Applied Acoustics 117 (2017) 94–105

Contents lists available at ScienceDirect

Applied Acoustics
journal homepage: www.elsevier.com/locate/apacoust

Broadband noise prediction using large eddy simulation and a frequency

domain method
Fan Tong a,⇑, Wei-Yang Qiao a, Wei-Jie Chen a, Liang-Feng Wang a, Xun-Nian Wang b
a
School of Power and Energy, Northwestern Polytechnical University, Xi’an 710072, China
b
State Key Laboratory of Aerodynamics, China Aerodynamics Research and Development Center, Mianyang 621000, China

a r t i c l e i n f o a b s t r a c t

Article history: A new LES-acoustic analogy method for accurate flow and broadband noise prediction is proposed. A fre-
Received 4 May 2016 quency domain method for the generalized Lighthill acoustic analogy theory is derived in detail and the
Received in revised form 26 July 2016 final equations for code is provided which can help to bridge the gap between the flow field prediction
Accepted 1 November 2016
and the acoustic field prediction for those who are interested in acoustic results but lack of acoustic pre-
Available online 9 November 2016
diction ability. The hybrid method (LES and acoustic analogy method) for broadband noise prediction is
validated using the rod-airfoil interaction problem. Both flow field results and acoustic field results are
Keywords:
compared with experimental results and other numerical results in detail. The flow field results agree
Broadband noise
Large eddy simulation
very well with the experimental results and other numerical results. The predicted acoustic results
Acoustic analogy and experimental results also reach good agreement both in far field acoustic pressure Power Spectral
Frequency domain method Density (PSD) and noise directivity. The method developed in the paper proves to be an effective tool
Span correction for broadband noise prediction. In addition, the different span correction methods for the acoustic pres-
sure spectrum are also discussed.
Ó 2016 Elsevier Ltd. All rights reserved.

1. Introduction fully numerical prediction methods [6] and hybrid prediction

methods [7–10]. Empirical prediction models are generally simple
The strong market demand for quieter aircraft encourages air- and fast but they cannot assess the acoustic performance between
plane and turbo engine manufacturers to provide more environ- different detailed designs of engine components since they can
mentally friendly and quieter aircraft and turbo engines. Noise only associate a few parameters. The analytical prediction models
prediction capability is of great importance for the design of future can provide satisfactory results fast. Nevertheless, they cannot
aircraft and turbo engines to make sure that the increasingly more guide the detailed designs due to the many simplifications. Fully
rigorous regulations can be met. numerical prediction methods can provide flow-to-far field simu-
Turbo engine noise involves tonal noise and broadband noise. lations with little simplification of the geometry. However, they
For modern ultra-high by-pass ratio engines, fan noise is becoming require huge computational resources. The hybrid prediction
more and more significant. The tonal part of the fan noise and the methods, which use CFD (like LES, DES, URANS) to obtain noise
broadband part of the fan noise both carry about half of the sound sources information and use acoustic analogy [11–14] to obtain
power. The underlying generation mechanism of fan tonal noise is far field acoustic information, can take detailed design parameters
relatively well understood. However, the fan broadband noise gen- into consideration and require much less computational resources
eration mechanism is much more complicated and only partly compared with fully numerical prediction methods. For these rea-
understood. And for military aircraft, the broadband noise from sons, the hybrid prediction methods are very promising methods
the exhaust jet often dominates all other sources. Therefore, broad- to predict aircraft/engine broadband noise.
band noise is of great academic and technical importance. In order to predict noise sources, especially broadband noise
There are several methods to predict broadband noise: empir- sources, highly accurate unsteady CFD solutions must be obtained.
ical prediction models [1–3], analytical prediction models [4,5], LES numerical method is adopted in this study. The rod-airfoil con-
figuration is particularly suitable for the assessment of CFD codes
in modeling broadband noise sources. Its relevance has been thor-
⇑ Corresponding author. oughly discussed by Jacob et al. [15]. The rod-airfoil configuration
E-mail address: [email protected] (F. Tong). combines the periodic vortex shedding with a random

http://dx.doi.org/10.1016/j.apacoust.2016.11.001
0003-682X/Ó 2016 Elsevier Ltd. All rights reserved.
 F. Tong et al. / Applied Acoustics 117 (2017) 94–105 95

perturbation due to the wake’s transition into turbulence. The 2. Methodology

airfoil undergoes a broadband perturbation which is dominated
by a preferred shedding frequency, somewhat like that observed 2.1. Numerical method for flow field
in turbomachinery applications. The ability of a combined CFD/
acoustic approach to predict the spectral broadening around the LES is used to compute the broadband noise sources with the
shedding frequency and its harmonics is a relevant measure of commercial code CFX [22]. In LES, the large three-dimensional
its ability to model broadband sources [15]. Due to these reasons, unsteady turbulent motions are directly represented, whereas
the rod-airfoil configuration is widely used for assessing the differ- the effects of the smaller-scale motions are modelled. The rationale
ent CFD/acoustic approaches. Casalina et al. [16] used 2D unsteady behind LES technique is a separation between large and small
RANS combined with FW-H equation to predict rod-airfoil interac- scales. Large scales of the flow contain the main part of the total
tion noise. The vortex shedding frequency is significantly over pre- fluctuating kinetic energy and characterize the flow. The driving
dicted. Magagnato [17] and Boudet et al. [8,18] performed the first physical mechanisms are carried by the large scales. The large
3D LES on rod-airfoil interaction noise. Boudet’s LES was performed scales of the flow are sensitive to the boundary conditions and so
with the Turb’Flow code and the near field is in good agreement are anisotropic [23]. In contrast, small scales of the flow contain
with experiment. However the far field acoustic results are poorly only a few percent of the total kinetic energy and have weak influ-
converged. Peth et al. [19] used LES and the Linearized Perturbed ence on the mean movement. Their main function is viscous dissi-
Compressible Equations (LPCE) to predict rod-airfoil interaction pation, however, they can also have an effect on higher scales. In
noise. The shedding frequency is slightly under-predicted and the LES, the unresolved small scales of the flow may be isotropic or ani-
far field pressure Power Spectral Density (PSD) peak is under pre- sotropic which depends on the user and how they decided to
dicted by 5–7 dB. More recently, Jacob et al. [20], Giret et al. [7] and model the unresolved scales. A filtering operation is defined to
Greschner [9] also carried out detailed analysis of rod-airfoil inter- decompose the flow variable U into the sum of a filtered (or
action noise prediction. resolved) component U  and a residual (or subgrid-scale, SGS) com-
Many researchers have validated their in-house codes using this ponent U0 . The equations for the filtered field can be derived from
benchmark case and most of their acoustic analogy methods are the Navier-Stokes equations, however a closure problem arises
based on time domain method [7–9,10,20]. At the same time, because of the SGS stress tensor. The closure can be obtained by
although many of the CFD solvers have demonstrated impressive modeling the SGS stress tensor.
ability in flow field prediction, the accurate prediction of aeroa- There are several methods to model the SGS stress. Smagorin-
coustic noise, especially broadband noise, is a more challenging sky proposed the original Smagorinsky model [24], Nicoud and
work compared to steady or unsteady flow field predictions. Ducros proposed the WALE model [25] (wall-adapted local eddy-
Because of this, only a few CFD solvers have integrated the basic viscosity model), Germano and Lilly presented the Dynamic
noise prediction ability, for example the familiar Fluent solver. Smagorinsky-Lilly model [26,27]. The Smagorinsky model and
Whereas there are also many CFD solvers whose noise prediction WALE model are algebraic models and the model coefficient is con-
ability is limited and not mature (e.g. CFX). Obviously, developing stant. In contrast, the Dynamic Smagorinsky-Lilly model uses
the broadband noise prediction tools will help to extend CFD sol- information contained in the resolved turbulent velocity field to
vers’ functionality and add our option and freedom. Epikhin [21] evaluate the model coefficient. Thus, the model coefficient is no
presented a Dynamic Library for computational aeroacoustics longer a constant value and adjusts automatically to the flow type.
applications using the OpenFOAM source package, however, the In this paper, the Dynamic Smagorinsky-Lilly model and the Sec-
details of the method are not available which may hinder its fur- ond Order Backward Euler transient scheme are adopted.
ther application and development by others. Moreover, the
Dynamic Library is specific for OpenFOAM source package. For
2.2. Numerical method for acoustic field
other codes the Dynamic Library does not work. Therefore, there
is a need to develop a broadband noise prediction tool for various
The acoustic prediction method is based on Goldstein’s general-
CFD codes which are not ready for direct broadband noise
ized Lighthill equation [14]. Goldstein extended Lighthill’s acoustic
prediction.
analogy to include the effects of solid boundaries and moving med-
The current study proposed a new LES-acoustic analogy
ium. The fundamental equation governing the generation of sound
method for accurate flow and broadband noise prediction. A fre-
in the presence of solid boundaries is presented below with slight
quency domain method for generalized Lighthill acoustic analogy
changes to its original form for the current application.
theory is derived in detail. Compared with time domain method,
Z Z Z Z
the frequency domain method has several advantages. For time !
T
DG !
T
@G !
domain methods, the determination of the retarded time is com- c20 q0 ðx ; tÞ ¼ q0 V 0N dAðy Þds þ fi dAðy Þds
T AðsÞ Ds T AðsÞ @yi
putationally intensive. In addition, the variables and their time Z Z
T
@2G !
derivatives must be interpolated to the retarded time for every þ T 0ij d y ds ð1Þ
grid point, retarded time and observer position, which also leads T vðsÞ @yi @yj
to intensive calculation. In contrast, frequency domain methods
where c0 is the ambient speed of sound and q0 is the acoustic den-
are more computationally efficient. The frequency domain meth-
sity disturbance. At sufficient distance from the source, c20 q0 is equal
ods can be carried out further faster when harmonic noise is of !
interest where only several selected frequencies need to be calcu- to the acoustic pressure p0 . x ¼ ðx1 ; x2 ; x3 Þ is the observation coordi-
!
lated. In this paper, the final equations of the frequency domain nate and y ¼ ðy1 ; y2 ; y3 Þ is the source coordinate. s is the source
method for code are also provided which can help to bridge the time (retarded time). V 0N is the velocity of the surface normal to
gap between the flow field prediction and the acoustic prediction itself relative to the fluid. The first term in Eq. (1) represents the
for those who are interested in acoustic results but lack of acous- sound generated due to volume displacement effects of the surface,
tic prediction ability. The code of the frequency domain method the second term represents the sound generated due to the exertion
is then successfully validated through the comparison with of unsteady forces by the boundaries on the fluid, and the last term
experimental results and other numerical results. represents the generation of the sound by volume sources. In each
96 F. Tong et al. / Applied Acoustics 117 (2017) 94–105

of the three terms, the outer integral is over a range of source time Since only f i eixs is dependent on source time s; Eq. (12) can be
T < s < T large enough to include all contributions to the noise rewritten as follows
signal at observer time t. For stationary surface, the first term
becomes zero. When the incoming flow Mach number is small, Z Z  
!
T
@ eixr=c0 !
the volume sources contribute to the noise relatively a little and p0 ðx ; xÞ ¼ f i eixs ds dAðy Þ ð13Þ
A T @yi 4pS
can be neglected. When the first term and last term in Eq. (1) are
ignored, the acoustic pressure at sufficient distance from the source The inner integration is the Fourier transform of the force f i ;
can be expressed as follows thus we can obtain the final expression of acoustic pressure in fre-
Z Z quency domain.
!
T
@G !
p0 ðx ; tÞ ¼ fi dAðy Þds ð2Þ Z  
T AðsÞ @yi ! ! @ eixr=c0 !
p0 ðx ; xÞ ¼ f i ðy ; xÞ dAðy Þ ð14Þ
where G is the free space, moving medium, time dependent Green A @yi 4pS
function [28] The Eq. (14) can be presented in a more simple form
dðt  s  r=c0 Þ Z
G¼ ð3Þ ! ! @Gx !
4pS p0 ðx ; xÞ ¼ f i ðy ; xÞ dAðy Þ ð15Þ
A @yi
where S is the amplitude radius
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
h iffi ixr=c
where Gx ¼ e 4pS 0 is the form of Green function in the
S¼ ðx1  y1 Þ2 þ b2 ðx2  y2 Þ2 þ ðx3  y3 Þ2 ð4Þ !
frequency domain. The force f i ðy ; xÞ can be obtained from LES
@Gx
and r is the phase radius with b2 ¼ 1  M2 and @y can also be acquired through the following equation
i

derivation.
Mðx1  y1 Þ þ S
r¼ ð5Þ Note that
b2
 
ixr=c0 @r @S
@Gx e i cx0 S @y  @y
M ¼ U 0 =c0 ð6Þ
¼ ðFor i ¼ 1; 2; 3Þ ð16Þ
i i

@yi 4pS2
In the current work, the stationary airfoil noise is of interest. So
the integral symbol AðsÞ in the area integration is independent of
Then take partial derivative of r and S with yi and we can obtain
source time s and AðsÞ can be changed to A. Eq. (2) can be reformed
the following equations
as follows
Z Z
!
T
@G ! @ r y1  x1 . 
p0 ðx ; tÞ ¼ fi dsdAðy Þ ð7Þ ¼ M 1  M2 ð17Þ
A T @yi @y1 S

Apply Fourier transform to Eq. (7), we can obtain the expression @ r ðy2  x2 Þ
of acoustic pressure in frequency domain ¼ ð18Þ
@y2 S
Z þ1 Z Z
!
T
@G !
p0 ðx ; xÞ ¼ fi dsdAðy Þeixt dt ð8Þ @ r ðy3  x3 Þ
1 A T @yi ¼ ð19Þ
@y3 S
Since only G is related to reception time t and A is
independent of reception time t or source time s , Eq. (8) can be @S ðy  x1 Þ
changed to ¼ 1 ð20Þ
@y1 S
Z Z Z þ1
!
T
@G ixt !
p0 ðx ; xÞ ¼ fi e dtdsdAðy Þ ð9Þ @S ðy  x2 Þð1  M 2 Þ
A T 1 @yi ¼ 2 ð21Þ
@y2 S
Insert the time dependent Green’s function from Eq. (3) into Eq.
(9) to obtain
@S ðy  x3 Þð1  M 2 Þ
¼ 3 ð22Þ
Z Z Z þ1   @y3 S
0
!
T
@ 1 !
p ðx ; xÞ ¼ fi dðt  s  r=c0 Þ eixt dtdsdAðy Þ
A T 1 @yi 4pS From Eqs. (17)–(22) and Eq. (16), we can obtain
ð10Þ n h i o
! @Gx e
ixr=c0
ðy1  x1 Þ ix=c0  ð1  M 2 Þ=S þ ixSM=c0
Note that f i is dependent on source time s and source location y ¼ ð23Þ
@y1 4pS2 ð1  M2 Þ
and independent on reception time t, so Eq. (10) can be written as
follows n h io
Z Z  Z þ1  @Gx e
ixr=c0
ðy2  x2 Þ ix=c0  ð1  M 2 Þ=S
!
T
@ 1 !
¼ ð24Þ
p0 ðx ; xÞ ¼ fi dðt  s  r=c0 Þeixt dt dsdAðy Þ @y2 4pS2
A T @y i 4 p S 1

ð11Þ n h io
@Gx e
ixr=c0
ðy3  x3 Þ ix=c0  ð1  M 2 Þ=S
Take consideration of the integration property of d function, we ¼ ð25Þ
can obtain
@y3 4pS2
Z Z   Finally, the acoustic pressure can be calculated by the following
!
T
@ 1 ixðsþr=c0 Þ !
p0 ðx ; xÞ ¼ fi e dsdAðy Þ ð12Þ equation
A T @yi 4pS
 F. Tong et al. / Applied Acoustics 117 (2017) 94–105 97

Z  !  @G 
! ! @Gx x ! @Gx ! Fig. 2 shows the sketch of the computational domain. The com-
p0 ðx ; xÞ ¼ f 1 ðy ; xÞ þ f 2 y; x þ f 3 ðy ; xÞ dAðy Þ
A @y1 @y2 @y3 putational domain extends 26c in the stream-wise direction, 20c in
2 n h i o
Z ðy1  x1 Þ ix=c0  ð1  M2 Þ=S þ ixSM=c0 the cross-stream direction and 2d in the spanwise direction. The
!
¼e ixr=c0 4f 1 ðy ; xÞ sketch of the computational mesh near the airfoil is shown in
A 4pS2 ð1  M2 Þ Fig. 3. Fig. 4 shows the dimensionless wall-cell sizes Dyþ along
n h io the airfoil. It can be seen that the Dyþ is below 1. In addition, there
! ðy2  x2 Þ ix=c0  ð1  M2 Þ=S
þ f 2 ðy ; xÞ are 240 grid points around the rod and 400 grid points around the
4pS2 airfoil resulting in the dimensionless wall-cell sizes Dxþ  90. In
n h io3
ðy3  x3 Þ ix=c0  ð1  M2 Þ=S the spanwise direction, there are 33 grid points and the dimension-
! !
þ f 3 ðy ; xÞ 5dAðy Þ ð26Þ less wall-cell sizes Dzþ  60. The total grid number is about
4pS2 5.15 million. At the same time, a coarser mesh which has also 33
!
grid points in the spanwise direction but with less grid points in
If the viscous stress tensor is ignored, f i ðy ; xÞ can be calculated
as follows

! ! !
f i ðy ; xÞ ¼ p0 ðy ; xÞn i ð27Þ

!
where p0 ðy ; xÞ is the airfoil surface pressure fluctuation spectrum
!
and n i is the unit normal vector in ith direction.
At last, it should be pointed out that the current approach uses a
free-space Green’s function and does not account for the effects of
scattering of sound by solid bodies or propagation/refraction
effects by the flow.
Fig. 3. Sketch of the computational Mesh.

3. Benchmark experiment and numerical setup

The rod-airfoil configuration in the benchmark experiment by

Jacob et al. [15] is shown in Fig. 1. The configuration involves a
symmetric NACA0012 (chord: c = 100 mm) airfoil located one
chord downstream of a rod (rod diameter: d = 10 mm). The airfoil
and rod both have a span of 300 mm. In the reference configura-
tion, the incoming flow velocity is 72 m/s and the free stream tur-
bulence intensity is about 0.8%. The Reynolds number based on rod
diameter is about 4:8  104 and the Reynolds number based on air-
foil chord is about 4:8  105 :

Fig. 1. Sketch of the experimental set-up [15]. Fig. 4. Dimensionless wall-cell sizes Dyþ distribution along the airfoil at mid-span.

Fig. 2. Sketch of the computational domain.

98 F. Tong et al. / Applied Acoustics 117 (2017) 94–105

the airfoil plane was also studied. The total grid number of the
coarse mesh is 3.46 million. Fig. 5 shows the velocity profiles of
the rod wake predicted by the two different meshes. It can be seen
that both predicted results agree reasonably well with the experi-
mental results. However, the coarse mesh slightly over predicts the
wake deficit while the fine mesh yields much better results. There-
fore, the fine mesh was used hereafter as it is expected that the fine
mesh can give more accurate prediction of the broadband noise
sources.
The highest frequency that can be resolved in the simulation is
determined by the grid. Michel [29] has proposed a grid Strouhal
number to estimate the highest frequency that can be resolved
by the grid. An axial grid Strouhal number is defined as

Sts ¼ f Dxs =U c < 0:25 ð28Þ

where the Dxs is the grid size in the mean flow direction, U c is the
convection speed of the disturbance in the flow and U c  0:8U 0 , f
is the frequency of the emitted sound. The max grid size around
the airfoil is Dxs  0:8 mm for the fine grid. It can be estimated that
the maximum resolvable frequency is f  16 kHz:
The inflow conditions and the flow parameters are based on the
experiment conditions. The incoming free stream velocity is Fig. 6. Residual curves during calculation.
72 m/s. The outlet static pressure is set to 98,900 Pa. For spanwise
direction, both use of periodic boundary conditions and slip (or
ison. The adiabatic no-slip conditions are imposed on the rod and
symmetry) boundary conditions have been documented
[7–9,17,19,20]. For the periodic boundary conditions, the periodic- the airfoil. The computational time step is 1  105 s and the total
ity is forced to be met for the boundaries and all the flow field acquired physical time for acoustic processing is about 0.22 s
quantities of the limiting planes are fully correlated. For slip (or which corresponds to a time interval during which the incoming
symmetry) boundary conditions, one component of velocity flow passes about 159 times of the airfoil chord. The simulation
(z-component) vanishes whereas the two planes in the spanwise was run on High Performance Cluster of Northwestern Polytechni-
direction are not forced to be correlated. The slip condition only cal University with 32 cores. The simulation time was about six
affects the vicinity of the boundary. It is expected that the slip weeks. The relatively long simulation time helps to increase the
(or symmetry) boundary condition will lead to a lower level of stability of the signal and is beneficial for the calculation of the
coherence than periodic boundary condition. Lockard [30] has acoustic results. The code employs a dual time-stepping algorithm
pointed out that only calculations based on the full length of the and we chose the number of sub-iterations to ensure that the vari-
model span were able to capture the complete decay in the span- ables plateaued and the residual dropped by 2–3 orders of magni-
wise correlation. However, in practical calculations, this is often tude within each time step. Fig. 6 shows the residual curves of
not feasible because of the huge computational resources it different variables during the calculation. The root mean square
demanded. Since most of the studies adopted the periodic bound- (RMS) residual for three momentum equations is about 1  105
ary conditions in the spanwise direction [7,17,19,20], the same to 2  105 and the RMS residual for continuity equation is about
boundary conditions are used in this paper for more direct compar- 1  106 to 2  106 . Fig. 7 shows the time history of the airfoil lift

Fig. 5. Velocity profiles of the rod wake of two different meshes. Fig. 7. Time history of the airfoil lift and drag coefficient.
 F. Tong et al. / Applied Acoustics 117 (2017) 94–105 99

and drag coefficient. It can be seen from Figs. 6 and 7 that the flow be found that the present numerical results agree well with the
field converges reasonably well. experimental results and other numerical results. The rod wake
In the present study, the solid surface was selected as the inte- mean velocity profile is well predicted by the current simulation.
gral surface. Therefore, only the dipole sources are considered since The minimum value of u/U0 is about 0.709 in experimental
the dipole sources dominate the sound sources in the current low results while the predicted minimum value of u/U0 is 0.699 and
Mach number condition. In the following part of the paper, it will the relative error is about 1.41%. It can be seen from Fig. 9(b) that,
be shown that the noise prediction results agree well with overall the turbulent intensity profile predicted by present LES
experimental results although only the dipole sources are reaches an agreement with experiment results and other numer-
considered. ical results. The maximum value of turbulent intensity of experi-
ment and the present simulation is about 0.176 and 0.193,
respectively. The predicted value is 9.66% higher than that of
4. Results and discussion
the experiment.
Fig. 9(c) shows the mean velocity profile at plane B. The present
Present numerical flow field results and far field acoustic results
LES performs similarly to Giret’s LES results and the mean profile is
are compared with experimental results by Jacob et al. [15] and
over predicted by about 5.33%. This over prediction is also found in
other numerical results. Both the steady flow field and unsteady
many LES results [8,20]. Fig. 9(d) shows the turbulent intensity
flow field are considered. The far field noise spectrum and noise
profile at plane B. The peak value of turbulent intensity of
directivity are also compared with experimental data.
experiment and the present simulation is about 0.164 and 0.223,
respectively. The present simulation over predicts the turbulent
4.1. Comparison positions intensity peak by about 36.0%. Part reason for this large
discrepancy is that the peak value of the turbulent intensity is hard
The profiles of the mean velocity and the rms (root mean to measure in the experiment due to the spatial resolution of the
square) value of velocity fluctuations obtained from LES are com- hot wire. Fig. 9(e) and (f) shows the mean velocity profile and tur-
pared with experimental results as well as other numerical results. bulent intensity profile at plane C, which represents the near wake
Fig. 8 shows the sketch of measurement positions. The profiles at of the airfoil. Compared with experimental results, the maximum
three different locations are compared, i.e. plane A at x/c = 0.25 wake deficit is under predicted by about 7.90% in Fig. 9(e). The
which corresponds to the far wake of the rod, plane B at maximum value of the turbulent intensity is 0.135 for experiment
x/c = 0.25 which is slightly upstream of the airfoil thickest point, and 0.152 for present simulation. The maximum value of the
plane C at x/c = 1.1 which corresponds to the near wake of the air- turbulent intensity is over predicted by about 12.6% in Fig. 9(f).
foil. It should be noted that x/c = 0 corresponds to airfoil leading There is a discrepancy between the experimental turbulent
edge. intensity results and numerical turbulent intensity results near
Streamwise velocity power spectrum density is also compared the location of y/c = 0.2, which can also be seen in Giret’s numerical
with available experimental results at two different locations, i.e. results.
P1 and P2 (Fig. 8). P1 is located at (0.25c, 0.08c) and P2 is Overall, the present LES gives satisfactory prediction of mean
located at (0.25c, 0.16c). P1 is 0.2d off the airfoil surface and velocity profiles and turbulent intensity profiles.
P2 is 1d off the airfoil surface. In addition, the far field acoustic
pressure spectrum at a distance of R = 1.85 m of 90° angular
angle (just above the airfoil) from the airfoil center is also com- 4.3. Unsteady flow field
pared with experimental results. Moreover, the Strouhal number
of the vortex shedding frequency is compared with experimental The unsteady flow field is also investigated. The Strouhal num-
results. ber is defined as St ¼ fd=U 0 . The experimental value of vortex shed-
ding Strouhal number is 0:19  0:002. The predicted value of
Strouhal number is 0:202. The predicted result agrees reasonably
4.2. Steady flow field
well with the experimental result which indicates that the physics
of the flow separation on the rod are correctly reproduced in the
Velocity profiles and turbulent intensity profiles are compared
current simulation.
with experiment results by Jacob et al. [15] and previous LES (or
The instantaneous span-wise vorticity and iso-surfaces of the
DES) results by Jacob [15], Giret [7] and Greschner [9] at three
Q-criterion are shown in Figs. 10 and 11, respectively. It can be
different locations in Fig. 8. Fig. 9(a) and (b) shows the mean
seen that the main vortices shedding from the rod form regular
velocity profile and turbulent intensity profile at plane A. It can
Karman vortex street which impinges onto the airfoil and partly
broken into small structures. Moreover, many smaller vortices
can be observed in the rod wake, due to the transition to
turbulence in the shear layer of the rod. The broken of large
vortices is successfully captured by the current simulation. It is
expected that the rod wake interacting with the leading edge of
the airfoil is the main mechanism of the turbulence airfoil interac-
tion noise.

4.4. Velocity spectrum

The power spectrum density of the streamwise velocity compo-

nent is compared with the experimental results in Fig. 12. The
agreement with experimental data is satisfactory. It can be seen
from Fig. 12(a) that the current simulation can capture the broad-
ened peak which means the interaction between the random part
Fig. 8. Sketch of measurement positions. of the flow and the quasi deterministic vortex shedding is well
100 F. Tong et al. / Applied Acoustics 117 (2017) 94–105

(a) Mean velocity proile at plane A (d) Turbulent intensity proile at plane B

(b) Turbulent intensity proile at plane A (e) Mean velocity proile at plane C

Fig. 9. Comparison of mean velocity profile and turbulent intensity profile with experiment and other numerical simulations.

predicted. However, the peak of the power spectrum density is is also well predicted by the current simulation. The peak of the
over predicted by about 1.6 dB and the vortex shedding frequency power spectrum density is over predicted by about 3.0 dB and
is over predicted by 6.3%. Fig. 12(b) shows the power spectrum the predicted power spectrum density is about 2 dB lower than
density of the streamwise velocity at P2, which is one rod diame- that of the experimental result in high frequency range (around
ter off the airfoil surface. It can be seen that vortex shedding fre- St = 0.8) where the predicted spectrum seems to decay slightly
quency is much less obvious than that at P1. The spectrum at P2 more quickly.
 F. Tong et al. / Applied Acoustics 117 (2017) 94–105 101

Fig. 10. Span-wise vorticity contour.

    Lc Lexp
Spp ðf Þ exp
¼ Spp ðf Þ sim þ 20 lg þ 10 lg ; Lsim < Lc 6 Lexp
Lsim Lc
ð32Þ

    Lexp
Spp ðf Þ exp
¼ Spp ðf Þ sim þ 20 lg ; Lexp < Lc ð33Þ
Lsim
Seo [32] assumes the coherence function of the Gaussian
Fig. 11. Iso-surfaces of the Q-Criterion (Q = 2:5  106 s2).
function form and the correction function can be simplified as
follows

4.5. Acoustic results     Lexp pffiffiffiffi

Spp ðf Þ exp
¼ Spp ðf Þ sim þ 10 lg Lc 6 Lsim = p ð34Þ
Lsim
The spanwise length of the LES computational domain is often
pffiffiffiffi
limited by computational resources. As a result, the spanwise     Lc pLexp
length in the numerical simulation Lsim is usually shorter than Spp ðf Þ
exp
¼ S pp ðf Þ sim
þ 10 lg þ 10 lg ;
Lsim Lsim
the experimental spanwise length Lexp . In order to compare the pre- pffiffiffiffi pffiffiffiffi
Lsim = p < Lc 6 Lexp = p ð35Þ
dicted sound pressure level with experimental data, the sound
pressure level obtained from numerical simulations must be cor-
rected. Different correction methods have been developed by Kato     Lexp pffiffiffiffi
Spp ðf Þ exp
¼ Spp ðf Þ sim þ 20 lg ; Lexp = p < Lc ð36Þ
et al. [31], Seo et al. [32], Boudet et al. [8], and Perot et al. [33]. The Lsim
correction of the power spectral density of the pressure in the far Perot [33] assumes the coherence function of the exponential
field can be expressed as follows [8], function form and gives a correction method independent on span-
R Lexp R Lexp ! wise coherence length but dependent of the distance between
  Cðjz2  z1 j; f Þdz1 dz2 observer point and the source point
ðSpp ðf ÞÞexp ¼ Spp ðf Þ sim þ 10 lg R L R L
0 0

0
sim
0
sim
Cðjz2  z1 j; f Þdz1 dz2    
Lexp Lsim
ð29Þ     arctan R
 arctan R
Spp ðf Þ exp
¼ Spp ðf Þ sim
þ 10 lg   ð37Þ
where Cðjz2  z1 j; f Þ is the coherence function at the frequency f arctan Lsim
R
between two points along the solid surface. The coherence length
Lc is defined as Actually, the coherence length Lc is a function of frequency and
the corrections should also be a function of frequency. Unfortu-
Z þ1
nately, the exact Lc as a function of frequency is often not available
Lc ¼ Cðjz2  z1 j; f Þdz2 ð30Þ so a reasonable estimate of coherence length is usually used in
z1
practice. According to Jacob’s experiment results, the spanwise
Simplifications of Eq.(29) can be made by assuming the coher- pressure correlation length Lc on the rod is about 6.5d [15].
ence function of specific mathematical form. Kato [31] assumes Fig. 13 shows the comparison of different correction methods. It
the coherence function of the rectangular function form and the should be noted that in Fig. 13 the following parameters are used,
correction function can be simplified as follows Lsim ¼ 2d, R ¼ 1:85 m and Lexp ¼ 30d. For the current problem,
Lc ¼ 6:5d and Lc =Lsim ¼ 3:25. So Kato’s method will lead to a correc-
    Lexp
Spp ðf Þ exp
¼ Spp ðf Þ sim þ 10 lg ; Lc 6 Lsim ð31Þ tion of 16.87 dB and Seo’s method will lead to a correction of
Lsim 19.37 dB while Perot’s method gives a correction of 11.42 dB.
102 F. Tong et al. / Applied Acoustics 117 (2017) 94–105

Fig. 13. Comparison of different spanwise correction methods.

(a) PSD at P1

Fig. 14. Comparison of SPL of the acoustic pressure at 90° and at 1.85 m corrected
Fig. 12. Comparison of PSD of the streamwise velocity component at P1 and P2. by different methods.

The sound pressure level (SPL) of the far field acoustic pressure
broadband part of the noise in high frequency range (near St = 1)
corrected by different methods at 1.85 m for the angular position
is also lower than that of experiment results by 3–4 dB. The Seo’s
90° is shown in Fig. 14 where the abscissa is scaled to Strouhal
correction method over predicts the peak value of SPL by 1.5 dB
number. The sound pressure level is defined as
and the broadband part of the noise by about 4 dB (near St = 1).
! The Kato’s correction method performs well both for the noise
PSDðf Þ
SPLðf Þ ¼ 10 log 10 ð38Þ peak level prediction and the broadband part of the noise predic-
p2ref tion in a large frequency range. The peak value of SPL is under pre-
dicted by 1 dB while the broadband part of noise is over predicted
where pref ¼ 2  105 Pa and the PSD is normalized to be per unit f by about 1.5 dB in high frequency range (near St = 1) and about
and Welch’s method was used to calculate the PSD in Eq. (27). 4 dB between St = 0.2–0.3. This big discrepancy is mainly due to
It can be seen from Fig. 14 that the peak value of SPL corrected the over-predicted peak frequency (vortex shedding frequency)
by Kato’s method is 90.44 dB, 92.94 dB by Seo’s method and which leads to the shift of the whole predicted spectrum. The vor-
84.99 dB by Perot’s method. The peak value of SPL is about tex shedding Strouhal number is 0:19  0:002 in experiment
91.44 dB in the experiment. Perot’s correction method obviously whereas the predicted value of Strouhal number is 0:202 with a
under predicts the noise peak level by 6.45 dB and the predicted 6.3% over predict. To address this effect, Fig. 15 shows the
 F. Tong et al. / Applied Acoustics 117 (2017) 94–105 103

Fig. 15. Comparison of SPL of the acoustic pressure at 90° and at 1.85 m corrected
by Kato’s method (with peak frequency corrected).

comparison of SPL after the predicted peak frequency are corrected

to the experimental peak frequency and the agreement improves
significantly. The difference between predicted results and
experimental results is within 2–3 dB over a wide frequency range.
The Kato’s correction method can give very good results if the peak
frequency could be predicted more accurately. In addition, the
Kato’s correction method is widely used and demonstrates
satisfactory prediction results [7,9,15]. As a result, the Kato’s
correction method [31], precisely Eq. (32), is used hereinafter to
correct the SPL of acoustic pressure obtained from the numerical
simulation.
Fig. 16 compares the SPL of the far field acoustic pressure at
1.85 m for angular position of 90° with the experimental data
[15] and other numerical results [7,9,20] available. It can be seen
from Fig. 16 that the present prediction results agree well with
experimental data. The discrepancy between the prediction and
experimental results is within 2 dB at the peak noise frequency
and the broadband part of the noise also agrees well with the
experimental results. It can be seen that the present prediction
method shows similar or improved results than other numerical Fig. 16. Acoustic pressure results compared with experimental data, at 90° and
results both in the main peak noise and the broadband part of 1.85 m.
the noise.
The acoustic directivity results are compared with experimen-
tal data in Fig. 17. The prediction results of the present method
agree fairly well with experimental results. The discrepancy 5. Conclusions
between predicted results and experimental results is within
3 dB for most of the angular positions. The maximum discrepancy A new LES-acoustic analogy method for accurate flow and
is about 4.5 dB near angular position of 110°. The acoustic direc- broadband noise prediction is proposed and assessed. A frequency
tivity results is over predicted by 2–4.5 dB between angular posi- domain method for the generalized Lighthill acoustic analogy the-
tion of 90° to 120° and 240° to 270°. Excellent agreement is ory is developed. The frequency domain method is derived in detail
obtained between angular position of 60° to 90° and 270° to and the final equations for code is provided which can help to
300°. Through the comparison of far field acoustic pressure spec- bridge the gap between the flow field prediction and the acoustic
trum and the acoustic directivity between the predicted results prediction for those who are interested in acoustic results but lack
and experimental results, it can be found that the acoustic results of acoustic prediction ability.
obtained from LES and the current acoustic analogy method The rod-NACA0012 airfoil configuration is selected because of
agree fairly well with the experimental results. This indicates its detailed experimental results and its relevance to model broad-
that the current LES and the frequency domain method of gener- band noise. The LES is performed by CFX and the acoustic predic-
alized Lighthill acoustic analogy theory are capable to predict tion method is based on the generalized Lighthill equation. Both
broadband noise sources of complex flows and the resulting the flow field results and the acoustic field results are compared
broadband noise. with the experimental results and other numerical results.
104 F. Tong et al. / Applied Acoustics 117 (2017) 94–105

Appendix A

Nomenclature

c chord length
c0 ambient speed of sound
d rod diameter
f frequency
G filter function or Green function
Lsim spanwise length in numerical simulation
Lexp spanwise length in experiment
Lc correlation length
R distance between observer point and the source point
S amplitude radius
St s grid Strouhal number
St Strouhal number
Spp power spectral density of the pressure in the far field
t observation time
C coherence function
s retarded time
U fluid variable

U filtered variable
U0 unresolved part of the variable
Fig. 17. Acoustic directivity compared with experimental data.
q0 acoustic density disturbance
p0 acoustic pressure
! observation coordinate
x
The flow field results show that the current LES can give accu- ! source coordinate
y
rate prediction of the broadband noise sources although the vortex
V 0N velocity of the surface normal to itself
shedding frequency is slightly over predicted. Both the mean veloc-
ity profiles and turbulent intensity profiles are well predicted.
Moreover, the predicted velocity spectrum agrees very well with
the experimental data.
The acoustic field results show that the predicted acoustic pres- References
sure spectrum results match very well with the experimental data.
The discrepancy between the predicted results and experimental [1] Heidmann MF. Interim prediction method for fan and compressor source
noise. Report, NASA TM X-71763, USA; 1975.
results is within 2 dB at the peak noise frequency and the broad- [2] Raney JP. Development of a new computer system for aircraft noise prediction.
band part of the noise also agrees well with the experimental In: Proceedings of 2nd aero-acoustics conference of American Institute of
results. Aeronautics and Astronautics, Hampton, VA, U.S.A., 24–26, March 1975, paper
no. 75–536.
Due to the discrepancy of span length between the numerical [3] Hough JW, Weir DS. Aircraft noise prediction program (ANOPP) fan noise
and experimental setup, the acoustic pressure spectrum prediction for small engines. Report, NASA CR 198300, USA; 1996.
obtained from numerical results needs to be corrected before [4] Ventres CS. Turbofan noise generation. Volume 1: analysis. Report, NASA CR-
167952, USA; 1983.
compared with experimental data. It is found that the span cor- [5] Meyer HD, Envia E. Aeroacoustic analysis of turbofan noise generation. Report,
rection methods have an important effect on the level of the NASA CR 4715, USA; 1996.
predicted acoustic results. It is shown that the assumption of [6] Clair V, Polacsek C, Garrec TL, et al. CAA methodology to simulate turbulence-
airfoil noise. In: 18th AIAA/CEAS aeroacoustics conference, Colorado Springs,
coherence function of the rectangular function form by Kato is
CO, USA, 04–06, June 2012, paper no. 2012–2189.
reasonable and gives better results than that of Seo’s and Perot’s [7] Giret JC, Sengissen A, Moreau S, et al. Prediction of the sound generated by a
method. rod-airfoil configuration using a compressible unstructured LES solver and a
FW-H analogy. In: 18th AIAA/CEAS aeroacoustics conference, Colorado Springs,
Overall, the LES and acoustic analogy method for broadband
CO, USA, 04–06, June 2012, paper no. 2012–2058.
noise prediction developed in this paper performs well and can [8] Boudet J, Grosjean N, Jacob MC. Wake-airfoil interaction as broadband noise
give a good prediction of the broadband noise. source: a large-eddy simulation study”. Int J Aeroacoust 2005;4:
93–116.
[9] Greschner B, Thiele F, Jacob MC, et al. Prediction of sound generated by a rod–
airfoil configuration using EASM DES and the generalised Lighthill/FW-H
Acknowledgements analogy. Comput Fluids 2008;37(37):402–13.
[10] Galdéano S, Barré S, et al. Noise radiated by a rod-airfoil configuration using
This study is supported by National Natural Science DES and the Ffowcs-Williams & Hawkings’ analogy. In: 16th AIAA/CEAS
aeroacoustics conference, Stockholm, Sweden, 7–9, June, 2010, paper no.
Foundation of China and the project number is 51276149, 2010-3702.
51476134. This research is also supported by State Key [11] Lighthill MJ. On sound generated aerodynamically. I. General theory. Proc R
Laboratory of Aerodynamics of China and the project number Soc 1952;211:564–87.
[12] Lighthill MJ. On sound generated aerodynamically. II. Turbulence as a source of
is SKLA20160201. Additional acknowledgements to Key
sound. Proc R Soc 1954;222:1–32.
Laboratory of Aerodynamic Noise Control of China Aerodynamics [13] Ffowcs Williams JE, Hawkings DL. Sound generated by turbulence and surfaces
Research and Development Center for the project of in arbitrary motion. Philos Trans R Soc 1969;264:321–42.
ANCL20160102. The simulation work was supported by Center [14] Goldstein ME. Unified approach to aerodynamic sound generation in the
presence of solid boundaries. J Acoust Soc Am 1974;56:497–509.
for High Performance Computing of Northwestern Polytechnical [15] Jacob MC, Boudet J, Casalino D, et al. A rod-airfoil experiment as benchmark for
University, China. broadband noise. Model Theor Comput Fluid Dyn 2005;19:171–96.
 F. Tong et al. / Applied Acoustics 117 (2017) 94–105 105

[16] Casalino D, Jacob MC, Roger M. Prediction of rod-airfoil interaction noise using [25] Nicoud F, Ducros F. Subgrid-scale stress modelling based on the square of the
the Ffowcs-Williams and Hawkings analogy. Am Inst Aeronaut Astronaut velocity gradient tensor. Flow Turbulence Combust 1999;62:
2003;41(2):182–91. 183–200.
[17] Magagnato F, Sorgueven E, Gabi M. Far field noise prediction by large eddy [26] Germano M, Piomelli U, Moin P, et al. A dynamic subgrid-scale eddy viscosity
simulation and Ffowcs-Williams Hawkings analogy. In: Proceedings of the 9th model. Phys Fluids A 1991;3(7):1760–5.
AIAA/CEAS aeroacoustics conference, Hilton Head, South Carolina, 12–14, May, [27] Lilly DK. A proposed modification of the Germano subgrid-scale closure
2003, paper no. 2003–3206. method. Phys Fluids A 1992;4(3):633–5.
[18] Boudet J, Casalino D, Jacob MC, et al. Prediction of sound radiated by a rod [28] Hanson DB, Parzych DJ. Theory for noise of propellers in angular inflow with
using large-eddy simulation. In: Proceedings of the 9th AIAA/CEAS parametric studies and experimental verification. Report, NASA CR 4499, USA;
aeroacoustics conference, Hilton Head, South Carolina, 12–14, May 2003, 1993.
paper no. 2003–3217. [29] Michel U, Eschricht D, Greschner B, et al. Advanced DES methods and their
[19] Peth S, Seo JH, Moon YJ, et al. Computation of aerodynamic noise from rod application to aeroacoustics. In: Peng SH, Doerffer P, Haase W, editors.
wake-airfoil interactions. In: Proceedings of the European conference on Progress in hybrid RANS-LES modelling. Berlin, Heidelberg: Springer; 2010. p.
computational fluid dynamics, Egmond aan Zee, The Netherlands, 5–8, 59–76.
September 2006. [30] Lockard DP, Khorrami MR, Choudhari MM, et al. Tandem cylinder noise
[20] Jacob MC, Ciardi M, Gamet L, et al. Assessment of CFD broadband noise predictions. In: Proceedings of the 13th AIAA/CEAS aeroacoustics conference,
predictions on a rod-airfoil benchmark computation. In: 14th AIAA/CEAS Rome, Italy, 23–25, May, paper no. 2007–3450.
aeroacoustics conference, Vancouver, British Columbia, Canada, 5–7, May [31] Kato C, Ikegawa M. Large eddy simulation of unsteady turbulent wake of a
2008, paper no. 2008–2899. circular cylinder using the finite element method. Adv Numer Simul Turbulent
[21] Epikhin A, Evdokimov I, Kraposhin M, et al. Development of a dynamic library Flows 1991;1:49–56.
for computational aeroacoustics applications using the OpenFOAM open [32] Seo JH, Moon YJ. Aerodynamic noise prediction for long-span bodies”. J Sound
source package. Proc Comput Sci 2015;66:150–7. Vib 2007;306(3–5):564–79.
[22] ANSYS CFX. Reference guide, release 14.0. ANSYS Inc.; 2011. [33] Perot F, Auger JM, Giardi H, et al. Numerical prediction of the noise radiated by
[23] Wagner CD, Hüttl T, Sagaut P. Large-eddy simulation for acoustics. 1st a cylinder. In: Proceedings of the 9th AIAA/CEAS aeroacoustics conference,
ed. Cambridge University Press; 2012. p. 90. Hilton Head, South Carolina, 12–14, May 2003, paper no. 2003–3240.
[24] Smagorinsky J. General circulation experiments with the primitive equations.
Month Weath Rev 1963;93:99–165.