Notes for:

Turbulence Modelling
(in 5C1218 and 5C5112)
Stefan Wallin
Dept. of Mechanics, KTH and Systems Dept, FOI
stefan.wallin@foi.se
March 8, 2006

Outline
• Introduction and background
• One-equation models
• Two-equation models
• Reynolds stress models

1
Introduction
Turbulent flow governed by Navier-Stokes equation. Why modelling?
• DNS: Direct Numerical Simulation
- Solution of the 3D time dependent Navier-Stokes equations
- Computational effort scales rapidly with Reynolds number
- Used for low Reynolds number generic flows
- Gives detailed knowledge about turbulence
- Not (yet?) practically useful for “CFD applications”
• Computational effort for wall bounded turbulence
- Number of grid points
L2δ L3( δ ⁄ L ) LU 3 δ u τ 3
N nodes ∼ --------- ∼ --------------------- ∼ ⎛ --------⎞ ⎛ ---⎞ ⎛ -----⎞ ∼ Re L3 Re L– 1 / 5 Re L– 3 / 10 ∼ Re L5 / 2
l3 * ( ν ⁄ u ) 3 ⎝ ν ⎠ ⎝ L⎠ ⎝ U ⎠
τ
- Number of time steps
T L ⁄ U LU u τ 2
N ∆t ∼ ---- ∼ ------------- ∼ -------- ⎛ -----⎞ ∼ Re L Re L– 1 / 5 ∼ Re L4 / 5
t* ν ⁄ u 2 ν ⎝ U ⎠
τ
- Computational effort
CPU ∼ N nodes N ∆t ∼ Re L33 / 10 ∼ Re L3.3

- Re L ∼ 10 4 can be simulated ( N nodes ∼ 10 8 , one CPU-year on a PC).

Full scale airplane with Re L ∼ 10 8 requires N nodes ∼ 10 18 and 10 13

times more CPU. Can be done in year 2070 if the computers con-
tinue to develop according to Moorse’s law (doubled every 18
month).

Thus, turbulent flows need to be modelled.

2
Turbulence modelling
Reynolds averaged Navier-Stokes equations (RANS)
• Navier-Stokes equations ( ũ, p̃ denote instantaneous values)
- Conservation of mass
∂ũ i
= 0 (1)
∂ xi
- Conservation of momentum
∂ũ i ∂ũ i 1 ∂p̃ ∂ ⎛ ∂ũ i ⎞
+ ũ k = – --- ------- + ⎜ν ⎟ (2)
∂t ∂ xk ρ ∂x i ∂ x k ⎝ ∂ x k⎠
- Conservation of a passive scalar quantity θ̃
∂θ̃ ∂θ̃ ∂ ⎛ ∂θ̃ ⎞
------ + ũ k -------- = D -------- (3)
∂t ∂x k ∂ x k ⎝ ∂x k⎠

• Reynolds decomposition
ũ i ( x, t ) = U i ( x ) + u i ( x, t )

where U i ( x ) = ũ i ( x, t ) and u i ( x, t ) = 0

• The “mean” is time average, ensemble average or averaging in homo-

geneous directions. U i ( x ) may actually vary in time ( U i ( x, t ) ) with a
time scale much longer than the turbulent time scale.
• Take the mean of the Navier-Stokes equations -> RANS
∂U i
= 0 (4)
∂ xi

DU i 1 ∂P ∂ ⎛ ∂U i ⎞
---------- = – --- ------- + ⎜ν – u i u k⎟ (5)
Dt ρ ∂x i ∂ x k ⎝ ∂ x k ⎠
∂ ⎛ ∂Θ
D -------- – u k θ⎞
DΘ
--------- = (6)
Dt ∂ x k ⎝ ∂x k ⎠
D ∂ ∂
where ------ ≡ ---- + U k --------
Dt ∂t ∂x k

3
• Reynolds stresses, u i u j (and scalar flux vector u i θ )
- Appears because of the non-linear term
- Not “small”
- Significant effects on the flow
- Needs to be modelled in terms of mean flow quantities
- Reduces the problem to steady (or slowly varying)
- 2D assumptions possible

The aim is to construct a closed system of equations for the one-point

quantities ( U i, P, u i u j ).
• Only one governing lengthscale
• Spectra are assumed to be self-similar

Equation for the Reynolds stress tensor, u i u j (and for u i θ )

• Can be derived from the Navier-Stokes equations
• Contains, however, higher order moments like u i u j u k
• Equations for u i u j u k contain even higher moments: the closure problem
• Modelling needed at some stage.

4
Equation for the Reynolds stress tensor, u i u j
Equation for the fluctuating velocity u i is derived by subtracting the RANS
equation (5) from the Navier-Stokes equation (2)
∂u i ∂ ( U i + u i ) ∂U i
------- = ------------------------- – --------- = … (7)
∂t ∂t ∂t
Then, the equation for u i u j is derived by the following
∂u i u j ∂u j ∂u i
------------- = u i ------- + u j ------- = … (8)
∂t ∂t ∂t
The equation for u i u j is then derived by taking the mean of (8)

Du i u j
-------------- = Pij – εij + Π ij + Dij (9)
Dt
where
• Production, transfers energi from mean flow to fluctuations
∂U ∂U
Pij = – ui uk ---------j – u j uk ---------i (10)
∂x k ∂x k

• Dissipation, transfers energy from fluctuations to heat

∂u i ∂u j
ε ij = 2ν -------- -------- (11)
∂x k ∂x k

• Pressure-strain, redistributes energy among the components ( Π ii = 0 )

1 ⎛ ∂u i ∂u j⎞
Π ij = --- p ⎜ ------- + -------⎟ (12)
ρ ⎝ ∂x j ∂x i⎠

• Diffusion, redistributes fluctuations in space

∂ 1 ∂u i u j
Dij = – -------- u i u j u + --- ( u i pδ jk + u j pδ ik ) – ν ------------- (13)
∂x k k ρ ∂x k
The equation for the turbulent kinetic energy, K ≡ u i u i ⁄ 2 , is derived by tak-
ing half of the trace of the equation (9)
DK
-------- = P–ε+D (14)
Dt

5
where
∂U i
P = – u i u k --------- (15)
∂x k

∂u i ∂u i
ε = ν -------- -------- (16)
∂x k ∂x k
∂ 1 1 ∂K
D = – -------- --- u i u i u + --- u k p – ν -------- (17)
∂x k 2 k ρ ∂x k

Example: Homogeneous shear flow

∂U
------- = S , V = W = 0 , Dij = D = 0
∂y
Gives equation for K
∂K
------- = – uvS – ε
∂t
and equations for the Reynolds stresses

∂u 2
--------- = – 2uvS + Π 11 – ε 11
∂t
∂v 2
-------- = Π 22 – ε 22
∂t
∂w 2
---------- = Π 33 – ε 33
∂t
∂uv
--------- = – v 2 S + Π 12 – ε 12
∂t
Note: all energy production in the u 2 component and redistributed to the
other components by Π ij to an asymptotic state (Tavoularis & Corrsin,
1981) where

u2 v2 w2 uv SK
------- = 0.53 ------- = 0.19 ------- = 0.28 ------- = – 0.15 ------- = 6
2K 2K 2K 2K ε
which give

P – uvS uv SK
---- = ------------ = – 2 ------- ------- = 2 × 0.15 × 6 = 1.8
ε ε 2K ε

6
Eddy-viscosity models (EVM)
Boussinesq (1877) made the assumption
∂U
uv = – ν T -------
∂y
This was later generalized to the eddy-viscosity, or Boussinesq, assump-
tion
2
u i u j – --- Kδ ij = – 2ν T S ij (18)
3
where the symmetric and antisymmetric parts of the velocity gradient are

1 ⎛ ∂U i ∂U j⎞ 1 ⎛ ∂U i ∂U j⎞
S ij ≡ --- ⎜ + ⎟ , Ω ij ≡ --- ⎜ – ⎟
2 ⎝ ∂ xj ∂ xi ⎠ 2 ⎝ ∂ xj ∂ xi ⎠
and the Reynolds stress anisotropy

ui uj 2
a ij ≡ --------- – --- δ ij (19)
K 3
The eddy-viscosity assumption may be written
2ν T
a ij = – --------- S ij (20)
K
The scalar flux is modelled accordingly
∂Θ
u i θ = – D T ------- (21)
∂x i
where the eddy diffusivity D T may be related to the eddy viscosity by a tur-
bulent Schmidt number (of the order of unity).
νT
D T = ------ (22)
σT
The RANS equation (5) can now be rewritten as
DU i ∂ P 2 ∂
---------- = – ------- ⎛ --- + --- K⎞ + [ 2 ( ν + ν T )S ij ] (23)
Dt ∂x i ⎝ ρ 3 ⎠ ∂ x k
where the turbulent kinetic energy, K , may be absorbed into an effective
pressure.

7
The following observations can be made:
• Zero strain ( S ij = 0 ) gives zero anisotropy ( a ij = 0 ). History effects are
not well described.
• S ij and a ij are not in general aligned and the deviation is large. Thus,
the eddy-viscosity relation is only valid for one component in thin shear
layers (as originally proposed by Boussinesq). This is also the case for
the scalar flux that is not aligned with the gradient.
• S ij is invariant of rotation, which will only enter into Ω ij . Since a ij
depends on S ij but not on Ω ij no rotational effects will enter into a ij .
Also the K equation is unaffected by rotation. So, rotation effects are
not well described by eddy-viscosity models.

The eddy viscosity is related to the large scale (most energetic) turbulent
scales as
νT ∼ V ⋅ L (24)
where V and L are the velocity and length scales respectivelly.

Classification of eddy-viscosity models:

• Algebraic models or zero-equation models:
- V and L related to mean flow field and global geometry (wall dis-
tance, wake thickness, et.c.)
- Works well for attached boundary layers and other thin shear layers
- Not very general
• One-equation models:
- One transport equation for K or ν T .
- Additional information from global conditions (typically wall distance)
- Works well for attached boundary layers and other thin shear layers
- Not very general, but more than algebraic models
- Example: Spalart-Allmaras (1992) (a reasonable and robust model
for external aerodynamics)

8
• Two-equation models:
- Two transport equations for the turbulence scales, typically K – ε or
K – ω.
- Completely determined in terms of local quantities (except near-wall
corrections which may be dependent on wall distance)
- Works well for attached boundary layers
- Somewhat more general than algebraic and one-equation models.
- Model transport equations loosely connected to the exact equations.
- Examples:
Standard K – ε model (Launder & Spalding 1974)
Wilcox (1988, ...) K – ω models
Menter (1994) SST K – ω model (performing reasonable well also in
separated flows)

9
Zero-equation models
Constant turbulent viscosity, ν T (and diffusivity D T )
• Fair assumption only in free shear flows
• elsewhere, no general constant ν T can be found.

Mixing length models

∂U and L ∼ l
V ∼ l m ------
- m
∂y
which gives:
∂U ∂U
2 ∂U or – uv = l 2 ------
νT = lm ------- m ∂y- ------- (25)
∂y ∂y
• Free shear flows: Constant l m ⁄ δ is a common choice. Ex:
- plane jet: l m ⁄ δ ≈ 0.9
- round jet: l m ⁄ δ ≈ 0.075
- plane wake: l m ⁄ δ ≈ 0.16
• Wall-bounded flows: l m chosen as a ramp function with κ = 0.4 and
λ = 0.1
lm l m = λδ
κy
=
lm

λ⁄κ y⁄δ

• Lack of generality!
• Example:
- Baldwin-Lomax (1978) (rather accurate for attached boundary lay-
ers)

10
One-equation models
The basic idea is to decouple the turbulent velocity scale, V , from the
mean strain field and determine it from a transport equation. If the trans-
port equation for the turbulent kinetic energy, K , is used the velocity scale
may be derived from that as V ∼ K (Kolmogorov & Prandtl)

ν T = c' µ KL (26)
where c' µ is an empirical constant and the equation for K (14) is
DK
-------- = P–ε+D (27)
Dt
where the terms are modelled as

P = 2ν T S ij S ij

K3 / 2
ε = c D ------------
L

∂ ⎛ ν T⎞ ∂K
D = -------- ⎜ ν + ------⎟ --------
∂x k ⎝ σ k⎠ ∂x k
with the model coefficients σ k = 1.0 and c' µ c D = 0.09
The length scale, L , needs to be determined. This is a problem, as with
zero-equation models.

Assume local equilibrium P≈ε and thin shear layers:

∂U 2
P = 2ν T S ij S ij = ν T ⎛ -------⎞
⎝ ∂y ⎠

K3 / 2 ( KL ) 3 ν T3
ε = c D ------------ = c D ------------------- = c D -------------
L L4 L 4 c' µ3

cD ν 2 c' µ3
∂U 2 ∂U
P ≈ ε => ⎛⎝ -------⎞⎠ = -------------
T
=> ν T = ------- L 2 -------
∂y L 4 c' µ3 cD ∂y
One-equation models when P≈ε is a mixing length model with
1/4
⎛ c' µ3 ⎞
l m = ⎜ -------⎟ L
⎝ cD⎠
Or, l m hypothesis only suitable when P ≈ ε . Fair in boundary layers.

11
Spalart & Allmaras one-equation model.
One attempt to try to avoid the problem with the length scale is to form a
transport equation for ν T directly. This has been proposed by Spalart & All-
maras (1994). The model reads (somewhat simplified)
Dν T
---------- = P ν T – ε ν + D ν T (28)
Dt T

The production of ν T is modelled as

PνT
= c b1 Sν T
where S is the absolute value of the vorticity (reduces to ∂U ⁄ ∂y in thin
shear layers). c b1 is a model constant.
The dissipation of ν T is modelled as
νT 2
ε ν = c w1 f w ------⎞
⎛
T ⎝ d⎠
where c w1 is a constant and f w is a complex function taking care of the
outer part of the boundary layer. Here, a length scale is actually introduced
with d , which is the wall distance.
The Spalart & Allmaras model consists of a lot of different empirical correc-
tions and has been carefully calibrated for different flows and performs
well in many different cases.

12
Two-equation models (eddy viscosity)
• Eddy-viscosity assumption
2ν T
a ij = – --------- S ij (29)
K
Remember:

ui uj 2 1 ⎛ ∂U i ∂U j⎞ 1 ⎛ ∂U i ∂U j⎞
a ij ≡ --------- – --- δ ij , S ij ≡ --- ⎜ + ⎟ , Ω ij ≡ --- ⎜ – ⎟
K 3 2 ⎝ ∂ xj ∂ xi ⎠ 2 ⎝ ∂ xj ∂ xi ⎠

• Model equations (standard K – ε model):

DK
-------- = P – ε + D
Dt
(30)
Dε
------- = Pε – εε + Dε
Dt

Modelling and calibration of the different terms

• The eddy-viscosity relation:
The standard form is derived from the natural velocity and length scales,
V ∼ K and L ∼ K 3 / 2 ⁄ ε , which gives

K2
ν T = C µ ------ (31)
ε
In thin shear flows P ≈ ε , which gives
∂U
– uv -------
P ∂y K ∂U K ∂U 1
---- = 1 = ----------------- = – a 12 ---- ------- => ---- ------- = -----------
ε ε ε ∂y ε ∂y – a 12
used in the eddy-viscosity relations (29, 31)
ν T ∂U K ∂U Cµ
2
– a 12 = ------ ------- = C µ ---- ------- = ----------- => C µ = a 12
K ∂y ε ∂y – a 12
In thin shear layers (e.g. boundary layers) – a 12 ≈ 0.3 (Bradshaw hypothe-
sis), which gives C µ = 0.09
• Production of K , P :
The exact form reduces to the modelled by applying the eddy-viscosity
relation (29)
∂U i
exact: P = – u i u k --------- , modelled: P = 2ν T S ij S ji
∂x k

13
• Destruction terms ε and ε ε :
ε in the K equation is modelled as a transport equation. The equation for
ε is derived from the Navier-Stokes equation and the production, destruc-
tion and diffusion terms all need modelling.
The destruction term in the ε equation, ε ε , is modelled by considering
homogeneous decaying turbulence
dK dε
------- = – ε , ----- = – ε ε
dt dt
The assumption is that the decay rates of K and ε should be self similar
and differ by a constant factor
dε
----- ⁄ ε
dt ε2
--------------- = C ε2 => ε ε = C ε2 -----
dK K
------- ⁄ K
dt
The evolution of K can then be derived as

K(t) ⎛ ε0 ⎞ –n 1
---------- = ⎜ 1 + ( C ε2 – 1 ) ------t⎟ , n ≡ ------------------
K0 ⎝ K0 ⎠ C ε2 – 1
and C ε2 is calibrated from experiments. C ε2 can also be derived from the-
oretical analyses of the energy spectra. “Standard” value C ε2 = 1.92 .
• Production of ε , P ε :
Mainly dimensional arguments (weak coupling to the exact term) gives
ε
Pε
= C ε1 ---- P
K
C ε1 calibrated from homogeneous ( D = D ε = 0 ) shear flow
dK
------- = P – ε
dt
(32)
Dε ε ε2
------- = C ε1 ---- P – C ε2 -----
Dt K K
For long times K ⁄ ε approaches a constant, thus, from (32) it can be
derived that

d K 1 dK K Dε P
----- ⎛ ----⎞ = --- ------- – ----- ------- = – ( C ε1 – 1 ) ---- + ( C ε2 – 1 ) = 0
dt ⎝ ε ⎠ ε dt ε 2 Dt ε
and from experiment (Tavoularis & Corrsin, 1981) P ⁄ ε ≈ 1.8 . With
C ε2 = 1.92 one can derive C ε1 ≈ 1.5 . “Standard” value C ε1 = 1.44 .

14
• Diffusion of K and ε , D and D ε
The simplest assumption is to assume gradient diffusion

∂ ⎛ ν T ⎞ ∂K ∂ ⎛ ν T⎞ ∂ε
D = ⎜ ν + -------⎟ , Dε = ⎜ ν + ------⎟
∂ xk ⎝ σ K⎠ ∂ x k ∂ xk ⎝ σ ε⎠ ∂ x k
The viscous diffusion is present also in the exact equations. The turbulent
diffusion has to be modelled, and the gradient diffusion is not a very good
approximation, but has been found to be a reasonable approximaion in
thin shear flows. This will be discussed lated.
The Schmidt number in the K equation should be close to unity, σ K = 1.0 .
The Schmidt number in the ε equation, σ ε , is calibrated from the log-layer
in boundary layers. The log-law

1 yu uτ
----- = --- ln -------τ- + C => ------- = ------
U dU
uτ κ ν dy κy
The RANS equation (23) in x-direction assuming ∂ ⁄ ∂t = ∂ ⁄ ∂x = V = 0
and constant P is

∂ ∂U
0 = ----- ( ν + ν T ) -------
∂y ∂y
integrating once give
∂U
( ν + ν T ) ------- = u τ2
∂y
and by applying the log law and neglecting ν

K2 uτ
C µ ------ ------ = u τ2 (33)
ε κy
P ≈ ε is assumed in the log law, which gives
⎛ ∂U⎞ 2 K2 ⎛ uτ ⎞ 2
ε = P = ν T ------- = C µ ------ ------ (34)
⎝ ∂y ⎠ ε ⎝ κy⎠
Combining (33) and (34) gives the log-law relations

u τ2 u τ3
K = ----------- and ε = ------ (35)
Cµ κy
D = 0 since K is constant, and the equation for K is fulfilled by P = ε.

15
The equation for ε reads

ε ε 2 d ⎛ ν T dε⎞
0 = C ε1 ---- P – C ε2 ----- + ------ ⎜ ------ ------⎟
K K dy ⎝ σ ε dy⎠
and by pluggin in P = ε and using (35) gives

κ2
σ ε = -----------------------------------------
( C ε2 – C ε1 ) C µ
Using the “standard” values and κ = 0.41 gives σ ε = 1.2 . The standard
value is σ ε = 1.3 .

Summary of the standard eddy-viscosity K – ε model

DK ∂ ⎛ ν T ⎞ ∂K
-------- = P–ε+ ⎜ ν + -------⎟
Dt ∂ xk ⎝ σ K⎠ ∂ x k

Dε ε ∂ ⎛ ν T⎞ ∂ε (36)
------- = ( C ε1 P – C ε2 ε ) ---- + ⎜ ν + ------⎟
Dt K ∂ xk ⎝ σ ε⎠ ∂ x k

K2
P = 2ν T S ij S ji ν T = C µ ------
ε
Model coefficients (standard values):
C µ = 0.09 , C ε1 = 1.44 , C ε2 = 1.92 , σ K = 1.0 , σ ε = 1.3

The eddy-viscosity K – ω model

Kolmogorov (1942), Wilcox (80:s and 90:s)
ω is intepreted as the inverse time scale of the large eddies, and the K – ω
model reads

DK ∂ ⎛ ν T ⎞ ∂K
-------- = 2ν T S ij S ji – C µ Kω + ⎜ ν + -------⎟
Dt ∂ xk ⎝ σ K⎠ ∂ x k

Dω ∂ ⎛ ν T ⎞ ∂ω (37)
-------- = 2αS ij S ji – βω +
2 ⎜ ν + -------⎟
Dt ∂ xk ⎝ σ ω⎠ ∂ x k

K
ν T = ----
ω

16
The model coefficients proposed by Wilcox (1988) are
C µ = 0.09 , α = 5 ⁄ 9 ≈ 0.56 , β = 0.075 , σ K = 2.0 , σ ω = 2.0
The equation for ω may also be derived from the ε equation using the
variable transformation ω = ε ⁄ C µ K
D ε 1 1 Dε ε DK
-------- = ------ ⎛ -----------⎞ = ------- ⎛ ---- ------- – ------ --------⎞ = …
Dω
(38)
Dt Dt ⎝ C µ K⎠ C µ ⎝ K Dt K 2 Dt ⎠
This transformation relates the α and β coefficients to C ε1 and C ε2 as
β
C ε1 = 1 + α = 1.56 and C ε2 = 1 + ------- = 1.83
Cµ
The turbulent diffusion term from the transformed K – ε model will, how-
ever, contain additional terms

∂ ν T ∂ω ω ∂ 1 ∂K 2ν T ∂K ∂ω
ν T ⎛ ------ – -------⎞
1
Dε = ------ + ----
⎝ σ σ ⎠ ∂x
+ ---------- (39)
∂ xk σε ∂ xk K ∂ xk ε K k Kσ ε ∂ x k ∂ x k
which is the major difference between the K – ε and K – ω models. Also
the coefficients are differently calibrated. Notably, is σ K = 2.0 compared
with σ K = 1.0 in the K – ε model.
One serious problem with the K – ω model is that the model is (unphysi-
cally) sensitive for the free stream conditions on K and ω . This can be
avoided by carefully calibration of the Schmidt numbers and by introducing
a “cross diffusion term” (the last term in eq. 39). Such modifications have
been proposed by Kok (1999) and Menter (1993).

Boundary conditions
The K – ε model cannot be applied all the way down to the wall. Since
K → 0 and ε is finite at the wall, the ε equation becomes singular at the
wall because of the ε ⁄ K term. There are two different solutions of this
• Log-law boundary conditions: The boundary condition is set in the log
layer away from the wall according to the log-law relations. The prob-
lem is that the log-law is only strictly valid in equilibrium boundary layers
and breaks down close to separation and reattachment points. How-
ever, the approach still works reasonable well also in separated flows.
• Near-wall corrections (or low-Reynolds number corrections): The K – ε
model equations are modified with “wall damping functions” based on
y + ≡ yu τ ⁄ ν , y * ≡ y K ⁄ ν or Re ≡ K 2 ⁄ νε active in the inner part up to
T

17
 y + ∼ 50 . The requirement on the resolution of the computational grid
close to the wall is high. ∆y + ≈ 1 for the first grid points at the wall. A lot
of variant are availabe in the litterature.
The K – ω model can be computed down to the wall and gives reasonable
mean velocity profiles without any modifications. The turbulence proper-
ties in the viscous sub-layer is, however, not correctly captured. The K – ω
model can also be used together with wall function boundary conditions.
Realizability considerations for two-equation eddy-viscosity models
A physical requirement on the Reynolds stresses is that the normal com-
ponents ( u 2 , v 2 and w 2 ) must be positive in all possible coordinate sys-
tems. This means that the Reynolds stress tensor must be positive definite
(all eigenvalues must be positive). Translated to the Reynolds stress
anisotropy, this means that
2 4
– --- ≤ a αα ≤ ---
3 3
(no summation over Greek indices) and
a αβ ≤ 1 ( α ≠ β )
The linear eddy-viscosity relation gives
K
a ij = – 2C µ ---- S ij
ε
and it is easily understood that the anisotropy can reach large values that
are physically unrealizable if S ij is large. In e.g. shear flows the a 12 com-
ponent becomes
1 K dU
a 12 = – 2C µ σ , σ = --- ---- -------
2 ε dy
which will give unrealizable a 12 for σ > 1 ⁄ ( 2C µ ) ≈ 5.6 .
The production term in the eddy-viscosity assumption is P = 2ν T S ij S ji ,
which is proportional to the strain rate squared, while the exact production
is P = – Ka ij S ji , linear in S ij . The eddy-viscosity model, thus, gives exces-
sive production in flows with strong strain rate.
In the Menter K – ω SST model (Menter 1993), the C µ coefficient is limited
for high shear rates (or for P ⁄ ε > 1 ) in order to limit a 12 to 0.3 according to
the Bradshaw assumption, preserving realizable values for a 12 . Moreover,
it has been observed that in adverse pressure gradient where P ⁄ ε > 1 the
limitation of the a 12 will be active, which will significantly improve the pre-
diction of adverse pressure gradient flows including flow separation.

18
Differential Reynolds stress models (DRSM)
The exact equation for the Reynolds stress tensor is derived in (9), and
reads

Du i u j
-------------- = Pij – εij + Π ij + D ij
Dt (40)
Dε ε
------- = ( C ε1 P – C ε2 ε ) ---- + D ε
Dt K
where the production term is exact and needs no further modelling
∂U j ∂U i
= – u i u k --------- – u j u k --------- = ε ⎛ --- S ij + a ik S kj + S ik a kj – a ik Ω kj + Ω ik a kj⎞
4
P ij ⎝3 ⎠
∂x k ∂x k
The production term is dependent on the strain- and rotation rate tensors
( S ij and Ω ij ). The dependency on Ω ij clearly improves the prediction of
the influence of rotation. Moreover, the production is linearly dependent on
the strain rate, and, thus the excessive production seen in eddy-viscosity
models are here avoided.

Modelling of the different terms

- Rotta (1951)
- LRR: Launder, Reece and Rodi (1975)
- SSG: Sarkar, Speziale and Gatski (1991)
• The dissipation rate tensor, ε ij
Introduce the dissipation rate anisotropy
ε ij 2
e ij ≡ ------ – --- δ ij (41)
ε 3
The dissipation rate ε ≡ ε ii ⁄ 2 and is derived by a transport equation for ε ,
very similar to the ε equation in the K – ε model. Alternativelly, an ω equa-
tion, similar to that in the K – ω model, can be used.
The simplest model is that the dissipation rate is isotropic, or e ij = 0 .
• The pressure strain rate term, Π ij
The exact form of Π ij is given by

1 ⎛ ∂u i ∂u j⎞
Π ij = --- p ⎜ ------- + -------⎟
ρ ⎝ ∂x j ∂x i⎠
A poisson equation for the pressure fluctuation is derived by taking the
divergence of the equation for the velocity fluctuations. This equation may

19
be formally solved giving the principal form of the model for Π ij . This gives
(s)
that the pressure strain term is divided into a slow and a rapid part, Π ij
(r)
and Π ij , and modelled separatley. In the LRR model, the slow part as
proposed by Rotta (1951) was modelled as
(s)
Π ij = – c 1 εa ij
where the Rotta constant c 1 = 1.5 – 1.8 . The rapid part proposed by LRR

(r) 9c 2 + 6
Π ij = --- εS ij + ------------------ ε ⎛ a ik S kj + S ik a kj – --- a kl S lk δ ij⎞
4 2
5 11 ⎝ 3 ⎠
– 7c 2 + 10
+ -------------------------ε ( a ik Ω kj + Ω ik a kj )
11
with the constant c 2 = 0.4 – 0.6 .
In later models, the connection to the exact Poisson equation is weaker,
and the pressure strain term is lumped together with the dissipation rate
anisotropy as
Π ij
-------- – e ij = f ij ⎛ a ij, ---- S ij, ---- Ω ij⎞
K K
ε ⎝ ε ε ⎠

• Turbulent diffusion terms, D ij and D ε :

In thin shear layers the vertical fluctuations v 2 are most responsible for the
turbulent mixing. A model that better responds to this, compared to the
gradient diffusion, is the generalized gradient diffusion model (GGD) pro-
posed by Daly & Harlow (1970)

∂ ⎛ ∂u k u l
νδ kl – c s ---- u k u l⎞ -------------
K
Dij =
∂ xk ⎝ ε ⎠ ∂x
l

∂ ⎛ ∂ε
νδ kl – c ε ---- u k u l⎞ -------
K
Dε =
∂ xk ⎝ ε ⎠ ∂x
l
In thin shear layers where ∂ ⁄ ∂y is dominating these will reduce to

∂ ⎛ K 2⎞ ∂u k u l
Dij = ν – c s ---- v -------------
∂y ⎝ ε ⎠ ∂y

∂ ⎛ ∂ε
ν – c ε ---- v 2⎞ -----
K
Dε =
∂y ⎝ ε ⎠ ∂y
Standar values are c s = 0.25 and c ε = 0.15 .

20
Explicit algebraic Reynolds stress models (EARSM)
The starting point is the DRSM

Du i u j
-------------- – D ij = Pij – εij + Π ij
Dt
Dε ε
------- – D ε = ( C ε1 P – C ε2 ε ) ----
Dt K
The equation for u i u j is rewritten in terms of the anisotropy and K

Da ij D ⎛ ui uj 2 ⎞ 1 Du i u j u i u j DK
----------- = ------ ⎜ --------- – --- δ ij⎟ = ---- -------------- – --------- -------- = …
Dt Dt ⎝ K 3 ⎠ K Dt K 2 Dt

DK 1 Du i u i 1
-------- = --- -------------- = --- ( P ii – ε ii + D ii )
Dt 2 Dt 2
which gives

K ⎛ Da ij (a) P Pij – εij + Π ij

---- ----------- – D ij ⎞ = – ⎛ a ij + --- δ ij⎞ ⎛ ---- – 1⎞ + ---------------------------------
2
(42)
ε ⎝ Dt ⎠ ⎝ 3 ⎠⎝ ε ⎠ ε
The model for Π ij that can be considered in EARSMs must be “quasi”-lin-
ear in a ij , that is tensorially linear in a ij (e.g. the LRR model)
Π ij 1P C3
-------- – e ij = – --- ⎛ C 1 + C 1 ----⎞ a ij + C 2 S ij + ------ ⎛ a ik S kj + S ik a kj – --- a kl S kl δ ij⎞ (43)
1 0 2
ε 2⎝ ε⎠ 2⎝ 3 ⎠
C4
– ------ ( a ik Ω kj + Ω ik a kj )
2
In this equation, and in the following concerning EARSM, the S ij and Ω ij
are normalized by the turbulent time scale K ⁄ ε

1 K ⎛ ∂U i ∂U j⎞ 1 K ⎛ ∂U i ∂U j⎞
S ij ≡ --- ---- ⎜ + ⎟ , Ω ij ≡ --- ---- ⎜ – ⎟
2 ε ⎝ ∂ xj ∂ xi ⎠ 2 ε ⎝ ∂ xj ∂ xi ⎠
The resulting equation for a ij

1 K Da ij (a) P
------ ---- ⎛ ----------- – D ij ⎞ = ⎛ A 3 + A 4 ----⎞ a ij + A 1 S ij – ( a ik Ω kj + Ω ik a kj ) (44)
A 0 ε ⎝ Dt ⎠ ⎝ ε⎠

+ A 2 ⎛ a ik S kj + S ik a kj – --- a kl S lk δ ij⎞
2
⎝ 3 ⎠
and the A coefficients are directly related to the C coefficients.

21
Weak equilibrium condition
So far, we have only rewritten the equation for u i u j as an equation for a ij
and an equation for K . The major assumption in obtaining an algebraic
relation for a ij is the so called weak equilibrium assumption. That is that
the anisotropy is assumed to be constant in time and space. Thus, the
l.h.s. in (44) vanishes, resulting in

Na = – A 1 S + ( aΩ – Ωa ) – A 2 ⎛ aS + Sa – --- { aS }I⎞
2
(45)
⎝ 3 ⎠
where

P
N = A 3 + A 4 ---- (46)
ε
Here, the bold matrix notation is used where e.g. aΩ ≡ a ik Ω kj , { aS }
denotes the trace of aS ( { aS } ≡ a kl S lk ) and I ≡ δ ij is the identity matrix.
The weak equilibrium assumption is:
• exact in steady homogeneous flows (homogeneous shear flows)
• good approximation in fully developed shear flows
• good approximation even in flows with (slowly) varying a ij with high S ij
and Ω ij since (44) is then dominating by the source terms on the r.h.s.
• bad approximation in flows with low S ij and Ω ij since (44) is then dom-
inating by the transport terms on the l.h.s.

Formal solution of (44) => EARSM

The algebraic relation (44) (ARSM) is implicit in a ij and also non-linear in
a ij since the ratio P ⁄ ε ≡ { aS } .
The first step is to solve the tensor equation by considering the P⁄ε ratio
(or N ) as unknown. The non-linearity will be consider later.
The tensor relation is, thus, that a = a ( S, Ω ) . The most general expres-
sion consists of ten tensorially independent terms
10
(i)
a = ∑ βi T (47)
i=1

22
where
(1) (2) 1
T = S T = S 2 – --- II S I
3
(3) 1 (4)
T = Ω 2 – --- II Ω I T = SΩ – ΩS
3
(5) (6) 2
T = S 2 Ω – ΩS 2 T = SΩ 2 + Ω 2 S – --- IVI (48)
3
(7) 2 (8)
T = S 2 Ω 2 + Ω 2 S 2 – --- VI T = SΩS 2 – S 2 ΩS
3
(9) ( 10 )
T = ΩSΩ 2 – Ω 2 SΩ T = ΩS 2 Ω 2 – Ω 2 S 2 Ω
The β i coefficients may be functions of the five independent invariants

II S = { S 2 } , II Ω = { Ω 2 } , III = { S 3 } , IV = { SΩ 2 } , V = { S 2 Ω 2 } (49)
and are determined by inserting the ansatz (47) in the ARSM equation (45)
and solving the resulting equation system for the β i coefficients.

The Wallin & Johansson model in 2D flows

In the model proposed by Wallin & Johansson (2000), the A 2 coefficient is
choosen to be 0 in (45), which corresponds to the specific choice of
C 2 = 2 in the pressure strain rate model (43). With A 2 = 0 only five of the
ten terms in (48) will remain, and the algebra is, thus, significantly
reduced.
The algebra is rather complex for 3D mean flows, and, thus, the method
will here be described in 2D mean flows, where the number of indepen-
dent tensor groups in general are three, but with A 2 = 0 they reduces to
(1) (4)
two: T and T . Also only two independent invariants remain, II S and
II Ω . The Wallin & Johansson model is also build on the LRR model,
slightly recalibrated, with c 1 = 1.8 and c 2 = 5 ⁄ 9 ≈ 0.56 , which gives the
following model coefficients
4 6 9 9
A 0 = – --- , A 1 = --- , A 2 = 0 , A 3 = --- ( c 1 – 1 ) , A 4 = ---
9 5 4 4
The implicit algebraic relation (45) becomes
6
Na = – --- S + ( aΩ – Ωa ) (50)
5
and with the ansatz (47)
a = β 1 S + β 4 ( SΩ – ΩS ) (51)

23
it becomes

N ( β 1 S + β 4 ( SΩ – ΩS ) ) = (52)
6
– --- S + β 1 ( SΩ – ΩS ) + β 4 ( SΩ 2 – 2ΩSΩ + Ω 2 S )
5
In 2D mean flows, the term

SΩ 2 – 2ΩSΩ + Ω 2 S = 2II Ω S
and (52) becomes
6
N ( β 1 S + β 4 ( SΩ – ΩS ) ) = – --- S + β 1 ( SΩ – ΩS ) + 2β 4 II Ω S (53)
5
sorting equal tensor terms gives
6
Nβ 1 = – --- + 2β 4 II Ω 6 N
5 => β 1 = Nβ 4 = – --- ------------------------ (54)
5 N 2 – 2II
Nβ 4 = β 1 Ω
and the solution for a ij is given with (54) in (52) as
6 ( NS + SΩ – ΩS )
a = – --- ----------------------------------------- (55)
5 N 2 – 2II Ω

The consistency condition

The production to dissipation ratio, P ⁄ ε , can be written as
P Ka ij S' ji
---- = ------------------ = a ij S ji = { aS } (56)
ε ε
Using the solution (52) for the anisotropy this gives that

P
---- = β 1 { S 2 } + β 4 ( { SΩS } – { ΩS 2 } ) = β 1 II S (57)
ε
since { SΩS } = { ΩS 2 } = 0 . Using the solution for β 1 in (54) and the rela-
tion for N in (46)

4 6 NII S
--- ( N – A 3 ) = – --- ------------------------ (58)
9 5 N 2 – 2II
Ω
and rearranged

N 3 – A 3 N 2 – ⎛ ------ II S + 2II Ω⎞ N + 2A 3 II Ω = 0
27
(59)
⎝ 10 ⎠

24
Closed solutions can be obtained from cubic equations

⎧ A3
⎪ ------ + ( P 1 + P 2 ) 1 / 3 + ( P 1 – P 2 ) 1 / 3 P2 ≥ 0
⎪ 3
N = ⎨
⎪ A3 ⎛1 ⎛ P1 ⎞ ⎞
⎪ ------ + 2 ( P 1 – P 2 ) cos ⎜ --- acos ⎜ -----------------------⎟ ⎟ P 2 < 0
2 1 / 6
3 ⎝3 ⎝ P 2 – P ⎠⎠ (60)
⎩ 1 2
3
⎛ A 32 9 2 ⎞ ⎛ A 32 9 2 ⎞
P 1 = ⎜ ------ + ------ II S – --- II Ω⎟ A 3 P 2 = P 1 – ⎜ ------ + ------ II S + --- II Ω⎟
2
⎝ 27 20 3 ⎠ ⎝ 9 10 3 ⎠

Effective C µ
The first term in the solution (51) corresponds to the eddy-viscosity
assumption where

eff β1 3 N
C µ = – ------ = --- ------------------------ (61)
2 5 N 2 – 2II
Ω
eff
here, the C µ is not a constant, thus the C µ , and contain rotational effects
as well as the possibility to adjust to the local flow conditions.

The boundary layer log region

Since P = ε , N can be determined directly as

9⎛ P P ⎞
9c 1
N = A 3 + A 4 = --- c 1 – 1 + ---- = -------- = 4.05
---- (62)
ε 4⎝ ε⎠ 4
In parallell flows

1 K dU 2
II S = – II Ω = --- ⎛ ---- -------⎞ (63)
2 ⎝ ε dy ⎠
and the P ⁄ ε ratio is derived to

P 6 NII S N2 K dU
---- = – β 1 II S = --- ----------------------- = 1 => II S = ---------------- ≈ 5.74 => ---- ------- = 3.39
ε 5 N 2 + 2II 6N ε dy
S ------- – 2
5
also
6 N 6 1 eff
β 1 = – --- ------------------------ ≈ – 0.174 , β 4 = – --- ------------------------ ≈ – 0.043 , C µ = 0.087 .
5 N 2 – 2II 5 N 2 – 2II
Ω Ω

25
The Reynolds stress anisotropy can now be derived

a 12 a 11 a 22 a 33 σ
1 K dU
, where σ ≡ --- ---- -------
DNS – 0.29 0.34 – 0.26 – 0.08 1.65 2 ε dy
WJ – 0.30 0.25 – 0.25 0 1.69
The a 12 and a 22 anisotropies and σ are well predicted, but the a 11 and
a 33 components are not that well. The reason is that c 2 = 5 ⁄ 9 which
gives that a 33 = 0 in thin shear flows. The approximation is not that prob-
lematic, since the a 12 and a 22 components are the only components that
enters into the balance equations.

High shear rates in parallell flow (homogeneous shear)

For high shear rates, σ → ∞ and the following holds

P1 → σ 2 , P2 → –σ 6 , N → σ , β 1 → σ –1 , β 4 → σ –2
The anisotropy goes like

a = β 1 S + β 4 ( SΩ – ΩS ) → σ – 1 σ + σ – 2 σ 2 = const
and the anisotropy will stay realizable (c.f. eddy viscosity models)
The production

P
= β 1 II S → σ – 1 σ 2 = σ
and has the correct development

Rotational flows (rotating channel flow)

The rotational effects enter mainly into β 1 through the dependency on II Ω
6 N
β 1 = – --- ------------------------
5 N 2 – 2II
Ω
In rotating channel flows the system rotation is added to the Ω ij tensor,
used for II Ω . On the unstable side, the system rotation will decrease II Ω
in magnitude (remember II Ω < 0 ) and β 1 will increase. The same holds in
curved flows, where a concave (unstable) curvature will decrease II Ω in
magnitude and β 1 will increase.

26
 KTH, Febr. 25, 2000

Production to dissipation ratio

3.0

eddy viscosity
2.5
Parallel
flow
2.0
hom. shear
P/ε

1.5
log-layer
1.0

0.5

Increased rotation
0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0
σ

Stefan Wallin FLYGTEKNISKA

Stefan.Wallin@ffa.se 10 of 27 .. ..
FORSOKSANSTALTEN
P.O. Box 11021, SE-161 11 Bromma, Sweden
Phone +46 8 5554 9318, Fax +46 8 5554 9071 THE AERONAUTICAL RESEARCH INSTITUTE OF SWEDEN
 KTH, Febr. 25, 2000

Rapidly sheared homogeneous flow

25

Eddy-viscosity model
20

15
K

W&J consistent EARSM

10

RDT
5
EARSM, fixed P ⁄ ε

0
0 2 4 6 8 10
T

Stefan Wallin FLYGTEKNISKA

Stefan.Wallin@ffa.se 12 of 27 .. ..
FORSOKSANSTALTEN
P.O. Box 11021, SE-161 11 Bromma, Sweden
Phone +46 8 5554 9318, Fax +46 8 5554 9071 THE AERONAUTICAL RESEARCH INSTITUTE OF SWEDEN
 KTH, Febr. 25, 2000

(i) 3D mean flows: Rotating pipe, swirl velocity

1.00

Eddy-viscosity k – ω
0.80
and non-cubic EARSM
Uθ / U wall

0.60

0.40

0.20

W&J 3D EARSM
0.00
0.0 0.2 0.4 0.6 0.8 1.0
r/R

Stefan Wallin FLYGTEKNISKA

Stefan.Wallin@ffa.se 14 of 27 .. ..
FORSOKSANSTALTEN
P.O. Box 11021, SE-161 11 Bromma, Sweden
Phone +46 8 5554 9318, Fax +46 8 5554 9071 THE AERONAUTICAL RESEARCH INSTITUTE OF SWEDEN
 KTH, Febr. 25, 2000

(ii) normal stresses

Homogeneous shear:

Model a 12 a 11 a 22 a 33 P⁄ε
Tavoularis & Corrsin 1981 -0.30 0.40 -0.28 -0.12 1.8
Recalibrated LRR (J&W) -0.30 0.31 -0.31 0 1.8
Linearized SSG -0.32 0.41 -0.30 -0.11 1.9
Boundary layers (log layer):

a 12 a 11 a 22 a 33 SK
Model -------
2ε
Moser, Kim & Mansour 1998 -0.29 0.34 -0.26 -0.08 1.65
Recalibrated LRR (J&W) -0.30 0.25 -0.25 0 1.69
Linearized SSG -0.32 0.36 -0.26 -0.10 1.59

Stefan Wallin FLYGTEKNISKA

Stefan.Wallin@ffa.se 15 of 27 .. ..
FORSOKSANSTALTEN
P.O. Box 11021, SE-161 11 Bromma, Sweden
Phone +46 8 5554 9318, Fax +46 8 5554 9071 THE AERONAUTICAL RESEARCH INSTITUTE OF SWEDEN
 1.80 0.25
deriving the measure of the co-ordinate system rotation rate
(r)
Ωij from the flow field. Only a few Galilean invariant meth-
ods have been proposed. Girimaji (1997) proposed to use the
1.20 -0.25
rotation rate of the acceleration vector and Gatski & Jon- U/Um uv+

gen (2000) used the rotation rate of the principal direction

of the strain-rate tensor. However, these methods were only
0.60 -0.75
derived for 2D flows.
Hellsten (2002) has found that the acceleration based
method leads to problems in some situations of mild curva-
0.00 -1.25
ture where the direction of the acceleration vector may vary -1.00 -0.33
y
0.33 1.00 -1.00 -0.33
y
0.33 1.00

rapidly. This was demonstrated in a U-bend flow, which

showed an almost singular behaviour. For the same case, Figure 1: Computed rotating channel flow for Ro = 0.77
the strain-rate based method behaves much better. compared to DNS of Alvelius & Johansson (2000). Curva-
Recently, Wallin & Johansson (2002) proposed a fully ture corrected original ( ) and recalibrated ( ) WJ
three-dimensional method based on the strain-rate tensor, EARSM compared to the non-corrected EARSM ( ).
which will be examined and tested in this study. In two- (r)
U 0 = 2ωz ( ) is also shown. (figure taken from Wallin
dimensional mean flows, the method by Wallin & Johansson & Johansson, 2002)
reduces to the Gatski & Jongen (2000) correction. In fully
developed swirling flows, which is fully three dimensional, lumped together
the proposed method has been shown to be identical to the
Πij
 
exact transformation. − eij = − 21 C10 + C11 P aij + C2 Sij
ε ε
Turbulent flows over curved surfaces, near stagnation (4)
+ C23 aik Skj + Sik akj − 23 akl Slk δij


and separation points, in vortices and turbulent flows in
− C24 aik Ωkj − Ωik akj .


rotating frames of reference are all affected by streamline
curvature effects. Strong curvature and/or rotational ef-
The A coefficients are related to the C coefficients through
fects form a major cornerstone problem also at the Reynolds
stress transport modelling level. Some of the effects of C4 3C2 −4
A0 = − 1, A1 = , A2 = C2A
3 −2
,
streamline curvature or local rotation is captured already 2 3A0 0
(5)
2−C10 1
−C1 −2
in standard EARSMs. E.g. the Coriolis term that appears A3 = 2A0
, A4 = 2A .
0
when transforming the Reynolds stress transport model
equation to a rotating frame is composed by two equal parts Introducing the curvature correction for the original
that originate from the transformation of the production and choice of the A0 coefficient in the Wallin & Johansson (2000)
the advection respectively. The production part is naturally model leads to a model that predicts rotational effects poorly
captured also in standard EARSMs while the advection part as shown by Wallin & Johansson (2002) and also observed
is not. by Wallin & Girimaji (2000) for the vortex flow. Wallin &
In cases with moderately curved streamlines the neglec- Girimaji found that the WJ model behaviour was improved
tion of the advection part has a rather minor effect, see e.g. by increasing A0 to a value closer to that of the linearized
Rumsey, Gatski & Morrison (1999) for flow over an airfoil, SSG.
but in other cases the inclusion of the curvature correction A more thorough analysis of the effect of the A0 coeffi-
is significant. That is the case in e.g. vortices (Wallin & cient was done by Wallin & Johansson (2002), where the long
Girimaji 2000), in rotating flows like rotating homogeneous time asymptotic behaviour in rotating homogeneous shear
shear and rotating channel (Wallin & Johansson 2002) and flow was considered. Depending on the rotation number the
in curved channels as well as in swirling flows, as will be seen turbulent kinetic energy grows exponentially with constant
in this paper. P/ε or follows a power-law solution where ε/(Uy K) → 0
(Speziale & Mac Giolla Mhuiris 1989). The bifurcation
CURVATURE-CORRECTED MODEL points between the two solution branches correspond to the
(eff)
points where Cµ becomes zero or where the flow is close
The curvature corrected EARSM (CC-EARSM) pro-
to neutral stability. Neutral stability occurs near rotation
posed by Wallin & Johansson (2002) will be repeated herein
number Ro = 0.5 and is also likely associated with the lin-
for clarity. It is based on a formal approximation of a
ear velocity profile in the core of a rotating channel (local
Reynolds stress transport model including an approximation
Ro≈0.5) according to Pettersson-Reif et al. (1999). Thus,
of the advection of the anisotropy.
the model coefficients are calibrated such that the required
General quasi-linear Reynolds stress transport models
bifurcation point Ro = 0.5 is obtained. The A coefficients
may be written in terms of a transport equation for the
proposed by Wallin & Johansson (2002) are A0 = −0.72,
anisotropy tensor
A1 = 1.20, A2 = 0, A3 = 1.80 and A4 = 2.25.


D aij (a)
  The recalibrated model is tested in rotating channel flow
τ Dt
− D ij = A 0 A3 + A 4 P
ε
aij

in figure 1. Both the curvature correction and the recalibra-
+A1 Sij − aik Ωkj − Ωik akj  (3) tion are of significant importance in predicting the slope of
2
+A2 aik Skj + Sik akj − 3 akl Slk δij

the mean flow velocity profile in the centre of the channel.
Also the relaminarization at the stabilized side is captured.
(a)
see Wallin & Johansson (2000). Dij is the diffusion of aij
and τ = K/ε is the turbulent time-scale. The strain and Strain rate based curvature correction
rotation rate tensors, Sij and Ωij , are normalized by τ . The advection of the strain rate tensor S may, similarly
This relation results from the general quasi-linear model for to the advection of the anisotropy tensor (1), be expressed as
the pressure-strain rate and dissipation rate anisotropy, eij , a differential plus an algebraic term arising from the trans-
Figure 2: The skin-friction coefficient along the convex wall, the velocity profile and turbulent shear stress at s = 71 in.
Experiment by So & Mellor (1973). The stresses are transformed into the local wall-tangential and -normal coordinate system,
and Upw is the theoretical potential velocity on the wall.

formation to the curvilinear co-ordinate system, esi where S is the control volume surface and n̂i is the unit
normal vector of the control surface. In this formulation,
s
D Spq
D Sij t

(r) (r)

the different components of DS/Dt may be computed di-
= Tip Tqj − Sik Ωkj − Ωik Skj (6)
Dt Dt rectly without the need of evaluating all components of the
The assumption made by Wallin & Johansson (2002) was gradient of S.
that, since the anisotropy and the strain rate tensors are re- The velocity gradient components are computed onto
lated, the co-ordinate system for which the differential part each face of a control-volume or cell using local staggered
of the advection of S is minimized is used also for transform- cells. The derivatives of the strain-rate components are then
ing the advection of the anisotropy. This may be obtained computed in the cell centrepoints. This way the numerical
by finding the solution for the Ω(r) tensor from (6) where error can, in principle, be kept small. However, spatially os-
the first term on the r.h.s. is set to zero. However, that cillating distribution of DS/Dt may be obtained, especially
equation system is overdetermined since there are five (two when high-resolution grids are employed. Presently, this
in 2D) independent equations for DS/Dt and three (one in problem is handled by applying a spatial filter for the com-
2D) independent components of Ω(r) . puted rotation vector ω (r) . A top-hat filter of the width of
(r) (r) three computational cells in each direction is employed. This
By using that Ωij ≡ −ijk ωk the equation for the ad-
turned out to be a sufficient remedy in the two-dimensional
vection of the transformed Sij in (6)) becomes
flows considered in this study. In the three-dimensional
s
D Spq swirling flow, some spatial oscillation still occurs in the recir-
t D Sij
(r)
eij ≡ Tip Tqj = − Sil ljk + Sjl lik ωk (7) culation zone, but this neither prevents the iteration from
Dt Dt
converging nor spoils the results. It is felt, however, that
eij may be minimized in a least square sense by minimiz- the numerical computation of DS/Dt still needs some fur-
ing the norm eij eij , which, for this case, is equivalent with ther attention.
Spl elq pqi = 0. That results in the following expression for
(r)
the rotation vector ωi
TEST CASES
(r) D Slq
ωi = Aij Spl pqj (8)
Dt
Two-Dimensional Boundary Layer on a Convex Wall
where A convex curved boundary layer experimentally studied
IIS2 δij + 12IIIS Sij + 6IIS Sik Skj by So & Mellor (1973) will be used for basic validation of
Aij = . (9) the curvature corrected EARS-modelling. The concave outer
2IIS3 − 12IIIS2
wall is contoured to obtain a nearly constant pressure dis-
The denominator in (9) may become zero when two of tribution on the inner wall. The CC-EARSM will be tested
(r)
the eigenvalues are equal or all eigenvalues are zero. The using two different ways to obtain ω3 : the Wallin & Johans-
singularities at these points may be avoided by adding a son (2002) strain-rate based method (8), which reduces to
small number to the denominator. Eq. (10) in two-dimensional mean flows, and the streamline
(S) (r)
In two-dimensional incompressible mean flows, ωi re- method in which ω3 is simply the rotation rate of the veloc-
duces to ity vector following a streamline. The latter method is not
(r) S11 Ṡ12 − S12 Ṡ11 generalizable due to its lack of Galilean invariance, but can
ω3 = 2 + 2S 2
(10)
2S11 12 be used as a reference here because the coordinate system
which is identical to the Spalart & Shur (1997) and the can be attached to the apparatus. The results will be com-
Gatski & Jongen (2000) corrections. pared with the experimental data, with the results obtained
with the standard EARSM derived in the inertial coordi-
nate system (iWJ), and with full differential Reynolds stress
Implementation aspects model (RSM) predictions using the corresponding pressure-
The strain-rate based curvature correction methods in- strain model, see Hellsten et al. (2002) and Salo (2003).
volve numerical approximation of DS/Dt. In steady-state The skin-friction coefficient along the convex wall is
problems (time-dependent problems not considered here) the shown in figure 2. Clearly, both CC-EARSMs as well as
derivative of the S tensor may be expressed in conservative the RSM agree well with the measurements while the stan-
form as dard EARSM slightly overestimates the wall shear-stress as
∂Sij 1 ∂ 1
I expected. The velocity and turbulent shear-stress distribu-
Uk = (ρSij Uk ) = ρSij Uk n̂k dS (11) tions are also shown in the figure. The differences in the
∂xk ρ ∂xk ρV S