Professur für

Thermofluiddynamik

Master

Impact of LES Subgrid-Scale Modeling on the

Acoustic and Flow Properties for Helmholtz
Resonators Without Mean Flow

Autor:
QIN Yupeng

Matrikel-No:
03293504

Betreuer:
Killian Förner, M. Sc.
Prof. Wolfgang Polifke, Ph. D.

April 29, 2016

Professur für Thermofluiddynamik

Prof. Wolfgang Polifke, Ph. D.
Erklärung
Hiermit versichere ich, die vorliegende Arbeit selbstständig verfasst zu haben. Ich habe
keine anderen Quellen und Hilfsmittel als die angegebenen verwendet.

Ort, Datum QIN Yupeng

iv
 Abstract

The impact of LES Subgrid-Scale (SGS) modeling on both acoustic and fluid dynamic
properties is numerically investigated for Helmholtz Resonators without mean flow. Dif-
ferent available LES SGS models for compressible flow in OpenFoam are first examined
theoretically. Nonlinear behavior in impedance and reflection coefficient are utilized for
acoustic evaluation, while correlations and turbulent kinetic energy spectra, based on az-
imuthal velocity, are used for turbulent evaluation. The final comparison show very similar
results for all implemented LES SGS models in OpenFoam in both acoustic and fluid
dynamic properties. Only minor better conformance with experimental data is shown us-
ing the SpalartAllmaras model. The acoustic energy proportion absorbed in turbulence
decreases with the growth of sound pressure level. These conclusions are confirmed for
different Helmholtz Resonator geometries.
ii
Contents

Nomenclature v

1 Introduction 1

2 Theoretical Basis 3
2.1 LES Principle and Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.1.1 Governing Equations . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.1.2 Correlations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.3 Energy Spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.4 Filtered Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . 8
2.1.5 Subgrid-Scale Modeling . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.6 Wall Treatment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2 Basis Of Acoustics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.1 Acoustic Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.2 Helmholtz Resonator . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3 Numerical Setup 24
3.1 PIMPLE with NSCBC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2 Test Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.2.1 Hersh’s Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.2.2 Förner’s Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

4 Simulation Results 31
4.1 SGS Modeling Influence Of Acoustic Behavior . . . . . . . . . . . . . . . . 31
4.1.1 Correction SPL to AMP Transformation . . . . . . . . . . . . . . . 33
4.1.2 Impact of SGS models . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.1.3 Nonlinear Resistance With Respect To SPL . . . . . . . . . . . . . 37
4.2 Influence of The SGS Modeling on Turbulent Kinetic Energy Spectra and
Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.2.1 Azimuthal Velocity . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.2.2 Correlations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.2.3 Setting of Energy Spectra . . . . . . . . . . . . . . . . . . . . . . . 45
4.2.4 Turbulent Kinetic Energy Spectra . . . . . . . . . . . . . . . . . . . 48

iii
 4.2.5 Energy Spectra with different SPLs . . . . . . . . . . . . . . . . . . 50
4.2.6 Position and Mesh Quality Effect in Energy Spectra . . . . . . . . . 51
4.3 SGS Comparison In Other Criteria . . . . . . . . . . . . . . . . . . . . . . 55
4.3.1 Visualization Through Vorticity and Q-Criterion . . . . . . . . . . . 55
4.3.2 kRES /kSGS = 4 Criterion . . . . . . . . . . . . . . . . . . . . . . . . 56
4.4 Turbulence Judgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

5 Conclusion and Outlook 60

5.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.2 Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

Appendices 62

A Theoretical Deviations 63
A.1 Deviation of Helmholtz Resonator Equations . . . . . . . . . . . . . . . . . 63
A.2 The Definition Of ”pwelch” Function . . . . . . . . . . . . . . . . . . . . . 66
A.2.1 Windowed PSD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

iv
Nomenclature

Roman Symbols
Ē Filtered Total Energy [kg·m2 /s2 ]

Dw Windowed Power Spectra Density [m2 /s]

Im Reactance [-]

P Production of Residual Kinetic Energy [m2 /s3 ]

Rf Reflection Coefficient [-]

Re Resistance [-]

T Helmholtz Resonator Coefficient [-]

K Filter Kinetic Energy [m2 /s2 ]

P Filtered Pressure [Pa]

Qi Redistribution and Dissipation of Kinetic Energy in Filtered Velocity Field

[m3 /s3 ]

S ij Filtered Strain Tensor Rate [s−1 ]

sij Strain Tensor Rate of Residual Velocity [s−1 ]

U Filter Velocity filed [m/s]

qbi Filtered Heat Flux Vector [kg·K/m2 ·s]

Tb Filtered Temperature [K]

U
b Filtered Velocity [m/s]

∆
e Test Filter Width [m]

U
e Double Filtered Velocity [m/s]

Af Pressure Amplitude [-]

v
Am Orifice Area of Helmholtz Resonator [m2 ]

As Coefficient in Sutherland’s Law [kg/(ms·K0 .5)]

Aca Cavity Area of Helmholtz Resonator [m2 ]

BUθ Uθ Correlation [m2 /s2 ]

BUnorm
θ Uθ
Normalized Correlation [-]

c Sound Velocity [m/s]

Ck Kolmogorov Constant [-]

Cp Specific Heat at Constant Pressure [J/K·mol]

dt Distance to Nearest Trip Point [m]

E Total Energy [kg·m2 /s2 ]

E(κ) Turbulent Kinetic Energy in Each Wavenumber Space [m3 /s2 ]

Eij Dynamic Error Tensor [m2 /s2 ]

F Mass Force [N]

f Frequency [Hz]

Fw Windowed Frequency in Matlab [Hz]

fµ Damping Function [-]

Hw Windowed Discrete Fourier Transformed Signal [-]

K Kinetic Energy per unit Mass [m2 /s2 ]

kRES Resolved Filtered Kinetic Energy Spectra [m2 /s2 ]

kSGS Modeled Filtered Kinetic Energy Spectra [m2 /s2 ]

L Characteristic Length [m]

lm Orifice Length of Helmholtz Resonator [m]

Ln o Window Length Without Overlapping in Frequency Space [Hz]

Lo Window Overlapping Length in Frequency Space [Hz]

Lp Sound Pressure Level [dB]

Lw Window Length in Frequency Space [Hz]

Lij Leonard Stress [m2 /s2 ]

vi
M Mach Number [-]

Msgs Mach Number in Subgrid-Scale [-]

N Windowed Frequency Number [-]

P Pressure [Pa]

P? Fourier Transormed Pressure [-]

Pw Windowed Power Spectra Density in Matlab [m2 /s]

Pr Prandtle Number [-]

Q0i Redistribution and Dissipation of Kinetic Energy in Residual Velocity Field

[m3 /s3 ]

R Gas Constant [J/(K·mol)]

Re Reynolds Number [-]

s Entropy [J/K]

Sij Strain Tensor [s−1 ]

t Time [s]

Ts Coefficient in Sutherland’s Law [K]

Tw hole Time Last In Simulation [Hz]

Tij Eddy Viscosity of Sub-Test-Grid [m2 /s2 ]

U Velocity [m/s]

U0 Velocity Fluctuation [m/s]

0
Uacou Velocity Fluctuation due to Acoustics [m/s]
0
Uturb Velocity Fluctuation due to Turbulence [m/s]

U? Fourier Transormed Velocity [-]

Uθ Azimuthal Velocity [m/s]

Uθ,rms Velocity Square-root of Variance [m/s]

Vcell Volume of Grid [m3 ]

W Window Number in Matlab [-]

WL Window Overlapping Number in Matlab [-]

vii
x Displacement [m]

Z Impedance [-]

Greek Symbols
σ̄ij Filtered Viscous Stress Tensor [kg/(m·s2 )]

χ Viscosity Ratio in Spalart Allamras Model [-]

∆t Time Step [s]

∆ Filter Width [m]

δ Characteristic Flow Width [m]

δt Time Step between Two Time Poins [s]

δU Velocity Difference [m/s]

δv Viscous Length Scale [m]

δw Spacing along Wall to Trip Point [m]

δx Two-points Distance in x Direction [m]

∆cell Grid Length [m]

δij Kronecker Delta [-]

 Viscous Dissipation in Unresolved Scale [m2 /s3 ]

nl Resonator Loss Factor [-]

η Kolmogorov Motion Scale [m]

γ Specific Heat Ratio [-]

κ, κn wavenumber [-]

λ Thermal Conductivity [kg·m/(K·s3 )]

µ Dynamic Viscosity [kg/(m·s)]

ν Kinetic Viscosity [m2 /s]

ν? Working Variable of Kinetic Viscosity [m/s2 ]

ω Angular frequency [rad/s]

ωt Vorticity at Trip Point [s−1 ]

ωeig Resonator Eigenfrequency [Hz]

viii
 Viscous Dissipation in Resolved Scale [m2 /s−3 ]

ρ Filtered Density [kg/m3 ]

ρ Density [kg/m3 ]

τijd Residual Stress Tensor [m2 /s2 ]

τijr Residual Stress Tensor [m2 /s2 ]

τw Wall Shear Stress [kg/(m2 ·s)]

τij Stress Tensor [kg/(m·s2 )]

θmaco Macrotemperature [K]

ξmaco Macropressure [Pa]

ζ Loss Coefficient [-]

uτ Friction Velocity [m/s]

Acronyms
AMP Amplitude Of Pressure

CFD Computational Fluid Dynamics

CFL Courant–Friedrichs–Lewy Number

DFT Discrete Fourier Transform

DNS Direct Numeric Simulation

EPSD Empirical Power Spectral Density

GS Grid Scale

LES Large Eddy Simulation

NWM Near Wall Modeling

NWR Near Wall Resolution

PSD Power Spectral Density

SGS Subgrid-Scale

SPL Sound Pressure Level

TKE Turbulent Kinetic Energy

ix
1 Introduction

Helmholtz Resonators were first used by Greeks to produce reverberation in open-air the-
aters [1]. As modern usage, it was utilized in turbofan aircraft engines to muffle inlet and
exhaust noise [2]. The scientific interest in this resonator was thus stated as to understand
the basic sound energy dissipation as well as to optimize the geometry for its utilized
condition.
At lower sound pressure levels (SPL), acoustic energy is absorbed particularly by the
resonator geometry through reflection within the resonator cavity. At higher SPL, the
introduced acoustic perturbation promote flow separation especially at resonator necks,
which can finally turn to turbulence by increasing the SPL. In such cases, turbulence also
plays an important role in absorbing the input acoustic energy, which makes the resonator
more efficient. This leads to the question how to quantify the influence of these structures
in a Helmholtz Resonator.
Keller and Zauner [3] derived their mathematical model for Helmholtz Resonator from
general resonators, where impedance and reflection coefficient serves to estimate acoustic
properties. Hersh et al. [1] also provided a mathematical model from a basic force analysis,
which is extremely complex to express. However, his experiment for that model can be
used as referential experimental data. Förner did a lot of improvements [4–6] for that topic
in Computational Fluid Dynamics (CFD) with OpenFoam based on Keller and Hersh’s
works, but mainly in lower SPL and laminar conditions. For turbulence in higher SPLs,
Zhang and Bodony [7] and Davidson [8] provided turbulent kinetic energy spectra and
correlations as estimation tools for turbulent properties.
This thesis is concerned mainly with turbulent behavior in Helmholtz Resonator with
higher SPLs with the simulation tool OpenFoam. Large Eddy Simulation (LES) approach
is selected for those turbulence calculations, because it represents turbulence with high
resolution and low computational cost in comparison to the Direct Numerical Simulation
(DNS). As one of the most important parts in this work, different LES Subgrid-scales
(SGS) models, that can be used in OpenFoam, will be investigated in detail. The setup of
Helmholtz Resonator includes the part of acoustic theory, since the simulated turbulence
is initially caused by acoustic perturbation. Discussion of result is therefore consist of
acoustics and fluid turbulence estimations. For the evaluation of acoustic characteristics,
the impedance and reflection coefficient with respect to frequency and SPL are used. As
turbulent properties, correlations and TKE spectra will be used to judge SGS models both
spatially and temporally.
In this thesis two different Helmholtz Resonator geometries are considered. The simu-
lation according to these geometries are set up as in Hersh et al. [1] and Förner et al. [4],
such that the given measurement can be used as experimental criteria to compare with the
simulation results. These two geometries are used to confirm final conclusions, which are

1
 Introduction

targeted as follows:

• Compare the acoustic effect of different LES SGS models with the experimental result
and conclude the optional choice for high SPL resonator.

• Compare the fluid dynamic effect of different LES SGS models and make judgment
of the turbulence phenomenon.

• Discuss the advantage and disadvantage of each LES SGS model in simulation of
Helmholtz Resonator.

• Make the judgment, whether the observed flow separation at resonator is turbulence.

2
2 Theoretical Basis

2.1 LES Principle and Modeling

Smagorinsky [9] first formulated the basic LES equations in early 1960’s. Unable to resolve
all the scales of motion due to limited computational resources, only large scales could
be correctly resolved, and the impact of small scales bas to be modeled following the
theory of Kolmogorov’s theory [10]. According to his result, the smallest motion scales
were uniform and only drain energy from larger scales through the cascade process. These
scales are thus successfully approximated rather than to be costly calculated. The large
scales, which contain most of the energy, in the meanwhile, do most of transportation and
are significantly affected by the boundary conditions should be calculated directly. This is
the basic idea of LES.

2.1.1 Governing Equations

Fluid motion is governed by the Navier-Stokes Equations, which mainly describes the con-
servation of mass and momentum. In this thesis, Newtonian compressible fluid are consid-
ered. A newtonian fluid is defined as the fluid whose viscous stress is linear to the velocity
gradient vertical to the mean flow direction. A compressible fluid is defined as a fluid
whose density can vary with time and space. The continuity equation can be formulated
as follows [11]:
dρ ∂Ui
+ρ = 0, (2.1)
dt ∂xi
where ρ stands for density, t for time, x for displacement, and U for velocity. The momentum
transport equation presents the fluid particle mass times the acceleration equals the sum
of acted forces. It can be expressed as [11]:
 
dUi ∂P ∂τji ∂P ∂ 2 ∂Uk
ρ =− + + ρFi = − + 2µSij − µ δij + ρFi , (2.2)
dt ∂xi ∂xj ∂xi ∂xj 3 ∂xk

where P presents the pressure, F the mass forces, and µ dynamic viscosity. For Newtonian
viscous fluids, the stress tensor τij and the rate of strain tensor Sij are defined as:

2
τij = 2µSij − µSkk δij , (2.3)
3
 
1 ∂Ui ∂Uj
Sij = + . (2.4)
2 ∂xj ∂xi

3
 Theoretical Basis

Here, δij denotes the Kronecker delta,


1, if i=j
δij = (2.5)
0, otherwise,
The energy transport equation for a compressible flow can be formulated as [11]:
d(I + K) ∂σji Ui ∂qi
ρ = − + ρUi fi , (2.6)
dt ∂xj ∂xi
where I donates internal energy which is usually ignored without large temperature change,
and K = (Ui Ui )/2 the kinetic energy. The stress term σij includes both the pressure stress
and viscous stress:
2
σij = −P δij + 2µSij − µSkk δij . (2.7)
3

2.1.2 Correlations
Two-points Correlation
The two-points correlation is used to evaluate some turbulent length characteristics. The
spatial power spectral density can be generated from two-points correlation with help of
Fourier transformation. Take two points xA and xB with a distance of δx from one of the
coordinate directions like x. The two-points correlation of velocity in azimuthal direction
Uθ0 can be written as [11]:
BUθ Uθ (xA , δx ) = hUθ0 (xA )Uθ0 (xB )i = hUθ0 (xA )Uθ0 (xA + δx )i, (2.8)
with h·i denoting an appropriate ensemble average. The normalized two-points correlation
can be expressed like:
hUθ0 (xA )Uθ0 (xA + δx )i
BUnorm
θ Uθ
(xA , δx ) = , (2.9)
Uθ,rms (xA )Uθ,rms (xA + δx )
where the index rms stands for square-root of variance:
Uθ,rms = hUθ02 i1/2 . (2.10)
By homogeneous direction, the two-points correlation does not depend on the distance be-
0 0
tween the two selected points instead of the location of xA . Thus Uθ,rms (xA ) = Uθ,rms (xA +
0
δx ) = Uθ,rms , Eq. (2.9) can be reduced to:
hUθ0 (xA )Uθ0 (xA + δx )i
BUnorm
θ Uθ
(xA , δx ) = 2
. (2.11)
Uθ,rms
Obviously, the correlation BUθ Uθ increases when selected points are close to each other,
and decrease when the distance become huge, see Fig. 2.1. The correlation BUθ Uθ tends
to zero when the distance between two points is large enough, such that the point xA
including eddies have no more influence of the velocity of point xB . With negative influence
in velocity, the correlation can also be smaller as zero. Thus, the normalized BUnorm
θ Uθ
varies
in a range of −1 to +1 with the distance between selected points as Fig. 2.1 shows.

4
2.1 LES Principle and Modeling

δx

Figure 2.1: Sketch of typical two-points correlation.

Auto Correlation
The auto-correlation can be computed in only one location with different time steps. Similar
as the derivation of two-points correlation, chose a two instances in time tA and tB with time
step δt in between. The auto correlation of the velocity fluctuation in azimuthal direction
Uθ0 presents:
BUθ Uθ (tA , δt ) = hUθ0 (tA )Uθ0 (tB )i = hUθ0 (tA )Uθ0 (tA + δt )i. (2.12)
In the situation of constant mean velocity, the homogeneous will also exist in ”time direc-
tion”, BUθ Uθ will therefore no dependend on time points, but only on time step δt . Thus a
similar normalized auto-correlation over time can be formulated as:
hUθ0 (t)Uθ0 (t + δt )i
BUnorm
θ Uθ
(δ t ) = 2
. (2.13)
Uθ,rms

2.1.3 Energy Spectra

Since a wide range of scales can be observed in turbulent fluctuations, it is easier to evaluate
the turbulence mechanisms with help of Fourier series. The presented discussion strictly
follows Davidson [11]. With a general Fourier series, periodic function g with a period of
2L can be formulated as [11]:
∞
1 X
g(x) = a0 + (an cos(κn x) + bn sin(κn x)) , (2.14)
2 n=1

where x acts as spatial coordinate and the Fourier coefficients are defined as:
1 L
Z
an = g(x) cos(κn x)dx,
L −L
1 L
Z
bn = g(x) sin(κn x)dx,
L −L
(2.15)

5
 Theoretical Basis

I
II

E(κ
) ∝κ
E(κ)

-5/3
III

κ
Figure 2.2: Spectrum for turbulent kinetic energy (TKE). I: Range for large energy
eddies. II: The inertial subrange. III: Range for small, isotropic scales

where n reads the nth harmonic wavenumber κn , with the definition:

nπ 2π
κn = or κ = . (2.16)
L L
The largest wavenumber, that can be resolved is determined by the two grid lengths:
2π/(2∆cell ). And the smallest wavenumber comes from the largest geometries scale, like in
this thesis the resonator cavity length lca , which then be defined as: 2π/lca .
The similar wavenumber transform is also available by analysis of turbulent kinetic
energy Ui0 Ui0 /2, which is more convenient to study in all eddy sizes in wave number space.
The energy of each eddy in wavenumber can be expressed as:

E(κ)dκ, (2.17)

where E(κ) denotes the turbulent kinetic Energy by each wavenumber space. Thus the
wavenumber can be treated as proportional to the inverse of an eddy’s diameter. And
a total turbulent kinetic energy turns to be the integrating of Eq. (2.17) of the whole
wavenumber space:
Z ∞ X∞
K= E(κ)dκ = (a2n + b2n ). (2.18)
0 n=0

The distribution of turbulence kinetic energy in regards of its wave number is also called
turbulent kinetic energy spectrum, as presented in Fig. 2.2. It can be separated to three
regions: I, II, III:

I Region of large eddies that includes most of the energy. In this region the energy
transfers from mean flow kinetic energy (in the situation of this thesis acoustic energy)

6
2.1 LES Principle and Modeling

to turbulent kinetic energy. This part of energy is called production term P and
defined in Eq. (2.40).
III Dissipation range. It is where the dissipation take place with small and isotropic
eddies. At this part the turbulent kinetic energy transfers to inner energy with a
governed parameter of , which is defined in Eq. (2.43).
II Inertial subrange. In this region, the middle sized eddy presents also as isotropic.
Production from region I will be transfered to the dissipation part in region III
through a transition process in this region. Thus a energy equivalence reads: P = .
As larger eddies size in this region connect those from region I, while the smaller
sizes connect those from region III, it can be imagine that turbulent kinetic energy
in this region can be characterized by wavenumber κ and dissipation rate .

E = κa b
[m3 /s2 ] = [1/m] [m2 /s3 ] (2.19)

A simple dimensional analysis gives b = 2/3 and a = −5/3. so that

2 5
E(κ) = CK  3 κ− 3 , (2.20)

where the CK is called Kolmogorov constant CK ≈ 1.5. Eq. (2.20) is always called
Kolmogorov spectrum law or simply −5/3 law. And the -5/3 decay in inertial region
acts as a very important criterion in evaluating turbulent behaviors.
Alike the derivation in wavenumber space, Fourier series can be expressed in frequency
space. Consider a periodic function h with a fundamental period of T and fundamental
frequency of f = 1/T . The periodic function can be expressed as:
∞
X
h(t) = cn ei2πnf t (2.21)
n=−∞

where n also reads the nth harmonic of fundamental frequency f0 . The coefficient cn can
be computed as: Z
1
cn = h(t)e−i2πnf t dt. (2.22)
T0 T0
Thus, the turbulent kinetic energy in frequency space can be computed as:
Z ∞ X∞ ∞
X
2
k= E(t)dt = |cn | = D(nf0 ). (2.23)
−∞ −∞ n=−∞

D(nf ) is the power Spectrum of h(t). It is a series of coefficient, with each of them describes
the contribution of the nth harmonic fundamental frequency f . Since the connection of wave
between space and time domain: wavenumber and frequency is proportional to each other:
f ∝ κ, the Kolmogorov’s −5/3 law is also available in frequency space.
There are two ways to calculate the TKE spectrum from simulated velocity data in
both wavenumber and frequency space:

7
 Theoretical Basis

• Calculate the discrete Fourier transform (like ”fft” in Matlab) of every single velocity
at the investigated points, and get the two-points covariant of these transformed
velocities, which is actually twice the value of TKE.
• Calculate the two-points covariant of real velocities the BUθ Uθ like in Sec. 2.1.2 , and
perform the discrete Fourier transform on this correlation. Still the final TKE value
should be half of the transformed value.
Overmore the function of ”pwelch” is usually preferred to calculate the power spectrum
density (PSD) instead of discrete Fourier transform with ”fft” function, since automatic
data optimization is also made through the function of ”pwelch” (detailed discussion in
Sec. 4.2.3), so that the presentation of TKE shows less fluctuation in curves like ”fft”
function does.

2.1.4 Filtered Navier-Stokes Equations

In Direct Numeric Simulation (DNS), governing equations are computed down to the Kol-
mogorov motion scale η, which is the smallest scale of motion. However, in LES a low-pass
filtering operation, which relates mainly to spatial grid properties is performed. The filter
width ∆ becomes thus a significant parameter and separates the whole flow into larger Grid
Scale (GS) and the smaller Sub-Grid Scale (SGS). As its scale is larger than the defined
filter width, the GS motion is resolved precisely, but the impact of SGS on GS has to be
modeled as well. In the meanwhile with motion scale smaller than the filter width, SGS
can be modeled without direct computation.
As mentioned above, every variable in LES consist of GS and SGS parts:
f¯ = f − f 0 . (2.24)
|{z} |{z}
GS SGS

The general filtering operation can be used to exactly define the GS:
Z
¯
f (x, t) = G(x, x0 , ∆)f (x0 , t) dx0 , (2.25)

where the filter function G(x, x0 , ∆) must satisfies the normalization condition:
Z
G(x, x0 , ∆) dx0 = 1. (2.26)

Figure 2.3 gives a comparison between filtered and unfiltered motion properties in both
spacial space and spectral space.
Some of the most common filters include:
• The Top-Hat Filter, with its definition also in real space:
, if |x0 | 6 ∆2
 1
G(x, ∆) = δ (2.27)
0, otherwise,
where δ presents characteristic flow width. This is the most common choice for finite
volume method, because the averages of flow variables over finite volume grid are
linear functions of x, with the filter width of ∆

8
2.1 LES Principle and Modeling

Figure 2.3: The filtered signal in spacial space (left) and spectral space (right) [12].

• The Sharp Fourier Cutoff Filter, with its definition in Fourier space:
 1 π
δ
, if κ 6 ∆
G(x,
b ∆) = (2.28)
0, otherwise,

where Gb donate the Fourier coefficient of the filter function and κ the wavenumber.
This model eliminates all the wave numbers above a chosen frequency, and hardly
be applied to inhomogeneous flows. As it also does not smooth the resolved scales
as the former two do, the Sharp Fourier Cutoff Filter is usually used in conjunction
with spectral methods [13].

In smooth filter like the Top-Hat filter, the boundary between resolved scale and un-
resolved scale is not clearly defined, which results the SGS stress presenting the influence
from both the resolved part and the unresolved part. In the meanwhile, the sharp Fourier
cutoff filter serves as a clear separation between unresolved scale and resolved scale, so that
the SGS stress presents only the effect from smaller scale on large ones. In this thesis a
smooth filter (which can be defined in OpenFoam) is preferred for a relatively homogeneous
calculated volume.

Filtered Incompressible Equations

At first, the filtered incompressible Navier-Stokes equations will be derived, with the ig-
noring of filter effects also in density ρ and pressure P . Only the velocity is considered.
Thus, if we apply the filter operation of Eq. (2.24) on the velocity field, it can be defined
as:
U (x, t) = U (x, t) + u0 (x, t), (2.29)
where U (x, t) denotes the filtered velocity part and u0 (x, t) the residual (Sub-Grid Scale)
part. In order to obtain the governing equations of filtered velocity field, the filter operation

9
 Theoretical Basis

should be applied on the Navier-Stokes Equation [12]. The continuity equation Eq. (2.1)
can be filtered as:
∂U i
= 0. (2.30)
∂xi
The filtered momentum equation can be rewrite from Eq. (2.2) as:
∂U j ∂Ui Uj ∂ 2U j 1 ∂P
+ =ν 2
− . (2.31)
∂xi ∂xi ∂xi ρ ∂xj
It is quite important to distinguish between the filtered product Ui Uj and the product
of filtered velocities U i U j here. The difference between these two is defined as residual
stress tensor [12]:
τijr = Ui Uj − U i U j . (2.32)
The residual stress tensor is related to the residual (Sub-Grid Scale) kinetic energy kSGS
as τiir = 2kSGS . The deviatoric stress tensor τijd can be derived from residual stress tensor
as:
2
τijd = τijr − kSGS δij , (2.33)
3
In tensor analysis, tensor Aij can be write as the summation of deviatoric part and
hydrostatic (residual) part as Aij = Adij + Arij , where the hydrostatic part is defined as
Arij = 31 Aii δij , and so the deviatoric part as Adij = Aij − 13 Aii δij [14]. In order to include the
hydrostatic residual stress into filtered momentum equation, a modified filtered pressure is
defined as P̄ = P + 23 kSGS . So that the filtered momentum equation formulates as:

∂U j ∂U j ∂ 2U j ∂τijr 1 ∂ P̄
+ Ui =ν 2
− − . (2.34)
∂t ∂xi ∂xi ∂xi ρ ∂xj
By comparison of the non-filtered (2.2) and the filtered (2.34) momentum equations, an
extra unknown quantity τijr is observed, that makes the whole equation a not closed one.
It therefore must be modeled for further computation. In this thesis, it will be discussed
how this term can be modeled with different methods. In deviation of filtered energy
conservation equation, kinetic energy K can be first filtered as:
1
K = Ui Ui = kRES + kSGS , (2.35)
2
where kRES and kSGS present the resolved part and Sub-Grid part of kinetic energy re-
spectively. They can be defined as:
1

kRES = U iU i , (2.36)
2
1
1
kSGS = K − kRES = Ui Ui − U i U i = τiir . (2.37)
2 2
By combination of Eqs. (2.6), (2.36) and (2.37), the energy conservation equations for kGS
and kSGS can be deduced as [12]:
∂kRES ∂kRES ∂Q
+ Ui = − i − P − , (2.38)
∂t ∂xi ∂xi

10
2.1 LES Principle and Modeling

∂kSGS ∂kSGS ∂Q0i

+ Ui =− + P − . (2.39)
∂t ∂xi ∂xi
Here, P presents the production of residual kinetic energy:

P = −τijd S ij , (2.40)

where S ij is the filtered strain tensor rate:

 
1 ∂U i ∂U j
S ij = + . (2.41)
2 ∂xj ∂xi

The production term acts as a sink term in resolved kinetic energy conservation (2.38),
and as a source term in unresolved kinetic energy conservation (2.39). Because the energy
in this term is transferred from the large scale (filtered and resolved velocities) to the small
scales (residual and unresolved velocities) [12]. The viscous dissipation terms are defined
respectively as:
 = 2νS ij S ij , (2.42)

 = 2νsij sij , (2.43)

where sij is the filtered strain tensor rate of residual velocities:

!
1 ∂u0i ∂u0j
sij = + . (2.44)
2 ∂xj ∂xi

The term , the viscous dissipation directly from the filtered velocity field is relatively small
compared with  the viscous dissipation from the residual velocity field. This means that
most energy is dissipated in the sub-grid scales [12]. The term Qi and Q0i represent the
redistribution and the dissipation of kinetic energy in filtered and residual velocity fields:
 
d P̄
Qi = U j 2νS ij − τij − δij , (2.45)
ρ

 
P̄
Q0i = U j 2νsij − τijd − δij , (2.46)
ρ
where the first terms represent diffusion while the other two represent energy redistribution
by velocity and pressure fluctuations.
As a criterion for the LES with near-wall resolution, 80% of the kinetic energy should
be resolved while other 20% be modeled, which means kRES /kSGS > 4 [15]. To fulfill this
requirement the grid width should not be too coarse to resolve 80% energy, otherwise the
simulation is more strongly dependent on the SGS modeling. That topic will be discussed
deeply later in Sec. 2.1.5.

11
 Theoretical Basis

Compressible Filtered Equations

With the above preparation of LES definition in incompressible flows, we now consider more
about the compressible flows. The main difference between compressible and incompressible
problem is the variable density. Especially, when we consider this variation in filtered set
of Navier-Stokes equations, three preliminary choice must be made [16]:
• The original set of unfiltered variables or equations.
• The filter.
• The filtered variables.
The first problem arises for huge numbers of formulations for the compressible Navier-
Stokes equations exist rely on different choice of basic variables to describe the flow. Ve-
locity is for example usually described with velocity and momentum. However, energy
conservation can be expressed using two independent variables among: density, entropy,
total energy, internal energy, enthalpy and so on. The difference of filter choice is already
discussed in Section 2.1.4. The last point is actually the most difficult and important. If
we consider the set of variables like (ρ, ρU, E), where E is the total energy, the direct
application of filtering procedure from incompressible flows leads to the filtered variables
(ρ, ρU , E). The term ρU can be rewritten as:
ρU ≡ ρU
b, (2.47)
where U b = ρU /ρ is also filtered velocity. b· only represents a change of variable. The
problem of basic variables choice arises because ρU , E, and P are nonlinear functions of
other variables which can be decomposed. Thus different sub-grid terms will arise from
the filtered equations as well as adding needs for specific Subgrid-modeling. Usually two
systems are used to define the filtered problem:
• System I: It was suggested by Vreman [17], that with the basic filtered variables
ρ, ρU
b , Ē, the macropresure ξ and the macrotemperature θ can be defined as:
1 3γ − 5
ξmaco = P + τkk = ρRθ − τkk , (2.48)
3 6
γ(γ − 1)M 2 τkk
θmaco = Tb + ρ, (2.49)
2
2 2
where M is Mach number. Since τkk is defined as τkk = γMsgs P , and Msgs repre-
sents the sub-grid scale Mach number, which is certainly very small, some authors
suggested that this term can be neglected.
• System II: Vreman [17] also imposed another system, in which the selected variables
are (ρ, U
b , Ē) The synthetic total energy is defined as:

P b2
Ē = + ρU (2.50)
γ−1
and correspond to the computational total energy [16]. With this system no sub-grid
contribution appears in the filtered equation of state.

12
2.1 LES Principle and Modeling

The choice of system for governing equations can change the closure problem by mod-
ifying the sub-grid terms. For application used in aeroacoustics and aero-optics system II
is preferred as direct pressure and temperature can be derived. But system I is simpler for
the energy equation, and is simpler implemented.
Using System II, and the similar method in derivation of incompressible filtered Navier-
Stokes equations, the compressible Navier-Stokes equations (Eqs. (2.1), (2.2), (2.6)) can
be filtered as:
∂ρ ∂(ρU cj )
+ = 0, (2.51)
∂t ∂xj
∂ρU
cj ∂(ρU
bi U
bj ) ∂P ∂ σ̄ij ∂ρτij
+ + − = , (2.52)
∂t ∂xj ∂xi ∂xj ∂xj
∂ Ē ∂((Ē + P )U
bj ) ∂(σ̄ij U
bj ) ∂b
qi
+ − + = −α − β − π + ε. (2.53)
∂t ∂xj ∂xj ∂xj
In the above Equations, the viscous stress tensor is defined as:

µ(Tb)
σ̄ij = S̄ij , (2.54)
Re

∂U
bi ∂ Ubj 2 ∂U bk
S̄ij = + − δij . (2.55)
∂xj ∂xi 3 ∂xk
The dynamic viscosity µ(Tb) can be derived from the Sutherland’s law of air. While the
filtered heat flux vector is defined as:

µ(Tb) ∂ Tb
q̄j = , (2.56)
(γ − 1)Re Pr M 2 ∂xj

where the Reynolds number Re and Prandtle number Pr are defined as:
ρUj L Cp µ
Re = , Pr = , (2.57)
µ λ
with L and λ represent characteristic length and thermal conductivity, respectively. The
filtered temperature Tb comes from the state equation:

P
Tb = γM 2 . (2.58)
ρ

The right-hand sides of Eqs. (2.52) and (2.53) contain sub-grid terms, that describe the
effect of unresolved sub-grid scales. Especially the terms of filtered energy equation can be
formulated as [18]:
α=U bi ρτij , (2.59)
xj
∂ h  i
β= P Uj − P Uj /(γ − 1) ,
b (2.60)
∂xj

13
 Theoretical Basis

∂Uj ∂U
bj
π=P −P , (2.61)
∂xj ∂xj
∂Ui ∂U
bi
ε = σij − σ̄ij . (2.62)
∂xj ∂xj

2.1.5 Subgrid-Scale Modeling

By LES varies models are used in order to make the simulation more accurate and efficient
in different geometries, flow and boundary conditions. In OpenFoam version 2.3.1, with the
limiting of compressible LES modeling, there are totally five available SGS models, which
will be discussed in following sections:
• Smagorinksy model

• One equation eddy model

• Homogeneous dynamic one equation eddy model

• Deardorff’s Model

• Spalart-Allmaras Model

Algebraic Smagorinsky Model

Algebraic Smagorinsky model [9] is the basic and simplest LES model which can be pro-
duced with the following steps:
1
τij − δij τkk = −2νSGS S ij , (2.63)
3
1 ∂U i ∂U j
S ij = ( + ), (2.64)
2 ∂xj ∂xi
νSGS = (CS ∆)2 |S|, (2.65)
∂U i ∂U j
where the scalar |S| is the norm of ∂xj
+ ∂xi
in the boussinesq assumption. It can be
presented as: q
|S| = 2S ij S ij , (2.66)
Parameter Cs , the Smagorinsky Constant varies with the isotropic turbulence decay in a
range between 0.18 and 0.23. ∆ is defined as the filter-width which takes the local grid
size:
∆ = (Vcell )1/3 . (2.67)
Close to the boundary condition , the SGS viscosity becomes huge for the large velocity
gradient. While the SGS turbulent fluctuations near wall go to zero, so as the SGS viscosity.
A Van driest damping function fµ is thus desirable:

fµ = 1 − exp −x+


2 /26 . (2.68)

14
2.1 LES Principle and Modeling

In OpenFoam, the damping function, which is only needed by Smagorinsky Model,

can be defined in the file ”LESProperties”. As a simple LES model, it has the lowest
calculation cost of all the models. However, the main disadvantage of this model related
to the variation of the defined Smagorinsky Constant Cs .

One Equation Eddy Model

Most one equation models are solving one equation for their own subgrid-scale quantity,
which also based on the eddy-viscosity concept. One such typical scale like kinetic energy
kSGS , is defined as the subtracting of kinetic energy and the resolved kinetic energy:
1
kSGS = K − kRES = τii . (2.69)
2
The transport equation for SGS turbulent energy can be derived by first subtracting the
filtered equations of motion from their resolved parts to give a relation to the fluctuating
component of velocity u0 . And then multiply the result by the sub-grid velocity vector to
give the equation, which was first formulated by Yoshizawa [19] as:
 
∂kSGS ∂ ∂ ∂kSGS 1 3/2
+ (uj kSGS ) = (ν + νSGS ) + 2νSGS S ij S ij − Ce kSGS , (2.70)
∂t ∂xj ∂xj ∂xj ∆
1/2
νSGS = Ck kSGS ∆. (2.71)
The production term 2νSGS S ij S ij in this equation is equivalent to the SGS dissipation
in the equation for the resolved turbulent kinetic energy kRES . Default values of model
coefficients Ce and Ck are given in OpenFoam as Ce = 1.048 and Ck = 0.094.
Actually under ideal equilibrium circumstances one equation models performs no better
than the Smagorinksy model [13]. They all have the main deficiency of not addressing the
discrepancy between the principal axes of the SGS stress and the rate of strain tensor,
which is common to most eddy viscosity models. However, one equation model provides
a more accurate time scale which is independent of velocity scale definition. This advan-
tage glittered by modeling transitional flows or flows with large scale unsteadiness like
Fureby [20] shows in channel flows.

Dynamic Models
As an improvement of Smagorinsky model, dynamic model was first formulated by Ger-
mano [21]. In his procedure, coefficients of SGS model should also be parts of calculation
from the energy of smallest resolved scales. The sub-grid scale model in grid filter:

τijr = Ui Uj − U i U j . (2.72)

Test filter with a width ∆

e larger than the original filter-width ∆ (typically ∆
e = 2∆).
Apply it to the filtered N-S equation, the in ”test filter” modeled sub-grid stress can be
written as:
]
Tijr = Ui Uj − U i U j . (2.73)
e e

15
 Theoretical Basis

The resolved part of SGS stress at the test filter, known as the Leonard stress can be
computed from LES fields with following relation:

]
Lij = U iU j − U iU j , (2.74)
e e

with above definition:

Lij = Tijr − τfr
ij . (2.75)
Eq. (2.75), known as Germano identity, forms the basis of dynamic models. This identity
was used by Piomelli [22] to derive model coefficients as shown in the following steps.
First parameters αij and βij were used to model generalized eddy-viscosity of subgrid and
sub-test-grid in the deviatoric part.

τijr = −2Ce αij , Tijr = −2Ce βij . (2.76)

Five independent equations are thus used to derive the single coefficient, which caused over-
determining problem. Based on the least-squares method Lilly [23] attempted to remedy
this problem with the help of minimizing an error, which was produced by two approxi-
mation Eqs. (2.75) and (2.76). The error can be formed as:

Eij = Ldij − Tijr − τfr d

ij = Lij + 2Ce Mij , (2.77)

where Mij = βij − αfij . The constant Cs is assumed to be smooth on the scale and can be
derived from the the filtering operation with the least-squares minimization:
 
∂hEij Eij i ∂Eij
= 2 Eij = 0, (2.78)
∂Ce ∂Ce

Combine Eqs. (2.77) and (2.78):

h(Ldij + 2Ce Mij )Mij i = 0, (2.79)

where Ce can be finally derived as:

1 hLdij Mij i
Ce = − . (2.80)
2 hMij Mij i
These model coefficients vary with time and space, and invalidate the original value,
which was assumed constant at first and then removed from the filtering operation. Some
stability problems are caused by the removing process. Integral formulation of the identity
Eq. (2.74) was introduced by Ghosal [24] to remove the inconsistency. The total viscosity
(molecular and SGS) must larger as zero for his improvement. The solution of this integral
through the use of least-square method leads to a constrained optimization problem, which
is known as the dynamic localization model. This model is reported to produce good results
with the cost of calculation time.
As the advantage of dynamic model, many problem with traditional ”static” methods,
especially the sub-grid scale modeling were successfully resolved. It adjusts the free shear
and channel flows by lowing the coefficient in areas of high shear and near the walls, where

16
2.1 LES Principle and Modeling

van Driest damping was used by other methods. In laminar part will the eddy viscosity
automatically be changed to zero.
As disadvantage,in order to avoide excessive fluctuations in the model coefficients, ad-
ditional averaging for all the homogeneous directions, along streamlines or in local regions
of the flow should also be iterated. Even the local average requires local homogeneous flow
or large averaging volume to smooth the model coefficients. Also, de Villar [13] reported
that some negative dissipation randomly occur in this procedure unless the scheme is some-
how artificially bounded. Piomelli [25] and Germano [21] interpret this as energy transport
from unresolved to resolved scale. However, it can also violate the conservation of energy
conservation and lead to unphysical results if unchecked.

Deardorff ’s Model
A major weakness for all the eddy-viscosity based models above is the assumption of
isotropy in the unresolved scales. This is true for only a large subset of flows and filters,
which actually unusual in practice. If the mesh spacing is too large, the unresolved motion
may be anisotropic. Even the smallest scales could not satisfy the isotropy assumption by
solid boundaries. In these cases, anisotropic grids will resolve isotropic eddies differently
with their own orientation. Further refinement of the mesh can reduce this problem, but a
better result can only be achieved if abandon the assumptions of isotropy.
By ignoring isotropy assumptions, the equation that is similar to the one-equation
turbulent energy model (Eq. (2.70)) can be formulated with the complete SGS stress tensor
τij (in OpenFoam defined in the file ”B”). The first model was derived by Deardorff [26]
and then developed by Fureby et al. [27] as the following:

∂τij ∂
+ (τij U i ) = 2νSGS S ij S ij +
∂t ∂xi
 
∂ ∂τij 2 1 1/2 2 1/2
(ν + νSGS ) + kSGS S ij − cM kSGS τij − Ce kSGS , (2.81)
∂x i ∂xj 5 {z∆ } |3∆ {z }
| {z } |
M Π E

where
νSGS = ck ∆K 1/2 , (2.82)

ck = 0.07, cM = 4.15, c = 1.05. (2.83)

In right hand-side of Eq. (2.81): M is an approximation for the triple correlation tensor,
Π models the pressure velocity-gradient and the dissipation tensor is approximated by
the E term, which is assumed to be isotropic. Fureby et al. [27] also proved that the
differential stress model produces backscatter similar to the linear combination model with
enough fine mesh, but with the inter-scale transfer limited by the transported SGS energy.
With decreasing spatial resolution, the modeled energy transfer behaves more like an eddy
viscosity model.
Since the stress tensor τ by Eq. (2.81) is symmetric (τij = τji ), only 6 tensor components
are needed to be defined in file ”B” for the simulation. But it is still more expensive as the

17
 Theoretical Basis

algebraic or one equation models. However, it costs only a small part of the overall calcula-
tion in consideration of the resolved pressure momentum calculation. The major advantage
of this model is proven to be his good performing in big range of flow conditions [27].

Spalart-Allmaras Model
The Spalart-Allmaras Model [28] is a one equation model which has been derived largely
with the help of empiricism. It gives the proper behavior in 2-D mixing layers, wakes, and
flat-plate boundary layers. A field equation for the working variable ν ? can be formulated
as:
  ? 2
∂ν ? ∂ν ?

? Cb1 ν
+ Uj = Cb1 (1 − ft2 )Sν − Cw1 fw − 2 ft2
∂t ∂xj κ0 d
? ? ?
   
1 ∂ ∂ν ∂ν ∂ν
+ (ν + ν ? ) + Cb2 , (2.84)
σ ∂xj ∂xj ∂xi ∂xi

the left-hand side represents advection along the particle path while the terms on the
right-hand side represent diffusion, anti-diffusion, production, destruction and a trip term,
respectively [29]. Boundary condition for the model are specified as: ν ? = 0 on solid walls;
ν ? = ν∞?
< 0.1ν∞ along inlet (u · n < 0) and ν ? = 0.1ν∞ as interior values along outlet.
This results in the low value of µSGS∞ = 2.79 × 10−7 µ∞ in the free stream. An initial
condition can be taken as ν ? = ν∞ ?
everywhere. Some coefficients above are defined using
the following functions:
ν?
χ= (2.85)
ν
χ3
fv1 = 3 (2.86)
χ + cv1 3
 −3
χ
fv2 = 1 + (2.87)
cv2
(1 + χfv1 )(1 − fv2 )
fv3 = (2.88)
χ
ν?
S ? = |Sij |fv3 + fv2 (2.89)
κ20 d2
where d is the distance to the closest wall. Since all these functions are also well calculated
as χ → 0, one can replace χ with max(χ, 0.001) to prevent division by zero in Eq. (2.88).
The destruction term is constructed with the aid of the following functions:
ν?
r= , (2.90)
S ? κ20 d2

g = r + cw2 (r6 − r), (2.91)

1/6
1 + c6w3

fw = g . (2.92)
g 6 + c6w3

18
2.1 LES Principle and Modeling

Function ft2 is used as a damper into the production and destruction terms in order to
make ν ? = 0 still a stable solution.It is formulated as:

ft2 = ct3 exp −ct4 χ2 .

(2.93)

The strip term ft1 δU U 2 allows one to specify explicitly the boundary transition locations.
The function ft1 is given by:
ωt 2 2
 
2 2
ft1 = ct1 gt exp −ct2 (d + gt )dt , (2.94)
∆U 2
where dt is the distance to the nearest trip point, ωt is the vorticity at the wall at the trip
point, δU is the norm of the difference between the velocity at the trip point and the field
point under consideration, and gt = min(0.1, δU /ωt δw ) where δw is the spacing along wall
at the trip point. With this help, the eddy viscosity is given by:

νSGS = ν ? fv1 . (2.95)

As fw stays almost constant for large g number, r can be replaced by min(r, 2) in order to
prevent overflow in Eq. (2.92), all the constants are defined as follows:
cb1 = 0.1355, cb2 = 0.622,
σ = 2/3, κ0 = 0.41,
cw1 = κ2 + 1+c
cb1
σ
b2
, cw2 = 0.3, cw3 = 2,
cv1 = 7.1, cv2 = 5,
ct1 = 1, ct2 = 2, ct3 = 1.2, ct4 = 0.5.
In the original version S ? could also be negative with the definition, which could then
disturb the value of r and result in stalled convergence. In the above improved version,
functions fv3 has been introduced an definition of fv2 has been changed, ensuring S ? to be
non-negative. It can however be zero, so that a protection against division by zero when
calculation r is needed.

2.1.6 Wall Treatment

Compared with the overall flow fields, the viscous effect become significant in the thin
near-wall sublayer. Since all terms like production, dissipation, kinetic energy, Reynolds
stress achieve their peak values when y + < 201 . In LES, generally there are two ways to
deal with the boundary condition problem [12]:
• LES-NWR (near wall resolution) → All parts in computational domain are resolved.
The grid spacing and filter width should be the same order as viscous length scale δv
which defined as: r
ρ ν
δv ≡ ν = , (2.96)
τw uτ
1 + yuτ
y is wall coordinate, that is defined as y + = ν with y the distance to wall and uτ the friction
velocity

19
 Theoretical Basis

where τw and uτ are named wall shear stress and friction velocity, respectively. There-
fore it is not suitable for high-Reynolds-number flows, for a high grid spacing demand
and simulation cost.

• LES-NWM (near wall modeling) → The near wall region be modeled.

The near wall progress is not resolved but modeled. It is usually used by high-
Reynolds-number flow, where the SGS model coefficient at boundary shows only low
accuracy. Transforms like Van Driest damper (Eq. (2.68)) turn out to be suitable.
Different kinds of wall functions can also be chose (even in OpenFoam) to model the
fluid profile in the boundary nearest grid point. Dynamic model can do even better
by iterating those coefficients within the simulation progress.

In this thesis, the Re number will respect to the velocity amplitude, which is very moderate.
Nevertheless, it can be observed that turbulence structures develop due to the pulsating
nature of the flow. Thus, for a more accuracy simulation first way (LES-NWR) is therefore
preferred. Accordingly, the grid spacing at the walls must be fine enough.

20
2.2 Basis Of Acoustics

2.2 Basis Of Acoustics

2.2.1 Acoustic Equations
With the assumption of zero mean flow velocity, i.e. U0 = 0, acoustic quantities P , ρ and
U can be separated into mean part (with index0 ) and fluctuation part (with superscript 0 )
as:
P = P0 + P 0 ,
ρ = ρ0 + ρ0 , (2.97)
U = U0 + U 0 .

Consider now the perfect gas. Set Eq. (2.97) in the continuity (2.1) and momentum
transport equations (2.2), and neglect the products of small quantities:

∂ρ0
+ ρ∇ · U 0 = 0, (2.98)
∂t
∂U 0
ρ + ∇P 0 = 0. (2.99)
∂t
Use the definition of sound speed c and Taylor expansion of P around P0 with respect to
ρ at constant entropy s yields:
 
∂P0  
2
P = P0 + ρ + O ρ0 .
0
(2.100)
∂ρ s
 
2 ∂P0
c ≡ , (2.101)
∂ρ s

The first-order approximation for pressure fluctuation can be formed as:

P 0 = c2 ρ 0 , (2.102)

By inserting Eq. (2.98) into Eq. (2.102), it can be rewritten as:

1 ∂P 0
+ ρ∇ · U 0 = 0. (2.103)
c2 ∂t
The time differentiating of Eq. (2.103) subtract the divergence of Eq. (2.99) yields the wave
equation respect to the pressure disturbance:

∂ 2P 0
− c2 ∇2 P 0 = 0. (2.104)
∂t2
In one dimension, wave equation can be factorized as:
  
∂ ∂ ∂ ∂
+c −c P 0 = 0. (2.105)
∂t ∂x ∂t ∂x

21
 Theoretical Basis

The general solution of the equation above can be written as:

P 0 (x, t)
= f (x − ct) + g(x + ct), (2.106)
ρc
where f and g describes the wave moving in positive and negative direction. In the one
dimensional cases, the so-called Riemann invariants f , g as the solution of Eq. (2.99), can
be expressed in terms of P 0 and U 0 :
1 P0
 
0
f= +U , (2.107)
2 ρ0 c
1 P0
 
0
g= −U . (2.108)
2 ρ0 c

2.2.2 Helmholtz Resonator

The Helmholtz Resonator consists of a cavity room and backing volume with the connection
of a orifice (see Fig. 3.2). Due to variable geometry, sound pressure level (SPL), the flow
in and out of resonator could be laminar or turbulent. A derivation of the flow motion in
Helmholtz Resonator is listed in App. A.1, with a final result of the motion in the neck of
the resonator can be approximately described by Keller and Zauner [3] as:
∂ 2U Am ρc2 ∂U ∂U ∂P
(1 + T )lm ρ − U = −ζρ |U | − T lm ρω − , (2.109)
∂t2 V ∂t ∂t ∂t
with T a coefficient, lm the orifice lenghth, Am the orifice area and ζ the loss coefficient
(see App. A.1)
As the acoustic resonator is always described with respect to frequency, parameter of
velocity and pressure can therefore be formulated in the form of U = U ? eiωt , P = P ? eiωt .
Eq. (2.109) can be rewritten as:
Aρc2 ?
(1 + T )lm ρωiU ? + T ρlm ωU ? + nl T ρdm ωU ? − iU = −P ? , (2.110)
| {z } | {z } | {z } |V ω{z }
A B1 B2
C

where the loss factor nl accounting for the nonlinear losses [30] can be expressed as
ζρU ? 2 ≈ nl T ρdm ωU ? , (2.111)
This factor depends both on the resonator geometry and especially the SPL [6].
In Eq. (2.110) the first term A on left hand-side accounts for air mass within the neck
region, while the second term represents the linear dissipation loss B1 due to friction of
fluid oscillation near wall and the nonlinear losses B2 due to vortex shedding. The third
term C reflected effect of compression. The right hand-side shows an acceleration caused
by pressure difference [6]. An very important criterion: the eigenfrequency of resonator can
be calculated from Eq. (2.110) as:
s
Am
ωeig = c , (2.112)
V (1 + T lm )

22
2.2 Basis Of Acoustics

Finally, the a normalized impedance can be expressed as:

P? lm √ Am ρc2
    
Aca γ−1 Aca
Z(ω) = ? = ρ 1+ √ 1 + nl + 2νair ω+i lm ρω(1 + T − ) ,
U Am Pr dm Am Vω
(2.113)
where the real part and imaginary part are defined as resistance Re and reactance Im:

lm √
  
Aca γ−1
Re(Z) = ρ 1+ √ 1 + nl + 2νair ω, (2.114)
Am Pr dm

Am ρc2
 
Aca
Im(Z) = lm ρω(1 + T − ) . (2.115)
Am Vω
By studying Eqs. (2.110) and (2.113), it can be observed that the resistance part is
constructed by pressure term and losses term, where nonlinearity may exists, and the
reactance includes mass term and compression. Another important factor: the reflection
coefficient Rf (ω) also relates to the normalized impedance Z, which can be defined as:

g? Z − ρc
Rf (ω) = ?
= , (2.116)
f Z + ρc
with ? indicates the Fourier transferred quantity.

23
3 Numerical Setup

In this chapter, the basic solver settings used for the presented simulations are presented.
Simulation algorithms and addiction smooth methods are first introduced in Sec. 3.1.
On the one hand, the target of algorithm choice is to make the whole simulation process
stable, and on the other hand, the chosen setup must ensure an accurate reproduction of the
physical processes. Geometries and meshes are presented in Sec. 3.2. All cases are simulated
with 3D structured meshes. On one side, different geometries and meshes will help to ensure
the conclusions that achieved are general, on the other side, different geometries and grid
sizes can also be treated as alternative parameters in the analysis of final results.

3.1 PIMPLE with NSCBC

All simulations presented this thesis are performed with PIMPLE algorithm in OpenFoam,
which combines the processes of SIMPLE and PISO, whose flow chart is depicted in Fig. 3.1.
In the inner PISO loop, a combination of pressure equation and velocity correction is
calculated, while the energy equation is solved in the outer SIMPLE loops. This process
suits especially to in high frequency fluctuating fluid fields and unstable fluid simulations.
Compared with the pure PISO algorithm, calculation for the inner time steps stays more
stable by the PIMPLE process, since the energy equation will not be calculated. And the
danger of instable break up can thus be enormously reduced.
In OpenFoam, the parameters of PIMPLE algorithm are defined in the file ”fvsolution”,
a sample code line can be defined as follows:

nOuterCorrector 20;
nCorrectors 3;

It means in each time step the outer loop will be computed till converged or 20 times if not
converged, and for each outer loop the pressure equation with velocity correction iterates
3 times.
In this thesis, PIMPLE algorithms that coupled with Navier-Stokes Characteristic
Boundary Condition (NSCBC) from the paper of Poinsot and Lele [31] will be used. With
this boundary condition, it is possible to insert a acoustic wave ”f ”, while ensuring low
acoustic reflections for the leaving wave ”g”.

24
3.1 PIMPLE with NSCBC

Initializing Fields

Temporal loop PISO loop

PIMPLE loop Pressure equation

Velocity equation Flux correction

Total energy
Velocity correction
equation

End PISO

End PIMPLE

Last time
step

End

Figure 3.1: PIMPLE Algorithms

25
 Numerical Setup

lt ln lca

lft lfc

dt dft dn dca
dfc

z
x

Figure 3.2: Helmholtz Resonator geometry in the simulations

3.2 Test Cases

In order to ensure mesh independence, different meshes are built up for the simulation,
which will be introduced in this section. Moreover two different geometries are used to
enable more general conclusions in the end. With actually the same construction (as in
Fig. 2.3 presented), Helmholtz Resonators from Hersh and Förner vary only in geometry
sizes. Hersh’s geometry will be investigated in special detail, because more series of exper-
imental data are given in his paper [1]. In the mean while, Förner’s geometry is used also
with measurement [4] to prove, that the conclusion from Hersh’s simulation is a general
one.

3.2.1 Hersh’s Geometry

A typical Helmholtz Resonator is structured as Fig. 3.2. In the simulated cases, it consists
of three parts: backing volume, the neck and impedance tube outside the resonator. Whole
structure is generated as a 3-D axis symmetric cylinder, whose geometrical size is tabulated
in Table 3.1.

Geometry Outside Tube Orifice Cavity

[mm] length lt diameter dt length ln diameter dn length lca diameter dca
Hersh 100 50.8 1.59 6.35 25.4 50.8
Förner 50 50 4 4.2 20 50

Table 3.1: Helmholtz Resonator geometries from Hersh et al. [1] and Förner et al. [4].

In part close to the neck, where the fluid separates and then developes into turbulence,

26
3.2 Test Cases

meshes need to be refined properly. However, over-sized refinement region will enormously
enhance the total computation cost without obviously improving the result since only
little computational region is filled with vorticity structures. As shown in Fig. 3.2, also the
sizes of refinement region differ with the resonator geometry, which can be formulated in
Table 3.2.

Geometry Tube Refinement Cavity Refinement

[mm] length lf t diameter df t length lf c diameter df c
Hersh 10 12 10 12
Förner 15 12 15 12

Table 3.2: Refinement of Hersh and Förner’s geometry.

Mesh Variation
In this section the mesh setting of each geometry is presented. From the direction of positive
symmetric axis (+X direction), the mesh can be divided in four parts like in Fig. 3.3a. Block
I and II are inside resonator orifice diameter, while sizes of Block III and IV can be referred
to the geometry defined in Sec. 3.2.1. This block separation validate both in Hersh’s and
Förner’s geometry.
IV

III

II

I
(a) Mesh view from +X direction. (b) Mesh sketch map from +X direction.

Figure 3.3: Mesh structure from +X direction with notation of blocks, I: H grid region
inside orifice; II: O grid region inside orifice; III: refinement region of O grid; IV: coarse
grid region of O grid.

Three meshes with different grid size (coarse[HC], middle[HM] and fine[HF]) are uti-
lized in the simulation of Hersh’s resonator. On the one hand, different grid scale means
also different computation node numbers, which results in different simulation quality and

27
 Numerical Setup

computational cost. In simulation of turbulence, the grid size plays even more important
role, as the ratio between grid scales and turbulence length scales can influence the simula-
tion result significantly. It can be evaluated through spatial TKE spectrum and two-points
correlation in Secs. 4.2.2 and 4.2.4. On the other hand, the suitable grid scale also varies
with different SGS-Models, compare the final results of different models in all of these
three meshes, the relationship between grid scale and SGS-models can also be evaluated
as will be discussed in Sec. 4.2.6. Last but not least, the results from different grids are
also to be compared with the Hersh’s experimental data, so that the grid influence can be
quantitatively appraised. Table 3.3 presents the detailed definition of these three grid sizes
in Hersh’s geometry.

Outline node numbers of meshes

and directions coarse (HC) middle (HM) fine (HF)
circumferential 64 80 120
orifice region in X 48 60 75
tube refinement in X 32 40 50
cavity refinement in X 32 40 50
tube in X 24 30 38
cavity in X 12 15 19
block II in radial 24 30 38
block III in radial 12 15 19
block IV in radial 12 12 19
space ratio 1.4 1.3 1.3
wall resolution 0.02[mm] 0.02[mm] 0.02[mm]

Table 3.3: Mesh data for Hersh’s resonator geometry.

Compared with normal mesh setting, an extremely fine grid separation is made in
circumferential direction and especially for the fine mesh. That is because of a utilization
of azimuthal velocity for the estimation of fluid dynamic properties (see Sec. 4.2.1). And a
refinement in that direction can help to enhance the analysis accuracy.

Boundary Conditions
As presented in Fig. 3.4, the whole boundary is divided into three part: inlet I, tube II
and wall III. For more general case inlet covers the boundary where the acoustic signal
is imposed, and the tube hold the mean fluid velocity, which flows along the z direction
queer to resonator orifice. In thesis, this velocity is set to be 0 to simplify the situation,
therefore no mean velocity is implemented in this simulation, but in the future the similar
simulation with mean flow still make sense. The wall boundary included all the solidity
boundaries in a real geometry. Thus, general parameters for all available models can be
defined in Tab. 3.4. By the definition of inlet velocity ”U”, an acoustic signal file ”inletf.txt”
will be included, which is generated through program Matlab following the definition of
NSCBC 3.1. In dynamic models a very small velocity field instead of value zero is needed

28
3.2 Test Cases

II

I III

Figure 3.4: Helmholtz Resonator with boundary: I–Inlet; II–Tube; III–Wall.

to initialize the SGS coefficient calculation. Boundary condition ”muSgs” and ”alphaSgs”
should also be defined for all SGS models besides laminar one. They describe the characters
of SGS thermal diffusivity and dynamic viscosity, respectively.

Parameter Inlet Tube Wall Initial field

U NSCBC symmetry
fixed velocity fixed velocity
[m/s] field (0 0 0) field (0 0 0)
p NSCBC symmetry Zero gradient fixed value
[Pa] 106
T fixed value symmetry Zero gradient fixed value
[K] 293 293
alphaSgs Zero gradient Zero gradient Zero gradient fixed value
[kg/(m · s)] 0
muSgs Zero gradient Zero gradient Zero gradient fixed value
[kg/(m · s)] 0

Table 3.4: Initial and general boundary conditions of all SGS Models.

Parameter Inlet Tube Wall Initial field

k Zerogradient symmetry Zerogradient fixed value
[m /s2 ]
2
10−8
nuTilda fixed value Zerogradient fixed value fixed value
[m2 /s] 4.44 × 10−5 0 4.44 × 10−5
B Zerogradient symmetry Zerogradient fixed field
[m2 /s2 ] (0 0 0 0 0 0 )

Table 3.5: Initial and boundary conditions of special SGS Models parameters.

29
 Numerical Setup

However, in some SGS models special parameters can be addicted, which are defined
in Tab. 3.5. ”k” is actual SGS kinetic energy kSGS in Sec. 2.1.5. It is required in one
Equation eddy model and homogeneous dynamic one Equation eddy model. Considering
the boundary effect, a very small initial value kSGS instead of 0 is desired to ease the danger
of simulation breaking up. ”nuTilda” is needes in SpalartAllamra model as ν ? in Sec. 2.1.5.
Referring Gomez [32], the value of ”nuTilda” can be defined as three times that of air
kinetic viscosity. At last, parameter ”B”, the complete SGS stress tensor is only needed by
Deardorff’s model. For a symmetric third-order matrix, only 6 items should be included
by the definition.

3.2.2 Förner’s Geometry

No big diversity between the geometrical structure of Förner’s and Hersh’s geometry can
be noticed, except the ratio of resonator orifice diameter to length value. That ratio is
1.05 by Förner’s geometry and 4 by Hersh’s. The resonator size changes as Tab. 3.1 and
Tab. 3.2 present. As the result, mesh data also varies as presented in Tab. 3.6. This mesh
is referred to as FM in the following discussions. With the geometrical diversity between
these two cases, a difference in eigenfrequency should also be concerned as about 380Hz in
Förner’s geometry and about 560Hz in all Hersh’s geometries.

Outline node number in

Förner’s Mesh(FM)
circumferential direction 80
orifice region in X direction 75
tube refinement in X direction 70
cavity refinement in X direction 70
tube in X direction 25
cavity in X direction 10
block II in radial direction 15
block III in radial direction 18
block IV in radial direction 25
space ratio 1.1
wall resolution 0.01[mm]

Table 3.6: Mesh data for Förner’s resonator geometry FM.

30
4 Simulation Results

In this chapter, the simulation results are discussed with the purpose to evaluate the
performance of different LES Subgrid-Scale models. This discussion consists of three parts
as considering of acoustic, fluid dynamic and other criteria. In first part, impedance and
reflection coefficient are considered to evaluate acoustic properties. While in fluid dynamic
part, spatial and temporal correlations and turbulent kinetic energy spectra are used. All
other criteria like Q-criterion and E/k criterion are described and utilized in the third part.

4.1 SGS Modeling Influence Of Acoustic Behavior

Acoustic analysis belongs to the most important contents in this work, since the mean
purpose of Helmholtz Resonator is to dampen the SPL and dissipate the acoustic energy.
According to the theoretical discussion in Sec. 2.2.2, the impedance Z and reflection co-
efficient Rf may act as useful criteria in the analysis of acoustic behaviors in Helmholtz
Resonator. In this section, most of the discussion is based on the impedance, while the
reflection coefficient is used in concerning about the SPL differences for his connection to
energy.
The presented impedance are normalized (see Eq. (2.113)) in this work to allow a better
comparison between different geometries. The real part of impedance is also called resis-
tance ”Re(Z)” and the imaginary part is called reactance ”Im(Z)”(see Eqs. (2.114) (2.115)).

2.5 15
125dB SPL
120dB SPL
10
90dB SPL
2 5

0
Re (Z) [-]

Im (Z) [-]

1.5
−5

−10
1
−15

−20
0.5
125dB SPL
−25
120dB SPL
90dB SPL
0 −30

0 100 200 300 400 500 600 700 800 0 100 200 300 400 500 600 700 800
Frequency [Hz] Frequency [Hz]

(a) Resistance. (b) Reactance.

Figure 4.1: Impedance: Nonlinear behavior in the resistance and linear behavior in the
reactance by geometry FM.

31
 Simulation Results

This strategy is widely used in analysis of acoustic properties in Helmholtz Resonator

(see [1, 33]). The resistance part and reactance part of impedance are therefore separated,
since the former one may presents a non-linearity behavior with the influence of SPL and
geometry, which is described in model of Eq. (2.111) with the nonlinear loss factor nl in
the resistance. However, with the including of the mass term and compression term, the
reactance part always shows a very good linearity with different SPLs. A typical impedance
series of Helmholtz Resonator with respect to frequency, where non-linearity in resistance
and linearity in reactance with the changed SPLs is presented in Fig. 4.1. It can be seen
that the resistance changes with the SPL and behaves nonlinear, while the reactance does
not.
Most helpful details can also be presented through the loss factor and the nonlinearity
caused by it. Since it is proven to be the most sensible to SPL difference as well as the
utilized SGS models. So that the difference between variety of SGS models can be figured
out more directly.
Compared with the impedance, the gain of reflection coefficient represents the absorp-
tion of acoustic energy more directly. According to Eq. (2.116), it is defined as the ratio of
Fourier transformed reflection wave g to the Fourier transformed initial wave f . Therefore,
larger gain of this coefficient means the the resonator damps inefficiently, as less acous-
tic energy are in resonator absorbed. While a smaller number of it implies more acoustic
energy is damped through resonator.
Figure 4.2 shows the reflection coefficient numbers of a certain SPL with respect to
the frequency. As expected close to eigenfrequency, the coefficient decreases to its smallest
value, where the turbulence is proved to be most strongly excited. However, since the
value at each frequency also changes with different SPLs, which will be further discussed
in Sec. 4.1.3, linear behavior of Resistance according to different SPLs is not possible. And
that is also the basic reason of Resistance nonlinearity.

1.2

0.8
Rf [-]

0.6

0.4

0.2 125dB SPL

120dB SPL
90dB SPL
0
100 200 300 400 500 600 700 800
Frequency [Hz]

Figure 4.2: Typical gain of the reflection coefficient with respect to frequency and SPL
with FM geometry.

32
4.1 SGS Modeling Influence Of Acoustic Behavior

4.1.1 Correction SPL to AMP Transformation

To make a meaningful comparison with the experimental data, the velocity conditions
should be set to match that of Hersh’s geometry [1] in the simulation. However, the SPL
value is measured at a certain position lt away from the neck (see Fig. 3.2) in experiment.
While in the simulation, this must be defined for the inlet boundary to match the exper-
iment. Therefore, a correction for the signal definition is needed. According to Förner et
al. [5], a transform of the acoustic signal definition from SPL to pressure amplitude (AMP)
of the wave Af with f = Af exp(iωt) can be used to set the acoustic input wave correctly.
As the characteristic wave can be expressed as:

P 0 (lt )
= [1 + Rf exp(−iω2lt /c0 )]f (lt ). (4.1)
ρ 0 c0
The reflection coefficient Rf must be calculated with a iterative process if it is not known
before. But in order to reduce the simulation cost, values are directly taken from Hersh’s
paper [1]. Thus, the relationship between the measured SPL of Hersh and the needed wave
pressure amplitude in simulation can be formulated as [4]:
√
Lp /20 2Pref
Af = 10 . (4.2)
|1 + Rf exp(−iω2lt /c0 )|ρ0 c0
where the SPL value Lp is defined as:
P0
 
Lp = 20 log10 (4.3)
Pref

Figure 4.3 makes the above discussed topic visible. Especially from the resistance part,
simulation results from through SPL numbers defined acoustic signals present huge inac-
curacy if compared with the experimental data. In the meanwhile, those through AMP

0.8 4
Hersh's data
Hersh's data With AMP Correction
With AMP Correction With SPL Signal
With SPL Signal 3
0.7

2
0.6
Im (Z) [-]
Re (Z) [-]

1
0.5
0

0.4
−1

0.3 −2

0.2 −3
350 400 450 500 550 600 650 700 750 800 350 400 450 500 550 600 650 700 750 800
Frequency [Hz] Frequency [Hz]

(a) Resistance by AMP correction. (b) Reactance by AMP correction.

Figure 4.3: Impedance comparison between the simulation results of SPL and corrected
AMP defined signals.

33
 Simulation Results

defined acoustic signals shows relative reasonable results compared with the experimental
data. Thus, the comparison between simulated and experimental results seem to be more
reliable if a AMP correction is acted. Therefore when the impedance curve of a special
SPL is simulated, a series of corresponding AMP values at different frequencies should be
computed at first for the setting of the NSCBC. In the following sections, all signals are
defined according to corrected AMP acoustic waves.

4.1.2 Impact of SGS models

With the previous preparations, the influence on the impedance and reflection coefficients
caused by SGS model differences will be discussed in this section as one of the most
important part of this work. A comparable laminar model without SGS modeling and all
available compressible LES SGS models that listed in Sec. 2.1.5 except Deardorff’s model
will be utilized in this comparison. Deardorff’s model is not discussed here due to its
unstable behavior for the setup consideration. Simulation of Deardorff’s model works out
only with quite small frequency, where the non-linearity performance can not be notice in
resistance even with different SPLs, which certainly make no sense for the evaluation. The
reason of this unstable behavior is still unclear. As authors assumption, at the time point
of reflected wave passing by the orifice region, an interference caused by initial acoustic
wave and the reflected wave may excited some unknown and unstable factor within the
complete SGS stress tensor τij , which dominants the iteration of Deardorff’s model. And
that instability can finally break the modeled NS-equation at the very same time point.
However, no other condition or setting changes except reducing frequency range can help
vanishing that error, so that this model has to be left for future discussion in other topics.
The comparison between experimental data and simulated results of geometry ”FH” is
presented in Fig. 4.4. The subfigures (a) and (b) show out result of the 130dB case, (c) and
(d) the result for a SPL of 140dB case. In general, the largest differences between experi-
mental and simulated results (Smagorinsky model as example) exist mostly in resistance
part by eigenfrequency range. The difference of measured and simulated reactance is quite
ignorable. Only in the resistance part, some larger variety presents by eigenfrequency range
(here about 600Hz) and high frequency ranges (here above 700Hz). Since all models present
with relatively similar behavior, only one curve of Smagorinsky model is printed and the
others are shown only with selected points. Even though these minimum differences in the
figure, Spalart Allmaras model will always be the one, that for most selected frequencies
closest to the experimental data, especially in higher frequency ranges. This is reasonable,
since besides the Deardorff’s model, Spalart Allmaras model is left to be the only one that
considers the complete SGS stress tensor. More empirical coefficients are in that model
also imposed to ensure his accuracy with different geometries. And in each flow condition
a special defined trip term ν ? can make a better adjustment for his environment. However,
even the simplest laminar model without any consideration of LES SGS also prints a rea-
sonable and very similar data series, which is unexpected. It means by case of no mean
flow, the simulation of by acoustics generated turbulence is relatively independent from the
turbulent model whatever be use. A complex SGS modeling utilization for acoustic analysis
is only needed by very precise simulations, otherwise laminar model is enough to fulfill the

34
4.1 SGS Modeling Influence Of Acoustic Behavior

1.6 4

1.4 2

1.2 0

1 −2
Re (Z) [-]

Im (Z) [-]
0.8 −4

0.6 Smagorinsky −6 Smagorinsky

Dynamic Dynamic
0.4 Laminar −8 Laminar
SpalartAllamra SpalartAllamra
0.2 OneEq −10 OneEq
Experiment-Hersch Experiment-Hersch
0 −12
100 200 300 400 500 600 700 800 100 200 300 400 500 600 700 800
Frequency [Hz] Frequency [Hz]
(a) Resistance for 130dB SPL. (b) Reactance for 130dB SPL.
3 4

2
2.5
0

2
−2
Im (Z) [-]
Re (Z) [-]

1.5 −4

−6
1
Smagorinsky
Smagorinsky −8 Dynamic
Dynamic
Laminar
0.5 Laminar
−10 Spalart Allamra
Spalart Allamra
OneEq
OneEq
Experiment-Hersch
Experiment-Hersch −12
0
100 200 300 400 500 600 700 800 900
100 200 300 400 500 600 700 800 900
Frequency [Hz] Frequency [Hz]
(c) Resistance for 140dB SPL. (d) Reactance for 140dB SPL.

Figure 4.4: Impedance with difference SGS models of SPL 130dB and 140dB in Hersh’s
geometry ”FH”.

estimation. However, it is not discussed in this work whether laminar modeling still works
in the case of larger mean flow, where the mean flow can also generate turbulence itself.
Similar comparison can also be made for Förner’s geometry ”FM” as presented in
Fig. 4.5. An Obvious distinction from the one of Hersh’s is that a larger difference between
the experimental and simulated results exists in relatively linear range (below 300Hz and
over 500Hz). However, as shown in Hersh’s case, different SGS models still result in only
a small impedance variety. For all comparable simulated frequencies and SPLs, different
SGS models present almost the same values. Therefore the conclusion can be made, as
the LES SGS models variety impacts non-linearity in impedance resistance in a Helmholtz
Resonator only with quite small range by means of CFD.
Besides the comparison of impedance, influence of different SGS models in reflection
coefficient can be fund in Fig. 4.6. Only the comparison of high SPLs in each geometry

35
 Simulation Results

1.8 15
Smagorinsky
Dynamic
1.6 Laminar 10
SpalartAllamra
1.4 OneEq
5
Experiment

1.2 0
Re (Z) [-]

Im (Z) [-]
1 −5

0.8 −10

Smagorinsky
0.6 −15 Dynamic
Laminar
0.4 −20 SpalartAllamra
OneEq
Experiment
0.2 −25
0 100 200 300 400 500 600 700 800 0 100 200 300 400 500 600 700 800
Frequency [Hz] Frequency [Hz]
(a) Resistance. (b) Reactance.

Figure 4.5: Impedance with different SGS models by SPL 120dB in Förner’s geometry
”FM”.

1 1.1
Smagorinsky
Dynamics 1
Laminar
0.9 Spalart Allamra
OneEq 0.9

0.8 0.8
Rf [-]

0.7
Rf [-]

0.7
0.6

0.6 0.5

0.4 Smagorinsky
Dynamics
0.5
Lamiar
0.3 Spalart Allamra
OneEq
0.4 0.2
200 300 400 500 600 700 800 900 100 200 300 400 500 600 700 800
Frequency [Hz] Frequency [Hz]
(a) Hersh’s geometry with SPL 140dB. (b) Förner’s geometry with SPL 125dB.

Figure 4.6: Reflection coefficient for different SGS models with meshes ”HF” and ”FM”.

is listed here, since the nonlinearity, according to previous discussion, is extremely strong
by those SPLs, and different SGS models varies more. As expected, in general only minor
differences can be observed in coefficient gain for different SGS Models. Relatively, some
differences by high frequency range in Hersh’s geometry is more obvious, which can also
be reflected from impedance discussions.

36
4.1 SGS Modeling Influence Of Acoustic Behavior

4.1.3 Nonlinear Resistance With Respect To SPL

Our discussion on LES SGS models depends fully on the assumption, that the turbulence
close to neck is already generated by acoustics. However, the flow does not separate by
lower SPLs, so that the turbulence close to neck will not be generated. As the result, the
nonlinearity behavior in resistance will also disappear. In this section, the above resistance
nonlinearity with its respect to SPL will be quantified. As comparison, experimental data
and simulated results (here the dynamic model) are selected. Both the results from Hersh’s
geometry ”FH” and Förner’s one ”FM” are presented in Fig. 4.7.
Consider first Hersh’s geometry in Fig. 4.7. The strongest nonlinear behavior occurs in
the frequency space between 560Hz and 640Hz. As the eigenfrequency of Hersh’s geometry
is around 580Hz, where also the highest nonlinear behavior occurs as expected. Considering
the resistance, the nonlinear behavior of resistance becomes stronger with the growth of

3 4
140dB Dynamic
140dB Experiment
2
2.5 130dB Dynamic
130dB Experiment
120dB Dynamic 0
2 120dB Experiment
−2
Re (Z) [-]

Im (Z) [-]

1.5 −4

−6 140 dB Dynamic
1
140 dB Experiment
−8 130 dB Dynamic
0.5 130 dB Experiment
−10 120 dB Dynamic
120 dB Experiment
0 −12
100 200 300 400 500 600 700 800 900 100 200 300 400 500 600 700 800 900
Frequency [Hz] Frequency [Hz]
(a) Resistance of Hersh’s ”FH”. (b) Reactance of Hersh’s ”FH”.
15
125dB Dynamic
2 125dB Experiment
10
120dB Dynamic
120dB Experiment
90dB Dynamic
5
1.5 90dB Experiment
0
Im (Z) [-]
Re (Z) [-]

−5
1 −10
125dB Dynamic
−15 125dB Experiment
0.5 −20 120dB Dynamic
120dB Experiment
−25 90dB Dynamic
90dB Experiment
0 −30
0 100 200 300 400 500 600 700 800 0 100 200 300 400 500 600 700 800
Frequency [Hz] Frequency [Hz]
(c) Resistance of Förner’s ”FM”. (d) Reactance of Förner’s ”FM”.

Figure 4.7: Impedance with different SPLs in Hersh and Förner’s geometries.

37
 Simulation Results

SPL, and the difference between experimental and simulated results also enlarges. Results
of simulation in this geometry is more reliable for lower SPL with a better conformity
with the experimental data if compared with the higher SPLs. However, the variety of
SPLs has not that strong impact on the reactance, where both of the three utilized SPL
curves behaves linearly with the similar form and value. No remarkable difference between
experiment and simulation because of SPL changes can be noticed if regarding only the
reactance part.
When the geometry of Förner is considered, the former conclusion about reactance
turns to be strictly true. Even the difference for change in SPLs is not noticeable. But
as discussed in previous section, relatively less similarity is expressed in the resistance
comparison between experiment and simulation in Förner’s geometry. Consider the curves,
where nonlinear behavior is triggered (with SPL of 120dB and of SPL 125dB), close to
the eigenfrequency, the resistance obtain from simulation is always higher than that from
experiment. Only away from the eigenfrequency, this prediction of the numerics vanishes
and the experiment measures higher resistance values. However, for SPL 90 dB, where
flow rarely separates, a linear relation can be followed in both experimental and simulated
curves, with the simulated value always smaller than the experimental ones.
Also for very high SPLs, the effective length decrease, which drive the eigenfrequency
shift to higher frequency range. This can be seen in Hersh’s geometry, where the neck
length is very small, such that a shift in the end correction has a huge impact. For Förner’s
geometry, the neck is larger and not so high SPLs are investigated. Therefore no shift in
eigenfrequency is visible.
As conclusion, simulated resistance values around the eigenfrequency are always larger
as the experimental data. But the comparison in relative linear parts depends on the
accuracy of SPL to AMP correction and geometry. No useful messages about SPL varieties
can be figured out from reactance curves.

1.2 1.2

1 1

0.8 0.8
Rf [-]
Rf [-]

0.6 0.6

0.4 0.4

0.2 140dB SPL 0.2 125dB SPL

130dB SPL 120dB SPL
120dB SPL 90dB SPL
0 0
200 300 400 500 600 700 800 900 100 200 300 400 500 600 700 800
Frequency [Hz] Frequency [Hz]
(a) Hersh’s geometry (b) Förner’s geometry

Figure 4.8: Reflection coefficient Rf by dynamic model simulation with variety of SPL in
mesh ”HF” and ”FM”.

38
4.1 SGS Modeling Influence Of Acoustic Behavior

As introduced in Sec. 4.1, nonlinearity of resistance can be traced back to the change of
resonator reflection coefficient ”Rf ” with SPL differences as shown in Fig. 4.8. By Förner’s
geometry, we find the value of reflection coefficients are usually larger by simulation of
linear case (90 dB SPL) than those by simulation of nonlinear case (like 120 dB SPL).
Since the generated turbulent structure will help to absorb acoustic energy to make the
reflection coefficients even smaller. However, by further increasing of SPL, only the growth
of absorbed absolute quantity of acoustic energy can be achieved, while the relative pro-
portion of absorbed parts to the whole acoustic energy parts decreases. This conclusion
can be summarized through the raising of reflection coefficients again with further growth
of SPL both in Förner’s geometry (from 120dB to 125db) and Hersh’s geometry (from
120dB to 140dB). Thus, with further growth of SPL, there should be a critical SPL value,
beyond which even the absolute quantity of absorbed acoustic energy will not change. This
assumption can be proven through turbulent kinetic energy spectra analysis in Sec. 4.2.5.
However, it is true only when the overdamped frequency for all SPLs are very close (can
be regarded as eigenfrequency) as the example in Förner’s geometry. A resonator is over-
damped when the normalized resistance Re > 1, where the gain of the reflection coefficient
increases strongly with the increasing of the SPL. However, if the overdamped frequency
also varies strongly (can not be treated as eigenfrequency) with the change of SPLs, as
Hersh’s geometry, this conclusion is not quite sure. As example we can consider the fre-
quency of 700Hz in Hersh’s geometry, since Rf is smaller for SPL 140dB case than for
130dB and 120dB ones, even by higher SPL, resonator can absorb both large proportion
and absolute quantity of acoustic energy.

39
 Simulation Results

4.2 Influence of The SGS Modeling on Turbulent Ki-

netic Energy Spectra and Correlation
In the previous section of the acoustic analysis, only minor differences resulting from of SGS
model utilization can be observed. Therefore, we concentrate more on the part of pure CFD
analysis to evaluate the turbulence behavior. According to Davidson’s investigation [8],
correlations and energy spectrum belong to the most popular and effective methods to
have a deep look into the turbulence. Because of that, correlations, including spatial two-
points and temporal auto-correlation, as well as turbulent energy spectra in frequency and
wave number space, will be the focus point in the following discussion of SGS modeling
comparison.

4.2.1 Azimuthal Velocity

Generally in full developed turbulent system, the selected velocity direction will not play a
very significant roll, since the turbulence is homogeneous. However, in our system, where
the original mean velocity is zero and the turbulence is generated by acoustic signals,
the velocity fluctuation includes therefore not only the part caused by turbulence but
also a part due to the acoustic signal fluctuation if consider only directions in axial and
radial directions. The complete velocity fluctuation can be expressed with both acoustic
0 0
component Uacou and pure turbulent component Uturb as:

U 0 = Uacou
0 0
+ Uturb . (4.4)

Thus, estimation criteria like correlation and turbulent kinetic energy, which base on
velocity fluctuation, will lose their accuracy or shows very unreasonable characters, if the
acoustic velocity fluctuation is still included in the analysis. To deal with this problem,
Zhang and Bodony has provided a clever strategy in their paper [7]. They transformed their
velocity in a cylindrical coordinate system and considered only the azimuthal velocity that
is vertical to the acoustic signal transfer direction instead of the axial and radial direction.

Uy

U
Uz

Uy
z

Uz y

Figure 4.9: Measurement of the azimuthal velocity in resonator.

40
4.2 Influence of The SGS Modeling on Turbulent Kinetic Energy Spectra and Correlation

axial velocity
1 azimuthal velocity

Ekin(f) [m2 /s] 0.01

0.0001

1e
06

1e
08

1e
10
10 100 1,000 10,000
f [Hz]

Figure 4.10: TKE spectra with azimuthal and axial velocity.

In our cases this strategy will also be used as illustrated in Fig. 4.9. Since the acoustics
is by definition irrotational, it does not influence the azimuthal velocity component in our
axis-symmetric cases. With its tight relationship to the location coordinates, the azimuthal
velocities must be calculated in each investigated point. Azimuthal velocity transform from
Cartesian coordinate can be expressed as:
Z Y
Uθ = Uy · sin θ + Uz · cos θ = Uy · √ + Uz · √ , (4.5)
Y 2 + Z2 Y 2 + Z2
where Uy and Uz donate velocity in Y and Z-direction, Y, Z the monitored locations.
Through azimuthal velocity transformation, acoustic factors, especially the resonance near
eigenfrequency range can be sharply reduced. A comparison of TKE spectra (see Sec. 4.2.4)
based on both axial velocity and cylindrical azimuthal velocity is presented in Fig. 4.10.
The acoustic signal in Fig. 4.10 is generated with a frequency of 400Hz, which is close
to the eigenfrequency of geometry ”FM” according to Sec. 4.1. From the comparison, an
obvious peak shaped form can be fund in the curve of axial velocity around the reso-
nance frequency of 400Hz, which shows a strong relationship with acoustic characters.
In the meantime, azimuthal velocity curve present a good reduction of acoustic effect in
vanishing the spike shaped form around the same region. The TKE spectra values also
varies strongly between that of axial velocity and azimuthal velocity. On the one hand,
an acoustic longitudinal wave excited fluctuation mainly in its propagating axial direction,
on the other hand, by reducing acoustic fluctuation part, some energy that shows acoustic
perturbation are therefore ignored. In the following sections only the estimation based on
azimuthal velocity is used to reduce the acoustic influence as much as possible.

41
 Simulation Results

4.2.2 Correlations
From the discussion of Sec. 2.1.2, two varieties of correlation can help to investigate the
turbulence simulation depending separately on spatial (grid size) and temporal (time step)
criteria. Their setting and analysis are discussed in this section.

Two-points correlation
According to the discussion in Sec. 4.2.1, the azimuthal velocity is utilized to study the
pure turbulent behavior in the situation of this work. As a result, the two-points correlation
should also present the turbulence characters in circumferential direction, which can also be
regarded as homogeneous turbulent direction. Therefore, the probe-points must be located
along a circle in the Y-Z plane.
1 2 3
ln 4
5
6
3o 3 o o

...
3

2ln
2 points in
ln each /60 grad
arc length
θ
of a half circle

3o

61 60

(a) Circle Θ at x = 3ln and diameter of ln . (b) Investigated points distribution in circle Θ.

Figure 4.11: Setting of investigated points and circle.

Figure 4.11 defines the circle and points that fulfill former discussed conditions. The
concerned circle Θ with radius of half orifice length is located at Y-Z plane with distance
of twice orifice length to the end of resonator neck like Fig. 4.11a shows. 61 investigated
points are distributed along the half circumference of circle Θ, which means one point
in each Ψ = 3◦ arc length as shown in Fig. 4.11b. Thus, the minimum resolved unit
is determined by the resonator geometry characteristic length lca as πRΨ/lca , while the
maximum resolved unit is determined by the minimum mesh grid size ∆cell as πRΨ/∆cell .
In the analysis for only two-points correlation of different SGS models, other conditions
are set as presented in Tab. 4.1 presents:

Conditions signal SPL signal frequency mesh

Value 120dB 600Hz HF

Table 4.1: Condition of two-points correlation analysis.

With the former settings, the normalized two-points correlation (see Eq. (4.8)) for dif-
ferent SGS models according to azimuthal velocity is calculated and presented in Fig. 4.12.

42
4.2 Influence of The SGS Modeling on Turbulent Kinetic Energy Spectra and Correlation

1 0 2 4 6 8 10
Smagorinsky 1 1
Dynamics Smagorinsky
Laminar Dynamic
Spalart Allamra Laminar
OneEq 0.8 Spalart Allamra 0.8
0.5 OneEq

0.6 0.6

0.4 0.4

−0.5 0.2 0.2

0 0
0 0.02 0.04 0.06 0.08
0 2 4 6 8 10
n R /lca [-] n R / cell [-]

(a) With corresponding to geometry scale. (b) With corresponding to mesh scale.

Figure 4.12: Two-points correlation with different models.

The left figure shows the normalized two-points correlation that corresponds to geometry
resonator scale. It presents a high fluctuation amplitude (between ±0.7) if compared with
the auto-correlation (between ±0.2) in later section. It actually means the spatial resolution
is not as good as the temporal one, so that the relationship between two selected spatial
points can express more unstable factors as the two from temporal points do. However,
if Fig. 4.12b on the right hand-side is considered, the spatial resolution still satisfies the
need to a accurate simulation. Since before the normalized two-points correlation falls to
0.1, which is often charactered as a turbulent length scale limitation (see [8]), there are
already 8 to 10 grids to cover the largest turbulent scale. These cell numbers varies only a
little within different SGS numbers, which means each model can provide a suitable spatial
resolving in the simulated case. Combining also the curves from Fig. 4.12a, the dynamic
model performs a little better for its relative smoother fluctuation than the other SGS
model. However, it is actually not so important in the discussing.

Auto-correlation

Similar settings can also be utilized to investigate the behavior of the velocity auto-
correlation. Different from the probe-points series in two-points correlation, only one probe-
point is needed for the auto-correlation. The Auto-correlation can be computed through
the velocities of each two different time steps in that point. As the two-points correlation
presents main characters in spatial mesh settings, auto-correlations depends more on time
step setting. Original time steps of simulations in this thesis are initialed to be 6 × 10−7 s
and the maximum time step is limited as 1×10−6 s. During the simulation process, a proper
time step between the initial value and the maximum value is chosen by the solver auto-
matically as to improve the simulation efficiency, while a certain maximal CFL-number is
chosen. The same condition setting as two-points correlation analysis is read from Tab. 4.1.
The single monitored point is chosen to be point 1 in Fig. 4.11b. Using the mathemati-
cal methods from Sec. 2.1.2, curves running of normalized auto-correlation at poin 1 for
different SGS models are computed as presented Fig. 4.13.
Different from two-points correlation in Fig. 4.12, more accurate curves is presented

43
 Simulation Results

1 Smagorinsky
Dynamics
laminar
Spalartallamra
0.8 OneEq

0.6

0.4

0.2

−0.2
0 0.0002 0.0004 0.0006 0.0008 0.001 0.0012 0.0014
time [s]
Figure 4.13: Auto-correlation with different models

in Fig. 4.13, as chronological velocity series including thousands of samplings in time is

generated automatically while the spatial local points have to be defined by hand. For most
curves, that corresponds to different SGS models, correlation fluctuations are between −0.2
and 0.2 for relatively bigger time steps, which also presents a better homogeneous character
in temporal series.
Comparing all the SGS models, Smagorinsky model and Spalart Allmaras model shows
the best behavior in this chronological valuating, since these two curve last the longest with
a correlation value of about 1, and the time integral scales also presents larger number of
6.23 × 10−4 s and 6.49 × 10−4 s, respectively. The value of other three models are around
2.5 × 10−4 s. It means the velocity influence in Smagorinsky and Spalart Allmaras model
can last longer, or with another expression: a tight relationship in velocities of same point
exists within a longer time distance in these two models as the other three. So that on the
one hand, if with the same time step setting these two models can resolve bigger time-scale
vortex in turbulence better. And on the other hand, the simulation time steps in these
two models can theoretically be defined relatively bigger as the other three models as to
seek better efficiency. However, considering that the former time step is limited between
6 × 10−7 and 1 × 10−6 , and the actual time step that chosen by the solver turns out to be
1 × 10−6 in all models by observing simulation ”log” file. Extra efficiency overage can thus
be used by rising the defined time step. By a roughly time step rising in all models to the
initial value 5 × 10−6 and the maximum value 1 × 10−5 only the Spalart Allmaras model
still converges. The model of Smagorinsky works without converging, which unexpectedly
enhances the simulation time cost as the result, while the simulation for other three model
even broken up after minutes of iteration. Consider only the successfully finished Spalart
Allmaras model, all the acoustic and turbulent properties present the same behavior as
the smaller time step case. In the mean while, the time cost changes promptly from about

44
4.2 Influence of The SGS Modeling on Turbulent Kinetic Energy Spectra and Correlation

23 × 56 per ”hour/CPU” in MPPII1 to 8.5 × 56 per ”hour/CPU”! Therefore a suitable time

step setting for different models can improve the simulation efficiency significantly, which
might be an interesting point to study further.

4.2.3 Setting of Energy Spectra

Based on the theoretical analysis in Sec. 2.1.3, energy spectra can be generated with
a Fourier transformation from velocity covariants. The spatial two-points covariant and
auto-covariant can be Fourier transformed to spatial TKE spectra in wavenumber space
and temporal TKE spectra in frequency space, respectively. Both energy spectra can be
used to evaluate the time step setting and mesh quality. The Komogolov’s −5/3 law (see
Sec. 2.1.3) provides a useful criterion to distinguish the turbulence behavior for the con-
sidered SGS models. The Fourier transformation is implemented in the post-processing
through obtained velocity series with help of a ”Matlab” program. There are dozens func-
tions to make Fourier transformation in ”Matlab”, and two of them: ”fft” and ”pwelch” is
chosen in the following to do the transformation job.

”fft” Function
fft is short for fast discrete Fourier transform(DFT), which transforms the data from
time domain to frequency domain. The transformed function is normalized through dividing
each transform by the square root of the signal (here velocity series) length, in order to
preserve the signal energy and noise level. It is the most effective way to achieve a discrete
Fourier transform, especially when the input data number can be written as 2m . However,
because of leakage phenomenon, which derives from an insufficient frequency resoluion of
DFT, this function always bring some problem by result analyzing. In signal or energy
spectrum evaluation, another function ”pwelch” is always preferred to smooth as well as
optimize the concluded data by Fourier transform processes.

”pwelch” Function
pwelch is short for Welch’s power spectral density. It calculats the PSD of a signal with-
out knowing the Fourier transformed value, in the meanwhile a windowing implementation
optimizes final data structure smoothly. Usually it performs better in a power signal, which
can be regarded as infinite long broadband. However, in a limited time domain, the PSD
can also be used to calculate the energy spectrum. A detailed derivation of the ”pwelch”
function is presented in Appendix A.2. The connection between the ”pwelch” function and
the ”fft” function in time and frequency space is formulated here as:

|Hw [n]|2
Dw (n) = , (4.6)
N2
where Dw (n) is a windowed PSD coefficient (”pwelch”), and Hw [n] denotes a windowed
discrete Fourier transformed signal and N is the windowed frequency number. Considering
1
MPPII is short for Massively Parallel Processing II, which is a cluster system of the LRZ.

45
 Simulation Results

0.0001
pwelch function
1e
05 fft function

1e
06
Ekin(f) [m2 /s]
1e
07

1e
08

1e
09

1e
10

1e
11

1e
12
10 100 1,000
f [Hz]

Figure 4.14: Temporal energy spectra with FFT and Pwelch function.

a so-called Equivalent Noise Band Width (ENBW):

PN −1
2
ki =0 r [ki ]
EN BW = PN N ∆f, (4.7)
( ki −1
=0 r[ki ])2

where r[ki ] is a window function. With the help of ENBW, a empirical PSD is defined as:

|Hw [n]|2
ED[n] = . (4.8)
EN BW
In LES simulation with influence of input acoustic signals like in this work, empirical PSD
in Eq. (4.8) suits better for reducing the effect of noise level.
The grammar of function pwelch for frequency space can following the simullar estima-
tion from Davidson [8] be written as:

[Pw Fw ] = pwelch(Uθ , W, WL , [], 1/∆t),

where Pw , Fw , Uθ , W , WL and ∆t donate the windowed PSD, windowed frequency, signal

series (here resolved azimuthal velocity), window number, window overlapping number and
the simulated time step as sample length, respectively.
Figure 4.14 presents a sample of the TKE curve computed by both ”fft” and ”pwelch”
function with the same chronological Uθ series. Within the frequency range between 10 and
1000 Hz, three TKE spectrum ranges (see Sec. 2.1.3) are best constructed. The curve gen-
erated directly through ”fft” function develops relatively flat without clearly showing the
range difference. Further more the TKE fluctuations is relatively stronger with the grow-
ing of frequency domain. Since the mean velocity in this thesis is actually zero, if every
monitored time points is considered, the velocity fluctuation switches between positive and
negative sign very frequently, however, by considering a finite range of time domain, the av-
eraging azimuthal velocity in each window space presents a comparatively less fluctuation,

46
4.2 Influence of The SGS Modeling on Turbulent Kinetic Energy Spectra and Correlation

and that is also why the window function can help to reduced the influence of background
noise. In comparison, ”pwelch” curve shows a definite improving tendency that reflect the
Kolmogolov’s Law (see Sec. 2.1.3) with a relative later curve beginning than the ”fft”. This
starting point of ”pwelch” function can be controlled by setting a suitable window length
with the thinking of the whole simulation time range Twhole (here 0.03s) and the minimum
detected frequency should be fmin = 1/Tmin , which should be calculated in this simulation
as 66.6 Hz (2/Twhole , 2 same period of sample can be recognized) in ideal cases. However,
different factors like grid quality and SGS models can make also sense. But compared with
the uncontrollability in ”fft”, it can be more reliable with ”pwelch” in this situation.

Window Setting of ”pwelch” Function

As discussed in the previous section, the ”pwelch” function is called in Matlab as:

[Pw Fw ] = pwelch(Uθ , W, WL , [], 1/∆t),

with W and WL controlling the window setting. The meaning of these two parameters
can be explained through Fig. 4.15, where Lw , Lo and Lno denote the window length,
window overlapping length (both through the former and next window overlapped) and
the not overlapping length, respectively. The complete data length Ldata = W · Lw , and the
overlapped data length Lodata = Lo · W = Lw · WL . In the algorithm of this function, once
Lw = Ldata /W is calculated, all the settings can be generated perfectly. Since Ldata < Lodata ,
a default condition included in the definition reads WL < W . The default value of WL can
be called with ”[]”, which means the half of W number is used.
Figure 4.16 provides a comparison between different window numbers and overlapping
number. Reducing the window number from 3750 step wise to 200, the curve changes as
in Fig. 4.16a. With the number Lw increasing, the minimal resolved frequency becomes
larger, while the curves turns out to be smoother. However, when the window number
becomes smaller than 500, the curve will not strictly follow original curve running in all
the three regions. Generally, the window number between 1000 and 2000 is to be preferred
for both smooth and precise curve. In Fig. 4.16b, the window number Lw is chosen as
1000, different curves express the variety of window overlapping setting. However, no big

Lw Lw

Lno

Lo/2 Lo/2

Lw

Figure 4.15: Window and the window overlapping in ”pwelch” function.

47
 Simulation Results

0.01 0.001
window number 200 window overlapping 50
window number 500 window overlapping 200
0.001 window number 1000 window overlapping 500
window number 2000 0.0001
window number 3750
0.0001
1e
05

Ekin(f) [m2 /s]

Ekin(f) [m2 /s]

1e
05
1e
06
1e
06

1e
07
1e
07

1e
08
1e
08

1e
09 1e
09
1,000 10,000 1,000 10,000
f [Hz] f [Hz]
(a) Change of window number by default (b) Change of window overlapping by window
overlapping in ”pwelch” function, number 1000 in ”pwelch” function,

Figure 4.16: TKE spectra with different window settings with mesh HF in 130 dB.

difference can be noticed with only the changing of overlapping, so that in all the following
analysis overlapping is set to Matlab default.

4.2.4 Turbulent Kinetic Energy Spectra

With preparations in the previous sections, analysis of both spatial and temporal spec-
tra can be generated by Fourier Transform in this section. However, as more data can
be achieved in a temporal velocity series, the further analysis in following sections are
performed in frequency domain.

Spatial Energy Spectra

Spatial TKE spectrum is derived from two-points covariance BUθ Uθ in Eq. (2.8) with a
Fourier Transform. The simple method of ”fft” will not be selected due to its bad prediction
because of leakage. Instead, a ”pwelch” transform is implemented, and the resulted series
follows a windowed Fourier Transform in Eq. (4.8).
Figure 4.17 shows how the curves of different SGS model vary in a spatial energy
spectra, where the X-axis is the wavenumber corresponding to grid length. According to
this figure, all SGS models present reasonable curve performance the region II between
about the wavenumber 5 × 103 to 2 × 104 in moderating the Komogolov’s −5/3 law, which
means all these models can accurately describe the turbulence behavior in this geometry as
the discussion in acoustic part already proved. However, some difference can still be fund
within these TKE values. According to the discussion in Sec. 4.2.2, as a full viscous stress
tensor τij is considered in Spalart Allmaras model, it presents with higher energy compared
with other models as expected. However, it is quite unexpected, why the most energetic
one turns out to be the one equation eddy model, while the similar dynamic model shows
the least energy level. Since only limited numbers of velocity data are available in spatial

48
4.2 Influence of The SGS Modeling on Turbulent Kinetic Energy Spectra and Correlation

0.01 Smagorinsky
Dynamics
laminar
Spalart Allamra
OneEq
0.001
Ekin( ) [m3/s2]

0.0001

1e
05

1e
06
100 1,000 10,000 100,000
κ π ∆cell [m−1]

Figure 4.17: Spatial energy spectra with different SGS models.

spectra, all SGS models draw only with quite smooth curves, where detailed differences
could actually be lost.

Temporal Energy Spectra

Different from spatial energy spectra, temporal energy spectra of TKE base theoretically on
the auto-covariance in Eq. (2.12) with a subsequent performance of the Fourier Transform.
In the algorithm the ”pwelch” function is used of for velocity U 0 (t) series. With adjusted
window settings, data visualization of temporal energy spectra for both Hersh and Förner’s
geometry are expressed in Fig. 4.18. In both geometry a frequency that close to their
resonator eigenfrequency (580Hz and 380Hz) are chosen to make the turbulence behavior
stronger with their resonance.
In Fig. 4.18a with a Hersh’s geometry ”FH”, the curves in region II and III can be
recognized from frequency range 500Hz to 4000Hz and after 4000Hz, respectively. Spalart
Allmaras, dynamic, Smagorinsky and even laminar model draw only similar curves with
−5/3 slop in region II. As expected, Spalart Allmaras model has the highest TKE compared
with the other three models. However, a irregular development can be observed in one
equation eddy line, where an unexpected turn point at about 1050Hz exists. In modeling
homogeneous turbulent scales region III, all models act similar without any noticeable
energy difference. Energy spectra in Förner’s geometry ”MF” are presented in Fig. 4.18b,
where all the three regions can be easily recognized. In region I from 500Hz to about
1000Hz, all models extend flatly with only lower energy in model one equation eddy and
Smagorinsky. In region II up to 8000Hz, five curves end up with almost the same TKE
value. Compared with Fig. 4.18a, unstable curves are presented in region III, which proves
stronger influence caused by resonance can exist when the orifice diameter to length ratio
becomes smaller.
Concluded from the correlation and the TKE spectra analysis, the variation of SGS
models also leads to a small difference in flow properties at a Helmholtz Resonator. How-

49
 Simulation Results

ever, in order to reduce the acoustic effect in analyzing, the azimuthal velocity is preferred,
such as transform error arises in one equation eddy based models. In TKE spectra running
some unexpected phenomena occur also special in one equation eddy model, which make it
unreliable. Compare the other three models, Spalart Allmaras model performs always bet-
ter in both correlations and TKE spectra analysis. In the following discussion, this model
is utilized without further emphasize.
Ignoring the difference caused by SGS models, there are still some factors that changes
the TKE spectra characters, which can have even more significant influence. Therefore con-
ditions like in following sections should be considered carefully by the setup of simulations.

4.2.5 Energy Spectra with different SPLs

In our discussion, turbulence is generated simply through the acoustic excitation from in-
let. No turbulence would even exist for low SPL numbers, in contrast however, turbulence
should becomes stronger by an enhanced SPL. A quantified analysis in this section in-
vestigates the SPL influence on turbulence character with the specialty of TKE spectra.

Figure 4.19 presents typical TKE spectra corresponding to different SPLs in both
Hersh’s and Föner’s geometries ”FH” and ”FM”. Focusing first on the ”FM” mesh in
Fig. 4.19b, for different SPL numbers, the available transferred energy from acoustic to
turbulent also differs. Therefore, in each discrete frequency number, a TKE difference can
be distinguished in this figure. Leaving out the 105 dB line, the region I&II as well as region
II&III separations can be recognized in 120 dB and 125 dB at about 1050 Hz and 8000 Hz,
respectively. Since a clear −5/3 slop are measurable in those two lines. However, from the
105 dB curve, no such slop can be found. It can be concluded that for that low SPL as 105
dB no turbulence structure develops. A comparable acoustic influence is also figured out in
Sec. 4.1.3. By Hersh’s geometry ”FH”, the ”pwelch” has windowed out the curves of region
I in all models. Only the separation between region II and III can be found out at about

1e
06 1e
08
Smagorinsky Dynamic
Dynamic Smagorinsky
1e
07 Laminar 1e
09 Laminar
SpalartAllamra Spalart Allamra
OneEq OneEq
1e
08 1e
10
Ekin(f) [m2/s]

Ekin(f) [m2/s]

1e
09 1e
11

1e
10 1e
12

1e
11 1e
13

1e
12 1e
14

1e
13 1e
15
1,000 10,000 1,000 10,000
f [Hz] f [Hz]
(a) Hersh’s geometry. (b) Förner’s geometry.

Figure 4.18: Temporal energy spectra with different SGS models in 120 dB SPL.

50
4.2 Influence of The SGS Modeling on Turbulent Kinetic Energy Spectra and Correlation

10500 Hz. As presented also in Förner’s case, the −5/3 slop can only be recognized in high
energy curves. In low energy curve as the 110 dB one, very flat slop means no homogeneous
full turbulence is structured by the acoustic signal, where the Kolmogolov’s ”−5/3” law
also vanishes. However, the most obvious difference between these two geometry is pre-
sented in high SPL curves. No markable difference between the 130 dB and 140 dB curves
can be found in Fig. 4.19a. An assumption is that by very high SPL of about 130 dB to
140 dB, homogeneous turbulence is generated in all directions in Helmholtz Resonator,
where the dissipation also reaches its highest level. So that the generated turbulence can
not absorb more acoustic energy except those to balance the dissipation. This can also be
proven in acoustic part of Fig. 4.8, where the absolute value of reflection coefficient Rf for
140 dB reads 0.16 and for 120 dB and 130 dB reads about 0.45.

4.2.6 Position and Mesh Quality Effect in Energy Spectra

Above discussions are only valuable when the same grid meshes and investigated points are
used. As an addiction part, how these two factors influence the simulation will be estimated
in this section. The influence of detected position will be discussed as the first part.

Position Variation

Variation in axial direction x = 3ln x = 6ln

Point names XV1 XV2

Table 4.2: Position variation in axial direction with R = 0.5ln .

As shown in Fig. 4.20 and Tabs. 4.2, 4.3, 4.4, three different position variations in
axial, radial and azimuthal directions are discussed in this section, respectively. All these
simulation above are compared with the same mesh ”HF”. Temporal TKE spectra are used
for comparison in-depth.

0.01
140 dB 125 dB
130 dB 1e
05
120 dB
120 dB 105 dB
0.0001
110 dB

1e
06
Ekin(f) [m2 /s]

Ekin(f) [m2 /s]

1e
08
1e
10

1e
10

1e
12

1e
14 1e
15

1,000 10,000 1,000 10,000

f [Hz] f [Hz]

(a) Hersh’s geometry. (b) Förner’s geometry.

Figure 4.19: TKE spectra with different SPLs.

51
 Simulation Results

5ln
RV3
RV3
1ln
RV2 RV2
2ln 3ln
XV1 XV2
0.5ln XV1
RV1 0.3ln
RV1 ALV1
r ALV2

ALV3

ALV4

(a) (b)

Figure 4.20: Position of detected points.

Variation in radial direction R = 0.3ln R = ln R = 5ln

Point names RV1 RV2 RV3

Table 4.3: Position variation in radial direction with x = 3ln .

Variation in azimuthal direction α = π/4 α = π/3 α = π/2 α = 3π/4

Point names ALV1 ALV2 ALV3 ALV4

Table 4.4: Position variation in azimuthal direction with x = 3ln , R = 0.3ln

Results of temporal TKE spectra variations according to different axial positions and
SPLs are presented in Fig. 4.21. As expected, the more distance to neck located point XV2
always turns to be the less energetic one compared with the spectra of point XV1 also with
different SPLs. By lower SPL, turbulent character is more obvious by points XV1 for his
better following of ”-5/3” law. By higher SPL, variation of axial direction becomes smaller
for the better development of turbulent structures.
Similar comparison can also be made for variations in radial and azimuthal directions
as shown in Fig. 4.22. SPL 120dB is chosen since from above axial comparison, more
difference can be expressed by lower SPL. However, variations of both directions are quite
small if compared with that of axial direction, which means the turbulence is therefore
more homogeneous in these two directions.

Mesh Quality Variation

Besides position variations, different mesh qualities also influence the final result. Since
fine mesh can express turbulence with small scales better as the coarse one.
Figure 4.23 present the result of temporal TKE spectra comparison for different mesh

52
4.2 Influence of The SGS Modeling on Turbulent Kinetic Energy Spectra and Correlation

of Hersh’s geometry. As expected, the spectrum with coarse mesh turns to be the least
energetic one since some smallest turbulent scales can’t be modeled. That part of energy
is therefore lost. However, the energetic spectra variation between middle and fine mesh of
Hersh’s geometries ”HM” and ”HF” is quite small. According to the discussion of spatial
correlation in Sec. 4.2.2, over eight cells are available for each smallest turbulent scale in
mesh ”FH”. Usually only four cells are required for a precise simulation, which may be the
case of middle sized mesh ”MH”.

0.1 0.1
x= 6 ln x= 6 ln
x= 3 ln x= 3 ln
0.01 0.01

0.001
0.001
Ekin(f) [m2/s]

Ekin(f) [m2/s]

0.0001
0.0001
1e
05
1e
05
1e
06

1e
06
1e
07

1e
08 1e
07

1e
09 1e
08
10 100 1,000 10,000 10 100 1,000 10,000
f [Hz] f [Hz]

(a) 140dB SPL. (b) 130dB SPL.

1e
05
x= 6 ln
1e
06 x= 3 ln

1e
07

1e
08
Ekin(f) [m2/s]

1e
09

1e
10

1e
11

1e
12

1e
13

1e
14
10 100 1,000
f [Hz]

(c) 120dB SPL.

Figure 4.21: Temporal TKE spectrum with different axial positions and SPLs.

53
 Simulation Results

1e
06
Beta = 3Pi/4
R= 5 ln 0.0001 Beta = Pi/2
1e
07 R= 1 ln
Beta = Pi/3
R= 0.3 ln
Beta = Pi/4

1e
08
1e
06

Ekin(f) [m2/s]
Ekin(f) [m2/s]

1e
09

1e
08
1e
10

1e
11
1e
10

1e
12

1e
13 1e
12
10 100 1,000 10 100 1,000
f [Hz] f [Hz]

(a) TKE spectra variation in radial direction (b) TKE spectra variation in azimuthal
with points RV1-3. direction with points ALV1-4.

Figure 4.22: Temporal TKE spectra with different radius (left) and angles (right).

1e
05

ne Mesh (HF)

1e
06 middle Mesh (HM)
coarse Mesh (HC)
1e
07
Ekin(f) [m2 /s]

1e
08

1e
09

1e
10

1e
11

1e
12

1e
13
1,000 10,000
f [Hz]

Figure 4.23: Temporal TKE spectra with different mesh qualities according to point XV1.

54
4.3 SGS Comparison In Other Criteria

4.3 SGS Comparison In Other Criteria

4.3.1 Visualization Through Vorticity and Q-Criterion
Above discussions are made through different acoustic and fluid dynamic criteria, which
includes a lot of indirect mathematical graphics. However, a visualization of turbulent
structures can also be presented through some turbulent characteristics like vorticity and
Q-criterion. These can also be treated as judgments to evaluate turbulent developments
through LES SGS models in this thesis.
Figure 4.24 shows the comparison of Q-criterion as well as vorticity among different
LES SGS models. The backgrounds of resonator are colored with logarithmic vorticity and
the bubble structures are defined through isosurface of Q-criterion with the value 1000. As
expected from above estimations, at the same instant in time the differences among all SGS
models are small. Considering the background of logarithmic spectrum expressed vorticity,
similar high values mainly located in resonator orifice range. More obvious differences can

(a) Dynamic one equation eddy model (b) One equation eddy model

(c) Smagorinsky model (d) SpalartAllmaras model

Figure 4.24: Turbulence structure visualization with different SGS models with
logarithmic vorticity background and Q-criterion isosurfaces of value 1000.

55
 Simulation Results

be found in Q-criterion, where OneEqEddy model shows less ”bubbles” also with smaller
bubble filled ranges. Among all the other models, Spalart Allmaras model represents more
details within Q-criterion also with a larger effected range. It means in this model, simulated
turbulence influences a larger flow range within the resonator. That may be caused by
more effective acoustic energy absorption within this model. Besides these two models,
Smagorinsky model is more like the Spalart Allmaras, while the dynamic model presents
more similarity to the OneEqEddy.

4.3.2 kRES /kSGS = 4 Criterion

Another criterion can be called as, kRES /kSGS = 4 criterion, where kRES the resolved
kinetic energy of flows and kSGS the modeled energy through SGS definition. According
to Pope’s book [12], where the ratio over 4 can be treated as range that is suitable for the
usage of LES SGS model. However, the ratio under 4 should be regarded as less effective.
As can be expected, this criterion can be influenced both by SGS models and mesh quality
in this case.
In Fig. 4.25, the kRES /kSGS = 4 criterion of one equation eddy models and Homoge-
neous dynamic one equation eddy model model are compared, since in the value of kSGS
is easier to obtain in these models. Also with logarithmic vorticity spectrum background,
blue lines as kRES /kSGS = 4 contour lines encircled the range, where this ratio smaller as
4. Generally the encircled area also covers the high vorticity value range, especially for the
dynamic model. However, the contour lines surrounds less area in OneEqEddy model as the
dynamic one, which means this model is more suitable for the simulation here according to
this criterion. Another difference that can be noticed here locates in the boundary region.
Some areas at boundary layer before and behind the resonator orifice are also surrounded
by the contour lines for the OneEqEddy model, which is caused by the unchangeable defi-
nition coefficient within the definition of model. Since the dynamic model also adapts these
coefficients during the simulation processes, smoother contour lines can thus be followed
along the boundary layer.
Similar comparison can also be made for different mesh qualities like Fig. 4.26. This
time Homogeneous dynamic oneEqEddy model is used, since OneEqEddy model already
shows good performance even for the middle sized mesh. However, according to the result
in Fig. 4.26, finer mesh will not help to make the case suitable for dynamic model, as only
a very small range of by contour lines surrounded area is freed. The only advantage is,
that turbulence in the backing volume seems to be improved for for dynamic model, with
smaller encircled areas.

56
4.3 SGS Comparison In Other Criteria

(a) OneEqEddy model. (b) Homogeneous dynamic oneEqEddy model.

Figure 4.25: kRES /kSGS = 4 criterion (blue contour lines) for OneEqEddy and
Homogeneous dynamic oneEqEddy model model.

(a) With middle mesh. (b) With fine mesh.

Figure 4.26: kRES /kSGS = 4 criterion (blue contour lines) by Homogeneous dynamic
oneEqEddy model model for different mesh sizes.

57
 Simulation Results

4.4 Turbulence Judgment

With the similarity in all LES SGS models, it is necessary to judge whether the generated
vorticity really means turbulence. A criterion already used is the Kolmogolov’s ”-5/3” law
in TKE spectra. Better following of that slope can be found in the region II of temporal
spectra as Fig. 4.18. However, the spatial spectra shows lacking quality by this criterion as
Fig. 4.17. Another criterion is the ratio of turbulent viscosity to molecular viscosity νSGS /ν.
The complete simulated flow viscosity can be expressed as sum of these two viscosities.
Turbulent viscosity νSGS is a modeled parameter in OpenFoam which is theoretically var-
ious defined according to different models like Eqs. (2.65), (2.71). The molecular viscosity
can be computed through Sutherland’s law [34] as:

As T 3/2
ν= , (4.9)
(T + Ts )ρ

(a) t=0.008s (b) t=0.012s

(c) t=0.018s (d) t=0.022s

Figure 4.27: Criterion of νSGS /ν for Spalart Allmaras model in neck range with different
time points.

58
4.4 Turbulence Judgment

where the constants As and Ts are defined as 1.4792 × 10−6 mskg√K and 116K, respectively.
The ratio can be very large to over 100 in some turbulent boundary layers, and quite low
in laminar flows. In turbulent open flow, the value should be between these two conditions.
As a general case in Spalart Allmaras model in Sec. 2.1.5, the ratio is 2.97 × 10−7 in free
stream.
Figure 4.27 shows how the ratio of νSGS /ν develops with time close to neck of Helmholtz
Resonator by Spalart Allmaras model. It can be found, that the ratio of expressed turbulent
structures is generally quite small at about 10−5 . It is larger than the value for the free
stream of 2.97×10−7 , which means turbulent structures in this region are indeed generated.
However, with such small number of ratio, the influence of SGS modeled structure to
the complete system becomes perspectively minimal, which results the high similarity in
simulation in all SGS models as well as the laminar one. The absolute value of this ratio also
shows no change over time even with an enlarged influenced ranges, so that the similarity
of all SGS models vary also quite small over time.

59
5 Conclusion and Outlook

5.1 Conclusion
Turbulent flow is such kind of three dimensional chaotic flows that includes a large number
of length and time scales. Especially in Helmholtz Resonator, the turbulent structures can
be strongly influenced by acoustic waves, which results also in inhomogeneity within the
structure. Therefore, a careful simulation setup in OpenFoam as well as suitable parameter
and criteria selection for estimation is needed.
The turbulence here is assessed through a LES approach, with different compressible
SGS models implemented in OpenFoam. Without mean flow, turbulence is completely
generated by acoustic waves in this work, which makes the turbulent behavior different
from usual ones. To deal with the inhomogeneity and reduce the acoustic perturbation
effect, azimuthal velocity is utilized instead of axial and radial ones for the data post-
processing.
In the acoustic part, parameter variation of different SGS models, SPLs and mesh
qualities are investigated to ensure the findings. Nonlinear behavior of resistance is used
to compare the difference between SGS models as well as the experimental data. However,
almost all models, including the laminar one, showed quite similar results. The Spalart
Allmaras model turns out to be the closest to the experimental data. This conclusion is
also true for changed SPLs. However, the difference between simulated and experimental
data increases by SPL growth. Reflection coefficient difference according to SGS models
is proven to be even minor. But with different SPLs, the amount of energy proportion
that absorbed through turbulence also changes, which resulted in the inconstant value of
reflection coefficient as well as the nonlinear resistance.
In the fluid dynamic part, both criteria of correlations and TKE spectra are based
on azimuthal velocity. Spatial correlation can only be calculated with by hand predefined
probe points, which shows whether the mesh is suitable for the used SGS model. As result,
all SGS models perform quite well in spatial correlation as to model the smallest turbulent
length scales properly. Compared with the spatial one, temporal correlations present the
simulated turbulence qualities in time scales, where the Spalart Allmaras model also shows
best characteristics. Both spatial and temporal correlations can be Fourier transformed to
TKE spectra, with spatial spectrum according to wave numbers and temporal spectrum
according to frequency. Komoglove’s ”-5/3” law helps to adjudicate the turbulence per-
formance in both spatial and temporal spectra. Such kind of slope can be found in all
simulations for the different SGS models. Additionally, Spalart Allmaras model shows to
be more energetic. SPLs influence TKE spectra also significantly, since by lower SPL no
”-5/3” slope can be found. However, at very high SPLs, since reflection coefficient rises and
absorbed acoustic energy proportion drops, the energy included in the spectrum changes

60
5.2 Outlook

minorly.

5.2 Outlook
Since no mean flow is regarded as the basic condition in this thesis, turbulence is simply
generated through acoustic waves. However, in a more general case, pruging or grazing
flow can also generate turbulence in a pure fluid dynamic way. In that case, turbulence
becomes a action that is influenced by both flow velocity and acoustic properties, where
the intensity of turbulence may also increase. How different LES SGS models perform in
that more industrial condition is still unknown. Besides, a connection between spatial and
temporal TKE spectra as the ”frozen turbulence hypothesis” might be validated, which
provides another fluid dynamic judgment criterion for future works.

61
Appendices

62
A Theoretical Deviations

A.1 Deviation of Helmholtz Resonator Equations

lt ln lca
lt ln lca lm lt

dt dm dca dn dt dt dn dca

(a) (b)

Figure A.1: General Helmholtz Resonator with original acoustic wave f (left) and
Helmholtz Resonator with original and reflected acoustic wave f , g (right)

Figure A.1a is a typical general resonator geometry, which consist of the outside tube,
neck part, and resonator cave. Physically, the resonator acts like a mass-spring-damper
system [6]. The mass in neck part and resonator cave is accelerated by the pressure fluc-
tuating from outside tube. The so caused flow motion, especially in near of the neck wall
region, results viscous losses, which turns to flow separation with increasing excitation
amplitudes. These separation converts the acoustic energy to hydrodynamic mode, which
means laminar flow fluctuation by moderate excitation level but turbulent vortices by large
levels [6]. When the acoustic wave frequency is close to the resonator eigenfrequency, the
strongest acoustical losses appears, where the turbulence also mostly exist.
In Figure. A.1a, lt , ln and lca represent the length of computed outside tube, the length
of two resonator necks and the height of resonator cave, respectively. In the mean while dt
reads the diameter of computed outside tube, dn and dm the diameter of orifice and dr the
diameter of resonator.
With the help of Fig. A.1a, we can derive a physical and mathematical describing for
the general resonator from basic mass and momentum conservation by following Keller
and Zauner’s model [3]. The mass equation between the region of necks n and m can be
described as:
∂ρ
V = (Un + un )ρAn − (Um + um )ρAm , (A.1)
∂t
63
 Theoretical Deviations

where Un and Um represent the mean flow velocities in neck regions n, m respectively, and
the un , um , An , Am the departure velocities and flow areas of them. This mass conservation
can theoretically also be formulated only for the mean flow:

Un An − Um Am = 0. (A.2)

Combine Eq. (A.1) and (A.2), the mass conservation for only departure velocities can be
written as:
∂ρ
V = Un ρAn − Um ρAm . (A.3)
∂t
For the momentum equation, only the orifice region m will be first considered. The equation
can then be written as:
∂um ρ
Am lm ρ = Am Pca − Am ζm (Um + um )|Um + um | − Am Pt , (A.4)
∂t 2
where the first and third terms on the right hand-side read the pressure difference between
the resonator cave and outside tube, the second term on right hand-side accounts for the
force due to the acceleration of fluid from pulsation system in resonator cave into the tube
element through orifice m. This force depends
  on the sign of velocity Um + um . ζm is a loss
2
coefficient, which defined as ζm = 1 − AAca
m
. ζm = 1 by loss-free cases,otherwise ζm < 1.
Now the Helmholtz Approximation can be introduced as: the effect from difference of
density in time serious is much smaller as that from difference of velocity. Mathematically
the definition can be expressed as:
   2 
∂ρ ∂um ∂ Um
O O ρ , (A.5)
∂t ∂t ∂t2
   
∂ρ ∂Um
O [Um + um ]  O ρ . (A.6)
∂t ∂t
Differentiating Eq. (A.4) with respect to time and then reduce it by using Eq. (A.5) (A.6)
yields:
∂ 2 um ∂Pca ∂um ∂Pt
lm ρ 2 = − ζm ρ (Um + um ) − . (A.7)
∂t ∂t ∂t ∂t
Suppose the acoustic wave f and g in Eq. (2.107) and (2.108) to be harmonic, which can
be expressed as:  
iωx
f = a exp iωt + + ϕx , (A.8)
c
 
iωx
g = b exp iωt − − ϕx . (A.9)
c
where ϕ  ω/c, and following Rott and Hartunian ϕ can be defined as:
√  
2ωνair γ−1
ϕ = [1 + i] 1 + 1/2 . (A.10)
Dt c Pr

64
A.1 Deviation of Helmholtz Resonator Equations

In this equation, D represents the diameter of tube and νair the kinetic viscosity of air. In
symmetric case a = b, U 0 and P 0 in Eq. 2.107 and (2.108) can be expressed as:
 
0 iωt iωx
U = f − g = ae cosh + ϕx ≈ aeiωt , (A.11)
c
   
0 iωt iωx iωt iωx
P = ρc(f + g) = −ae sinh + ϕx ρc ≈ −ae + ϕx ρc. (A.12)
c c
With the consideration of thermal boundary layer, a coefficient T can be write as:
γ−1 r
1 + Pr1/2 2νair
T = . (A.13)
Dt ω
For the orifice region m, Combination of Eq. (A.10), (A.11) and (A.12) yields:

0 ∂U 0
Pca − Pt0 = [1 + T ]ρlm + T ρlm ωU 0 (A.14)
∂t
Now turn back to the Eq. (A.4). Consider also the thermal boundary layer with Eq. (A.14),
it can be reduced as:
∂ 2 um ∂Pca ∂um ∂um ∂Pt
[1 + Tm ]lm ρ 2
= − ζm ρ |Um + um | − Tm lm ρω − , (A.15)
∂t ∂t ∂t ∂t ∂t
 r
γ−1 πνair
Tm = 1 + 1/2 . (A.16)
Pr 2ωAm
The same strategy can also be used in orifice region n, which leads:
∂ 2 un ∂Pt ∂un ∂un
[1 + Tn ]ln ρ 2
= − ζn ρ |Un + un | − Tn ln ρω , (A.17)
∂t ∂t ∂t ∂t
As we supposed the time-dependent part of pressure in pulsating system (the tube) is
harmonic oscillated:
Pt = bPt c − Pt? sin(ω[t − t0 ]), (A.18)
where b·c refers to time average and t0 accounts for a shift of original time coordinate which
should be chosen suitably. Suppose the time changing of average pressure is incredibly
tc
small: ∂bP
∂t
= 0, Eq. (A.15) is reduced to:

∂ 2 um ∂Pca ∂um ∂um

[1 + Tm ]lm ρ 2
= − ζm ρ |Um + um | − Tm lm ρω − ωPt? cos(ω[t − t0 ]). (A.19)
∂t ∂t ∂t ∂t
Another translation can be included here but actually only strictly valid for isotropic
flow:
∂Pca ∂ρ
= c2 . (A.20)
∂t ∂t
Then the combination of Eq. (A.3) and (A.20) yields:
∂Pca An un ρc2 Am um ρc2
= − . (A.21)
∂t V V
65
 Theoretical Deviations

As a direct outcome from pressure oscillation, the velocity in orifice m and n can also be
represented as harmonic oscillation form. However, the time shift t0 and position difference
finally results a time delay ς for the velocity in orifice n. All these two velocities can be
formulated as:
um = u?m sin(ωt), un = −u?n sin(ω[t − ς]) (A.22)
Now combine all the strategies and derivations above and eliminates the Pt term, the
momentum equation will finally transfered for the orifice m and n like:

∂ 2 um Am ρc2 An ρc2 ∂um

(1 + Tm )lm ρ 2
+ um − un = − ζm ρ |Um + um |
∂t V V ∂t
∂um ∂P
− Tm lm ρω + , (A.23)
∂t ∂t
∂ 2 un An ρc2 Am ρc2 ∂un ∂un
(1 + Tn )ln ρ 2
+ u n − um = −ζn ρ |Un + un | − Tn ln ρω . (A.24)
∂t V V ∂t ∂t
Now we can turn back to the special case which this thesis really interested in, the
Helmholtz Resonator like Figure. A.1b shows. Compare with the general resonator, Helmholtz
Resonator can be mathematically simplified as: An = 0 t0 = 0 Un = Um = 0, so that the
momentum equation can be reduced as:

∂ 2 u Am ρc2 ∂u ∂u ∂P
(1 + T )lm ρ 2
− u = −ζρ |u| − T lm ρω − . (A.25)
∂t V ∂t ∂t ∂t
As the acoustic resonator is always described in the form of frequency, the Eq. (A.18)
and (A.22) from above deviation can also be formulated in the form of u = u? eiωt , P =
P ? eiωt . The Eq. (A.25) should be rewritten as:

Aρc2 ?
(1 + T )lm ρωiu? + (T ρlm ω + ζρu? )u? − iu = −P ? . (A.26)
Vω

A.2 The Definition Of ”pwelch” Function

The following derivation of ”pwelch” function according to Miranda [35] writes in frequency
space since more velocity data in time domain is available, so that in the analysis part more
detailed comparison runs easily. Considering the discrete periodic time signal h[ki ]:
∞
X
h[ki ] = x(ki ts )δ(t − ki ts ), (A.27)
k=−∞

where the ts samples the time, and ki the time instants, such as the Fourier transform in
Eqs. (2.22) 2.21 is written in discrete meaning:
N 0 −1
nk
i2π N i
X
h[ki ] = cn e 0 , (A.28)
n=0

66
A.2 The Definition Of ”pwelch” Function

N0 −1
1 X nk
−i2π N i
cn = h[ki ]e 0 (A.29)
N0 k =0
i

with N0 acts as the discrete fundamental period. In more common non-periodic signal, a
frequency space is preferred as:
∞
X
H(f ) = h[ki ]e−i2πf k . (A.30)
ki =−∞

Different from the h[ki ], Hf definite mainly in continuous space, which make the PSD
definition possible like:
|H(f )|2
Π(f ) = lim . (A.31)
N →0 2N
However Eq. (A.30) is also true in a infinite numbers of data series N < ∞, where the
Discrete Fourier Transform (DFT) acts:
∞
X nki
H(n) = h[ki ]e−i2π N . (A.32)
ki =−∞

Combined with Eq. (A.29), the coefficient cn and Power Spectrum D(n) can be simply
expressed as:
H[n] |H[n]|2
cn = , D(n) = . (A.33)
N N2
Mathematically it is hard to define a PSD through the DFT series H[n], since it is only
defined in discrete frequencies. However PSD estimation is helpful in some discrete problems
like in this thesis, a concept of Empirical Power Spectral Density (EPSD) ED[n] which
donates:
N
X −1 N
X −1 N
X −1
2 2
|h[ki ]| = |H[n]| = ED[n]∆f, (A.34)
ki =−∞ n=−∞ n=0

and the EPSD can be computed as:

|H[n]|2
ED[n] = . (A.35)
N 2 ∆f

Eq. (A.35) is a general definition of PSD in ”Matlab”, however, for big data series, a
optimized and smoothed process can help to make the estimation clear as well as more
efficient. That is why the windowing is introduced in the function of ”pwelch”.

A.2.1 Windowed PSD

Consider that a finite range discrete time signal equals an infinite time signal h[ki ] multi-
plied by a rectangular time window function r[k] as:

hw [ki ] = h[ki ]r[ki ] (A.36)

67
 Theoretical Deviations

where the window function writes for example:


1 if 0 ≤ ki ≤ N − 1
r[ki ] = (A.37)
0 if ki ≥ N

Use the Fourier Transform for all the terms in Eq. (A.36), then discrete Fourier series can
be expressed as:
N
X −1
w
H [n] = H[l]R[n − l], (A.38)
l=0

which generated an finite range of signals. With a variety in non zero range of window
function like r[ki ], the finite windowed range moves, that covers the whole range of to be
analyzed signal. Using different window functions r[ki ] can reduce the effect of leakage,
which is named as the distribution between the adjacent frequencies. Available window
functions like Hanning, Flattop, Blackman-Herris are not discussed here, but can be tried
in defining the structure of ”pwelch”. Overlapping of two neighboring windowed range also
turns to be a parameter that controls the final results.
With help of Eq. (A.38), the Power Spectrum of the windowed signal hw [k] can be
transformed from Eq. (A.33) to:

|H w [n]|2
Dw (n) = . (A.39)
N2
Also by normal computation a EPSD is used, like in this thesis:

|H w [n]|2
ED[n] = . (A.40)
EN BW
where the ENBW called Equivalent Noise Band Width and defines as:
PN −1 2
ki =0 r [ki ]
EN BW = PN N ∆f, (A.41)
( ki −1
=0 r[ki ])
2

68
Bibliography

[1] A.S. Hersh, B.E. Walker, and J.W. Celano. Helmholtz resonator impedance model,
part 1: Nonlinear behavior. AIAA journal, 41(5):795–808, 2003.

[2] P.K. Tang and W.A. Sirignano. Theory of a generalized Helmholtz Resonator. Journal
of Sound and Vibration, 26(2):247–262, 1973.

[3] J.J. Keller and E. Zauner. On the use of Helmholtz Resonators as sound attenuators.
Zeitschrift für angewandte Mathematik und Physik ZAMP, 46(3):297–327, 1995.

[4] K. Förner, M.A. Temiz, W. Polifke, and I.L. Arteaga. On the non-linear influence of
the edge geometry on vortex shedding in Helmholtz Resonators. 2015.

[5] K. Förner, J. Tournadre, P. Martı́nez-Lera, and W. Polifke. Characterization of the

nonlinear response of a Helmholtz Resonator. In N. A. Adams, R. Radespiel, T. Sat-
telmayer, W. Schröder, and B. Weigand, editors, Annual Report, pages 33 – 45. Son-
derforschungsbereich/Transregio 40, 2015.

[6] K. Förner and W. Polifke. Aero-Acoustic Characterization of Helmholtz Resonators

in the Linear Regime with System Identification. In 22nd International Congress on
Sound and Vibration (ICSV22), Florence, Italy, 12 - 16 July 2015.

[7] Q. Zhang and D.J. Bodony. Numerical investigation and modelling of acoustically
excited flow through a circular orifice backed by a hexagonal cavity. Journal of Fluid
Mechanics, 693:367–401, 2012.

[8] L. Davidson. How to estimate the resolution of an les of recirculating flow. In Quality
and Reliability of Large-Eddy Simulations II, pages 269–286. Springer, 2011.

[9] J. Smagorinsky. General circulation experiments with the primitive equations: I. the
basic experiment*. Monthly weather review, 91(3):99–164, 1963.

[10] A.N. Kolmogorov. The local structure of turbulence in incompressible viscous fluid
for very large reynolds numbers. Proceedings: Mathematical and Physical Sciences,
pages 9–13, 1991.

[11] L. Davidson. Fluid mechanics, turbulent flow and turbulence modeling, lecture notes
in msc courses. Technical report, Chalmers University of Technology, 2010.

[12] S.B. Pope. Turbulent flows. Cambridge university press, 2000.

69
 BIBLIOGRAPHY

[13] E. De Villiers. The potential of large eddy simulation for the modelling of wall bounded
flows. 2006.

[14] G.T. Mase, R.E. Smelser, R.M. Smelser, and G.E. Mase. Continuum mechanics for
engineers. CRC series in computational mechanics and applied analysis. Taylor and
Francis, 2009.

[15] A. Cabrena. Linear system identification of the heat transfer behavior of a cylinder
in pulsating cross-flow at Re = 3900 using large eddy simulation. Master’s thesis,
Technische Universität München, 2015.

[16] C. Wagner, T. Hüttl, and P. Sagaut. Large-eddy simulation for acoustics, volume 20.
Cambridge University Press, 2007.

[17] A. W. Vreman. Direct and large-eddy simulation of the compressible turbulent mixing
layer. University of Twente, 1995.

[18] B. Vreman, B. Geurts, and H. Kuerten. Subgrid-modelling in les of compressible flow.

Applied scientific research, 54(3):191–203, 1995.

[19] A. Yoshizawa. Statistical theory for compressible turbulent shear flows, with the
application to subgrid modeling. Physics of Fluids (1958-1988), 29(7):2152–2164,
1986.

[20] C. Fureby, A.D. Gosman, G. Tabor, H.G. Weller, N. Sandham, and M. Wolfshtein.
Large eddy simulation of turbulent channel flows. Turbulent shear flows, 11:28–13,
1997.

[21] U. Germano, M.and Piomelli, P. Moin, and W.H. Cabot. A dynamic subgrid-scale
eddy viscosity model. Physics of Fluids A: Fluid Dynamics (1989-1993), 3(7):1760–
1765, 1991.

[22] U. Piomelli. Large-eddy simulation: achievements and challenges. Progress in

Aerospace Sciences, 35(4):335–362, 1999.

[23] D.K. Lilly. A proposed modification of the germano subgrid-scale closure method.
Physics of Fluids A: Fluid Dynamics (1989-1993), 4(3):633–635, 1992.

[24] S. Ghosal, T. S. Lund, P. Moin, and K. Akselvoll. A dynamic localization model for
large-eddy simulation of turbulent flows. Journal of Fluid Mechanics, 286:229–255,
1995.

[25] U. Piomelli, W.H. Cabot, P. Moin, and S. Lee. Subgrid-scale backscatter in turbulent
and transitional flows. Physics of Fluids A: Fluid Dynamics (1989-1993), 3(7):1766–
1771, 1991.

[26] J.W. Deardorff. A numerical study of three-dimensional turbulent channel flow at

large reynolds numbers. Journal of Fluid Mechanics, 41(02):453–480, 1970.

70
BIBLIOGRAPHY

[27] C. Fureby, G. Tabor, H.G. Weller, and A.D. Gosman. Differential subgrid stress models
in large eddy simulations. Physics of Fluids (1994-present), 9(11):3578–3580, 1997.

[28] P.R. Spalart and S.R. Allmaras. A one equation turbulence model for aerodinamic
flows. AIAA Journal, 94-0439, 1992.

[29] G.A. Ashford and K.G. Powell. An unstructured grid generation and adaptive solution
technique for high-reynolds-number compressible flows. Lecture series-van Kareman
Institute for fluid dynamics, 6:J1–J110, 1996.

[30] M.A. Cárdenas. Influence of Enhanced Heat Transfer in Pulsating Flow on the Damp-
ing Characteristics of Resonator Rings. Dissertation, Technische Universität München,
München, 2014.

[31] T.J. Poinsot and S.K. Lele. Boundary conditions for direct simulations of compressible
viscous flows. Journal of computational physics, 101(1):104–129, 1992.

[32] G. Sebastian. Changes and settings for standard turbulence model implementation in
openfoam (special in spalartallmara),
http://turbmodels.larc.nasa.gov/changestoopenfoam.pdf. 2014.

[33] U. Ingard and H. Ising. Acoustic nonlinearity of an orifice. The journal of the Acous-
tical Society of America, 42(1):6–17, 1967.

[34] W. Sutherland. The viscosity of gases and molecular force. The London, Edinburgh,
and Dublin Philosophical Magazine and Journal of Science, 36(223):507–531, 1893.

[35] A. Miranda. Identification of Sound Sources in Duct Sigularities. Dissertation, Tech-

nische Universität München, München, 2016.

71