You are on page 1of 19

Broadband slat noise prediction based on CAA and stochastic

sound sources from a fast random particle-mesh (RPM) method


Roland Ewert
*
DLR, Deutsches Zentrum fu r Luft- und Raumfahrt, Institut fu r Aerodynamik und Stro mungstechnik, Technische Akustik,
Lilienthalplatz 7, 38108 Braunschweig, Germany
Available online 20 February 2007
Abstract
The application of a low-cost computational aeroacoustics (CAA) approach to a slat noise problem is studied. A fast and ecient
stochastic method is introduced to model the unsteady turbulent sound sources in the slat-cove of a high-lift airfoil. It is based on
the spatial convolution of spatiotemporal white-noise and can reproduce target distributions of turbulence kinetic energy and length
scales, such as that provided by a RANS computation of the time-averaged turbulent ow problem. The computational method yields
a perfectly solenoidal velocity eld. For homogeneous isotropic turbulence, the complete second-order two-point velocity correlation
tensor is realized exactly. Two RANS turbulence models are applied to the slat noise problem to study how sensitive the aeroacoustics
predictions depend on turbulence kinetic energy predictions. Results for the sound generation at the slat are given for a Menter SST
turbulence model with and without KatoLaunder modication. The aeroacoustic simulations yield a characteristic narrow band spec-
trum that compares very well with the experimental data. The directivities found point toward an edge noise mechanism at the slat as the
main cause for slat noise sound generation.
2007 Elsevier Ltd. All rights reserved.
1. Introduction
Aircraft noise reduction, as achieved through the devel-
opment and application of high bypass low noise turbofan
engines, has shifted the focus of interest to airframe noise
as an equally important noise source during the approach
phase. Experimental studies have identied deployed slats
as prominent noise contributors [1,2]. To reduce noise lev-
els further, numerical tools will become essential in the
future to achieve an optimized low noise design of airframe
components. Therefore, the development of ecient and
quick computational methods, which can be used in an
optimization process, is crucial. Over recent years, compu-
tational methods have been improved for computational
aeroacoustics (CAA) purposes (in particular, high-order
non-dispersive spatial and temporal discretization schemes
and high-quality non-reecting boundary conditions have
been introduced). A key point, however, is how an acoustic
solver based on the linearized Euler equations (LEE) or
derivatives of it can be used for designing low-noise air-
frame components (design-to-noise approach). A presum-
ably accurate noise prediction methodology arises if
appropriate acoustic sources of the propagation equations
are computed from time accurate large eddy simulation
(LES) or direct numerical simulation (DNS) of the turbu-
lent near eld. However, such approaches are too time con-
suming to be used in a design process that requires
numerous evaluations of design modications.
In this paper, a computational approach with consider-
ably increased eciency is studied for a slat noise problem.
It involves CAA techniques in conjunction with a stochas-
tic method to set up turbulence velocities for broadband
sound sources. The random particle-mesh (RPM) method
is introduced as a new technique to generate uctuating
turbulence velocities in the time domain that realize the
local statistical features as provided by a steady state
0045-7930/$ - see front matter 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compuid.2007.02.003
*
Tel.: +49 531 295 2177; fax: +49 531 295 2320.
E-mail address: roland.ewert@dlr.de
www.elsevier.com/locate/compuid
Available online at www.sciencedirect.com
Computers & Fluids 37 (2008) 369387
RANS solution. The RPM method is based on the spatial
ltering of white-noise. The method is a further develop-
ment of the technique used by Klein et al. [3] to set up uc-
tuations at a LES inow boundary. Their technique
accurately reproduces rst- and second-order one-point
statistics of the turbulent velocity components. The method
is extended in the RPM framework to prescribe a diver-
gence-free (solenoidal) turbulent velocity eld on a source
patch within the CAA domain, which avoids the occur-
rence of spurious monopolar sound sources. The solenoidal
extension is shown to reproduce exactly the second-order
two-point correlation tensor of homogeneous isotropic tur-
bulence in two or three dimensions.
The rst stochastic method to generate spectra and cor-
relation functions of turbulence was introduced by Kraich-
nan in the early 1970s [4], which is based on random
Fourier modes. A modied version of this approach was
applied by Bechara et al. [5] to model noise sources of free
turbulent ows. Further improvements of the stochastic
noise generation and radiation (SNGR) method were
introduced by Bailly and Lafon [6,7] and Billson et al.
[8]. Kalitzin et al. [9] applied the SNGR approach in a rst
attempt to predict trailing edge noise. Ewert and Bauer [10]
derived stochastic surface sources from SNGR volume
sources for the prediction of broadband trailing edge noise.
The new stochastic RPM method presented herein has
some advantages over random Fourier mode methods.
Besides being strictly solenoidal, it is time- and memory
ecient, and can be easily applied to highly non-uniform
mean-ow elds (e.g. as apparent in the slat-cove region
of a high-lift airfoil) using local results for the turbulence
kinetic energy and the corresponding turbulent length scale
from a RANS solution of the turbulent mean-ow. Fur-
thermore, it avoids the occurrence of shear decorrelations
and resolves broadband spectra continuously. A further
extension to non-homogeneous anisotropic solenoidal
ows is possible with the transformation proposed by
Smirnov et al. [11].
This paper is organized as follows. Section 2 gives an
overview about the stochastic method used in this paper.
Section 3 describes some details about the CAA method
employed to integrate the acoustic eld. Section 4 gives
computational results for simulated homogeneous isotro-
pic turbulence and for the sound generation at the slat of
a high-lift airfoil in comparison to experimental ndings.
The slat-noise computations utilize two turbulence models
and are carried out for a Mach number Ma = 0.10. A mod-
ied high-lift conguration with retracted ap is used in
order to avoid additional sound sources in the experiments.
Finally, conclusions are drawn in Section 5.
2. Stochastic sound source modeling
2.1. Statistical sound source modeling
Sound generation and propagation due to turbulent
sources is described by acoustic analogies, which have the
form of a forced linear acoustic wave equation with wave
operator L, acoustic variable p
0
, and source Q
Lp
0
x; t Qx; t: 1
The source Q usually depends on time and position and
is a function of the turbulent velocity uctuations for tur-
bulence related sound. Starting from Eq. (1) the far-eld
spectral density of the acoustic variable can be expressed
in terms of the statistical properties of the source [12]
Sx; x
_
y
_
r

x; y; x

Gx; y r; x

Q
12
y; r; xdr dy:
2
Here, S(x,x) stands for the power spectrum of the acoustic
variable p
0
in (1), i.e.,
p
02

1
2p
_
1
1
Sx; x dx:
The quantities

G and

G

denote, respectively, the appro-


priate Greens function and its conjugate complex of the
problem

Lpx; x

Qx; x; 3
which is the frequency domain Fourier transform of Eq.
(1). The Greens function is dened by the linear wave-
operator of Eq. (3), i.e.

L

Gx; y; x dx y, and the


appropriate boundary conditions. In the presence of imper-
meable surfaces, for example, it is the exact Greens func-
tion that satises o

G=on 0 on the surface, whereas for


sound radiation into unbounded space the free space
Greens function would be the appropriate choice.

Q
12
is
a cross spectral density which is obtained by Fourier trans-
form of the two-point spacetime correlation of source Q
in Eq. (1) between source points y
1
y and y
2
y r
and time separation s t
1
t
2

Q
12
y; r; x
_
1
1
Ry; r; se
ixs
ds; 4
Ry; r; s Qy; tQy r; t s: 5
Eq. (2) is the basis of all statistical noise theories in the fre-
quency domain. It evidences that the far-eld spectrum is
completely determined if just the two-point spacetime cor-
relation of the sources, Eq. (5), is known. For turbulence
related noise sources this means a considerable reduction
in the amount of essential statistical information, since
the underlying turbulent ow eld is determined by an in-
nite set of m-point nth-order correlations.
As discussed by Tam and Auriault, one useful model
spacetime correlation function characterized by three
parameters may be expressed by Gaussian spatial and
exponential temporal correlations. It closely approximates
the measurements of Davies et al. [13] for the uctuating
axial velocity component in jets. This generic two-point
spacetime correlation can be written as
370 R. Ewert / Computers & Fluids 37 (2008) 369387
Ry; r; s

R exp
s
s
s

pr u
c
s
2
4l
2
s
_ _
: 6
The parameters s
s
and l
s
dene, respectively, the correla-
tion time- and length scales and

R denotes the mean-square
value of the correlated quantity for vanishing separation
space r and time s. Taylors hypothesis is taken into
account by the convection velocity u
c
. For inhomogeneous
turbulence s
s
, u
c
, l
s
, and

R depend on position y. In a
comoving frame of reference, just considering frozen con-
vection through s
s
! 1, the correlation takes on a Gauss-
ian form. Although Eq. (6) is intended to approximate
axial velocity correlations, it is also used with great success
by Tam and Auriault [14,15] to model the statistical fea-
tures of a full jet noise source term, which describes ne
scale (broadband) turbulence noise. Such a source term is
in general a non-linear function of all turbulent velocity
components. Tam and Auriault solve problem (2) with
the statistical source features expressed through Eq. (6),
using a CAA method to compute the ow Greens function
in Eq. (2). It utilizes the adjoint Euler equations in spectral
space. Agarwal and Morris recently extended the statistical
source description through Eq. (6) to an airframe noise
problem [16]. On a near-eld grid that resolves the noise
sources, they compute for each grid point, frequency
band, and observer position the exact Greens function
with a boundary element method (BEM). Convection
eects are considered on the basis of a uniform mean-
ow.
Solving the acoustic analogy Eq. (1) with a synthetic
uctuating source, which satises the constraint to match
the two-point correlations of a statistical approach, pro-
vides as a consequence of Eq. (2) an identical far-eld spec-
trum. In other words, a synthetic uctuating source term
that realizes the proper two-point spacetime correlations
gives an alternative way to solve Eq. (2). It yields an exact
solution of the acoustic far-eld spectrum if the underlying
second-order two-point correlations are realized through
the uctuating source.
The random particle-mesh (RPM) method, which will
be introduced herein, is capable of generating spatially
and temporally uctuating quantities that reproduce target
two-point spacetime correlations of the type described by
Eq. (6), whereby local target values for the parameters are
realized. In other words, sources with the statistical fea-
tures corresponding to that used by Tam and Auriault
and subsequently by Agarwal and Morris in statistical
noise source models can be generated with the RPM
method for a direct time-domain approach. A direct
CAA approach to solve Eq. (2) via Eq. (1) with uctuating
sources has the advantage that only one computation is
necessary to determine broadband acoustic wave genera-
tion and propagation through non-uniform mean-ow
for all frequency bands and observer positions, whereas
an adjoint approach in the frequency domain needs n m
computations for n observer positions and m frequency
bands.
2.2. Improved turbulent velocity modeling
The source Q in Eq. (1) generally is a function of the tur-
bulent velocity uctuations if vortex sound is considered.
To model vortex sound sources by prescribing the underly-
ing velocity uctuations, some further physical features
may have to be introduced. For instance, the two-point
correlations of all velocity components form a second rank
tensor, which is inadequately modeled by one scalar corre-
lation function of the type given in Eq. (6) (which is
intended to resolve just the axial velocity correlations).
Furthermore, velocity uctuations originating from vortic-
ity are divergence-free (solenoidal). Deviations of the mod-
eled uctuations from this constraint may give rise to
spurious sound sources. Both constraints are taken into
account in this work by dening the uctuating quantity
of the stochastic model to be a scalar streamfunction
wx
i
; t (not e.g. the source full source Q), from which
velocity components are deduced subsequently. Hence,
the 2D uctuating velocities are
u
0
i

ij
ow
oy
j
; 7
where
ij
denotes the 2D -tensor. The dierentiation with
respect to one spatial direction, which appears in Eq. (7),
causes a reduction of the integral length-scale of the veloc-
ity uctuations in that particular direction (due to the
amplication of higher wave-numbers in wave-number
space). Hence, a uctuation model that realizes for the uc-
tuating streamfunction wx
i
; t (isotropic) correlations of
the type described in Eq. (6) apparently causes an aniso-
tropic distortion of the resulting velocity eld. It is worth
to note and to show that these distorted velocity elds actu-
ally exhibit rather physical characteristics. That is, in the
case of a homogeneous problem (i.e. in the case of constant
parameters in Eq. (6)) the velocity correlations that follow
from Eq. (7)
R
ij
r; s u
0
i
r
1
; t
1
u
0
j
r
2
; t
2
; 8
match perfectly the complete velocity correlation tensor of
isotropic turbulence for s = 0,
R
ij
r; 0
2
3

k
f r gr
r
2
r
i
r
j
gr
^
d
ij
_ _
: 9
In the above expressions, r r
1
r
2
and s jt
1
t
2
j
are the spatial and temporal separations between points 1
and 2, and
^
d
ij
is the Kronecker symbol. The separation dis-
tance is r jrj and r
i
is the ith component of vector r. The
turbulence kinetic energy is denoted by

k and f(r) and g(r)
denote the longitudinal and lateral correlation functions,
respectively, which are normalized such that f = 1 and
g = 1 for r = 0. Both correlations are connected for a
two-dimensional problem through
gr f r r
df r
dr
: 10
R. Ewert / Computers & Fluids 37 (2008) 369387 371
The feature of a scalar streamfunction with isotropic
correlation to perfectly describe the two-point correlations
of homogeneous isotropic turbulence in 2D was shown by
Careta et al. [17]. It can be proven by considering the
spacetime correlations of the uctuating streamfunction
in this homogeneous case to be a pure function of the sep-
aration vector r
R
ww
r; s wr
1
; t
1
wr
2
; t
2
: 11
The velocity correlations equation (8) are connected via
Eq. (7) with the correlations equation (11). By taking
r r
1
r
2
into account, the relationship is
R
ij

ik

jl
o
2
R
ww
or
k
or
l
: 12
Now, let us consider the correlation equation (11) to de-
pend for a vanishing time separation s = 0 only on the ra-
dial distance r jrj, i.e., R
ww
r; 0 Cr. Then, by
expressing R
ww
in Eq. (12) through Cr, the R
ij
can be ver-
ied to agree formally with Eq. (9). Furthermore, the lon-
gitudinal and lateral correlation functions f(r) and g(r) can
be identied to be related to the scalar eld Cr via
f r
1
C
00
0
C
0
r
r
; gr
C
00
r
C
00
0
; 13
where C
n0
r d
n
Cr=dr
n
and C
00
0
C
00
0. Thereby, C
00
0
is
used to normalize f(r) and g(r) such that f 0 g0 1.
Note that C
00
0 is also used for f(r), since
lim
r!0
C
0
r
r
C
00
0:
It is easy to prove that Eq. (10) is satised when insert-
ing expressions (13) into it. Hence, the only constraint to
perfectly realize the correlation tensor of homogeneous iso-
tropic turbulence (HIT) in 2D is that the correlation of the
uctuating scalar streamfunction depends for s = 0 just on
radial distance. This constraint is satised if the uctuating
streamfunction is generated by a procedure that realizes
correlations of the type described in Eq. (6). It is evident
that in the homogeneous case, where the parameters in
Eq. (6) are position independent constants, the constraint
R
ww
r; 0 Cr is fullled. Realizing correlations R
ww
described in Eq. (6), the amplitude

R has to be scaled such
that 2=3

k C
00
0

R is satised to match quantitatively Eq.


(9). This yields

R
4l
2
s

k
3p
: 14
The resulting longitudinal correlation function that fol-
lows from Eq. (6) is a Gaussian,
f r exp
p
4
r
2
l
2
s
_ _
; 15
with an integral length scale directly determined by the
parameter l
s
,
L
_
1
0
f r dr l
s
:
Hence, the parameters

R and l
s
in Eq. (6) are directly
linked via Eqs. (14) and (15) to the turbulence kinetic
energy

k and length scale L.
An extension of the method of Careta et al. to 3D
homogeneous isotropic turbulence is straightforward by
introducing a uctuating vector streamfunction w
k
x
i
; t,
which denes the uctuating velocities via
u
0
i

ijk
ow
k
ox
j
; 16
with
ijk
being the third rank permutation tensor. It is obvi-
ous that the u
0
i
satises the divergence-free condition. Fur-
thermore, if the components of the vector streamfunction
are statistically independent, each having similar spatial
correlations, or equivalently
R
w
i
w
j
r; 0 w
i
r
1
; tw
j
r
2
; t
^
d
ij
Cr; 17
it can be shown that the velocity correlations also satisfy
Eq. (9). By extending the 2D concept, the velocity correla-
tions equation (8) follow to be related through Eqs. (16) to
(17) via
R
ij

ikl

jmn
o
2
w
l
w
n
or
k
or
m
: 18
By inserting
^
d
ij
Cr for the streamfunction correlations in
Eq. (18), straightforward algebra proves the formal shape
of Eq. (9) to be realized. The longitudinal and lateral cor-
relation functions are related in the 3D case to the correla-
tion function of the streamfunction components by
f r
1
C
00
0
C
0
r
; gr
C
00
C
0
=r
2C
00
0
: 19
These expressions dier slightly from their 2D counter-
part, Eq. (13). Note, however, that in 3D the relationship
between the longitudinal and lateral correlation function
changes from Eq. (10) to [18]
gr f r
r
2
df r
dr
: 20
This relation between the longitudinal and lateral correla-
tion function in 3D is perfectly satised, which can be
shown by inserting the expressions (19) into Eq. (20).
Hence, the uctuating 3D velocity elds, which are dened
by Eq. (16) in conjunction with three uctuating stream-
functions satisfying Eq. (17), realize the correlation tensor
of HIT in 3D. The additional term for g(r) in Eq. (19)
causes the 3D lateral correlation function to have a less
pronounced negative dip compared to the 2D case, Eq.
(13). In 3D the amplitude in Eq. (6) has to match
2=3

k 2C
00
0

R, which yields

R
2l
2
s

k
3p
: 21
372 R. Ewert / Computers & Fluids 37 (2008) 369387
The turbulence properties of technical ow problems
often deviate from homogeneous isotropic turbulence.
However, modelling these features serves also as a starting
point to consider anisotropy in a next step. For instance,
anisotropic turbulence based on modeled 3D homogeneous
isotropic turbulence can be achieved with the additional
velocity transformation proposed by Smirnov et al. [11].
The transformation is applied to initially homogeneous iso-
tropic velocities, e.g. as provided in Eq. (16). The outcome
of the transformation is a time-dependent ow-eld v
0
i
with
one-point correlation functions v
0
i
v
j
equal to that of the
local (anisotropic) Reynolds stress tensor s
ij
R
ij
0; 0
and turbulent length/time scales equal to l
s
and s
s
. The
ow-eld is also exactly divergence free for homogeneous
turbulence and to a high degree divergence-free for inho-
mogeneous turbulence.
2.3. Stochastic uctuating elds via RPM
The random particle mesh (RPM) method delivers an
unsteady uctuating streamfunction whose correlation
function matches the formal shape prescribed by Eq. (6).
A uctuating streamfunction is generated by spatially l-
tering a white-noise eld. The procedure is the discrete real-
ization of a convolution or spatial ltering integral, which
reads in 2D
wx; t
_ _
A
S

Ax
0
G
0
jx x
0
j; l
s
x
0
Ux
0
dx
0
: 22
In Eq. (22), G
0
is a lter kernel, Udenotes a spatiotemporal
white-noise eld with unity spectrum, and w is the uctuat-
ing streamfunction, which determines the pseudo-turbulent
velocity uctuations via Eq. (7). The lter kernel is normal-
ized such that wx; twx; t 1 for

A 1. The integration
is carried out on a source patch, A
S
. The argument of the
lter kernel indicates that it is a function of the separation
distance jx x
0
j, and of the locally x
0
-dependent kernel
width l
s
and kernel peak amplitude

A. For the simulation
of frozen turbulence considered in this work, the spatio-
temporal white-noise eld is uniquely dened by the
properties
Ux; t 0; 23
Ux; tUx r; t dr; 24
D
0
Dt
U 0; 25
where dr denotes a multi-dimensional Dirac d-function,
which reads in 2D dr dr
1
dr
2
. Eq. (25) introduces
the convection property into the uctuation model through
the passive convection of the white-noise eld in the mean-
ow u
0
. It is to be understood such that in a locally comov-
ing frame of reference the spatiotemporal white-noise eld
remains locally static. This condition can be satised even
for an ideal white-noise realization (which is non-dieren-
tiable) although the substantial time derivative D
0
=Dt
o=ot u
0
$ involves spatial and temporal derivatives. The
denition equation (24) has to be satised for each point
in the domain where U is dened (in the source region),
whereas the convection property equation (25) determines
the solution to be completely prescribed by the values at
the inow boundary after a transient time period. It can
be rigorously proven that both constraints can be satised
simultaneously.
The method is sketched in the following for a homoge-
neous one-dimensional problem. In this case the uctuating
eld w results from the one-dimensional convolution of the
unity white-noise eld U with a lter kernel G(r):
wx; t
_
1
1
Gx x
0
Ux
0
; t dx
0
: 26
G(r) is supposed to be a real and even function, i.e.,
Gr Gr. Denition (24) becomes
Rr Ux r; tUx; t dr: 27
Inserting Eq. (26) into the denition of the correlation of
the uctuating streamfunction Rr : wx r; twx; t
and using the properties of Eq. (27), the correlation of
the ltered eld becomes the self convolution of the lter
kernel
Rr
_
1
1
Gr nGn dn GHG: 28
In general, the lter kernel can be weighted with an arbi-
trary constant. In the following the constant is assumed
to be chosen such that the correlation Rr of Eq. (28) be-
comes the normalized correlation R
0
r of w with
R
0
0 1 since
R
0
r :
wx r; twx; t
wx; twx; t
:
The related lter kernel is labeled G
0
r. Due to Eq. (28)
the spatial Fourier transform of the normalized correlation
function reads

R
0
k

G
0
k
2
; 29
where k denotes the wave-number and relates to the spatial
separation r. Note, since G
0
r is a real and even function,

G
0
k is also real and even. In the last step, it follows that
the lter kernel is specied for a given correlation R
0
r by

G
0
k

R
0
k
_
: 30
Practically, the lter kernel of any meaningful correlation
can be achieved through Eq. (30). However, the lter ker-
nel, which is used in this paper, is Gaussian. It is easy to
show using Eq. (29) that a spatial Gaussian kernel
G
0
r exp
p
2
r
2
l
2
s
_ _
31
yields for the correlation function again a Gaussian, but
with a width a factor

2
p
larger, i.e.
R. Ewert / Computers & Fluids 37 (2008) 369387 373
R
0
r exp
p
4
r
2
l
2
s
_ _
: 32
The extension of the scalar streamfunction approach to
two or three dimensional source domain is straightforward.
Integral (26) becomes for a n-dimensional problem
wx; t
_

_
..
n

AG
0
x x
0
Ux
0
; t dx
0
: 33
Due to the white-noise eld denition equation (24), the
normalized correlation becomes
R
0
r
_

_
G
0
r nG
0
n dn: 34
Since the Gaussian lter is separable, i.e. for the n-
dimensional problem it can be split according to
G
0
jxj

n
i1
G
0
x
i
;
the right-hand side of Eq. (34) decomposes into n products
of decoupled one-dimensional convolutions, eventually
leading with the one-dimensional ndings equations (31)
and (32) to a correlation
R
0
r exp
p
4
jrj
2
l
2
s
_ _
35
based on the lter kernel
G
0
r exp
p
2
jrj
2
l
2
s
_ _
: 36
The separation property of a Gaussian lter kernel
allows to carry out the integration in Eq. (33) in a sequence
of one-dimensional lter operations, which makes the l-
tering procedure very ecient. The correlation equation
(35) is a function of radial distance necessary to realize
homogeneous isotropic turbulence, Section 2.2.
Assuming a constant mean-owu
0
in (25), the convection
property equation (25) is solved by Ux; t Ux u
0
s.
Hence, Ux r; t s can be expressed as Ux r
u
0
s; t. Accordingly, the white-noise property (24) becomes
for s 6 0
Ux; tUx r; t s dr u
0
s;
which subsequently yields a normalized spatial correlation
of the form
R
0
r; s exp
pr u
0
s
2
4l
2
s
_ _
: 37
Note that Eq. (6) takes on the form realized through Eq.
(37) for u
c
u
0
and s
s
! 1, i.e., for frozen turbulence. To
introduce an additional exponential temporal correlation
as in (6) the homogeneous convection equation (25) has
to be modied into a Langevin equation by adding a
white-noise source term to the right-hand side, whose pre-
factor determines the correlation time s
s
.
A 3D extension of the streamfunction approach is pos-
sible by determining the components of the 3D vector
streamfunction w through
w
i
x; t
_ _
V
S
_

AG
0
x x
0
U
i
x
0
dx
0
: 38
Here, the U
i
denote three independent stochastic white-
noise elds of the type dened by Eqs. (23)(25) and G
0
is a scalar lter kernel. The scalar amplitude

A controls
the amplitude of the uctuations. Due to their denition
and independence, the white-noise elds satisfy
U
i
x; tU
j
x r; t
^
d
ij
dr: 39
Hence, applying the lter kernel equation (36) the condi-
tion (17) holds for the 3D realization of a uctuating vector
streamfunction through Eq. (38) (independent correlations
that spatially depend only on radial distance), such that
this extended ltering procedure is capable of providing
3D homogeneous isotropic turbulence.
Rigorously, the derivation presented so far is restricted
to non-uniform mean-ows u
0
but homogeneous lter
kernels that realize constant correlations and length-scales
throughout the source domain. Local kernels that realize
inhomogeneous correlations and length-scales can also be
deduced. However, the variation of these stationary quan-
tities is usually small compared to the turbulent length-
scale itself. Numerical test indicated the analytical ndings
for homogeneous lter kernels to also be valid with good
accuracy for length scales l
s
x
0
and kernel amplitudes

Ax
0
not locally varying too strong. The amplitude has
to be chosen such that the locally generated turbulence
kinetic energy 2=3

k R
ii
equals the target value. The
appropriate RANS based scaling will be discussed in Sec-
tion 2.5.
2.4. Numerical discretization of the RPM method
In the 2D discretization of the RPM method the contin-
uous integral equation (22) is approximated through the
nite sum
wx; t

i;j
G
ij
xr
ij
t: 40
Eq. (40) follows by splitting the source domain A
S
in Eq.
(22) into M subdomains DA
ij
and by approximating the
integral through the summation over all subdomains. Here
the indices i and j dene discrete points on the 2D source
patch. The amplitude

A is absorbed in the lter kernel,
i.e. G
ij
x Gx; x
0
ij


Ax
0
ij
G
0
x; x
0
ij
. It is advantageous
to evaluate the lter kernel at the cell center
x
s
ij

1
DA
ij
_ _
DAij
x
0
dx
0
41
374 R. Ewert / Computers & Fluids 37 (2008) 369387
of subdomain DA
ij
, i.e. G
ij
x :

A
ij
x
s
ij
G
0
ij
x; x
s
ij
.
The quantity r
ij
in Eq. (40) is a random value dened
through
r
ij
t
_ _
DAij
Ux
0
; t dx
0
: 42
Let hU
ij
i : r
ij
=DA
ij
be the average of the white noise
eld over DA
ij
, then Eq. (40) is
wx; t

i;j
G
ij
xhU
ij
iDA
ij
: 43
Since hU
ij
i !U
ij
holds in the limit of innitely small sub-
domains DA
ij
! 0, Eq. (43) and thus Eq. (40) is a con-
sistent approximation to Eq. (22). Two basic
approximations are introduced in Eq. (43):
(1) hU
ij
i is a ltered approximation to U
ij
.
(2) The summation is a fourth-order accurate approxi-
mation to the integral based on hU
ij
i as the underly-
ing white-noise eld.
Proof. To show the integral approximation to be fourth-
order accurate, rewrite Eq. (22) using an averaged white-
noise eld hU
ij
i instead of U
ij
wx; t

i;j
_ _
DA
ij
Gx; x
0
hU
ij
idx
0

i;j
hU
ij
i
_ _
DAij
Gx; x
0
dx
0
: 44
Now, inserting a Taylor expansion of the kernel around
x
s
ij
, i.e. G
0
G
ij
$G
ij
x
0
x
s
ij
Oh
2
(with G
ij
:
Gx; x
s
ij
, G
0
: Gx; x
0
, and h
2
/ DA
ij
), and using the def-
inition equation (41), it follows that
wx; t

i;j
G
ij
xhU
ij
iDA
ij
Oh
4
:
A further simplication is introduced by truncating the
summation over all M subdomains through omission of
all contribution with
jG
0
ij
j < ; 45
i.e. the summation in Eq. (40) is carried out for the N 6 M
subdomains closest to a point x. A value of
10
2
. . . 10
4
was found to yield a sucient accurate
approximation to the integral equation (22).
The local integration of the white-noise eld over DA
ij
in
Eq. (42) can be deemed a low-pass lter applied to the eld,
which causes a spectral cut-o of the corresponding spec-
trum. As long as this cut-o wave number is larger than
the highest wave-number to be resolved by the uctuating
streamfunction, such an approximation has only little eect
on the resolved scales. All further properties of r
ij
can be
derived from the denitions equations (23)(25) of the spa-
tiotemporal white-noise eld U. Due to Eq. (23) and de-
nition equation (42) the random value exhibits the
property r
ij
0. Using denition equation (24), the corre-
lation of the random values becomes
r
ij
r
kl

_ _
DAij
_ _
DA
kl
dx x
0
dxdx
0

0 if i 6 k _ j 6 l;
DA
ij
if i j ^ k l:
_
46
That is, r
ij
is a uctuating quantity with zero mean and
mean-square value r
2
ij
DA
ij
. According to Eq. (25) the
white-noise eld remains locally static in a comoving frame
of reference. That is, if DA
ij
describes a convecting control
volume, whose boundary curve is drifting in the mean-ow,
Fig. 1, the random value r
ij
dened by Eq. (42) is indepen-
dent of time in incompressible mean-ow
1
. In a second-
order consistent approximation the cell center convects
with its local mean-ow velocity
_ x
s
ij
u
0
x
s
ij
Oh
2
: 47
Proof. Consider a small element dA in subdomain DA
ij
at
initial position x
0
t, Fig. 1. Taylor expansion gives its
position a time s later with x
0
t s x
0
t dx
0
t=dt
s
1
s
2
. The element drifts with the mean-ow, i.e.
dx
0
=dt u
0
0
. The velocity u
0
0
at x0 follows from a spatial
Taylor expansion around the subdomain center x
s
ij
:
u
0
0
u
s
ij
$u
s
ij
x
0
t x
s
ij
t
2
h
2
, with u
s
ij
: u
0
x
s
ij
t).
Then the cell center location at t s follows by introducing
x
0
t s x
0
t u
s
ij
s $u
s
ij
x
0
t x
s
ij
ts
1
s
2

2
h
2
s
into Eq. (41). Evaluation of the integral (in incompressible
ow dA and DA
ij
are invariants) and dierentiation with re-
spect to s yields the convection velocity equation (47) of the
cell center at time level t (i.e. at s 0). h
The discretization can be interpreted such that the
source domain A
S
is resolved by M discrete particles at
locations x
s
ij
. Each particle carries (in low Mach number
Fig. 1. Sketch of one drifting control volume DA
ij
;
1
and
2
denote
constants that limit the leading temporal and spatiotemporal error terms.
1
For compressible 2D mean-ow the control volume changes its
magnitude (q
0
DA
ij
is conserved) such that q
0
r
ij
remains constant.
R. Ewert / Computers & Fluids 37 (2008) 369387 375
ow) a frozen random value r
ij
and convects with its local
mean-ow velocity through the source domain. In this pic-
ture the random values r
ij
represents, the white-noise eld
for the surrounding control volume DA
ij
.
The source patch is constructed by following the paths
of the mean-ow streamlines, which start along an
upstream seeding line, to a user-dened downstream posi-
tion. See e.g. Fig. 2a that shows a bundle of streamlines
to resolve the slat-cove shear layer of a high-lift airfoil,
whose initial streamlines are equidistantly distributed along
the seeding line. The complete source domain is depicted in
Fig. 2b.
Random particles are introduced along each streamline,
whereby a constant drift time separation, Dt, between the
particles is realized. Since the maximal time to reach the
downstream border of a source patch depends on the con-
sidered streamline, the number of discrete particles also
varies accordingly, see e.g. Fig. 3. The sketch furthermore
highlights the area DA
ij
surrounding each discrete particle.
The drift separation time and the number of streamlines
determine the total number of random particles involved
as well as the size of the subdomain DA
ij
. The drift separa-
tion usually is chosen to be larger than the CAA time step.
An approximation to white-noise with a root-mean
square (RMS) value normalized to one can be realized
through a sequence of random numbers in the range

3
p
, generated with a constant clock rate. The highest
resolved frequency of this realization is linked to the seed-
ing clock-rate Dt through the sampling theorem, i.e.
f
max
1=2Dt. To achieve a mean-square value r
2
ij
DA
ij
through a sequence of random numbers, each element ij
has to take on a random value in the range

3DA
ij
_
. As
an example Fig. 4a shows the time history of the random
values r
ij
at a xed location inside the source domain for
a convection velocity normalized to one. The correspond-
ing spectrum is presented in Fig. 4b. It evidences a good
realization of a unity spectrum up to approx. 50% of the
Nyquist frequency 2Dt
1
.
The complete algorithm to compute the streamfunction
becomes
for each CAA time increment convects the random par-
ticles downstream;
if a particle crosses the downstream border delete it and
update the rst upstream position with a new random
particle with a new random value in the range

3DA
ij
_
;
lter and simultaneously interpolate the random eld
onto the CAA grid.
The local value of the turbulence kinetic energy at the
random particle location scales the local amplitude

A of
the lter kernel. The exact value of

A based on the RANS
mean-ow eld will be given in Section 2.5. The ltered val-
ues are directly computed for the relevant CAA grid points.
The lter kernel is computed in a sequence of one-
dimensional lter operations, see Fig. 5. It takes typically
101% of the time the direct evaluation of the full lter ker-
nel would need in 2D and 3D, respectively. First, the ran-
dom eld is ltered along the streamline for each discrete
point on the streamline, using the length scale at each par-
ticle location for the kernel scaling. Next, the intermediate
Fig. 2. Resolution of the slat shear-layer in the 2D test problem: (a) streamlines in the slat-cove and (b) resolved source domain and curvilinear multi-
block CAA mesh.
Fig. 3. Sketch of streamlines and discrete particles in non-uniform mean-
ow.
376 R. Ewert / Computers & Fluids 37 (2008) 369387
ltered values are distributed onto the CAA grid, Fig. 5.
Let us denote D
1
the length-scale at a given stochastic par-
ticle position A on the streamline, P a point on the stream-
line with smallest distance to the CAA grid point B, s the
distance along the streamline to the base-point, D
2
the
length-scale at the base-point P, and d the distance to
the grid point B, Fig. 5. Then the contribution of a random
element to the grid point due to the lter kernel reads
G
0
AB
: G
0
x
B
; x
A
G
0
dG
0
s
exp
p
2
s
2
D
2
1
_ _
exp
p
2
d
2
D
2
2
_ _
: 48
For uniform ow with constant length-scale this is identical
with the Gaussian kernel equation (31). For curved stream-
lines a small distortion of the kernel shape occurs.
2.5. Scaling of the lter kernel parameters from RANS
The lter kernel parameters l
s
and

A are scaled using a
steady RANS computation. In general, based on the turbu-
lence kinetic energy

k and the mean dissipation rate , an
integral length scale of turbulence can be deduced shows
that
l
s
c
l

k
3=2

: 49
Following the discussion of Bailly and Juve [7], the con-
stant can be estimated to be c
l
% 0:54 for a modied von
Karman spectrum. Although a Gaussian spectrum is used
in this work, the value for c
l
is also adopted here. Equating
the turbulent viscosity m
T
of the

k with that of a

kx
model, the relation C
l

kx follows, where C
l
0:09.
Then the length scale in terms of the

kx-model becomes
l
s

c
l
C
l

k
1=2
x
% 6:00

k
1=2
x
: 50
The amplitude

A is scaled by relating the solution for
R
ww
0; 0 to the amplitude

R, Eq. (14),
R
ww
0; 0
_ _

A
2
G
0
n
2
dn
4l
2
s

k
3p
:
Inserting the Gaussian kernel and neglecting the week
spatial dependence of

A, the amplitude relates to the value
of the turbulence kinetic energy via

4
3p
_

k
1=2
% 0:651

k
1=2
: 51
Note that in 2D the amplitude is independent of the
local value of l
s
.
3. Acoustic simulation techniques
3.1. Acoustic perturbation equations
For the acoustic simulations acoustic perturbation equa-
tions (APE-4) as introduced by Ewert and Schro der [19]
are used. They are a modication of the genuine linearized
Fig. 4. White noise representation at xed position and corresponding spectrum; ratio of drift time and time step Dt
drift
=Dt
CAA
10, convection velocity
u
0
1:0, 10
5
CAA time steps sampled, Dt
CAA
10
3
. (a) Random data r
ij
t at a xed probe location and (b) corresponding spectrum.
Fig. 5. Interpolation onto CAA grid points.
R. Ewert / Computers & Fluids 37 (2008) 369387 377
Euler equations (LEE). The system solved for the pressure
and velocity perturbations p
0
; u
0
is
op
0
ot
c
2
0
$ q
0
u
0
u
0
p
0
c
2
0
_ _
c
2
0
q
c
52
ou
0
ot
$u
0
u
0
$
p
0
q
0
_ _
q
m
; 53
and excludes the non-acoustic modes otherwise apparent in
the LEE. In Eqs. (52) and (53) q
0
, p
0
, and u
0
denote the
density, pressure and velocity of the time averaged ow,
respectively. Furthermore, c
0
is the local speed of sound.
By taking the curl of Eq. (53) the vorticity equation of
the APE system becomes
ox
0
ot
$ q
m
: 54
Hence, the perturbation vorticity x
0
$ u
0
on the left-
hand side is completely prescribed by the right-hand side
(RHS) source term. For the homogeneous system with all
sources removed, the vorticity equation reduces to the
statement that the vorticity remains constant (zero), i.e.,
the convective vorticity mode is not present.
The major source terms for turbulence related vortex
sound can be identied [19,20] to be the Lamb vector
q
m
x u
0
x
0
u
0
x
0
u
0
x
0
u
0

0
; 55
where . . .
0
: . . . . . . denotes the perturbation of
terms. A similar vortex source term appears in the acoustic
analogies of Powell, Howe, and Mo hring [2123]. The
source term is computed from the velocity uctuations pro-
vided by the stochastic method. The equations are inte-
grated with the DLR CAA code PIANO applying the
fourth-order DRP scheme of Tam and Webb in space
[24] and a LDDRK method [25] in time on block struc-
tured meshes.
3.2. Evaluation of the source terms via RPM
The RPM method provides the value of the streamfunc-
tion, Eq. (22) and, via fourth-order numerical dierentia-
tion, its time derivative on the CAA grid. The velocity
components are found by computing the gradients of the
streamfunction with the DRP scheme on the CAA grid.
Accordingly, time derivatives of the velocities are found
from the gradient of the time dierentiated streamfunction.
The uctuating Lamb vector equation (55) can be split
into an irrotational and a solenoidal part. Airframe noise
source mechanisms based on the interaction of vorticity
with surfaces are simulated directly with the APE, whereby
the involved vorticity is completely prescribed by the sole-
noidal part of the source term, Eq. (54). The remaining
irrotational part can be attributed to turbulence self and
shear noise, which can be deemed to be in low Mach num-
ber ow negligibly small compared to airframe noise
sources. Since the turbulent velocity uctuations equation
(7) of the RPM method completely determine the uctuat-
ing vorticity, i.e. $ u
0
x
0
, it is easy to proof the source
q
m

ou
0
ot
56
to satisfy Eq. (54). Since $ u
0
0, also $ q
m
0 holds.
That is, the source equation (56) describes the solenoidal
part of the Lamb vector equation (55) and as such neglects
turbulence self and shear noise. The source equation (56) is
used for the slat-noise computations presented in this
work.
4. Computational results
4.1. Homogeneous isotropic turbulence
As rst test-case the stochastic realization of 2D isotro-
pic turbulence with the RPM method has been considered.
Fig. 6 shows a snapshot of the x
3
vorticity component of
the uctuating pseudo-turbulent eld. The mean-ow is
uniform and runs from left to right. Using this uctuating
velocity eld the three independent components of the ten-
sor R
0
ij
r; s are processed. No temporal separation is con-
sidered, i.e., s 0.
Fig. 7 and 8 juxtapose the analytical solution Eq. (9) to
the statistically evaluated stochastic RPM data. In the g-
ures Dx and Dy denote the x and y components of the sep-
aration distance r

Dx
2
Dy
2
_
. Almost perfect
agreement is obtained. The analytical solution is based
on a longitudinal correlation f r : exp
p
4
x
2
l
2
s
_ _
, Eq.
(15), and its related lateral correlation gr f f
0
=r,
Eq. (10), whereby the lter kernel Eq. (31) is used.
Fig. 9 depicts the correlations along a horizontal cut
Dy 0. The black lines indicate the target distributions
for the longitudinal correlation function f, the 2D lateral
correlation function g and the theoretical correlation of
the vorticity x
3
ov
0
=ox ou
0
=oy. The colored triangles
represent the respective solutions from the statistical anal-
ysis of the uctuating velocities. Again almost perfect
quantitative agreement is achieved.
Fig. 6. Homogeneous isotropic turbulence (HIT); snapshot of x
3
-
vorticity.
378 R. Ewert / Computers & Fluids 37 (2008) 369387
Fig. 7. Two-point velocity correlation maps of HIT from RPM (i): (a) R
0
12
Dx; Dy analytical, Eq. (9) with f r : exp
p
4
x
2
L
2
_ _
and (b) R
0
12
from stochastic
simulation.
Fig. 8. Two-point velocity correlation maps of HIT from RPM (ii): (a) R
0
11
Dx; Dy analytical, Eq. (9) with f r : exp
p
4
x
2
L
2
_ _
and (b) R
0
11
from stochastic
simulation. (c) R
0
22
Dx; Dy analytical, Eq. (9) with f r exp
p
4
x
2
L
2
_ _
. (d) R
0
22
from stochastic simulation.
R. Ewert / Computers & Fluids 37 (2008) 369387 379
4.2. Slat-noise simulations
In this paper, the RPM/CAA method is applied for con-
venience and as a rst feasibility study not in full rigor to a
slat-noise problem. For airframe noise problems like wing
slat- or trailing edge noise, only the vorticity vector compo-
nent in spanwise direction will contribute to the sound gen-
eration. The velocity eld related to this single vorticity
component can be resolved by one scalar streamfunction.
If taken into account, anisotropy would appear through
a local factor that determines the auto-correlation of the
spanwise vorticity component with respect to the trace of
all vorticity correlations. Neglecting anisotropy tacitly
means to use a constant anisotropy factor 1/3. We believe
that anisotropy eects are of minor importance if not the
absolute sound pressure levels are of interest but rather dif-
ferences of levels resulting from design variations are stud-
ied. It is esteemed that the basic noise generating
mechanism at a slat is the interaction of vortical distur-
bances with geometrical inhomogeneities. Hence, the tem-
poral correlation (which is responsible for turbulent self-
noise generation) is neglected for the slat-noise problem
by just considering frozen turbulence with s
s
! 1. Conse-
quently, 2D CAA simulations are carried out in this paper
with a source just realized through one scalar uctuating
streamfunction. One has to bear in mind that an additional
2D to 3D correction has to be applied to the 2D acoustic
eld [19]. Except for a level o-set, which is deemed con-
stant for all ow parameter variations, the 2D approach
can be used to determine the shape of acoustic spectra,
directivities, Mach number scaling laws, and the inuence
of design parameter variations.
Fig. 10 shows the reference CAA grid used for the slat-
noise computations. It consists of 25 blocks with about
250k mesh points. It solves the acoustic eld in a 5c 5c
box, where c denotes the main-element chord length. Based
on a dimensional chord length c 0:4 m, the grid is su-
cient to resolve frequencies up to 12 kHz. The topology is
of a mixed H/C-type. An H-topology is used in the outer
region, which yields almost Cartesian cells in the far-eld.
A C-grid is used in the vicinity of the airfoil to resolve
the airfoil with the slat. The mesh resolution is enhanced
in the slat-cove to resolve the estimated length scale of
the source term properly.
The RANS mean-ow eld is computed with the DLR
ow-solver TAU on an unstructured mesh with approxi-
mately 30k elements. The free stream Mach number is con-
sidered in this paper to be Ma = 0.10. Furthermore, a
Menter SST turbulence model [26] with and without
KatoLaunder modication [27], respectively, is used.
Fig. 11 depicts the distribution of turbulence kinetic energy
in the slat-cove region. The solution with KatoLaunder
modication [27], Fig. 11a, exhibits a physically meaning-
ful solution, whereas at the main element nose the Menter
SST model yields an articial turbulence kinetic energy
spot, which is one order of magnitude larger than the
KatoLaunder solution. Since

k directly determines the
strength of the acoustic source terms via Eq. (51), the arti-
Fig. 9. Reconstruction of spatial two-point correlations from stochasti-
cally generated homogeneous isotropic turbulence; black line: target
function, colored triangles: reconstruction.
Fig. 10. CAA grid, every other grid line shown: (a) full resolved airfoil and (b) slat cove.
380 R. Ewert / Computers & Fluids 37 (2008) 369387
cial

k-spot will probably yield an overamplied acoustic
response. However, CAA computations are carried out
based on both RANS turbulence models to identify the dif-
ference that occurs in the predicted acoustic sound pressure
levels. This way the sensitivity of the turbulence kinetic
energy distribution on the sound pressure levels predicted
by the CAA/RPM model can be studied and an error esti-
mate for the expected absolute errors of the method can be
given.
Fig. 12 depicts the integral length scale distribution due
to Eq. (50) for both models with c
l
=C
l
6:0. They yield
the same magnitudes in the slat-cove region. The black cir-
cles in Fig. 12 indicate the size of the length-scale l
s
, where
the radius of the circles is equal to l
s
.
Fig. 13 shows a snapshot of the v
0
-velocity eld, gener-
ated from the RPM procedure using the RANS solutions
of

k and the length scale given by the KatoLaunder
model. A bundle of 20 streamlines is considered for the
source computation. Their extension in the computational
domain is shown in Fig. 2. On average, about 800 discrete
points along the streamlines are used, corresponding to a
drift time separation of Dt 4 10
3
. The method has a
memory overhead of about 3% compared to the memory
requirement for the baseline CAA solver without source
model. In principle it would be possible to incorporate also
the recirculation bubble in the slat-cove into the source
model by extending the model to closed loops of stream-
lines. However, the conjecture is to be tested whether the
major source of slat-noise is due to the vortical structures
that are generated in the slat-cove shear-layer and that con-
vect along the slat trailing edge. The scaling factors are
c
l
=C
l
2:0 in Fig. 13a and c
l
=C
l
6:0 in Fig. 13b, respec-
tively. Both results of the pseudo-turbulent ow eld exhi-
bit similar features: from the slat hook the ow structures
emerge, as one would expect it for instabilities growing in
a shear layer. However, note that this solution is a com-
plete reconstruction from a steady state RANS solution,
which for acoustic purposes does not have to model the
turbulent ow eld with all its details.
Fig. 14 compares the contour levels of the turbulence
kinetic energy with that reconstructed from RPM, using
the RANS reference solution to scale the lter parameters.
Fig. 11. Turbulence kinetic energy in the slat-cove, M=0.10: (a) KatoLaunder modication and (b) Menter SST baseline.
Fig. 12. Length scale in the slat-cove, c
l
=C
l
6:0, M = 0.10; the circle indicates size of the length-scale (the radii are equal to the local length-scale at the
circle origin): (a) KatoLaunder modication and (b) Menter SST baseline.
R. Ewert / Computers & Fluids 37 (2008) 369387 381
A qualitatively good agreement in the eld distribution of
the target and the stochastically induced levels is visible.
Fig. 15 highlights the eect of the number of streamlines
and the drift time separation Dt on the quality of the recon-
structed turbulence kinetic energy. Both parameters deter-
mine the number of random particles involved, Section 2.4.
Fig. 15e presents the turbulence kinetic energy levels of the
RANS target solution (Menter SST with KatoLaunder
modication). Solutions for three dierent numbers of
streamlines (10, 50, 100) as well as two dierent drift time
separations (2 10
2
, 4 10
3
) are shown, Fig. 15ad. It
is evident that the reconstructed turbulence kinetic energy
levels are unaected by the parameter variations and that
the major features of the turbulence kinetic energy distribu-
tion are recovered by the uctuation model in comparison
to the target solution, Fig. 15e.
Fig. 16 shows snapshots of the unsteady acoustic pres-
sure eld that is generated with the RPM source model.
Fig. 16a presents a solution for the turbulence model with
KatoLaunder modication, Fig. 16b for the Menter SST
baseline case. Each (sequential) computation took about
5CPUh on a NEC-SX6 computer. Both acoustic elds
exhibit the same directivities. The acoustic eld roughly
corresponds to the one expected for a dipole source placed
at the slat trailing edge with its axis normal to the slat
chord. Also the apparent wave-lengths have similar magni-
tudes. This might be due to the fact that both turbulence
models reproduce the same length-scales in the slat-region.
Fig. 16c is a magnication of the resolved domain that
highlights the near-eld structures of the perturbation
pressure in the vicinity of the slat. Fig. 16d depicts narrow
band spectra for a point 1.5c below the slat trailing edge for
both turbulence models. The spectra are related to a
dimensional frequency, which corresponds to a dimen-
sional model chord length of 0.4 m. As expected, due to
the excessive articial

k-spot at the leading edge, the Men-
ter SST baseline model clearly overpredicts the sound pres-
sure levels. The dierence in the sound pressure levels is
Fig. 13. Snapshot of the v-component of the generated stochastic eld, KatoLaunder: (a) KatoLaunder, c
l
=C
l
2:0 and (b) KatoLaunder,
c
l
=C
l
6:0.
Fig. 14. Turbulence kinetic energy in the slat-cove, M = 0.10: (a) Menter SST solution with KatoLaunder modication and (b) reconstruction from
stochastic eld.
382 R. Ewert / Computers & Fluids 37 (2008) 369387
about 13 dB, which conforms to the dierence in the

k-distributions that exceeds one order of magnitude. The


decay of the spectrum is predicted similarly by both mod-
els. The narrow band decay is conned to a range between
f
2
and f
3
, which roughly corresponds to the 12th octave
decay between f
1
and f
2
that was found by Choudhari
et al. [28] in measurements.
Fig. 17 presents narrow band spectra based on the SST
model with KatoLaunder modication for two grids with
dierent resolution and an observer position above the slat.
The reference grid and a ner grid with the grid density
doubled in each direction, which consists of approximately
1M mesh points, is used for the simulations. The computed
narrow band spectra are juxtaposed to an acoustic far-eld
measurement that has been conducted by EADS Corporate
Research Centre in DLRs Acoustic Windtunnel Braun-
schweig (AWB) as part of the German national project
FREQUENZ. A qualitative good agreement in the spectral
Fig. 15. Turbulence kinetic energy levels on the source strip; variation of the number of streamlines (s) and drift time separation along the streamline (Dt):
(a) s = 10, Dt 2 10
2
; (b) s = 10, Dt 4 10
3
; (c) s = 50, Dt 4 10
3
; (d) s = 100, Dt 4 10
3
; and (e) RANS target solution.
R. Ewert / Computers & Fluids 37 (2008) 369387 383
distributions is obtained. As expected, the coarse (refer-
ence) grid solution falls o the spectral trend at rst, reach-
ing a highest frequency of approximately 12 kHz. This
nding is a conrmation of the resolution estimate initially
made for the reference grid. The ner grid solution follows
the measured narrow band solution almost over the whole
range of meaningful frequencies up to 20 kHz.
The ne grid spectrum in Fig. 17 shows two resonance-
like bumps around 6.5 and 9 kHz, which are also present in
the coarse grid spectrum, however less amplied. Chances
are that especially the peak around 9 kHz is slightly
reduced due to the coarse grid cut-o that starts before
10 kHz. Perhaps their origin can be devoted to open cavity
resonances [29], which in principle are resolved as part of
the CAA simulations. Note that the eect on the spectrum
for a full 3D simulation depends on the coherence length
scale in spanwise direction, which could clearly dier
between the broadband airframe noise sources and those
of open cavity resonances. The peak accentuation therefore
will presumably dier in a full 3D simulation and will be
the subject of some future studies involving full 3D
CAA/RPM simulations of the slat-noise problem.
Fig. 18 compares an APE/RPM based 1/3-octave spec-
trum with the empirical slat-noise model of Dobrzynski
[1,30]. The empirical slat-noise model is based on a generic
spectrum that shows a typical peak frequency and falls o
Fig. 16. Sound pressure eld and narrow band spectra, M = 0.10: (a) with KatoLaunder modication; (b) Menter SST baseline; (c) KatoLaunder near-
eld; and (d) comparison of spectra.
Fig. 17. Comparison of APE/RPM based narrow band spectra on ne/
coarse CAA meshes with acoustic windtunnel measurement.
384 R. Ewert / Computers & Fluids 37 (2008) 369387
for higher frequencies according to a f
1:8
law. In this
model the peak frequency is a function of the slat deection
angle. The simulation reproduces these two distinct fea-
tures, i.e., the peak frequency in a similar frequency range
and the rate of decay of the spectrum for higher
frequencies.
The inuence of the length-scale parameter c
l
, Eq. (50),
on the acoustic far-eld and the turbulent near-eld spectra
is studied in Fig. 19. Acoustic simulations have been car-
ried out for two values of c
l
, i.e., a theoretically motivated
value of c
l
6:0, Section 2.5, and a clearly reduced value
of c
l
2:0. The dashed lines in Fig. 19 correspond to the
longitudinal velocity correlations induced by the model in
a representative point in the slat-cove shear layer. Accord-
ing to the underlying lter kernel, their spectral distribu-
tions follow the shape of a Gaussian. Obviously, due to
the three times smaller structures in case c
l
2:0, the
related spectrum is stretched to (three times) higher fre-
quencies. Accordingly, the characteristic spectral fallo
occurs at higher frequencies. A surprising eect is found
for the acoustic far-eld spectra, plotted with solid lines.
They exhibit a common narrow band fallo in the range
between f
2
and f
3
, with just a small absolute dierence
of about 3 dB over a large range of frequencies. Appar-
ently, the shape of the acoustic spectrum is not directly
inuenced by the shape of the turbulence spectrum and
supports a completely dierent decay law. Furthermore,
the length-scale parameter has almost no eect on the
acoustic spectra.
An explanation for this unexpected eect could be based
on the incompressible upwash velocity concept introduced
by Howe in [31]. Howe argues that the upwash velocity in the
vicinity of a at plate trailing edge, which is determined by
the BiotSavart induction formula applied to the boundary
layer vorticity outside the viscous sublayer, is the appropri-
ate metric that characterizes structure born sound genera-
tion in low Mach number ows. Since a kinematic
boundary condition holds at the surface, the surface will
respond withawall normal velocity that cancels the upwash
velocity, eventually giving rise to sound radiation. Fig. 20
sketches the upwash velocity eld v
n
induced by a vortex of
characteristic size D that passes a plate in distance h at con-
vection speed u
c
. It is evident that the characteristic fre-
quency induced by moving the upwash pattern with u
c
over
the trailing edge is determined by the miss distance h, i.e.,
f /
u
c
h
;
and not by the characteristic eddy size D.
Fig. 18. Comparison of APE/RPM based 1/3-octave spectrum with the
empirical model spectrum of Dobrzynski [1,30].
Fig. 19. Inuence of length-scale parameter c
l
, Eq. (50), on the acoustic
far-eld and the turbulent near-eld spectra.
Fig. 20. Incompressible upwash velocity concept [31]; the characteristic frequency is determined by the miss distance h as characteristic length scale, i.e.,
f / u
c
=h not the eddy size D.
R. Ewert / Computers & Fluids 37 (2008) 369387 385
Finally, Fig. 21 compares directivities for both models
for four dierent frequencies. The directivities are com-
puted for a circle with center at the slat trailing edge and
radius 1.5c. Similar shapes are obtained for both underly-
ing turbulence models. For the lowest frequency of
1 kHz, Fig. 21a, the directivity in the upper half plane
exhibits a directivity pattern, which is well known for trail-
ing edge noise [20], see also Fig. 22 that depicts analytical
solutions to the trailing edge directivity for dierent
wave-numbers. However, compared to Fig. 22 the upper
half plane pattern of Fig. 21a is mirrored at a vertical axis.
In other words, it agrees with the directivity of an edge
noise source located at the leading edge of a thin prole.
Obviously, for the small frequencies the directivity is dom-
inated by the scattering of the acoustic waves at the main
element. At the highest frequency, Fig. 21d, the directivity
shows again a typical trailing edge pattern. But in this fre-
quency range, the pattern is rotated such that the usual
upstream direction points towards the ground. The direc-
tion roughly agrees with the slat chord deection angle.
Hence, for the higher frequency range, the directivity is
mostly aected by the slat geometry.
The ndings so far can be used to estimate the potential
of the stochastic noise prediction method to accurately pre-
dict absolute sound pressure levels for airframe noise prob-
lems. The acoustic simulations evidence a dependence of
Fig. 21. Normalized directivities for the Menter SST model with/without KatoLaunder modication; the directivity circle is centered at the slat trailing
edge and has radius r = 1.5c: (a) 1 kHz; (b) 2 kHz; (c) 4 kHz; and (d) 8 kHz.
Fig. 22. Typical trailing-edge directivity of a at plate for dierent wave-
numbers.
386 R. Ewert / Computers & Fluids 37 (2008) 369387
the far-eld narrow band spectra roughly proportional to

k
p
, whereby the typical spectral distribution remains
almost unchanged, Fig. 16d. Experience expects RANS
solutions to prescribe the turbulence kinetic energy for a
wide range of technical applications within a range of say
30%. Hence, taking these values into account, the acous-
tic simulations can be expected to simulate the narrow
band spectra with an absolute level o-set around 1 dB
DSPL 20 lg

1 0:3
p
over all frequencies.
5. Conclusions
The random particle-mesh (RPM) method presented in
this paper is capable of generating a pseudo-turbulent
velocity eld, which exactly reproduces the second-order
two-point correlation tensor of homogeneous isotropic tur-
bulence in two or three dimensions. The method is time-
and memory ecient, strictly solenoidal, and can be easily
applied to the highly non-uniform mean-ow eld in the
slat-cove region using local results for the turbulence
kinetic energy and the corresponding turbulent length scale
from a RANS solution. Furthermore, it avoids the occur-
rence of shear decorrelations and resolves broadband spec-
tra continuously. A further extension to non-homogeneous
anisotropic solenoidal ows is possible with the transfor-
mation proposed by Smirnov et al. [11]. The simulations
of slat-noise with the RPM/CAA method yield directivi-
ties, which can be explained on the basis of a noise source
located at the slat trailing edge. The simulated narrow band
spectra follow over a wide range of frequencies related
measurements and the simulated 1/3-octave spectrum
exhibits the main features of an empirical slat-noise model.
Acknowledgement
The author thanks Professor Peter Ko ltzsch of the Tech-
nical University of Dresden for kindly providing the pic-
ture showing the analytical trailing edge directivities of a
at plate based on the edge Greens function of Howe.
References
[1] Dobrzynski W, Pott-Pollenske M. Slat noise studies for fareld noise
prediction. AIAA Pap. 2001-2158.
[2] Singer B, Lockard D, Brentner K. Computational aeroacoustics
analysis of slat trailing-edge ow. AIAA J 38(9).
[3] Klein M, Sadiki A, Janicka J. A digital lter based generation of
inow data for spatially developing direct numerical or large eddy
simulations. J Comp Phys 2003;186:65265.
[4] Kraichnan R. Diusion by a random velocity eld. Phys Fluids
1970;13:2231.
[5] Bechara W, Bailly C, Lafon P. Stochastic approach to noise
modelling for free turbulent ows. AIAA J 1994;32(3):45563.
[6] Bailly C, Lafon P. Computation of noise generation and propagation
for free and conned turbulent ows. AIAA Pap. 96-1732.
[7] Bailly C, Juve D. A stochastic approach to compute subsonic noise
using linearized eulers equations, AIAA Pap. 99-1872.
[8] Billson M, Eriksson L, Davidson L. Jet noise prediction using
stochastic turbulence modeling, AIAA Pap. 2003-3282.
[9] Kalitzin G, Kalitzin N, Wilde A. A factorization scheme for rans
turbulence models and sngr predictions of trailing edge noise, AIAA
Pap. 2000-1982.
[10] Ewert R, Bauer M. Towards the prediction of broadband trailing
edge noise via stochastic surface sources. AIAA Pap. 2004-2861.
[11] Smirnov A, Shi S, Celik I. Random ow generation technique for
large eddy simulations and particle dynamics modeling. J Fluids Eng
2001;123:35971.
[12] Khavaran A, Bridges J. Modelling of ne-scale turbulence mixing
noise. J Sound Vibr 2005;279:113154.
[13] Davies P, Fisher M, Barratt M. The characteristics of the turbulence
in the mixing region of a round jet. J Fluid Mech 1963;15:33767.
[14] Tam C, Auriault L. Jet mixing noise from ne-scale turbulence.
AIAA J 1999;37(2):14553.
[15] Morris P, Farassat F. Acoustic analogy and alternative theories for
jet noise prediction. AIAA J 2002;40(4):67180.
[16] Agarwal A, Morris P. Prediction method for broadband noise from
unsteady ow in a slat cove. AIAA J 2006;44(2):30110.
[17] Careta A, Sagues F, Sancho J. Stochastic generation of homogeneous
isotropic turbulence with well-dened spectra. Phys Rev E
1993;48(3):227987.
[18] Batchelor G. The theory of homogeneous turbulence. Cambridge
University Press; 1960.
[19] Ewert R, Schro der W. Acoustic perturbation equations based on ow
decomposition via source ltering. J Comput Phys 2003;188:36598.
[20] Ewert R, Schro der W. On the simulation of trailing edge noise with a
hybrid LES/APE method. J Sound Vibr 2004;270:50924.
[21] Powell A. Theory of vortex sound. J Acoust Soc Am 1964;36/
1:17795.
[22] Howe MS. Contributions to the theory of aerodynamic sound, with
application to excess jet noise and the theory of the ute. J Fluid
Mech 1975;71(4):62573.
[23] Mo hring W. Modelling low mach number noise. In: Mu ller E-A,
editor. Mechanics of sound generation in ows. Springer; 1979.
[24] Tam C, Webb J. Dispersion-relation-preserving nite dierence
schemes for computational acoustics. J Comp Phys 1993;107:26281.
[25] Hu FQ, Hussaini MY, Manthey JL. Low-dissipation and low-
dispersion RungeKutta schemes for computational acoustics.
J Comp Phys 1996;124:17791.
[26] Menter F. Zonal two equation kx turbulence models for aerody-
namic ows. AIAA Pap. 93-2906.
[27] Kato M, Launder B. The modelling of turbulent ow around
stationary and vibrating square cylinders, in: Proceedings of the ninth
symposium on turbulent shear ows, Kyoto, Japan, August 1618th,
1993; 1993.
[28] Choudhari M, Khorrami M, Lockard D, Atkins H. Slat cove noise
modeling: a posteriori analysis of unsteady rans simulations, AIAA
Pap. 2002-2468.
[29] Koch W. Acoustic resonances in rectangular open cavities, AIAA
Paper 2001-2134.
[30] Pott-Pollenske M, Dobrzynski W, Buchholz H, Gehlhar B, Walle F.
Validation of a semiempirical airframe noise prediction method
through dedicated a319 yover noise measurements, AIAA Pap.
2002-2470.
[31] Howe M. Trailing edge noise at low mach numbers. J Sound Vibr
1999;225(2):21138.
R. Ewert / Computers & Fluids 37 (2008) 369387 387

You might also like