This action might not be possible to undo. Are you sure you want to continue?
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 41
Large Eddy Simulations February 23, 2006
Time averaging and ﬁltering
In CFD we time average our equations to get the equations
in steady form. This is called Reynolds time averaging:
Φ =
1
2T
_
T
−T
Φ(t)dt, Φ = Φ + Φ
(67)
(note that we use the notation . for time averaging). In
LES we ﬁlter (volume average) the equations. In 1D we
get:
¯
Φ(x, t) =
1
∆x
_
x+0.5∆x
x−0.5∆x
Φ(ξ, t)dξ
Φ =
¯
Φ + Φ
no filter
one filter
2 2.5 3 3.5 4
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
PSfrag replacements
u
,
¯u
x
Since in LES we do not average in time, the ﬁltered vari
ables are functions of space and time.
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 42
The equations for the ﬁltered variables have the same
form as NavierStokes, i.e.
∂¯ u
i
∂t
+
∂
∂x
j
(¯ u
i
¯ u
j
) = −
1
ρ
∂¯ p
∂x
i
+ ν
∂
2
¯ u
i
∂x
j
∂x
j
−
∂τ
ij
∂x
j
∂ ¯ u
i
∂x
i
= 0
(68)
where the subgrid stresses are given by
τ
ij
= u
i
u
j
− ¯ u
i
¯ u
j
(69)
Contrary to Reynolds time averaging where u
i
= 0, we
have here
u
i
= 0
¯ u
i
= ¯ u
i
Note that for the spectral cutoff ﬁlter ¯ u
i
= ¯ u
i
, see p. 46.
However, in ﬁnite volume methods, box ﬁlters are always
used. In this course we use box ﬁlters, if not otherwise
stated.
Differences between timeaveraging (RANS) and space
ﬁltering (LES)
In RANS, if a variable is time averaged twice (u), it is
the same as time averaging once (u). This is because u
is not dependent on time. From Eq. 67 we get
u =
1
2T
_
T
−T
udt =
1
2T
u2T = u
This is obvious if the ﬂow is steady, i.e. ∂u/∂t = 0. If
the ﬂow is unsteady, we must assume a separation in time
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 43
scales, i.e. T t, where t is the time scale for the mean
ﬂow. In practice this requirement is rarely satisﬁed.
In LES, ¯ u = ¯ u (and since u = ¯ u + u
we get u
= 0).
PSfrag replacements
u , ¯ u
x
I
I + 1
I −1
x
x x
A
B
∆x
Let’s ﬁlter ¯ u
I
once more (ﬁlter size ∆x, see ﬁgure above).
For simplicity we do it in 1D. (Note that subscript I denotes
node number.)
¯ u
I
=
1
∆x
_
∆x/2
−∆x/2
¯ u(ξ)dξ =
1
∆x
_
_
0
−∆x/2
¯ u(ξ)dξ +
_
∆x/2
0
¯ u(ξ)dξ
_
=
=
1
∆x
_
∆x
2
¯ u
A
+
∆x
2
¯ u
B
_
.
The trapezoidal rule, which is secondorder accurate, was
used to estimate the integrals. ¯ u at locations A and B (see
ﬁgure above) is estimated by linear interpolation, which
gives
¯ u
I
=
1
2
__
1
4
¯ u
I−1
+
3
4
¯ u
I
_
+
_
3
4
¯ u
I
+
1
4
¯ u
I+1
__
=
1
8
(¯ u
I−1
+ 6¯ u
I
+ ¯ u
I+1
) = ¯ u
I
(70)
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 44
Resolved & SGS scales
•The basic idea in LES is to resolve (large) grid scales (GS),
and to model (small) subgridscales (SGS).
The limit (cutoff) between GS and SGS is supposed to
take place in the inertial subrange (II), see ﬁgure below. PSfrag replacements
u , ¯ u
x
I
I + 1
I −1
x
x
A
B
∆x
I
I
I
I
I
I
κ
E(κ)
cutoff
I: large, energycontaining scales
II: inertial subrange (Kolmogorov −5/3range)
III: dissipation subrange
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 45
The boxﬁlter and the cutoff ﬁlter
The ﬁltering is formally deﬁned as (1D)
¯ u(x) =
_
∞
−∞
G(r)u(x −r)dr
_
∞
−∞
G(r)dr = 1
(71)
It is often convenient to study the ﬁltering process in the
spectral space using Fourier transforms deﬁned as
ˆ u(κ) =
1
2π
_
∞
0
u(r) exp(−ıκr)dr (72)
and its inverse
u(r) =
_
∞
0
ˆ u(κ) exp(ıκr)dκ (73)
where κ denotes the wavenumber and ı =
√
−1. Using the
convolution theorem (saying that the integrated product of
two functions is equal to the product of their Fourier trans
forms) the ﬁltering in Eq. 71 is conveniently written (as
suming homogeneous turbulence)
ˆ u(κ) =
´
¯ u(κ) =
1
2π
_
∞
0
¯ u(η) exp(−ıκη)dη
=
1
2π
_
∞
0
_
∞
0
exp(−ıκη)G(ρ)u(η −ρ)dρdη
=
1
2π
_
∞
0
_
∞
0
exp(−ıκρ) exp(−ıκ(η −ρ))G(ρ)u(η −ρ)dρdη
=
1
2π
_
∞
0
_
∞
0
exp(−ıκρ) exp(−ıκξ)G(ρ)u(ξ)dξdρ =
ˆ
G(κ)ˆ u(κ)
(74)
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 46
The ﬁlter used in spectral space is the cutoff ﬁlter, which
ﬁlters out all scales with wavenumber larger than the cut
off wavenumber κ
c
= π/∆. This is deﬁned as
ˆ
G(κ) =
_
1 if κ ≤ κ
c
0 otherwise
(75)
If we ﬁlter twice with the cutoff ﬁlter we get
ˆ u =
ˆ
G
ˆ
Gˆ u =
ˆ
Gˆ u = ˆ u (76)
using Eqs. 74 and 75. Thus, contrary to the boxﬁlter (see
Eq. 70), nothing happens when we ﬁlter twice in spectral
space.
We have mentioned the volume ﬁlter (box ﬁlter), which
ﬁlters out ﬂuctuations in space with scales smaller than ∆.
It reads
G(r) =
_
1/∆ if r ≤ ∆/2
0 otherwise
(77)
Note that the box ﬁlter is sharp in physical space but not
in wavenumber space; for the cutoff ﬁlter it is vice versa.
In ﬁnite volume methods box ﬁltering is always used.
Furthermore implicit ﬁltering is employed. This means
that the ﬁlering is the same as the discretization (=inte
gration over the control volume which is equal to the ﬁlter
volume, see Eq. 79).
Subgrid model
We need a subgrid model to model the turbulent scales
which cannot be resolved by the grid and the discretization
scheme.
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 47
The simplest model is the Smagorinsky model (Smagorin
sky, 1963):
τ
ij
−
1
3
δ
ij
τ
kk
= −2ν
sgs
¯ s
ij
ν
sgs
= (C
S
∆)
2
_
2¯ s
ij
¯ s
ij
≡ (C
S
∆)
2
¯ s
¯ s
ij
=
1
2
_
∂¯ u
i
∂x
j
+
∂¯ u
j
∂x
i
_
(78)
and the ﬁlterwidth is taken as the local grid size
∆ = (∆V
IJK
)
1/3
(79)
Because the SGS turbulent ﬂuctuations near a wall go
to zero, so must SGS viscosity. A damping function f
µ
is
added to ensure this
f
µ
= 1 −exp(−y
+
/26) (80)
Disadvantage of Smagorinsky model: the “constant” C
S
is not constant, but it is ﬂowdependent. It is found to vary
in the range from C = 0.065 (Moin & Kim, 1982) to C =
0.25 (Jones & Wille, 1995).
Smagorinsky model vs. mixinglength model
The eddy viscosity according to the mixing length theory
reads in boundarylayer ﬂow(Hinze, 1975; Schlichting, 1979)
ν
t
=
2
¸
¸
¸
¸
∂U
∂y
¸
¸
¸
¸
.
Generalized to three dimensions, we have
ν
t
=
2
__
∂U
i
∂x
j
+
∂U
j
∂x
i
_
∂U
i
∂x
j
_
1/2
=
2
(2S
ij
S
ij
)
1/2
≡
2
S.
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 48
In the Smagorinsky model the SGS turbulent length scale
corresponds to = C
S
∆ so that
ν
sgs
= (C
S
∆)
2
¯ s
which is the same as Eq. 78
Smagorinsky model derived from the k
−5/3
law
Another way to derive Eq. 78 is as follows. The subgrid
model is supposed to model the turbulence with scales
II
,
where
II
is the inverse of the wave number
II
= 1/κ
II
in
the inertial subrange, see the ﬁgure below.
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
u
,
¯u x I
I
+
1
I
−
1 x x
A B
∆
x
I
I
I
I
I
I
κ
E
(
κ
)
c
u
t

o
f
f
I
I
I
I
I
I
κ
E(κ)
Spectrum for k. I: Range for the large, energy containing
eddies. II: the inertial subrange for isotropic scales, inde
pendent of the large scales () and the dissipative scales (ν).
III: Range for small, isotropic, dissipative scales.
In the inertial subrange the turbulence is at large Reynolds
number independent of the large scale as well as of the
small scales. This range is characterized by the dissipa
tion ε and the wavenumber ∆ = 1/κ
II
. Thus we can write
the SGS viscosity as
ν
sgs
= ε
a
(C
S
∆)
b
(81)
Dimensional analysis yields a = 1/3, b = 4/3 so that
ν
sgs
= (C
S
∆)
4/3
ε
1/3
. (82)
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 49
The scales in the inertial subrange are isotropic, and thus
they are in local equilibrium, i.e. in the k
sgs
equation we
have that production balances dissipation
2ν
sgs
¯ s
ij
¯ s
ij
= ε. (83)
Eq. 83 substituted into Eq. 82 gives
ν
3
sgs
= (C
S
∆)
4
ε = (C
S
∆)
4
ν
sgs
(2¯ s
ij
¯ s
ij
)
⇒ ν
sgs
= (C
S
∆)
2
¯ s
¯ s = (2¯ s
ij
¯ s
ij
)
1/2
(84)
Derivation of the Smagorinsky constant
In the inertial subrange production and dissipation are, as
mentioned above, in balance, and using Eq. 84 in Eq. 83
gives
ε = (C
S
∆)
2
¯ s
3
(85)
Let’s compute ¯ s
2
in the Kolmogorov −5/3 range, where the
turbulence is isotropic and homogeneous. We can write
¯ s
2
= 2¯ s
ij
¯ s
ij
=
∂¯ u
i
∂x
j
∂¯ u
i
∂x
j
= −
∂
2
Q
i,i
∂r
j
∂r
j
¸
¸
¸
¸
r=0
(86)
where the ﬁrst line is valid because the turbulence is isotropic,
and the second line is due to that the turbulence is homo
geneous (see derivation in Lecture Notes on RSM
www.tfd.chalmers.se/doct/comp turb model/). The
general twopoint correlation Q
i,j
of u
i
and u
j
can be ex
pressed by the energy spectrum tensor as (Hinze, 1975,
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 50
Chapter 3) (cf. Eq. 73)
Q
i,j
(r) =
_
∞
0
E
i,j
(κ) exp(ıκ
m
r
m
)dκ
1
dκ
2
dκ
3
(87)
Taking the trace and derivating twice gives
∂
2
Q
i,i
∂r
j
∂r
j
= −
_
∞
0
κ
2
E
i,i
(κ) exp(ıκ
m
r
m
)dκ
1
dκ
2
dκ
3
(88)
Integrating over a spherical shell yields
∂
2
Q
i,i
∂r
j
∂r
j
= −4π
_
∞
0
κ
4
E
i,i
(κ)
sin(κr)
κr
dκ (89)
The threedimensional spectrum E(κ) is deﬁned as
E(κ) = 2πκ
2
E
i,i
(κ) (90)
Using Eqs. 86, 89 & 90 and letting r → 0 gives (note that
¯
E(κ) is equal to the square of the Fourier coefﬁcient, i.e.
¯
E(κ) =
ˆ
¯ u
2
(κ))
¯ s
2
= 2
_
∞
0
κ
2
¯
E(κ)dκ
This is the ﬁltered spectrum. Introduce the ﬁltering func
tion
ˆ
G for the cutoff ﬁlter (Eq. 75) so that
¯ s
2
= 2
_
∞
0
κ
2
ˆ
G
2
(κ)E(κ)dκ
where Eq. 74 was used. Using Kolmogorov spectrumE(κ) =
Cκ
−5/3
ε
2/3
and a cutoff ﬁlter (see Eq. 75) we get
¯ s
2
= 2
_
κ
c
0
κ
2
Cκ
−5/3
ε
2/3
dκ =
3
2
π
4/3
Cε
2/3
∆
−4/3
(91)
Inserting this in Eq. 85 gives
ε = (C
S
∆)
2
_
3
2
C
_
3/2
π
2
ε∆
−2
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 51
and ﬁnally we get
C
S
=
1
π
_
2
3C
_
3/4
= 0.17 (92)
with the standard value on the Kolmogorov constant C =
1.5.
SGS kinetic energy
The SGS kinetic energy k
sgs
can be estimated from the Kol
mogorov −5/3 law. The total turbulent kinetic energy is
obtained from the energy spectrum as
k =
_
∞
0
E(κ)dκ
Changing the lower integration limit to wavenumbers larger
than cutoff (i.e. κ
c
) gives the SGS kinetic energy
k
sgs
=
_
∞
κ
c
E(κ)dκ
The Kolmogorov −5/3 law now gives
k
sgs
=
_
∞
κ
c
Cκ
−5/3
ε
2/3
dκ
(Note that for these high wavenumbers, the Kolmogorov
spectrumought to be replaced by the KolmogorovPau spec
trum in which an exponential decaying function is added
for high wavenumbers (Hinze, 1975, Chapter 3)). Carrying
out the integration and replacing κ
c
with π/∆ we get
k
sgs
=
3
2
C
_
∆ε
π
_
2/3
(93)
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 52
LES vs. RANS
LES can handle many ﬂows which RANS (Reynolds Averaged
Navier Stokes) cannot; the reason is that in LES large, tur
bulent scales are resolved. Examples are:
o Flows with large separation
o Bluffbody ﬂows (e.g. ﬂow around a car); the wake of
ten includes large, unsteady, turbulent structures
o Transition
•In RANS all turbulent scales are modelled ⇒inaccurate
•In LES only small, isotropic turbulent scales are mod
elled ⇒ accurate
LES is very much more expensive than RANS.
The dynamic model
In this model of Germano et al. (1991) the constant C is not
arbitrarily chosen (or optimized), but it is computed.
If we apply two ﬁlters to NavierStokes [grid ﬁlter and
a second, coarser ﬁlter (test ﬁlter, denoted by
¸¸
. )] where
¸¸
∆ = 2∆ we get
∂
¸¸
¯ u
i
∂t
+
∂
∂x
j
_
¸¸
¯ u
i
¸¸
¯ u
j
_
= −
1
ρ
∂
¸¸
¯ p
∂x
i
+ ν
∂
2
¸¸
¯ u
i
∂x
j
∂x
j
−
∂T
ij
∂x
j
(94)
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 53
where the subgrid stresses on the test level now are given
by
T
ij
=
¸ ¸
u
i
u
j
−
¸¸
¯ u
i
¸¸
¯ u
j
(95)
If we apply the second ﬁlter to the gridﬁltered equations
(Eq. 68) we obtain
∂
¸¸
¯ u
i
∂t
+
∂
∂x
j
_
¸¸
¯ u
i
¸¸
¯ u
j
_
= −
1
ρ
∂
¸¸
¯ p
∂x
i
+ ν
∂
2
¸¸
¯ u
i
∂x
j
∂x
j
−
∂
¸¸
τ
ij
∂x
j
−
∂
∂x
j
_
¸ ¸
¯ u
i
¯ u
j
−
¸¸
¯ u
i
¸¸
¯ u
j
_
(96)
Identiﬁcation of Eqs. 94 and 96 gives
¸ ¸
¯ u
i
¯ u
j
−
¸¸
¯ u
i
¸¸
¯ u
j
+
¸¸
τ
ij
= T
ij
(97)
The dynamic Leonard stresses are now deﬁned as
L
ij
≡
¸ ¸
¯ u
i
¯ u
j
−
¸¸
¯ u
i
¸¸
¯ u
j
= T
ij
−
¸¸
τ
ij
(98)
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 54
PSfrag replacements
u , ¯ u
x
I
I + 1
I −1
x
x
A
B
∆x
I
II
III
κ
E(κ)
cutoff
I
I
I
I
I
I
κ
E(κ)
cutoff, grid ﬁlter test ﬁlter
κ
c
= π/∆
κ = π/
¸¸
∆
In the energy spectrum, the test ﬁlter is located at lower
wave number than the grid ﬁlter, see the ﬁgure above.
The test ﬁlter
The test ﬁlter is twice the size of the grid ﬁlter, i.e.
¸¸
∆ = 2∆.
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 55
E
W
P
PSfrag replacements
u , ¯ u
x
I
I + 1
I −1
x
x
A
B
∆x
I
II
III
κ
E(κ)
cutoff
I
II
III
κ
E(κ)
cutoff, grid ﬁlter
test ﬁlter
κ
c
= π/∆
κ = π/
¸¸
∆
∆x ∆x
grid ﬁlter
test ﬁlter
The testﬁltered variables are computed by integration
over the test ﬁlter. For example, in the 1D example above
¸¸
¯ u is computed as (
¸ ¸
∆x = 2∆x)
¸¸
¯ u =
1
2∆x
_
E
W
¯ udx =
1
2∆x
__
P
W
¯ udx +
_
E
P
¯ udx
_
=
1
2∆x
(¯ u
w
∆x + ¯ u
e
∆x) =
1
2
_
¯ u
W
+ ¯ u
P
2
+
¯ u
P
+ ¯ u
E
2
_
=
1
4
(¯ u
W
+ 2¯ u
P
+ ¯ u
E
)
(99)
For 3D, ﬁltering at the test level is carried out in the
same way by integrating over the test cell assuming linear
variation of the variables (Zang et al., 1993), i.e. (see ﬁgure
below)
¸¸
¯ u
I,J,K
=
1
8
(¯ u
I−1/2,J−1/2,K−1/2
+ ¯ u
I+1/2,J−1/2,K−1/2
+¯ u
I−1/2,J+1/2,K−1/2
+ ¯ u
I+1/2,J+1/2,K−1/2
+¯ u
I−1/2,J−1/2,K+1/2
+ ¯ u
I+1/2,J−1/2,K+1/2
+¯ u
I−1/2,J+1/2,K+1/2
+ ¯ u
I+1/2,J+1/2,K+1/2
)
(100)
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 56
o
o
o
o
o
o
PSfrag replacements
u , ¯ u
x
I
I + 1
I −1
x
x
A
B
∆x
I
II
III
κ
E(κ)
cutoff
I
II
III
κ
E(κ)
cutoff, grid ﬁlter
test ﬁlter
κ
c
= π/∆
κ = π/
¸¸
∆
∆x
∆x
grid ﬁlter
test ﬁlter
I, J −1, K
I, J + 1, K
I, J, K
I −1, J, K
I + 1, J, K
x
y
z
Stresses on grid, test and intermediate level
The stresses on the grid level, test level and intermediate
level (dynamic Leonard stresses) have the form
τ
ij
= u
i
u
j
− ¯ u
i
¯ u
j
stresses with < ∆
T
ij
=
¸ ¸
u
i
u
j
−
¸¸
¯ u
i
¸¸
¯ u
j
stresses with <
¸¸
∆
L
ij
= T
ij
−
¸¸
τ
ij
stresses with ∆ < <
¸¸
∆
Thus the dynamic Leonard stresses represent the con
tribution to the stresses with scales in the range between
∆ and
¸¸
∆.
Assume now that the same functional form for the sub
grid stresses that is used at the grid level (τ
ij
) also can be
used at the test ﬁlter level (T
ij
). If we use the Smagorinsky
model we get
τ
ij
−
1
3
δ
ij
τ
kk
= −2C∆
2
¯ s¯ s
ij
(101)
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 57
T
ij
−
1
3
δ
ij
T
kk
= −2C
¸¸
∆
2

¸¸
¯ s 
¸¸
¯ s
ij
(102)
where
¸¸
¯ s
ij
=
1
2
_
∂
¸¸
¯ u
i
∂x
j
+
∂
¸¸
¯ u
j
∂x
i
_
, 
¸¸
¯ s  =
_
2
¸¸
¯ s
ij
¸¸
¯ s
ij
_
1/2
Applying the test ﬁlter to Eq. 101 (assuming that C varies
slowly), substituting this equation and Eq. 102 into Eq. 98
gives
L
ij
−
1
3
δ
ij
L
kk
= −2C
_
¸¸
∆
2

¸¸
¯ s 
¸¸
¯ s
ij
−∆
2
¸ ¸
¯ s¯ s
ij
_
(103)
Note that the “constant” C really is a function of both space
and time, i.e. C = C(x
i
, t).
Equation 103 is a tensor equation, and we have ﬁve (¯ s
ij
is symmetric and traceless) equations for C. Lilly (1992)
suggested to satisfy Eq. 103 in a leastsquare sense, deﬁn
ing the error Q as the difference between the lefthand side
and the righthand side of Eq. 103 raised to the power of
two, i.e.
Q =
_
L
ij
−
1
3
δ
ij
L
kk
+ 2CM
ij
_
2
M
ij
=
_
¸¸
∆
2

¸¸
¯ s 
¸¸
¯ s
ij
−∆
2
¸ ¸
¯ s¯ s
ij
_
and requiring ∂Q/∂C = 0, which gives
C = −
L
ij
M
ij
2M
ij
M
ij
(104)
It turns out the the dynamic coefﬁcient C ﬂuctuates wildly
both in space and time. This causes numerical problems,
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 58
and it has been found necessary to average C in homoge
neous direction(s). Furthermore, C must be clipped to en
sure that the total viscosity stays positive (ν + ν
sgs
≥ 0).
In real 3D ﬂows, there is no homogeneous direction. This
makes it very difﬁcult (read: impossible) to use this model,
without introducing some arbitrary averaging and clipping.
Use of oneequation models solve these numerical prob
lems (see p. 66).
Scalesimilarity Models
In the models presented in the previous sections (the Smagorin
sky and the dynamic models) the total SGS stress τ
ij
=
u
i
u
j
− ¯ u
i
¯ u
j
was modelled with an eddyviscosity hypothe
sis. In scalesimilarity models the total stress is split up
as
τ
ij
= u
i
u
j
− ¯ u
i
¯ u
j
= (¯ u
i
+ u
i
)(¯ u
j
+ u
j
) − ¯ u
i
¯ u
j
= ¯ u
i
¯ u
j
+ ¯ u
i
u
j
+ ¯ u
j
u
i
+ u
i
u
j
− ¯ u
i
¯ u
j
= (¯ u
i
¯ u
j
− ¯ u
i
¯ u
j
) +
_
¯ u
i
u
j
+ ¯ u
j
u
i
_
+ u
i
u
j
where the termin brackets is denoted the Leonard stresses,
the term in square brackets is denoted cross terms, and the
last term is denoted the Reynolds SGS stress. Thus
τ
ij
= L
ij
+ C
ij
+ R
ij
L
ij
= ¯ u
i
¯ u
j
− ¯ u
i
¯ u
j
C
ij
= ¯ u
i
u
j
+ ¯ u
j
u
i
R
ij
= u
i
u
j
.
(105)
Note that the Leonard stresses L
ij
are computable, i.e. they
are exact and don’t need to be modelled.
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 59
In scalesimilarity models the main idea is that the tur
bulent scales just below (smaller than) ∆ are similar to the
ones just above ∆ (hence the word ”scalesimilar”). Looking
at Eq. 105 it seems natural to assume that the cross term is
responsible for the interaction between resolved scales (¯ u
i
)
and modelled scales (u
i
), since C
ij
includes both scales.
The Bardina Model
In the Bardina model the Leonard stresses L
ij
are com
puted explicitly, and the sum of the cross term C
ij
and
the Reynolds term is modelled as (Bardina et al., 1980;
Speziale, 1985)
C
M
ij
= c
r
(¯ u
i
¯ u
j
− ¯ u
i
¯ u
j
) (106)
and R
ij
= 0 (superscript M denotes Modelled). It was found
that this model was not sufﬁciently dissipative, and thus a
Smagorinsky model was added
C
M
ij
= c
r
(¯ u
i
¯ u
j
− ¯ u
i
¯ u
j
)
R
M
ij
= −2C
S
∆
2
¯ s¯ s
ij
(107)
Galilean invariance
Speziale (1985) found that the Leonard term L
ij
and the
cross term C
ij
are not Galilean invariant by themselves,
but only the sum L
ij
+ C
ij
is. As a consequence, if the
cross term is neglected, the Leonard stresses must not be
computed explicitly, because then the modelled momentum
equations do not satisfy Galilean invariance.
Below we repeat some of the details of the derivation
given in Speziale (1985). Galilean invariance means that
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 60
the equations do not change if the coordinate system is
moving with a constant speed V
k
. Let’s denote the moving
coordinate system by ∗, i.e.
x
∗
k
= x
k
+ V
k
t, t
∗
= t, ¯ u
∗
k
= ¯ u
k
+ V
k
(108)
By differentiating a variable φ we get
∂φ
∂x
k
=
∂x
∗
j
∂x
k
∂φ
∂x
∗
j
+
∂t
∗
∂x
k
∂φ
∂t
∗
=
∂φ
∂x
∗
k
∂φ
∂t
=
∂x
∗
k
∂t
∂φ
∂x
∗
k
+
∂t
∗
∂t
∂φ
∂t
∗
= V
k
∂φ
∂x
∗
k
+
∂φ
∂t
∗
.
(109)
From Eq. 109 is it easy to show that the NavierStokes
(both with and without ﬁlter) is Galilean invariant (Speziale,
1985; Panton, 1984). Transforming the material derivative
from the (t, x
i
)ccordinate system to the (t
∗
, x
∗
i
)ccordinate
system gives
∂φ
∂t
+ u
k
∂φ
∂x
k
=
∂φ
∂t
∗
+ V
k
∂φ
∂x
∗
k
+ (u
∗
k
−V
k
)
∂φ
∂x
∗
k
=
∂φ
∂t
∗
+ u
∗
k
∂φ
∂x
∗
k
,
as it should.
Now, let’s look at the Leonard term and the cross term.
Since the ﬁltering operation is Galilean invariant (Speziale,
1985), we have ¯ u
∗
k
= ¯ u
k
+V
k
and consequently also u
∗
k
= u
k
.
For the Leonard and the cross term we get (note that since
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 61
V
i
is constant V
i
=
¯
V
i
=
¯
V
i
)
L
∗
ij
= ¯ u
∗
i
¯ u
∗
j
− ¯ u
∗
i
¯ u
∗
j
= (¯ u
i
+ V
i
)(¯ u
j
+ V
j
) −(¯ u
i
+ V
i
)(¯ u
j
+ V
j
)
= ¯ u
i
¯ u
j
+ ¯ u
i
V
j
+ ¯ u
j
V
i
− ¯ u
i
¯ u
j
− ¯ u
i
V
j
−V
i
¯ u
j
= ¯ u
i
¯ u
j
− ¯ u
i
¯ u
j
+ V
j
(¯ u
i
− ¯ u
i
) + V
i
(¯ u
j
− ¯ u
j
)
= L
ij
−V
j
u
i
−V
i
u
j
C
∗
ij
= ¯ u
∗
i
u
∗
j
+ ¯ u
∗
j
u
∗
i
= (¯ u
i
+ V
i
)u
j
+ (¯ u
j
+ V
j
)u
i
=
= ¯ u
i
u
j
+ u
j
V
i
+ ¯ u
j
u
i
+ u
i
V
j
= C
ij
+ u
j
V
i
+ u
i
V
j
(110)
From Eq. 110 we ﬁnd that the Leonard term and the cross
term are different in the two coordinate systems, and thus
the terms are not Galilean invariant. However, note that
the sum is, i.e.
L
∗
ij
+ C
∗
ij
= L
ij
+ C
ij
. (111)
The requirement for the Bardina model to be Galilean
invariant is that the constant must be one, c
r
= 1 (see
Eq. 107). This is shown by transforming both the exact
C
ij
(Eq. 105) and the modelled one, C
M
ij
(i.e. Eq. 106). The
exact form of C
ij
transforms as in Eq. 110. The Bardina
term transforms as
C
∗M
ij
= c
r
(¯ u
∗
i
¯ u
∗
j
− ¯ u
∗
i
¯ u
∗
j
)
= c
r
_
(¯ u
i
+ V
i
)(¯ u
j
+ V
j
) −(¯ u
i
+ V
i
)(¯ u
j
+ V
j
)
_
= c
r
[¯ u
i
¯ u
j
− ¯ u
i
¯ u
j
−(¯ u
i
− ¯ u
i
)V
j
−(¯ u
j
− ¯ u
j
)V
i
]
= C
M
ij
+ c
r
_
u
i
V
j
+ u
j
V
i
¸
.
(112)
As is seen, C
∗M
ij
= C
M
ij
, but here this does not matter, be
cause provided c
r
= 1 the modelled stress C
M
ij
transforms
in the same way as the exact one C
ij
. Thus, as for the exact
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 62
stress C
ij
(see Eq. 111), we have C
∗M
ij
+L
∗
ij
= C
M
ij
+L
ij
. Note
that in order to make the Bardina model Galilean invariant
the Leonard stress must be computed explicitly.
Redeﬁned terms in the Bardina Model
The stresses in the Bardina model can be redeﬁned to make
it Galilean invariant for any value c
r
. A modiﬁed Leonard
stress tensor L
m
ij
is deﬁned as (Germano, 1986)
τ
m
ij
= τ
ij
= C
m
ij
+ L
m
ij
+ R
m
ij
L
m
ij
= c
r
(¯ u
i
¯ u
j
− ¯ u
i
¯ u
j
)
C
m
ij
= 0
R
m
ij
= R
ij
= u
i
u
j
(113)
Note that the modiﬁed Leonard stresses is the same as the
“unmodiﬁed” one plus the modelled cross term C
ij
in the
Bardina model with c
r
= 1 (righthand side of Eq. 106), i.e.
L
m
ij
= L
ij
+ C
M
ij
In order to make the model sufﬁciently dissipative a Smagorin
sky model is added, and the total SGS stress τ
ij
is modelled
as
τ
ij
= ¯ u
i
¯ u
j
− ¯ u
i
¯ u
j
−2C∆¯ s¯ s
ij
(114)
Belowwe verify that the modiﬁed Leonard stress is Galilean
invariant.
1
c
r
L
m∗
ij
= ¯ u
∗
i
¯ u
∗
j
− ¯ u
∗
i
¯ u
∗
j
= (¯ u
i
+ V
i
)(¯ u
j
+ V
j
) −(¯ u
i
+ V
i
) (¯ u
j
+ V
j
)
= ¯ u
i
¯ u
j
+ ¯ u
i
V
j
+ ¯ u
j
V
i
− ¯ u
i
¯ u
j
− ¯ u
i
V
j
−V
i
¯ u
j
= ¯ u
i
¯ u
j
− ¯ u
i
¯ u
j
=
1
c
r
L
m
ij
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 63
(115)
Numerical method
A numerical method based on an implicit, ﬁnite volume
method with collocated grid arrangement, central differ
encing in space, and CrankNicolson (α = 0.5) in time is
brieﬂy described below. The discretized momentum equa
tions read
¯ u
n+1/2
i
= ¯ u
n
i
+ ∆tH
_
¯ u
n
, ¯ u
n+1/2
i
_
−α∆t
∂¯ p
n+1/2
∂x
i
−(1 −α)∆t
∂¯ p
n
∂x
i
(116)
where H includes convective, viscous and SGS terms. In
SIMPLE notation this equation reads
a
P
¯ u
n+1/2
i
=
nb
a
nb
¯ u
n+1/2
+ S
U
−α∆t
∂¯ p
n+1/2
∂x
i
∆V
where S
U
includes all source terms except the implicit pres
sure. The face velocities ¯ u
n+1/2
f
= 0.5(¯ u
n+1/2
j
+ ¯ u
n+1/2
j−1
) do not
satisfy continuity. Create an intermediate velocity ﬁeld by
subtracting the implicit pressure gradient fromEq. 116, i.e.
¯ u
∗
i
= ¯ u
n
i
+ ∆tH
_
¯ u
n
, ¯ u
n+1/2
i
_
−(1 −α)∆t
∂¯ p
n
∂x
i
⇒ ¯ u
∗
i
= ¯ u
n+1/2
i
+ α∆t
∂¯ p
n+1/2
∂x
i
(117)
Take the divergence of this equation and require that ∂¯ u
n+1/2
f,i
/∂x
i
=
0 so that
∂
2
¯ p
n+1
∂x
i
∂x
i
=
1
∆tα
∂¯ u
∗
f,i
∂x
i
(118)
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 64
The Poisson equation for p
n+1
is solved with an efﬁcient
multigrid method (Emvin, 1997). In the 3D MG we use
an planebyplane 2D MG. The face velocities are corrected
as
¯ u
n+1
f,i
= ¯ u
∗
f,i
−α∆t
∂¯ p
n+1
∂x
i
(119)
Afewiterations (typically two) solving the momentumequa
tions and the Poisson pressure equation are required each
time step to obtain convergence. More details can be found
Davidson & Peng (2003)
1. Solve the discretized ﬁltered NavierStokes equation
for ¯ u, ¯ v and ¯ w.
2. Create an intermediate velocity ﬁeld ¯ u
∗
i
from Eq. 117.
3. The Poisson equation (Eq. 118) is solved with an efﬁ
cient multigrid method (Emvin, 1997).
4. Compute the face velocities (which satisfy continu
ity) from the pressure and the intermediate velocity from
Eq. 119
5. Step 1 to 4 is performed till convergence (normally
two or three iterations) is reached.
6. The turbulent viscosity is computed.
7. Next time step.
Since the Poisson solver is a nested MG solver, it is dif
ﬁcult to parallelize with PVM. Hence, when we do large
simulations (> 20M cells) we use a traditional SIMPLE
method.
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 65
SIMPLE solver on Linux clusters
CALCPVM: Pressure correction code (SIMPLE) code par
allelized based on domain decomposition (Nilsson, 1997;
Nilsson & Davidson, 1998)
Central differencing, CrankNicolson
Speedup on the IBM SP using MPI (Message Passing
Interface).
No. of proc. speedup
1 1
8 8
32 32
Note that the speedup is deﬁned as elapsed time to reach
convergence at each time step. This is a ﬂow where the par
allelization is unusually efﬁcient.
Oneequation k
sgs
model
A oneequation model can be used to model the SGS tur
bulent kinetic energy. The equation can be written on the
same form as the RANS kequation, i.e.
∂k
sgs
∂t
+
∂
∂x
j
(¯ u
j
k
sgs
) =
∂
∂x
j
_
(ν + ν
sgs
)
∂k
sgs
∂x
j
_
+ P
k
sgs
−C
ε
k
3/2
sgs
∆
ν
sgs
= c
k
∆k
1/2
sgs
, P
k
sgs
= 2ν
sgs
¯ s
ij
¯ s
ij
(120)
Note that the production term, P
k
sgs
, is equivalent to the
SGS dissipation in the equation for the resolved turbulent
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 66
kinetic energy (look at the ﬂow of kinetic energy discussed
at the of Davidson & Billson (2006)).
From the Kolmogorov −5/3 law we can obtain an value
of the C
ε
coefﬁcient. Using ε = C
ε
k
3/2
sgs
/∆ in Eq. 93 gives
k
sgs
=
3
2
C
_
C
ε
k
3/2
sgs
π
_
2/3
(121)
so that C
ε
= π(
3
2
C)
−3/2
which with C = 1.5 gives C
ε
= 0.93.
To derive a value on the c
k
constant, we use the expres
sion on ν
sgs
in Eq. 120 together with Eqs. 83 and 84
c
k
k
1/2
sgs
∆¯ s
2
= ε
Replacing ¯ s
2
with Eq. 91 and k
sgs
with Eq. 93 gives
c
k
_
3
2
C
_
∆ε
π
_
2/3
_
1/2
∆
3
2
π
4/3
Cε
2/3
∆
−4/3
= ε
so that c
k
=
_
3
2
C
_
−3/2
/π = 0.094 with C = 1.5.
A dynamic oneequation model
One of the drawbacks of the dynamic model of Germano
et al. (1991) (see p. 52) is the numerical instability associ
ated with the negative values and large variation of the C
coefﬁcient. Usually this problem is ﬁxed by averaging the
coefﬁcient in some homogeneous ﬂow direction. However,
in real applications, no such ﬂow direction exists. Below a
dynamic oneequation model is presented. The main object
when developing this model was that it should be applica
ble to real industrial ﬂows. Furthermore, being a dynamic
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 67
model, it has the great advantage that the coefﬁcients are
computed rather than being prescribed.
The equation for the subgrid kinetic energy reads (David
son, 1997b,c) (see also Sohankar et al. (2000b,a))
∂k
sgs
∂t
+
∂
∂x
j
(¯ u
j
k
sgs
) = P
k
sgs
+
∂
∂x
j
_
ν
eff
∂k
sgs
∂x
j
_
−C
∗
k
3/2
sgs
∆
P
k
sgs
= −τ
a
ij
¯ u
i,j
, τ
a
ij
= −2C∆k
1
2
sgs
¯ s
ij
(122)
with ν
eff
= ν + 2C
hom
∆k
1
2
sgs
. The C in the production term
P
k
sgs
is computed dynamically (cf. Eq. 104). To ensure nu
merical stability, a a constant value (in space) of C (C
hom
) is
used in the diffusion term in Eq. 122 and in the momentum
equations. C
hom
is computed by requiring that C
hom
should
yield the same total production of k
sgs
as C, i.e.
2C∆k
1
2
sgs
¯ s
ij
¯ s
ij
xyz
= 2C
hom
∆k
1
2
sgs
¯ s
ij
¯ s
ij
xyz
The dissipation term ε
k
sgs
is estimated as:
ε
k
sgs
≡ νT
f
(u
i,j
, u
i,j
) = C
∗
k
3/2
sgs
∆
. (123)
Now we want to ﬁnd a dynamic equation for C
∗
. The equa
tions for k
sgs
and K read in symbolic form
T(k
sgs
) ≡ C
k
sgs
−D
k
sgs
= P
k
sgs
−C
∗
k
3/2
sgs
∆
T(K) ≡ C
K
−D
K
= P
K
−C
∗
K
3/2
¸¸
∆
(124)
Since the turbulence on both the grid level and the test
level should be in local equilibrium (in the inertial −5/3
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 68
region), the lefthand side of the two equations in Eq. 124
should be close to zero. An even better approximation should
be to assume T(k
sgs
) = T(K), i.e.
¸¸
P
k
sgs
−
1
∆
¸ ¸
C
∗
k
sgs
3/2
= P
K
−C
∗
K
3/2
¸¸
∆
,
so that
C
n+1
∗
=
_
P
K
−
¸¸
P
k
sgs
+
1
∆
¸ ¸
C
n
∗
k
3/2
sgs
_
¸¸
∆
K
3
2
. (125)
The idea is to put the local dynamic coefﬁcients in the source
terms, i.e. in the production and the dissipation terms of
the k
sgs
equation (Eq. 122). In this way the dynamic coef
ﬁcients C and C
∗
don’t need to be clipped or averaged in
any way. This is a big advantage compared to the standard
dynamic model of Germano (see discussion on p. 57).
A Mixed Model Based on a OneEq. Model
Recently a new dynamic scalesimilarity model was pre
sented by Krajnovi´ c & Davidson (2002b). In this model a
dynamic oneequation SGS model is solved, and the scale
similarity part is estimated in a similar way as in Eq. 114.
Applied LES
At the Department we used LES for applied ﬂows such
as ﬂow around a cube (Krajnovi´ c & Davidson, 2002a; Kra
jnovi´ c, 2002), the ﬂow and heat transfer in a square rotat
ing duct Pallares & Davidson (2000, 2002), the ﬂow around
a simpliﬁed bus (Krajnovi´ c & Davidson, 2003; Krajnovi´ c,
Large Eddy Simulations
Lars Davidson: MTF270 Turbulence Modelling
http://www.tfd.chalmers.se/doct/comp turb model 69
2002), a simpliﬁed car (Krajnovi´ c & Davidson, 2005a,b,c)
and the ﬂowaround an airfoil Dahlstr¨ om&Davidson (2003);
Dahlstr¨ om(2003). We have also done some work on buoyancy
affected ﬂows Peng & Davidson (2002, 2001); Peng (1998).
Large Eddy Simulations