Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
Aeroacoustic Noise Prediction and Acoustic Optimization of Mufflers
by
Jobin Xaviour Puthuparampil
A thesis submitted in conformity with the requirementsfor the degree of Master of Applied Science
Department of Mechanical and Industrial EngineeringUniversity of Toronto
© Copyright 2018 by Jobin Xaviour Puthuparampil
Abstract
Aeroacoustic Noise Prediction and Acoustic Optimization of Mufflers
Jobin Xaviour Puthuparampil
Master of Applied Science
Department of Mechanical and Industrial Engineering
University of Toronto
2018
Noise control of large diesel and natural gas generators is achieved through industrial mufflers. Design
of such mufflers relies heavily on general guidelines. But these guidelines are not suitable for complex
mufflers; instead, automated optimization provides an effective means of design. Optimization of a
plug flow muffler (PFM) is conducted in this work with two different approaches: 1) a relatively simple
gradient-descent algorithm (L-BFGS-B) maximizing the transmission loss of the PFM, and 2) a multi-
objective (transmission loss and pressure drop) simulation-based optimization using the Efficient Global
Optimization (EGO) algorithm. The EGO algorithm is shown to be well suited for muffler optimization,
performing vastly better than the commonly used NSGA-II algorithm. In addition, the initial steps
towards the prediction of aeroacoustic noise (self-noise) in mufflers is accomplished through the CFD
simulation of the tandem cylinder benchmark experiment.
ii
This work is dedicated with love to my parents, my siblings, and my precious little niece.
iii
Acknowledgements
I acknowledge and thank Safety Power Inc. (SPI), especially Bob Stelzer, Robert Desnoyers, and Henry
Pong, for all the support, monetary and otherwise, that has led to the successful completion of my
master’s degree. I am especially grateful for SPI’s openness to my ideas and willingness to pursue them.
I would also like to acknowledge MITACS and NSERC for monetary support, as well as CMC and
SOSCIP for providing software and computational resources.
I thank Prof. Pierre Sullivan for the years of guidance and support, through my master’s degree and
prior to it. You were able to simultaneously nurture my desire to explore new ideas while providing the
right amount of supervision to coax me onto the right path. Thank you for providing me the opportunity
to do this master’s and for introducing me to the wonderful folks at SPI. For this, and for everything
not mentioned here, I will always be grateful.
By the end of my two years of master’s, I am in ever greater and unrequitable debt to my parents
and siblings. I thank my brother for giving me the confidence to tackle grad school. Thanks to my
sister for her gracious efforts in proof-reading my work many times during the duration of my master’s.
I thank my parents for all the sacrifices they’ve made for me, including those within the last two years;
they’ve shown me how to love.
I also thank my friends who have kept me sane throughout the last two years, especially those I’ve
grown closer to in this time.
Finally and most importantly, I would like to thank God for His love and blessings. I especially thank
Him for my curious mind and drive in seeking answers. A. M. D. G.
iv
Contents
Acknowledgements iv
Table of Contents v
List of Tables vii
List of Figures viii
Nomenclature x
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2 Background and Theory 3
2.1 Aeroacoustic Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.1.1 Acoustic Analogy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Acoustic Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2.1 Plug Flow Muffler (PFM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2.2 Transfer Matrix Method (TMM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.3 Finite Element Method (FEM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.4 Perforate Impedance Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Pressure Drop Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4 Global Optimization Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4.1 Single-Objective Optimization (SOO) . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4.2 Multi-Objective Optimization (MOO) . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.4.3 Dimensionality Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.5 Muffler Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.6 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3 Methodology 24
3.1 Aeroacoustic Predictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.1.1 Simulation Verification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.1.2 Receiver Noise Calculation - Acoustic Analogy . . . . . . . . . . . . . . . . . . . . 27
3.2 Single-Objective Optimization (SOO) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
v
3.2.1 TMM Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2.2 FE Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2.3 SOO Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.3 Multi-Objective Optimization (MOO) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3.1 FE Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.3.2 Pressure Drop Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.3.3 Dimensionality Reduction - Feature Selection . . . . . . . . . . . . . . . . . . . . . 33
3.3.4 MOO Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.4 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4 Results and Discussion 37
4.1 Aeroacoustic Predictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.1.1 Simulation Verification - Hydrodynamic Results . . . . . . . . . . . . . . . . . . . 37
4.1.2 Receiver Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.2 Single-Objective Optimization (SOO) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.2.1 TMM & FE Model Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.2.2 NLO Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.3 Multi-Objective Optimization (MOO) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3.1 Empirical Pressure Drop Model Validation . . . . . . . . . . . . . . . . . . . . . . 52
4.3.2 Dimensionality Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.3.3 EGO’s Performance in MOO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.3.4 Analysis of EGO Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.4 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5 Conclusion and Future Work 61
5.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.2.1 Aeroacoustics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.2.2 Muffler Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
Appendices 63
A Mann-Whitney U (MWU) Test 63
B Tandem Cylinder Simulations - Additional Results 65
B.1 2D TKE - Line Extractions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
B.2 Span-wise Correlation of Pressure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
Bibliography 67
vi
List of Tables
2.1 Taxonomy of multi-objective EGO algorithms used in present work . . . . . . . . . . . . . 21
3.1 CFD setup overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2 Hydrodynamic metrics used to compare simulation to experimental data . . . . . . . . . . 26
3.3 Parameters defining the PFM and associated values for the Wu et al. muffler [89] . . . . . 28
3.4 Octave band limits and associated engine noise, noise criterion, and NLO weights for SOO 30
3.5 Bounds and constraints applied on SOO . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.6 Geometric parameters defining the PFM, recast for multi-objective optimization as ratios 32
3.7 1/3 octave band limits and TLO weights for MOO . . . . . . . . . . . . . . . . . . . . . . 34
3.8 Bounds and constraints applied on MOO . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.1 Microphone A tonal peaks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.2 Comparison of the simulation in the present work to a selection of those in literature [59] 45
4.3 Pearson correlation coefficient - prediction vs. experimental data . . . . . . . . . . . . . . 46
4.4 Optimized muffler dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.5 Summary of dimensionality reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
A.1 IHV and Iε+ for Objective 1; bolded algorithm is the baseline in MWU test . . . . . . . . 63
A.2 IHV and Iε+ for Objective 2; bolded algorithm is the baseline in MWU test . . . . . . . . 64
A.3 IHV and Iε+ for Objective 3; bolded algorithm is the baseline in MWU test . . . . . . . . 64
A.4 IHV and Iε+ for Objective 4; bolded algorithm is the baseline in MWU test . . . . . . . . 64
vii
List of Figures
2.1 Acoustic sources present in a flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 Sound waves around a muffler element; the pressures used in the TL calculation (Equation
2.10 and 2.12) is highlighted in red. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Cross-section of the plug flow muffler . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.4 Dimensions defining the perforated expansion . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.5 Dimensions defining the perforated contraction . . . . . . . . . . . . . . . . . . . . . . . . 8
2.6 Non-zero mean flow caused by grazing flow and/or bias flow on a perforated surface . . . 11
2.7 Electrical flow resistance network depicting the two components of a plug flow muffler . . 12
2.8 Optimization taxonomy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.9 Pareto front in a multi-objective minimization problem . . . . . . . . . . . . . . . . . . . . 16
2.10 NSGA-II algorithm, image based on [35] . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.11 EGO algorithm as presented by Jones et al. [46] . . . . . . . . . . . . . . . . . . . . . . . 18
2.12 Multi-objective EGO taxonomy, based on [40]; grayed out boxes and dotted lines represent
aspects not explored in the present work. . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.13 Hypervolume calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.14 Additive epsilon indicator (Iε+) calculation . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.1 Tandem cylinder setup, with microphone locations (not to scale) . . . . . . . . . . . . . . 25
3.2 Tandem cylinder meshes - URANS, LES, and DES, respectively . . . . . . . . . . . . . . . 26
3.3 Cross-section of a plug flow muffler and associated dimensional parameters . . . . . . . . 27
3.4 Single-objective optimization (SOO) methodology: evaluation and optimization . . . . . . 28
3.5 Multi-objective optimization (MOO) methodology: evaluation and optimization . . . . . . 32
3.6 Dimensionality reduction through feature selection methodology . . . . . . . . . . . . . . 33
4.1 Mean x-velocity with streamlines from the DES simulation . . . . . . . . . . . . . . . . . 38
4.2 Cp along the surface of the cylinders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.3 Cp′ RMS along the surface of the cylinders . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.4 u along y = 0 in gap between the two cylinders, and after second cylinder . . . . . . . . . 39
4.5 Contours of normalized 2D TKE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.6 PSD of surface pressure at 135° on C1 and 45° on C2 . . . . . . . . . . . . . . . . . . . . . 42
4.7 SPL at the three microphone locations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.8 TMM evaluation of transmission loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.9 FEM evaluation of transmission loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.10 Optimized muffler transmission loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
viii
4.11 Original engine noise, NC-60 criterion, and the output sound pressure levels of the opti-
mized muffler . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.12 Parametric study - output sound pressure levels . . . . . . . . . . . . . . . . . . . . . . . . 49
4.13 Histogram of NLO evaluations at the final result of each optimization . . . . . . . . . . . 50
4.14 Dimension-reduced visualization of one cluster of 883 unique solutions (local minima),
colored by their associated NLO evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.15 Comparison of pressure drop calculated via the empirical pressure drop model and CFD,
for 6 random PFM geometries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.16 Median attainment surfaces comparing the 4 EGO variants . . . . . . . . . . . . . . . . . 54
4.17 Median attainment surfaces comparing different Nevals . . . . . . . . . . . . . . . . . . . . 54
4.18 Median attainment surfaces comparing different sets of initial conditions . . . . . . . . . . 55
4.19 Median attainment surfaces comparing the ParEGO algorithm and NSGA-II . . . . . . . 56
4.20 Hypervolume metric (IHV ) comparing the ParEGO algorithm and NSGAII . . . . . . . . 57
4.21 Binary additive epsilon metric (Iε+) comparing the ParEGO algorithm and NSGAII . . . 57
4.22 Visualization of kriging model developed by the ParEGO algorithm for one set of TLO
and DP weights, with the optimal points on the Pareto front approximation overlaid as
white points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
B.1 2D TKE along y = 0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
B.2 2D TKE along x = 1.5 D and x = 4.45 D . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
B.3 Span-wise correlation of surface pressure at 135° on C1 and C2 . . . . . . . . . . . . . . . 66
ix
Nomenclature
∆P Pressure drop
νo Kinematic viscosity of air
ω Angular frequency
ρo Ambient density of air
σ Perforated plate area porosity
ξ Perforate impedance
co Ambient speed of sound in air
Cp Coefficient of pressure
Cp′ RMS Root-mean-square of coefficient of pressure
Dh Perforated hole diameter
H Dynamic head
Iε+ Binary additive epsilon indicator/metric
IHV Hypervolume indicator/metric
k Angular wave number
LW Acoustic power
LW Sound power
M Mach number
OAR Open area ratio
p Acoustic pressure
Q Volume flow rate
R Flow resistance
Rp Span-wise correlation of pressure
Re Reynolds number
x
S Cross-sectional surface area
To Ambient temperature of air
Tp Perforated plate thickness
U Stream-wise velocity
V Flow velocity
1D One-dimensional
2D Two-dimensional
2DA Two-dimensional axisymmetric
3D Three-dimensional
BFGS Broyden-Fletcher-Goldfarb-Shanno (algorithm)
CFD Computational fluid dynamics
CFS Correlation-based feature selection
DES Detached eddy simulation
DNC Direct noise computation
DP Pressure drop objective
EGO Efficient global algorithm
EIF Expected improvement function
FEM Finite element method
FFT Fast Fourier transform
FWH Ffowcs-Williams and Hawkings
GA Genetic algorithm
LES Large eddy simulation
LHS Latin hypercube sampling
MOO Multi-objective optimization
MWU Mann-Whitney U
NC Noise criterion
NLO Noise level objective
NSGA-II Non-dominated sorting genetic algorithm-II
NSPD Normalized stagnation pressure drop
xi
PCA Principal component analysis
PFM Plug flow muffler
PSD Power spectral density
PSO Particle swarm optimization
QFF Quiet Flow Facility
SA Simulated annealing
SOO Single-objective optimization
SPL Sound pressure level
TKE Turbulent kinetic energy
TL Transmission loss
TLO Transmission loss objective
TMM Transfer matrix method
URANS Unsteady Reynolds-averaged Navier-Stoked
xii
Chapter 1
Introduction
1.1 Motivation
Large-scale (500 kW and above) diesel and natural gas generators produce up to 105 dBA of noise at
a distance of 7 m from the engine [4]. With peak allowable noise levels based on regulations ranging
anywhere from 57 dBA to 72 dBA in suburban to industrial areas respectively, there is a clear need for
acoustic control [4]. Generators have many use cases ranging from providing backup power, to offsetting
energy demand peaks from the grid (peak shaving), to continuous power generation. This versatility
has led to generators being installed in residential zones in recent years, necessitating strict acoustic
attenuation in order to operate legally with minimal disruption to the public.
Noise control of the exhaust of an internal combustion engine is typically achieved with a muffler.
Mufflers are acoustic filters designed to reduce the magnitude of audible noise (20 Hz to 20 000 Hz).
There are two main mechanisms of passive noise attenuation used in mufflers:
1. Reactive silencing: structural elements within the muffler passively create sound waves that are out
of phase with the incoming sound, causing destructive interference that reduces the amplitude of
the sound. This mechanism is most effective in the low-frequency range (< 500 Hz). It is achieved
by tuning the physical dimensions of the components inside the muffler to change the phase of
internally reflected waves.
2. Absorptive or dissipative silencing: viscous effects dampen the sound energy through the use of
fibrous or foam-like acoustic material. Its effect is more pronounced in the mid to high frequency
range, and the use of more acoustic material generally results in better attenuation. Forcing the
flow to pass through perforated surfaces (solid surface with an array of small holes with a diameter
of a few millimeters) is another method of dissipative silencing as viscosity is quite effective at noise
attenuation at these scales. The perforation geometry can also be tuned to affect performance.
Tuning the physical dimensions of the acoustic components inside the muffler is crucial to the operation
of a muffler. General guidelines exist for such tuning, but only for basic muffler designs. In order to
truly design the best muffler for a given generator, a computer-based automated optimization process
is necessary. Such optimization aims to maximize or minimize an objective by varying a set of pa-
rameters, given a set of constraints. Thus, tuning a muffler’s geometry to produce maximum acoustic
1
Chapter 1. Introduction 2
performance is well suited for an automated optimizer. In addition to maximizing acoustic performance,
a muffler must be designed to create the least amount of back-pressure on the engine as possible, as
higher back-pressures reduce engine efficiency. To account for this, optimization which simultaneously
tries to maximize acoustic performance while minimizing back-pressure can be conducted. Additionally,
complex mufflers require computer simulation in order to accurately determine its acoustic and flow
characteristics; mathematical descriptions will generally be ineffective when a design is of industrial
complexity and size. Such simulations can take a considerable amount of computational time, but are
necessary in the optimization process to retain accuracy. Thus, it becomes crucial to use an efficient
optimization algorithm when conducting simulation-based optimization. All aforementioned ideas are
explored in the single-objective and multi-objective optimization sections of this work.
As the turbulent exhaust gasses from the engine flow through the muffler, it can produce flow-induced
noise. This phenomenon, also referred to as self-noise, usually produces broadband (across a large por-
tion of the frequency spectrum) noise. However, if the flow interacts with sharp edges or perforations
within the muffler, it has the potential of creating tonal noise as well. The first steps towards the predic-
tion of such noise is conducted in the aeroacoustic modelling sections of the thesis through a benchmark
problem.
1.2 Objectives
This work has two major objectives:
1. Accurate and computationally cost-effective means of predicting aeroacoustic noise
2. Develop a methodology for the automated optimization of large industrial mufflers, simultaneously
emphasizing speed and accuracy of the optimization
Chapter 2
Background and Theory
2.1 Aeroacoustic Modelling
Aeroacoustic modelling assists the prediction of sound produced from flow turbulence. This sound is
generally broadband, with sound energy spread over a large range of audible frequencies. However,
certain flow mechanisms, such as vortex shedding, can add a tonal component to the noise generated.
There are various geometries in a muffler that can lead to tonal flow generated noise, including sharp
edges with perpendicular flow or perforated surfaces. This is in addition to broadband aeroacoustic noise
generated by fast moving flow which can be approximated using empirical relationships, and is a strong
function of the flow velocity [1].
Direct noise computation (DNC) can simulate the pressure variation as a function of time by solving
the compressible flow equations. The pressure field is then decomposed into a hydrodynamic component
and an acoustic component, following which the acoustic component can be analyzed in the frequency
domain to reveal the sound pressure level (SPL) at various frequencies. DNC proves to be extremely
challenging because of the implementation of:
1. high order spatial discretization schemes to account for the large mismatch in energy scales of the
hydrodynamic and acoustic components (acoustic energy is smaller than the hydrodynamic energy
by an order O(M4) where M is the Mach number) [88]
2. large domains, required to capture the relatively large lengths of acoustic waves, compared to
hydrodynamic length scales; computational requirement is proportional to Re3t/M
4 where Ret is
the turbulent Reynolds number [74]
3. Absorbent boundary conditions, which produce minimal spurious reflections into the domain [88]
2.1.1 Acoustic Analogy
Acoustic analogies, originally proposed by Lighthill [57, 58] and developed by Curle [24] and Ffowcs-
Williams & Hawkings (FWH) [32], aim to rectify the challenges seen with DNC by decoupling the
hydrodynamic and acoustic fields. Traditional time-resolved flow simulations are used to estimate the
turbulent properties of the flow. Then, the strengths and locations of acoustic source terms such as
3
Chapter 2. Background and Theory 4
Figure 2.1: Acoustic sources present in a flow
monopoles, dipoles, and quadrupoles are determined based on the turbulence in the flow. These three
sources radiate acoustic energy with different directivity and efficiencies (Figure 2.1). The contribution
of these acoustic sources are assessed to estimate the acoustic pressure fluctuations at far-field acoustic
listeners locations, which can subsequently be converted to the frequency domain to show the SPL at
various frequencies.
Lighthill manipulated basic flow equations to convert them into a form similar to the inhomogeneous
wave equation, creating the aeroacoustic analogy [57, 58]. It is derived beginning with the conservation
of mass and momentum
∂ρ
∂t+∂ (ρui)
∂xi= m (2.1)
∂ (ρui)
∂t+∂ (ρuiuj)
∂xj= fi −
∂p
∂xi+∂τij∂xj
(2.2)
where ρ is the density and u is the velocity, m is a mass source term, p is the hydrostatic pressure, and
τij is the shear stress tensor. Subtracting the divergence of Equation 2.2 from the time derivative of
Equation 2.1 to eliminate the term with ρui yields
∂2ρ
∂t2=∂m
∂t− ∂fi∂xi
+∂2 (ρuiuj)
∂xixj+∂2p
∂x2i
− ∂2τij∂xixj
(2.3)
∂2ρ
∂t2=∂m
∂t− ∂fi∂xi
+∂2
∂xixj(ρuiuj + pδij − τij) (2.4)
Subtracting c2o∇2ρ = c2o∂2
∂x2i
(ρ) = ∂2
∂xixj
(c2oρδij
)from both sides, where co is the ambient speed of sound,
gives∂2ρ
∂t2− c2o∇2ρ =
∂m
∂t− ∂fi∂xi
+∂2
∂xixj
(ρuiuj +
(p− c2oρ
)δij − τij
)(2.5)
which has a similar form to the inhomogeneous wave equation. The right hand side of Equation 2.5
consists of the source term (S), which has three terms, the latter of which is simplified by Lighthill by
Chapter 2. Background and Theory 5
an argument of practicality [57]:
S = Smonopole + Sdipole + Squadrupole (2.6)
Smonopole =∂m
∂t(2.7)
Sdipole = − ∂fi∂xi
(2.8)
Squadrupole =∂2
∂xixj
(ρuiuj +
(p− c2oρ
)δij − τij
)≈ ∂2
∂xixj(ρuiuj) (2.9)
Lighthill’s original formulations focused on quadrupoles only, such that S = Squadrupole [57, 58]. Lighthill
provided an exact solution to Equation 2.5 using Green’s functions with the simplified Equation 2.9.
Curle expanded Lighthill’s exact solution to Equation 2.5 while considering S = Sdipole + Squadrupole
(i.e., sum of Equations 2.8 and simplified 2.9) where the dipole sources, distributed on the surface of a
stationary object in flow, contain the noise created by the force exerted by the fluid on the solid [24].
The exact solution provided with the FWH analogy [32] includes monopoles and thus uses the complete
source term (Equation 2.6, with simplified Equation 2.9). Monopoles are generated through the physical
displacement of fluid volumes by solid surfaces.
The relative importance of each source varies based on the type of flow encountered; the dipole is
dominant when periodic flow separation occurs [37], and will be the focus of the aeroacoustic modelling
in the present work (Sections 3.1 and 4.1).
2.2 Acoustic Modelling
Transmission loss (TL) is the difference in acoustic power between the sound travelling toward the muffler
on the inlet side (LWi) and sound travelling away from the muffler on the outlet side (LWo); direction of
sound travel is important as any reflections will transport sound power in the opposite direction (Figure
2.2).
TL = LWi − LWo (2.10)
Sound power (LW ) normal to a surface, such as the inlet or outlet cross-section of a muffler, is
LW =Sp2
ρoco(2.11)
where S is the cross-sectional area of the surface, p is the acoustic pressure, and ρo and co are the density
and speed of sound of the medium. Expressing TL as a decibel (dB) using Equations 2.10 and 2.11,
with the direction specified, results in
TL = 10 log10
∣∣∣∣ Sip2i+
Sop2o+
∣∣∣∣ (2.12)
Chapter 2. Background and Theory 6
Figure 2.2: Sound waves around a muffler element; the pressures used in the TL calculation (Equation2.10 and 2.12) is highlighted in red.
Figure 2.3: Cross-section of the plug flow muffler
where pi+ is the acoustic pressure of the incoming incident wave, po+ is the acoustic pressure of the
outgoing transmitted wave, and Si and So are the inlet and outlet cross-sectional areas, respectively.
Various analytical and numerical methods exist to predict the TL of a given muffler. Analytical meth-
ods include the transfer matrix method (TMM) [63], modal meshing approach [62], and exact 2D/3D
analytical solutions [e.g., 42, 75], while numerical alternatives are typically the finite element method
(FEM) and boundary element method (BEM). The present work utilizes TMM and FEM for analytical
and numerical predictions, respectively.
2.2.1 Plug Flow Muffler (PFM)
This work focuses on the plug flow muffler (PFM) (Figure 2.3), a muffler consisting of a perforated
expansion followed by a perforated contraction. Engine exhaust gas enters the first section of the inner
tube, is forced out through the perforations on the inner tube to the surrounding annular chamber due
to the presence of the flow plug, rejoins in the second portion of the inner tube after passing through the
second perforated surface, and finally exits the muffler. This type of muffler has a relatively high pressure
drop due to the presence of the plug as well as the perforated surfaces, but yields high transmission losses
in a small form factor. It utilizes both reactive and dissipative mechanisms of noise control through the
various tubes and perforated surfaces, respectively. The PFM can be used as a component within a larger
muffler, as it is well suited for attenuating the mid frequencies (500 Hz to 2000 Hz); this frequency range
is generally not addressed easily with a traditional industrial muffler’s reactive or dissipative sections.
Chapter 2. Background and Theory 7
2.2.2 Transfer Matrix Method (TMM)
The transfer matrix method (TMM) requires the calculation of a 2 x 2 transfer matrix, also known as
the “four-pole parameters”. Sullivan and Crocker proposed the first mathematical model of a perforated
muffler and analyzed a non-plug concentric tube muffler [85]. Jayaraman and Yam modified Sullivan and
Crocker’s model with a decoupled solution applied to the two cross-flow elements found in a plug flow
muffler: a perforated expansion (Figure 2.4) and a perforated contraction (Figure 2.5) [43, 86]. The key
improvement was that each muffler element only required one set of calculations. However, a limitation
in their derivation defined the mean flow velocity in the inner tube and outer annulus be the same. It
is clear that the change in cross-sectional area from the inner to the outer section results in a change in
mean flow velocity, except for the zero mean flow case. Munjal et al. [65] addressed this limitation and
produced an analytical decoupling method which accounted for the difference in mean flow velocity in the
interior and exterior annulus. This approach is well suited for the analysis of the plug flow muffler shown
in Section 2.2.1. TMM splits the muffler into a sequence of its sub-components, analyzed individually,
and then combined to reveal their total effect. For a given frequency, each sub-component results in a
transfer matrix (T ), (p1
ρocou1
)= T
(p2
ρocou2
)=
[TA TB
TC TD
](p2
ρocou2
)(2.13)
The transfer matrix relates the acoustic pressure (p) and the velocity (u) at the inlet of the sub-
component, subscript 1, to the outlet condition, subscript 2; ρo and co are the ambient density and
speed of sound. The muffler’s complete transfer matrix is
Tmuffler = T1T2 . . . Tn =
[TA TB
TC TD
]1
[TA TB
TC TD
]2
. . .
[TA TB
TC TD
]n
(2.14)
where the subscript represents the sub-component under consideration. In the case of a plug flow muffler,
two acoustics sub-components need to be considered (n = 2): a perforated expansion and a perforated
contraction. Munjal shows that the transfer matrix for a perforated expansion chamber (Figure 2.4) is
[64]
TA = P12 +A1A2 TB = P14 +B1A2 (2.15)
TC = P32 +A1B2 TD = P34 +B1B2
A1 = (X1P22 − P42) /F1 B1 = (X1P24 − P44) /F1
A2 = P11 +X2P13 B2 = P31 +X2P33
F1 = P41 +X2P43 −X1 (P21 +X2P23)
X1 = −i tan (koLa1) X2 = i tan (koLb1)
where La1 and Lb1 refer to the lengths of the solid sections of tube before and after the expansion’s
perforated section. The transfer matrix for a perforated contraction chamber (Figure 2.5) is [64]
Chapter 2. Background and Theory 8
Figure 2.4: Dimensions defining the perforated expansion
Figure 2.5: Dimensions defining the perforated contraction
TA = P21 +A1A2 TB = P23 +B1A2 (2.16)
TC = P41 +A1B2 TD = P43 +B1B2
A1 = (X1P11 − P31) /F1 B1 = (X1P13 − P33) /F1
A2 = P22 +X2P24 B2 = P42 +X2P44
F1 = P32 +X2P34 −X1 (P12 +X2P14)
X1 = −i tan (koLa2) X2 = i tan (koLb2)
Chapter 2. Background and Theory 9
where La2 and Lb2 refer to the lengths of the solid sections of tube before and after the contraction’s
perforated section. In Equations 2.15 & 2.16, the matrix P is
P =
P11 P12 P13 P14
P21 P22 P23 P24
P31 P32 P33 P34
P41 P42 P43 P44
= [A(0)][A(1)]−1 (2.17)
where the A(x) matrix’s elements are (for i = 1, 2, 3, 4)
A1,i = ψ3,ieβix A2,i = ψ4,ie
βix (2.18)
A3,i = − eβix
iko +M1βiA4,i = − ψ2,ie
βix
iko +M2βi
ψ and β are the eigenmatrix and eigenvector, respectively, of the following matrix−α1 −α3 −α2 −α4
−α5 −α7 −α6 −α8
1 0 0 0
0 1 0 0
(2.19)
α1 = − iM1
1−M21
(k2a + k2
o
ko
)α2 =
k2a
1−M21
α3 = − iM1
1−M21
(k2a − k2
o
ko
)α4 = −
(k2a − k2
o
1−M21
)α5 = − iM2
1−M22
(k2b − k2
o
ko
)α6 = −
(k2b − k2
o
1−M22
)α7 = − iM2
1−M22
(k2b + k2
o
ko
)α8 =
k2b
1−M22
ko = ω/co
M1 = V1/co M2 = V2/co
k2a = k2
o −4ikoD1ξ
k2b = k2
o −4ikoD1
(D22 −D2
1) ξ
where ω is the angular frequency, ξ is the specific perforate impedance (Section 2.2.4), and V1 and V2
are the mean flow velocities in the inner tube and outer annulus, respectively.
Once the complete transfer matrix is calculated (Equation 2.14), the transmission loss is [65]
TL = 10 log10
[SiSo
(1 +M1
1 +M2
)2 ∣∣∣∣TA + TB + TC + TD2
∣∣∣∣]
(2.20)
TMM is a 1D method. It only considers the propagation of plane waves, and so, is only valid below the
Chapter 2. Background and Theory 10
cutoff frequency (fc) for the first circumferential mode [11, 31], above which higher wave modes occur:
fc =1.84coπD2
(2.21)
2.2.3 Finite Element Method (FEM)
An alternate method to predict the TL of a muffler is the finite element numerical discretization of the
Helmholtz equation. This equation can be written with pressure as the scalar of interest as
∇ ·[− 1
ρo(∇p− qd)
]− ω2p
ρoc2o= Qm (2.22)
where ρo is the density, co is the speed of sound, ω is the angular frequency, Qm is the monopole source
term, and qd is the dipole source term, and the pressure p is a time-harmonic wave of the form
p (x, t) = p (x) eiωt (2.23)
Equation 2.22 can be parametrically solved for a range of frequencies to determine the pressure dis-
tribution within a domain at each frequency. Following this, the transmission loss is calculated as a
decibel:
TL = 10 log10
(LWi
LWo
)(2.24)
where LWi and LWo correspond to the acoustic power at the muffler inlet and outlet respectively.
Acoustic power is calculated using the pressure field solution integrated over the surface (S) of interest
(cross-sectional surface of the inlet or outlet):
LWs = 2
∫S
|p|2
2ρoco(2.25)
Note that accuracy is not limited to a cutoff frequency as with TMM; it is limited by the size of the
mesh used. However, the computational cost of FE is at least 120 times TMM, in the case of the plug
flow muffler, discouraging the use of FE in muffler optimization.
2.2.4 Perforate Impedance Model
An important feature of a plug flow muffler is the perforated surfaces that the flow moves through.
Accurate descriptions of the acoustic effects of perforated surfaces exist primarily in empirical work.
In general, acoustic impedance of perforated surfaces have been classified in (a) zero mean flow and
(b) non-zero mean flow. Zero mean flow, allowing only for the oscillatory acoustic flow, is a physically
unrealistic condition as mufflers are generally always subject to some amount of exhaust flow. However,
it is still a valuable tool in developing the understanding behind perforated surfaces. Non-zero mean
flow can be caused due to one or both of the following flows: bias flow or grazing flow. Bias flow involves
flow passing perpendicularly through the perforations, whereas grazing flow considers flow parallel to
the perforated plate (Figure 2.6).
Chapter 2. Background and Theory 11
Grazing FlowBias Flow
Figure 2.6: Non-zero mean flow caused by grazing flow and/or bias flow on a perforated surface
Zero Mean Flow
Three models for zero mean flow perforate impedance (ξ) are considered in this work:
1. Sullivan [84]
ξ =p
ρocou=
6× 10−3
σ+ik (Tp + 0.75Dh)
σ(2.26)
2. Bento Coelho [25]
ξ =p
ρocou=
1
ρoco(R+ iX) (2.27)
R =1
σ
[ρo
(d′
d′′
)√8νoω +
(ρo8co
)(ωDh)
2
]X =
(ωρoσ
)(d′′ +
(d′
Dh
√8νoω
))d′ = Tp +Dh
d′′ = Tp +
(8
3π
)Dh
(1− 0.7
√σ)
3. Bauer [9]
ξ =p
ρocou=
√8νokco
(1 +
Tp
Dh
)σ
+ik (Tp + 0.25Dh)
σ(2.28)
where k is the angular wave number, Tp is the thickness of the perforated plate, Dh is the diameter of
the perforated holes, σ is the area porosity of the perforated sections, ρo is the ambient air density, co
is the speed of sound in ambient conditions, νo is the ambient kinematic viscosity, and ω is the angular
frequency.
Non-Zero Mean Flow
A unified model for perforate impedance is used in this work to account for non-zero mean flow [30]:
ξ = θ + iχ (2.29)
θ = Re
{jk
σCD
[Tp
F (µ′o)+
δreF (µo)
fint
]}+
1
σ
[1− 2J1 (kDh)
kDh
]+
0.3
σMg +
1.15
σCDMb
χ = Im
{jk
σCD
[Tp
F (µ′o)+
0.5Dh
F (µo)fint
]}
Chapter 2. Background and Theory 12
Figure 2.7: Electrical flow resistance network depicting the two components of a plug flow muffler
F (KDh) = 1− 4J1 (KDh/2)
KDhJ0 (KDh/2)K =
√− iωνo
=
√− iωρo
µoK ′ =
√− iων′o
=
√− iωρo
µ′o
δre = 0.2Dh + 200D2h + 16000D3
h fint = 1− 1.47√σ + 0.47
√σ3
where Mg is the grazing flow Mach number, Mb is the bias flow Mach number, J0 and J1 are Bessel
functions of the first kind (first and second order respectively), CD is the orifice discharge coefficient as
given by Elnady [30], and µ′o = 2.179µo.
2.3 Pressure Drop Modelling
Flow networks can be modelled using using the circuit and resistance network analogy. The plug flow
muffler is modelled as a simple resistance network (Figure 2.7). The flow resistance (R) of any component
in a flow path is
R =∆P
Q|Q|(2.30)
where ∆P is the pressure drop across the component and Q is the volume flow rate. Equation 2.30 can
be simplified for a unidirectional flow to be
R =∆P
Q2(2.31)
The total equivalent resistance (Req) of the network in Figure 2.7 is given as
Req = Rexp +Rcon (2.32)
Therefore, the pressure drop across the plug flow muffler (PFM) is
(∆P )PFM = ReqQ2 = (Rexp +Rcon)Q2 (2.33)
Munjal et al. [66] experimentally developed relationships to describe the flow resistance of both per-
forated expansions (Rexp) and perforated contractions (Rcon) based on geometrical parameters. The
flow resistances are represented by normalized stagnation pressure drops (NSPD = ∆PH ), where H is the
dynamic head of the incoming flow,
R = (NSPD)H
Q2(2.34)
Chapter 2. Background and Theory 13
Combining equations 2.33 and 2.34, where Q is constant throughout the muffler:
(∆P )PFM =
[NSPDexp
(H
Q2
)+ NSPDcon
(H
Q2
)]Q2
= (NSPDexp + NSPDcon)H
= (NSPDexp + NSPDcon)
(1
2ρoV
2inlet
)(2.35)
where ρ is the density of air or combustion gas passing through the muffler and Vinlet is the area-averaged
velocity at the inlet. Munjal provides the following relationships for calculating the NSPD [66]
NSPDexp =
(∆P
H
)exp
= 3.136(OARexp)−1.391 0.3 < OARexp < 2.2 (2.36)
NSPDcon =
(∆P
H
)con
= 2.208(OARcon)−1.5818 0.31 < OARcon < 1.66 (2.37)
Open area ratio (OAR) compares the open area present on the surface of the perforated tube to the
cross-sectional area of the perforated tube. OAR is
OAR =πD1Lσπ4D
21
=4Lσ
D1(2.38)
where D1 is the diameter of the perforated tube, L is the length of the perforated section of tube, and
σ is the open area porosity of the perforated tube.
2.4 Global Optimization Methodology
Sections 2.2 and 2.3 provide a means of analyzing the plug flow muffler’s two main characteristics: trans-
mission loss and pressure drop. Aside from general design trends (e.g., increased OAR leads to lower
pressure drop), a specific design methodology is not directly discernible from the analysis above, due
to the complex nature of the transmission loss prediction. In order to design the best possible muffler
for a given set of operating conditions and constraints, automated design via optimization is the solution.
Optimization is the process of finding the best performing state of an objective function, applied in
this work to find designs that maximize acoustic performance. Formally, it is
minimizex
f(x)
subject to gi(x) ≤ 0, k = 0, . . . ,m,
hj(x) = 0, k = 0, . . . , p
(2.39)
where f(x) is the objective function to be optimized (traditionally, a minimization) with an n-variable
long vector x, gj(x) and hi(x) are sets of inequality and equality constraints respectively, and m ≥ 0
and p ≥ 0 [13]. There are multiple ways to look at the taxonomy of optimization problems as the field
is very large and there exists many interrelated ideas, but Figure 2.8 provides an overview from the
perspective relevant to this work. The five major attributes of an optimization problem are:
Chapter 2. Background and Theory 14
1. Variables: The goal of any optimization problem is to determine the variable value(s) that provide
the optimal conditions for f(x). There can be just one variable to optimize (n = 1), in the
uni-variate case, or as more often seen in engineering, multiple variables (n > 1), leading to multi-
variate optimization. These variables can be continuous or discrete or a combination of both.
2. Objectives: In single-objective optimization, there is only one objective function to optimize,
leading to a clearly defined optimality criterion. However, in multi-objective optimization, the
objectives may conflict; there may not be a clear optimality criterion leading to one ideal x where
all objectives are optimized. Instead, there is a set of well performing solutions that satisfy each
objective to a different degree - i.e., a Pareto front.
3. Constraints: If constraints exist, they come in two forms, inequality (gi) or equality (hj), that
can be applied as linear constraints or non-linear constraints. Linear constraints are imposed on
the variable x while non-linear constraints are applied on the evaluation of the function, f(x). If
m = 0 and p = 0, it becomes an unconstrained optimization problem. In addition to linear and
non-linear constraints, upper and lower bounds on the variables (x) can be established to limit the
optimization within a predefined boundary; in an engineering context, these may be boundaries
dictated by physical limitations or other costs.
4. Linearity: A linear optimization problem requires both the objective function and the constraints
to be linear functions. If this condition is not met, the problem is termed non-linear. In a linear
problem, any local optima discovered is always the global optima as well, allowing for simpler and
more robust optimization algorithms, and the ability to solve large problems.
5. Convexity: A function is convex when any line drawn between two points on the function (chord)
lies above the function [13]. A convex minimization problem requires all objective function as well
as the constraints to be convex. In a convex minimization problem, any local minima discovered is
also the global minima. However, in a non-convex problem, there is no such guarantee, requiring
the search of all local minima to truly guarantee the global minima. Thus, if a problem has multiple
local minima, it can be classified as non-convex.
The present problem is bounded, multi-variate (continuous variables), non-linear, and non-convex. Thus,
a global optimization strategy able to balance local (narrowing down on optima) and global (exploration)
searches is required. Both single-objective and multi-objective formulations were considered in this work,
each achieving separate goals.
2.4.1 Single-Objective Optimization (SOO)
A single-objective formulation of the present problem involves the optimization of transmission loss alone.
Many algorithms exist for non-linear single-objective optimization, including the Broyden-Fletcher-
Goldfarb-Shanno (BFGS) algorithm [14, 33, 36, 77], non-linear conjugate gradient methods such as
Fletcher-Reeves [34], Levenberg-Marquardt algorithm [56, 61], and the Nelder-Mead method [67]. The
present work utilized the BFGS algorithm to solve the single-objective optimization problem.
Chapter 2. Background and Theory 15
Figure 2.8: Optimization taxonomy
Broyden-Fletcher-Goldfarb-Shanno (BFGS)
The BFGS algorithm was discovered independently by Broyden [14], Fletcher [33], Goldfarb [36], and
Shanno [77]. It is a second-order iterative quasi-Newton method that calculates the gradient of a func-
tion and estimates the inverse of the Hessian to guide itself towards a local minima.
A function f(x) can be estimated by its second-order Taylor approximation
f (xk + ∆x) ≈ f(xk) +∇f(xk)T∆x+1
2∆xT∇2f(xk)∆x (2.40)
where xk is the current approximation of the minima, ∆x is a small step to explore the function at a
nearby value, ∇f(xk) is the gradient at xk, and ∇2f(xk) is the Hessian. The gradient of Equation 2.40
is
∇f (xk + ∆x) ≈ ∇f(xk) +∇2f(xk)∆x (2.41)
To determine the minima, the gradient (Equation 2.41) is set to 0 and rearranged to solve for ∆x,
resulting in
∆x = −(∇2f(xk)
)−1∇f(xk) (2.42)
In the case of a quasi-Newton method, the Hessian is approximated as B and Equation 2.42 is written
as
∆x = −B−1∇f(xk) (2.43)
Iterative updates of B (or B−1) to reveal better approximations of the true Hessian (or inverse Hessian)
is the key aspect of BFGS, as is achieved via
Bk+1 = Bk +yky
Tk
yTk ∆xk− Bk∆xk (Bk∆xk)
T
∆xTkBk∆xk(2.44)
where ∆xk = αkB−1k ∇f(xk) with αk determined from the Armijo rule [5] and yk = ∇f(xk+1)−∇f(xk).
Chapter 2. Background and Theory 16
Figure 2.9: Pareto front in a multi-objective minimization problem
Memory efficient versions of BFGS have been developed, known as limited-memory BFGS (L-BFGS),
which limit the amount of data used in the Hessian approximation [68, 16]. Additionally, BFGS was
originally formulated as an unconstrained optimization algorithm; variable bounds constraints have been
implemented on the limited memory BFGS in an algorithm referred to as L-BFGS-B [15]. It is important
to note that the BFGS algorithm does not guarantee global minimums, but a local one. To increase the
probability of discovering the global minima, a common technique is to employ a multi-start framework,
where multiple independent optimizations are run in parallel, with each starting at a different random
initial condition [28].
2.4.2 Multi-Objective Optimization (MOO)
Attempting to optimize both transmission loss and the pressure drop through the muffler leads to a
multi-objective problem with competing objectives. The optimal solution to a multi-objective problem
is a Pareto front when nobj = 2 or Pareto surface for nobj > 2. The Pareto front consists of a set of
Pareto efficient solutions to the optimization problem; Pareto efficient (equivalently, non-dominated)
solutions are those where any improvement in one objective must be met with a deterioration in another
objective (Figure 2.9). In the engineering context, a Pareto front is a trade-off curve, representing the
interplay between competing objectives.
Genetic Algorithm
Evolutionary algorithms such as the genetic algorithm are well suited for the discovery of Pareto fronts
as multiple solution candidates are considered and improved through the optimization process. Non-
dominated sorting genetic algorithm-II (NSGA-II) has been shown to be an effective variant of the genetic
algorithm in multi-objective optimization [27], outperforming other established evolutionary algorithms
such as Pareto archived evolution strategy (PAES) [52] and strength pareto evolutionary algorithm
(SPEA) [98]
NSGA-II’s algorithm flowchart is seen in Figure 2.10. It begins with a randomly generated population of
size npop whose objective functions are first evaluated. This population is ranked via the non-dominated
sort; the population is organized into multiple levels of non-domination, resulting in multiple fronts.
After the initial sorting, the child population is created (i.e., next generation) using standard genetic
algorithm processes such as selection, crossover, and mutation, which are described extensively in other
Chapter 2. Background and Theory 17
Figure 2.10: NSGA-II algorithm, image based on [35]
sources [98]. Additionally, to promote exploration of the entire Pareto front, the crowding of solutions
is discouraged via calculation of a crowding distance metric.
The elitist aspect of NSGA-II combines the parent and children populations together and applies the
non-dominated ranking procedure and crowding distance calculations on the combined set, from which
the best performing npop individuals are chosen, with higher preference to non-domination - if two so-
lutions share the same non-domination rank, the less crowded solution is chosen. The process loops
back to the creation of children, resulting in he next generation, and continues until a stopping criterion
is satisfied, typically maximum number of generations ngen. This results in Nevals = npop · ngen total
evaluations of the objective function.
Efficient Global Optimization
Efficient global optimization (EGO) is a surrogate model based optimization technique first proposed as
a single-objective optimization algorithm by Jones et al. [46]. EGO uses kriging, the process of fitting
a Gaussian process model, to create a surrogate model to estimate where the optima lies, along with
an expected improvement function (EIF) to determine where the next evaluation point must be. The
algorithm for EGO as developed by Jones et al. is shown in Figure 2.11; the single objective EGO will
be described before moving onto the multi-objective formulation. EGO begins with the generation of a
set of initial points using Latin hypercube sampling (LHS) and subsequent calculation of the objective
function value at each point. This information is used to create the Gaussian process-based surrogate
model, providing both an estimate of the objective value at a given location along with the uncertainty
of the estimate. The process only continues if the magnitude of the cross-validated standardized error
residuals of the surrogate model (or the surrogate model of the log of the objective space) is less than 3
- a sufficient range for the error as stated by Jones et al. The next step is to determine the location for
the next objective function evaluation.
EGO determines the location of the next evaluation by balancing local and global searches through
Chapter 2. Background and Theory 18
Figure 2.11: EGO algorithm as presented by Jones et al. [46]
the use of an expected improvement function (EIF). EIF is defined as:
EIF(x) = (fmin − y)D
(fmin − y
s
)+ sP
(fmin − y
s
)(2.45)
where fmin is the lowest value of the objectives that are evaluated directly thus far, y is the surrogate
model’s prediction of the objective value, s is the standard error (uncertainty) of the surrogate model’s
estimate, D is the normal distribution function, and P is the normal density function. EGO achieves
a natural balance between local and global search since small values of (fmin − y) and large values of s
promote local and global searches, respectively. Maximizing the EIF will reveal the next evaluation point
which would maximize information gain about the actual minima and the objective surface, f(x). In
EGO, if the maximized EIF is above 1%, this point is evaluated and the algorithm iterates, fitting a new
surrogate model with the new information. If the EIF < 1%, it signals an accuracy based termination
condition.
EGO has been extended to a multi-objective optimization algorithm by multiple researchers in recent
years; Horn et al. provides a framework to understand the variations that exist [40]. A visual repre-
sentation of the taxonomy of multi-objective EGO algorithms developed in [40] is shown in Figure 2.12.
The initial design, as with single-objective EGO is typically generated using LHS. There are two options
available when fitting a kriging model: 1) each objective is fit with a separate surrogate model, or 2) the
multiple objectives are scalarized (weighted sum of each objective), after which one surrogate model is
fit over the scalarized objective space. Much work has been done in utilizing a variety of infill criteria for
the generation of an evaluation point (candidate generation), such as hypervolume contribution [69] and
additive ε-indicator [87], in addition to the EIF. In general, these methods involve 1) multi-objective op-
timization of a given criterion when individual models have been fit for each objective, 2) single-objective
optimization of a criterion, or 3) multiple single-objective optimizations of a criterion. If required, mul-
tiple candidates can be generated using either 1) or 3). Following the generation of multiple candidates,
a selection strategy can be used to narrow down the number of candidates to a desired value based on
computational budget or other requirements. After the evaluation of the candidate(s), the process can
return to the fitting of the surrogate model and iterate, typically until a computational budget is met.
Chapter 2. Background and Theory 19
Figure 2.12: Multi-objective EGO taxonomy, based on [40]; grayed out boxes and dotted lines representaspects not explored in the present work.
The present work explores 4 multi-objective EGO variants, with the taxonomy of each presented in
Table 2.1:
1. ParEGO [51]: This is a scalarization based method which randomizes the weights used for each
objective at each iteration. Thus, the EIF is calculated and optimized a single surrogate model,
while still having a general idea of all objectives involved without focusing down on any single
objective.
2. SMS-EGO [69]: After developing individual surrogate models for each objective, SMS-EGO con-
verts the MOO into a SOO by calculation of the hypervolume (S-metric). Hypervolume is the
volume (or area in a bi-objective optimization) encompassed by the Pareto front (or Pareto front
approximation), with reference to a prescribed point [6]. In practice, the hypervolume is calculated
through rectangular approximations as shown in Figure 2.13. Candidate points are generated and
chosen based on highest hypervolume contribution potential as this would push out the Pareto
front.
3. ε-EGO [87]: Similar to SMS-EGO, ε-EGO fits individual surrogate models to each objective and
then converts the MOO into a SOO by calculating the additive epsilon indicator Iε+. This is the
minimum distance that a solution set needs to be translated to just weakly dominate another set,
with positive numbers signifying a shift towards the ideal point (Figure 2.14) [97] (a more detailed
explanation of Iε+ is given in Section 3.3.4). By minimizing Iε+, a candidate point which pushes
out the Pareto front can be selected for evaluation.
Chapter 2. Background and Theory 20
Figure 2.13: Hypervolume calculation
Figure 2.14: Additive epsilon indicator (Iε+) calculation
4. MSPOT [94]: This conducts an NSGA-II multi-objective optimization directly on the model pre-
diction using multiple surrogate models fit to each objective, resulting in a Pareto set of viable
points for evaluation in the next iteration. The single point for evaluation is chosen based on it’s
hypervolume contribution.
2.4.3 Dimensionality Reduction
Dimensionality reduction is the task of reducing the the degrees of freedom (number of dimensional
parameters being optimized) an optimization algorithm needs to consider in an effort to mitigate the
curse of dimensionality [70]. In addition to optimization speedup, dimensionality reduction can result in
better solutions as unimportant variables that add noise to the optimizer’s model of the objective space
(i.e., over-fitting) are removed. In general, it can be achieved through: 1) feature extraction, where m
new variables are generated from the original n variables with m < n, or 2) feature selection, where a
subset of m variables are chosen from the original n variables, or 3) a combination of feature extraction
and selection. This work utilizes feature selection, as the physical meaning of the variables are retained
- a fact that proves useful in an engineering context.
Chapter 2. Background and Theory 21
Table 2.1: Taxonomy of multi-objective EGO algorithms used in present work
Algorithm Initial Design Model FittingCandidateGeneration
CandidateSelection
StoppingDecision
ParEGO
LHS
Model ofscalarization
SOO of EIF1 point,generation criterion Budget
SMS-EGO Individualmodels foreach objective
SOO ofhypervolume
ε-EGOSOO ofε indicator
MSPOTMOO of modelpredictions
1 point,hypervolume contribution
There are three types of feature selection methods:
1. Filter: Ranks the variables or subsets of variables by a directly calculated metric, and is very quick
[78].
2. Wrapper: This method is more applicable in the machine learning context. The metric used
to evaluate a given subset is determined by the accuracy of a trained model, generally leading to
effective dimensionality reduction at the cost of the resulting subset’s performance being specifically
tied to the machine learning algorithm used [78].
3. Hybrid: Combination of filter and wrapper methods [78].
Filter methods are well suited for dimensionality reduction in optimization as they are not tied to any
specific algorithm’s ability to learn underlying structures. Correlation-based feature selection (CFS) is
a well-established metric for the filter method that aims to choose variables with a high correlation to
the desired output (transmission loss of the muffler, in this case) but have low correlation with other
variables, reducing redundancy in the input variable subset [39]. A variety of methods and heuristics
exist for the selection of the m-variable subset including best first [72], greedy forward selection [49], and
genetic algorithms. However, as the number of variables are relatively low in the present case (n = 11),
an exhaustive search was conducted, looking through every possible subset combination (211 = 2048
combinations). In summary, dimensionality reduction prior to optimization was conducted using a
filter-based feature selection using CFS to evaluate an exhaustive list of all possible variable subsets.
2.5 Muffler Optimization
Engineering design optimization generally falls under two categories: topology optimization and shape
(parameter) optimization. The former assumes very little about the design and the process is tasked
with generating the details of the design - in the case of a muffler, the inputs would just be the bounding
dimensions of the muffler and the algorithm would be responsible for filling in all the internal compo-
nents such as baffles and tubes. On the other hand, shape optimization involves the tuning of a set
of parameters that control the dimensions or properties of components within a predetermined design.
Topology optimization has the potential of generating radically new designs, but its prevalence is limited
Chapter 2. Background and Theory 22
in the field of muffler optimization. Due to the general nature of this method, its designs are limited to
simple mufflers such as simple expansion chambers [90], expansion chambers with offset inlet and outlet
[54], and dual expansion chambers [55, 26, 53].
The bulk of the existing work on muffler optimization involves shape optimization. Yeh et al. initiated
this work on a simple expansion chamber with an offset inlet and outlet configuration [93], followed by a
multi-chamber expansion chamber [92, 91], all using 1D TMM models. While [93] used algorithms such
as exterior penalty, interior penalty, and method of feasible directions with success, [92] and [91] used
genetic algorithms (GA). Barbieri and Barbieri focused on 2D FEM-based optimization of an expansion
chamber with extended inlet and outlet using the method of feasible directions [7]; this algorithm doesn’t
see much use in later literature. Chiu and Chang used GA in the optimization of a multi-chambered
muffler with perforated tubes, reverting back to 1D TMM [22]. They expanded the work by consider-
ing pressure drop as a constraint (calculated via empirical relationships) while employing the simulated
annealing (SA) algorithm in the optimization of 1D TMM models of perforated mufflers [23]. SA based
optimization of a plug flow muffler using 1D TMM without considering pressure drop is conducted in
[17]. Much of the work following this included perforated elements in the mufflers, modelled with 1D
techniques, relying on GA [19, 81] or SA [18, 20]. Huang et al.’s work was the first to consider pres-
sure drop in a truly multi-objective fashion, using optimization by design of experiments (DOE) [41].
Siano et al. continued using pressure drop in a multi-objective setting, utilizing genetic algorithms for
optimization [81]. Additionally, single-objective muffler optimization has incorporated pressure drop via
scalarization (weighted sum of acoustic and pressure drop objectives) [80, 79]. Barbieri et al. introduced
the use of particle swarm optimization (PSO) using a 2D FEM model of a simple expansion chamber
with extended inlet and outlet [8].
Analyzing recent literature at large (92 papers with ‘muffler optimization’ since 2003) shows a few
striking observations:
1. Majority of the work (70.7%) used SA or GA. These methods use anywhere from 2000 function
evaluations [23] to 64000 function evaluations [22] for SA and GA, respectively.
2. Majority of the work (81.5%) used 1D TMM to model the muffler [e.g., 96, 71, 48, 21]. This is
tied to observation #1, as SA and GA require a large number of function evaluations in order to
be successful, necessitating low-fidelity 1D models for the sake of computational time.
3. Multi-objective work considering TL and pressure drop as two distinct objectives is scarce (3.3%).
This is a crucial consideration as high back-pressure on engines lead to loss of performance and
efficiency.
Sections 3.2 & 4.2 address the first observation by using a much simpler algorithm (L-BFGS-B) in order
to optimize TL, removing the complexity associated with implementing algorithms such as GA or SA.
The latter two observations are addressed in Sections 3.3 & 4.3 by using a truly multi-objective method
capable of optimization using a relatively small number of function evaluations (EGO).
Chapter 2. Background and Theory 23
2.6 Chapter Summary
This chapter provided the theoretical foundations for various aspects of the design of a muffler. Section
2.1 detailed the prediction of turbulence-induced noise in a muffler. Section 2.2 covered the analysis of the
plug flow muffler’s acoustic performance and Section 2.3 presented a means of characterizing the pressure
drop across the muffler. Finally, Sections 2.4 and 2.5 addressed methods related to the automated design
of an optimal muffler and surveyed the current muffler optimization literature, respectively.
Chapter 3
Methodology
3.1 Aeroacoustic Predictions
The present work was an introductory investigation into using the acoustic analogy (Section 2.1.1) to
predict the SPL caused by turbulence-induced noise at specific receiver locations. NASA’s tandem
cylinder benchmark for aeroacoustic noise [60, 59] was replicated using incompressible computational
fluid dynamics (CFD) using three established methods for time-resolved simulation: unsteady Reynolds-
averaged Navier-Stokes (URANS), large eddy simulation (LES), and detached eddy simulation (DES).
The tandem cylinder experiment was conducted at NASA’s Quiet Flow Facility (QFF) - an open jet
wind tunnel in an anechoic room. Two cylinders (C1 and C2) of equal diameter (D = 0.057 15 m) with
a 16 D span were placed in oncoming flow with a uniform flow velocity (Uo = 43.4 m/s;ReD = 166, 000),
one downstream of the other, separated by a center-to-center distance L = 3.7 D (Figure 3.1). The aeroa-
coustic noise created through turbulent vortex shedding was measured by an array of microphones in the
far-field. Three microphones, A, B, and C, were considered in this work, located at (−8.33 D, 27.815 D),
(9.11 D, 32.49 D), and (26.55 D, 27.815 D), respectively, from the centroid of C1.
In the present work, the tandem cylinder experiment was simulated using ANSYS Fluent 17.1 on an
8-core computer with three methods, URANS, LES, and DES, in order to determine the accuracy of each
method, while considering the computational cost of each. The setup parameters are shown in Table
3.1. The LES was a low-fidelity simulation due to the coarse span-wise resolution. The three meshes
used are pictured in Figure 3.2. The DES mesh has RANS-based meshing (high aspect ratio) near the
walls and transitions into an LES-based mesh in the ”focus region” as defined by Spalart [83].
3.1.1 Simulation Verification
The accuracy of the simulations was verified through the comparison of various hydrodynamic values
and statistics. Table 3.2 shows the list of flow parameters that were compared against experimental
values for URANS, LES, and DES.
Section 3.2 was previously published in J. Puthuparampil, H. Pong, and P. Sullivan, Modelling and Optimization of PlugFlow Mufflers in Emission Control Systems, in SAE 2017 Noise and Vibration Conference and Exhibition, 2017. [71].
24
Chapter 3. Methodology 25
Figure 3.1: Tandem cylinder setup, with microphone locations (not to scale)
Table 3.1: CFD setup overview
Mesh Setup URANS LES DES
Mesh Type Planar, extruded Planar, extruded Planar, extrudedFirst Layer Thickness 100 y+ 57 y+ 1 y+Num. of Elements (Plane) 15,292 81,609 88,842Span 16 D (33 elements) 16 D (120 elements) 3 D (120 elements) †
Span Resolution [mm] 27.7 7.62 1.42875Total Num. of Elements 504,636 9,793,080 10,661,040
Modelling Setup
Turbulence Model k-ε Dynamic Smagorinsky-Lilly SST k-ωP-V Coupling SIMPLE Fractional Step (NITA) Fractional step (NITA)Momentum Discretization 2nd order upwind Bounded central difference Bounded central differencek Discretization 2nd order upwind – 2nd order upwindε Discretization 2nd order upwind – –ω Discretization – – 2nd order upwindPressure Discretization 2nd order PRESTO 2nd orderTransient Discretization 2nd order implicit 2nd order implicit Bounded 2nd order implicitTime Step [s] 2.5× 10−5 1.5× 10−5 1.5× 10−5
Time Sampled [s] 0.30 0.15 0.2658Num. of Flow-Through 9.47 4.75 8.06
Boundary Conditions
Inlet Velocity [m/s] (43.4, 0, 0) (43.4, 0, 0) (43.4, 0, 0)Pressure Outlet [gauge Pa] 0 0 0Cylinder Surfaces No-slip No-slip No-slipTop and Bottom Free-slip Free-slip Free-slipSides (Span-wise) No-slip No-slip Periodic
† Full-span noise contribution estimated using long-span correction [76]
Chapter 3. Methodology 26
Figure 3.2: Tandem cylinder meshes - URANS, LES, and DES, respectively
Table 3.2: Hydrodynamic metrics used to compare simulation to experimental data
Parameter Description URANS LES DES
Cp Coefficient of pressure (Cp) on the surface of cylinders Yes Yes YesCp′ RMS Root-mean-square of Cp on the surface of cylinders No Yes Yesu Mean stream-wise velocity along y = 0 Yes Yes Yes2D TKE Contour plots of 2D turbulent kinetic energy (TKE) No Yes YesPSD Power spectral density of pressure at 135° on C1 and 45° on C2 No No YesRp Span-wise correlation of pressure at 135° on both cylinders † No No Yes
† Span-wise correlation results in Appendix B.2
Chapter 3. Methodology 27
Figure 3.3: Cross-section of a plug flow muffler and associated dimensional parameters
3.1.2 Receiver Noise Calculation - Acoustic Analogy
Fluent implements the full FWH analogy, but since the cylinders were stationary in flow, the FWH
analogy reduced to the Curle analogy. Furthermore, dipole sources were dominant as periodic flow sepa-
ration through vortex shedding was the primary flow characteristic (Section 2.1.1). Thus, the quadrupole
contributions were not considered, simplifying the computation.
The cylinder surfaces were used as noise sources to estimate the time-resolved sound pressures at mics A,
B, and C. The sound pressure at each mic was transformed into the frequency domain with a fast Fourier
transform (FFT) using a Hamming window with a 0.50 overlap (10 Hz resolution for URANS, 15 Hz for
LES, and 10 Hz for DES). The pressure levels were then converted to decibels using the standard pressure
reference value of pref = 2× 10−5 Pa, arriving at SPL (dB) at a range of frequencies.
3.2 Single-Objective Optimization (SOO)
In general, the plug flow muffler can be fully defined by 14 design parameters, shown in Figure 3.3
and Table 3.3. Prior to optimization, the TMM and FE models were evaluated against a baseline plug
flow muffler geometry presented by Wu et al. [89]; its dimensions are seen in Table 3.3. TMM and FE
were conducted with all three zero mean flow impedance models (Section 2.2.4), and the accuracy of
each was evaluated against the experimental data [89]. The best performing TMM impedance model
was used in the TMM-based SOO, conducted at room temperature (To = 22 ◦C, ρo = 1.1965 kg/m3,
co = 344.82 m/s, νo = 1.5295× 10−5 m2/s). The final optimized design’s TL prediction was then verified
using the FE model. The general methodology for SOO is seen in Figure 3.4.
Chapter 3. Methodology 28
Table 3.3: Parameters defining the PFM and associated values for the Wu et al. muffler [89]
Parameter Description Wu et al. Muffler [89]
D1 Inner diameter 0.0508 mD2 Outer diameter 0.1076 mLa1 Length of first solid tube section (expansion section) 0.0317 mL1 Length of perforated section (expansion section) 0.0952 mLb1 Length of second solid tube section (expansion section) 0.0255 mLa2 Length of first solid tube section (contraction section) 0.0063 mL2 Length of perforated section (contraction section) 0.0952 mLb2 Length of second solid tube section (contraction section) 0.0953 mσ1 Area porosity of perforated expansion tube 21.68 %σ2 Area porosity of perforated contraction tube 13.54 %Tp1 Thickness of perforated expansion tube 1.1938 mmTp2 Thickness of perforated contraction tube 1.1938 mmDh1 Perforate hole diameter of expansion tube 6.35 mmDh2 Perforate hole diameter of contraction tube 6.35 mm
Figure 3.4: Single-objective optimization (SOO) methodology: evaluation and optimization
Chapter 3. Methodology 29
3.2.1 TMM Setup
The TMM for a plug flow muffler (Section 2.2.1) was implemented in Python 2.7 utilizing the numpy
module for matrix operations. The Sullivan, Bento Coelho, and Bauer zero mean flow impedance models
(Section 2.2.4) were implemented on the perforated surfaces. TMM is only accurate below the cutoff
frequency (Equation 2.21); D2 was held as a constant in the optimization process at the same value as
the Wu et al. muffler, leading to a cutoff frequency of 1877 Hz for all mufflers.
3.2.2 FE Setup
The FE simulation was conducted through COMSOL Multiphysics 5.2a using the Acoustics Module with
a parametric frequency sweep study. The mufflers were implemented with both two-dimensional axisym-
metric (2DA) and three-dimensional (3D) geometries. First, the muffler was parametrically modelled.
Plane wave boundary conditions were applied at the inlet and outlet, along with hard wall boundary
conditions at all relevant surfaces. The perforated boundaries were defined via custom functions for the
Sullivan, Bento Coelho, and Bauer impedance models (Section 2.2.4). An incident pressure wave of 1
Pa amplitude was applied normal to the inlet as the initial condition. A free triangular mesh (for 2DA)
and a tetrahedral mesh (for 3D) were used, ensuring the maximum element size was limited by
Max. Size =co
10 · fmax(3.1)
where fmax is the maximum frequency of interest in the simulation. Equation 3.1 ensured that the
model always had enough spatial resolution to resolve the smallest wavelength under consideration [3].
TL solutions were obtained for a parametric frequency sweep from 10 Hz to 4000 Hz (fmax = 4000 Hz),
with a resolution of 10 Hz.
In the SOO process, FE simulations are only used to check the accuracy of the TL predictions pro-
duced by TMM; FE is not used in the actual optimization.
3.2.3 SOO Setup
Objective Function
The objective function used, noise level objective (NLO), was defined as
NLO =∑i
|ESLi − TLi|wi (3.2)
for i = 63, 125, 250, 500, 1000, 2000, 4000 Hz
where i denotes the set of octave band center frequencies of interest, ESLi is the excess sound level
at a given octave band, TLi is the averaged muffler transmission loss over the octave band under
consideration, and wi is a weight applied to each octave band that determines the band’s contribution
to the NLO. ESLi is
ESLi = ENi −NCi (3.3)
where ENi is the measured engine noise (sound pressure level at 1 m) in decibels and NCi is the noise
criterion standard required for the application [10]. The difference between the two quantities provides
Chapter 3. Methodology 30
Table 3.4: Octave band limits and associated engine noise, noise criterion, and NLO weights for SOO
Octave Band [Hz] Lower Limit [Hz] Upper Limit [Hz] EN [dB] NC-60 [dB] wi
63 40 90 98.3 77 0.2125 90 180 107.1 71 0.2250 180 360 102.5 67 0.2500 360 710 102.5 63 0.18751000 710 1420 104.0 61 0.15252000 1420 2840 100.8 59 0.0554000 2840 4000 96.5 58 0.005
a suitable TL target for the design. This work used the NC-60 limits as the desired noise criterion; the
maximum permissible output noise level for each octave band is provided in Table 3.4, along with the
measured sound pressure level (ENi) of a CAT C27 750 kW generator at 1 meter. It is important to
note that Equation 3.2 uses transmission loss where insertion loss would be a more technically correct
value. However, insertion loss requires knowledge of the engine (source) impedance, muffler outlet (ter-
mination) impedance, and the impedance of any ducting used between the engine and the muffler. Since
transmission loss and insertion loss are correlated, the use of transmission loss in Equation 3.2 has no
bearing on the optimization methodology described.
The muffler’s average transmission loss over an octave band (TLi) is
TLi = 10 log10
Upper Limit of i∑
j=Lower Limit of i
10TLj/10
ni
(3.4)
where ni is the number of elements summed for a given octave band. The upper and lower frequency
limits for a given octave band is listed in Table 3.4. The frequency limits are rounded to the nearest
10 Hz, as the solution is calculated with a 10 Hz resolution over the frequency range from 10 Hz to
4000 Hz. Note that Equation 3.4 simply calculates an average of the decibel values of TL.
The NLO weights (wi) in Table 3.4 were chosen (a) to prioritize the frequency ranges where a reactive
muffler is traditionally used as an effective method of noise attenuation and (b) to account for the
inherent inaccuracies that occur above the cutoff frequency (fc = 1877 Hz). The higher order modes
that occur above the cutoff frequency are not included in TMM, but the TL contribution of the plane
wave mode that occurs in the high frequency regime is included; thus, the overall TL predicted in the
high-frequency range is useful even though not entirely accurate, leading to non-zero weights in these
octave bands.
Implementation
The L-BFGS-B algorithm (Section 2.4.1) from the scipy.optimize module [47] was used for the SOO, aim-
ing to minimize Equation 3.2. The algorithm was implemented in a randomized multi-start framework
[28] with parallel computation accomplished using the pp module. The dimensional variables defining the
PFM were optimized subject to the bounds and constraints seen in Table 3.5. The limits were informed
Chapter 3. Methodology 31
Table 3.5: Bounds and constraints applied on SOO
Parameter Lower Bound Upper Bound
rD = D1/D2 0.4 0.8La1 [m] 0.0 0.15L1 [m] 0.05 0.25Lb1 [m] 0.0 0.15σ1 0.14 (14 %) 0.25 (25 %)
Constraint Equality
La2 La1
L2 L1
Lb2 Lb1σ2 σ1
by physical and manufacturing constraints. The perforated expansion and perforated contraction were
defined as symmetrical using equality constraints, to simplify the manufacturing process. The variables
D2, Tp1, Tp2, Dh1, and Dh1 were held constant as they were seen to have a comparatively low effect
on the TL in a sensitivity study. Therefore, these variables were set to their respective values from the
baseline Wu et al. muffler [89].
The optimization was performed 2000 times in parallel on a 4-core Intel i7-4790 processor. Each ini-
tialization of the optimization algorithm began at random values between the prescribed limits in Table
3.5, striving towards a minima, terminating when
NLOk−1 −NLOk
max(NLOk,NLOk−1, 1.0)< 1× 10−4 (3.5)
where NLO (Equation 3.2) is evaluated at the current iteration, k, and the previous iteration, k − 1.
Research Objective
The SOO work primarily aimed to answer the following objective: can L-BFGS-B, a relatively sim-
ple algorithm, be used in place of complex stochastic algorithms such genetic algorithm or simulated
annealing to successfully conduct muffler optimization?
3.3 Multi-Objective Optimization (MOO)
In the MOO process, an FE simulation was used for the TL prediction, removing any limitation as-
sociated with the TMM cutoff frequency. Firstly, the 14 variables defining the plug flow muffler were
recast to those presented in Table 3.6 (Figure 3.3 for visual reference). Lt and D2 were held con-
stant at 17 inches and 12 inches respectively, as these bounding dimensions are generally prescribed by
space limitations. Since larger mufflers are generally more effective, a good muffler design will always
take advantage of all the space available. The importance of the remaining 12 variables in predicting
the acoustic performance of a PFM was discovered using feature selection; the most important subset
of variables were used in the MOO process, with the other variables held constant at their nominal values.
Chapter 3. Methodology 32
Table 3.6: Geometric parameters defining the PFM, recast for multi-objective optimization as ratios
Paremeter Description
Lt Total lengthD2 Outer diameterrD = D1/D2 Diameter ratiorexp = Lexp/Lt Expansion chamber length ratiorL1 = L1/Lexp Expansion perforated tube length ratiorL2 = L2/(Lt − Lexp) = L2/Lcon Contraction perforated tube length ratior1 = La1/(Lexp − L1) Expansion solid tube length ratior2 = La2/Lcon − L2 Contraction solid tube length ratioσ1 Area porosity of perforated expansion tubeσ2 Area porosity of perforated contraction tubeTp1 Thickness of perforated expansion tubeTp2 Thickness of perforated contraction tubeDh1 Perforate hole diameter of expansion tubeDh2 Perforate hole diameter of contraction tube
Figure 3.5: Multi-objective optimization (MOO) methodology: evaluation and optimization
The pressure drop of the muffler was estimated to provide a competing objective to TL in MOO pro-
cess. The empirical model was validated through CFD before its use. The optimization was conducted
at typical operating temperatures for a muffler (To = 400 ◦C, ρo = 0.5245 kg/m3, co = 520.07 m/s,
νo = 6.1961× 10−5 m2/s) and a typical operating flow rate (Q = 800 ACFM). The MOO algorithm
aimed to simultaneously optimize two aforementioned objectives, resulting in a Pareto front showing the
trade-off between TLO and DP for the PFM. The MOO methodology is depicted in Figure 3.5.
3.3.1 FE Setup
The FE simulation was set up in a similar manner as in Section 3.2. A few key differences are:
• the simulation’s frequency range was from 80 Hz to 4470 Hz with a resolution of 10 Hz
• only a 2DA model was implemented, as any inaccuracies caused by neglecting 3D modes are
minimal in the frequency range considered
Chapter 3. Methodology 33
Figure 3.6: Dimensionality reduction through feature selection methodology
• only the Elnady perforate impedance model (Section 2.2.4, Equation 2.29) was used, as it is able
to account for the effects of flow through the muffler
3.3.2 Pressure Drop Model
The empirical pressure drop model (Section 2.3) for the PFM was implemented in in Python 2.7. The
output of the calculation was converted from Pa to inH2O as it is a more industrially relevant unit of
measure. A model, rather than a CFD simulation, was used to calculate pressure drop due to the large
savings in computational time. However, the model was first checked against a 2DA steady state RANS
CFD conducted in Fluent for 6 random PFM geometries. The perforated plate was implemented in
Fluent as a porous surface using tabulated discharge coefficients [82]. Once validated, the model was
used in the MOO.
3.3.3 Dimensionality Reduction - Feature Selection
To determine the importance of each of the 12 variables from Table 3.6 (excluding Lt and D2 as they are
held constant), dimensionality reduction through feature selection was performed. 2400 (= 200·nvariables)
randomly generated mufflers (i.e., 2400 randomized sets of the 12 geometric parameters) were evaluated
via a 2DA FE simulation. The TL curve for each muffler was transformed into a single-number rep-
resentation, TLO (Equation 3.6). 10 different TLO were calculated for each muffler using 10 different
randomly generated sets of weighting factors (wi) from Equation 3.6 to generalize the findings of the
feature selection process.
The feature selection process (Section 2.4.3) was conducted using an exhaustive search of all subsets
of the 12 variables ranked through the correlation-based feature selection (CFS) metric, implemented in
Weka 3.8 [38]. For a given wi, 10-fold cross-validated feature selection was performed, which resulted
in a set of percentages representing how often each variable was chosen in the 10-fold cross-validation;
these values are associated with the importance of each variable. This was repeated for all 10 wi and
the results were averaged, to arrive at a generalized and statistically robust subset of most important
variables to be used in optimization. The dimensionality reduction methodology is pictured in Figure
3.6.
Chapter 3. Methodology 34
Table 3.7: 1/3 octave band limits and TLO weights for MOO
1/3 Octave BandTLO Weight (wi)
Band Number Lower Limit [Hz] Center Frequency [Hz] Upper Limit [Hz]
9 89 100 112 0.00759387310 112 125 141 0.00759387311 141 160 178 0.01012516412 178 200 224 0.01518774613 224 250 282 0.02278161914 282 315 355 0.03290678415 355 400 447 0.04556323916 447 500 562 0.06075098517 562 630 708 0.07847002318 708 800 891 0.09872035119 891 1000 1122 0.13668971720 1122 1250 1413 0.15187746321 1413 1600 1778 0.14276481522 1778 2000 2239 0.10033809823 2239 2500 2818 0.05807544424 2818 3150 3548 0.02296693325 3548 4000 4467 0.007593873
3.3.4 MOO Setup
Objective Functions
The acoustic objective function used, transmission loss objective (TLO), was defined as
TLO = −25∑i=9
TLi · wi (3.6)
where i denotes the 1/3 octave band number from 100 Hz (band number 9) to 4000 Hz (band number
25), TLi is the averaged transmission loss over the 1/3 octave band, and wi is the weight applied to each
1/3 octave band. The negation allows the minimization of Equation 3.6 to be associated with higher
acoustic performance. TLi is calculated using Equation 3.4, but with 1/3 octave band resolution using
the upper and lower frequency limits shown in Table 3.7 (TL calculated by FE at a resolution of 10 Hz
is interpolated to a 1 Hz resolution to match the 1/3 octave band limits closely). The weights used are
also listed in Table 3.7. The weights were chosen to maximize the PFM’s performance between 891 Hz
and 1778 Hz (bands 19 through 21); however, any set of arbitrary weights summing to 1.0 is sufficient.
The pressure drop objective function (DP) was calculated in units of inH2O. As the pressure drop
model is empirically derived, it should not be considered valid outside the range of experimental pa-
rameters. Thus, linear constraints were introduced to limit the OAR as per Equations 2.36 and 2.37:
0.3 < OARexp < 2.2 and 0.31 < OARcon < 1.66.
The complete set of variable bounds and constraints used in MOO is seen in Table 3.8. Based on
the results of feature selection, the less important variables were be held constant at their nominal
values during the MOO.
Chapter 3. Methodology 35
Table 3.8: Bounds and constraints applied on MOO
Parameter Lower Bound Nominal Value Upper Bound
Lt [in] – 17.0 –D2 [in] – 12.0 –rD 0.1 0.5 0.9rexp 0.1 0.5 0.9rL1 0.1 0.5 0.9rL2 0.1 0.5 0.9r1 0.1 0.6 † 0.9r2 0.1 0.4 † 0.9σ1‡ 0.1 (10 %) 0.5 0.9 (40 %)
σ2‡ 0.1 (10 %) 0.5 0.9 (40 %)
Tp1 [mm] 0.5 1.1938 2.0Tp2 [mm] 0.5 1.1938 2.0Dh1 [mm] 2.0 5.0 8.0Dh2 [mm] 2.0 5.0 8.0
Constraint Lower Bound Nominal Value Upper Bound
OARexp 0.3 – 2.2OARcon 0.31 – 1.66
† Asymmetric r1 and r2 shown to boost PFM performancein a sensitivity study; peak values chosen as nominal‡ 10 % to 40 % porosity linearly mapped from 0.1 to 0.9
Implementation
The EGO algorithm (Section 2.4.2) from the mlrMBO package for the statistical programming language
R was used in MOO [12]. All 4 EGO variants (SMS-EGO, ε-EGO, MSPOT, and ParEGO) are available
in mlrMBO. In addition, the NSGA-II algorithm (Section 2.4.2) from the mco package was used for GA-
based MOO (MSPOT and NSGA-II). A constant set of initial points (Ninit = 40) was used for all EGO
runs to remove any difference in performance associated with initial conditions. A maximum evaluation
budget Neval was used as the stopping criterion for all MOO; the baseline value for Neval = 240, which
is in addition to Ninit. Each method was repeated n = 8 times to produce statistics for the results.
Research Objectives & Metrics Used
The MOO work aimed to address the following objectives:
1. Which EGO variant performs best for multi-objective muffler optimization?
• Comparison of ParEGO, SMS-EGO, ε-EGO, and MSPOT
2. How sensitive is the EGO algorithm to Nevals?
• Comparison of Nevals = 240, 300, 360
3. How sensitive is the EGO algorithm to different initial conditions?
• Comparison of four randomly generated LHS of size (Ninit = 40) labeled A, B, C, and D
4. How does EGO compare to the NSGA-II algorithm (standard algorithm for MOO)?
Chapter 3. Methodology 36
The Pareto front approximation resulting from each method is visualized using an attainment surface
[50]. An attainment surface is a single summary curve of the n = 8 repetitions of a given optimization
method. This work utilizes the median attainment surface for all visualizations - the 4th attainment sur-
face (p = 4) when n = 8. The median attainment surface is obtained by (1) querying the n Pareto front
approximations at regular intervals, (2) choosing the Pareto front approximation that weakly dominates
at least n− p+ 1 = 5 approximations at each queried position.
Pareto front approximations are numerically compared using two different metrics - hypervolume and
the additive binary epsilon indicator. Together, they provide a holistic look at the quality of a Pareto
front [51]:
1. Hypervolume (IHV (A)): The hypervolume is the volume (or area in bi-objective optimization)
encompassed by the Pareto front approximation, with reference to a prescribed point (Figure
2.13). A larger value indicates a front of higher quality. However, it should be noted that different
fronts can have the same hypervolume; thus, other metrics are also necessary for a more complete
understanding of the quality of a Pareto front approximation. The reference point chosen was
(TLO = 0,DP = 10).
2. Additive binary epsilon indicator (Iε+(A,B)): This metric, being a binary indicator, compares two
Pareto front approximations (A,B) and produces a value IA. IA is the minimum distance that
all the points in A must be translated such that A weakly dominates B (i.e., A is equal or better
than B across the entire front). A translation towards the ideal point is positive. Translations are
conducted independently in each objective’s direction, and the maximum value is IA (Figure 2.14).
The reverse, (Iε+(B,A)), produces IB . The lower numerical value between IA and IB indicates the
better front, as it needs to be shifted less to weakly dominate the other set. If (IA ≤ 0, IB > 0),
then A fully dominates B (i.e., A is strictly better).
To answer the objectives in a statistical fashion, IHV and Iε+ from multiple runs of one method was
compared against another method using the Mann-Whitney U (MWU) test. Details are seen in Appendix
A.
3.4 Chapter Summary
This chapter provided the methodology pertaining to the acoustic analysis and optimization of a plug
flow muffler. Section 3.1 presented a means of predicting turbulence-induced noise using the aeroacoustic
analogy. Section 3.2 detailed the single-objective optimization process where a relatively simple gradient-
based algorithm was used in a multi-start framework to optimize the acoustic performance of a plug
flow muffler. Lastly, Section 3.3 covers an efficient multi-objective optimization of acoustics and pressure
drop across a muffler, utilizing an FE simulation to calculate the acoustic objective function.
Chapter 4
Results and Discussion
4.1 Aeroacoustic Predictions
The tandem cylinder experiment (Section 3.1) was recreated using CFD as the first steps in exploring
turbulence-induced noise produced within mufflers. This section first evaluates the ability of different
low computational-cost CFD methods to accurately depict the flow observed in the experiment, followed
by their ability to recreate the measured noise. The acoustic predictions are then compared against
those from literature.
4.1.1 Simulation Verification - Hydrodynamic Results
The mean stream-wise velocity and flow streamlines around the tandem cylinders are seen in Figure
4.1. Like a single cylinder in cross-flow, vortex shedding occurs off C1 (front cylinder). The turbulent
structures associated with vortex shedding are then convected downstream and impinge on C2 (rear
cylinder). Two pairs of counter-rotating recirculation cells are observed, one behind each cylinder. The
recirculation cells behind C1 are larger and promote a significant amount of flow in the y-direction as
shown by the streamlines, leading to a high level of turbulent energy between the two cylinders.
The accuracy of each simulation was verified using a set of hydrodynamic parameters (Table 3.2). The
results for URANS are span-averaged, while the results for LES and DES are both span and time-
averaged, when applicable. All results were captured after initial transience in the flow exited the
domain and a periodic behaviour was observed in the coefficients of lift and drag of the cylinders.
Coefficient of Pressure
Figure 4.2 shows the mean coefficient of pressure (Cp) on the surface of each cylinder. The QFF data
shows excellent symmetry along the y = 0 plane. On C1, a stagnation point is seen at 0°, and Cp reduces
as the flow speeds up around the curvature of the cylinder. Figure 4.2 shows a region of flow separation
on C1 between 110° and 250°. The separated flow impinges on C2, reattaching at 0°. The Cp profile
on C2 is similar to C1, but with lower magnitudes of pressure. The separation region on C2 extends
between 130° and 230°. The smaller angular span of the separation region is associated with the smaller
pair of counter-rotating recirculation cells seen behind C2, as compared to C1 (Figure 4.1).
37
Chapter 4. Results and Discussion 38
−1
−0.5
0
0.5
1
0 1 2 3 4 5 6
y/D
x/D
-0.2
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
u/Uo
Figure 4.1: Mean x-velocity with streamlines from the DES simulation
In general, the experimental trends are captured by all methods. There is excellent agreement ob-
served between the LES and DES predictions for both cylinders. The LES and DES Cp match the QFF
experiment for both C1 and C2, except for the separated zone behind C2 - LES and DES over-predict
the amount of suction. URANS over-predicts the peak suction location and magnitude, and under-
predicts the Cp in the separation region on C1. However, it is able to quite accurately determine the Cp
on C2. The asymmetry seen in the URANS results for C2 is associated with a lack of time-averaged data.
Figure 4.3 shows the root mean square of the pressure fluctuations (Cp′) on the surface of each cylinder.
As with Cp, Cp′ RMS is also symmetric about y = 0. On C1, two small peaks are seen in the QFF data
at 90° and 270°, and an increased level is seen in between these angles. These fluctuations are associated
with the unsteadiness in the separation points, as well as interactions between C1’s surface and the
vortex shedding that occurs [44]. The front surface of C2 sees the highest levels of Cp′ RMS due to the
impingement of the separated turbulent shear layer from C1. The fluctuation magnitude reduces on the
downstream side of C2, and reaches approximately the same levels as C1 at 180°, within the separated
region.
DES successfully predicts the Cp′ RMS on C1, except for the over-predictions at the 90° and 270°peaks. LES shows significant over-prediction for all of C1. The C2 profile is much more complex, but
LES and DES are able to predict the overall levels well and capture most of the variation on the surface
of C2 with reasonable accuracy, well within expected engineering needs.
Mean Velocity
Figure 4.4 shows the mean velocity in the x-direction (u) normalized by the Uo, along the y = 0 plane.
The negative velocity in the QFF data from 0.5 < x/D < 1.6 occurs due to two counter-rotating recircu-
lation cells visible in Figure 4.1. The stream-wise velocity then recovers to approximately 0.4Uo before
impinging onto the second cylinder at x/D = 3.2. The smaller pair of counter-rotative recirculation cells
lead to the the negative stream-wise between 4.2 < x/D < 4.6. The flow then regains momentum as it
exits the domain.
Chapter 4. Results and Discussion 39
−2.5
−2
−1.5
−1
−0.5
0
0.5
1
1.5
0 45 90 135 180 225 270 315 360
Cp
Cylinder 1, θ [°]
−2.5
−2
−1.5
−1
−0.5
0
0.5
1
1.5
0 45 90 135 180 225 270 315 360
Cp
Cylinder 2, θ [°]
QFFURANS
LESDES
Figure 4.2: Cp along the surface of the cylinders
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0 45 90 135 180 225 270 315 360
Cp′
RM
S
Cylinder 1, θ [°]
QFFLESDES
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0 45 90 135 180 225 270 315 360
Cp′
RM
S
Cylinder 2, θ [°]
Figure 4.3: Cp′ RMS along the surface of the cylinders
−0.4
−0.2
0
0.2
0.4
0.6
0.5 1 1.5 2 2.5 3
u/U
o
x/D
QFFURANS
LESDES
−0.4
−0.2
0
0.2
0.4
0.6
4.5 5 5.5 6 6.5
u/U
o
x/D
Figure 4.4: u along y = 0 in gap between the two cylinders, and after second cylinder
Chapter 4. Results and Discussion 40
The general trend of the velocity in the gap region is captured by all three methods. However, URANS
significantly under-predicts the amplitude of the velocity. LES shows inaccuracies in predicting the
recirculation region in the gap, but is successful in the recovery zone prior to the second cylinder. In
contrast, the DES simulation predicts the entire gap region’s stream-wise velocity with high accuracy.
LES is the only method that predicts the velocity after the second cylinder with good accuracy; however,
it over-predicts the velocity recovery after x/D = 4.8. Both URANS and DES are able to predict the
general trends, but not the absolute values.
2D Turbulent Kinetic Energy (TKE)
The 2D turbulent kinetic energy (TKE) is calculated as
2D TKE =1
2
(u′u′ + v′v′
)/U2
o (4.1)
where u′ and v′ are the fluctuating components of the x and y velocities respectively (U = u + u′ and
V = v + v′ ), and Uo is the inlet velocity. TKE represents the energy of turbulent fluctuations per
unit mass. Figure 4.5a is the experimentally measured 2D TKE. It shows a maxima between the two
cylinders, centered approximately where the two counter-rotating cells meet (x/D ≈ 1.7 as seen from
Figure 4.1). Smaller peaks of TKE are seen just upstream and downstream of C2; the upstream TKE is
associated with turbulent structures from C1 passing around C2 [45], and the downstream peak is most
probably a combination of the convected structures from C1 and any additional structures created at C2.
Figures 4.5b and 4.5c show the 2D TKE predictions from the LES and DES simulations respectively.
Both simulations are able to capture the general trends. However, LES struggles with the TKE around
C2; it under-predicts the upstream section of C2 and over-predicts the wake TKE. In contrast, DES
over-predicts the TKE maxima between C1 and C2 while having much higher accuracy upstream and
downstream of C2. These trends are further examined by the line extractions presented in Appendix
B.1.
Power Spectral Density (PSD) of Pressure
The power spectral density (PSD) of surface pressure at 135° on C1 and 45° on C2 are shown in Figure
4.6. A sharp peak is present at 178 Hz on C1 due to the vortex shedding. This peak is also seen on
C2, as the point of measurement at 45° lies in the path of the impinging vortex. Additionally, a smaller
peak is seen at the second harmonic (356 Hz) on on C1 and very faintly on C2. Furthermore, it should
be noted that the broadband “hump” seen between 500 Hz < f < 2000 Hz on C1 is attributed to the
trip wire used in the experiment to induce turbulent behavior [59]. Additionally, the energy on C2 is
significantly higher than C1 due to the separated turbulent shear layer from C1 impinging on C2 [44].
The DES simulation is able to accurately predict the PSD on both cylinders with a high accuracy
(Figure 4.6). It captures all features of the experimental data, except for the “hump”, as no trip wire
was used in the simulation. The only other deviation is that DES predicts a quicker drop-off in energy
toward the higher frequency region (f > 2000 Hz).
Chapter 4. Results and Discussion 41
−1
−0.5
0
0.5
1
0 1 2 3 4 5 6
y/D
x/D
0.00
0.05
0.10
0.15
0.20
2D
TK
E/U
2 o
(a) QFF experiment
−1
−0.5
0
0.5
1
0 1 2 3 4 5 6
y/D
x/D
0.00
0.05
0.10
0.15
0.20
2DT
KE/U
2 o
(b) LES simulation
−1
−0.5
0
0.5
1
0 1 2 3 4 5 6
y/D
x/D
0.00
0.05
0.10
0.15
0.20
2DT
KE/U
2 o
(c) DES simulation
Figure 4.5: Contours of normalized 2D TKE
Chapter 4. Results and Discussion 42
60
70
80
90
100
110
120
130
50 200 500 2000 5000100 1000
PS
D[d
B/H
z]
Frequency [Hz]
QFFDES
80
90
100
110
120
130
140
150
50 200 500 2000 5000100 1000
PS
D[d
B/H
z]
Frequency [Hz]
Figure 4.6: PSD of surface pressure at 135° on C1 and 45° on C2
Summary of Hydrodynamic Results
URANS shows the ability to capture general trends well, but lacks accuracy on the magnitude of the
hydrodynamic quantities. The LES conducted here, being of relatively low-fidelity, has difficulty sim-
ulating unsteady metrics such as Cp′ RMS and TKE. However, it still shows higher accuracy than the
URANS simulations. DES most closely predicts the QFF experiment and shows a high level of accuracy
in the pressure PSD. The PSD correlates to the unsteady forces applied onto the cylinders, which is used
for the acoustic source strength calculations in the aeroacoustic analogy.
4.1.2 Receiver Noise
The noise produced by the tandem cylinders was a combination of broadband noise and tonal peaks, as
seen in Figure 4.7. The oscillatory vortex shedding phenomenon at 178 Hz, seen in Figure 4.6 produced
a strong tonal noise at 178 Hz, and less prominent peaks at its harmonics (mainly 356 and 536 Hz). This
noise was observed at all microphones in the QFF experiment (Figure 4.7), with the relative strength of
the harmonics varying between microphones while the fundamental tone remained loudest. Variations in
the SPL of the harmonics can be attributed to varying directivity at different frequencies. The following
analysis will focus primarily on microphone A, as the other SPL predictions follow similar trends.
The URANS predictions show significant under-prediction of the overall SPL and it over-emphasizes
the relative importance of the peaks above the third harmonic. However, it is able to predict the fre-
quency of the 3 major peaks relatively well with a maximum error of 8 %, as seen in Table 4.1. URANS
is only capable of capturing the major structures in the flow as the high turbulent viscosity used in
RANS models dissipate the finer details of the flow. Thus, primarily tonal information is recovered, with
very little broadband noise in the prediction.
In contrast, SPL predictions via LES shows excellent agreement of broadband noise levels till approxi-
mately 2000 Hz; the time-step used is too large to fully resolve the high-frequency content beyond this.
Since LES resolves (as opposed to models, in the case of URANS) the turbulent flow field and energy
cascade up till the limit of mesh size, detail can be recovered throughout the entire resolved range of
Chapter 4. Results and Discussion 43
Table 4.1: Microphone A tonal peaks
Harmonic QFF (3.125 Hz Resolution) URANS (10 Hz Res.) LES (15 Hz Res.) DES (10 Hz Res.)
1 178 Hz, 94.1 dB 192 Hz, 82.2 dB 153 Hz, 83.3 dB 173 Hz, 94.7 dB2 356 Hz, 69.6 dB 384 Hz, 44.0 dB – 356 Hz, 70.3 dB3 536 Hz, 63.7 dB 571 Hz, 37.0 dB – 539 Hz, 65.5 dB
frequencies. However, the three tonal peaks are not reproduced by LES - there is only one spread-out
peak visible near the first harmonic. This peak’s frequency and amplitude are under-predicted by 14 %
and 11.5 %, respectively (Table 4.1). Neither the URANS or LES meshes are detailed enough in the
span-wise direction, as the entire 16 D span of the experiment was modelled while aiming to maintain
a reasonable level of computational cost (<10 million elements). Authors reproducing this experiment
have noted the importance of a fine span-wise resolution [60].
Increased span-wise resolution was achieved without significantly higher computational cost by using
periodic boundary conditions in the DES simulation. The SPL predicted by DES is in excellent agree-
ment with the experimental measurements up till approximately 2000 Hz; like LES, the time-step used
was not sufficient to resolve higher frequency. The three peak frequency predictions have a maximum
error of 2.8 %. The significance of this error reduces even further when the 10 Hz FFT resolution for the
DES data is considered; it is very likely that a finer FFT resolution, achieved simply through a longer
simulation period, would reduce this error. In addition to the excellent peak frequency prediction, the
SPL amplitude error is only 0.6 % at the first harmonic and increases to 2.8 % at higher harmonics.
Table 4.2 compares the maximum errors seen in the sound level predictions of the three models used
in the present work to a selection of published work recreating this experiment, using a wide range of
turbulence models and simulation size. It is clear that it is not necessary to model the complete span in
order to reach high accuracy - both ELAN S-A DDES and the SST k-ω DES conducted in the present
work achieve some of the lowest errors observed using a span of only 3 D. Simulations using above
11 million elements rarely see any substantial increase in accuracy. However, Table 4.2 only considers
the capability of aeroacoustic noise prediction; the higher fidelity models may be able to predict the
hydrodynamic aspects of the flow to higher accuracies.
In summary, the three simulation methods URANS, LES, and DES yielded increasing levels of success,
respectively, in recreating the complex noise produced in the tandem cylinder experiment. URANS,
designed to recreate the most dominant features of the flow, significantly under-predicts the broadband
noise levels. This is rectified by LES, where much more turbulence is resolved rather than modelled.
However, LES still showed difficulty in predicting the tonal portion of the noise. This may be due
to the coarse span-wise resolution used. DES, with a much finer span-wise resolution, produces SPL
with excellent agreement to the QFF experiment. In addition, the accuracy of the DES conducted in
this work is comparable to the the better-performing results in literature, generally achieved with much
higher computational cost. The learnings from simulating this benchmark study can be applied to more
complex geometries in the future, such as the components used in a muffler (e.g., perforated plates and
sharp edges in flow).
Chapter 4. Results and Discussion 44
0
20
40
60
80
100
50 200 500 2000 5000100 1000
SP
L[d
B/H
z]
Frequency [Hz]
QFFURANS
LESDES
(a) Mic A
0
20
40
60
80
100
50 200 500 2000 5000100 1000
SP
L[d
B/H
z]
Frequency [Hz]
(b) Mic B
0
20
40
60
80
100
50 200 500 2000 5000100 1000
SP
L[d
B/H
z]
Frequency [Hz]
(c) Mic C
Figure 4.7: SPL at the three microphone locations
Chapter 4. Results and Discussion 45
Table 4.2: Comparison of the simulation in the present work to a selection of those in literature [59]
Turbulence ModelNum. Elements(millions)
Span / DPeak FrequencyMax. Error
Peak AmplitudeMax. Error
PowerFLOW k-ε DES 66 16 2.0 % 5.5 %CEDRE SST DDES 17 4 14.3 % 7.0 %UPACS-LES Zonal S-A 70 18 1.6 % 1.3 %CFL3D Zonal SST 60 18 7.6 % 4.0 %ELAN S-A DDES 10 3 1.1 % 3.8 %
Present Work
Fluent k-ε RANS 0.5 16 7.9 % 41.9 %Fluent LES 9.8 16 14.0 % 11.5 %Fluent SST k-ω DES 10.7 3 2.8 % 2.8 %
4.2 Single-Objective Optimization (SOO)
The primary goal when designing a muffler is to maximize its transmission loss in the desired portions
of the audible frequency spectrum. This was conducted through single-objective optimization (SOO)
of the plug flow muffler as detailed in Section 3.2. Prior to the optimization process, the prediction
methods and perforate impedance models used were validated using experimental data from literature.
As outlined in Figure 3.4, the optimization process used the TMM with the best performing impedance
model. This section also analyzes the limitations in the optimization process stemming from the use of
TMM. Finally, the solution space is visualized in order to reveal the underlying structure.
4.2.1 TMM & FE Model Evaluation
The first step was an evaluation of the 3 zero mean flow perforate impedance models using both TMM
and FE, by comparison against experiments conducted by Wu et al. [89]. The TL calculated via TMM
are presented in Figure 4.8 alongside the experimental data. The experimental data up to 3200 Hz shows
three major TL peaks occurring at 747 Hz, 1770 Hz, and 2062 Hz. The Sullivan and Bento Coelho models
are in qualitative agreement with the experimental data up to roughly 2200 Hz. Substantial deviations
only begin after the cutoff frequency, fc = 1877 Hz, and are associated with the plane wave assumption
inherent to the TMM. However, both the Sullivan and Bento Coelho models predict the general decrease
in TL above 2200 Hz. The Bauer model, while containing the 3 major peaks, exhibits an increasing fre-
quency shift at higher frequencies; the first peak is shifted by 42 Hz, the second peak by 220 Hz, and the
third peak by 270 Hz. This behavior is currently unexplained. Of the three impedance models used in
the TMM, the Sullivan model performs best, as corroborated by the high Pearson correlation coefficient
listed in Table 4.3.
Both 2DA and 3D FE simulation results are shown in Figure 4.9. The 3D simulations took 70 times the
computational time when compared to the 2DA model, primarily because of the 100 times increase in
the number of elements. The 2DA results lie almost coincident with the 3D results, implying that there
are no 3D modes of pressure distribution in the frequency range considered.
Section 4.2 was previously published in J. Puthuparampil, H. Pong, and P. Sullivan, Modelling and Optimization of PlugFlow Mufflers in Emission Control Systems, in SAE 2017 Noise and Vibration Conference and Exhibition, 2017. [71].
Chapter 4. Results and Discussion 46
Table 4.3: Pearson correlation coefficient - prediction vs. experimental data
Impedance Model TMM 2DA 3D
Sullivan 0.8084 0.6683 0.6681Bento Coelho 0.7445 0.7880 0.7878Bauer 0.2696 0.8496 0.8495
0
10
20
30
40
50
60
0 500 1000 1500 2000 2500 3000
fc = 1877 Hz
Tra
nsm
issi
on
Loss
[dB
]
Frequency [Hz]
Experimental DataTMM - Sullivan
TMM - Bento CoelhoTMM - Bauer
Figure 4.8: TMM evaluation of transmission loss
0
10
20
30
40
50
60
0 500 1000 1500 2000 2500 3000
fc = 1877 Hz
Tra
nsm
issi
onL
oss
[dB
]
Frequency [Hz]
Experimental Data2DA FE - Sullivan
2DA FE - Bento Coelho2DA FE - Bauer3D FE - Sullivan
3D FE - Bento Coelho3D FE - Bauer
Figure 4.9: FEM evaluation of transmission loss
Chapter 4. Results and Discussion 47
Table 4.4: Optimized muffler dimensions
Parameter Optimized Value
rD 0.4000La1 = La2 0.1498 mL1 = L2 0.1144 mLb1 = Lb2 0.060 03 mσ1 = σ2 18.24 %Objective Function (NLO) 12.86 dB
Of the three impedance models used in the FE simulations, the Bauer model shows the highest level
of qualitative accuracy, contrary to the results from the TMM. These findings are further verified by
the high correlation coefficient for the FE Bauer model seen in Table 4.3. Both the Sullivan and Bento
Coelho models show significant deviation from approximately 1400 Hz onwards. However, certain sim-
ilarities exist across TMM and FE - the order in which each impedance model peaks stays consistent.
For example, when considering the second peak in the experimental data (1770 Hz), the Sullivan model
peaks first, followed by the Bento Coelho, and finally the Bauer model. This trend is seen for all TL
peaks in both TMM and FE.
The observations in this section lead to adopting the Sullivan impedance model for TMM and the
Bauer model for 2DA FE simulations, as they provide the highest correlation and qualitative agreement
with experimental data for their respective modelling methods. In the SOO work, TMM was used in
the optimization process and the FE simulation was used only to check the final optimized result.
4.2.2 NLO Optimization
Optimized Solution
The optimization process was conducted to design a plug flow muffler that reduced the noise output of
a generator to the NC-60 standard. 2000 random initializations of the optimization were conducted in 5
hours and 40 minutes on a 4-core Intel i7-4790 processor, with only 3 instances of the optimizer failing
(due to exceeding the 15,000 iterations limit). The dimensions of the best performing design along with
its objective function value is given in Table 4.4. Figure 4.10 shows the associated TMM and 2DA FE
transmission loss predictions. Based on the results from Section 4.2.1, the FE Bauer model is expected
to predict the true TL with a much higher accuracy than the TMM Sullivan model. Thus, the current
section compares the TMM results against FE simulations, in lieu of experimental data.
All 7 major TL peaks, especially the 4 up to 2550 Hz, are predicted equally well by both the TMM and
FE. Additionally, general TL levels agree across both modelling methods, with a region of 3 peaks up to
approximately 1570 Hz, after which there is relatively low TL to 2400 Hz, following which 4 peaks occur.
The accuracy is reflected in the high correlation coefficient of 0.7520 between the two prediction meth-
ods. Due to the high level of agreement between TMM and FE, one can expect the true experimental
TL to be very similar.
Chapter 4. Results and Discussion 48
0
10
20
30
40
50
60
70
0 500 1000 1500 2000 2500 3000 3500 4000
fc = 1877 HzT
ran
smis
sion
Loss
[dB
]
Frequency [Hz]
TMM - SullivanFE - Bauer
Figure 4.10: Optimized muffler transmission loss
Figure 4.11 presents the final noise output from the engine and muffler system based on the TMM
and FE predictions for the optimized muffler. The output noise sound pressure level (SPL) is defined as
SPLi = ENi − TLi (4.2)
and both the TMM and FE results agree extremely well. This is because the Noise Criterion standard
only considers an average over an octave band calculated through Equation 3.4, disregarding small
frequency mismatches.
However, a large discrepancy is observed between the acceptable output SPL determined by NC-60
and the optimized muffler systems output SPL in the 63 Hz, 125 Hz, and 250 Hz octave bands (the
low-frequency range). In fact, this difference is the major contributor to the non-zero evaluation of
the objective function seen in Table 4.4. The inability for the optimized muffler to meet the low-
frequency targets is related to the rD variable and its chosen lower limit of 0.4; TMM exhibits prediction
inaccuracies below this value. However, a strong inverse proportionality between rD and low-frequency
muffler TL was found with an FE parametric study where rD was varied from 0.1 to 0.8, keeping all
other parameters constant at their values from Table 4.4. Figure 4.12 shows the output SPL for the
parametric study; a muffler with rD = 0.1 easily outperforms the NC-60 criterion. However, choosing
such a small diameter for the inner tube introduces manufacturing challenges, lack of structural integrity
if porosity is held constant, and increased pressure drop if porosity is reduced to maintain structural
integrity [66].
Chapter 4. Results and Discussion 49
0
20
40
60
80
100
120
63 125 250 500 1000 2000 4000
Ou
tpu
tS
ou
nd
Pre
ssu
reL
evel
[dB
]
Octave Band [Hz]
Engine NoiseNC-60
TMM - SullivanFE - Bauer
Figure 4.11: Original engine noise, NC-60 criterion, and the output sound pressure levels of the optimizedmuffler
0
20
40
60
80
100
120
63 125 250 500 1000 2000 4000
Ou
tpu
tS
oun
dP
ress
ure
Lev
el[d
B]
Octave Band [Hz]
NC-60rD = 0.1rD = 0.2rD = 0.3rD = 0.4rD = 0.5rD = 0.6rD = 0.7rD = 0.8
Figure 4.12: Parametric study - output sound pressure levels
Chapter 4. Results and Discussion 50
0
200
400
600
800
1000
(12.0
,14
.5]
(14.
5,17.
0]
(17.0
,19
.5]
(19.5
,22.0
]
(22.0
,24.5
]
(24.5
,27.0
]
(27.0
,29.5
]
(29.
5,32.
0]
(32.
0,34.
5]
Cou
nt
NLO
36
325
900
399
179116
21 20 1
Figure 4.13: Histogram of NLO evaluations at the final result of each optimization
Optimization Analysis
A histogram of NLO evaluations of the final result of all 1997 successful optimizations is seen in Figure
4.13. It reveals that only 1.8 % of optimizations reached the lowest recorded bin, while 45.1 % of op-
timizations terminated between NLO = (17, 19.5]. Thus, the use of a multi-start algorithm is justified
and necessary to find the best performing local minima.
Of the 1997 solutions discovered, 1756 of them were unique local minima. This 5-dimensional solution
space can be visualized through dimensionality reduction. Principal component analysis (PCA) was
conducted on a z-score standardized dataset of the unique solutions to map the space onto 2 indepen-
dent variables which maximized the variance of the projection [73]. Due to the correlation between
the La and Lb values, two major clusters with very similar behavior are seen. Thus, focusing on one
of these clusters in the PCA provides a clearer visualization of underlying trends, as shown in Figure
4.14. A region of well-performing solutions is observed surrounding the point (0, -1), as marked by the
circle labelled Optimal Solution Cluster. The majority of the low-NLO solutions lie within this region,
including the previously discussed best-performing design.
Identifying a region of significance such as this provides an effective method to check whether a given
set of muffler dimensions will perform well, simply based on the X1 and X2 values it maps on to. This
check can be executed using the PCA transformation coefficients on the z-score standardized muffler
Chapter 4. Results and Discussion 51
−4
−3
−2
−1
0
1
2
3
4
−3 −2 −1 0 1 2 3 4
X2
X1
12.0
14.5
17.0
19.5
22.0
24.5
NL
O
Optimal Solution Cluster
Figure 4.14: Dimension-reduced visualization of one cluster of 883 unique solutions (local minima),colored by their associated NLO evaluation
dimension values, given as
X1 =
0.6440
−0.08345
−0.3862
0.6166
−0.2213
·
z-score(rD)
z-score(La)
z-score(L)
z-score(Lb)
z-score(σ)
(4.3)
X2 =
0.3263
−0.6727
−0.1524
−0.3273
0.5573
·
z-score(rD)
z-score(La)
z-score(L)
z-score(Lb)
z-score(σ)
(4.4)
4.3 Multi-Objective Optimization (MOO)
For a more holistic design of mufflers, the pressure drop across a muffler must be considered in conjunction
with its acoustic properties. Thus, multi-objective optimization (MOO) as described in Section 3.3
was conducted. As per Figure 3.5, two steps precede the optimization process, 1) validation of the
empirical pressure drop model, and 2) dimensionality reduction in order to identify the variables of
higher importance. Following this, the optimization was conducted with the EGO algorithm and the
Pareto front approximation was compared against results from the standard NSGA-II algorithm.
Chapter 4. Results and Discussion 52
0
2
4
6
8
10
A B C D E F
Pre
ssu
reD
rop
[in
H2O
]
Muffler ID
Empirical ModelCFD
Figure 4.15: Comparison of pressure drop calculated via the empirical pressure drop model and CFD,for 6 random PFM geometries
4.3.1 Empirical Pressure Drop Model Validation
The accuracy of the empirical pressure drop model was evaluated against 2DA steady state RANS
CFD using 6 randomly chosen PFM geometries (Figure 4.15). The empirical model is able to predict
the pressure drop relatively well, with a mean absolute percent error of 12 %. An average error of
this magnitude is acceptable for industrial applications. Importantly, there is a high correlation (0.97)
between the empirical and CFD-based pressure drops, meaning that the multi-objective optimization
would behave similarly regardless of the method of pressure drop estimation.
4.3.2 Dimensionality Reduction
10-fold cross-validated feature selection was performed on the 12 variables defining the PFM, for 10
different sets of TLO weights, as described in Section 3.3.3. The average frequency of a variable being
part of the reduced subset is presented in Table 4.5. rD, rexp, rL1, rL2, σ1, and σ2 are chosen in all
cross-validations for all weights. Only two other variables were chosen at all - r2 and Tp2 at an average of
18 % and 33 % of all cross-validations, respectively. Given the dominating importance of the 6 variables
at a 100 % selection rate, those were selected as the subset to be used in MOO.
4.3.3 EGO’s Performance in MOO
EGO’s effectiveness as a MOO algorithm for muffler optimization is evaluated through the four objectives
outlined in Section 3.3.4. The following naming convention is used to differentiate various optimizations:
Algorithm – Nevals – Initial Condition Set.
Chapter 4. Results and Discussion 53
Table 4.5: Summary of dimensionality reduction
Parameter Average Selection Rate Across Different wi Part of Reduced Subset
rD 100 % ± 0 % Yesrexp 100 % ± 0 % YesrL1 100 % ± 0 % YesrL2 100 % ± 0 % Yesr1 0 % ± 0 % Nor2 18 % ± 7.9 % Noσ1 100 % ± 0 % Yesσ2 100 % ± 0 % YesTp1 0 % ± 0 % NoTp2 33 % ± 24.5 % NoDh1 0 % ± 0 % NoDh2 0 % ± 0 % No
Objective 1 - Comparison of EGO Variants
Figure 4.16 shows the median attainment surfaces of the four EGO variants using Nevals = 240 and
initial condition set A. No significant difference is observed between the four algorithms. MWU tests on
the distribution of the hypervolume (IHV ) and binary additive epsilon (Iε+) metrics show no statistically
significant difference either (Table A.1, Appendix A). ParEGO performs equally as well as any other
variant even though it relies on a kriging model fit to a scalarized version of the objectives. Thus,
unlike the other three methods, only one model-fitting operation is conducted, leading to a nominal
2 to 4 % decrease in execution time. However, a scalarization-based method might prove less effective
if the optimization involved a large number of objectives. It has been shown that for standard multi-
objective test functions, ParEGO performs worse than the other variants [12]. For the purposes of
muffler optimization, where there are only two primary well-behaved objectives, ParEGO proves to
be an effective choice of algorithm. The rest of the present work will use ParEGO when conducting
EGO-based optimization.
Objective 2 - EGO’s Sensitivity to Nevals
The median attainment surfaces comparing ParEGO using Nevals = 240, 300, and 360 total evaluations
show no distinct differences (Figure 4.17). Additionally, the MWU tests (Table A.2, Appendix A) present
no significant statistical difference either. ParEGO’s success at relatively low Nevals provides simulation-
based optimization at industrially relevant timescales. The ParEGO-240-A simulation takes 2.9 hours on
average. If this simulation was conducted with genetic algorithms, it would require anywhere from 2000
to 64000 evaluations (Section 2.5), taking between 20 hours to 640 hours, respectively. Additionally,
GA-based muffler optimization work state that larger evaluation budgets led to better optimized results,
with convergence seen only after a large number of evaluations [22]. GA also requires the tuning of
multiple parameters including crossover probability and mutation probability, for each problem solved
[29]. Slow convergence and the need for meta-optimization (i.e., finding the best parameters required
for GA to optimize effectively) is not ideal for an optimization process - especially not for industrial use.
In contrast, ParEGO’s relative indifference to Nevals makes it well suited for such uses.
Chapter 4. Results and Discussion 54
0
2
4
6
8
10
−30 −25 −20 −15 −10 −5 0
DP
[in
H2O
]
TLO [dB]
ParEGO-240-ASMS-240-A
ε-240-AMSPOT-240-A
Figure 4.16: Median attainment surfaces comparing the 4 EGO variants
0
2
4
6
8
10
−30 −25 −20 −15 −10 −5 0
DP
[in
H2O
]
TLO [dB]
ParEGO-240-AParEGO-300-AParEGO-360-A
Figure 4.17: Median attainment surfaces comparing different Nevals
Chapter 4. Results and Discussion 55
0
2
4
6
8
10
−30 −25 −20 −15 −10 −5 0
DP
[in
H2O
]
TLO [dB]
ParEGO-240-AParEGO-240-BParEGO-240-CParEGO-240-D
Figure 4.18: Median attainment surfaces comparing different sets of initial conditions
Objective 3 - EGO’s Sensitivity to Initial Conditions
ParEGO is seen to be a robust algorithm due to it’s insensitivity to initial conditions. Four different
sets of random initial conditions (A, B, C, and D) resulted in median attainment surfaces that are very
similar (Figure 4.18). The only discernible difference is ParEGO-240-C’s deviation from the other curves
near (-18.5, 2.5). This was due to the presence of an uncharacteristically well-performing muffler in this
region of the attainment surface within the set of 40 initial points for set C. Although ParEGO-240-C’s
mean statistical performance of IHV and Iε+ is marginally better than the other initial condition sets, the
MWU tests show no statistically significant difference (Table A.3, Appendix A). This robust behaviour
is very advantageous in an industrial setting as repetitions of the optimization is not necessary if the
computational resource and time are not available.
Objective 4 - Comparison of EGO Against NSGA-II
The performance of the NSGA-II algorithm using Nevals = 240, 300, and 360 show significantly worse
results than ParEGO-240-A (Figure 4.19). The IHV and Iε+ metrics in Figures 4.20 and 4.21 depict the
vast difference in performance between the ParEGO and the NSGA-II algorithms. Table A.4 (Appendix
A) contains the MWU test on the two metrics and it corroborates the statistically significant differences
seen in Figures 4.20 and 4.21. In addition, due to the Nevals = npop ·ngen restriction, NSGA-II produces
Pareto fronts of very low resolution. The resolution is determined by npop; increasing it necessitates a
reduction in ngen, for a given Nevals. Increased resolution would result in solutions that are not very
different from the random initial conditions since the population would not have had time to evolve into
a better state. The EGO algorithm circumvents this issue entirely as it is free to add or remove candidate
solutions from the final Pareto front approximation at each iteration, based only on the performance of
a given solution.
Chapter 4. Results and Discussion 56
0
2
4
6
8
10
−30 −25 −20 −15 −10 −5 0
DP
[in
H2O
]
TLO [dB]
ParEGO-240-ANSGAII-240NSGAII-300NSGAII-360
Figure 4.19: Median attainment surfaces comparing the ParEGO algorithm and NSGA-II
In addition, there is no correlation seen in the NSGA-II results between the quality of a Pareto front and
Nevals; others have reported a strong positive correlation between evaluation budget and the quality of
results prior to convergence when using GA [22]. The lack of correlation here is attributed to the fact
that GA is simply not designed to perform at such low Nevals; the evolutionary strategies in the algo-
rithm are fruitful only under large populations with many generations - both of which are not possible
with low Nevals.
4.3.4 Analysis of EGO Model
A major advantage of EGO (or any model-based optimization algorithm) over evolutionary methods
is that the underlying model developed by the algorithm can be visualized and analyzed, providing
an understanding of the relationships between different optimization variables and how they affect the
objective function. Figure 4.22 is the model visualization of one ParEGO-240-A run. The contours are
coloured based on the sum of the normalized values of TLO and DP
DP + TLO = wDP
(DP−DPlow
DPhigh −DPlow
)+ wTLO
(TLO− TLOlow
TLOhigh − TLOlow
)(4.5)
where the tilde represents normalized values, DPlow = 0.0, DPhigh = 40, TLOlow = −50, TLOhigh = −10,
and wDP and wTLO are the weights for DP and TLO, both set equal to 1.0. Equation 4.5 can be thought
of as a simplified version of the scalarized objective function that is minimized by ParEGO. Minimizing
Equation 4.5 leads to better performing mufflers with low DP and high TLO. Figure 4.22 has each opti-
mization variable is plotted against every other, in order to reveal the inter-relations between variables.
In addition, the final Pareto front approximation from this run is plotted as white points on each plot.
Chapter 4. Results and Discussion 57
60
80
100
120
140
160
180
200
220
ParEGO-240-A NSGAII-240 NSGAII-300 NSGAII-360
I HV
Figure 4.20: Hypervolume metric (IHV ) comparing the ParEGO algorithm and NSGAII
−5
0
5
10
15
20
25
(ParEGO-240-A,
NSGAII-240)
(ParEGO-240-A,
NSGAII-300)
(ParEGO-240-A,
NSGAII-360)
I ε+
Figure 4.21: Binary additive epsilon metric (Iε+) comparing the ParEGO algorithm and NSGAII
Chapter 4. Results and Discussion 58
In general, these optimized points lie closer to the lower contour values (darker colours), validating
the visualization. However, some points are seen to stray away to higher contour values. This is because
the visualization only used one set of weights (wDP = 1.0 and wTLO = 1.0); the ParEGO algorithm uses
a randomized pair of weights in each iteration to develop it’s true optimization model. Thus, points
seen to lie outside the darker zones in Figure 4.22 most probably lie within dark zones under different
visualization weights.
Further observation and analysis of Figure 4.22 provides some insight into the behaviour of each variable:
1. rD: The graphs in column A show a maxima for low rD values, regardless of what the other
variable is. This shows the high sensitivity of rD, since at low rD the pressure drop grows very
large, while TLO doesn’t increase as dramatically. We see optimized points spanning [0.3, 0.9],
showing that an effective muffler can be constructed using a variety of expansion ratios.
2. rexp: All plots associated with rexp (A-I and column B) show a relatively symmetric trend as rexp
controls the geometric symmetry of the muffler, across the plug in the PFM. Too low would be
a short expansion section leading to high DP, and too high would be a short contraction section,
also leading to high DP. Thus, all the optimized points lie within the middle range of [0.4, 0.85].
3. rL1: As with rexp, the optimized points lie in the middle range of [0.3, 0.7], where the contour
values are seen to be quite low (A-II, B-II, and column C). Only three solutions are seen to be
outside this range into high rL1 values (the three points are seen in the top-right of plot A-II). The
high rD values of these three solutions allow for a lower pressure drop, balancing out the larger
pressure drop incurred by moving to higher rL1 region.
4. rL2: It is similar to rL1 in trends, except the optimized results seem to have a slightly larger spread
(row III, D-IV, and D-V). This could indicate that the muffler performance is marginally less
sensitive to the contraction section of the PFM. Additionally, both rL1 and rL2 show a bifurcating
behaviour when compared against rD. The optimized points greater than rD = 0.7 seem to split
into two separate paths in plots A-II and A-III. This indicates that for rD > 0.7, two separate
sets of solutions exist. Furthermore, the rL1 vs rL2 graph (C-III) shows the minima of the contour
offset from the center of the plot, signifying that a slight asymmetry between the lengths of the
perforated sections is important (too asymmetric would incur significant pressure drop penalties).
5. σ1 and σ2: The optimized points exhibit a large spread across both these variables; it is an indicator
of the relatively lower sensitivity of σ1 and σ2. Plot E-V shows that a higher number of optimized
points are skewed toward higher σ1. This could be because the pressure drop associated with the
expansion section is higher than the contraction section, for a given OAR (Equations 2.36 and
2.37). Thus, the optimizer is able to achieve lower DP by using higher σ1 values, while sacrificing
some TL performance. As σ2 has lower DP for a given OAR, the optimizer has more freedom in
exploring a larger range of this variable, resulting in a relatively even distribution of optimized
points along the σ2 axis, from approximately [0.2, 0.8].
Chapter 4. Results and Discussion 59
0
1
Sum of normalized TLO and DP
A B C D E
I
II
III
IV
V
0
1
0
1
0
1
0
1
0 1 0 1 0 1 0 1 0 1
rexp
0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2
rL1
rL2
σ1
σ2
rD rexp rL1 rL2 σ1
Figure 4.22: Visualization of kriging model developed by the ParEGO algorithm for one set of TLO andDP weights, with the optimal points on the Pareto front approximation overlaid as white points
Chapter 4. Results and Discussion 60
4.4 Chapter Summary
Section 4.1 contained the aeroacoustic noise prediction of the tandem cylinder experiments. The short-
span DES work conducted produced an excellent prediction of the measured noise, at significantly
lower computational cost than most of the work in literature. The single-objective muffler optimization
in Section 4.2 showed the effectiveness of L-BFGS-B (a relatively simple gradient-based optimization
algorithm) in a multi-start framework. In addition to arriving at a successfully optimized muffler, this
section contained analysis and visualization of the solution space. Finally, Section 4.3 demonstrated
that the EGO algorithm is robust and extremely capable of multi-objective muffler optimization, and
performs significantly better than the commonly used NSGA-II algorithm. Additionally, the models of
the solution space developed by EGO were visualized and analyzed.
Chapter 5
Conclusion and Future Work
5.1 Conclusion
The ideal muffler is found at the intersection of high acoustic performance, low pressure drop through
the muffler, and a low amount of flow-induced aeroacoustic noise. This work addressed muffler design
in a holistic manned by incorporating all three of the aforementioned aspects.
As the first steps towards aeroacoustic noise predictions of a muffler, the geometrically simpler tandem
cylinder benchmark experiment was studied. Testing the ability of URANS, LES, and DES to predict
the complex noise in a computationally cost-effective manner revealed that DES is an effective tool for
aeroacoustic prediction. DES was able to reproduce the hydrodynamics of the flow with an industrially
sufficient level of accuracy. The relatively low computational-cost DES conducted in this work produced
acoustic predictions as accurate as much more computationally expensive simulations found in literature.
In order to design a muffler with the optimal acoustic performance, the general guidelines and strategies
used by muffler manufacturers is not enough; automated optimization is necessary. The single-objective
optimization portion of this work showed that relatively simple gradient-based optimization algorithms
were successful in optimizing a plug flow muffler’s transmission loss. This is in contrast to the complex
algorithms, such as genetic algorithms or simulated annealing, typically used in such work.
However, in order to consider pressure drop alongside acoustic transmission loss in large mufflers,
simulation-based multi-objective optimization was conducted using the EGO algorithm. Surrogate
model-based algorithms such as EGO are very efficient when working with real engineering problems
with computationally expensive objective functions, as the surrogate model provides an extremely cost-
effective view of the objective space. EGO was shown to outperform the commonly used NSGA-II
genetic algorithm in the multi-objective optimization of a plug flow muffler.
61
Chapter 5. Conclusion and Future Work 62
5.2 Future Work
5.2.1 Aeroacoustics
The aeroacoustic noise predictions conducted in the present work was only an introductory exploration.
Future work can expand by:
1. Porting the simulation into an open source CFD platform such as OpenFOAM in order to allow
for computation in a high performance computing cluster
2. Simulating more complex and industrially relevant 3D geometry; several such geometries exist
as part of AIAA’s Second Workshop on Benchmark problems for Airframe Noise Computations
(BANC-II) [2] (the tandem cylinder experiment used in this work is part of this database)
3. Visualization of the noise sources to better understand the effects of different geometries and noise
producing mechanisms
5.2.2 Muffler Optimization
Optimization work using EGO can be expanded to:
1. Calculate pressure drop via a CFD simulation, removing any limitations imposed by the empirical
pressure drop model and allowing the exploration of more varied designs
2. Include a third objective which aims to minimize the strength of the aeroacoustic noise sources,
calculated through a CFD simulation (low-fidelity aeroacoustic predictions based on steady-state
RANS is available in the form of Broadband Noise Source (BNS) models [95])
3. Different styles of mufflers, including the combination of reactive and absorptive silencing
Appendix A
Mann-Whitney U (MWU) Test
To answer the multi-objective optimization objectives (Section 3.3.4) in a statistical fashion, IHV and
Iε+ from multiple runs of one method was compared against another method using the Mann-Whitney U
(MWU) test. MWU aims to discern if two independent samples were selected from the same population
(null hypothesis), but without requiring a normal distribution (unlike the Student’s t-test). Proving the
null hypothesis false would show that there is a statistical difference between two sets of data.
The MWU test relies on calculating the U statistic. If U ≤ Ucrit, where Ucrit is obtained from tabulated
values based on sample size, the null hypothesis is rejected - i.e., the two sets of data are from different
populations. In this work, the MWU test was conducted with a significance threshold (p-level) of 0.05
(two-tailed); this yields Ucrit = 13.
Table A.1: IHV and Iε+ for Objective 1; bolded algorithm is the baseline in MWU test
ParEGO-240-A SMS-240-A EPS-240-A MSPOT-240-A
IHV 186.0± 5.7 185.8± 4.4 187.8± 4.3 188.0± 4.4U Value – 30 26 27Similar to Baseline? – Yes Yes Yes
Iε+ – (1.095, 1.169) (1.022, 0.975) (0.945, 1.037)U Value – 30 25 27Similar to Baseline? – Yes Yes Yes
63
Appendix A. Mann-Whitney U (MWU) Test 64
Table A.2: IHV and Iε+ for Objective 2; bolded algorithm is the baseline in MWU test
ParEGO-240-A ParEGO-300-A ParEGO-360-A
IHV 186.0± 5.7 190.3± 6.1 189.1± 3.1U Value – 22 22Similar to Baseline? – Yes Yes
Iε+ – (1.188, 0.909) (1.376, 0.869)U Value – 16 13Similar to Baseline? – Yes No (nominally)
Table A.3: IHV and Iε+ for Objective 3; bolded algorithm is the baseline in MWU test
ParEGO-240-A ParEGO-240-B ParEGO-240-C ParEGO-240-D
IHV 186.0± 5.7 187.5± 5.4 193.5± 5.8 187.6± 2.6U Value – 29 10 23Similar to Baseline? – Yes No Yes
Iε+ – (1.178, 1.578) (1.675, 1.272) (1.254, 1.508)U Value – 30 20 029Similar to Baseline? – Yes Yes Yes
Table A.4: IHV and Iε+ for Objective 4; bolded algorithm is the baseline in MWU test
ParEGO-240-A NSGAII-240 NSGAII-300 NSGAII-360
IHV 186.0± 5.7 134.9± 46.4 123.6± 31.6 108.6± 38.4U Value – 16 0 0Similar to Baseline? – Yes No No
Iε+ – (0.927, 8.609) (0.481, 8.441) (0.396, 13.246)U Value – 4 0 0Similar to Baseline? – No No No
Appendix B
Tandem Cylinder Simulations -
Additional Results
B.1 2D TKE - Line Extractions
Horizontal line extractions (y = 0 plane) of 2D TKE in the gap between the two cylinders and aft of the
second cylinder are seen in Figure B.1. Similarly, vertical line extractions at x = 1.5 D and x = 4.45 D
are presented in Figure B.2. Both extractions are created from the 2D TKE contour plots (Figure 4.5).
B.2 Span-wise Correlation of Pressure
Figure B.3 shows the span-wise correlation of pressure along both cylinders at θ = 135° for one half
of the domain (z/D ≥ 0). The correlation is calculated between the central point at z/D = 0 and
other span-wise locations. The span-wise correlation in the QFF data reduces towards 0 as one moves
closer to the no-slip condition at the walls of the wind tunnel (z/D = 8). DES data for C1 shows a
slow decrease of span-wise correlation which does not reach 0, primarily due to the periodic boundary
conditions used. The span-wise correlation on C2 is in higher agreement with experimental data, but it
too suffers from the inaccuracies of the periodic boundary condition (in reality, it is a no-slip boundary
condition at z/D = ±8).
65
Appendix B. Tandem Cylinder Simulations - Additional Results 66
0
0.05
0.1
0.15
0.2
0.25
0.3
0.5 1 1.5 2 2.5 3
2D
TK
E/U
2 o
x/D
0
0.05
0.1
0.15
0.2
0.25
0.3
4.5 5 5.5 6 6.5
2DT
KE/U
2 o
x/D
QFFLESDES
Figure B.1: 2D TKE along y = 0
0
0.05
0.1
0.15
0.2
0.25
0.3
−0.8−0.6−0.4−0.2 0 0.2 0.4 0.6 0.8
2DT
KE/U
2 o
y/D
0
0.05
0.1
0.15
0.2
0.25
0.3
−0.8−0.6−0.4−0.2 0 0.2 0.4 0.6 0.8
2DT
KE/U
2 o
y/D
QFFLESDES
Figure B.2: 2D TKE along x = 1.5 D and x = 4.45 D
0
0.2
0.4
0.6
0.8
1
0 1 2 3 4 5 6 7
Cor
rela
tion
z/D
QFFDES
0
0.2
0.4
0.6
0.8
1
0 1 2 3 4 5 6 7
Cor
rela
tion
z/D
Figure B.3: Span-wise correlation of surface pressure at 135° on C1 and C2
Bibliography
[1] Acoustics - Guidelines for noise control by silencers - ISO 14163:1998.
[2] BANC-II Workshop, 2012.
[3] COMSOL - The Acoustics Module Users Guide, 2013.
[4] Dennis Aaberg. Generator set noise solutions: Controlling unwanted noise from on-site power
systems. Technical report, Cummins Power Generation Inc., 2007.
[5] Larry Armijo. Minimization of functions having Lipschitz continuous first partial derivatives. Pacific
Journal of Mathematics, 16(1):1–3, 1 1966.
[6] Anne Auger, Johannes Bader, Dimo Brockhoff, and Eckart Zitzler. Theory of the hypervolume
indicator. In Proceedings of the tenth ACM SIGEVO workshop on Foundations of genetic algorithms
- FOGA ’09, page 87, New York, New York, USA, 2009. ACM Press.
[7] Renato Barbieri and Nilson Barbieri. Finite element acoustic simulation based shape optimization
of a muffler. Applied Acoustics, 67(4):346–357, 4 2006.
[8] Renato Barbieri, Nilson Barbieri, and Key Fonseca De Lima. Some applications of the PSO for
optimization of acoustic filters. Applied Acoustics, 89:62–70, 2015.
[9] Andrew B. Bauer. Impedance Theory and Measurements on Porous Acoustic Liners. Journal of
Aircraft, 14(8):720–728, 8 1977.
[10] Leo L. Beranek. Balanced noisecriterion (NCB) curves. The Journal of the Acoustical Society of
America, 86(2):650–664, 8 1989.
[11] Rmy Binois, Nicolas Dauchez, Jean-Michel Ville, Emmanuel Perrey-Debain, and Ghislain Beillard.
Modeling of cylindrical baffle mufflers for low frequency sound propagation. In Acoustics, 2012.
[12] Bernd Bischl, Jakob Richter, Jakob Bossek, Daniel Horn, Janek Thomas, and Michel Lang. ml-
rMBO: A Modular Framework for Model-Based Optimization of Expensive Black-Box Functions. 3
2017.
[13] Stephen P. Boyd and Lieven. Vandenberghe. Convex optimization. Cambridge University Press,
2004.
[14] C. G. Broyden. The Convergence of a Class of Double-rank Minimization Algorithms 1. General
Considerations. IMA Journal of Applied Mathematics, 6(1):76–90, 3 1970.
67
BIBLIOGRAPHY 68
[15] Richard H. Byrd, Peihuang Lu, Jorge Nocedal, and Ciyou Zhu. A Limited Memory Algorithm
for Bound Constrained Optimization. SIAM Journal on Scientific Computing, 16(5):1190–1208, 9
1995.
[16] Richard H. Byrd, Jorge Nocedal, and Robert B. Schnabel. Representations of quasi-Newton matrices
and their use in limited memory methods. Mathematical Programming, 63(1-3):129–156, 1 1994.
[17] Ying Chun Chang and Min Chie Chiu. Shape optimization of one-chamber perforated plug/non-
plug mufflers by simulated annealing method. International Journal for Numerical Methods in
Engineering, 74(10):1592–1620, 2008.
[18] Min-Chie Chiu. Optimal Design of Multichamber Mufflers Hybridized With Perforated Intruding
Inlets and Resonating Tubes Using Simulated Annealing. Journal of Vibration and Acoustics,
132(5):054503, 10 2010.
[19] Min Chie Chiu. Shape optimization of multi-chamber mufflers with plug-inlet tube on a venting
process by genetic algorithms. 71(6):495–505, 6 2010.
[20] Min-Chie Chiu. Numerical assessment for a broadband and tuned noise using hybrid mufflers and
a simulated annealing method. Journal of Sound and Vibration, 332(12):2923–2940, 6 2013.
[21] Min-Chie Chiu. Numerical assessment of two-chamber mufflers hybridized with multiple parallel
perforated plug tubes using simulated annealing method. Journal of Low Frequency Noise, Vibration
and Active Control, 36(1):3–26, 3 2017.
[22] Min Chie Chiu and Ying Chun Chang. Numerical studies on venting system with multi-chamber
perforated mufflers by GA optimization. Applied Acoustics, 69(11):1017–1037, 11 2008.
[23] Min Chie Chiu and Ying Chun Chang. Shape optimization of multi-chamber cross-flow mufflers by
SA optimization. Journal of Sound and Vibration, 312(3):526–550, 2008.
[24] N. Curle. The Influence of Solid Boundaries upon Aerodynamic Sound. Proceedings of the Royal
Society of London A: Mathematical, Physical and Engineering Sciences, 231(1187), 1955.
[25] Balasubramanian Datchanamourty. Detailed Modelling of Mufflers With Perforated Tubes Using
Substructure Boundary Element Method. PhD thesis, University of Kentucky, 2004.
[26] Key Fonseca de Lima, Arcanjo Lenzi, and Renato Barbieri. The study of reactive silencers by shape
and parametric optimization techniques. Applied Acoustics, 72(4):142–150, 3 2011.
[27] Kalyanmoy Deb, Amrit Pratap, Sameer Agarwal, and T Meyarivan. A Fast and Elitist Multi-
objective Genetic Algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation, 6(2),
2002.
[28] L. C. W. Dixon and M. Jha. Parallel algorithms for global optimization. Journal of Optimization
Theory and Applications, 79(2):385–395, 11 1993.
[29] A.E. Eiben and S.K. Smit. Parameter tuning for configuring and analyzing evolutionary algorithms.
Swarm and Evolutionary Computation, 1(1):19–31, 3 2011.
BIBLIOGRAPHY 69
[30] Tamer Elnady. Modelling and Characterization of Perforates in Lined Ducts and Mufflers. PhD
thesis, The Royal Institute of Technology (KTH), 2004.
[31] L.J. Eriksson. Higher order mode effects in circular ducts and expansion chambers. The Journal of
the Acoustical Society of America, 68(2):545, 1980.
[32] J. E. Ffowcs Williams and D. L. Hawkings. Sound Generation by Turbulence and Surfaces in
Arbitrary Motion. Mathematical and Physical Sciences, 264(1151):321–342, 1969.
[33] R. Fletcher. A new approach to variable metric algorithms. The Computer Journal, 13(3):317–322,
3 1970.
[34] R. Fletcher and C. M. Reeves. Function minimization by conjugate gradients. The Computer
Journal, 7(2):149–154, 2 1964.
[35] Tushar Goel. Elitist Non-dominated Sorting Genetic Algorithm: NSGA-II, 2011.
[36] Donald Goldfarb. A family of variable-metric methods derived by variational means. Mathematics
of Computation, 24(109):23–23, 1 1970.
[37] Johan Gustafsson. Mathematical modelling and solutions to Flow Acoustical problems Master’s
Thesis in the Master’s programme in Sound and Vibration. PhD thesis, Chalmers University of
Technology, 2016.
[38] Mark Hall, Eibe Frank, Geoffrey Holmes, Bernhard Pfahringer, Peter Reutemann, and Ian H.
Witten. The WEKA data mining software. ACM SIGKDD Explorations Newsletter, 11(1):10, 11
2009.
[39] Mark A. Hall. Correlation-based Feature Selection for Machine Learning. PhD thesis, The University
of Waikato, 1999.
[40] Daniel Horn, Tobias Wagner, Dirk Biermann, Claus Weihs, and Bernd Bischl. Model-Based Multi-
objective Optimization: Taxonomy, Multi-Point Proposal, Toolbox and Benchmark. pages 64–78.
Springer, Cham, 2015.
[41] Hua Huang, Jimin Ni, Qianying Du, and Nan Xie. Multi-objectives Optimization on Exhaust
Muffler Based on DoE. In 2012 Fourth International Conference on Computational and Information
Sciences, pages 29–32. IEEE, 8 2012.
[42] J.-G. Ih and B.-H. Lee. Theoretical prediction of the transmission loss of circular reversing chamber
mufflers. Journal of Sound and Vibration, 112(2):261–272, 1 1987.
[43] Krishnamurthy Jayaraman and Kit Yam. Decoupling approach to modeling perforated tube muffler
components. The Journal of the Acoustical Society of America, 69(2):390, 1981.
[44] Luther Jenkins, Dan Neuhart, Cathy McGinley, Mehdi Khorrami, and Meelan Choudhari. Measure-
ments of Unsteady Wake Interference Between Tandem Cylinders. In 36th AIAA Fluid Dynamics
Conference and Exhibit, Reston, Virigina, 6 2006. American Institute of Aeronautics and Astronau-
tics.
BIBLIOGRAPHY 70
[45] Luther N. Jenkins, Mehdi R. Khorrami, Meelan M. Choudhari, and Catherine B. McGinley. Char-
acterization of unsteady flow structures around tandem cylinders for component interaction studies
in airframe noise. AIAA paper, 2812(May):2005, 5 2005.
[46] Donald R. Jones, Matthias Schonlau, and William J. Welch. Efficient Global Optimization of
Expensive Black-Box Functions. Journal of Global Optimization, 13(4):455–492, 1998.
[47] Eric Jones, Travis Oliphant, and Pearu Peterson. SciPy: Open source scientific tools for Python,
2001.
[48] Abdelkader Khamchane, Youcef Khelfaoui, and Brahim Hamtache. Shape optimization of reactive
mufflers using threshold Acceptance and FEM. The Journal of the Acoustical Society of America,
141(5):3881–3881, 5 2017.
[49] Josef Kittler. Feature set search algorithms. Pattern recognition and signal processing, 1978.
[50] J. Knowles. A summary-attainment-surface plotting method for visualizing the performance of
stochastic multiobjective optimizers. In 5th International Conference on Intelligent Systems Design
and Applications (ISDA’05), pages 552–557. IEEE, 2005.
[51] J. Knowles. ParEGO: a hybrid algorithm with on-line landscape approximation for expensive
multiobjective optimization problems. IEEE Transactions on Evolutionary Computation, 10(1):50–
66, 2 2006.
[52] J. Knowles and D. Corne. The Pareto archived evolution strategy: a new baseline algorithm
for Pareto multiobjective optimisation. In Proceedings of the 1999 Congress on Evolutionary
Computation-CEC99 (Cat. No. 99TH8406), pages 98–105. IEEE.
[53] Jin Woo Lee. Optimal topology of reactive muffler achieving target transmission loss values: Design
and experiment. Applied Acoustics, 88:104–113, 2 2015.
[54] Jin Woo Lee and Gang-Won Jang. Topology design of reactive mufflers for enhancing their acoustic
attenuation performance and flow characteristics simultaneously. International Journal for Numer-
ical Methods in Engineering, 91(5):552–570, 8 2012.
[55] Jin Woo Lee and Yoon Young Kim. Topology optimization of muffler internal partitions for im-
proving acoustical attenuation performance. International Journal for Numerical Methods in En-
gineering, 80(4):455–477, 10 2009.
[56] Kenneth Levenberg. A method for the solution of certain non-linear problems in least squares.
Quarterly of Applied Mathematics, 2(2):164–168, 7 1944.
[57] M. J. Lighthill. On Sound Generated Aerodynamically. I. General Theory. Proceedings of the Royal
Society of London A: Mathematical, Physical and Engineering Sciences, 211(1107), 1952.
[58] M. J. Lighthill. On Sound Generated Aerodynamically. II. Turbulence as a Source of Sound. Proceed-
ings of the Royal Society of London A: Mathematical, Physical and Engineering Sciences, 222(1148),
1954.
BIBLIOGRAPHY 71
[59] David Lockard. Summary of the Tandem Cylinder Solutions from the Benchmark problems for
Airframe Noise Computations-I Workshop. 49th AIAA Aerospace Sciences Meeting including the
New Horizons Forum and Aerospace Exposition, (January):1–22, 1 2011.
[60] David P Lockard, Mehdi R Khorrami, Meelan M Choudhari, Florence V Hutcheson, and Thomas F
Brooks. Tandem Cylinder Noise Predictions. In AIAA Aeroacoustics Conference, volume 2007-3450,
pages 1–26, Reston, Virigina, 5 2007. American Institute of Aeronautics and Astronautics.
[61] Donald W. Marquardt. An Algorithm for Least-Squares Estimation of Nonlinear Parameters. Jour-
nal of the Society for Industrial and Applied Mathematics, 11(2):431–441, 6 1963.
[62] M L Munjal. A Simple Numerical Method for Three-Dimensional Analysis of Simple Expansion
Chamber Mufflers of Rectangular as well as Circular Cross-Section With a Stationary Medium.
Journal of Sound and Vibration, 116(1):71–88, 1987.
[63] M. L. Munjal. Acoustics of Ducts and Mufflers With Application to Exhaust and Ventilation System
Design. John Wiley & Sons, New York, 1987.
[64] M. L. Munjal. Muffler Acoustics. In Fridolin P. Mechel, editor, Formulas of Acoustics, chapter K.
Springer, New York, 2nd edition, 2008.
[65] M. L. Munjal, K. N. Rao, and A. D. Sahasrabudhe. Aeroacoustic analysis of perforated muffler
components. Journal of Sound and Vibration, 114(2):173–188, 1987.
[66] M.L. Munjal, N.K. Vijayasree, and Chaitanya Paruchuri. Flow resistance network analysis of
the back-pressure of automotive mufflers. Indian Journal of Engineering and Materials Sciences,
20(5):336–349, 2013.
[67] J. A. Nelder and R. Mead. A Simplex Method for Function Minimization. The Computer Journal,
7(4):308–313, 1 1965.
[68] Jorge Nocedal. Updating quasi-Newton matrices with limited storage. Mathematics of Computation,
35(151):773–773, 9 1980.
[69] Wolfgang Ponweiser, Tobias Wagner, Dirk Biermann, and Markus Vincze. Multiobjective Optimiza-
tion on a Limited Budget of Evaluations Using Model-Assisted S-Metric Selection. In G. Rudolph,
T. Jansen, N. Beume, S. Lucas, and C. Poloni, editors, Parallel Problem Solving from Nature PPSN
X, pages 784–794. Springer, Heidelberg, 2008.
[70] Pavel Pudil and Jana Novovicova. Novel Methods for Feature Subset Selection with Respect to
Problem Knowledge. In Feature Extraction, Construction and Selection, pages 101–116. Springer
US, Boston, MA, 1998.
[71] Jobin Puthuparampil, Henry Pong, and Pierre Sullivan. Modelling and Optimization of Plug Flow
Mufflers in Emission Control Systems. In SAE 2017 Noise and Vibration Conference and Exhibition,
6 2017.
[72] Elaine. Rich and Kevin. Knight. Artificial intelligence. McGraw-Hill, 1991.
[73] Markus Ringner. What is principal component analysis? NATURE BIOTECHNOLOGY, 26(3),
2008.
BIBLIOGRAPHY 72
[74] R D Sandberg. Compressible-Flow DNS with Application to Airfoil Noise. Flow Turbulence Com-
bust, 95:211–229, 2015.
[75] A. Selamet and Z.L. Ji. Acoustic Attenuation Performance of Circular Expansion Chambers with
Offset Inlet/Outlet: I. Analytical Approach. Journal of Sound and Vibration, 213(4):601–617, 6
1998.
[76] Jung H. Seo and Young J. Moon. Aerodynamic noise prediction for long-span bodies. Journal of
Sound and Vibration, 306(3-5):564–579, 2007.
[77] D. F. Shanno. Conditioning of quasi-Newton methods for function minimization. Mathematics of
Computation, 24(111):647–647, 9 1970.
[78] Matthew Shardlow. An Analysis of Feature Selection Techniques. Technical report, The University
of Manchester, 2016.
[79] Chao Shen and Liang Hou. Comparison of various algorithms for improving acoustic attenuation
performance and flow characteristic of reactive mufflers. Applied Acoustics, 116:291–296, 1 2017.
[80] Hui Shen, Hua Huang, and Zhen Bin Chen. Establishment of Exhaust Muffler Comprehensive
Evaluation System and Muffler Performance Optimization. Advanced Materials Research, 538-
541:2789–2793, 6 2012.
[81] Daniela Siano, Fabio Bozza, and Fabio Auriemma. Acoustic and fluid-dynamic optimization of an
automotive muffler. Proceedings of the Institution of Mechanical Engineers, Part D: Journal of
Automobile Engineering, 227(5):735–747, 5 2012.
[82] P. L. Smith and Matthew Van Winkle. Discharge coefficients through perforated plates at reynolds
numbers of 400 to 3,000. AIChE Journal, 4(3):266–268, 9 1958.
[83] P. R. Spalart. Young-Person’s Guide to Detached-Eddy Simulation Grids. Technical report, NASA,
Hampton, Virginia, 2001.
[84] Joseph W. Sullivan. A method for modeling perforated tube muffler components. II. Applications.
The Journal of the Acoustical Society of America, 66(3):779, 1979.
[85] Joseph W. Sullivan and Malcolm J. Crocker. Analysis of concentric-tube resonators having unpar-
titioned cavities. The Journal of the Acoustical Society of America, 64(1):207, 1978.
[86] Prakash T. Thawani and Krishnamurthy Jayaraman. Modeling and applications of straight-through
resonators. Journal of the Acoustical Society of America, 73(4):1387–11389, 1983.
[87] Tobias Wagner. Planning and multi-objective optimization of manufacturing processes by means of
empirical surrogate models. PhD thesis, TU Dortmund, 2013.
[88] Meng Wang, Jonathan B Freund, and Sanjiva K Lele. Computational Prediction of Flow-Generated
Sound. Annu. Rev. Fluid Mech, 38:483–512, 2006.
[89] T.W. Wu, P. Zhang, and C.Y.R. Cheng. Boundary Element Analysis of Mufflers With an Improved
Method for Deriving the Four-pole Parameters. Journal of Sound and Vibration, 217(4):767–779,
1998.
BIBLIOGRAPHY 73
[90] Esubalewe Lakie Yedeg, Eddie Wadbro, and Martin Berggren. Interior layout topology optimization
of a reactive muffler. Structural and Multidisciplinary Optimization, 53(4):645–656, 4 2016.
[91] Long Jyi Yeh, Ying Chun Chang, and Min Chie Chiu. Numerical studies on constrained venting
system with reactive mufflers by GA optimization. International Journal for Numerical Methods in
Engineering, 65(8):1165–1185, 2 2006.
[92] Long Jyi Yeh, Ying Chun Chang, Min Chie Chiu, and Gaung Jer Lai. GA optimization on multi-
segments muffler under space constraints. Applied Acoustics, 65(5):521–543, 5 2004.
[93] Long Jyi Yeh, Ying Chun Chang, Min Chie Chiu, and Gaung Jer Lay. Computer-aided optimal
design of a single-chamber muffler with side inlet/outlet under space constraints. Journal of Marine
Science and Technology, 11(4):189–196, 2003.
[94] Martin Zaefferer, Thomas Bartz-Beielstein, Boris Naujoks, Tobias Wagner, and Michael Emmerich.
A Case Study on Multi-Criteria Optimization of an Event Detection Software under Limited Bud-
gets. In R.C. Purshouse, P.J. Fleming, C.M. Fonseca, S. Greco, and J. Shaw, editors, Evolutionary
Multi-Criterion Optimization, pages 756–770. Springer, Heidelberg, 2013.
[95] Lei Zhang, Rui Wang, and Songling Wang. Simulation of Broadband Noise Sources of an Axial Fan
under Rotating Stall Conditions. Advances in Mechanical Engineering, 6:507079, 1 2014.
[96] Qi Zhou, Chaoqun Wu, Weijing Zhao, Weijie Hua, and Linghao Liu. A Novel Auto Design Method
of Acoustic Filter Based on Genetic Programming. pages 661–690. Springer, Singapore, 10 2018.
[97] E. Zitzler, L. Thiele, M. Laumanns, C.M. Fonseca, and V.G. da Fonseca. Performance assess-
ment of multiobjective optimizers: an analysis and review. IEEE Transactions on Evolutionary
Computation, 7(2):117–132, 4 2003.
[98] Eckart Zitzler. Evolutionary Algorithms for Multiobjective Optimization: Methods and Applications.
PhD thesis, Swiss Federal Institute of Technology (ETH), 1999.