Center for Turbulence Research 17

Annual Research Briefs 2009

Post-processing of large-eddy simulations for jet

noise predictions
By S. Mendez, M. Shoeybi, A. Sharma, S. K. Lele AND P. Moin

1. Motivation and objectives

Despite almost 60 years of active research on jet noise, the noise generated by expulsion
of hot gases at the exhaust of aircraft engines remains an important part of the total
aircraft noise. This is particularly true in the case of supersonic aircrafts, where the
high speed of the exhaust gases yields extremely high noise. In the framework of the
development of future civil supersonic aircrafts, it is of major importance to decrease the
noise from supersonic jets in order to comply with strict noise regulations enforced in
civil air transportation.
To assist the experimental efforts in finding innovative solutions for jet noise reduction,
numerical methods have been developed. Among them, Large-Eddy Simulation (LES) has
the potential to become a tool of choice to perform predictions of the noise generated by
turbulent jets (Bodony & Lele 2008). It does not have the same limitation in Reynolds
number as Direct Numerical Simulation and thus can handle cases relevant to industrial
applications. From the physical point of view, LES is well adapted to jet noise compu-
tations, because important contribution to jet noise comes from the large scales of the
turbulent flow (Tam 1995, 1998; Bogey & Bailly 2006), for which Reynolds-Averaged
Navier-Stokes methods fail to provide a good description in the general case.
One of the outputs of interest in jet noise study is the far-field noise. While LES is
well suited for the computation of the flow, it would be inefficient to use it to propagate
acoustic waves to the far field. In order to compute the far-field noise, one relies on
acoustic analogies: volumetric methods such as Lighthill analogy (Lighthill 1952; Freund
2001; Uzun et al. 2004) allow modeling of noise sources and computation of the far-
field sound. Surface integral methods like Kirchhoff (Farassat & Myers 1988) or Ffowcs-
Williams & Hawkings (Ffowcs Williams & Hawkings 1969) methods (with volumetric
source terms, corresponding to the presence of quadrupoles outside the surface, neglected)
rely on near-field information gathered over a surface enclosing as much as possible the
noise sources. The latter methods, owing to their limited cost and their success (coupled
with LES) in predicting the noise emitted by high-speed jets, are now very popular in
the jet noise LES community.
However, the diversity of results presented in the literature suggests that details of
implementation are important, and that a thorough study of the capacity of such methods
has to be undertaken, as done by Rahier et al. (2003) and Shur, Spalart & Strelets (2005),
among others, before using it routinely. To post-process the jet noise results obtained by
the compressible version of the unstructured LES solver CDP (Shoeybi et al. 2009), a
solver using the Ffowcs-Williams & Hawkings (FWH) surface integral method has been
implemented. The purpose of this report is to document and analyze far-field acoustic
results obtained from a set of LES to determine best practices when using a FWH solver
for jet noise predictions.
18 S. Mendez et al.
2. Numerical method
In the present study, we use the frequency domain permeable surface FWH formulation
(Ffowcs Williams & Hawkings 1969), already described by Ham et al. (2009). In this
formulation, the volume integral of the original FWH equation is omitted, which generates
errors. One of the challenges is to minimize the errors due to the use of the inexact FWH
formulation.
This section describes the procedure employed to calculate the far-field sound. The
time history of conservative variables is saved over a given surface S (referred to as FWH
surface) at a specified sampling frequency f and for a total time τ . f is associated with
the Nyquist Strouhal number Stmax = f D/2Uj , where D = 2R is the nozzle diameter
at the exit (R the radius) and Uj is the jet velocity at the nozzle exit. τ determines the
minimum frequency accessible by this post-processing Stmin = D/τ Uj .
For each surface element of S, the time history of source terms F1 and F2 are con-
structed from the conservative variables using the following expressions:
p′ n̂j r̂j + ρuj un r̂j ρun p′ n̂j r̂j + ρuj un r̂j
F1 = + and F2 = . (2.1)
c0 r r r2
n̂j is the jth component of the unit surface normal vector, and rr̂j represents magnitude
and direction of the vector from the surface element location y to the observer location x.
In the expression of F1 and F2 , p′ is the fluctuating pressure (p′ = p−p∞ , where subscript
∞ refers to the ambient value), uj is the jth component of the velocity vector, and
un = ui ni . Note that the effect of viscous stresses has been neglected. Two formulations
are used in this report (section 3.4): in the original formulation (Ffowcs Williams &
Hawkings 1969), ρ is the density. A second formulation based on pressure is used (Spalart
& Shur 2009). In the absence of volume integral, the only difference with the original
formulation is that ρ = ρ∞ + p′ /c∞ 2 , where ρ∞ is the ambient density.
F1 and F2 and then windowed (see section 3.2) after subtracting the mean and time-
Fourier-transformed. The time derivative of F1 is calculated in the frequency space. The
retarded time (exp(−iωr/c∞ )) is also applied in the frequency space. The integral of the
source terms over the surface then yields the Fourier transform R ∞ of the pressure at the
observer location (with the Fourier transform of g, ĝ(ω) = −∞ g(t)e−iωt dt):
Z Z
4π p̂(x, ω) = iω F̂1 (y, ω)exp(−iωr/c∞ )dy + F̂2 (y, ω)exp(−iωr/c∞ )dy. (2.2)
S S

The Narrowband Sound Pressure Level (SP L) level (in dB) is calculated as:
!
2 p̂(x, ω)p̂∗ (x, ω)
SP L(x, St) = 10 log10 , (2.3)
Stmin p2ref

where p̂∗ is the complex conjugate of p̂, pref = 20µP a and St = ω D/2π Uj .
The Overall Sound Pressure Level (OASP L, in dB) is calculated as

St
!
max
X 2 p̂(x, ω)p̂∗ (x, ω)
OASP L(x) = 10 log10 (2.4)
p2ref
St=0
The two FWH surfaces used below are shown in Fig. 9 (section 3.6). The surfaces
more or less follow the growth of the jet, their section increasing downstream. A vertical
outflow disk is located downstream, at x = 25 D in Fig. 9. For x < 0, the surface follows
the shape of the nozzle. Technically, all surfaces used are then open at the inflow, but
 Post-processing of LES for jet noise predictions 19
this has no consequence on the calculated sound. Note that in the following, we use the
terms closed surface and open surface to refer to the fact that the outflow disk is included
or excluded from the FWH surface (see section 3.5).
In terms of outflow disk, a third option is considered: averaging over outflow disks. This
technique was first used by Shur et al. (2005). It consists in computing p̂ using Eq. 2.2
for several surfaces having the exact same shape, but with the outflow disks located at
different axial locations. Results in p̂ from the different surfaces are then averaged. The
spurious noise generated by the passage of turbulent eddies through the outflow disk is
not consistent from one surface to another. It is thus partially or totally cancelled when
doing the averaging (section 3.5). The outflow disks will be regularly spaced. The spacing
between two consecutive outflow disks is denoted by ∆min , and the distance between the
first and the last outflow disks is denoted by ∆max . When outflow disk averaging is used,
the first outflow disk is always located at x = 25 D, so that the set of surfaces over which
averaging is performed is entirely described by ∆min and ∆max .

3. Results
As the focus is not on the LES themselves, only a short presentation is given. LES are
performed using a compressible version of the unstructured LES solver CDP, CDP-C,
developed by Shoeybi et al. (2009). CDP-C is a second-order finite volume solver, with
a hybrid implicit-explicit time advancement scheme. Small-scale turbulence is modeled
using the dynamic modeling procedure of Moin et al. (1991) with Lilly’s modification
(Lilly 1992). The artificial bulk viscosity method (see Cook & Cabot 2005) in a generalized
form is used to capture shock waves on unstructured grids in CDP-C code. This model
has further been improved by Mani, Larsson & Moin (2009) to minimize the effect of
artificial bulk viscosity on turbulence as well as dilatational motions. In this study, the
model has been adapted for unstructured grids. More details about the solver and the
LES themselves are given by Shoeybi et al. (2009) and Mendez et al. (2010).
Four different large-eddy simulations will be used in this report (Table 1). Simulation
S1 is an unheated perfectly expanded jet at Mach 1.4. The operating point of simulations
S2, S3 and S4 is identical: a perfectly expanded heated jet is considered. The difference
between the three simulations is the grid. A direct comparison will be shown in this
report between S2 and S3. S4 is mainly used because different time sampling was used
for this calculation (see section 3.1). For more comparisons between the three grids, the
reader is referred to Mendez et al. (2010).
In the following, multiple options of post-processing will be tested. However, not all the
used options are reminded for each test. Unless otherwise specified, the default options are
the following: FWH acoustic calculations are performed using the time record specified in
Table 1, and time records are windowed using the Hanning window. The sound at 100 D
from the nozzle exit is presented, but by default it is calculated with observers located
at 50 D from the nozzle exit and rescaled using the assumption that the amplitude of
acoustic waves decays like the inverse of the distance from the nozzle exit (as in the
experiment). The pressure-based formulation of the FWH equation is used, from data
recorded over surface s01 (see Fig. 9) closed at x = 25 D. Results are compared with
experimental data provided by Dr Bridges (Bridges & Wernet 2008). Angles are defined
from the jet axis, pointing upstream (maximum noise is observed around 150◦ ).
20 S. Mendez et al.

Simulation Mj Ma M∞ TR Rej Mesh size ∆tc∞ /R Stmin Stmax

S1 1.4 1.4 0.008 1.0 150,000 17 × 106 0.005 0.0040 7.2

S2 1.4 1.86 0.008 1.765 76,600 17 × 106 0.005 0.0030 5.4

S3 1.4 1.86 0.008 1.765 76,600 20 × 106 0.005 0.006 5.4

S4 1.4 1.86 0.008 1.765 76,600 28 × 106 0.005 0.0031 5.4

Table 1. Operating conditions and numerical characteristics of the simulations performed.

Subscripts j and ∞ refer to the jet at the nozzle exit and the ambient free-stream, respec-
tively. Notations are the following: Mj = Uj /cj , Ma = Uj /c∞ , M∞ = U∞ /c∞ , T R = Tj /T∞ ,
Rej = ρj Uj D / µj (µj is the dynamic viscosity at the nozzle exit), and ∆tc∞ /R is the time
step. Stmin and Stmax are the default values of Strouhal number used in the plots displayed.

120 140
(a) (b)
SP L at 150◦ (in dB)
SP L at 60◦ (in dB)

110 130

120
100

110
90
100
80
90
2 4 6 2 4 6 2 4 6 2 4 6 2 4 6 2 4 6
0.01 0.1 1 10 0.01 0.1 1 10
St St
◦ ◦
Figure 1. Sound pressure level at 60 (a) and 150 (b) for simulation S4 calculated with
three different time records: 0.0062 < St < 10.8 ( ), 0.0031 < St < 5.4 ( ), and
0.0124 < St < 5.4 ( ). Experimental results are also displayed ( ).

3.1. Time sampling and length of time record

First, the influence of the time record used is shown. Two parameters are considered, the
time sampling and the length of the time record, determining the maximum and minimum
Strouhal numbers available for acoustic post-processing. Sound spectra for case S4 are
shown in Fig. 1. Concerning the time sampling, it is shown that aliasing is visible at
upstream angles when Stmax = 5.4 (thin and thick solid lines in Fig. 1). When sampling
frequency is doubled, results differ at high frequencies (Fig.1a). This is not observed for
downstream angles. In the calculations presented here, aliasing is visible for upstream
angles only, and for frequencies that do not seem well resolved by the current numerical
setup. The time sampling is then considered sufficient and does not alter meaningful
results.
The length of the time record is also varied in Fig. 1 to test the convergence at low
frequencies. It is shown that spectra depart from a more converged spectrum if St <
10 Stmin : at a given St, results are converged if 10 periods have been accumulated.
 Post-processing of LES for jet noise predictions 21
140
(a) (b)
SP L at 120◦ (in dB)

SP L at 150◦ (in dB)

120
130

110 120

100 110

100
90
90
2 4 6 2 4 6 2 4 6 2 4 6 2 4 6 2 4 6
0.01 0.1 1 10 0.01 0.1 1 10
St St
Figure 2. Sound pressure level at 120◦ (a) and 150◦ (b) for simulation S2 calculated with three
different windowing procedures: no windowing ( ), hyperbolic tangent windowing ( ),
and Hanning windowing ( ). Experimental results are also displayed ( ).

3.2. Windowing
To avoid spectral leakage, signals are windowed before post-processing. Figure 2 compares
results without windowing, results using a hyperbolic tangent (HT) type window, or using
a Hanning window. Window functions using hyperbolic tangent have been used before to
maximize the time record effectively taken into account (Freund 2001; Ham et al. 2009).
Sound spectra presented in Fig. 2 show that windowing is necessary. HT and Hanning
windows give very similar results, except at low frequencies, where Hanning window gives
slightly lower sound. However, the range of frequencies where HT and Hanning windowed
results differ is the range where results are not fully converged. In the following, Hanning
window is systematically used.

3.3. Observer distance

In the formulation of the FWH equation employed here, no far-field assumption has been
made. Acoustic results thus depend on the distance of the observer to the nozzle exit.
This dependance is addressed by computing the acoustic far-field at different observer
distances and rescaling them using the r−1 decay of acoustic waves to obtain the acoustic
far-field at 100 D. Note that experimental results are obtained by measuring the acoustic
field at 50 D and projecting it to 100 D.
Figure 3 shows the acoustic far-field at 100 D. Overall sound pressure level is presented
rescaling results calculated at 50 D (as in the experiment), 100 D, and 125 D. It can be
seen that 50 D is not sufficient to be in the region where acoustic waves decay as r−1 . On
the contrary, results calculated at 100 D and 125 D are very close. A consequence of this
figure is that results will be calculated at 50 D and rescaled to match the experimental
procedure. This also shows that a far-field assumption can change the results.

3.4. Formulation of the FWH equation

Spalart & Shur (2009) have shown that the original FWH formulation is inferior to an-
other formulation using pressure instead of density fluctuations (see section 2). They ar-
gue that the density formulation is not well suited to situations where important entropy
fluctuations cross the FWH surface. In that case, density fluctuations ρ′ significantly
differ from their acoustic estimation from pressure fluctuations: p′ /c2 . Spalart & Shur
(2009) show that volumetric terms (quadrupoles) in the FWH equations are more com-
pact when computed in the pressure formulation. As a consequence, the error made by
neglecting their effect outside the FWH surface is lower in the pressure formulation. Us-
22 S. Mendez et al.

OASP L at 100 D (in dB)

130

125

120

115

110

40 80 120 160
Angle (◦ )
Figure 3. Overall sound pressure level for simulation S2 calculated at three different observer
distances, 50 D ( ), 100 D ( ), and 125 D ( ), and rescaled at 100 D. Experimental
results are also displayed ( ). They are calculated at 50 D and rescaled.

120 120
(a) (b)
SPL at 120◦ (in dB)
SPL at 90◦ (in dB)

110 110

100
100

90
90
80
80
2 4 6 2 4 6 2 4 6 2 4 6 2 4 6 2 4 6
0.01 0.1 1 10 0.01 0.1 1 10
St St
Figure 4. Sound pressure level at 90◦ (a) and 120◦ (b) for simulation S1 (unheated) calculated
with the density formulation ( ) and the pressure formulation ( ). The outflow disk of
the FWH surface is located 27.5 D from the nozzle exit. Experimental results are also displayed
( ).

ing a pressure-based formulation is thus expected to improve results, particularly for hot
jets. As observations on FWH results differ depending on authors, tests were performed
to verify the conclusions from Spalart & Shur (2009).
Density and pressure formulations were first compared in the case of an unheated jet,
the outflow disk of the FWH surface being included (Fig. 4). Sound spectra were almost
identical, consistently with expected low entropy fluctuations.
The same observation was made for heated jets when the outflow disk was excluded
from the computation (not shown). Vortical motions only rarely cross the surface, as it is
located apart from the jet and open at the downstream end. Differences between ρ′ and
p′ /c2 are small, so results from the density and the pressure formulations were almost
identical.
Figure 5 shows sound spectra for the heated simulation S2, calculated using a FWH
surface with the outflow disk located 25 D downstream of the nozzle exit. In this case,
improvement from the use of the pressure form of the FWH equation is substantial.
The over-prediction at low frequencies is significantly reduced, especially at downstream
angles. Improvement is probably limited to low frequencies due to the grid stretching in
the axial direction: the grid cannot sustain high-wavenumber waves, thus limiting the
frequency content at the end of the FWH surface to low frequencies.
In conclusion, based on our experience, the pressure-based formulation is better than
 Post-processing of LES for jet noise predictions 23
120
(a) (b)

SP L at 120◦ (in dB)

SP L at 90◦ (in dB)

120
110

110
100

100
90

80 90

2 4 6 2 4 6 2 4 6 2 4 6 2 4 6 2 4 6
0.01 0.1 1 10 0.01 0.1 1 10
St St
Figure 5. Sound pressure level at 90◦ (a) and 120◦ (b) for simulation S2 calculated with the
density formulation ( ) and the pressure formulation ( ). The outflow disk of the FWH
surface is located 25 D downstream of the nozzle exit. Experimental results are also displayed
( ).

the original density formulation. All the results obtained in this study are consistent with
the results shown by Spalart & Shur (2009).

3.5. Outflow disk

All studies using the FWH equation to compute the sound generated by high-speed
jets have debated the question of the outflow disk. Articles often concluded that open
surfaces yield better results than closed surfaces (among them Rahier et al. 2003; Uzun
et al. 2004; Eastwood et al. 2009). They stated that vortical motions crossing the surface
create strong unphysical noise. On the contrary, Shur et al. (2005) showed results where
the open surface provides highly inaccurate predictions at low frequencies. Moreover, they
argued that using open surfaces leads to a loss of information at shallow angles, close
to the downstream axis. From the results they obtained, closed surfaces are superior to
open surfaces. Note that Shur et al. (2005) also used ‘outflow disk averaging’, explained
in section 2. Results were substantially improved, in particular at upstream angles, where
the amplitude of the physical noise is not sufficient to mask the spurious noise related to
the outflow disk. This summary shows that the question of the outflow disk is complex
and obviously depends on the implementation, as results differ depending on authors. As
a consequence, it is essential to test the different options with the implementation chosen
here.
Note first that results using closed surfaces with the outflow disk at different distances
from the nozzle exit (25 D, 27.5 D, 30 D, 32.5 D) have been compared, without observing
major differences. Turbulent eddies do cross the outflow disks of all these surfaces, but
the spurious noise is similar for all surfaces. Of course, this conclusion would be different
if the outflow disks were too close to the nozzle exit.
Figure 6 compares the sound spectra obtained from simulation S1 with a closed surface
with the outflow disk at x = 30 D, an open surface (the same without outflow disk), and
averaging the results from 11 surfaces with their outflow disks located between x = 25 D
and x = 30 D (∆max = 5.0 D) and spaced by ∆min = 0.5 D. Note first that the effect of
the closure is limited to low frequencies (St < 0.3), which depends directly on the grid
resolution at the outflow disk location. For both angles considered, results using outflow
disk averaging are more accurate. In particular, the experimental shape is recovered.
Closed surface results are much noisier at low frequencies for the station located at 60◦
(Fig. 6a), while they are close to the averaged results at 120◦ (Fig. 6b). At 60◦ (Fig. 6a),
24 S. Mendez et al.
120
(a) (b)

SP L at 120◦ (in dB)

120
SP L at 60◦ (in dB)

110
110
100
100
90
90
80
80
2 4 6 2 4 6 2 4 6 2 4 6 2 4 6 2 4 6
0.01 0.1 1 10 0.01 0.1 1 10
St St
Figure 6. Sound pressure level at 60◦ (a) and 120◦ (b) for simulation S1 (unheated) changing the
outflow disk closure: surface with outflow disk at x = 30 D ( ), same surface without outflow
disk ( ), and results averaged using 11 surfaces with ∆min = 0.5 D and ∆max = 5.0 D
( ). Experimental results are also displayed ( ).

spurious results are obtained for St < 0.3. Results using the open surface depart from
the averaged results at lower frequency (St < 0.1). However, the errors at low frequencies
are impressive. Note that exactly the same types of errors are shown for open surfaces
by Shur et al. (2005), who identified them as pseudo-sound generated by the passage of
vortices near the downstream end of the open FWH surface.
The same test has been performed on the heated case S2 with the original formulation
(not shown), with exactly the same observations. However, as discussed previously in
section 3.4, the formulation has no influence when using open surfaces, whereas the
spurious sound obtained using closed surfaces with the original formulation is higher than
with the pressure formulation. Note also that outflow disk averaging is not sufficient to
completely correct these errors.
Finally, the same comparison is shown for simulation S4 using the pressure formulation,
at four different angles with, again, the same conclusions (Fig. 7). Note also that among
the four angles displayed in Fig. 7, only the upstream angle predictions (Fig. 7a) are
really inaccurate using a closed surface. On the contrary, open surfaces show spurious
artifacts at all angles.
From the results shown here, one can propose an explanation for the differences in the
conclusions of previous studies regarding the effect of the outflow disk treatment:
(a) Using a pressure-based formulation reduces the spurious noise caused by closed
surfaces. On the contrary, the formulation has no effect on open surfaces. Authors choos-
ing to use open surfaces calculate the far-field sound with the original density-based
formulation,
(b) Shur et al. (2005) show comparisons of open and closed surfaces at sideline and
downstream angles, where spurious effect of closed surfaces are less visible than at up-
stream angles,
(c) In the present study, the spurious effect of closed surfaces is limited to low fre-
quencies, because of the grid stretching. It appears that studies in favor of open surfaces
generally use an axial grid stretching that is either less aggressive than in the present
study and in Shur et al. (2005) or no stretching at all (see for example Rahier et al. 2003;
Uzun et al. 2004). In the latter cases, outflow disk errors contaminate all the spectra and
are much more noticeable.
However, it is not clear why all studies do not show important errors resulting from
pseudo-sound at low frequencies using open surfaces. Nevertheless, best results are ob-
 Post-processing of LES for jet noise predictions 25
120 120
(a) (b)

SP L at 90◦ (in dB)

SP L at 60◦ (in dB)

110 110

100 100

90 90

80 80

2 4 6 2 4 6 2 4 6 2 4 6 2 4 6 2 4 6
0.01 0.1 1 10 0.01 0.1 1 10
St St
140
(c) (d)
SP L at 120◦ (in dB)

SP L at 150◦ (in dB)

120
130

110 120

100 110

100
90
90
2 4 6 2 4 6 2 4 6 2 4 6 2 4 6 2 4 6
0.01 0.1 1 10 0.01 0.1 1 10
St St
Figure 7. Sound pressure level at 60◦ (a), 90◦ (b), 120◦ (c) and 150◦ (d) for simulation S4
(heated) changing the outflow disk closure, using the FWH pressure formulation: surface with
outflow disk at x = 25 D ( ), same surface without outflow disk ( ) and results averaged
using 11 surfaces with ∆min = 0.5 D and ∆max = 5.0 D ( ). Experimental results are also
displayed ( ).

tained using outflow disk averaging. However, additional parameters are involved when
outflow disk averaging is used. The following figures aim at clarifying how to chose the
number of surfaces for averaging, and the distance between their outflow disks.
A simple reasoning is used to estimate Sts , the Strouhal number for which outflow
disk averaging has maximum effect. Let us consider that the passage of a vortex at time
t through an outflow disk is seen by an observer as a spurious acoustic wave that we
consider sinusoidal, of period ts . If this vortex is frozen and convected at the convection
speed Uc , it reaches a second outflow disk with the time delay ∆/Uc , where ∆ is the
streamwise distance between the outflow disks of the two surfaces. When reaching the
second outflow disk, the frozen vortex will produce the same sinusoidal spurious wave.
Using a far-field hypothesis, the two spurious signals are seen by a far-field observer
shifted by the same time delay, ∆/Uc . In this case, pure cancellation of the two signals
occurs if ∆/Uc = ts /2 or Sts = Uc D/2Uj ∆. The relevance of this expression is addressed
in the following.
Figure 8 shows how sound spectra obtained by outflow disk averaging vary by chang-
ing the distance between two consecutive outflow disks ∆min (Fig. 8a) or the distance
between the first and the last outflow disks ∆max (Fig. 8b). An upstream angle (60◦ )
is shown, where the effect of averaging is maximum. In Fig. 8(a), outflow disks are
separated by ∆min = 0.5 D, 1.0 D, and 2.0 D. The first two cases are compared: aver-
aging with ∆min = 0.5 D or ∆min = 1.0 D gives very close results, except in the range
0.15 < St < 0.3. If time-averaged velocity is considered as a good estimate for Uc , then
Uc ≈ 0.3Uj in S1 in the region where outflow disks are located. With ∆min = 0.5 D,
26 S. Mendez et al.

110 (a) 110 (b)

SP L at 60◦ (in dB)

SP L at 60◦ (in dB)

105 105

100 100

95 95

90 90
2 3 4 5 6 2 3 4 5 6 2 3 4 5 6 2 3 4 5 6
0.01 0.1 1 0.01 0.1 1
St St
Figure 8. Sound pressure level at 60◦ for simulation S1 (unheated) calculated using outflow
disk averaging. (a) ∆max = 6.0 D and ∆min = 0.5 D ( ), ∆min = 1.0 D ( ) and
∆min = 2.0 D ( ). (b) ∆min = 0.5 D and ∆max = 2.5 D ( ), ∆max = 5.0 D ( )
and ∆max = 7.5 D ( ). Results from a closed surface with the outflow disk at x = 25 D
( ) is shown for a. and b. Experimental results are also displayed ( ).

Figure 9. Vorticity magnitude (non-dimensionalized by c∞ /R) in an instantaneous solution

from case S2. Surfaces s01 (loose) and s02 (tight) are shown. The nozzle (not shown) is located
between x = −7.55 D and x = 0.

the maximum effect for outflow disk averaging should be seen for Sts = 0.3. The simple
formula used above gives a good approximation, although high, of the range of maximum
effect of averaging. This is also confirmed by comparing the cases where ∆min = 1.0 D and
∆min = 2.0 D. Maximum efficiency for ∆min = 1.0 D should be seen around Sts = 0.15,
and the two cases indeed differ in the range 0.06 < St < 0.15.
Figure 8(b) shows how results obtained by averaging vary by changing ∆max . When
using outflow disk averaging, ∆max appears to determine the low-frequency limit of
effectiveness of averaging. Spurious effects are moved down to lower and lower frequencies
as ∆max increases. Three cases are compared: ∆max = 2.5 D (Sts = 0.06), ∆max = 5.0 D
(Sts = 0.03), and ∆max = 7.5 D (Sts = 0.02). Again, the relevance of the formula to
evaluate Sts is observed. It is shown in Fig. 8(b) that averaging over surfaces with
∆max = 5.0 D (resp. 7.5 D) mainly reduces the spurious sound around St = 0.03 (resp.
St = 0.02) compared to the case where ∆max = 2.5 D (resp. 5.0 D).

3.6. Surface location

The question of the surface location is of particular importance when stretched grids are
used. Figure 9 shows the location of surfaces s01 and s02, used in Fig. 10, together with
an instantaneous field of vorticity magnitude (case S2). Both surfaces (but especially s02)
are placed very close to the jet and are regularly crossed by vortices. Surfaces s01 and
s02 are closed, with their outflow disk located at x = 25 D.
Figure 10 shows how sound predictions can be modified by the surface location (solid
 Post-processing of LES for jet noise predictions 27
120 120
(a) (b)

SP L at 90◦ (in dB)

SP L at 60◦ (in dB)

110 110

100 100

90 90

80 80

2 4 6 2 4 6 2 4 6 2 4 6 2 4 6 2 4 6
0.01 0.1 1 10 0.01 0.1 1 10
St St
140
(c) (d)
SP L at 120◦ (in dB)

SP L at 160◦ (in dB)

120
130

110 120

100 110

100
90
90
2 4 6 2 4 6 2 4 6 2 4 6 2 4 6 2 4 6
0.01 0.1 1 10 0.01 0.1 1 10
St St
Figure 10. Sound pressure level, averaged over 1/3rd octave, at 60◦ (a), 90◦ (b), 120◦ (c), and
160◦ (d) calculated for simulation S3 using surface s01 (see Fig. 9) ( ) and for simulation
S2 using s01 ( ) and s02 ( ), compared to experimental measurements ( ).

lines). To be able to distinguish the results accurately, SPL have been bin-averaged
over 1/3rd octave for this figure. A third set of numerical results (discussed below) is
also shown (dashed lines). Several observations can be made on the comparison of the
noise results using either s01 or s02 to post-process simulation S2. Note first that at
low frequencies, the sound is lower using s02, because of a too small radial extension of
the surface for large distances from the nozzle: part of the sound at low frequencies is
generated outside s02. This sound is then missed by the FWH calculation, which omits
the volume integration.
Another unwanted consequence of using s02 can be observed in Fig. 10(d). At this
shallow angle, the sound predicted using s02 is generally 3 dB higher than using s01.
This is obviously spurious, as the difference in the grid refinement cannot explain this
over-estimation. Surface s02 actually crosses the source region too much, as shown in
Fig. 11. The root mean square of the quadrupole term of the FWH equation (integrand
∂ 2 Tij
of the volume integration) is displayed: ∂xi ∂x j
, where Tij = ρui uj + (p′ − c∞ 2 ρ′ )δij . Note
that, as classically done, viscous stresses are ignored in this expression of Tij . Ideally,
one would use a FWH formulation without volume integral only if this term is zero
outside the surface. From Fig. 11, it can be seen that surface s02 cuts the source zone
quite severely in the region 1 D < x < 2 D. In additional tests (not shown), shifting the
surface slightly away from the jet in this zone has significantly decreased the spurious
noise at shallow downstream angles.
A striking difference between surfaces s01 and s02 is for high frequencies. For all angles
considered, using s02 leads to higher sound at high frequencies, recovering values closer
to experimental ones. Using a tight surface also decreases the spurious trend at high
28 S. Mendez et al.

Figure 11. Cutting plane z = 0 showing the root mean square of the quadrupole term in
the FWH equation (double divergence of the Lighthill tensor Tij ), non-dimensionalized by
ρ∞ c∞ 2 /R2 . Results are azimuthally averaged except in the core, where the grid is unstruc-
tured.

-30 -20
(a) (b)
-40
P SD(p′ ) (in dB)

P SD(p′ ) (in dB) -40

-50
-60
-60
-80
-70

-80 -100

-90 -120
2 4 6 8 2 4 6 8 2 4 6 2 4 6 8 2 4 6 8 2 4 6
0.01 0.1 1 0.01 0.1 1
St St
Figure 12. Power spectral density (PSD) of pressure fluctuations, in dB (the reference pressure
is p∞ ) at x = 5 D (a) and x = 15 D (b), on the surfaces used to calculate the sound in Fig. 10:
simulation S3, surface s01 ( ) and simulation S2, surfaces s01 ( ) and s02 ( ).
Experimental results are also displayed ( ).

frequency, which results in a violent drop in the intensity of the noise, preceded by a
small bump, as observed for example in Fig. 10(b) for St ≈ 1.5 using surface s01. This
is directly related to the grid resolution. The grid along surface s02 is finer than along
s01. Surface s02 thus supports higher wavelengths. This is confirmed by comparing power
spectral density of pressure fluctuations along the two surfaces (Fig. 12). At x = 5 D and
x = 15 D, pressure from the simulation is extracted at the location of surfaces s01 and
s02 and spectra are displayed. Figure. 12(a) has to be compared with the view of the
grid used in S2 (Fig. 13a). As the grid is stretched in the radial direction, the grid cutoff
wavelength is smaller for s01 than for s02, resulting in a drop in turbulent fluctuations
for smaller values of the Strouhal number (Fig. 12a). This has direct consequences on
the far-field noise.
This shows the importance of the grid resolution in the high Mach number flows
considered here. In order to make the sound calculation less sensitive to the surface
location, a new grid has been used for simulation S3. The difference between cases S2
and S3 lies only in the radial grid refinement in the intermediate region between the
main noise sources and the surface location (see Fig. 13). By improving the radial grid
resolution for acoustic propagation, one can increase the grid cutoff wavelength at the
location of surface s01(Fig. 12) and recover better sound predictions without moving the
surface, as shown in Fig. 10.
Note that improvement in sound predictions thanks to better radial resolution is not
the same for all angles. At upstream angles (Fig. 10a), the refinement is obviously not
 Post-processing of LES for jet noise predictions 29

Figure 13. Cutting plane z = 0 showing the grid and vorticity magnitude (non-dimensionalized
by c∞ /R) near the nozzle exit in instantaneous solutions from cases S2 (left) and S3 (right).
Surfaces s01 (loose) and s02 (tight) are shown, as in Fig. 9.

Figure 14. Schematic of the influence of the surface location on the sound calculated in the
far-field. The solid line represents the optimal surface. The dotted-line surface cuts the jet
severely while the dashed line displays a surface too far from the jet, in a zone where the grid
is coarsened.

sufficient to improve significantly the results. On the contrary, predictions at 90◦ and
120◦ are significantly better at high frequencies for case S3 than for case S2, using the
same surface s01. Results remain unchanged at shallow downstream angles, where no
drop is observed for case S2.
As a conclusion on the surface location, it is obvious that the surface should be located
as close as possible from the jet when stretched grids are used. However, defining a simple
and universal criterion is difficult. In order to make the results less sensitive to the surface
location, the grid has to be designed with particular care in the region where the acoustic
waves propagate before reaching the FWH surface.
Even if the location where the surface should be placed is difficult to determine, trends
in terms of the change in the results with the surface location can be easily defined (see
Fig. 14). A hypothetical optimal surface location is considered. Using a surface away
from this location will make the sound at high frequencies drop, the cutoff frequency
associated with the grid being lower away from the jet, caused by grid stretching. For
frequencies just lower than the cutoff frequency, increase of the noise will be seen: under-
resolved fluctuations are artificially dispersed to lower frequencies. Using a radially too
tight surface will produce two types of spurious effects:
(a) The surface is too tight far from the nozzle exit. A loss of information can occur,
observed in particular at low frequencies. Far from the nozzle exit, sound sources associ-
ated with low frequencies have a large radial extension. Sound deficit at low frequencies
caused by too tight surface has been observed for all angles (Fig. 10).
(b) The surface is too tight near the nozzle exit and crosses the zone of intense sources.
In our calculations, this spurious effect has been seen for very shallow downstream angles
and for a wide range of frequencies (Fig. 10(d) when surface s02 is used). It contaminates
other angles when the surface crosses the jet even more aggressively. An interesting
30 S. Mendez et al.
feature is that the radial gradient of noise sources is high in the first few diameters of
the jet. This permits having a surface very close to the jet axis without crossing the jet.

4. Conclusions and future work

This work focused on the post-processing of LES of supersonic jets for the extraction
of acoustic far field. The FWH integration, with omission of the volume integral, has
been used for this purpose. It should allow for optimal use of the FWH solver developed
for the jet noise project for the future calculations.
A set of best practices can be deduced from this study:
(a) The pressure-based formulation is better than the original density-based formula-
tion for FWH predictions with omission of the volume integral. Improvement is observed
in terms of spurious noise through the use of FWH surfaces cutting the noise source
region (typically at the outflow disk).
(b) In conjunction with the FWH equation based on pressure, closed surfaces give
better results than open surfaces. However, spurious noise is still visible for upstream
observers. The use of outflow disk averaging significantly decreases this spurious noise
and should be used routinely. The combination of pressure-based formulation and outflow
disk averaging provides the best results.
(c) The study of outflow disk averaging has shown that the number and the spacing
of the outflow disks can be estimated from a simple expression, given in this report.
(d) Classical windowing should be used; the Hanning window is used here.
(e) To obtain converged acoustic results at a given period, the time record used for
FWH computations has to be at least 10 times longer.
(f) Since the experimental results available here are probably not measured in the far
field, the far-field assumption is not used. Note that the running time that could be saved
using a far-field assumption is not considered significant, as the main time-consuming
step of the calculations is the LES.
(g) The FWH surface can be intermittently crossed by the edge of the jet without
generating noticeable spurious noise. However, no clear criterion has been determined to
define the optimal location of the FWH surface. To make noise results less sensitive to
the FWH surface location, the grid has to be designed with care, to ensure the proper
propagation of acoustic waves from the location where they are generated to the surface.

Acknowledgments
This study is funded by NASA in the framework of the NASA Research Opportunities
in Aeronautics. Dr. J. Bridges and Dr. J. De Bonis, NASA Glenn Research Center, are
gratefully acknowledged for their help and for sharing the experimental database.

REFERENCES
Bodony, D. J. & Lele, S. K. 2008 Current status of jet noise predictions using large-
eddy simulation. AIAA Journal 46 (3), 364–380.
Bogey, C. & Bailly, C. 2006 Investigation of downstream and sideline subsonic jet
noise using large eddy simulation. Theoret. Comput. Fluid Dynamics 20 (1), 23–40.
Bridges, J. & Wernet, M. P. 2008 Turbulence associated with broadband shock
noise in hot jets, AIAA 2008-2834. In 14th AIAA/CEAS Aeroacoustics Conference,
Vancouver, CANADA, 5-7 May 2008 .
 Post-processing of LES for jet noise predictions 31
Cook, A. W. & Cabot, W. H. 2005 Hyperviscosity for shock-turbulence interactions.
J. Comput. Phys. 203, 379–385.
Eastwood, S., Tucker, P. & Xia, H. 2009 High-performance computing in jet
aerodynamics. In Parallel Scientific Computing and Optimization (ed. S. N. York),
Springer Optimization and Its Applications, vol. 27, pp. 193–206.
Farassat, F. & Myers, M. K. 1988 Extension of Kirchhoff’s formula to radiation from
moving surfaces. J. Sound Vib. 123 (3), 451–460.
Ffowcs Williams, J. E. & Hawkings, D. L. 1969 Sound generation by turbulence
and surface in arbitrary motion. Phil. Trans. R. Soc. Lond. 264, 321–342.
Freund, J. B. 2001 Noise sources in a low-Reynolds-number turbulent jet at Mach 0.9.
J. Fluid Mech. 438, 277–305.
Ham, F. E., Sharma, A., Shoeybi, M., Lele, S. K. & Moin, P. 2009 Noise prediction
from cold high-speed turbulent jets using large-eddy simulation. AIAA 2009-9. In
47th AIAA Aerospace Sciences Meeting Including The New Horizons Forum and
Aerospace Exposition, Orlando, Florida, 5-8 January 2009 .
Lighthill, M. J. 1952 On sound generated aerodynamically : I. General theory. Proc.
R. Soc. Lond. A 211, 564–581.
Lilly, D. K. 1992 A proposed modification of the Germano sub-grid closure method.
Phys. FluidsA 4 (3), 633–635.
Mani, A., Larsson, J. & Moin, P. 2009 Suitability of artificial bulk viscosity for
large-eddy simulation of turbulent flows with shocks. J. Comput. Phys. 228 (19),
7368–7374.
Mendez, S., Shoeybi, M., Sharma, A., Ham, F. E., Lele, S. K. & Moin, P. 2010
Large-eddy simulations of perfectly-expanded supersonic jets: Quality assessment
and validation. AIAA 2010-271. In 48th AIAA Aerospace Sciences Meeting Including
The New Horizons Forum and Aerospace Exposition, Orlando, Florida, 4-7 January
2010 .
Moin, P., Squires, K., Cabot, W. & Lee, S. 1991 A dynamic subgrid-scale model
for compressible turbulence and scalar transport. Phys. FluidsA 3 (11), 2746–2757.
Rahier, G., Prieur, J., Vuillot, F., Lupoglazoff, N. & Biancherin, A. 2003
Investigation of integral surface formulations for acoustic predictions of hot jets start-
ing from unsteady aerodynamic simulations. AIAA 2003-3164. In 9th AIAA/CEAS
Aeroacoustics Conference, Hilton Head, South Carolina, May 12-14, 2003 .
Shoeybi, M., Svärd, M., Ham, F. & Moin, P. 2009 An adaptive hybrid implicit-
explicit scheme for DNS and LES of compressible flows. Submitted to J. Comput.
Phys. .
Shur, M. L., Spalart, P. R. & Strelets, M. K. 2005 Noise prediction for increas-
ingly complex jets. Part I: Methods and tests. Int. J. Aeroacous. 4 (3&4), 213–246.
Spalart, P. R. & Shur, M. L. 2009 Variants of the Ffowcs Williams–Hawkings equa-
tion and their coupling with simulations of hot jets. Int. J. Aeroacous. 8 (5), 477–492.
Tam, C. K. W. 1995 Supersonic jet noise. Annu. Rev. Fluid Mech. 27, 17–43.
Tam, C. K. W. 1998 Jet noise: since 1952. Theoret. Comput. Fluid Dynamics 10,
393–405.
Uzun, A., Lyrintsis, A. S. & Blaisdell, G. A. 2004 Coupling of integral acoustics
methods with LES for jet noise prediction. Int. J. Aeroacous. 3 (4), 297–346.