Full Text 01

DEGREE PROJECT IN MECHANICAL ENGINEERING,
SECOND CYCLE, 30 CREDITS

STOCKHOLM, SWEDEN 2021
Shape optimization of axial

cooling fan via 3D CFD
simulation and surrogate
modeling
MARTIN GRANLÖF
KTH ROYAL INSTITUTE OF TECHNOLOGY

SCHOOL OF ENGINEERING SCIENCES
Shape optimization of axial cooling fan via 3D CFD simulation and
surrogate modeling / Formoptimering av axiel kylningsfläkt via
3D CFD-simulering och surrogatmodellering
Degree Programme in Mechanical Engineering

Date: June 24, 2021
Supervisor:
Mats Åbom, KTH
Mihai Mihaescu, KTH
Chenyang Weng, Volvo Car Group
Asuka Gabriele Pietroniro, Volvo Car Group
Examiner:
Mihai Mihaescu
Mats Åbom
School of Engineering Sciences
Host company: Volvo Car Group
Collaboration between VCG and the Competence Center for Gas
Exchange (CCGEx), KTH
Swedish title: Formoptimering av axiel kylningsfläkt via 3D
CFD-simulering och surrogatmodellering
© 2021 Martin Granlöf

|i
Abstract
Due to legislative reasons and environmental concerns the automotive and
transport sector are shifting their focus from traditional internal combustion
engine (ICE) vehicles to development of battery electric vehicles (BEVs). This
brings new challenges to design of cooling systems where axial fans are one of
the key components. Axial fans are usually designed with regards to a certain
operating condition and outside this region the efficiency of the fan drops
drastically. Due to difficulty in specifying the exact operational parameters
when placed in a car, post-design optimization may be necessary to ensure
maximized performance.
This thesis focuses on fan blade shape optimization through mesh morphing
using the surrogate based optimization algorithm called Efficient Global
Optimization (EGO). The target fan was a 9 bladed prototype fan by Johnson
Electric with uneven blade spacing. The optimization uses steady state
Reynolds-averaged Navier-Stokes (RANS) simulations to evaluate the fan
designs and a Bezier curve parametrization in order to change the fan blade
shape together with mesh morphing. The simulation setup was evaluated
before preceding with the optimization, and showed good agreement close to
intended operational conditions. Differences in turbulence modeling treatments
were also evaluated in order to have a satisfactory agreement with measurement
data.
The EGO algorithm manages to provide fan designs with higher total-to-
static efficiency at several different operational conditions. Evaluation of
the optimized fan designs was limited to comparison with the provided
measurement data and corresponding simulations.
Acoustic evaluation of selected fan designs is also attempted, but further work
is required in order for the study to result in a quantitative comparison.
Keywords— Fan noise, shape optimization, mesh morphing, computational

fluid dynamics, computational aeroacoustics, Efficient Global Optimization,
Reynolds-averaged Navier-Stokes.
ii |
Sammanfattning
På grund av lagstifting och miljöpåverkan har bil- och transportidustrin
börjat skifta fokus från traditionella förbränningsfordon till utveckling av
batteridrivna elbilar. Med detta medfäljer nya utmaningar kring kylsystemsdesign
där axiella fläktar är en av huvudkomponenterna hos systemet. Axiella fläktar
är vanligtvis designade kring ett specifikt driftstillstånd och utanför detta har
fläkten avsevärt lägre verkningsgrad. På grund av svårigheter att specificera
detta driftstillstånd med hög precision, speciellt när fläkten monteras i en bil,
kan efterdesigns-optimering vara nödvändigt för att uppnå maximal prestanda.
Denna avhandling fokuserar på form-optimering av fläkt blad via mesh morphing

med hjälp av den surrogat-baserade optimeringsalgoritmen Efficient Global
Optimization (EGO). Fläkten som optimerades var en prototypfläkt designad
av Johnson Electric med 9 fläktblad och icke-symmetriska mellanrum mellan
bladen. I optimeringsprocessen användes icke-tidsberoende Reynolds-averaged
Navier-Stokes (RANS) simuleringar för att utvärdera fläktdesignerna och
parametrisering med hjälp av Bezier kurvor och mesh morphing för att ändra
fläktbladen. Simulerings-uppställningen utvärderades innan optimeringen och
bra överensstämning nära avsett driftstillstånd kunde påvisas. Skillnader i
turbulens-modelering utvärderades även för att få en tillfredställande överensstämning
med mätdata.
EGO-algoritmen klarar att förse fläktdesigner med högre total-till-statisk
verkningsgrad vid flera olika driftstillstånd. Uvärdering av fläktdesignerna var
dock begränsad till jämförelse med mätdata och motsvarande simuleringsdata.
En akustisk utvärdering av utvalda fläkt-designer försöktes, men mer arbete
krävs för att studien ska erhålla en kvantitativ jämförelse.
Keywords— Fan noise, shape optimization, mesh morphing, computational

fluid dynamics, computational aeroacoustics, Efficient Global Optimization,
Reynolds-averaged Navier-Stokes.
| iii
Acknowledgments
I would like to thank my supervisors Chenyang Weng and Asuka Gabriele
Pietroniro at VCG. Chenyang Weng for his tenacity through the ups and downs
of this project, for sharing his knowledge and for always being positive and
encouraging. Asuka Gabriele Pietroniro for sharing his knowledge regarding
turbo machinery, unsteady RANS simulations and all things CFD.
I would like to thank Magnus Knutsson at VCG for proposing the thesis
subject, allowing me to work on it and for being a great enabler throughout
the project.
I would like to thank my supervisors at KTH Mihai Mihaescu and Mats Åbom
for their constructive criticism and guidance through out the project. I am
also very thankful for being included in the research meetings at KTH by
Mihai. Being able to partake and listen to the discussions was inspiring and
interesting.
I would like to thank Roberto Adorno and Paolo Cavallo from Johnson Electric
for providing the flow measurement data for the fan.
I would like to thank Emil Ljungskog and Martin Lindstroem from Simcenter
STAR-CCM+ for their help solving issues and answering questions throughout
the thesis work.
I would like to thank Brian Fechner from the thermal management team, Tore
Bark, Jacob Vikström, Karthik Narendra Babu, Dalibor Cuturic, Umut Cirik
from the thermal efficiency team, and all the team members from the GasEx,
Cool & Climate team at Volvo Cars for the discussion regarding the simulation.
Stockholm, June 2021

Martin Granlöf
iv |
CONTENTS | v
Contents
1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 Original problem and definition . . . . . . . . . . . . 4
1.2.2 Available methods for problem solving . . . . . . . . 4
1.3 Purpose and objectives . . . . . . . . . . . . . . . . . . . . . 7
1.4 Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.5 Structure of the report . . . . . . . . . . . . . . . . . . . . . . 7
2 Theory 9
2.1 Axial flow fans . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.1 Initial design process . . . . . . . . . . . . . . . . . . 11
2.1.2 Blade construction . . . . . . . . . . . . . . . . . . . 14
2.1.3 Fan performance metrics . . . . . . . . . . . . . . . . 15
2.2 Fan noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3 Efficient global optimization . . . . . . . . . . . . . . . . . . 19
3 Methodology 23
3.1 Simulation method . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 Mesh morphing . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.3 Optimization application . . . . . . . . . . . . . . . . . . . . 25
4 Simulation setup and testing 27

4.1 2D testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.1.1 NACA0012 . . . . . . . . . . . . . . . . . . . . . . . 27
4.1.2 Testing discussion . . . . . . . . . . . . . . . . . . . 31
4.2 Fan Simulation . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.2.1 Comparison with supplier data . . . . . . . . . . . . . 37
4.3 Fan Optimization setup . . . . . . . . . . . . . . . . . . . . . 41
4.4 Fan Acoustic setup . . . . . . . . . . . . . . . . . . . . . . . 43
vi | Contents
5 Results and discussion 45

5.1 ISO5801 Fan Optimization . . . . . . . . . . . . . . . . . . . 45
5.1.1 RP Mopt . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.1.2 ET Aopt1.1 . . . . . . . . . . . . . . . . . . . . . . . . 48
5.1.3 ET Aopt1.2 . . . . . . . . . . . . . . . . . . . . . . . . 49
5.1.4 ET Aopt2 . . . . . . . . . . . . . . . . . . . . . . . . 51
5.1.5 ET Aopt3 . . . . . . . . . . . . . . . . . . . . . . . . 53
5.1.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . 56
5.2 Acoustic evaluation . . . . . . . . . . . . . . . . . . . . . . . 57
6 Conclusions and Future work 61

6.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
6.2 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
6.3 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
References 65
A NACA0012 result plots 71

LIST OF FIGURES | vii
List of Figures
1.1 Axial cooling fan mounted to radiator. . . . . . . . . . . . . . 2

1.2 Performance chart for fan at different percentages of the
maximum fan rotational speed (solid lines), and operational
curves for different scenarios (dotted lines). The efficiency
(dashed line) is normalized with the maximum efficiency in
the range. Data reproduced from [3]. . . . . . . . . . . . . . . 3
2.1 Targeted fan. . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2.2 Fan test categories ISO5801. . . . . . . . . . . . . . . . . . . 11
2.3 Inflow schematic for fan blade section. . . . . . . . . . . . . . 14
2.4 Geometric definition of sweep and dihedral. . . . . . . . . . . 15
3.1 Overview of the EGO optimization process connected to

STAR-CCM+. . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.1 NACA0012 airfoil. (a), unperturbed shape and (b), an example

of perturbed shape. The “dots” in the figure are the point sets
used for the control of mesh morphing. . . . . . . . . . . . . . 29
4.2 EGO optimization of NACA0012 airfoil for setup 1. . . . . . . 30
4.3 Optimized airfoil from setup 1 with pink outline being the
baseline airfoil shape. . . . . . . . . . . . . . . . . . . . . . . 31
4.4 Local Cp for optimized setup 1 compared with baseline. . . . . 31
4.5 ISO5801 simulation domain. . . . . . . . . . . . . . . . . . . 33
4.6 Wall y + on the fan blades for the Baseline mesh in the
ISO5801 simulation setup at Qv = 1.01 m3 /s. Specifics of
the mesh can be seen in table 4.4. . . . . . . . . . . . . . . . . 34
4.7 Wall y + on the fan blades for the “Low y + 2” mesh in the
ISO5801 simulation setup at Qv = 1.01 m3 /s. Specifics of
the mesh can be seen in table 4.4. . . . . . . . . . . . . . . . . 35
viii | LIST OF FIGURES
4.8 Cross section of the fan region of "Baseline" mesh (a) showing
the blade surface mesh and (b) the prism layer mesh on the
blades. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.9 Cross section of the fan region of Low y + 2 mesh (a) showing
the blade surface mesh and (b) the prism layer mesh on the
blades. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.10 Relative difference in pressure rise in percent for the Baseline
mesh compared with measurement data. . . . . . . . . . . . . 38
4.11 Relative difference in pressure rise in percent for the “Low y +
2” mesh compared with measurement data. . . . . . . . . . . 38
4.12 Pressure rise for the “Low y + 2” mesh compared with measurement
data. Data on the vertical axis are hidden in the public version
of this report. . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.13 Relative difference in pressure rise in percent for the Baseline,
“Low y + ” and “Low y + 2” mesh compared with measurement
data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.14 Total to static efficiency for the Baseline, “Low y + ” and “Low
y + 2” mesh compared with measurement data. Simulation
data is calculated using equation (4.2) and measurement data
using equation (4.3). Data on the vertical axis are hidden in
the public version of this report. . . . . . . . . . . . . . . . . 40
4.15 Control surface and fan blade mesh with (a) being the un-
perturbed and (b) the perturbed control surface and mesh. . . . 42
4.16 Probe placement around the fan. . . . . . . . . . . . . . . . . 44
5.1 Performance of the designs evaluated in the RP Mopt optimization.

The baseline fan at the evaluated flow rate runs at 2914 RPM. . 47
5.2 Performance of the designs evaluated in the RP Mopt optimization.
The baseline fan at the evaluated flow rate has an efficiency
ηts = 0.5224. . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.3 Convergence curve and evaluated designs in the RP Mopt
optimization run. Designs within the red region are part of
the initial DOE. . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.4 Control surface and fan blade mesh with (a) being the baseline
design and (b) the optimized from the ET Aopt1.1 run. . . . . . 49
5.5 Performance of the designs in the ET Aopt1.2 optimization.
The baseline fan at the evaluated flow rate runs at 2914 RPM . 50
LIST OF FIGURES | ix
5.6 Performance of the designs in the ET Aopt1.2 optimization.

The baseline fan at the evaluated flow rate has an efficiency
ηts = 0.5224. . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.7 Convergence curve and evaluated designs in the ET Aopt1.2
the initial DOE. . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.8 Performance of the designs in the ET Aopt2 optimization. The
baseline fan at the evaluated flow rate runs at 2935 RPM. . . . 52
5.9 Performance of the designs that reach above ∆pts,target in the
ET Aopt2 optimization. The baseline fan at the evaluated flow
rate has an efficiency ηts = 0.4882. . . . . . . . . . . . . . . . 52
5.10 Convergence curve and evaluated designs in the ET Aopt2
the initial DOE. . . . . . . . . . . . . . . . . . . . . . . . . . 53
baseline fan at the evaluated flow rate runs at 2992 RPM . . . 54
baseline fan at the evaluated flow rate has an efficiency ηts =
0.2161. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.13 Convergence curve and evaluated designs in the ET Aopt3
the initial DOE. . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.14 Control surface and fan blade mesh with (a) being the baseline
design and (b) the optimized from the ET Aopt3 run. . . . . . . 56
5.15 Directivity of the OASPL for the baseline, ET Aopt1.1 and
ET Aopt1.2 fan designs. . . . . . . . . . . . . . . . . . . . . . 58
5.16 Single sided FFT of the acoustic pressure for the baseline,
ET Aopt1.1 and ET Aopt1.2 fan designs. Data on the vertical
axis are hidden in the public version of this report. . . . . . . . 58
5.17 Single sided FFT of the acoustic pressure at the first blade
passing frequency for the baseline, ET Aopt1.1 and ET Aopt1.2
fan designs. Data on the vertical axis are hidden in the public
version of this report. . . . . . . . . . . . . . . . . . . . . . . 59
5.18 Single sided FFT of the acoustic pressure at the second blade
passing frequency for the baseline, ET Aopt1.1 and ET Aopt1.2
fan designs. Data on the vertical axis are hidden in the public
version of this report. . . . . . . . . . . . . . . . . . . . . . . 59
A.1 EGO optimization of NACA0012 airfoil for setup 2 . . . . . . 71

x | LIST OF FIGURES
A.2 EGO optimization of NACA0012 airfoil for setup 3 . . . . . . 72

A.3 Optimized shape of the airfoil in setup 2 with original airfoil
in pink. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
A.4 Optimized shape of the airfoil in setup 3 with original airfoil
in pink. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
A.5 Local Cp for optimized setup 2 compared with baseline. . . . . 73
A.6 Local Cp for optimized setup 3 compared with baseline. . . . . 74
LIST OF TABLES | xi
List of Tables
4.1 Important flow quantities used in the optimization performed

using OpenFoam [41] and the one using STAR-CCM+. . . . . 28
4.2 Results from the optimized airfoils compared with the original
together with the number of input parameters and total design
evaluations. . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.3 Simulation physics and schemes for ISO5801 fan simulation. . 33
4.4 Cell count, thickness of the first cell on the surface of the fan
blades and growth rate of the prism layer for the three tested
meshes in the ISO5801 setup. . . . . . . . . . . . . . . . . . . 34
4.5 Total-to-static pressure rise at three different Qv for the three
meshes from full convergence simulations and sweep. . . . . . 41
5.1 Volume flow rate, control surface points, weighting factors

and input parameter intervals for the different optimization runs. 46
5.2 Summary of the performance of the best designs from the
different optimization runs, the Bezier curve inputs for these
designs and the number of the optimized design. . . . . . . . . 57
xii | List of acronyms and abbreviations
List of acronyms and abbreviations

ANN Artificial Neural Network
AoA Angle of Attack
BEV Battery Electric Vehicle
CAD Computer-Aided Design
CFD Computational Fluid Dynamics
DOE Design of Experiment
EGO Efficient Global Optimisation
EST Enhanced Stability Treatment
FFD Free-Form Deformation
FFT Fast Fourier Transform
FW-H Ffowcs Williams Hawkings
GA Genetic Algorithm
GBM Gradient-Based Methods
HM Hybridised method
ICE Internal Combustion Engine
KRG Kriging
LHS Latin Hypercube Sampling
MLP Multi-Layer Perception
MOEA Multi-Objective Evolutionary Algorithm
MOGA Multi-Objective Genetic Algorithm
NSGA Non-dominated Sorting Genetic Algorithm

List of acronyms and abbreviations | xiii
NURBS Non-Uniform Rational Basis Spline
OA Orthogonal Arrays
OASPL Overall Sound Pressure Level
PDE Partial Differential Equation
PSO Particle Swarm Optimisation
RANS Reynolds-averaged Navier-Stokes
RBF Radial Basis Function
RBNN Radial Basis Neural Network
RSA Response Surface Approximation
SA Simulated Annealing
SPL Sound Pressure Level
VOF Volume of Fluid

xiv | List of acronyms and abbreviations
Introduction | 1
Chapter 1
Introduction
In the year 2017 a motion containing a ban on newly produced combustion

engine cars by the year 2030 was brought up in Swedish parliament [1].
This can be seen as a direct response to the Paris Agreement, which entered
into force in 2016, where goals were set in order to limit global warming
effects. Since the domestic transport in Sweden and many other countries
is accountable for a large portion of the carbon dioxide emissions (around
32% of Carbon dioxide equivalents in Sweden), this seems like a reasonable
legislation initiative [2].
One way to fulfill the needs for transportation of goods and people is through
electrification of vehicles, either through battery electric vehicles (BEV)
or hybrid solutions containing standard internal combustion engine (ICE)
systems. These solutions present new challenges to manufacturers and designers
in many different areas.
One of these areas is cooling of BEV systems such as, electric motors,
circuitry and batteries. These cooling systems vary in complexity, but main
staple parts are usually heat exchangers and fans, which inevitably leads to
design challenges for the aforementioned components in these new operating
conditions. The focus in this report will be on fan optimization.
1.1 Background
The start of a fan design process is usually motivated by a need to fulfill some
system requirement. This is usually a requirement of a provided volumetric
flow at a given system resistance. If the fan is placed in a system with complex
flow phenomena, for example on top of a radiator as shown in figure 1.1, the
specification of these two system parameters can be difficult. Still the design
2 | Introduction
around these operational parameters is paramount in order to also maximize

fan efficiency. Most fans have a limited operational area where high efficiency
is guaranteed, and outside this range the efficiency drastically drops. An
example of a fan performance chart can be seen in figure 1.2, where the
pressure rise is normalized by the maximum pressure rise at maximum fan
speed, the efficiency by the maximum efficiency, and the volumetric flow by
the volumetric flow at 100% fan speed and zero pressure rise.
Figure 1.1 – Axial cooling fan mounted to radiator.

Introduction | 3
1.4
Normalized pressure rise and efficiency 1.2 38%

Fan efficiency@100% 58%
1
79%
0.8 100%
Stand-still
0.6
City
0.4
Highway
0.2
-0.2
-0.4
-0.6
0 0.2 0.4 0.6 0.8 1 1.2
Normalized volume flow
Figure 1.2 – Performance chart for fan at different percentages of the maximum
fan rotational speed (solid lines), and operational curves for different scenarios
(dotted lines). The efficiency (dashed line) is normalized with the maximum
efficiency in the range. Data reproduced from [3].
Apart from meeting criteria with regards to cooling and efficiency there is
also an ever-growing demand for less noisy fans, driven both by legislation
and customer satisfaction. This means that fans also need to be designed
and optimized with regards to acoustic performance, i.e. reduction of noise
at operation.
1.2 Problem
The problem with fan design relates to the aforementioned operational conditions
and requirements, but also the tools available. Firstly there is some difficulty in
accurately estimating the resistance of the system due to the complex geometry
of the parts involved and their interaction with one another. Secondly there are
many vastly different fan designs that can satisfy the same requirements, but
the best design can be hard to quantify. Depending on dimensions, shape of
the blades, and many other parameters, different design methods are also used,
which makes finding the best fan design even harder.
4 | Introduction
1.2.1 Original problem and definition

The problem investigated in this thesis is the optimization of cooling fan for
the charging system of a BEV. When a BEV is charging, especially during fast
charging, a lot of excess heat is built up in the batteries and related circuitry.
In order to dispose of this heat an efficient cooling system is needed. Due to
legislative and customer satisfactory reasons, the fan needs to both supply the
necessary volumetric flow at a high efficiency and produce as little noise as
possible.
The problem therefore boils down to aerodynamic and aeroacoustic optimization.
This optimization contains both accurately representing the fan geometry and
choosing an optimization method.
1.2.2 Available methods for problem solving

The methods of representing the fan geometry through parametrisation can be
split into three main categories: constructive, deformative and volume based
according to Skinner et al. [4].
Constructive parametrisation methods define basic body geometries. Examples
of these are spline methods, such as Bezier splines and basis splines, non-
uniform rational basis spline (NURBS), and partial differential equations
(PDEs) [4].
Deformative parametrisation methods use computational mesh points in order
to change the geometry. Examples of these are free-form deformation (FFD)
and radial basis function (RBF) utilization, which is sometimes used in mesh
morphing [4].
Volume based parameterization methods uses the volume fractions of computational
cells and conservation of volume in order to model the geometry. An example
of volume based parameterization is the volume of fluid (VOF) method,
which is more commonly used to model free surfaces in computational fluid
dynamics (CFD) simulations [5].
The optimization methods usually all start from a similar problem formulation,
which is described in Skinner et al. [4]. That is minimizing (or maximizing)
an objective function with regards to the design variables, which are provided
partly by the geometry parameterization. This function is then minimized
through an iterative process. Common method categories are:
• Gradient-based methods (GBM), which utilize derivatives of the

objective function with regards to the provided parameters. The derivative
Introduction | 5
information is then used in order change the parameters between iterations

to reach a function minimum [4].
• Genetic algorithms (GAs), that are based on evolutionary theory.

These methods are population-based and utilize mutation, cross-population
and selection phases that are iterated over generations in order to meet
convergence criterion. Different algorithms can be used for generation
of initial population, evaluation and selection, and reproduction [4].
• Particle swarm optimization (PSO), which is a stochastic population

based method, that emulates the behavior of herd animals or swarms.
This means that the evaluation of designs is carried out both with regards
to single individuals within the population, and with regards to the
population as a whole [4].
• Simulated annealing (SA) is a stochastic point based method, which

emulates the physical behavior of a cooling molten metal. The idea is
that at the start of optimisation, parameters are allowed to perturb quite
far from the current design, but as the process continues, the "cooling
effect" constrains the amount that parameters can be perturbed [6].
• Hybridised methods (HMs), are combinations of the previously mentioned

optimisation methods. Common hybridisation is GA combined with
GBM, and PSO combined with GBM [4].
Apart from the optimization methods mentioned above surrogate modelling is

becoming increasingly popular. Surrogate modelling methods are usually used
together with one of the listed optimization methods above when evaluation
of the objective function is expensive. Surrogate models try to construct a
correlation between input parameters and the objective function and output
this as a continuous function. The approximation is then used to guide the
optimization process [7].
These surrogates are sometimes paired with statistical optimization tools
in order to quantify uncertainties withing the objective function modelling to
ensure that minimums are not missed due to faulty initial modelling. This can
be especially helpful when few function evaluations are available due to high
computational cost as described by Jones et al. [8].
Apart from the vast amount of optimization done on aerodynamic bodies such
as wings and aeronautical hulls summarized by Skinner et al. [4], there is also
optimization focused on fans specifically.
6 | Introduction
Kim et al. [9] used a B-spline and Bezier curve combination for geometry
representation and manipulation in order to optimize aerodynamic and aeroacoustic
performance of an axial flow fan. They used a weighted average of a
response surface approximation (RSA) model, a Kriging (KRG) model and a
radial basis neural network (RBNN) model for the aerodynamic optimization,
and a combination of RSA, multi-objective evolutionary algorithm (MOEA)
and non-dominated sorting genetic algorithm (NSGA) was used for the
aeroacoustic optimization.
Öksüz et al. [10] used a multi-fidelity multi-objective genetic algorithm
(MOGA) to optimize blade shape of a low speed turbine with regards to torque
and total adiabatic efficiency. They used several Bezier curves to represent
and alter the blade geometry. With this numerical setup they outperformed
a standard MOGA both with regards to overall performance of the optimized
blade shape and computational cost.
Bamberger et al. [11][12] used multiple GAs to maximize the total to static
efficiency of low pressure axial fans given certain requirements for pressure
rise and volumetric flow. They use a multi-layer perception (MLP) artificial
neural network (ANN) that had been trained using data from RANS CFD
simulations to evaluate the performance of the fan designs. A total of 26
geometric parameters were included, varying from blade section parameters
such as chord length, thickness and position of maximum camber, to overall
geometric parameters such number of fan blades and hub-to-tip ratio. Further
CFD analysis and experimental evaluation of the optimized designs was also
carried out and showed good agreement.
Bamberger and Carolus [13] used and modifies an adjoint optimizer
algorithm available in the OpenSource CFD software OpenFoam to maximize
the total-to-static efficiency and pressure rise of a low pressure axial fan,
separately. Using the surface sensitivity of the fan blades to the objective
functions calculated through the adjoint solver, they change the blade CAD
using spline interpolation.
Cho et al. [14] used a GBM to optimize the efficiency of a low pressure
axial fan. They considered 12 design parameters for the fan rotor, such as
stagger, blade angle and sweep, and one for the stator that is part of the fan
shroud. The fan design was altered through the CAD and then re-meshed. The
results were also compared with experiments and showed good agreement at
design operating condition.
Apart from geometry parametrization and optimization scheme/method
the choice of simulation models within CFD softwares is also important for
the outcome of the optimisation.
Introduction | 7
1.3 Purpose and objectives

The purpose of this work is to optimise the fan that cools the battery system
during charging with regards to aerodynamic performance, operating efficiency,
and noise. Included in this is to extend Volvo’s current in-house shape
optimization framework so that it can provide fan design that has the optimal
performance for the actual requirements when placed in the car.
It should be emphasised, that the application of the current work is not to
design a fan from scratch, but to tune an existing fan so its performance is
optimum at a given operating point.
1.4 Scope
The work is limited to optimising a 9 bladed prototype fan made by Johnson
Electric using the Efficient Global Optimisation (EGO) surrogate-model-
based algorithm, outlined by Jones et al. [8]. A Python implementation of the
EGO algorithm by Bouhlel et al. [15] has been adopted in the optimization
framework. Bezier and sinusoidal curves/surfaces are used for geometry
parametrisation and mesh morphing is used to change the geometry. CFD
simulations are performed in Simcenter STAR-CCM+ [16].
1.5 Structure of the report

In chapter 2, theory for axial flow fans, fan noise generation and sources,
and the EGO surrogate model method will be presented. In chapter 3,
the simulation method and models used in STAR-CCM+ will be described
together with the details of the mesh morphing used and how it is carried
out in the software. Included in this chapter will be how the EGO model
is implemented and what it will look like for the specific case. In chapter
4, the ability of the EGO model to produce valid results, the fan simulation
setup and validation, and the fan optimization will be presented. In chapter 5,
results from the optimization using the simulation that emulates ISO5801 [17]
standard will be presented and discussed. In chapter 6 conclusions from the
work will be brought up. This will include limitations of the modeling, how
the results might be interpreted, and also suggestions for future work.
8 | Introduction
Theory | 9
Chapter 2
Theory
2.1 Axial flow fans

Axial flow fans can be divided into many different sub-categories but the
purpose of an axial flow fan is to create an airflow parallel to its axis of rotation.
The possible components and their function are listed below:
• Prerotator generate swirl, i.e. azimutal velocity about the fan axis, in
order to control the angle of incidence to the different sections of the fan
rotor.
• Nose and tail fairings aim to reduce the drag created by the hub of the
fan and also the losses due to separation behind the fan unit.
• Rotor is the part that generates the pressure difference that induces
the flow. The rotor can be made out of cambered blades with uniform
thickness or from blades with airfoil shapes.
• Straightener aim to remove the swirl from the flow behind the fan,
making it purely axial again.
Usually both prerotator and straightener are referred to as stators [18].

The fan in question for this thesis can be seen in figure 2.1. The fan has no
prerotator or straightener and negligible amount of nose and tail fairings. It
consists of 9 complex shaped and unevenly spaced fan blades.
10 | Theory
Figure 2.1 – Targeted fan.
The aerodynamic performance of an axial flow is usually evaluated according

to certain national or international standard. In this thesis we refer to ISO5801
[17] which is commonly used in Europe. The ISO5801 fan standard tests fans
according to five separate categories. These can be seen below in figure 2.2.
The categories are as follows:
• A: Free inlet and free outlet with partition
• B: Free inlet, ducted outlet
• C: Ducted inlet, free outlet
• D: Ducted inlet, ducted outlet
• E: Free inlet and free outlet without partition
The fan setup for simulations and experiments in this thesis are compliant
with Category A.
Theory | 11
Fan Fan Fan
A: Free inlet, free outlet B: Free inlet, ducted outlet C: Ducted inlet, free outlet
with partition
Fan Fan
D: Ducted inlet, ducted outlet E: Free inlet and free outlet without partition
Figure 2.2 – Fan test categories ISO5801.
2.1.1 Initial design process

As has been emphasized earlier, the application of the current project is not
initial fan design (i.e., designing a fan from scratch for a given operating point).
Even though, the initial design process is summarised in this section, because
it helps to understand the basic fan aerodynamics, e.g., how an axial fan can
provide pressure rise, etc..
Following the methodology outlined by Carolus and Starzmann [19] the
design of an axial fan can be split into the following four steps:
Step 1: Choosing fan diameter, hub diameter and fan speed
The first step of the design process is, as mentioned, to decide the fan diameter,
hub diameter and fan speed. The total diameter of the fan is usually decided
by the available space in the location that the fan will operate. The choice of
hub diameter is more free but can also be restricted by the size and placement
of the fan motor. The fan speed can be chosen more freely and changed during
the design process, but as suggested by Carolus and Starzmann [19] a Cordier
diagram can be used to guide this choice.
Step 2: Calculating flow rate and total pressure rise at the rotor
With the three quantities decided above the flow rate at the fan rotor and the
total pressure rise at the fan can be determined in the second step. These two
quantities are determined through the required system flow rate and the system
resistance at that the given flow rate [19].
Firstly, by assuming that the radial velocity is zero, the axial flow at
an annulus of the rotor can be calculated through continuity as seen in
12 | Theory
equation (2.1).
qv
us = up = (2.1)
2πrdr
where us and up are respectively the axial velocities at the suction and pressure
side of the rotor annulus at position r, qv is the flow rate at the annulus, and
dr is the thickness of the annulus at r.
Secondly, the theoretical total pressure rise of the fan at an annulus of
position r is given by Euler’s equation of turbomachinery
∆pth = ρΩr(uθ,s − uθ,p ) (2.2)
where Ω is the rotor rotational speed in radians per second, and uθ,s and uθ,p
are respectively the circumferential velocities at the suction and pressure side
of the rotor annulus at position r.
The non-dimensional form of this equation reads
2
hth = (s − p ) (2.3)
λ
with h = ∆p/ 12 ρU 2 being the non-dimensional pressure rise, where U is the
mean axial velocity through the fan, λ = U/Ωr the speed ratio, and = uθ /U
the swirl coefficient.
Step 3: Choosing load distribution and calculating velocity triangles in
order to obtain the total pressure rise at the given flow rate
Moving on to the third point, a load distribution along the span of the blades of
the fan rotor needs to be specified. There are several different design methods
for axial fans but two common design approaches are free vortex flow and
arbitrary, also called non-free, vortex flow. Free vortex flow assumes that the
swirl velocity through the fan is inversely proportional to the radial position,
this also has the benefit that the total pressure distribution and axial velocity
is constant with radial position. For arbitrary vortex flow the swirl velocity is
independent of the radial position.
A general expression for the spanwise distribution of the swirl velocity can
be written as
B
(2.4)
X
ruθ,p = ab r b
b=0
where ab is a constant and B = 0 would be give a flow distribution of a free

vortex flow loading. The axial velocities up and us can then be calculated
using a radial equilibrium condition given by Dixon [20], which assumes the
radial velocity to be zero. Using the total pressure and assuming the flow to be
Theory | 13
incompressible, the following equation for the radial equilibrium is obtained
∂up ∂ruθ,p uθ,p ∂ruθ,p

up =Ω − . (2.5)
∂r ∂r r ∂r
Note here that the constants ab in equation (2.4) needs to be specified in a
way that the axial and swirl velocity distributions satisfy the following global
relations for volumetric flow and pressure rise
Z rt
Qv = 2πrup (r)dr (2.6)
rh
Z rt
1
∆pth = ∆p(r)dqv . (2.7)
Qv rh
In the above equations Qv is the total volumetric flow rate, ∆pth is the total
theoretical pressure rise across the rotor, ∆p(r) is local total pressure rise, and
rh and rt are the hub radius and tip radius respectively.
Given the knowledge of the variables mentioned above the velocity triangles
can be calculated. The angle ϕr of the resulting velocity wr at a radial position
of a fan blade can be calculated as
λ
tan ϕr = 1 (2.8)
1− (
2 s
− p )λ
where the second term in the denominator is the average swirl coefficient
across the rotor.
Step 4: Choosing blade sections (airfoils) and calculating solidity
Given the inflow angle, ϕr , and the axial and swirl velocities together with the
rotational speed of the rotor the blade forces can be calculated in the fourth
step outlined above. The rotor normal and parallel forces for an annulus of
thickness dr, denoted as Y and X in figure 2.3, can be calculated as
Y = s∆pdr (2.9)
X = sρU 2 (s − p )dr (2.10)

where s is the gap between the fan blades.
Given this it is straightforward to calculate the drag and lift force, D and L,
using trigonometry, and also the drag and lift coefficients, CD and CL written
as
1
CD = hR sin3 ϕr (2.11)
σ
14 | Theory
2
CL = (s − p ) sin ϕr − CD cot ϕr (2.12)
σ
where σ = c/s is the solidity, and c is the camber length.
Through equations (2.9) to (2.12) we have a connection among the global
quantities, such as the pressure rise and volumetric flow, and the local blade
quantities such as lift, drag and solidity. With these equations one can design
the rotor blades using blade element theory outlined by, for example Wallis,
using the isolated airfoil method for solidities below 0.7 and the cascade
method above one [18]. The free choice of the designer is to choose the
optimum angle of attack, α of the blade section, whose relation to ϕr can
be seen in figure 2.3 and equation (2.13). This will provide the lift and drag
coefficients enabling the calculation for the solidity σ. The angle β in equation
(2.13) is the angle between the rotor plane and the chord line of the blade
section which can be seen in figure 2.3.
L
Ch Y ϕr
ord
lin
α e
β wr X
ϕr
D
Figure 2.3 – Inflow schematic for fan blade section.
β = ϕr + α (2.13)
2.1.2 Blade construction

Apart from the choice of blade shape and stagger angle along the blade the
blade sweep and dihedral are also blade design aspects to be considered. A
simple geometric definition can be seen in figure 2.4. Worth noting is the
illustration of sweep and dihedral in figure 2.4 are of a simple form, just linear
displacement along the blade proportional to the radial position. Other more
Theory | 15
advanced forms are also possible.
No sweep No dihedral
Backward Forward
sweep sweep - dihedral + dihedral
Inflow
Hub Hub
Figure 2.4 – Geometric definition of sweep and dihedral.
A detailed review of the effects of sweep and dihedral on axial fans and
compressors can be found in the article by Vad [21]. Although the effects
of these parameters are complex and depend on the operational conditions of
the fan they can be generalized a bit. Well designed forward sweep increases
efficiency at both intended loading and at partial load. This seems to be mainly
due to the decrease in low energy flow accumulation near the blade tip that is
a result of radial flow along the fan blades. Lower discharge loss and noise
generation have also been observed [22]. Note that forward sweep unloads
the tip region and moves the load towards inboard blade sections. Depending
on initial loading, this region may experience higher losses due to this load
redistribution. Backward sweep seems to decrease performance in almost all
operating conditions and is therefore undesirable [21].
Positive dihedral at the hub and tip seems to be beneficial to unload these
regions where low energy fluid can accumulate. Similar to forward sweep, this
moves the load to the mid-span part of the fan blades and therefore additional
evaluation might be needed when this is used [23].
2.1.3 Fan performance metrics

Apart from fulfilling the requirements for volumetric flow rate at a given
system resistance the efficiency of the fan is also important. When evaluating
the efficiency of a fan the total-to-total or the total-to-static is usually used. The
choice between these often depends on the importance of the exhaust flow of
the fan. If the fan is in a closed loop for example, the total-to-total efficiency
16 | Theory
could be useful. Since the fan in this case is not in a loop the total-to static
efficiency will be used. Apart from these two distinctions of efficiency, the
total input power used in the efficiency formulation can vary. Common choices
for the input power are shaft power at the fan, electrical input power to the fan
motor and total mechanical power. In this report a similar expression to that
used by Kim et. al [9] will be used, which utilizes the mechanical power at the
fan blades as the total input power. This is written as
∆pts Qv
ηts = , (2.14)
Ωτ
where ∆pts = pp − (ps + us ) is the total-to-static pressure difference between
the suction side and the pressure side of the fan, Qv is the volumetric flow
through the fan, Ω is the angular speed of the fan and τ is the torque on the fan
blades in the axial direction.
In the book by William K. Blake [24] the possibility of a more efficient fan
also having better acoustic performance is discussed. If this is the case then
optimizing towards efficiency could also optimize acoustics.
2.2 Fan noise

The noise generated by flow phenomenon is usually referred to as aeroacoustics.
The aeroacoustic noise generation originates from small fluctuations in flow
quantities and can be formulated as deviations from an equilibrium state
as shown in equation (2.15). The zero sub-scripted quantities refer to the
equilibrium state conditions of pressure p, density ρ and velocity ui , while
the primed quantities refer to the acoustic contribution.
p(x, t) = p0 (x) + p0 (x, t)

ρ(x, t) = ρ0 (x) + ρ0 (x, t) (2.15)
ui (x, t) = u0,i (x) + u0i (x, t)
From this one can insert the quantities of equation (2.15) into the mass and
momentum equations and linearize. Assuming that the acoustic propagation
is isentropic one now arrives at the inhomogeneous wave equation, given in
equation (2.16). The first term on the right hand side is the contribution to
the acoustic field due to an unsteady external force, the second term is due to
Theory | 17
unsteady mass injection and c0 is the adiabatic speed of sound [25].
1 ∂ 2 p 0 ∂ 2 p0 ∂fi ∂Qm
2 2
− 2
= + (2.16)
c0 ∂t ∂xi ∂xi ∂t
By splitting space into two parts, the source part and the sound part, where the
source field satisfies the inhomogeneous wave equation (2.16) and the sound
field the homogeneous wave equation we arrive at Lighthill’s aeroacoustic
analogy. Making the same assumptions [25] as for the inhomogeneous wave
equation and using the mass and momentum equations without linearization,
one gets the following for the source field
1 ∂ 2 p0 ∂ 2 p0 ∂ 1 ∂ 0 2 0
∂f
i ∂2
− = Qm + (p − c 0 ρ ) − + (ρui uj − τij ).
c20 ∂t2 ∂x2i ∂t c20 ∂t ∂xi ∂xi ∂xj
| {z }| {z } | {z }
s1 s2 s2
(2.17)
In equation (2.17), which is the famous Lighthill’s equation, the three terms on
the right hand side are usually referred to as, from left to right, the monopole,
the dipole and the quadrupole source term, labeled as s1 , s2 and s3 . The
monopole source s1 field is due to unsteady mass injection and deviations
form adiabatic change of state, the dipole s2 source field is due to an unsteady
external force, and the quadrupole source field s3 is due to the viscous stresses
τij and the Reynolds stresses ρui uj . The quadrupole source field is usually
summarized to be the noise generated by turbulence since the viscous stresses
mostly have a dampening effect on the noise production. The dipole source
term is the one that is responsible for the tonal noise related to moving
fan blades and the quadrupole is responsible for the more broadband noise
generated by the fan.
The two equations above are valid for situations with zero mean flow outside
the source region, i.e. u0,i = 0. For most applications involving fans the mean
flow Mach number is small, around 0.1, and the effects of the flow on the
sound propagation can be neglected. A common method used for acoustic
calculations for propellers and fans is the Ffowcs Williams Hawkings FW-H
equation [26]. This utilizes the Heaviside function in order to write the mass
and momentum equations in a way that is valid for the whole fluid domain
18 | Theory
including the propeller or fan. The resulting equation is then
1 ∂ 2 p0 ∂ 2 p0 ∂2
− = (Tij H)
c20 ∂t2 ∂x2i ∂xi ∂xj
∂ ∂H
+ (ρ0 vn + ρ(un − vn )) (2.18)
∂t ∂xn
∂
0 ∂H
− (ρui (un − vn ) + p δij nˆj ) ,
∂xi ∂xn
with H being the Heaviside function, being zero inside the rotating body and
one in the fluid domain, vn being the normal surface velocity of the body,
un the normal velocity of the fluid and nˆj the outwards facing normal of the
surface. The variable inside the first term parenthesis is the Lighthill stress
tensor Tij = p0 δij + ρui uj − c0 ρ0 δij . Often the assumption of an impermeable
surface is used, i.e. un = vn , further neglecting the first term on the right hand
side of equation (2.18), which is a quadrupole term, reduces the equation to
the following form
1 ∂ 2 p0 ∂ 2 p0 ∂ ∂H ∂ 0 ∂H
− = ρ0 vn − p δij nˆj , (2.19)
c20 ∂t2 ∂x2i ∂t ∂xn ∂xi ∂xn
where the right hand side is only non-zero on the surface of the moving body
due to the gradient of the Heaviside function ∂H/∂xn [27]. The removed
quadrupole term can also be modeled instead of directly calculated and then
added back to equation (2.19) in order to capture the contribution. Looking at
the right hand side of equation (2.19) the first term is referred to as thickness
noise and is generated by the fluid displacement of the moving body. The
second term is called loading noise and is related to the pressure distribution
on the propeller or fan blades [25]. Similarly for the second and third term of
equation (2.18), but without the assumption of an impermeable surface.
In CFD software such as STAR-CCM+ the integral form of the FW-H equation
shown in equation (2.18) is used. This is written as
p0 = p0T + p0L + p0Q (2.20)

Theory | 19
where p0T and p0L are the contributions to the acoustic pressure due to thickness
and loading, as discussed earlier, and are written as

 
Zρ0 U̇n + Uṅ
4πp0T =   dS
r (1 − Mr )2
 ret
 (2.21)
Z ρ0 Un rṀr + c0 (Mr − M 2 )
+   dS,
r2 (1 − Mr )3
ret
Z " #
Lr − Li Mi
Z
1 L̇r
4πp0L = dS + dS
c0 r(1 − Mr )3 r2 (1 − Mr )2 ret
ret

Z Lr rṀr + c0 (Mr − M 2 )
 (2.22)
1  dS.
+ 
c0 r2 (1 − Mr )3
ret
These integrals are evaluated at the surface of the FW-H surface with
Ui = (1 − (ρ/ρ0 )) + (ρui /ρ0 ), Li = p0 δij n̂j + ρui (un − vn ) and Mi = vi /c0 .
Subscripts r and n indicates that the variable is taken in the radiation direction
and surface normal direction respectively, doted variables are differentiated
with regards to source time, and subscript ret means that the quantities are
evaluated at retarded or source time. The formulation in equation (2.22) is
known as Farassat 1A or just formulation 1A and is given by Farassat et al
[28]. In this the quadropole term is left out because of the low Mach number
[25], but formulations of the term can be found by, for example, Brentner [29].
The acoustic pressure fluctuations are commonly expressed using sound
pressure level SPL as shown in equation (2.23), where p0ref = 2 · 10−5 Pa.
p0 2
SP L = 10 log10 ( ) (2.23)
p0ref
This gives the acoustic pressure fluctuations in dB.
2.3 Efficient global optimization

To optimize the blade shape of a given fan design, so as to increase the fan’s
efficiency and hopefully decrease its noise, the efficient global optimization
20 | Theory
(EGO) is employed in this study as the optimization algorithm.

EGO is a surrogate optimization method based on the Kriging modeling
technique developed by Jones et al. [8]. Kriging uses statistical methods and
the assumption that the function of interest can be seen as a stochastic process.
In its simplest form the Kriging modeling technique approximates a function,
f (χ) with the following model:
F = µ + ε(χ) (2.24)
where µ is the average of this the aforementioned stochastic process, χ herein
refers to the design parameter(s) of the optimization, and ε(χ) is the error,
which is assumed to belong to a normal distribution with mean 0 and variance
σ 2 (χ) [30]. Unlike linear regression, which assumes ε(χ) to be independent
between two points in design space, EGO assumes correlation between ε(χ)
to be proportional to the distance between points. The distance in this case
is weighted with regards to the influence a parameter has on the objective
function [8]. This correlation method is usually referred to as the DACE
method, from the name of the paper by Sacks et al. [31]. The parameters of
the DACE model are fitted with the initial data using a maximum likelihood
estimate.
After an initial design population, so called design of experiment (DOE), has
been selected, using for example orthogonal arrays (OA) or Latin hypercube
sampling (LHS), the exploration and exploitation of the design space given by
the parameter limits is evaluated using an expected improvement estimate [7].
The algorithm seeks to find the point that maximizes equation (2.25)
fmin − µ(χ) fmin − µ(χ)

EI(χ) = (fmin − µ(χ))Φ + σ(χ)ϕ ,
σ(χ) σ(χ)
(2.25)
where fmin is the best solution that has been found and Φ(χ) and ϕ(χ)
are the cumulative and probability density function of the standard normal
distribution respectively [30]. The new design point is then added to the data
set and the parameters of the DACE model are recalculated and the new data
is cross-validated [8].
The EGO algorithm was initial designed for unconstrained optimization, but
can be extended to handle inequality constraints as done by Schonlau et al.
[32]. This is done by extending the expected improvement estimate shown in
equation (2.25) using a penalty function for the feasibility of the given design.
Theory | 21
This is written as
m
(2.26)
Y
EIp (χ) = EI(χ)F (χ) = EI(χ) P (gi (χ) < 0)
i=1
where F (χ) is the feasibility function and P (gi (χ) < 0) is the probability
of gi (χ) to be feasible, quantified by the Kriging modeling of the ith constraint.
This penalized expected improvement function can be further alternated in
order to fit the problem at hand. This is done by Bagheri et al. [30] in order to
improve performance when handling more than two constraints.
22 | Theory
Methodology | 23
Chapter 3
Methodology
3.1 Simulation method

The simulations performed in Simcenter STAR-CCM+ [16] utilizes the Reynolds-
averaged Navier-Stokes (RANS) equations for solving the flow field, shown
below in equation (3.1) written with tensor notation. The continuity equation
is shown in (3.2).
∂Ui ∂Ui 1 ∂P ∂ h ∂U ∂Uj 2 i

(3.1)
i
+ Uj =− + (ν + νt ) + − kδij
∂t ∂xj ρ ∂xi ∂xj ∂xj ∂xi 3
∂ρ ∂ρUi
+ =0 (3.2)
∂t ∂xi
Here Ui and P are the ensemble-averaged velocity and pressure, respectively.
They are represented by lower-case letters, ui and p, in the rest of the thesis.
The turbulence behavior is modeled through the eddy viscosity νt often
involving the turbulent kinetic energy k. Most commercial CFD codes include
turbulence modeling through transport equations for different turbulence
quantities. Some of the common ones are the standard k- model with
constants from the article by Launder and Spalding [33], the realizable k-
model by Shih et al. [34] and the low-Reynolds number version by Launder
and Sharma [35].
Other common turbulence models are the k-ω model, using the specific
turbulence dissipation ω instead of . Common versions of the k-ω model are
the standard k-ω model by Wilcox [36] and the SST k-ω model by Menter [37].
Both of these are two equation models, i.e. they use two equations in order to
24 | Methodology
quantify νt . The Spalart-Allmaras one equation model is also a turbulence

model that is commonly found in commercial codes [38].
In this report the focus is mainly on the realizable k- model and the SST k-
ω model. The reasoning being that both these perform reasonably in turbo
machinery applications and rotating flows, where we have high mean shear
rate and separation phenomena [34][37]. For the realizable k- model a two
layer approach is also applied described by Chen and Patel [39]. This allows
for a different mixing length approach close to walls within the computational
domain. The scope of the report does not include any comprehensive
comparison between turbulence models in the given simulation setup, but a
quantitative reasoning will be put forth for the choice of turbulence model in
a given situation.
Most of the simulation and optimization process is performed using steady
simulations. For the time dependent simulations an implicit time stepping
method available in Simcenter STAR-CCM+ is used. The time step itself
is chosen so that it represents roughly one degree of rotation, which is
recommended by the user guide in STAR-CCM+ and also used by previous
Volvo Cars master thesis student Omar Fares [40].
3.2 Mesh morphing

Simcenter STAR-CCM+ version 2020.3 offers a few different alternatives for
mesh morphing, such as preset meshing motions using auto-generated RBF
parameters or Bsplines. Control points within the mesh and their displacement
can be specified from tables, auto generated or derived from surfaces [16].
The alternative used in this report is the RBF with control points, where both
the original position and the displacement of the points are imported from
CSV files through tables. The control points and displacements are placed
into point sets within STAR-CCM+ and the displacement setting used is the
total displacement. The recompute interfaces setting is also used in order to
have a better mesh close to the morphed surface. Apart from these changes
to the settings the default ones for the mesh morpher and morpher motion are
used.
The position and displacement interval of the control points are decided using
measurements within STAR-CCM+ and trial and error.
Methodology | 25
3.3 Optimization application

The application of the EGO algorithm in conjunction with the CFD software
STAR-CCM+ is done in several steps and can be seen summarized in figure
3.1. First the user chooses the input parameters and their bounds together with
the objective function. This information is then used to generate the DOE
and the designs are evaluated in STAR-CCM+ in order to make a connection
between the input parameters and the objective function. This evaluation can
be ran in parallel (multiple simulations simultaneously) or in series (one at a
time).
This data is then used to train the surrogate model and the next design is chosen
using the expected improvement function as discussed in Section 2.3. The new
design is evaluated and the surrogate model is retrained. This is repeated until
a certain threshold or an amount of design evaluation is reached.
The brain of the optimization is the open-source surrogate modeling
toolbox code, written in python by Bouhlel et al.[15]. This code together with
the geometry parameterization of the fan blade geometry used for the mesh
morphing is all situated in python scripts. The communication between these
scripts and STAR-CCM+ and the automation of mesh morphing and parameter
changes is done through Java macros and python servers.
User inputs
• Upper and lower bounds of design parameters
• Definition of objective function
Job #1
Design of experiment (DOE)
Train
Job #2 EGO
. Surrogate model
Simulation job
Update
.
manager Expected improvement
.
New design parameters
Job #N
Optimum design
Figure 3.1 – Overview of the EGO optimization process connected to STAR-

CCM+.
26 | Methodology
Simulation setup and testing | 27
Chapter 4
Simulation setup and testing
4.1 2D testing
In order to get a better understanding of how the optimization algorithm
behaves, several 2D simulations were performed on airfoil sections. Since
the blades of a fan are usually made from airfoil shapes, the simulation results
might also be extrapolated to the fan simulations.
The 2D simulations presented below are based on flow parameters and airfoil
shapes investigated in the tutorial on DaFoam’s web page [41].
4.1.1 NACA0012
This simulation is based on the NACA0012 airfoil and the parameters presented
for the incompressible NACA0012 airfoil tutorial on the web page of DAFoam
[41]. Since the original simulation is performed in the CFD software OpenFoam,
the default fluid parameters are slightly different, but the chord Reynolds
number, Rec has been matched. In the tutorial simulation the aim is to
reduce the Drag coefficient, CD while keeping the lift coefficient, CL fixed
at 0.375. To do this they use an adjoint optimization method with 40 free
form deformation (FFD) points, with freedom to move in one of the coordinate
axis, and the angle of attack (AoA) as input parameters for the algorithm. The
starting angle of attack is 3.579107°, and then this is free to change in order
to keep CL constant. The important quantities can be seen in table 4.1, with c
being the chord length of the airfoil.
28 | Simulation setup and testing
Table 4.1 – Important flow quantities used in the optimization performed using
OpenFoam [41] and the one using STAR-CCM+.
OpenFoam STAR-CCM+
c [m] 1 1
Uin [m/s] 35 36
ρ [kg/m3 ] 1 1.18415
Rec 2.3e6 2.3e6
The simulation in STAR-CCM+ uses the Kriging based EGO algorithm

and the coordinates of two point sets as an input for two separate Bezier curves,
one located above the upper surface and one below the lower surface of the
airfoil. Three different surface control setups were tested. The first setup uses
one control point for the upper line segment and one for the lower that is free to
move between 5% and 90% of the cord and is free to perturb the surface in the
local wall normal direction. The second setup uses two control points for both
the upper and lower line segment, where one of them is free to move between
5% and 40% of the chord and the other between 40% and 90%. The points are
free to perturb the surface in the wall normal direction. The third case uses
two points on the upper line segment and one on the lower. An example of the
perturbed and the unperturbed airfoil can be seen in figure 4.1.
All cases use the control points and the angle of attack as the input for the
EGO algorithm, with the objective to reduce the drag while keeping the lift at
the baseline value or above. This is done through a penalty method shown in
equation (4.1)
Fobj = νp (min(0, CL − CL,b ))2 + CD (4.1)

where νp is a penalty coefficient, usually a large value, and CL,b is the
baseline lift coefficient. For the simulations presented in this section νp = 10.
(a) Unperturbed
(b) Perturbed
Figure 4.1 – NACA0012 airfoil. (a), unperturbed shape and (b), an example of
perturbed shape. The “dots” in the figure are the point sets used for the control
of mesh morphing.
Using the suggestion of 10 times the design dimensions for the initial
design amount, proposed by Jones et al. [8], the initial designs are selected
using the Latin hypercube method. In this case each control point has two
degrees of spatial freedom and the angle of attack has one degree of freedom.
Setup one therefore has 50 initial designs, setup two has 90 and setup three has
70. The number of designs investigated after the initial designs was manually
set to 30.
The CD , CL , angle of attack α and the ratio between lift and drag for
the original and the optimized airfoils can be seen below in table 4.2, the
performance of the design samples in setup 1 can be seen in figure 4.2, with
the leading edge of the airfoils corresponding to the data point. The doted
line represents the CL value of the baseline airfoil. The optimized shape of
the airfoil in setup 1 can be seen compared with the original airfoil in figure
4.3 and the local pressure coefficient Cp can be seen in figure 4.4 with points
labeled opt being the optimized airfoil shape. Similar plots shown in figure
4.2, 4.3 and 4.4 can be found for setup 2 and 3 in Appendix A.
Table 4.2 – Results from the optimized airfoils compared with the original
together with the number of input parameters and total design evaluations.
Original Setup 1 Setup 2 Setup 3

CD 0.011236 0.01092246 0.01101979 0.01116956
CL 0.38395 0.38725109 0.39135579 0.3861878
CL /CD 34.17 35.45 35.51 34.59
α 3.58 3.60 2.84 2.07
# inputs - 5 9 7
# evaluated designs - 80 120 100
0.7
3.50
0.6 3.25
0.5 3.00
0.4 2.75
AoA
CL
2.50
0.3
2.25
0.2
2.00
0.1
1.75
0.0 1.50
0.0095 0.0100 0.0105 0.0110 0.0115 0.0120 0.0125 0.0130 0.0135 0.0140
CD
Figure 4.2 – EGO optimization of NACA0012 airfoil for setup 1.
Figure 4.3 – Optimized airfoil from setup 1 with pink outline being the baseline
airfoil shape.
1.5
Suction side opt
Pressure side opt
1.0 Suction side base
Pressure side base
0.5
Cp
0.0
0.5
1.0
0.0 0.2 0.4 0.6 0.8 1.0
x /c
Figure 4.4 – Local Cp for optimized setup 1 compared with baseline.
4.1.2 Testing discussion

For the NACA0012 airfoil there is an improvement in the drag coefficient of
2.8% and 1.9% for setup 1 and 2 respectively, and 0.59% for setup 3. All the
setups achieved a higher lift coefficient than the baseline simulation too, even
though this was not the main objective of the simulation. This could be due the
way the penalty oriented objective function is designed and most importantly
the penalty coefficient νp .
Looking at these results in the context of fan performance through Euler’s
law of turbomachinery and equation (2.12), assuming other quantities to be
similar in between the blade designs, the improved lift and drag coefficients
would equate to around 2% higher pressure rise at the same flow rate.
The effect of relative size difference between the lift and coefficient could
influence the optimization as discussed by Marler et al. [42] and therefore
the choice of objective function is paramount for the outcome of the process.
Therefore the penalty coefficient νp should both account for the relative
difference between the included parameters and the importance to fulfill the
lift constraint.
4.2 Fan Simulation

The simulation setup for the fan was constructed to resemble the ISO5801
Category A test as much as possible, in order to compare with data supplied
by the manufacturer. The simulation domain, sliced through the middle, can
be seen below in figure 4.5. The boundary condition for the walls marked
with red is the one called Walls in Star-CCM+, which is a no-slip and no-
penetration condition. For the upper and lower wall on the suction side this is
modified to a slip wall condition. The inlet, marked as green, uses a velocity
inlet boundary condition with the direction of the velocity being normal to the
inlet surface. The outlet, marked as blue, is a pressure outlet with a boundary
normal backflow specification and the environmental pressure specified at the
surface. The region containing the fan, in the middle of the domain, uses
a moving reference frame with the same wall boundary conditions as for
the walls dividing the suction side and pressure side marked with red. To
reduce computational cost, parts that do not influence the fan performance
significantly, such as the motor and struts, are excluded in the simulation.
Pressure side
Suction side
Figure 4.5 – ISO5801 simulation domain.
The choice of physics and schemes for the ISO5801 fan simulations can
be found below in table 4.3. The setup is a three dimensional steady state
simulation using air at constant density and viscosity as the fluid with the
segregated flow solver. Three turbulence modeling treatments were evaluated,
all using the all y + treatment in Star-CCM+. These treatments were the
realizable k − , SST k − ω and the SST k − ω with enhanced stability control
EST.
Table 4.3 – Simulation physics and schemes for ISO5801 fan simulation.
Space Three dimensional

Time Steady
Fluid Gas (air)
Fluid solver Segregated flow
Equations of state Constant density, constant viscosity
Viscous regime Turbulent
Turbulence modeling Realizable k − / SST k − ω / SST k − ω EST
Wall treatment All y + wall treatment
Spatial discretization Second order upwind

Pressure-velocity coupling SIMPLE
Three meshes where evaluated using the setup described above. These
meshes only differ inside the fan region (the moving reference frame) and some
overall mesh quantities are can be found in table 4.4. The mesh called Baseline
and the one called “Low y + ” have the same fan blade surface resolution and
the mesh called “Low y + 2” has a lower surface resolution compared with
these two.
Table 4.4 – Cell count, thickness of the first cell on the surface of the fan blades
and growth rate of the prism layer for the three tested meshes in the ISO5801
setup.
Cell count Cell thickness NW Growth Rate

Baseline 42.2E6 1E-4 m 1.73
Low y + 82.6E6 2.5E-6 m 1.48
Low y + 2 44.9E6 2.5E-6 m 1.48
The setup were run close to intended operating conditions until convergence
to check the wall y + at the fan blades. The resulting wall y + can be seen for
the Baseline mesh and the “Low y + 2” mesh in figure 4.6 and 4.7 respectively.
The Baseline mesh has a wall y + significantly higher than one and is therefore
not compliant with the recommendations of the SST k − ω turbulence model,
while the “Low y + ” and “Low y + 2” mesh has a y + of one or lower. Figure
4.8 shows the surface mesh on the blades and prism layer mesh of the Baseline
mesh and figure 4.9 shows the same for the “Low y + 2” mesh.
Figure 4.6 – Wall y + on the fan blades for the Baseline mesh in the ISO5801
simulation setup at Qv = 1.01 m3 /s. Specifics of the mesh can be seen in
table 4.4.
Figure 4.7 – Wall y + on the fan blades for the “Low y + 2” mesh in the ISO5801
simulation setup at Qv = 1.01 m3 /s. Specifics of the mesh can be seen in table
4.4.
(a) Surface mesh
(b) Prism layer mesh
Figure 4.8 – Cross section of the fan region of "Baseline" mesh (a) showing
the blade surface mesh and (b) the prism layer mesh on the blades.
(a) Surface mesh
(b) Prism layer mesh
Figure 4.9 – Cross section of the fan region of Low y + 2 mesh (a) showing the
blade surface mesh and (b) the prism layer mesh on the blades.
4.2.1 Comparison with supplier data

The simulation setup outlined above was compared to supplier while evaluating
the three meshes and the three turbulence modeling treatments. For the
comparison 13 operating points were chosen from the supplier data. These
operating points were run in a so called sweep, using an asymptotic stopping
criterion on the total to static pressure rise across the fan and a maximum
iteration count of 5000 for each point. This meant that if the simulation gave
less than 1% change in pressure rise over 1500 iterations it could move on to
the next point in the sweep. The relative difference in pressure rise, compared
with the measurement data provided by the supplier, for the Baseline mesh
and for the “Low y + 2” mesh can be seen in figure 4.10 and 4.11 respectively.
Note here that the sweep is carried out going from high volume flow rate to
low. The actual pressure data from the sweep ran on the “Low y + 2” mesh can
be seen compared with the measurement data in figure 4.12
2.0
Simulation Sweep k −ω
Simulation Sweep k− ω EST
1.5
Simulation Sweep k−
Rel. diff. Pressure rise
1.0
0.5
0.0
0.5
1.0
0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00
Volume Flow [m3 /s ]
Figure 4.10 – Relative difference in pressure rise in percent for the Baseline
mesh compared with measurement data.
2.0
Simulation Sweep k −ω
1.5
Simulation Sweep k−
1.0
0.5
0.0
0.5
1.0
0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00
Figure 4.11 – Relative difference in pressure rise in percent for the “Low y +
2” mesh compared with measurement data.
Pressure rise [Pa]
JE measurement 50 Pa
Simulation Sweep k− ω
Simulation Sweep k −
0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00
Figure 4.12 – Pressure rise for the “Low y + 2” mesh compared with
measurement data. Data on the vertical axis are hidden in the public version
of this report.
The flow rate interval that is most crucial is that close to Qv = 1m3 /s and
this area shows a discrepancy of less than 6% for the Baseline mesh and less
than 5% for the “Low y + 2” mesh, compared to the measurements.
Since simulations on the "Low y + mesh" are considerably more computationally
expensive, a sweep was only performed using the SST k − ω with EST. A
comparison between all meshes using this turbulence model can be seen in
figure 4.13. What can be noted is that the Baseline mesh under-predicts the
pressure rise, for the most part, compared to the “Low y + ” and the “Low y + 2”
mesh.
2.0
Simulation Sweep k− ω EST Baseline
Simulation Sweep k− ω EST Low y +
1.5 Simulation Sweep k− ω EST Low y+ 2
1.0
0.5
0.0
0.5
1.0
0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00
Figure 4.13 – Relative difference in pressure rise in percent for the Baseline,
“Low y + ” and “Low y + 2” mesh compared with measurement data.
The total to static efficiency from each of the evaluated points during the
sweep can be seen below in figure 4.14. The efficiency in the simulation is
calculated using equation (4.2) and the efficiency from the fan experimental
data is calculated using equation (4.3). The difference between these two
equation is that the simulation uses the fan angular speed Ω times the fan blade
torque τ as the total work per unit time done by the fan, and the experiment
uses the electrical power input for the fan motor, i.e., the voltage U times the
current I. This means that the experimental efficiency accounts for both fan
motor losses and bearing losses, which should explain the slight offset in the
peak efficiency shown in figure 4.14. Similar behavior is presented also by
Cho et al. [14].
Qv ∆pts
ηts,sim = (4.2)
Ωτ
Qv ∆pts
ηts,exp = (4.3)
UI
JE measurement
Simulation Sweep k− ω EST Baseline
Simulation Sweep k− ω EST Low y +
Simulation Sweep k− ω EST Low y + 2
ts
η
0.2
0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00

Volume Flow [m3 /s]
Figure 4.14 – Total to static efficiency for the Baseline, “Low y + ” and “Low
y + 2” mesh compared with measurement data. Simulation data is calculated
using equation (4.2) and measurement data using equation (4.3). Data on the
vertical axis are hidden in the public version of this report.
Three out of the 13 evaluated points from the sweep were also ran to full
convergence, with regards to pressure rise, to check the accuracy of the sweep
data. These points were all ran using the SST k − ω with EST and the results
can be seen in table 4.5. Close to intended operation, at Qv = 1.01m3 /s, all
the meshes are within 1% of the fully converged results
Table 4.5 – Total-to-static pressure rise at three different Qv for the three
meshes from full convergence simulations and sweep.
Qv [m3 /s] ∆pts Full Conv. [Pa] ∆pts Sweep [Pa]

Baseline 0.81 309.39 310.97
1.01 260.44 261.81
1.60 62.20 62.33
Low y + 0.81 307.18 306.40
1.01 262.13 264.57
1.60 65.59 65.31
Low y + 2 0.81 306.60 308.71
1.01 259.11 260.25
1.60 63.95 64.06
Given the following investigations regarding the turbulence models and

meshes it was decided that the “Low y + 2” mesh and the SST k − ω with EST
would be used for further simulations. This decision was based on accuracy
with regards to the measurement data, stability, where the SST k − ω with
EST performed better than the one without, and computational cost. The SST
k − ω should be better suited to adverse pressure gradient flows than the k-
turbulence model, which further solidifies the decision.
4.3 Fan Optimization setup

The fan optimization setup uses the “Low y + 2” mesh, the SST k − ω with
EST and simulation setup described in Section 4.2. The fan blade geometry
is controlled using a Bezier curve in the chord wise or azimutal direction,
similarly to that shown in Section 4.1, and a sinusoidal function with amplitude
equal to one in the span wise or radial direction. The span wise shaping of the
control surface is such that the peak perturbation is at the mid span of the fan
blade and is then attenuated towards the tip and the hub of the fan. Each blade
has two control surfaces, one on the suction side and one on the pressure side
of the fan. An example of the un-perturbed and perturbed control surfaces can
be seen below in figure 4.15
(a) Un-perturbed
(b) Perturbed
Figure 4.15 – Control surface and fan blade mesh with (a) being the un-
perturbed and (b) the perturbed control surface and mesh.
Two objective functions will mainly be considered, one targeting the total-
to-static efficiency ηts and the other the fan RPM. These objective function are
written as
Fobj = ν1 (min(0, ∆pts − ∆pts,target ))2 − ν2 ηts (4.4)
Fobj = ν1 (min(0, ∆pts − ∆pts,target ))2 + ν2 RP M, (4.5)

where ν1 and ν2 are weighting coefficients and ∆pts,target is the total-to-
static pressure difference for the baseline fan design at the given volume flow
rate and RPM. Through the structure of the objective functions we aim to
punish designs that do not reach the baseline total-to-static pressure rise while
increasing efficiency and reducing RPM.
As seen in the 2D testing in Section 4.1 the required design evaluations
increase with the number of input parameters. Therefore a setup with one
free control point in the Bezier curve in the cord wise direction for each of the
control surfaces will be used. This, together with the RPM, constitutes 5 input
parameters in total, since the control point has two degrees of freedom. This
means that the initial DOE will be 50 designs and then 30 more designs will
be evaluated as was done for the NACA0012 optimization.
Each of the designs are run for 7000 iterations in order to ensure decent
stability and convergence. Promising designs are run until full convergence
and will presented as such.
4.4 Fan Acoustic setup

The acoustic simulation setup uses the “Low y + 2” mesh with some clean
up done between prism layers on neighboring surfaces, the SST k − ω with
EST and a similar setup to that presented in Section 4.2. The change to the
setup was mainly to make all the wall surfaces marked with red in figure 4.5
slip walls, and changing to an ideal gas treatment with Sutherland’s law for
viscosity instead of the constant density used for the steady state simulations.
These changes were made due to fix stability issues seen on the suction side
of the domain.
Apart from these changes the switch to an implicit unsteady time stepping
solver and addition of an FW-H equation solver was also made. The time
stepping solver uses a first order temporal discretization and, as mentioned in
Section 3.1, a time step equivalent to the time it takes for the fan to rotate one
degree. Before this the simulation is run using the steady state solver and then a
larger time step in order to initialize the flow field and reduce simulation time.
For the FW-H solver the hub, fan blades and ring connecting the blades are
used as the impermeable FW-H surface and the quadrupole term is neglected
in the evaluation. The partition splitting the inlet and outlet region and the side
walls of the inlet chamber are not included in the FW-H surface even though
these surfaces scatter the acoustic field. This could bring some error to the data
compared to measurements made in a similar setup, but since all simulation
cases are treated equally they should still be comparable. Receiver probes are
placed on a circle at a 2.5m distance from the fan origin with 5 degrees in
between and uses the Farassat 1A integral formulation of the FW-H presented
in Section 2.2. Figure 4.16 shows the placement of the acoustic probes around
the fan.
Figure 4.16 – Probe placement around the fan.

Results and discussion | 45
Chapter 5
Results and discussion
5.1 ISO5801 Fan Optimization

One optimization was run using the RPM focused objective function shown
in equation (4.5) and four were run using the efficiency focused objective
function shown in equation (4.4). Three different operating points were
investigated, where the volume flow rate was the fixed quantity. These flow
rates were Qv = 0.81, 1.01 and 1.6 m/s, which are also the cases ran until full
convergence for the baseline fan presented in table 4.5 in Section 4.2.
The volume flow rate, input parameter ranges and the control surface details for
all the optimization runs can be seen in table 5.1. RP Mopt is the optimization
using the RPM focused objective function and ET Aopt1.1 to ET Aopt3 all use
the efficiency based objective function. CSP Span wise and CSP Chord wise
are the points in the control surface in the span wise and chord wise direction
respectively. The quantity npert is the movement of the control point normal
to the control surface in meters and tpert is the movement of the control point
tangentially to the control surface, i.e. in the chord wise direction, where 0 is
the trailing edge of the fan blade and 1 is the leading edge. The maximum
RPM for each of the optimization runs are equal to the RPM that the baseline
fan is evaluated at for the specified volume flow rate Qv .
46 | Results and discussion
Table 5.1 – Volume flow rate, control surface points, weighting factors and
input parameter intervals for the different optimization runs.
RP Mopt ET Aopt1.1/1.2 ET Aopt2 ET Aopt3

Qv 1.01 1.01 0.81 1.60
CSP Span wise 11 34 34 34
CSP Chord wise 11 11 11 11
ν1 1000 1 1 1
ν2 1 100 100 100
min npert -0.01 -0.01 -0.01 -0.01
max npert 0.03 0.03 0.03 0.03
min tpert 0.1 0.1 0.1 0.1
max tpert 0.85 0.85 0.85 0.85
min RPM 2500 2700 2800 2700
max RPM 2914 2914 2935 2992
5.1.1 RP Mopt
The results from the RP Mopt optimization can be seen below in figures 5.1
to 5.3. The best design with regards to efficiency and pressure rise was run
to full convergence and has a total-to-static pressure rise ∆pts = 257.81 Pa
and efficiency ηts = 0.53904 with the same RPM as the baseline design at the
evaluated volume flow rate. This corresponds to 0.5% lower pressure rise, but
a 3.2% higher total-to static efficiency than the baseline. The design itself is
pushed inwards close to the leading edge on the suction side and similarly on
the pressure side but at the leading edge.
Figure 5.1 – Performance of the designs evaluated in the RP Mopt

optimization. The baseline fan at the evaluated flow rate runs at 2914 RPM.
Figure 5.2 – Performance of the designs evaluated in the RP Mopt

optimization. The baseline fan at the evaluated flow rate has an efficiency
ηts = 0.5224.
250000 Simulation samples
EGO convergence curve
200000
150000
Fobj
100000
50000
0
0 10 20 30 40 50 60 70 80
Sample No.
(a) Full
3200
3100
3000
Fobj
2900
2800
Simulation samples
2700
0 10 20 30 40 50 60 70 80
Sample No.
(b) Close-up
Figure 5.3 – Convergence curve and evaluated designs in the RP Mopt

optimization run. Designs within the red region are part of the initial DOE.
5.1.2 ET Aopt1.1
The optimization data for this run was sadly lost in a backup error, but luckily
the best performing design was saved and therefore will be presented. This
design has a pressure rise ∆pts = 255.124 Pa and an efficiency ηts = 0.5337
with the same 2914 RPM as the baseline design at the evaluated volume
flow rate. This corresponds to 1.54% lower pressure rise and 2.16% higher
efficiency than the baseline fan design. The optimized design can be seen
below in figure 5.4, where one can note that the fan blade is pushed down close
to the trailing edge on the suction side and pushed up close to the leading edge
on the pressure side.
(a) Baseline
(b) Optimized
Figure 5.4 – Control surface and fan blade mesh with (a) being the baseline
design and (b) the optimized from the ET Aopt1.1 run.
5.1.3 ET Aopt1.2
The results from the ET Aopt1.2 optimization can be seen below in figure 5.5
to 5.7. The best design in this case had a pressure rise ∆pts = 256.96 Pa
and an efficiency ηts = 0.53411 with the same RPM as the baseline design
at the evaluated volume flow rate. This corresponds to 0.83% lower pressure
rise and 2.24% higher efficiency than the baseline fan design. This design is
almost unchanged on the suction side and on the pressure side the mid chord
section is slightly pushed outwards.
Target pts
300 Optimization data
250
pts
200
150
2700 2750 2800 2850 2900 2950 3000

RPM
Figure 5.5 – Performance of the designs in the ET Aopt1.2 optimization. The
baseline fan at the evaluated flow rate runs at 2914 RPM
Target pts
250
pts
200
150
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8

ts
Figure 5.6 – Performance of the designs in the ET Aopt1.2 optimization. The

baseline fan at the evaluated flow rate has an efficiency ηts = 0.5224.
15000
Fobj
10000
5000
0
0 20 40 60 80
Sample No.
(a) Full
20
20
Fobj
40
60
Simulation samples
80
0 20 40 60 80
Sample No.
(b) Close-up
Figure 5.7 – Convergence curve and evaluated designs in the ET Aopt1.2

5.1.4 ET Aopt2
The results from the ET Aopt2 can be seen below in figure 5.8 to 5.10. The best
design had a pressure rise ∆pts = 303.2 Pa and an efficiency ηts = 0.4952
with the same RPM as the baseline design at the evaluated volume flow rate.
This corresponds to 1.1% lower pressure rise and 1.4% higher efficiency than
the baseline fan. The optimized design shares some of the same features to
that of the ET Aopt1.1 fan. The suction side surface is pushed inwards close to
the trailing edge and the pressure side surface is pushed inwards close to the
leading edge.
Note here that this part of the operational map was quite hard to evaluate and
if more time was available then more iterations per design evaluation would
be allocated. This could also explain why the optimum design is not that much
better than the baseline.
Target pts
300
pts
250
200
150
2700 2750 2800 2850 2900 2950 3000
RPM
Figure 5.8 – Performance of the designs in the ET Aopt2 optimization. The
baseline fan at the evaluated flow rate runs at 2935 RPM.
Target pts
300
pts
250
200
150
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
ts
Figure 5.9 – Performance of the designs that reach above ∆pts,target in the
ET Aopt2 optimization. The baseline fan at the evaluated flow rate has an
efficiency ηts = 0.4882.
Simulation samples
25000 EGO convergence curve
20000
15000
Fobj
10000
5000
0
0 10 20 30 40 50 60 70 80
Sample No.
(a) Full
20
25
30
35
Fobj
40
45
50
60
0 10 20 30 40 50 60 70 80
Sample No.
(b) Close-up
Figure 5.10 – Convergence curve and evaluated designs in the ET Aopt2

5.1.5 ET Aopt3
The results from the ET Aopt3 optimization run can be seen below in figure
5.11 to 5.13. the best design had a pressure rise ∆pts = 98.14 and an efficiency
ηts = 0.2450 with RPM=2986.2 at the evaluated volume flow rate. This is
54.5% higher pressure rise and 13.4% higher efficiency than the baseline at
a lower RPM. The optimized design seems to have increased the camber of
the baseline design by pushing the suction side surface outward close to mid
chord and the pressure side surface inwards around closer to the trailing edge
of the blade. The design can be seen in figure 5.14.
100 Target pts

Optimization data
75
50
25
pts
0
25
50
75
100
2700 2750 2800 2850 2900 2950 3000
RPM
baseline fan at the evaluated flow rate runs at 2992 RPM
100 Target pts

Optimization data
75
50
25
pts
0
25
50
75
100
0.75 0.50 0.25 0.00 0.25 0.50 0.75
ts

baseline fan at the evaluated flow rate has an efficiency ηts = 0.2161.
20000
15000
Fobj
10000
5000
0
0 10 20 30 40 50 60 70 80
Sample No.
(a) Full
20
10
0
Fobj
10
20
40
0 10 20 30 40 50 60 70 80
Sample No.
(b) Close-up
Figure 5.13 – Convergence curve and evaluated designs in the ET Aopt3

(a) Baseline
(b) Optimized
Figure 5.14 – Control surface and fan blade mesh with (a) being the baseline
design and (b) the optimized from the ET Aopt3 run.
5.1.6 Summary
In table 5.2 the best performing designs from each of the optimization runs
are shown. Total-to-static pressure rise and efficiency together with the input
for the Bezier curve which modifies the blade shape are shown. The negative
value of npert means that the control surface and the blade surface are pushed
inwards. The number of the optimized design is also shown, which indicates
where in the process it was found. The abbreviations "SS" and "PS" stand for
suction side and pressure side respectively.
Table 5.2 – Summary of the performance of the best designs from the different
optimization runs, the Bezier curve inputs for these designs and the number of
the optimized design.
RP Mopt ET Aopt1.1 ET Aopt1.2 ET Aopt2 ET Aopt3

RPM 2914 2914 2914 2935 2986.2
∆pts 257.81 255.124 256.96 303.2 98.14
ηts 0.53904 0.5337 0.53411 0.4952 0.2450
SS npert -0.01 - -0.000479 -0.00257 0.00517
SS tpert 0.85 - 0.446 0.850 0.423
PS npert -0.01 - 0.649 -0.000927 -0.00925
PS tpert 0.1 - 0.00225 0.139 0.847
design number 52 82 63 59 56
As can be noted in table 5.2, the design from the ET Aopt1.1 has a number
higher than the 80 design evaluations mentioned in Section 4.3. The maximum
allowed design evaluations was increased to 90 for ET Aopt1.1 and ET Aopt1.2
since it proved harder to find a good design, using this setup, than previously
thought. The ET Aopt3 run ended up at 82 evaluated designs in total due to the
way it was parallelized. The exact data for the ET Aopt1.1 best design is left
out since the data from the optimization was lost, as previously mentioned.
Note also that the best designs are not usually the designs shown in the EGO
convergence plots for the optimization runs shown above. This could also
indicate that the objective functions are not the optimized for the goal, that is
increasing the efficiency.
5.2 Acoustic evaluation

The optimized designs from the ET Aopt1.1 and ET Aopt1.2 runs were evaluated
using the simulation setup described in Section 4.4. The results can be seen
in figure. In figure 5.15 the directivity of the overall sound pressure level
(OASPL) can be seen, where 0 is on the axis of the fan at the pressure side
and 180 is on the fan axis at the suction side. In figure 5.16 to 5.18 the fast
Fourier transform FFT of the acoustic pressure can be seen for with zoom in
views at the first and second blade passing frequencies.
180
Base
150 210
Opt1.1
Opt1.2
120 240
90 40 20 270
80 60
60 300
30 330
0
Figure 5.15 – Directivity of the OASPL for the baseline, ET Aopt1.1 and
ET Aopt1.2 fan designs.
Baseline
Opt1.1
Single-sided FFT of pressure (dB)
Opt1.2
10 dB
2 3
10 10
Frequency (Hz)
Figure 5.16 – Single sided FFT of the acoustic pressure for the baseline,
ET Aopt1.1 and ET Aopt1.2 fan designs. Data on the vertical axis are hidden
in the public version of this report.
Baseline
Opt1.1

Opt1.2
2 dB
435 435.5 436 436.5 437 437.5 438 438.5 439 439.5
Frequency (Hz)
Figure 5.17 – Single sided FFT of the acoustic pressure at the first blade
passing frequency for the baseline, ET Aopt1.1 and ET Aopt1.2 fan designs. Data
on the vertical axis are hidden in the public version of this report.
Baseline
Opt1.1
Opt1.2
1 dB
970 970.5 971 971.5 972

Frequency (Hz)
Figure 5.18 – Single sided FFT of the acoustic pressure at the second blade
passing frequency for the baseline, ET Aopt1.1 and ET Aopt1.2 fan designs. Data
on the vertical axis are hidden in the public version of this report.
Note here that the frequency resolution is quite coarse, around 4.86 Hz,
which is the first blade passing frequency multiplied by a factor of 1/90. This
is due to lack of time and also some stability issues with the setup. The flow
quantities are converging within the time step, but the residuals are slowly
diverging. Therefore a shorter sampling time was selected to have some data
that can be used for this comparison.
Using this data one can conclude that the ET Aopt1.1 design seems to be
slightly better than the baseline and the ET Aopt1.2 design in the OASPL and
also at the first and second blade passing frequencies. Note that unevenly
spaced fans shift some of the tonal noise to fractions of the blade passing
frequencies, as discussed by Peng et al.[43]. This can also be seen in figure
5.16 where there is a substantial peak at 2/9, around 97 Hz, of the blade passing
frequency.
The lack of time and stability issues is also the reason why the promising
design from ET Aopt3 was not simulated. Looking at scaling laws for the
pressure, RPM and sound power level from Harris [44] there could be a
possible reduction in RPM in-order to bring the pressure rise down to the same
level as the baseline fan design. This would then equate to a reduction in the
sound power level equal to roughly 10 · log10 (1.55/2 ) ≈ 4.4 dB.
Conclusions and Future work | 61
Chapter 6
Conclusions and Future work
6.1 Conclusions
The EGO can provide an aerodynamically improved fan blade design at the
intended volume flow rate, i.e. a design with higher total-to-static efficiency
ηts . It can also be concluded that the toolbox created for manipulating the
fan blade geometry is capable of changing the original design in a suitable
way for post design optimization. It is highly versatile both in adhering to the
original geometry and creating new ones with correct user input. It is also
rather straightforward to add functionality to the toolbox if one would like to.
Considering the results from the different objective function use, for the case
with Qv = 1.01 m3 /s one can conclude that the efficiency based objective
functions is not constructed in the best way. The RPM based objective function
does not intend to increase efficiency directly, but it still manages to provide a
better design than the efficiency-based optimization.
For the ET Aopt3 case that is run at Qv = 1.60 m3 /s the performance gain is a
lot higher than for the lower volume flow rates. This is most likely due to the
fact that the fan was designed for lower flow rates.
The two fan designs that were evaluated acoustically and compared with the
baseline seem to perform quite similarly to the baseline fan. This is most likely
due to them not being significantly improved compared to the baseline. To
conclude that improving efficiency also improves acoustic performance the
design from ET Aopt3 should be investigated and compared with the baseline.
62 | Conclusions and Future work
6.2 Limitations
The limiting factor with optimization and simulations is almost always the
time, and that was the case for this work as well. That is also the reasoning
behind the choice of using steady RANS simulations for the optimization
simulations. A large eddy simulation or a direct numerical simulation would
most likely give more accurate results, being setup correctly, the computational
cost, however, is also a lot higher. An unsteady RANS could also have
been used for the optimization, but since we were interested in averaged
quantities for efficiency and pressure rise this would also be a waste of
computational resources since this is also more computationally demanding.
The time consumption for one design evaluation for the acoustic evaluation
simulations is more than 30 times that of the steady state simulation used in
the optimization.
Another aspect that may have limited the maximum performance of the
optimized design is that the focus of the morphing was placed on the mid span
section of the fan blades. This was a decision based on the literature available
and also because it is the furthest away from any other surface, simplifying the
mesh morphing.
The mesh used for the optimization simulations could also be improved if
more time was available. It was made early in the project and with knowledge
gathered throughout the process it could be better when it comes to the prism
layers connecting to the blade surfaces.
The optimization and the acoustic evaluations were not carried out in exactly
the same simulation setups. Therefore there is some ambiguity how to
interpret the results and connections between the efficiency and the acoustic
performance. The setup itself is not completely stable and is slowly diverging.
The evaluation is carried out with a quite coarse frequency resolution.
6.3 Future work

One thing that could be done in the future is to use the simulation setup that
was used for the unsteady RANS simulation and re-run the optimization to see
if there are any differences. This mesh should be better structured given that
it was made later in the project with more experience in 3D meshing.
Further development of the fan blade control surfaces used for mesh morphing
could be done in order to have non-fixed parametrization in the span-wise
direction. This could possibly help identify regions of the fan blade that
Conclusions and Future work | 63
effect performance the most. Investigating objective function formulations and

tuning weighting factors could also be added into this in order to maximize the
potential performance gain.
Implementing a non-uniform inlet velocity and running the optimization under
those conditions could also be done. This could provide useful designs for
more realistic operational conditions when the fan is placed in situ, since inlet
conditions are usually not uniform.
Implementing acoustic evaluation into the optimization would also be very
good. For this to be possible a lot more computational power would be needed.
The simulation time for the unsteady RANS used to evaluate the acoustic
performance compared to the steady RANS used for the design evaluation
within the optimization is more than 30 times higher.
More work needs to be done on the unsteady RANS simulation as well in order
to have a reliable comparison of the fan designs. Once a stable setup is created
it could also be used for the steady state simulations, so that one has continuity
throughout the optimization and evaluation.
Apart from these suggestions it would also be interesting to see how the
optimization works for other baseline fan designs. The shape of the fan
optimized in this thesis was quite complex and it might be possible to get more
improvements from a less complex baseline design.
64 | Conclusions and Future work
REFERENCES | 65
References
[1] Sveriges Riksdag. (2017) Förbud mot försäljning av nya

fossildrivna bilar år 2030. Accessed: 2021-01-21. [Online].
Available: https://www.riksdagen.se/sv/dokument-lagar/dokument/
motion/forbud-mot-forsaljning-av-nya-fossildrivna-bilar_H5022106t
[2] Statistiska Centralbyrån. (2020) Totala utsläpp växthusgaser, år 1990-

2018. Accessed: 2021-01-27. [Online]. Available: https://www.scb.se/
hitta-statistik/sverige-i-siffror/miljo/utslapp-av-vaxthusgaser/
[3] M. Karlsson and S. Etemad, “Installation effects on the flow generated

noise from automotive electrical cooling fans,” SAE Technical Paper,
Tech. Rep., 2020.
[4] S. N. Skinner and H. Zare-Behtash, “State-of-the-art in aerodynamic

shape optimisation methods,” Applied Soft Computing, vol. 62, pp. 933–
962, 2018.
[5] J. Hall, T. Rendall, C. Allen, and D. Poole, “A volumetric geometry and

topology parameterisation for fluids-based optimisation,” Computers &
Fluids, vol. 148, pp. 137–156, 2017.
[6] S. Kirkpatrick, C. D. Gelatt, and M. P. Vecchi, “Optimization by

simulated annealing,” science, vol. 220, no. 4598, pp. 671–680, 1983.
[7] N. V. Queipo, R. T. Haftka, W. Shyy, T. Goel, R. Vaidyanathan, and

P. K. Tucker, “Surrogate-based analysis and optimization,” Progress in
aerospace sciences, vol. 41, no. 1, pp. 1–28, 2005.
[8] D. R. Jones, M. Schonlau, and W. J. Welch, “Efficient global

optimization of expensive black-box functions,” Journal of Global
optimization, vol. 13, no. 4, pp. 455–492, 1998.
66 | REFERENCES
[9] J.-H. Kim, B. Ovgor, K.-H. Cha, J.-H. Kim, S. Lee, and K.-Y. Kim,
“Optimization of the aerodynamic and aeroacoustic performance of an
axial-flow fan,” AIAA Journal, vol. 52, no. 9, pp. 2032–2044, 2014.
[10] Ö. Öksüz and İ. S. Akmandor, “Multi-objective aerodynamic

optimization of axial turbine blades using a novel multilevel genetic
algorithm,” Journal of turbomachinery, vol. 132, no. 4, 2010.
[11] K. Bamberger and T. Carolus, “Design guidelines for low pressure

axial fans based on cfd-trained meta-models,” in Proc. European
Turbomachinery Conference, vol. 11, 2015.
[12] K. Bamberger, T. Carolus, and M. Haas, “Optimization of low-pressure

axial fans and effect of subsequent geometrical modifications,” in
Proceedings of the Fan 2015 Conference, 2015.
[13] K. Bamberger and T. Carolus, “Aerodynamic optimization of axial fans

using the adjoint method,” in Turbo Expo: Power for Land, Sea, and
Air, vol. 50985. American Society of Mechanical Engineers, 2018, p.
V001T09A014.
[14] C.-h. Cho, S.-y. Cho, and C. Kim, “Development of an axial-type

fan with an optimization method,” Frontiers of Energy and Power
Engineering in China, vol. 3, no. 4, pp. 414–422, 2009.
[15] M. A. Bouhlel, J. T. Hwang, N. Bartoli, R. Lafage, J. Morlier, and J. R.

Martins, “A python surrogate modeling framework with derivatives,”
Advances in Engineering Software, vol. 135, p. 102662, 2019.
[16] Siemens PLM Software, “STAR-CCM+ 2020.3,” 2020.
[17] “ISO 5801. Fans – Performance testing using standardized airways,”

EUROPEAN COMMITTEE FOR STANDARDIZATION, EN ISO
5801:2017 E.
[18] A.R. Wallis, Axial flow fans: design and practise. Academic press,
1961. ISBN 1-4832-7422-5
[19] T. H. Carolus and R. Starzmann, “An aerodynamic design methodology

for low pressure axial fans with integrated airfoil polar prediction,” in
Turbo Expo: Power for Land, Sea, and Air, vol. 54648, 2011, pp. 335–
342.
REFERENCES | 67
[20] S.L. Dixon, Fluid mechanics and thermodynamics of turbomachinery,

7th ed. Butterworth-Heinemann, 2014. ISBN 0-12-391410-8
[21] J. Vad, “Aerodynamic effects of blade sweep and skew in low-speed

axial flow rotors at the design flow rate: an overview,” Proceedings of
the Institution of Mechanical Engineers, Part A: Journal of Power and
Energy, vol. 222, no. 1, pp. 69–85, 2008.
[22] M. B. Wilkinson, “The design of an axial flow fan for air-cooled heat
exchanger applications.” Ph.D. dissertation, 2017.
[23] L. H. Smith Jr and H. Yeh, “Sweep and dihedral effects in axial-flow

turbomachinery,” Journal of fluids engineering, vol. 85, pp. 401–414,
1963.
[24] W. K. Blake, Mechanics of flow-induced sound and vibration, Volume 2:

Complex flow-structure interactions. Academic press, 2017.
[25] M. Åbom, An introduction to flow acoustics. Skolan för

teknikvetenskap, Kungliga Tekniska högskolan, 2006.
[26] J. Ffowcs-Williams and D. Hawkings, “Sound generated by turbulence

and surfaces in aroitrary motion,” Philosophical Transactions of the
Royal Society (London) A, vol. 264, pp. 321–342, 1969.
[27] K. S. Brentner and F. Farassat, “Analytical comparison of the acoustic

analogy and kirchhoff formulation for moving surfaces,” AIAA journal,
vol. 36, no. 8, pp. 1379–1386, 1998.
[28] F. Farassat and G. P. Succi, “The prediction of helicopter rotor discrete

frequency noise,” In: American Helicopter Society, pp. 497–507, 1982.
[29] K. S. Brentner, “An efficient and robust method for predicting helicopter
high-speed impulsive noise,” Journal of Sound and Vibration, vol. 203,
no. 1, pp. 87–100, 1997.
[30] S. Bagheri, W. Konen, R. Allmendinger, J. Branke, K. Deb, J. Fieldsend,

D. Quagliarella, and K. Sindhya, “Constraint handling in efficient
global optimization,” in Proceedings of the Genetic and Evolutionary
Computation Conference, 2017, pp. 673–680.
[31] J. Sacks, W. J. Welch, T. J. Mitchell, and H. P. Wynn, “Design and

analysis of computer experiments,” Statistical science, pp. 409–423,
1989.
68 | REFERENCES
[32] M. Schonlau, W. J. Welch, and D. R. Jones, “Global versus local

search in constrained optimization of computer models,” Lecture Notes-
Monograph Series, pp. 11–25, 1998.
[33] B.R. Launder and D.B. Sharma, “The numerical computation of

turbulent flows,” Computer methods in applied mechanics and
engineering, vol. 3, pp. 269–289, 1974.
[34] T.-H. Shih, W. W. Liou, A. Shabbir, Z. Yang, and J. Zhu, “A new k- eddy
viscosity model for high reynolds number turbulent flows,” Computers
& fluids, vol. 24, no. 3, pp. 227–238, 1995.
[35] B. E. Launder and B. I. Sharma, “Application of the energy-dissipation

model of turbulence to the calculation of flow near a spinning disc,”
Letters in heat and mass transfer, vol. 1, no. 2, pp. 131–137, 1974.
[36] D. C. Wilcox, “Reassessment of the scale-determining equation for

advanced turbulence models,” AIAA journal, vol. 26, no. 11, pp. 1299–
1310, 1988.
[37] F. R. Menter, “Influence of freestream values on k-omega turbulence

model predictions,” AIAA journal, vol. 30, no. 6, pp. 1657–1659, 1992.
[38] P. Spalart and S. Allmaras, “A one-equation turbulence model for

aerodynamic flows,” in 30th aerospace sciences meeting and exhibit,
1992, p. 439.
[39] H. Chen and V. Patel, “Near-wall turbulence models for complex flows
including separation,” AIAA journal, vol. 26, no. 6, pp. 641–648, 1988.
[40] O. Fares, “Numerical investigation of noise generation by automotive

cooling fans,” 2019.
[41] DAFoam. (2019) Naca0012 airfoil incompressible. Accessed: 2021-

02-08. [Online]. Available: https://dafoam.readthedocs.io/en/latest/
Tutorial_Aerodynamics_NACA0012_Incompressible.html
[42] R. T. Marler and J. S. Arora, “The weighted sum method for multi-
objective optimization: new insights,” Structural and multidisciplinary
optimization, vol. 41, no. 6, pp. 853–862, 2010.
[43] Z. Peng, H. Ouyang, Y. Wu, and J. Tian, “Tonal noise control of cooling
fan module by using modulation principles on both rotor and stator,” in
REFERENCES | 69
Turbo Expo: Power for Land, Sea, and Air, vol. 50985. American
Society of Mechanical Engineers, 2018, p. V001T09A007.
[44] C. M. Harris and G. C. Tocci, “Handbook of acoustical measurements

and noise control,” Acoustical Society of America Journal, vol. 93, no. 6,
p. 3532, 1993.
70 | REFERENCES
Appendix A: NACA0012 result plots | 71
Appendix A
NACA0012 result plots
0.7
3.50
0.6 3.25
0.5 3.00
0.4 2.75
AoA
CL
2.50
0.3
2.25
0.2
2.00
0.1
1.75
0.0 1.50
0.0095 0.0100 0.0105 0.0110 0.0115 0.0120 0.0125 0.0130 0.0135 0.0140
CD
Figure A.1 – EGO optimization of NACA0012 airfoil for setup 2
72 | Appendix A: NACA0012 result plots
0.7
3.50
0.6 3.25
0.5 3.00
0.4 2.75
AoA
CL
2.50
0.3
2.25
0.2
2.00
0.1
1.75
0.0 1.50
0.0095 0.0100 0.0105 0.0110 0.0115 0.0120 0.0125 0.0130 0.0135 0.0140
CD
Figure A.2 – EGO optimization of NACA0012 airfoil for setup 3
Figure A.3 – Optimized shape of the airfoil in setup 2 with original airfoil in
pink.
Appendix A: NACA0012 result plots | 73
Figure A.4 – Optimized shape of the airfoil in setup 3 with original airfoil in
pink.
Suction side opt

1.0 Pressure side opt
Suction side base
Pressure side base
0.5
Cp
0.0
0.5
1.0
0.0 0.2 0.4 0.6 0.8 1.0
x /c
Figure A.5 – Local Cp for optimized setup 2 compared with baseline.

74 | Appendix A: NACA0012 result plots
Suction side opt

1.0 Pressure side opt
Suction side base
Pressure side base
0.5
Cp
0.0
0.5
1.0
0.0 0.2 0.4 0.6 0.8 1.0
x/ c
Figure A.6 – Local Cp for optimized setup 3 compared with baseline.

TRITA TRITA-SCI-GRU 2021:354
www.kth.se

Full Text 01

Uploaded by

Copyright:

Available Formats

Full Text 01

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Full Text 01

Uploaded by

Copyright:

Available Formats

DEGREE PROJECT IN MECHANICAL ENGINEERING,

SECOND CYCLE, 30 CREDITS

Shape optimization of axial

KTH ROYAL INSTITUTE OF TECHNOLOGY

Degree Programme in Mechanical Engineering

© 2021 Martin Granlöf

Keywords— Fan noise, shape optimization, mesh morphing, computational

Denna avhandling fokuserar på form-optimering av fläkt blad via mesh morphing

Keywords— Fan noise, shape optimization, mesh morphing, computational

Stockholm, June 2021

4 Simulation setup and testing 27

5 Results and discussion 45

6 Conclusions and Future work 61

A NACA0012 result plots 71

1.1 Axial cooling fan mounted to radiator. . . . . . . . . . . . . . 2

2.1 Targeted fan. . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

3.1 Overview of the EGO optimization process connected to

4.1 NACA0012 airfoil. (a), unperturbed shape and (b), an example

5.1 Performance of the designs evaluated in the RP Mopt optimization.

5.6 Performance of the designs in the ET Aopt1.2 optimization.

A.1 EGO optimization of NACA0012 airfoil for setup 2 . . . . . . 71

A.2 EGO optimization of NACA0012 airfoil for setup 3 . . . . . . 72

4.1 Important flow quantities used in the optimization performed

5.1 Volume flow rate, control surface points, weighting factors

List of acronyms and abbreviations

AoA Angle of Attack

BEV Battery Electric Vehicle

CAD Computer-Aided Design

CFD Computational Fluid Dynamics

DOE Design of Experiment

EGO Efficient Global Optimisation

EST Enhanced Stability Treatment

FFD Free-Form Deformation

FFT Fast Fourier Transform

FW-H Ffowcs Williams Hawkings

GBM Gradient-Based Methods

ICE Internal Combustion Engine

LHS Latin Hypercube Sampling

MLP Multi-Layer Perception

MOEA Multi-Objective Evolutionary Algorithm

MOGA Multi-Objective Genetic Algorithm

NSGA Non-dominated Sorting Genetic Algorithm

NURBS Non-Uniform Rational Basis Spline

OASPL Overall Sound Pressure Level

PDE Partial Differential Equation

PSO Particle Swarm Optimisation

RANS Reynolds-averaged Navier-Stokes

RBF Radial Basis Function

RBNN Radial Basis Neural Network

RSA Response Surface Approximation

SPL Sound Pressure Level

VOF Volume of Fluid

In the year 2017 a motion containing a ban on newly produced combustion

around these operational parameters is paramount in order to also maximize

Figure 1.1 – Axial cooling fan mounted to radiator.

Normalized pressure rise and efficiency 1.2 38%

1.2.1 Original problem and definition

1.2.2 Available methods for problem solving

• Gradient-based methods (GBM), which utilize derivatives of the

information is then used in order change the parameters between iterations

X = sρU 2 (s − p )dr (2.10)

∂Ui ∂Ui 1 ∂P ∂ h ∂U ∂Uj 2 i