Fluctuation value φ’ Δt k= =
1
∫
– Definition of mean value ϕ= ϕdt 2 2
– Basic result: Δt • Turbulent intensity = u’rms/|V|
0
ϕ = ϕ + ϕ' '
urms =
1
3
[ ]
(u ')2 + (v')2 + (w')2 3 = 2k
3
– Applied to velocity components sometimes
use Ui for mean velocity component
V = (u )2 + (v )2 + (w)2
ui = u i + ui ' = U i + ui ' 9 10

Turbulent Energy Transfer More on Turbulent Fluctuations

• Turbulent flows have a series of length • Definition shows that average
scales which transfer energy fluctuation is zero Δt
1
– At largest scales, turbulent structures get
kinetic energy from main flow
ϕ = ϕ + ϕ' ϕ=
Δt
ϕdt ∫
– Kinetic energy transfers from larger to Δt Δt Δt0

∫ (ϕ − ϕ )dt = Δt ∫ ϕdt − Δt ∫ ϕdt = ϕ − ϕ = 0

1 1 1
smaller length scales ϕ' =
Δt
– Energy is dissipated by viscous effects at 0 0 0
the smallest length scales • The mean value is a constant and the
• Large Reynolds numbers or Grashof average of a constant is just that
numbers mark transition to turbulence constant
11 12

ME 692 – Computational Fluid Dynamics 2

Turbulence Models February 1–3, 2010

Average of a Product Average of a Product

• The mean of the product of two flow • Two flow properties φ and ψ
Δt Δt Δt
properties φ and ψ, written as ψϕ, is the
∫ (ϕ − ϕ')(ψ − ψ')dt =
1 1 1
sum of two terms:
ψϕ =
Δt Δt ∫ ϕ ψ dt −
Δt ∫ ϕψ' dt
0 0 0
– The product of the means of each Δt Δt Δt Δt
individual term ϕ and ψ 1 1 1 1
– The mean of the product of the two
−
Δt ∫ ϕ' ψ dt +
Δt ∫ ϕ' ψ ' dt = ϕ ψ
Δt ∫ dt − ϕ
Δt ∫ ψ' dt
0 0 0 0
fluctuation quantities ϕ'ψ' (correlation term) Δt Δt Δt
1 1 1
ψϕ = ϕ ψ + ϕ'ψ ' −ψ
Δt ∫ ϕ' dt +
Δt ∫ ϕ' ψ ' dt = ϕ ψ − 0 − 0 +
Δt ∫ ϕ' ψ' dt
0 0 0
– Although ϕ' and ψ' are zero ϕ'ψ ' is not zero
– Derivation on next slide
13
ψϕ = ϕ ψ + ϕ'ψ ' 14

More on Averages Reynolds-Average

• Root-mean square properties based on • Reynolds average transport equation
turbulent fluctuations (including Navier Stokes, called RANS)
Δt – Start with general transport equation
1
Δt ∫0
ϕ rms = (ϕ ' ) 2 = (ϕ ' ) 2 dt
⎡ ∂ρφ ∂ρ ui ϕ ⎤ ∂ (ϕ ) ∂ϕ
c⎢ + = Γ + S (ϕ )
⎣ ∂t ∂xi ⎥⎦ ∂xi ∂xi
• Average of a space derivative is space
derivative of average – Look at steady-state, zero-source, constant
properties
∂ϕ 1 ∂ϕ
Δt
∂ ⎡1 ⎤ ∂ϕ
Δt
Γ (φ )
= ∫
∂xi Δt 0 ∂xi
dt =
∂xi
⎢ ∫ ϕdt ⎥ =
⎣ Δt 0 ⎦ ∂xi
∂ui ϕ
= γ (ϕ )
∂ ∂ϕ γ (φ ) =
ρc
15
∂xi ∂xi ∂xi 16

What is γ(φ) = Γ(φ)/ρc? Reynolds Average II

• Recall definition of Γ(φ) as general • Take time average of last equation
transport coefficient
Δt Δt
1 ⎡ ∂ui ϕ ⎤ 1 ⎡ (ϕ ) ∂ ∂ϕ ⎤
Δt ∫0 ⎣ ∂xi ⎦ Δt ∫0 ⎣
φ u v w e h T T ⎢ ⎥ dt = ⎢γ ⎥ dt
Γ(φ) μ μ μ k/cv k/cp k k ∂xi ∂xi ⎦
c 1 1 1 1 1 cv cp • Time average of derivatives are
γ(φ) μ/ρ μ/ρ μ/ρ k/ρcv k/ρcp k/ρcv k/ρcp derivatives of time average
• Dimensions for γ(φ) are L2/T ∂ ui ϕ ∂ ∂ϕ
– Kinematic viscosity, ν = μ/ρ k is thermal = γ (ϕ )
– Thermal diffusivity α = k/ρcp
conductivity ∂xi ∂xi ∂xi
17 18

ME 692 – Computational Fluid Dynamics 3

Turbulence Models February 1–3, 2010

Reynolds Average III Reynolds Average IV

• Use expression for average of a product • Combine fluctuation product term with
to compute average of uiφ diffusive flux
(
∂ ui ϕ ∂ ui ϕ + ui ' ϕ '
=
)
= γ (ϕ )
∂ ∂ϕ ∂ ui ϕ
=
∂ ⎡ (ϕ ) ∂ϕ
⎢γ
⎤
− ui ' ϕ ' ⎥
∂xi ∂xi ∂xi ∂xi ∂xi ∂xi ⎣ ∂xi ⎦
• Compare to equation before averaging • Boussinesq approximation: the turbulent
– Overall values replaced by averages fluctuation is proportional to the gradient
– Add a new term: the average of the of the mean property with empirical
product of two fluctuations turbulent transport coefficient, γt(φ)
• Have to compute this product term ∂ϕ
ui ' ϕ' = − γ t(ϕ)
19 ∂xi 20

Reynolds Average IV Momentum Terms

• Combine fluctuation product term with • Apply general formulation to momentum
diffusive flux equation where φ = uj in j direction
∂ui ϕ ∂ ⎡ (ϕ ) ∂ϕ ∂ϕ ⎤ ∂ ⎡ (ϕ ) (ϕ ) ∂ϕ ⎤
– For momentum equations, γ(φ) = ν
= ⎢γ l + γ t(ϕ ) ⎥= ⎢(γ l + γ t ) ⎥
∂xi ∂xi ⎣ ∂xi ∂xi ⎦ ∂xi ⎣ ∂xi ⎦ • Also have time averaged pressure
• Turbulent transport coefficient, now γt(φ) gradient ∂p/∂xj which is ∂p /∂xj
contains our ignorance about turbulence • Time average of momentum equation
• Seek ways to model this term gives
• Usually much larger than laminar term ∂ ui u j ∂ ui ' u j ' 1 ∂p ∂ ⎡ ∂u j ⎤
+ =− ⎢ν − ui ' u j ' ⎥
so that latter is often neglected ∂xi ∂xi ρ ∂x j ∂xi ⎣⎢ ∂xi ⎦⎥
21 22

Momentum Terms II Turbulent Viscosity

• Move fluctuation product to right side • Arguments based on dimensional
∂ ui u j ∂p ∂ ⎛⎜ ∂ u j ρui ' u j ' ⎞⎟
analysis
=− + ν − • Define characteristic velocity scale, ,
∂xi ∂x j ∂xi ⎜⎝ ∂xi ρ ⎟⎠
and length scale, l
• Resulting terms are called the Reynolds
stress terms • Consider kinematic viscosity, ν, with
• Nine such terms but only six are unique dimensions of L2/T
• Dimensions of  and l and L/T and L
• Simple models look at only one term
• Advanced models try to compute all six • Dimensionally correct choice for n is
product of  and l or νT = C  l
23 24

ME 692 – Computational Fluid Dynamics 4

Turbulence Models February 1–3, 2010

Turbulence Modeling Mixing Length Theory

• Reynolds averaging terms like ρu’φ’
modeled by a turbulent transport
coefficient γt(φ) , e.g., turbulent viscosity ν
• To use this approach we have to find
ways to compute ν and the general γt(φ)
• Various turbulence models proposed
– Some use simple concepts
– Others require numerical solution of one or
more partial differential equations (PDE)
• PDEs have same form as other CFD equations
25
• Originally proposed by Prandtl in 1925
• Basic idea
– Consider a turbulent fluid with a mean
temperature gradient
– A fluid particle that moves from the cold
region to the hot region will take on the
characteristics of the hot region after it
moves through one mixing length, l
– The temperature fluctuations are related to
the mean gradient as ∂T
T ' = −l
∂y 26

Mixing Length Mixing Length II

• By dimensional arguments νt = μt/ρ = • If νt = μt/ρ = Cl, and  = cl|∂U/∂y|, then
Cl, where C is dimensionless constant νt = Ccl|∂U/∂y|l = Kl2|∂U/∂y|
• For simple flows have only one important • This gives the Reynolds stress as
Reynolds stress, -ρu’v’ follows 2 ∂u ∂u
− ρ v ' u ' = − ρl
• If length scale is measure of largest eddy ∂y ∂y
size, one possible dimensionally correct • Other flow properties use turbulent
velocity scale is  = cl|∂U/∂y| transport ratio, σt = μt/Γt(φ)
– c is dimensionless constant which is
different from constant C in νt = Cl 1 2 ∂u ∂φ
− ρv' φ' = − ρl
27
σt ∂y ∂y 28

Boundary Layers Boundary Layer Profile

• Turbulent flow next to a wall has laminar • Flow in x direction (velocity u) with y as
sublayer and transition to fully turbulent distance perpendicular to wall y
u, x
flow – Wall shear stress = τw
• Experimental data and simplified – uτ is called friction velocity wall

analysis give empirical equations for • Define dimensionless variables

velocity profile of turbulent boundary – y+ is dimensionless distance from wall
layer – u+ is dimensionless velocity parallel to wall
• Simplified theory used for turbulent “wall τw u y uτ y
functions” to give boundary conditions uτ = u+ = y+ = =
for turbulent flows in CFD ρ uτ yτ ν
29 30

ME 692 – Computational Fluid Dynamics 5

Turbulence Models February 1–3, 2010

Boundary Layers II y+ = 30

• Law of the wall: u+ = f(y+)

• No equation 5 < y+ < 30
y+ = 5
• In other regions
⎧
⎪ y+ y+ < 5
⎪ + +
⎪ ln( y ) ln( Ey )
u+ = ⎨ +B= 30 < y + < 500
⎪ κ κ
⎪ u + − 1 ln⎛⎜ y ⎞⎟ − A y + > 500
⎪⎩ max κ ⎝ δ ⎠

• B = 5.5; E = 9.8, κ = 0.41

31 C. Delo and A.J. Smits, Volumetric visualization of coherent structure in a low Reynolds number turbulent
32
boundary layer, International Journal of Fluid Dynamics (1997), Vol. 1, Article 3, Figure 6, downloaded
1/31/2010 from http://www.eng.nus.edu.sg/ijfd/1997_vol1/paper3/figures/Figure6.html

Spalart-Allmaras Model k-epsilon (k-ε) Model

• Solves one PDE for ~ ⎛~
3
ν ~
ν⎜ ⎟
ν⎞ • Most popular model of turbulence
which is used to find νt νt = ⎝ν⎠ – Original and modified versions
⎛~
3
from equation at right ν⎞ • Works well in confined flows with no
⎜ ⎟ + Cv1
3
– Default value: Cv1 = 7.1 ⎝ν⎠ rotating parts
– In most of flow away from walls νt = ~
ν • Less applicable in external flows and
• Developed for aerospace applications oscillating flows
to external flows with walls • Requires solution of two PDEs one for
• Has recently been applied to other flow turbulent kinetic energy, k, and one for
situations turbulent dissipation, ε
33 34

What is ε? Computing νt from k and ε

• Dissipation rate, ε, is rate of kinetic • Back to basic equation that νt is the
energy transfer from smallest eddies product of a length scale and a velocity
working against the viscous forces scale: νt = μt/ρ = Cμ  l,
• Defined in terms of deformation rates, eij • Use  = k1/2 as velocity scale
∂u ∂uj 2 • Length scale, l = k3/2/ε
eij = i + τ ij = μeij + (κ − μ )Δδ ij
∂x j ∂xi 3 • Result for viscosity is νt = μt/ρ = Cμ k2/ε
• Definition of ε is ε = 2ν e'ij e'ij • Solve PDEs for k and ε
• Dimensions of ε are energy divided by • Can derive form of equation for k
[(mass)(time)], e.g. m2/s3
35 36

ME 692 – Computational Fluid Dynamics 6

Turbulence Models February 1–3, 2010

k Equation Modeling k Equation Terms

• Can derive following balance equation • Turbulent diffusion of kinetic energy
∂ρk ∂ρ ui k ∂ ∂k ∂ ⎛ρ ∂u ⎛ρ ⎞ μ ∂k
+ = μ −
⎞
⎜ ui ' u j ' u j ' + p' ui ' ⎟ + ρ ui ' u j ' i − ρε − ⎜ u i ' u j ' u j ' + p ' ui ' ⎟ ≈ t
∂t ∂xi ∂xi ∂xi ∂xi ⎝ 2 ⎠ ∂x j ⎝2 ⎠ σ k ∂xi
• Can identify transient, convection and • Empirical constant: σk
diffusion terms; remainder is source • Production term, Pk , uses model for
• Terms with three fluctuating Reynolds stresses
components cannot be computed ∂ui ⎛∂u ∂uj ⎞ ∂ui
Pk = ρ ui ' u j ' = −μt ⎜ i + ⎟
without introducing terms with four ∂x j ⎜ ∂x j ∂xi ⎟ ∂x j
⎝ ⎠
fluctuating components
• Implied summation over two indices
37 38

The k Equation The ε Equation

• Final form with modeled terms • Derivation similar to that for k
– Obtain partial differential equation with
∂ρk ∂ρ ui k ∂ ⎛ μ ⎞ ∂k
+ = ⎜μ + t ⎟⎟ + Pk − ρε terms that have to be modeled
∂t ∂xi ∂xi ⎜⎝ σk ⎠ ∂xi
∂ρε ∂ρ ui ε ∂ μ t ∂ε ε ε2
+ = + Cε 1 Pk − ρCε 2
• Usually ignore laminar viscosity ∂t ∂xi ∂xi σ ε ∂xi k k

• Empirical constants in k-ε model

∂ρk ∂ρ ui k ∂ μt ∂k
+ = + Pk − ρε
∂t ∂xi ∂xi σ k ∂xi Cμ = 0.09 Cε 1 = 1.44 Cε 2 = 1.92 σ k = 1.0 σ ε = 1.3

39 40

Boundary Conditions What is uτ?

• Must specify values of k and ε at all • Friction velocity, uτ = (τw/ρ)1/2
boundaries (walls, inlets, outlets, etc.) • Need a value for τw
• Solid walls use “wall functions” to • Assume production and dissipation of
provide correct result at first node turbulence are nearly equal near wall
• Based on boundary layer results for flat • Derive following equation: τw = (Cμk2)1/4
walls • Combine with u+ equation to get result
– u is velocity parallel to wall for u at first node in from the wall
– y is direction normal to wall
⎛ EyC 14 k ⎞
u ln( Ey + ) 1 ⎛ Eyuτ ⎞ ⎛ Eyuτ ⎞ 1
⎜ μ ⎟
+ u = uτ ln⎜ ⎟ = Cμ 4 k ln⎜
u = = = ln⎜ ⎟ ⎝ ν ⎠ ν ⎟
uτ κ κ ⎝ ν ⎠ 41 ⎝ ⎠ 42

ME 692 – Computational Fluid Dynamics 7

Turbulence Models
Inlet and Outlet Conditions
• Ideally have data on similar flows that
relate k and ε to inlet properties
• Failing that use the following equations
3 k
k = Cuinlet
2
l = 0.07 L ε = Cμ 4
• C, proportional to square of turbulence
intensity, typically about 0.01 to 0.05
• L is a characteristic length of the inlet
(e.g., the hydraulic diameter)
• Use zero gradient conditions for outlet
• k = ε = 0 for free stream
43
Reynolds Stress Models
• Solve a differential equation for each
unique Reynolds stress
– PDEs for u’u’, u’v’, u’w’, v’v’, v’w’ w’w’
– Find k = (u’u’ + v’v’ + w’w’)/2
– Also solve ε equation
• Should work well for strongly anisotropic
flows, but modeling assumptions may
limit accuracy of model
• Usually applied to flows with rotating
flow or swirl
45
Large Eddy Simulation (LES)
• Use small grid scale to compute largest
eddies as part of the calculation
• Use sub grid scale models to get results
for the finest turbulence scales
• Now available in production codes
• Generally not worth the extra
computational cost except for complex
flows
47
ME 692 – Computational Fluid Dynamics
February 1–3, 2010
Low Reynolds Numbers
• Special wall functions are required for
“low” Reynolds numbers flows
• Define wall damping functions, f
– Use fμ to compute μt = Cμfμk2/ε
l – Use f1 and f2 to modify production and
dissipation terms in ε equation
– Use laminar viscosity in addition to
turbulent viscosity in both k and ε
equations
– Different forms for damping functions
44
Reynolds Stress Models II
• Other transport equations compute
turbulent viscosity, νt, and use Prandtl
number, σφ, to get γ(φ) = νt / σφ
• Algebraic Reynolds stress model is
simplification that solves k and ε
equations then gets Reynolds stress by
six simultaneous algebraic equations
– Equations for u’u’, u’v’, u’w’, v’v’, v’w’ w’w’
– Other turbulence quantities: γ(φ) = νt / σφ
46
Direct Numerical Simulation (DNS)
• Use a grid fine enough to resolve the
smallest turbulence structures which are
10-4 to 10-5 m
• A volume of 30 m3 for flow around a
motorcycle would require 3x1012 nodes
• DNS not practical for engineering
calculations, but an important research
tool for examining turbulence properties
and testing other turbulence models
48
8
Turbulence Models
Model Guidance
• Use model which has been used
previously for your problem
– Previous work at your organization or from
literature research
• Consult user’s manual for CFD code
regarding increased computer time and
memory use for more complex models
• Use default constants in model unless
you have specific data to justify
alternative values
49
Model Guidance III
• Whatever model you use, make sure
that you have proper boundary
conditions
– Use special wall functions for non-
equilibrium turbulence when laminar
sublayer is not resolved
– Use correct grid spacing for first node from
wall for choice of wall functions or resolving
laminar sublayer
• Check this after calculations
51
Conclusions II
• Spalart-Allmaras model developed
especially for aerospace applications
with wall-bounded flows
– One equation model
• Relatively new model now seeing
applications in areas other than its initial
aerospace applications
– Adverse pressure gradients
– Turbomachinery
53
ME 692 – Computational Fluid Dynamics
February 1–3, 2010
Model Guidance II
• Review material on turbulence models
to see if they can handle unusual
features of the flow you are modeling
– Low Reynolds number (non-equilibrium)
turbulence
– High strain rates
– Adverse pressure gradients
– Rotating machinery
– Compressible flows
– Other complex flows
50
Conclusions
• k-ε most common turbulence model for
non-aerospace engineering applications
– widely regarded as having many
shortcomings in representing turbulence
– most widely validated model
– probably best choice for applications
without strong directional effects or
rotational flows
– Renormalizable group and realizable k-ε
models can give better results
52
Conclusions III
• Reynolds stress model has had success
for flows with directional effects and
rotational flows
– Requires solution of seven partial
differential equations to compute turbulent
viscosity
– Algebraic version has been used
• Other models available which may have
better accuracy for limited range of
flows
54
9
Turbulence Models February 1–3, 2010

Conclusions IV
• LES used for complex flows particularly
transient and oscillating flows
• Not usually required for common
engineering problems
• Choice of turbulence model should be
based on previous success of model in
similar applications
• No one right model to choose

55

ME 692 – Computational Fluid Dynamics 10