Professional Documents
Culture Documents
A DISSERTATION
SUBMITTED TO THE INSTITUTE FOR COMPUTATIONAL
AND MATHEMATICAL ENGINEERING
AND THE COMMITTEE ON GRADUATE STUDIES
OF STANFORD UNIVERSITY
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
Abstract
Nonlinear fluid-structure interaction (FSI) is a dominating feature in many important
engineering applications. Examples include underwater implosions, pipeline explosions, flapping wings for micro aerial vehicles, and shock wave lithotripsy. Due to the
inherent nonlinearity and system complexity, such problems have not been thoroughly
analyzed, which greatly hinders the advance of related engineering fields.
This thesis focuses on the development, verification, and validation of a fluidstructure coupled computational framework for the solution of nonlinear multi-phase
FSI problems involving high compressions and shock waves, large structural displacements and deformations, self-contact, and possibly the initiation and propagation of
cracks in the structure.
First, an embedded boundary method for solving 3D multi-phase compressible
inviscid flows on arbitrary (i.e. structured and unstructured) non body-conforming
CFD grids is presented. Key components include: (1) robust and efficient computational algorithms for tracking open, closed, and cracking fluid-structure interfaces
with respect to the fixed, non body-conforming CFD grid; (2) a numerical algorithm
based on the exact solution of local, one-dimensional fluid-structure Riemann problems to enforce the no-interpenetration transmission condition at the fluid-structure
iv
interface; and (3) two consistent and conservative algorithms for enforcing the equilibrium transmission condition at the same interface.
Next, the multi-phase compressible flow solver equipped with the aforementioned
embedded boundary method is carefully coupled with an extended finite element
method (XFEM) based structure solver, using a partitioned procedure and provably
second-order explicit-explicit and implicit-explicit time-integrators. In particular, the
interface tracking algorithms in the embedded boundary method are adapted to tracking embedded discrete interfaces with phantom elements and carrying implicitly represented cracks.
Finally, the resulting fluid-structure coupled computational framework is applied
to the solution of several challenging FSI problems in the fields of aeronautics, underwater implosions and explosions, and pipeline explosions to assess its performance.
In particular, two laboratory experiments are considered for validation purpose: the
first one concerns the implosive collapse of an air-filled aluminum cylinder; the second
one studies the dynamic fracture of pre-flawed aluminum pipes driven by detonation
waves. In both cases, the numerical simulation correctly reproduces in a quantitative
sense the important features in the experiment.
Acknowledgements
First and foremost, I offer my deepest gratitude to my advisor, Professor Charbel
Farhat, for his constant encouragement, support, and guidance through the last five
years. Without him, this work would not have been possible. I also thank him
for providing me with unique opportunities for interaction with industry and the
academic world.
I would like to thank the other members of my thesis committee, Professor Adrian
Lew and Professor Gianluca Iaccarino, for reviewing this work and providing me with
helpful advice and comments.
I wish to express my warm and sincere thanks to Professor Ron Fedkiw and
Professor Stelios Kyriakides for their detailed suggestions and constructive comments
on the development and validation of the computational framework proposed in this
work. I am also grateful to Professor Jean-Frederic Gerbeau and Dr. Michel Lesoinne
for their precious help during their various visits to Stanford.
I would like to express thanks to my former and current research collaborators:
Dr. Arthur Rallu, Dr. Phil Avery, Jon Gretarsson, Alex Main, Dr. Jeong-Hoon Song,
Patrick Lea, and Dr. Liang-Hai Lee. By working with them, I have been able to
significantly extend my knowledge of computational mathematics and mechanical
engineering. I am also very grateful to all my colleagues in the Farhat Research Group,
vi
with whom I have spent five years of joyful time. Special thanks to Dr. Charbel BouMosleh for generating numerous high-quality computational meshes and Dr. Julien
Cortial for proofreading this document.
I want to acknowledge the Office of Naval Research (ONR), which sponsored this
work.
Finally, I wish to thank my friends and family for their unconditional support.
vii
Contents
Abstract
iv
Acknowledgements
vi
1 Introduction
1.1
Motivations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1.1
Underwater implosions . . . . . . . . . . . . . . . . . . . . . .
1.1.2
Pipeline explosions . . . . . . . . . . . . . . . . . . . . . . . .
1.1.3
1.1.4
1.1.5
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2
Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
1.3
11
1.3.1
Thesis accomplishments . . . . . . . . . . . . . . . . . . . . .
11
1.3.2
Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
2 Mathematical Models
2.1
14
Fluid model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
2.1.1
15
Governing equations . . . . . . . . . . . . . . . . . . . . . . .
viii
2.1.2
2.2
2.3
2.4
Equations of state . . . . . . . . . . . . . . . . . . . . . . . . .
17
Structure model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
18
2.2.1
Governing equations . . . . . . . . . . . . . . . . . . . . . . .
19
2.2.2
Constitutive laws . . . . . . . . . . . . . . . . . . . . . . . . .
19
2.2.3
20
Interface conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
2.3.1
22
2.3.2
23
24
2.4.1
24
2.4.2
27
2.4.3
28
3 Computational Framework
31
3.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
31
3.2
. . . .
37
3.3
39
3.3.1
39
3.3.2
42
3.3.3
Time integration . . . . . . . . . . . . . . . . . . . . . . . . .
43
44
3.4.1
44
3.4.2
Time integration . . . . . . . . . . . . . . . . . . . . . . . . .
46
3.5
47
3.6
49
3.4
ix
54
4.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
4.2
A projection-based approach . . . . . . . . . . . . . . . . . . . . . . .
59
4.2.1
4.2.2
60
Signed distance between a CFD grid point and its closest point
on the embedded interface . . . . . . . . . . . . . . . . . . . .
61
Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
63
A collision-based approach . . . . . . . . . . . . . . . . . . . . . . . .
67
4.3.1
68
4.3.2
Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
69
4.3.3
70
4.4
72
4.5
Numerical examples . . . . . . . . . . . . . . . . . . . . . . . . . . . .
74
4.5.1
75
4.5.2
78
4.5.3
78
4.2.3
4.3
88
5.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
88
5.2
90
5.2.1
Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
90
5.2.2
95
5.2.3
96
5.3
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
x
5.3.2
5.3.3
5.3.4
6 Applications
139
6.1
6.2
6.3
6.4
. . 145
6.5
6.6
6.4.1
Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
6.4.2
Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
6.5.2
6.5.3
196
7.1
7.2
200
Bibliography
209
xi
List of Tables
4.1
6.1
76
6.2
Statistics of five non body-conforming CFD grids created for the simulation of a steady flow around an aircraft. . . . . . . . . . . . . . . . 148
xii
List of Figures
1.1
1.2
Ruptured pipe sections from the Hamaoka Nuclear Power Station accident (left) and the San Bruno natural gas transmission pipeline accident (right).
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3
1.4
1.5
1.6
2.1
2.2
26
L and EOS-R refer to the EOS that governs the fluid medium on the
left and right of the contact discontinuity, respectively. . . . . . . . .
2.3
28
xiii
30
3.1
3.2
4.1
39
49
Domain setting of an Eulerian embedded method for fluid-structure interaction: extended fluid domain F , structural domain S , embedded
surface E , and outward normal n~E to E . . . . . . . . . . . . . . . .
4.2
4.3
65
4.5
61
4.4
58
66
Illustration of Algorithm 4.1 and Algorithm 4.2 for three distinctive cases. A point in green (blue) color represents a CFD grid point
lying the fluid (structure) region of the computational domain. An
edge in red represents a fluid-structure intersecting edge. . . . . . . .
71
4.6
72
4.7
4.8
73
4.9
xiv
. . . . . . . .
81
4.10 Tracking an embedded AGARD wing: scope decomposition of the embedded discrete interface for 32 fluid computational subdomains (each
color designates a relevant component of the interface for a specific
subdomain for which a bounding box hierarchy is computed on the
assigned CPU). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
81
82
4.12 A circular cylinder: DhE (magenta color, top) and a cut-view of the non
body-conforming CFD grid at z = 0 of Dh (bottom). . . . . . . . . .
83
85
86
86
4.17 A pair of ultra-thin triangular wings: intersecting edges and node statuses identified by Algorithm 4.2. . . . . . . . . . . . . . . . . . .
5.1
Two control volumes on the left and right sides of a region of an embedded discrete interface (two-dimensional case, quadrilateral mesh).
5.2
94
5.3
87
. . . . . . . . . . . . . . . . . . . . . . .
96
xv
98
5.4
A one-dimensional stiffened gas thin shell perfect gas problem: analytical and numerical solutions for the density field at time
t = 0.007 s. Top: the entire computational domain; bottom-left: the
vicinity of the rarefaction; bottom-right: the vicinity of the shock. . .
5.5
99
A one-dimensional stiffened gas thin shell perfect gas problem: analytical and numerical solutions for the velocity field at time
t = 0.007 s. Top: the entire computational domain; bottom-left: the
vicinity of the rarefaction; bottom-right: the vicinity of the shock. . . 100
5.6
A one-dimensional stiffened gas thin shell perfect gas problem: analytical and numerical solutions for the pressure field at time
t = 0.007 s. Top: the entire computational domain; bottom-left: the
vicinity of the rarefaction; bottom-right: the vicinity of the shock. . . 101
5.7
A one-dimensional barotropic liquid thin shell perfect gas problem: analytical and numerical solutions for the density field at time
t = 0.007 s. Top: the entire computational domain; bottom-left: the
vicinity of the rarefaction; bottom-right: the vicinity of the shock. . . 102
5.8
A one-dimensional barotropic liquid thin shell perfect gas problem: analytical and numerical solutions for the velocity field at time
t = 0.007 s. Top: the entire computational domain; bottom-left: the
vicinity of the rarefaction; bottom-right: the vicinity of the shock. . . 103
5.9
A one-dimensional barotropic liquid thin shell perfect gas problem: analytical and numerical solutions for the pressure field at time
t = 0.007 s. Top: the entire computational domain; bottom-left: the
vicinity of the rarefaction; bottom-right: the vicinity of the shock. . . 104
xvi
5.10 A one-dimensional perfect gas solid body problem: two non bodyconforming CFD grids with 40 and 80 grid points in length-direction.
105
106
xvii
5.21 Realistic case: coarsest mesh in set S2 (arbitrary, triangular, not satisfying Assumption 5.1) and reconstructed (left) and surrogate (right)
interfaces. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
5.22 Ideal case: performance of Algorithm 5.3 and Algorithm 5.4 for
load computation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
5.23 Realistic case: performance of Algorithm 5.3 and Algorithm 5.4
for load computation. . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
6.1
6.2
6.3
6.4
. . . . . . . . . . 146
6.5
Two different views of a surface grid with 853 grid points and 1, 687
triangular elements for an aircraft with high aspect ratio. . . . . . . . 147
6.6
6.7
6.8
6.9
. . . 149
6.11 Ultra-thin flexible flapping wings: description and structural modeling. 153
6.12 A pair of ultra-thin triangular wings: computational fluid domain and
the embedded discrete interface (magenta). . . . . . . . . . . . . . . . 154
6.13 A pair of ultra-thin triangular wings: non body-conforming CFD grid
cutview at y = 5 mm (left) and cutview at z = 0 mm (right). . . . 154
6.14 Ultra-thin flexible flapping wings: time-history of a tip displacement
along the z-axis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
6.15 Ultra-thin flexible flapping wings: snapshots of the fluid pressure (cutview
at y = 20 mm) and structural deformation at six different time-instances.157
6.16 Schematic drawing of a cylindrical implodable with end caps designated
by stripes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
6.17 Photographs of the collapsed cylinder (courtesy of Stelios Kyriakides). 161
6.18 Pressure time-history recorded by sensor 1 (courtesy of Stelios Kyriakides). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
6.19 CSD grid for Model 1.1 and 1.2. The elasto-plastic shell elements for
the cylinder and the rigid shell elements for the plug are distinguished
by color. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
6.20 Model 1.1 for an aluminum cylinder and a steel plug. . . . . . . . . . 164
6.21 Model 1.2 for an aluminum cylinder and a steel plug. . . . . . . . . . 165
6.22 The structure grid for Model 2. . . . . . . . . . . . . . . . . . . . . . 168
6.23 Model 2 for an aluminum cylinder and a steel plug. . . . . . . . . . . 169
6.24 Underwater implosion problem: the CFD domain and the embedded
discrete interface. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
6.25 A cut-view at z = 0 of CFD grid Dh for the underwater implosion
problem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
xix
. 179
xx
6.38 The initial crack prescribed on the CSD model for a pre-flawed aluminum pipe. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
6.39 The embedding CFD domain and the embedded CSD model. . . . . . 188
6.40 Two cut-views of the non body-conforming CFD grid for a fluid-structure
coupled pipe explosion simulation. Left: cut-view at y = 0 m; right:
cut-view at x = 0.457 m . . . . . . . . . . . . . . . . . . . . . . . . . 188
6.41 Snapshots of the fluid pressure (cut-view at y = 0 m) and structural
deformation at six different time-instances. Top-left: t = 0 s; topright: t = 4.0 105 s, middle-left: t = 8.0 105 s; middle-right: t =
1.6 104 s; bottom-left: t = 3.2 104 s; bottom-right: t = 4.0 104 s.189
6.42 Time-history of overpressure at a sensor location (0.24 m above the
notch): experimental signal (left) and simulation result (right). . . . . 190
6.43 The peak overpressure at a sensor location (0.24 m above the notch)
as a function of initial notch length: experimental and simulation results.190
6.44 The CFD domain and the embedded CSD model for an underwater
explosion simulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
6.45 A cut-view (y = 0 m) of the non body-conforming CFD grid for an
underwater explosion simulation. . . . . . . . . . . . . . . . . . . . . 193
6.46 The structural deformation and effective plastic strain at T = 1.0
103 sec. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
6.47 The fluid pressure field (left) and the domains of water and air (right)
at 1.78 104 sec (top), 6.25 104 sec (middle), and 1.0 103 sec
(bottom). A cut-view for y = 0 is shown. The structure geometry is
shown in wireframe. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
xxi
xxii
Chapter 1
Introduction
Fluid-structure interaction describes a large class of physical problems where a fluid
flow causes the deformation of a structure, and the resulting change of configuration
in turn influences the fluid flow, ultimately leading to a two-way coupling. It is a fairly
common phenomenon both in nature and around human world. Examples include
the flapping flight of insects, such as fruit fly [2], and the generation of sound by reed
instrument, such as harmonica. The first scientific study of fluid-structure interaction
dates back to the 1820s when Friedrich Bessel investigated experimentally the motion
of a pendulum in a fluid. He discovered that a pendulum moving in a fluid had longer
period than in a vacuum even if the buoyancy effects were taken into account. This
phenomenon was then described as added mass as it seemed like the surrounding
fluid increased the mass of the pendulum. It had drawn immediate attention from
many contemporary mathematicians including Simeon Poisson, George Green, and
George Stokes. In 1853 George Stokes published a theoretical study [1] and concluded
that the effective mass of the cylinder moving in the fluid increased due to the
dynamic effect of surrounding fluid by the amount of hydrodynamic mass equal to
CHAPTER 1. INTRODUCTION
CHAPTER 1. INTRODUCTION
1.1
1.1.1
Motivations
Underwater implosions
The implosive collapse of an air-filled underwater structure can lead to high compressions and strong shock waves which form a potential threat to a nearby structure.
In particular, the collapse of an air-backed implodable volume1 external but close
to a submarine, such as a universal modular mast (see Figure 1.1left) or an unmanned undersea vehicle (see Figure 1.1right), may lead to the damage or failure of
the submarine. With the number of these volumes expected to increase, underwater
implosion has become a concern to the Navy, which now requires an improved capability to design and qualify submarine external payloads for implosion avoidance and
platform survivability.
From a physical point of view, the implosion of an air-filled underwater structure is
a transient, high-speed, nonlinear, multi-phase fluid-structure interaction problem. It
is characterized by ultrahigh compression and shock waves, large structural displacements and deformations, self-contact, and possibly the initiation and propagation
of cracks in the structure. The development of a computational framework for this
problem is a formidable challenge. It requires not only incorporating in the computations material failure models, but also accounting for all possible interactions of the
external liquid namely, water the internal gas namely, air and the given
nonlinear elasto-plastic structure.
1
CHAPTER 1. INTRODUCTION
1.1.2
Pipeline explosions
As a common type of accidents in gas transmission systems and nuclear power plants,
pipeline explosions can, and usually do, lead to disastrous consequences including
both casualty and property losses. For example, on November 7, 2001, a segment of
the steam condensing pipeline at the Hamaoka Nuclear Power Station (Japan) Unit-1
ruptured (see Figure 1.2left), most likely due to the detonation of hydrogen accumulated in the pipe [7]. Fortunately there were no injuries or loss of life. However, the
surrounding area was severely damaged, and the reactor had to be shut down temporarily. More recently, on September 9, 2010, a segment of a natural gas transmission
pipeline ruptured in a residential area in San Bruno, California (see Figure 1.2right).
CHAPTER 1. INTRODUCTION
The released natural gas exploded, killing eight and injuring more [8].
Figure 1.2: Ruptured pipe sections from the Hamaoka Nuclear Power Station accident
(left) and the San Bruno natural gas transmission pipeline accident (right).
In these accidents, a lot of damages and harms to structures (e.g. buildings) and
personnel are caused by the high blast overpressure in explosion waves. To predict the
blast overpressure in such events, a computational approach must take into account
the interaction between a nonlinear fluid flow carrying strong shocks and a structure
undergoing large, plastic deformations and cracking.
1.1.3
During the last decade, there have been growing interests in the design and analysis
of aeronautical systems with flapping or flexible wings for both military and civilian
applications.
Demonstrated by flying birds and insects, flapping wings can be advantageous
over fixed or rotary wings in maneuverability and lift generation, particularly for
small vehicles and at low speed. Therefore, they have great potential in micro aerial
CHAPTER 1. INTRODUCTION
vehicles (MAV) designed for surveillance and reconnaissance purposes. As an example, DARPAs2 hummingbird (Figure 1.3) was recently recognized by TIME
Magazine as one of The 50 Best Inventions of 2011. Because the wing structures are
often flexible and tend to deform during flight, the fluid and structure dynamics of
such systems are closely linked to each other [9]. As a result, fluid-structure coupled
analysis is necessary for studying the aerodynamics, lift/drag generation, and control
of flapping wing MAVs.
CHAPTER 1. INTRODUCTION
aspect ratio and low weight. As a result, they exhibit high flexibility in flight. One
example of such aircrafts is the Helios prototype vehicle, a propeller-driven, remotely
piloted aircraft developed by NASA (see Figure 1.4). Its wing has a span of 247 ft and
a chord length of 8 ft, correspondingly to a very high aspect ratio (30.9) 3 . On June
26, 2003, during a test flight, this aircraft encountered turbulence and morphed into
an unexpected, persistent, high dihedral configuration [29] (depicted in Figure 1.5
left). Consequently, it became unstable and finally crashed (Figure 1.5right). A
post-mortem investigation of this mishap [29], conducted later by NASA, conclude
that one of the root causes is
Lack of adequate analysis methods led to an inaccurate risk assessment of the
effects of configuration changes leading to an inappropriate decision to fly an
aircraft configuration highly sensitive to disturbances.
Indeed, the physics behind this mishap is characterized by a challenging fluid-structure
interaction problem involving large structural deformations induced by a turbulent
flow. To analyze this problem, the investigation committee recommend:
Develop more advanced, multidisciplinary (structures, aeroelastic, aerodynamics, atmospheric, materials, propulsion, controls, etc) time-domain analysis
methods appropriate to highly flexible, morphing vehicles.
Develop multidisciplinary (structures, aerodynamic, controls, etc) models, which
can describe the nonlinear dynamic behavior of aircraft modifications or perform
incremental flight-testing.
3
For comparison, the aspect ratio of the wing of Boeing 777-200 is 8.68.
CHAPTER 1. INTRODUCTION
1.1.4
Shock wave lithotripsy (SWL) is a non-invasive treatment of kidney stones. Extracorporeally generated shock waves are focused on the stone in order to pulverize it
into grains, which can then travel through the urinary tract and pass from human
body (see Figure 1.6). Currently, a significant percentage of kidney stones (69% in
US in 2000 [10]) are treated with this procedure. Given that a significant fraction of
human body is water, SWL highlights a complex fluid-structure interaction problem,
which involves shock waves and cavitation in the fluid medium, as well as cracking
and comminution of the structure (kidney stone). Due to the lack of knowledge in the
physics behind it, particularly the mechanism of fluid-induced stone comminution, no
fundamental improvements in SWL technology have been accomplished in the past
two decades towards better treatment efficiency and reduced tissue injury [11]. A
statistical study even reveals an unfortunate trend of moving away from SWL to
CHAPTER 1. INTRODUCTION
Figure 1.5: NASA Helios Prototype Aircraft (HP03) at high wing dihedral (left) and
falling towards the pacific ocean (right). Mishap happened on June 26, 2003.
more invasive therapies [10].
Currently, there exists no fluid-structure coupled computational method which
can simultaneously capture all the aforementioned features involved in SWL. Indeed,
recent computational studies on SWL have been focusing on either the fluid part of
the problem, particularly the flow dynamics of cavitation bubbles interacting with
shock waves and/or rigid structures [1114]; or the structure part of the problem,
particularly the stress growth and distribution in a kidney stone hit by acoustic and
shock waves [15, 16].
1.1.5
Summary
At the time of this writing, all the problems presented above are active research topics.
More generally, they represent a large class of challenging fluid-structure problems
which involve one or more of the following features:
large, possibly plastic, structural deformations;
initiation and propagation of cracks in the structure medium;
CHAPTER 1. INTRODUCTION
10
Figure 1.6: Extracorporeal shock wave lithotripsy (left) and a kidney stone (right).
strong shock waves in the fluid medium;
multiple fluid media directly in contact.
To the authors best knowledge, there currently exists no computational approach in
either academia or industry that simultaneously addresses all these challenges.
1.2
Objectives
CHAPTER 1. INTRODUCTION
11
performed:
develop an efficient and accurate embedded boundary method for solving on
fixed, non body-conforming CFD grids multi-phase compressible fluids interacting with structures undergoing large deformations and cracking;
develop a fluid-structure coupled computational framework by coupling the
multi-phase compressible flow solver equipped with the aforementioned embedded boundary method, with a state-of-the-art, extended finite element method
(XFEM) based structural solver4 .
demonstrate the effectiveness and robustness of the aforementioned computational framework for challenging applications;
validate the proposed computational framework for challenging applications using relevant laboratory experimental data.
1.3
1.3.1
Thesis accomplishments
this solver is developed and provided by Prof. Ted Belytschkos research group at Northwestern
University.
CHAPTER 1. INTRODUCTION
12
(a) robust and efficient computational algorithms for tracking open, closed,
and cracking embedded interfaces with respect to a fixed, three-dimensional,
structured or unstructured non body-conforming CFD grid;
(b) a numerical algorithm based on the solution of local, one-dimensional fluidstructure Riemann problems for enforcing the no-interpenetration transmission condition at the fluid-structure interface;
(c) two consistent and conservative algorithms for enforcing the equilibrium
transmission condition at the fluid-structure interface.
Assessment of the order of spatial accuracy of the developed embedded boundary method for academic examples.
Development of a fluid-structure coupled computational framework using a
partitioned procedure and provably second-order explicit-explicit and implicitexplicit time-integrators. This framework links two major components: (1)
the multi-phase compressible flow solver equipped with the aforementioned embedded boundary method; and (2) a state-of-the-art, extended finite element
method (XFEM) based structural solver. In particular, the interface tracking algorithms in the embedded boundary method are adapted to tracking embedded
discrete interfaces with phantom elements and carrying implicitly represented
cracks.
Validation of the developed computational framework for two laboratory experiments:
(a) the implosive collapse of an air-filled aluminum cylinder submerged in water;
CHAPTER 1. INTRODUCTION
13
1.3.2
Outline
The document is organized as follows. Chapter 2 describes the mathematical formulations considered in this work for modeling the nonlinear fluid-structure interaction
problems of interest. Chapter 3 presents the design of the fluid-structure coupled computational framework as well as the numerical methods underlying each component.
Chapter 4 details two robust, efficient, and accurate numerical algorithms for tracking
an embedded fluid-structure interface with respect to a fixed, non body-conforming
CFD grid. In addition, the performance of these algorithms are demonstrated and assessed. Chapter 5 presents the numerical algorithms used by the embedded boundary
method for the enforcement of transmission conditions which hold at the embedded
fluid-structure interface. Numerical accuracy analysis of these algorithms are also
included. In Chapter 6, the computational framework is assessed in the context of
six realistic fluid and fluid-structure problems in the fields of aeronautics, underwater
implosion, underwater explosion, and pipeline explosion. Finally, Chapter 7 provides
conclusions as well as perspectives for future work.
Chapter 2
Mathematical Models
The present chapter describes the mathematical models employed in this thesis for
analyzing highly nonlinear fluid-structure problems in general, and the physical problems described in Section 1.1 in particular. The organization of this chapter is
straightforward. A fluid-structure interaction problem can be conceptually divided
into two sub-problems: a fluid sub-problem and a structure sub-problem. The mathematical models for solving the fluid and structure sub-problems are presented in
Sections 2.1 and 2.2, respectively. Section 2.3 deals with modeling the interaction of
different materials across an interface. Finally, three one-dimensional models for fluid
and fluid-structure problems are described in Section 2.4. These simple and wellunderstood models play an important role in the design of several key algorithms in
the computational framework, to be discussed in Chapter 3, 4 and 5.
14
15
2.1
Fluid model
2.1.1
Governing equations
Let F (t) R3 be the time-dependent flow domain of interest. Within this thesis,
the fluids dynamics are assumed to be compressible and inviscid, hence governed by
the Euler equations, which account for conservations of mass, momentum, and energy.
Their strong conservative form can be written as:
W
~ F(W
~
+
) = 0,
t
in F (t)
(2.1)
where
W = (, vx , vy , vz , E)T ,
T
~ =
,
,
,
x y z
~
F(W
) = (Fx (W ), Fy (W ), Fz (W ))T ,
vx
p + v 2
x
Fx =
vx vy
vx vz
vx (E + p)
vy
vx vy
2
Fy =
p + vy
vy vz
vy (E + p)
vz
vx vz
Fz =
vy vz
p + vz2
vz (E + p)
(2.2)
(2.3)
(2.4)
, (2.5)
denotes the fluid density, E is its total energy per unit volume and is given
1
by E = e + (vx2 + vy2 + vz2 ), where e denotes the internal energy per unit mass, p
2
denotes the fluid pressure, and ~v = (vx , vy , vz ) is the velocity vector.
16
The closure of the Euler equations is obtained with an equation of state (EOS)
which relates the thermodynamic variables , p and e. Three EOSs are used in this
thesis for modeling different fluid media. They will be presented in Section 2.1.2.
Remark.
1. The Euler equations described above are an appropriate model for fluids involved in underwater implosions and pipeline explosions described in Section 1.1.1
and 1.1.2. Indeed, the violence of these phenomena leads to large motions,
shock waves, and strong acoustic waves in the fluid media (liquid or gas). Hence
the nonlinearity and compressibility of the fluids (even liquid water) must be
properly modeled. On the other hand, the time scale in these problems (103 sec
to 101 sec) is sufficiently small such that the diffusion of heat and momentum
can be neglected.
2. For flows involved in the flights of highly flexible aeronautical systems, such as
those described in Section 1.1.3, the Euler equations are only a coarse model,
as the viscous effects in these problems can be strong and even dominate the
flow dynamics. However, they can still reproduce some key features in these
problems, such as fluid-induced, large structural deformations, which bring significant challenges to numerical simulations. In the present work, this aspect
of the model is emphasized, as two applications involving highly flexible aeronautical systems are considered in Chapter 6 to demonstrate the capability of a
computational framework for handling complex geometry and large structural
deformations.
2.1.2
17
Equations of state
Three EOS are presented here for modeling different types of fluids involved in this
thesis, namely perfect gas, stiffened gas, and Tait barotropic liquid.
Perfect Gas: The EOS for a perfect gas can be written as
p = ( 1)e
(2.6)
where is the heat capacity ratio defined as the ratio of the specific heat capacity
at constant pressure (cp ) to the specific heat capacity at constant volume (cv ). For a
perfect gas, cp and cv are assumed to be constant.
The underlying physical model assumes that the gas molecules have a negligible
volume and the potential energy associated with intermolecular forces is also negligible. This EOS is widely used for modeling real gases at low to normal pressures and
normal to high temperatures. For the modeling of air, is usually set to 1.4.
Stiffened Gas: The EOS for a stiffened gas can be written as
p = ( 1)e pc
(2.7)
where is an empirical constant which is usually chosen in the way that a desired
shock speed is reproduced correctly by the EOS [31]. pc is another empirical constant
representing the intermolecular attractions. This EOS is a generalization of the EOS
for perfect gas as the latter can be obtained by setting pc = 0. It has been used
for modeling liquids and even solid materials, and to propagate waves in such stiff
media [32]. In this thesis, this EOS is used for modeling the high pressure water
in underwater implosion problems. and pc are set to 7.15 and 2.89 108 Pa,
respectively.
18
Tait barotropic liquid: The EOS for a Tait barotropic liquid can be written as
1
0
(2.8)
k1
,
k2
(2.9)
p = p0 +
where
= p0 +
= k2
p0 and 0 are the pressure and density of a reference state, usually chosen to be the
far-field state. k1 and k2 are two constants chosen to fit the relation of bulk modulus
dp
and pressure using an affine function: k1 + k2 p = d
. For water, k1 and k2 are most
2.2
Structure model
The present thesis does not involve the design of mathematical models for the structure sub-problem, nor the development of computational methods for analyzing them.
However, since they represent one of the two sub-problems in any fluid-structure interaction problem, a brief summary of the structure model used in this thesis is provided
here.
19
2.2.1
Governing equations
Let S (t) R3 be the structural domain of interest. The governing equations of the
dynamic equilibrium of the structural system are written in a Lagrangian formulation
as [35]
s x s (us , u s ) = fsext
s u
s nt = t
u = u
in S (0)
(2.10)
on t
on u
where us denotes the displacement of the structure with respect to the reference configuration (S (0)), s and s denote its density and Cauchy stress tensor, respectively,
and fsext is the external force acting on it. The dot denotes the time derivative. t is
is the applied displacement
the applied traction on the Neumann boundary t and u
on the Dirichlet boundary u . A constitutive law that expresses the stress tensor in
terms of displacement and velocity is required to close this system.
2.2.2
Constitutive laws
A structure constitutive law consists of two aspects: the relation between strain and
displacement (kinematics), and the relation between stress and strain in the structure
material (kinetics).
Geometric nonlinearity is always considered, as this thesis focuses on physical
problems involving large structural deformations (see Section 1.1 for examples). Consequently, the strain-displacement relation of a structure medium is formulated as
1
s = (u + uT + u uT ).
2
(2.11)
20
where s denotes the second-order Greens symmetric strain tensor in the structure,
and denotes the tensor product.
As for kinetics, both linear elastic and nonlinear elasto-plastic materials are considered in this thesis. Indeed, the structures in certain highly flexible aeronautical
systems, such as the ones described in Section 1.1.3, can be modeled using linear
elasticity, as permanent deformations are usually negligible. However, in underwater implosions and pipeline explosions, where irreversible deformations and cracking
appear as dominant features in the structure medium, a nonlinear elasto-plastic material law is more appropriate. In particular, the J2 -flow theory plasticity is used in
this thesis as it was specifically developed for metals [35].
2.2.3
As mentioned in Section 1.1, the present work is motivated to a large extent by fluidstructure interaction problems in which dynamic cracking appears in the structure
medium. The mathematical tools for modeling crack propagations are summarized
here. First, a cohesive crack model is employed to ensure an accurate dissipation of
energy due to cracking. It is assumed that across a crack, the normal component of
the stress tensor satisfies the following cohesive law:
s+ nc = s nc = c (JuK),
(2.12)
in which nc is the unit normal to the crack surface, and c is the cohesive traction
across it, expressed as a function of the jump of displacement across the crack in the
normal direction (nc ). Superscript plus and minus signs refer to the two sides of the
discontinuity (crack). In this thesis, a piecewise linear cohesive model is prescribed,
21
i.e.
k, 0 max
c () =
0, >
max
(2.13)
where max is the maximum crack opening displacement, and k is a constant chosen
such that the dissipated energy due to the crack propagation is equal to the fracture
energy, i.e.
max
c ()d,
Gf =
(2.14)
4
=
w(r)drd
0
where r and are the distance from the crack tip and the angle with the tangent to
the crack, respectively. w(r) is a weight function.
2.3
Interface conditions
Within the scope of this thesis, interactions of different fluid and/or structure materials are assumed to be localized at their common interfaces. Governing equations
representing these localized interactions, or interface conditions, are described in this
section together with their physical interpretations. Two types of material interfaces
22
are considered in this thesis, namely the impermeable interface between a fluid and a
structure, and the immiscible interface between two fluids such as water and air. Their
underlying interface conditions are presented in Section 2.3.1 and 2.3.2 respectively.
2.3.1
Within this thesis, the fluid and structure media are assumed to be impermeable. In
other words, interpenetration of fluid and structure particles is not allowed. To enforce
this assumption in the mathematical model, the following Dirichlet and Neumann
T
transmission conditions are imposed at F (t) S (t), the interface between the fluid
and structure media:
u n = v n
(2.15)
s n = pn
(2.16)
u n = 0,
which is well-known as the slip-wall boundary condition for an inviscid flow. Eq. 2.16
implies the equilibrium of the interaction force between the fluid and the structure,
and is referred to as the equilibrium condition in the remainder of this thesis.
23
2.3.2
Within the scope of this thesis, any fluid-fluid interface is assumed to be immiscible,
meaning they do not mix when put in contact. This assumption holds perfectly for
the underwater implosion problems described in Section 1.1.1, as the interface occurs
between water and air. In addition, effects of evaporation and surface tension are
ignored. Again, this is valid for underwater implosions. On one hand, the time scale
is sufficiently small (103 to 101 sec) such that mass diffusion (evaporation) at the
water-air interface is negligible. In addition, the effects of surface tension are also
negligible, as the surface tension is very small compared to the high water pressure
in these problems.
Consequently, in this thesis a fluid-fluid interface is modeled by a contact disconti(1)
(2)
(2.17)
p(1) = p(2) .
(2.18)
In other words, the pressure fields and the normal component (with respect to the
interface) of the velocity fields of the two fluids are continuous across the interface.
24
2.4
Three one-dimensional models for single-phase fluid, two-phase fluid, and fluid-structure
problems are discussed here. These initial value problems are characterized by piecewise constant initial state with a single discontinuity. Simple yet revealing, they have
been throughly studied [37, 38, 40] and considered as fundamental tools for understanding the nonlinear behavior of compressible inviscid flows, as well as designing
and developing numerical methods for analyzing them. In particular, they play an
important role in the design of several key algorithms in the computational framework
presented in this thesis, which is discussed in Chapter 3, 4, and 5.
2.4.1
The strong conservative form of the three-dimensional Euler equations are stated in
Eq. 2.1. For a time-independent one-dimensional fluid domain F = R, they reduce
to
w
+
F(w) = 0,
t
x
in F
(2.19)
where
w = v ,
F(w) = v 2 + p
v(E + p)
(2.20)
Assuming a single EOS holds in the entire domain, the single-phase fluid Riemann
problem (P1) is defined as finding the solution w(x, t) to Eq. 2.19, given the following
25
wL , x 0
w(x, 0) =
,
w , x>0
R
(2.21)
26
1-wave
(rarefaction)
L
vL
p
L
2-wave
(contact)
L
v
p
R
v
p
3-wave
(shock)
R
vR
p
R
x
(0, 0)
27
For a perfect gas or a stiffened gas, the existence and uniqueness of solution to
P1 is guaranteed by the Bethe-Weyl theorem, provided that the formation of
voids is allowed.
For most EOS (including the ones stated in Section 2.1.2), the exact solution
of intermediate states can not be obtained in closed-form expression. However,
approximate solutions can be computed using numerical nonlinear equation
solvers such as Newtons method (also known as the Newton-Raphson method).
2.4.2
The two-phase fluid Riemann problem (P2) is defined as finding the solution of the
one-dimensional Euler equations (Eq. 2.19), together with initial condition
wL , x 0
w(x, 0) =
,
w , x>0
R
(2.22)
with two different fluid media on the left and right of the origin. These two fluids
can be governed by either the same EOS with different parameters, or different EOS.
In both cases, the solution still has the same structure as discussed in Section 2.4.1,
except that at any given time, the fraction of the computational domain on the left or
right of the contact discontinuity is governed by its respective EOS. As an example,
the structure of solution involving a rarefaction traveling in the left fluid and a shock
traveling in the right fluid is shown in Figure 2.2
Problem P2 can still be solved by the procedure outlined in Section 2.4.1. However, in Step 2 and 3 the relation between two states across the 1 or 3-wave must be
formulated using the EOS (and parameters therein) for the fluid medium in which
28
1-wave
(rarefaction)
EOS-L
L
vL
p
L
2-wave
(contact)
EOS-L
EOS-R
L
v
p
R
v
p
3-wave
(shock)
EOS-R
R
vR
p
R
x
(0, 0)
Figure 2.2: Typical solution structure of a two-phase fluid Riemann problem. EOS-L
and EOS-R refer to the EOS that governs the fluid medium on the left and right of
the contact discontinuity, respectively.
this wave travels. Details of the solution procedure are provided in [26, 39]. All the
EOS presented in Section 2.1.2 are covered.
2.4.3
The fluid-structure Riemann problem (P3) considers a one-dimensional compressible and inviscid flow with a moving wall boundary. A constant initial condition is
prescribed for the flow, while a constant velocity is prescribed for the wall boundary. Also known as the piston problem [40], this problem is simple yet powerful.
First, it can be solved analytically and for some EOS the solution even exists in
closed-form expression. Second, it can reveal the fundamental features in the interaction of a three-dimensional compressible flow and a dynamic structure, such as the
nonlinearity of the flow in the forms of shocks and rarefactions, and a moving flow
29
boundary. Within this thesis, P3 is used in the design and development of a computational technique for enforcing the no-interpenetration condition and simultaneously
recovering the fluid pressure at the interface for the enforcement of the equilibrium
condition (Section 5.2).
Given a constant fluid initial state wL and a constant wall velocity vwall , P3
can be formulated as follows. Considering a one-dimensional time-dependent fluid
domain F (t) = (, vwall t ], find the solution w(x, t) to the one-dimensional Euler
equations stated as Eq. 2.19, together with initial condition
w(x, 0) = wL
v(vwall t, t) = vwall
30
t
moving wall
rarefaction
x vwall t
vwall
p
L
vL
p
L
not involved
x
(0, 0)
Chapter 3
Computational Framework
3.1
Introduction
The present thesis focuses on the design, development, verification, and validation of a
computational framework for highly nonlinear multi-phase fluid-structure interaction
problems. To a large extent, this work is motivated by the simulation of physical
problems described in Section 1.1, which include underwater implosions, pipeline
explosions, and the flights of certain types of highly flexible aeronautical systems.
They all share a common feature: the deformation of the structure is large, and
induced either partially or fully by the dynamic fluid flow. Other features appearing
in at least one of these problems include:
1. strong acoustic and shock waves in the fluid medium (in underwater implosions
and pipeline explosions);
2. multiple fluid media in contact (in underwater implosions and pipeline explosions);
31
32
3. strong viscous effects, laminar or turbulent, in the fluid medium (in the flights
of aeronautical systems with flapping wings or highly flexible wings);
4. plastic deformations in the structure medium (in underwater implosions, pipeline
explosions, and aircrafts for HALE);
5. initiation and propagation of cracks in the structure medium (in underwater
implosions and pipeline explosions);
The present computational framework accounts for all the features mentioned
above except the viscous effects that occur in the flights of aeronautical systems
involving flapping or flexible wings. Adding high-fidelity models for capturing viscous
effects is feasible and in progress but considered outside the scope of this thesis (see
Chapter 7). Mathematical models used in the present framework are presented in
Chapter 2.
The design of the present computational framework is summarized here. First of
all, it is necessary to solve the dynamic fluid and structure systems simultaneously
as a coupled system. Indeed, the aforementioned features in the fluid and structure
media are responses to the interaction of the fluid and structure sub-systems. For
example, in an underwater implosion problem, the strong acoustic and shock waves
propagating in water are generated by the self-contact of the collapsed structure; while
the collapse and self-contact of the structure are caused by the high water pressure.
Similarly, in a pipeline explosion problem, the initiation and propagation of cracks
in the structure medium lead to strong acoustic and shock waves entering the fluid
outside the pipe, whereas the cracking of the pipe is caused by the high dynamic fluid
pressure inside it.
33
Depending on whether the semi-discretized fluid and structure systems are timeintegrated jointly as one system or separately as two systems, computational formulations for transient fluid-structure coupled systems can be categorized in two classes,
namely monolithic approaches [4143] and partitioned approaches [4446, 48, 49].
In a monolithic approach, the fluid and structure governing equations are combined and solved using a single preferred time-integrator. This can be appealing as
it is relatively easy to achieve the desired numerical properties, including accuracy,
stability, and conservation, for the entire coupled system. However, it necessitates
writing a fully-integrated fluid-structure solver, thus precluding the use of existing
fluid and structure softwares. Additionally, in order to integrate the fluid and structure systems together, modeling assumptions are commonly needed in at least one
system, which inevitably reduces model fidelity.
By contrast, in a partitioned approach, the fluid and structure systems are semidiscretized and time-integrated by different schemes carefully tailored to their corresponding mathematical models, and coupled through explicit inter-system communications. This strategy enables the exploitation of off-the-shelf software components
specifically developed for each system, and makes replacements relatively painless
when better mathematical models and numerical methods emerge. However, it is
relatively difficult to achieve some numerical properties, particularly conservation,
for the entire coupled system. The loss of conservation can be troublesome for some
problems, as discussed in [51,52]. A more detailed comparison, or debate, over monolithic and partitioned approaches can be found in [48] and [43], with advocates on
both sides.
34
35
at Lawrence Livermore National Laboratory (LLNL) for structural/continuum mechanics problems [65]. Both solvers are capable of modeling geometric and material
nonlinearities, as well as contact. AERO-S supports both explicit and implicit timeintegrators whereas DYNA3D is limited to explicit ones. The main algorithms used
by these two solvers are summarized in Section 3.4. Fracture dynamics is modeled via
the extended finite element method, or XFEM, following the phantom node formulation [36]. Underlying algorithms are summarized in Section 3.5. This capability is
incorporated into the in-house version of DYNA3D by Prof. Ted Belytschkos team.
Popular numerical techniques designed for handling the interaction of a fluid and
a structure in general, and enforcing the fluid-structure transmission conditions (see
Section 2.3.1) in particular, can be divided into two classes depending on whether
the CFD grid moves/deforms according to the motion/deformation of the structure.
The first class of methods, including the well-known Arbitrary Lagrangian-Eulerian
(ALE) method [64, 67, 68, 95], the closely-related dynamic mesh method [69], and
the co-rotational approach [79,80], operate on dynamic, body-conforming CFD grids.
Equipped with a mesh motion computation algorithm (for examples, see [7072]),
this type of methods allow the CFD grid to move or deform in the way that it is
always conformal to the dynamic wet surface of the structure. These methods can be
appealing as they naturally accommodate spatially high-order schemes at the fluidstructure interface, and allow a high mesh resolution for boundary layers in viscous
flows. However, achieving efficiency, robustness, and high mesh quality requires extreme care when the motion/deformation of the structure is large and not known a
priori [26]. Moreover, as soon as the structure undergoes some kind of topological
change, such as cracking, the challenge becomes practically insurmountable. The
36
second class of methods, collectively referred to in this thesis as embedded boundary methods, operate on fixed, non body-conforming CFD grids. During the last
four decades, embedded boundary methods have gained tremendous popularity in
CFD and fluid-structure interaction, under different names. These include immersed
boundary [73, 74], embedded boundary [49, 75], immersed interface [85, 86], fictitious
domain [19], and Cartesian [76] methods. They can be attractive because they simplify a number of issues ranging from meshing the fluid domain, to formulating and
implementing Eulerian-based algorithms for difficult fluid-structure applications such
as those involving very large structural motions and deformations [49], or topological
changes [78]. However, because a non body-conforming CFD grid does not contain a native representation of the wet surface of the structure, embedded boundary
methods also tend to complicate other issues such as the treatment of fluid-structure
transmission conditions.
In the present computational framework, an embedded boundary method is developed and used, since large deformation and cracking appear as main features of interest. This method is equipped with novel algorithms for tracking the fluid-structure
interface with respect to the CFD grid, as well as enforcing the fluid-structure transmission conditions. The fundamental concepts of this method are summarized in
Section 3.6, while the design and development of specific algorithms are detailed in
Chapter 4 and 5.
3.2
37
In a partitioned procedure, the fluid and structure systems are semi-discretized and
time-integrated by different numerical schemes that are tailored to their different
mathematical models, and coupled through explicit inter-system communications.
These numerical schemes can be advanced in time either simultaneously, or one at
a time in a staggered manner. In the former case, the resulting partitioned analysis
procedure is usually referred to as a strongly coupled solution algorithm. In the latter
case, each fluid or structure time-step can be completed either in one step or through
a loop of sub-iterations. A staggered solution procedure with carefully designed subiterations is sometimes also considered as a strongly coupled algorithm, whereas a
staggered, sub-iteration free procedure is referred to as a loosely coupled algorithm.
Within the domain of partitioned procedures, loosely coupled algorithms can be
advantageous compared to their strong counterparts because of reduced computational cost and simplified software development. In the past, they have been suspected for inherent issues in accuracy and stability. Indeed, for some schemes [27],
the time-accuracy of the coupled scheme is found to be one order lower than the
ones of the underlying fluid and structure time-integrators; whereas its stability limit
tends to be more restrictive than the ones of the fluid/structure time-integrators.
However, recently it has been shown that these deficiencies can be overcome with
carefully designed prediction/correction approaches, without substantial increase of
computational cost. First, provably second-order explicit-explict, implicit-explict,
and implicit-implict schemes have been developed using second-order time-integrators
for both the fluid and structure systems [27, 48]. Second, it is shown in [27] that
38
at least for a specific numerical example, the time-step stability limit of a carefully
designed explicit-explicit scheme is equal to the lowest of the time-step stability limits
of the individual fluid and structure schemes.
Loosely coupled partitioned procedures developed in [48] and [27] are adopted
in the present thesis. It should be noted that the appealing accuracy and stability
properties of these algorithms reported there are mathematically proved and experimentally verified for an ALE framework. Whether they hold for the present computational framework, in which an embedded boundary method is used instead, has not
been rigorously investigated yet. This work is considered as a possible future research
direction and discussed in Chapter 7.
Finally, the basic cycle of a loosely coupled partitioned procedure within one timestep can be summarized as follows:
1. transfer the motion of the wet surface of the structure to the fluid system;
2. advance in time the fluid system with a specific time-integrator;
3. compute the fluid-induced force load acting on the wet surface of the structure,
and send it to the structure system;
4. advance in time the structure system with a time-integrator that may be different from the one used for the fluid.
3.3
39
3.3.1
As discussed previously in Section 2.1.1, the strong conservative form of the threedimensional Euler equations
W
~ F(W
~
+
) = 0,
t
in F (t).
(3.1)
is chosen as a fundamental mathematical model. The exact definitions of the con~ are provided in Section 2.1.1 and
servative state vector W and the flux function F
omitted here. In the present computational framework, Eq. 3.1 is semi-discretized by
a classical finite volume method. For the sake of completeness, the basic steps of this
method are outlined below.
Vj
nij
@ Cij
@ Ci
Vi
Ci
Figure 3.1: Definition of a control volume Ci , its boundary surface Ci , and a facet
Cij of Ci (view of half entities for an hexahedral discretization).
40
Ci = Dh .
(3.2)
i=1
In other words, the union of the cells exactly covers the computational domain.
Using the standard characteristic function associated with a control volume Ci , a
standard variational approach, and integration by parts, Eq. (2.1) can be transformed
into its weaker form1
Z
Ci
X Z
Wh
~ h ) ~nij d = 0,
F(W
d +
t
(3.3)
jK(i)C
ij
where Wh denotes a discrete approximation of the fluid state vector W in a semidiscrete space, K(i) denotes the set of adjacent vertices of Vi , Cij denotes a segment
T
of Ci that is defined as Cij = Ci Cj , and ~nij is the unit outward normal
to Cij . The weaker form above suggests that in practice, the computations can
be performed in a one-dimensional manner, essentially by evaluating fluxes along
normal directions to boundaries of the control volumes. For this purpose, Ci is split
1
Eq. 3.3 can also be obtained by directly applying conservations of mass, momentum, and energy
within control volume Ci [40].
41
in control volume boundary facets Cij connecting the centroids of the hexahedra
sharing vertices Vi and Vj (Figure 3.1). The total flux across the boundaries of Ci ,
X Z
~ h ) ~nij d in (3.3), is approximated as follows
expressed as
F(W
jK(i)C
ij
X Z
~ h ) ~nij d =
F(W
jK(i)C
ij
(3.4)
jK(i)
Z
Wh d,
1
Wj =
|Cj |
Ci
Z
Wh d.
(3.5)
Cj
Fij is the numerical flux function evaluated using Roes approximate Riemann
solver [81] and MUSCL (Monotonic Upwind Scheme Conservation Law) [84] with
linear reconstruction. Combining Eqs. 3.4 and 3.5 with Eq. 3.3 yields
X
dWi
+ |Ci |
Fij (Wi , Wj , EOS, ~nij ) = 0.
dt
(3.6)
jK(i)
(3.7)
where W and F denote the cell-averaged state vector and the numerical flux vector
for the entire grid.
3.3.2
42
As mentioned in Section 2.3.2, different fluids are assumed immiscible and every fluidfluid interface is modeled by a contact discontinuity, or free surface. Therefore, the
velocity and pressure of the flow are continuous across the interface, whereas the
density is typically discontinuous. Such multi-phase flows with immiscible fluid-fluid
interfaces require extra computational efforts in a flow solver in order to:
locate the dynamically evolving fluid-fluid interface;
apply the interface conditions (Eqs. 2.17 and 2.18).
Numerical approaches for this type of problems abound in the literature. A review
of the popular options is provided in [26]. In the present framework, the level set
method [82, 83] is employed to locate the dynamic fluid-fluid interface, whereas a
two-phase Riemann solver based ghost fluid method [25] is used to apply the interface
conditions.
Using the level set method, the dynamic interface is implicitly represented by
the zero iso-value of the level set function , which satisfies the following advection
equation:
+ v = 0,
t
in F
(3.8)
where v denotes the velocity field of the flow, is initialized at each point in by
the signed distance from this point to the initial location of the interface.
The two-phase Riemann solver based ghost fluid method applies the interface
condition by modifying the usual numerical flux (Eq. 3.4) locally between two grid
points residing in different fluid media. More specifically, if a pair of adjacent grid
points, namely Vi and Vj , reside in different fluid media, the numerical flux across Cij
43
is defined as
ij = ij (Wi , Wj , EOSi , EOSj , ~nij ),
(3.9)
where Wi and Wj are the two constant states in the exact solution (see Section 2.4.2
for details) of the one-dimensional fluid-fluid Riemann problem, initialized by Wi and
Wj .
3.3.3
Time integration
Two second-order accurate time-integrators for the fluid system are considered.
Second-order explicit Runge-Kutta: to advance the semi-discretized fluid system
(Eq. 3.7) from time tn to tn+1 , the explicit second-order Runge-Kutta scheme can be
written as
K1 = tn F(Wn ),
(3.10)
W = Wn K1 ,
(3.11)
K2 = tn F(W ),
(3.12)
1
Wn+1 = Wn (K1 + K2 ),
2
(3.13)
(3.14)
44
Eq. 3.14 is an implicit system of nonlinear equations in Wn+1 . In the present computational framework, it is solved numerically using Newtons method.
3.4
Within this section, the displacement field of the structure system is assumed to
be continuous in both space and time. In particular, this means fracture, which is
characterized by a strong spatial discontinuity in displacement, is not considered here.
The special treatment for modeling dynamic fracture will be presented in Section 3.5.
Under the above assumption, the equations governing the dynamic equilibrium of the
structure subsystem (Eq. 2.11) are semi-discretized using a standard finite element
method. The time-integration is completed using schemes of the Newmark family.
They can be either explicit or implicit depending on the choice of two parameters.
Basic ideas of this process are summarized in this section.
3.4.1
The standard finite element method is applied to the spatial discretization of the
structure governing equations, whose strong formulation has been presented in Section 2.2.1 and is reminded here
= f ext
u x (u, u)
in S
(3.15)
First, a weak formulation is obtained by multiplying Eq. 3.15 by a test function, followed by an integration by parts. It can be stated as follows.
45
on u }, such that
Find solution u(x, t) U = {u(x, t)|u(x, t) H1 (S ), u = u
u(x) U0 = {v(x, t)|v(x, t) H1 (S ), v = 0 on u }, it satisfies
Z
u : (u, u)d
=
u
ud +
S
u f
S
ext
Z
d +
u td.
(3.16)
Following the standard finite element approach [88], the computational domain
S is subdivided into non-overlapping elements denoted by ei , i = 1, 2, ..., N e . The
union of all the elements defines a discretization of S , which satisfies
e
N
[
ei = S .
i=1
Basing on this discretization, approximate solution of Eq. 3.16 is sought in a finitedimensional subspace of U, defined as
n
U = {u(x, t) =
N
X
on u },
NI (x)uI (t), u(x, t) = u
I=1
U0 = {v(x) =
N
X
NI (x)vI , v(x, t) = 0 on u .}
I=1
n
Plugging u(x, t) =
N
X
I=1
Eq. 3.16, it can be shown that the discrete nodal displacement vector u = [u1 (t), ..., uN n (t)]T
46
(3.17)
where M denotes the mass matrix, f int denotes the vector of internal forces, and f ext
denotes the vector of external forces.
3.4.2
Time integration
Schemes from the Newmark family [87] are employed to integrate the semi-discretized
structure governing equation (Eq. 3.17). Eq. 3.17 is linearized at every time step. The
general form of the linearized system can be written as
M
u(t) + Cu(t)
+ Ku = f ,
(3.18)
where matrices M, C, K, and vector f are constant within each time-step. To integrate this system from tn to tn+1 , the family of Newmark schemes can be written
as
n)
(M + tn C + (tn )2 K)
un+1 = f n+1 C(u n + (1 )tn u
u n+1
1
n (3.19)
K un + tn u n + ( )(tn )2 u
2
n
n n
+ tn u
n+1
= u + (1 )t u
(3.20)
1
un+1 = un + tn u n + (tn )2 ( )
un
2
n+1
+ (tn )2 u
(3.21)
47
where and are two constant parameters. For example, with = 0.5 and = 0,
the explicit central difference scheme is recovered, whereas the implicit midpoint rule
is obtained by setting = 0.5 and = 0.25.
3.5
Dynamic fracture in the structure medium is modeled by the extended finite element
method (XFEM) [89,90], which has been implemented in the local version of DYNA3D
by Prof. Ted Belytschkos team, following the phantom node formulation [36, 91].
This thesis does not involve any work on the design and development of specific
algorithms in this area. However, given the significance of fracture modeling in the
present computational framework, a brief summary of involved numerical methods is
provided here.
A crack in a structure medium is represented as a strong discontinuity in its
displacement field. It cannot be captured automatically in classical finite element
analysis which relies on continuous shape functions. The key idea of XFEM is to
enrich the approximation basis with shape functions that are discontinuous across
the crack. More specifically, the XFEM approximation of the displacement field is
given by [36]
n
u(x, t) =
=
N
X
I=1
Nn
X
I=1
NI (x)uI (t) +
N
X
I=1
NI (x)H(f (x))qI ,
(3.22)
48
1, x > 0
H(r) =
.
0, x 0
(3.23)
f (x) is a level set function used to define the location of the crack [92]. It is positive
on one side of the crack and negative on the other side. Therefore, the enrichment
shape functions NI (x)H(f (x)) are discontinuous across the crack. uI and qI are
the regular and enrichment nodal variables, respectively. In practice, the enrichment
shape functions and associated nodal variables need to be introduced only in elements
traversed by the crack, or so-called cracked elements.
The phantom node formulation employed in the present computational framework [36] uses a transformation of the nodal variables which leads to a superposed
element formalism. More specifically, within a cracked element, let
and
uI
if f (xI ) < 0
u1I =
u q if f (x ) > 0
I
I
I
(3.24)
uI + qI if f (xI ) < 0
u2I =
u
if f (xI ) > 0
I
(3.25)
be the new nodal variables, the displacement field expressed in Eq. 3.22 can be rewritten as
u(x, t) =
X
IS1
(3.26)
IS2
where S1 and S2 are the index sets of the nodes of superposed element 1 and 2,
49
V1
V2
f(x) > 0
V1
V2
f
crack
e(1)
V6
f(x) < 0
V7
V4
V5
V8
f(x) > 0
V3
e(2)
f(x) < 0
real node
phantom node
V4
V3
Figure 3.2: The phantom node formulation: each cracked element is replaced by two
phantom elements with additional phantom nodes.
3.6
Debuted in 1972 for the coupled fluid-structure simulation of blood flows through elastic heart valves [73], immersed/embedded boundary methods have gained tremendous
popularity during the last four decades in Computational Fluid Dynamics (CFD),
50
under different names. These include immersed boundary [73, 74], embedded boundary [49,75], immersed interface [85,86], fictitious domain [19], and Cartesian [76] methods. All of these and other related methods which are popular nowadays for a large
variety of flow simulations around fixed [17, 19, 49, 7476], moving [18, 20, 21, 49, 96],
and deformable [22, 49, 53, 54, 73, 78] bodies, are collectively referred to in this thesis
as embedded boundary methods. They are particularly attractive for dynamic fluidstructure interaction (FSI) problems characterized by large structural motions and
deformations [27] or topological changes [78], for which most arbitrary LagrangianEulerian (ALE) methods [63, 64, 68] are often unfeasible.
Embedded boundary methods simplify the gridding task as they operate on non
body-conforming grids. Most of them are designed for computations on Cartesian
grids [18, 2022, 53, 54, 73, 74, 78], but some have also been tailored for computations
on unstructured meshes [49, 93]. However, embedded boundary methods complicate
the treatment of slip or no-slip wall boundary conditions in general [1721, 75, 76],
and fluid-structure transmission conditions (see Section 2.3.1) in particular [22, 49,
53, 54, 73, 78]. This is essentially because a non body-conforming CFD grid does not
contain a native representation of the wet surface of the body of interest. Indeed,
recent developments in embedded boundary methods have focused mostly on these
two issues, albeit primarily on the treatment of the velocity wall boundary condition
for incompressible viscous flows past rigid and motionless obstacles (for example, see
the review paper [74]). In this context, recently proposed algorithms for interface
treatment have focused either on some form of interpolation [55] with particular
attention to numerical stability [56] or higher-order accuracy [20, 55, 57], or on the
concept of a ghost cell [58, 59], some variant of the penalty method [60], and the
mirroring technique [61].
51
52
velocity at a wall and recovers the value of the fluid pressure at this wall via the
exact solution of local, one-dimensional, fluid-structure Riemann problems. This approach is detailed in Section 5.2. In addition, two novel, consistent and conservative
methodologies for evaluating flow-induced forces and moments on rigid and flexible
embedded interfaces are also incorporated in this computational framework. One of
them is based on the local reconstruction of the embedded discrete interfaces. The
other one is based on the level set concept. It is particularly attractive because it
rigorously allows the substitution of an embedded discrete interface by a simpler surrogate, which simplifies implementation and at the same time reduces computational
cost. These algorithms are presented in Section 5.3.
Typically, all interface treatment techniques require tracking the position of the
embedded interface (or the collection of interfaces representing the entire wet surface or the body of interest) with respect to the non body-conforming CFD grid.
In this aspect, most computational methods described in the literature have focused
primarily on closed embedded interfaces (e.g. surfaces of solid bodies) and Cartesian grids [17, 18, 2022]. However, the closed interface assumption is limiting as
many FSI problems, such as flapping wings and cracked pipes, involve open thin
shell surfaces. Furthermore, the Cartesian grid assumption forbids the use of most
state-of-the-art flow solvers, such as AERO-F, which operate on unstructured grids.
The present computational framework is equipped with two robust and efficient interface tracking algorithms capable of operating on structured as well as unstructured
three-dimensional CFD grids. The first one is based on a projection approach. It
is faster but still restricted to closed interfaces and resolved enclosed volumes. The
second algorithm is based on a collision approach with motivation from the computer
53
graphics community [53, 94]. While reasonably slower, it can handle open shell surfaces and underresolved enclosed volumes. Both computational algorithms exploit
the bounding box hierarchy technique and its parallel distributed implementation
to efficiently store and retrieve the elements of the discretized embedded interface.
These algorithms are presented in Chapter 4 in details.
Chapter 4
Tracking the Embedded
Fluid-Structure Interface
The embedded boundary method introduced in Section 3.6 operates on fixed, non
body-conforming CFD grids which do not have a native representation of the static
or dynamic wall boundary in general, and the fluid-structure interface in particular1 .
Instead, this wall boundary or fluid-structure interface is represented as a Lagrangian
surface grid, commonly referred to as the discrete embedded surface, which is generated independently from the CFD grid. This surface grid can be static or dynamic.
In the latter case, its motion/deformation can be either prescribed, as in a forced
motion/deformation simulation, or computed by interpolating or extrapolating the
displacement field of a dynamic structure, as in a two-way coupled fluid-structure
simulation. In both cases, in order to discretize and enforce the wall boundary conditions or the fluid-structure transmission conditions (see Section 2.3.1 for details),
special computational techniques must be employed to find the location of the discrete
1
54
4.1
Introduction
In general, item (1) above is needed in most embedded boundary methods for
the purpose of fluid-structure demarcation. Item (2) is commonly used in ghost-cell
based computational methods [21] for finding the image of a ghost node in the
real fluid domain. Item (3) has been needed at least for two purposes: to define a
surrogate fluid-structure interface where to enforce the transmission conditions [49],
and to reshape boundary cells in a cut-cell based computational method [18].
So far, the computational methods described in the literature for computing the
outputs highlighted above have focused primarily on closed embedded interfaces (e.g.
surfaces of solid bodies) and Cartesian CFD grids [17,18,2022]. The closed interface
assumption has led to several algorithms for determining item (1) highlighted above
that are based on inspecting the outward normals of simplices in the vicinity of
a given grid point [17, 20, 21]. This assumption is limiting however as many FSI
problems such as flapping wings and parachutes involve open thin shell surfaces. The
Cartesian grid assumption simplifies the task associated with item (2) highlighted
above. It has led to the development of both geometric [17, 18, 20, 21] and nongeometric [22] closest point transform algorithms. However, unstructured grids can
be computationally advantageous even for embedded boundary methods, particularly
for complex geometries. Furthermore, most computational algorithms that have been
proposed in the literature for computing the intersections of a given CFD grid and
a given embedded interface are based on identifying the penetrating edges of the
CFD grid as those edges whose end-points lie in different media [18]. However, such
algorithms fail when the embedded interface is closed and intersected twice by one
or several edges of the CFD grid (see Figure 4.5, Case II). This scenario is likely to
occur when the body is thin, and the volume enclosed by its surface is underresolved
by the CFD grid (see Section 4.5.1). It is detrimental to accuracy as it leads to the
Figure 4.1: Domain setting of an Eulerian embedded method for fluid-structure interaction: extended fluid domain F , structural domain S , embedded surface E ,
and outward normal n~E to E .
Also, both interface trackers presented in this section take advantage of the bounding box hierarchy [23]. A parallel distributed implementation of this concept is described in Section 4.4. This technique is a popular form of spatial partitioning. It is
used to rapidly determine whether a grid point of Dh lies near or far from the embedded discrete interface DhE . The interface tracker based on the collision approach relies
on a robust edge-triangle collision detection algorithm [24] to determine whether a
given edge of Dh intersects a particular triangle of DhE .
Finally, in Section 4.5, both proposed computational algorithms are illustrated,
and their respective performances are assessed and contrasted, in three numerical
examples arising from the fields of aeronautics and underwater implosion.
4.2
A projection-based approach
This interface tracking approach is based on the principle that only the edges of the
CFD grid connecting a node lying in the fluid medium F (t) and a node lying in
the structure medium S (t) can be classified as an intersecting fluid-structure edge.
Following this principle and assuming that (1) the embedded discrete interface has
an interior and an exterior which is equivalent to assuming that the embedded
interface E is a closed surface, and (2) the interior volume it encloses is resolved by
the CFD grid that is, any edge of Dh that intersects DhE intersects it at a single
point, the medium in which each node of a given edge of the CFD grid can be identified at each time-step of a simulation using orthographic projections and flood-fill
as described in Algorithm 4.1 below:
Algorithm 4.1
1. For each grid point Vi Dh , construct an axis-aligned bounding box bi defined
as the smallest axis-aligned box containing Vi and its adjacent nodes.
2. Construct an axis-aligned bounding box hierarchy BE which stores the triangles
of DhE .
3. For each grid point Vi Dh , determine if it lies close to DhE , and if it does, find
the location of its closest point on DhE and determine the status si of Vi . This
step can be performed as follows:
3.1. Set si = 1 to indicate that, so far, the status of Vi is unknown.
3.2. Using the axis-aligned bounding box hierarchy BE , find the set of candidate
triangles C(Vi ) DhE that may contain the closest point to Vi denoted here
by Vi0 . These are the triangles in DhE whose bounding boxes intersect bi .
3.3. If C(Vi ) 6= , find the location of Vi0 and compute (Vi0 , Vi ), the signed
distance from Vi0 to Vi . If (Vi0 , Vi ) > 0, set si = 0 to indicate that Vi is
inside the fluid domain ?F (t). If (Vi0 , Vi ) 0, set si = 1 to indicate that
Vi is inside S and thus outside ?F (t).
4. Determine the status of all remaining CFD grid points using a flood-fill algorithm as follows. For each grid point Vi Dh , if si 6= 1, loop through its
adjacent nodes N (Vi ). For each Vk N (Vi ), if sk = 1, set sk = si .
5. Compute the intersections between the edges of Dh and elements of DhE as
follows. For each edge (Vi , Vj ), if si 6= sj , identify this edge as a fluid-structure
intersecting edge. Then, cast a ray from Vi to Vj and find the intersection point
on DhE .
The critical components of Algorithm 4.1 described above are the search for
a given grid point Vi of its closest point Vi0 DhE , and the calculation of (Vi0 , Vi ).
These are explained next in details.
4.2.1
To find Vi0 , the closest point to Vi lying in each triangle Tk C(Vi ) is first determined.
(k)
all three coordinates are non-negative, the projection point is Vi . If one or two
(k)
coordinates are negative, the projection point lies outside Tk . In this case, Vi
is
located either on the line containing an edge of Tk , or at one of its vertices. Therefore, Vi is reprojected in this case onto each edge of Tk corresponding to a negative
coordinate. For example, if A < 0, Vi is projected onto the line determined by BC.
(k)
[A > 0 [B < 0
[C < 0
[A > 0
[B > 0
[C < 0
[A>0
[B>0
[C>0
V1
[A > 0
[B < 0
[C > 0
[A < 0 [B > 0 B
C
[C < 0
[A < 0 [B > 0 [C < 0
V1
[A < 0
[B < 0
[C > 0
V3
V3
V2
V2
Figure 4.2: Signs of the barycentric coordinates of the projection point in different
regions.
4.2.2
Signed distance between a CFD grid point and its closest point on the embedded interface
The magnitude of (Vi0 , Vi ) is simply kVi Vi0 k2 , which is trivial to compute once Vi0
has been determined. Hence, it remains only to find the sign of this distance. To this
effect, three different cases must be treated separately:
(a) Vi0 is inside a triangle Tk C(Vi ).
(b) Vi0 is not inside a triangle Tk C(Vi ) but is the projection of Vi onto an edge of
a triangle Tk C(Vi ).
(c) Vi0 is neither according to case (a) or case (b), but is a vertex of C(Vi ).
If Vi0 falls in case (a), it is the projection of Vi onto the plane containing a triangle
Tk C(Vi ). In this case, sign ((Vi0 , Vi )) is determined as the sign of the dot product
n~k Vi0 Vi , where n~k is the unit outward normal to Tk and Vi0 Vi denotes the spatial
vector connecting the points Vi0 and Vi .
If Vi0 falls in case (b), it is the projection of Vi onto an edge of Tk C(Vi ). In this
case, the sign of (Vi0 , Vi ) is determined using information from the two triangles of
DhE sharing this edge. Two examples are shown in Figure 4.3. Using the notation
shown in this figure, this case is handled as follows. First, Vi is projected onto the
plane determined by one of the aforementioned two triangles that is not coplanar
with it 2 . For example, suppose that ABD is that triangle, and PVi is the projection
point. Then, vertex C in ABC is projected onto the same plane. The projection
point is denoted by PC . Next, the following scalar quantities sv and sc are introduced
sv = PV i Vi n~r ,
sc = PC C n~r ,
where n~r is the unit outward normal to triangle ABD . Then if sv sc > 0 (see Fig
ure 4.3-top), sign (Vi0 , Vi ) is determined as sign(sv ). Otherwise if sv sc < 0 (see
Figure 4.3-bottom), sign (Vi0 , Vi ) is determined as sign(sv ).
Finally, consider case (c) graphically depicted in Figure 4.4. The set of triangles
adjacent to Vi0 is denoted by Ni and shown in dark blue in Figure 4.4. Extended
2
away from Vi0 , these triangles form an infinite open surface that is denoted here by
Ni and colored in light blue in the same figure. Consider the plane crossing Vi and
orthogonal to Vi~0 Vi . For any point on this plane sufficiently far from Vi , its closest
point on Ni is either on a face or on an edge (see Figure 4.4). In other words, this
point, denoted here by Vi , falls into case (a) or case (b), and therefore (Vi0 , Vi ) can
be computed as discussed above. Finally, the sign of (Vi0 , Vi ) is determined as the
sign of (Vi0 , Vi ).
4.2.3
Remarks
C
nk
Tk
Vi
Tr
D
Vi
PC
PVi
B
nr
C
nk
Tk
D
Tr
nr
Vi
PVi PC
Vi
Figure 4.3: Determination of the signed distance (Vi0 , Vi ) when Vi0 , the closest point
to Vi on DhE , lies on an edge of a triangle.
Vi
Vi
Ni
Vi
Ni
Vi
Vi
Vi
Ni
Ni
Figure 4.4: Determination of the signed distance (Vi0 , Vi ) when Vi0 , the closest point
to Vi on DhE , is the vertex of a triangle.
4.3
A collision-based approach
When the embedded discrete interface DhE is a closed surface and the volume it
encloses is not resolved by the fluid grid for example, when the embedded interface
is the wet surface of a thin wing and a single element of the CFD grid embeds pieces
of both the upper and lower surfaces of the wing or when DhE is an open surface like
a membrane sheet, a different approach for tracking the interface becomes necessary.
Indeed in the first case, the intersecting fluid-structure edge can connect two points
that lie in the same fluid medium, and in the second case, there is no inside and
outside to the embedded surface. To address the above issues, a robust geometric
approach based on the point-simplex collision algorithm [24] is described here to
determine whether an edge intersects DhE or not and provide the related interface
tracking information.
The point-simplex collision algorithm is robust in the sense that it is guaranteed
to never report a false negative that is, to never have an edge erroneously miss
DhE . The trade-off for this robustness, however, is that such an algorithm occasionally
reports false positives, either when an edge passes close to a sharp feature in DhE , or
when one grid point of an edge lies infinitesimally close to DhE . The former case can
be addressed by casting rays back and forth along the edge; if only one of these two
rays intersects DhE , it can be safely treated as a false positive and ignored. In the
latter case where a grid point lies so close to DhE that determining numerically on
which side of the interface it lies becomes challenging, the grid point can be flagged
as occluded. Then, every edge connected to this grid point can be automatically
flagged as having intersected the embedded surface. In this work, occluded vertices
are always considered to be inside S even in the case of a membrane sheet.
4.3.1
4.3.2
Remarks
4.3.3
SE
Dh
Dh
SE
Dh
SE
SE
Dh
Dh
SE
Dh
SE
Figure 4.5: Illustration of Algorithm 4.1 and Algorithm 4.2 for three distinctive
cases. A point in green (blue) color represents a CFD grid point lying the fluid
(structure) region of the computational domain. An edge in red represents a fluidstructure intersecting edge.
implicitly using a local distance function. In a fluid-structure coupled computation,
the connectivity of each new phantom element is sent to the fluid solver, which then
updates DhE accordingly. The values of the local distance function at its nodes are
also sent to the fluid solver and store there.
Algorithm 4.2 is then modified as follows. In Step 4, if the detected intersection
point is located in a phantom element in DhE , the local distance function, denoted
by , is interpolated at this point. If 0, the intersection point is located in
the real region of the phantom element, and considered as a valid intersection point
(Figure 4.6left). If < 0, this intersection is ignored as it is located in the phantom
Vi
f* > 0
f* < 0
Vj
a valid intersection
Vj
4.4
Both Algorithm 4.1 and Algorithm 4.2 described above call for the computation
of the bounding box hierarchy BE for the embedded discrete interface. For dynamic
problems, BE is time-dependent. For complex geometries or sufficiently refined discretizations, the computational cost associated with its construction is not negligible.
Most importantly, for a massively parallel distributed CFD simulation where the
computational domain is typically decomposed into subdomains, triangles in the embedded discrete interface lying far outside a given CFD subdomain are irrelevant for
that subdomain. Therefore, adding them to the hierarchy is only detrimental to computational efficiency. Hence, the key idea here is to implement a distributed bounding
box hierarchy for massively parallel computations where the hierarchy is constructed
only over triangles near a given fluid subdomain, thereby creating a scope consisting
only of triangles that are relevant to this particular subdomain.
Consider the simple two-dimensional example shown in Figure 4.7. Here, the
10
2
3
9
16
15
11
7
13
14
10
12
scope 10
5
8
Figure 4.7: CFD subdomains and distributed bounding box hierarchy: scoping for
subdomain 10 (right).
Constructing a distributed bounding box hierarchy for either Algorithm 4.1 or
Algorithm 4.2 begins with constructing a global bounding box hierarchy on each
processor where all elements of DhE are assumed to be stored. Then, this hierarchy
is used in the first time-step to identify in each subdomain by the processor assigned to this subdomain the triangles of DhE that are near each grid point of this
subdomain (the candidate triangles). Next, the candidate triangles are gathered on
a subdomain basis, each forming a scope or relevant component of the embedded discrete interface for that subdomain. In the second time-step, the global bounding box
hierarchy stored in each processor is replaced by its distributed counterpart where
only the triangles within the scope of the subdomain mapped to this processor are
stored.
In a dynamic fluid-structure simulation, the embedded discrete interface moves
and/or deforms in time. Therefore in such a simulation, the scope of each subdomain must be updated, in principle, at every time-step. Assuming that the coupled
fluid-structure analyzer obeys a time-step restriction that typically restrains DhE from
crossing more than one layer of elements of the CFD grid Dh in a given time-step, the
updated subdomain scope needs to include only: (1) the current candidate triangles
for each grid point in the interior of this subdomain, and (2) any additional candidate
triangle for each grid point at the boundary of this subdomain. To obtain the latter
information, interprocessor communication is needed only locally that is, between
processors assigned to neighboring subdomains.
4.5
Numerical examples
The purpose of this section is to assess and compare the performances of the presented fluid-structure interface tracking algorithms (Algorithm 4.1 and 4.2). To
this end, three numerical examples arising from fluid-structure interaction problems
in aeronautics and underwater implosions are considered. Each one is designed to
represent a different case discussed in Section 4.3 and illustrated in Figure 4.5. In the
first example, the surface of a thin airplane wing is embedded in a relatively coarse
CFD grid. This example falls into Case II of Figure 4.5 as the volume inside the wing
is not well resolved by the CFD grid. The second example, in which a circular cylinder with low aspect ratio is embedded in a fine CFD grid, arises from an underwater
implosion problem and illustrates Case I of Figure 4.5. The last example corresponds
to Case III of Figure 4.5. In this example, an open surface discretized with one sheet
of triangular elements is embedded in the CFD grid. It represents the surface of a
pair of ultra-thin triangular wings designed for the study of large amplitude flapping
motion. Finally, it is noted that in all three examples, the three-dimensional CFD
grid Dh is unstructured and constructed using tetrahedral elements.
4.5.1
The AGARD3 445.6 wing is considered here. It has a root chord length Lc = 22.0 in,
a semi-span Ls = 30.0 in, a tip chordlength Lt = 14.5 in, and a quarter-chord sweep
angle of 45o . Its panel aspect ratio is equal to 1.65 and its taper ratio is equal to 0.66.
Its airfoil section is the NACA4 65A004, with a maximum thickness of approximately
0.9 in (4% of root chord length). Hence, this wing is quite thin.
The surface of the wing is discretized with 20, 721 grid points and 41, 438 triangles (Figure 4.8-top). This discrete representation DhE is embedded in a non bodyconforming CFD grid Dh with 105, 030 grid points and 609, 576 tetrahedra (Figure 4.8-bottom).
Both Algorithm 4.1 and Algorithm 4.2 are tested here to track DhE in Dh .
Figure 4.9 displays the edges of Dh that are flagged as fluid-structure intersecting
edges by Algorithm 4.1 and Algorithm 4.2. It can be observed that, as anticipated, Algorithm 4.1 misses large swaths of the structure, essentially because the
3
AGARD stands for Advisory Group for Aerospace Research and Development, an agency of
North Atlantic Treaty Organization (NATO).
4
NACA stands for National Advisory Committee for Aeronautics, a US federal agency.
performance results on 32
Average
0.050 sec
0.20 sec
CFD grid does not resolve the structure everywhere. Hence, this is the same scenario
as Case II of Figure 4.5. More specifically, Algorithm 4.1 does not recognize the
CFD grid edges that intersect the embedded discrete interface twice at both the
upper and lower surfaces of the wing as intersecting edges, because both vertices
of each edge share the same status (see Step 5 of Algorithm 4.1). It is shown in
Section 6.1 that the failure in capturing these double intersections leads to large
numerical error in the flow solution for a fluid-structure interaction problem.
Next, the concept of scoping for constructing a distributed bounding box hierarchy
presented in Section 4.4 is illustrated. To this end, Figure 4.10 shows the scope
decomposition of the embedded discrete interface for a partitioning of the CFD grid
in 32 subdomains. Compared to the 41, 438 triangles in DhE , the largest subdomain
scope has only 4, 278 triangles.
Based on the same partition of the CFD grid, timing information is reported in
Table 4.1 for the two numerical tests using 32 MPI processors, in which each processor
operates on one subdomain. More specifically, the minimum, maximum, and average
CPU time consumed by the 32 processors are shown.
As expected, Algorithm 4.2 is slower than Algorithm 4.1. However, it is
shown in Section 6.1 and 6.3 that in a fluid-structure interaction simulation, interface
tracking using either algorithm consumes only a small fraction ( usually less than 5%
of the total CPU time.
4.5.2
4.5.3
In this example, an open surface discretized with 650 grid points and 1, 152 triangular
elements (see Figure 4.14) is embedded in a non body-conforming CFD grid with
2, 086, 997 grid points and 12, 462, 983 tetrahedra (see Figure 4.15 and Figure 4.16).
It represents the surface of a pair of ultra-thin triangular wings designed for studying
the unsteady aerodynamics of flapping wing MAVs. The objective here is to verify
the fluid-structure interface tracking algorithms, whereas an aeroelastic simulation
using this model is presented in Section 6.3 in details.
Because the embedded discrete interface is now an open surface, an outward normal direction to each triangular element, required by Algorithm 4.1 as an input,
can not be defined. As a result, Algorithm 4.1 is not capable of handling this
example. Figure 4.17 displays the intersecting edges, shown as gray bars, identified
by Algorithm 4.2 near a wing tip. The embedded discrete interface is shown in
red, while the status of the end-points of the intersecting edges is also color coded.
The reader can observe that all end points of intersecting edges share the same green
color, which indicates that these grid points are correctly identified as residing in the
same flow medium (air).
Figure 4.8: The AGARD 445.6 wing: embedded discrete interface (top) and a cutview
at z = 0 of inviscid non body-conforming CFD grid Dh (bottom).
Figure 4.9: The AGARD 445.6 wing: fluid-structure intersecting edges found by
Algorithm 4.1 (Left) and Algorithm 4.2 (Right).
Figure 4.10: Tracking an embedded AGARD wing: scope decomposition of the embedded discrete interface for 32 fluid computational subdomains (each color designates
a relevant component of the interface for a specific subdomain for which a bounding
box hierarchy is computed on the assigned CPU).
Figure 4.12: A circular cylinder: DhE (magenta color, top) and a cut-view of the non
body-conforming CFD grid at z = 0 of Dh (bottom).
Algorithm 1
Algorithm 2
y
z
81 mm
(0,0,0)
140 mm
140 mm
Figure 4.15: A pair of ultra-thin triangular wings: computational fluid domain and
the embedded discrete interface.
Figure 4.16: A pair of ultra-thin triangular wings: non body-conforming CFD grid
cutview at y = 5 mm (left) and cutview at z = 0 mm (right).
Figure 4.17: A pair of ultra-thin triangular wings: intersecting edges and node statuses identified by Algorithm 4.2.
Chapter 5
Enforcing the Fluid-Structure
Transmission Conditions
5.1
Introduction
88
motionless obstacles (for example, see the review paper [74]). In this context, recently
proposed algorithms for interface treatment have focused either on some form of interpolation [55] with particular attention to numerical stability [56] or higher-order
accuracy [20, 55, 57], or on the concept of a ghost cell [58, 59], some variant of the
penalty method [60], and the mirroring technique [61].
This chapter focuses on this aspect of the problem, assuming that all required
information about the location of the interface with respect to the CFD grid is available. (See Chapter 4 for interface tracking algorithms.) The first objective is to
propose a numerical method for discretizing the no-interpenetration condition (2.3.1)
and as a particular case the slip boundary condition in the context of compressible inviscid flows and static and dynamic rigid or flexible embedded interfaces. This
method is a departure from the methods outlined above and related published works
in that it treats the velocity and pressure boundary conditions on the embedded interfaces simultaneously, rather than disjointly. Furthermore, instead of relying for this
purpose exclusively on interpolation or extrapolation, the proposed method enforces
the appropriate value of the fluid velocity at a fluid-structure interface and recovers
the value of the fluid pressure at this interface via the exact solution of local, onedimensional, fluid-structure Riemann problems (see Section 2.4.3 and Appendix A).
The second objective is to present, in the same context, two alternative approaches
for discretizing the equilibrium transmission condition (2.3.1 and computing the generalized and total flow-induced forces and moments on both rigid bodies and flexible
structures. One of them is based on the local reconstruction of the embedded discrete
interfaces. The other one is based on the level set concept. It is particularly attractive
because it rigorously allows the substitution of an embedded discrete interface by a
simpler surrogate, which simplifies computations. All of these proposed algorithms
are applicable independently from the type and topology of the embedding CFD grid.
They are described here in the context of inviscid flows and the finite volume (FV)
method. However, the underlying ideas are equally applicable to viscous flows and
the finite element (FE) method.
5.2
In [25], a FV method based on the exact solution of local, one-dimensional, twophase Riemann problems was proposed for the solution of compressible multi-fluid
problems. Here, the idea of this method is adapted for the solution of fluid-structure
interaction problems. More specifically, it is tailored to the enforcement of the nointerpenetration condition (2.15) at a wall boundary surface and the recovery of the
value of the fluid pressure at this boundary surface.
5.2.1
Algorithm
Let the superscripts L and R designate the grid points on the left and right sides
of a small region of a discrete interface DhE embedded in a given Eulerian CFD grid
Dh . If ViL denotes a first-layer grid point on the left side of DhE , then ViR denotes
a vertex on its right side that is connected to ViR by an edge ViL ViR traversing DhE
(see Figure 5.1). Let also MiL iR denote the point where ViL ViR intersects the control
volume boundary facet
CiL iR = CiL CiR .
(5.1)
Consider first the case where the flow occurs only on one side of the region of
interest of DhE for example, the left side. Then, the following four-step algorithm
is proposed for treating a fluid-structure interface in an embedded boundary method.
Algorithm 5.1
For each control volume CiL and for each edge ViL ViR intersecting the embedded
discrete interface DhE :
1. If the considered region of DhE corresponds to a dynamic embedded interface,
F
W
) = 0
+
(W
t
s
(s, 0) = W
iL , if s 0
W
v(v0 t, t) = v0 , 0 t t
(5.2)
=W
(s, t) = (, v, E)T .
W
(5.3)
where
~ is defined by
origin at the point MiL iR (see Figure 5.1), and the flux function F
~ W
) = v, v 2 + p, (E + p)v
F(
T
(5.4)
iL
iL =
,
viL
(5.5)
where viL = ~viL ~niL iR is the normal component of the fluid velocity at ViL and
EiL = iL eiL + 21 (viL )2 . Finally,
v0 = u M ~nEM
(5.6)
is the normal component of the structural velocity at the point MiL iR and is
assumed to be constant for 0 t t, where t denotes the computational
time-step. Therefore, as far as this one-dimensional problem is concerned, the
fluid-structure interface is located at s0 (t) = v0 t at any time t [0, t].
The exact solution of the above one-sided, one-dimensional Riemann problem,
which is discussed in details in Section 2.4.3, contains a constant (in time) state
at the fluid-structure interface which is denoted here by
W = ( , v , E )T .
(5.7)
as follows
LM =
L
~vM
= v ~nEM
+ ~viL (~viL ~nEM )~nEM
pLM = p
pLM
( 1)LM
1 L
L
= LM eLM + ~vM
~vM
2
eLM =
L
EM
L
~ h ) ~niL iR d = F (WiL , WM
F(W
, EOSL , ~niL iR ).
iL iR
(5.8)
CiL iR
and the pressure value pLM , and (2) a second time for each control volume CiR and each
edge ViR ViL intersecting DhE with the left one-dimensional fluid-structure Riemann
R
problem replaced by its right counterpart to compute instead FiR iL (WiR , WM
, EOSR , ~niR iL )
and pR
M . This procedure does not require the same equation of state for flows on two
SE
nEM
s
CiL
ViL
MiLiR
@CiLiR
niLiR Vi
CiR
Figure 5.1: Two control volumes on the left and right sides of a region of an embedded
discrete interface (two-dimensional case, quadrilateral mesh).
sides of the interface. In fact, the flows can be modeled by the same equation of state
with the same parameters, the same equation of state with different parameters, or
different types of equations of state.
Remark. It is noted that Assumption 5.1 described above introduces a firstorder geometric (and therefore spatial) error in the computation of the numerical flux
function across an embedded interface and the recovery of the value of the pressure
field on that surface. Extension of this algorithm to achieve second-order accuracy is
possible and in progress [77], but considered outside the scope of this thesis.
5.2.2
Algorithm 5.1 can be seamlessly adapted to any time-integration scheme. Its implementation in a conventional FV-based CFD solver is straightforward, provided that
computational tools are identified for: (a) determining the status of a grid point, and
(b) computing the intersection of the CFD grid and the embedded discrete interfaces
of interest. (See Chapter 4 for applicable computational algorithms.)
For the computational framework presented in Chapter 3, the implementation of
Algorithm 5.1 is performed essentially by modifying the flux computation loop as
follows. For each control volume boundary facet Cij associated to edge Iij connecting
grid points Vi and Vj , the intersection information associated to Iij is first inspected.
If it does not intersect the embedded discrete interface (DhE ), the calculation of
the numerical flux function Fij (3.4) is not altered. More specifically, if Vi and Vj
reside in the same fluid medium (Figure 5.2, Case 1), the usual single-phase flux
function (eg. Roes flux [81]) is applied. If Vi and Vj reside in different fluid media,
which indicates a fluid-fluid interface in between (Figure 5.2, Case 2), the two-phase
Riemann solver based approach [25, 26] is applied.
On the other hand, if Iij does intersect DhE , one or two fluid-structure interface
fluxes Fij are calculated as Eq. (5.8). More specifically, if fluid occurs only on one
side of the interface (Figure 5.2, Case 3), Fij is only calculated for the fluid side.
If fluid occurs on both sides of interface (Figure 5.2, Case 4), two interface fluxes,
namely iFij and jFij , are calculated independently for each side.
fluid-fluid interface
fluid 1
fluid 1
I ij
fluid 1
Cij
Cij
Case 1:
no intersection, same fluid
Case 2:
no intersection, different fluids
fluid-structure interface
fluid 1
covered by
structure
I ij
Cij
Case 3:
fluid-solid
fluid 2
I ij
fluid-structure interface
fluid 1
fluid 2
I ij
Cij
Case 4:
fluid-shell-fluid
Figure 5.2: Fluid media and material interfaces near a control volume boundary facet
(Cij ): different cases.
5.2.3
In this section, the present method for enforcing the no-interpenetration condition is
verified on three one-dimensional academic problems involving either one fluid interacting with a solid body, or two different fluids interacting with a thin-shell structure in between. The structural motion is prescribed, which makes these problems
particularly favorable for testing the numerical treatment of the no-interpenetration
transmission condition. Indeed, in this case the numerical techniques employed for
the approximate solution of the structure system and the enforcement of the equilibrium transmission condition are not involved, thus would not influence the solution.
Although these benchmark problems are one-dimensional by nature, they are solved
on three-dimensional CFD grids using the 3D implementation of the computational
framework summarized in Chapter 3, which is also used in the realistic applications
presented in Chapter 6.
The first two examples involve different fluid media separated by a thin-shell structure with a prescribed constant velocity. Each fluid medium is initialized by a uniform
state. Consequently, these problems are essentially fluid-structure Riemann problems
for which analytical solutions are available (see Section 2.4.3 and Appendix A for the
detailed solution procedure). The analytical solutions are used to verify the numerical approximations obtained from a 3D CFD grid with 800 grid-points in the length
direction. Finally, an order of spatial accuracy is assessed in the third example which
involves one fluid medium and a solid structure body undergoing prescribed harmonic
motion.
5.2.3.1
This one-dimensional example involves two fluid media, namely water and air, separated by a rigid wall which moves at a constant speed. Located to the left(x
direction) of the wall, water is modeled as stiffened gas with = 7.15 and Pc =
2.89 108 Pa. Located to the right(+x direction) of the wall, air is modeled as perfect
gas with = 1.4. Initial conditions are set to l = 1, 000 kg/m3 , vl = 0.0 m/s, and
Pl = 3.0 107 Pa for water and r = 1.3 k/m3 , vr = 0.0 m/s, and Pr = 1.0 105 Pa
for air. The rigid wall, initially located at the origin, moves to the right (i.e. towards air) with a constant velocity vwall = 10.0 m/s, which leads to a shock wave
traveling in air and at the same time a rarefaction/expansion wave traveling in water. The thickness of this rigid wall is assumed to be zero. As discussed in Section 2.4.3, analytical solution to this problem is available. More specifically, the
shock wave in air travels at speed s = 334.2 m/s, leaving behind it a uniform flow
characterized by r = 1.3401 kg/m3 , vr = 10.0 m/s, and Pr = 1.0434 105 Pa.
The rarefaction/expansion fan traveling in water extends from x1 = 1.5102 t m to
x2 = 1.4695 t m for any time instance t > 0. A uniform flow is left between x2 and
the structure, with l = 993.3 kg/m3 , vl = 10.0 m/s, and Pl = 1.5101 107 Pa.
Figure 5.3: A 3D non body-conforming CFD grid for the simulation of 1D fluidstructure problems.
The computational domain for the three-dimensional simulation is a square tube
with dimensions set to 15.0 m x 5.0 m, 0.05 m y 0.05 m, and 0.05 m
z 0.05 m. It is discretized by a non body-conforming CFD grid with 12, 800 grid
points and 43, 146 tetrahedral elements (Figure 5.3). In the vicinity of the fluidstructure interface, the fluid-structure Riemann problem based approach presented in
Section 5.2.1 is applied to enforce the no-interpenetration condition which degenerates
to a moving wall boundary condition in this example. Away from the interface, the
second-order semi-discretization approach presented in Section 3.3 is applied. The
Analytical
Numerical
Density (kg/m3)
1000
800
600
400
200
0
15
10
5
x (m)
1001
1.35
Analytical
Numerical
1000
Analytical
Numerical
1.34
Density (kg/m3)
Density (kg/m3)
999
998
997
996
1.33
1.32
1.31
995
1.3
994
993
12
11
10
x (m)
1.29
1
3
x (m)
Figure 5.4: A one-dimensional stiffened gas thin shell perfect gas problem:
analytical and numerical solutions for the density field at time t = 0.007 s. Top: the
entire computational domain; bottom-left: the vicinity of the rarefaction; bottomright: the vicinity of the shock.
The analytical and numerical solutions at time t1 = 0.007 s are shown in Figure 5.4, 5.5, and 5.6 for fluid density, velocity, and pressure fields, respectively.
Three snapshots are provided for each field. The first one displays the solution over
the entire computational domain, whereas the other two highlight the rarefaction in
water and the shock in air, respectively. It can be observed that for all three fields,
the numerical solutions agree very well with the analytical solutions. In particular,
Analytical
Numerical
12
Velocity (m/s)
10
8
6
4
2
0
15
10
Analytical
Numerical
12
10
10
6
4
Analytical
Numerical
6
4
2
0
12
12
Velocity (m/s)
Velocity (m/s)
5
x (m)
11
10
x (m)
3
x (m)
Figure 5.5: A one-dimensional stiffened gas thin shell perfect gas problem:
analytical and numerical solutions for the velocity field at time t = 0.007 s. Top: the
entire computational domain; bottom-left: the vicinity of the rarefaction; bottomright: the vicinity of the shock.
for both water and air, the relative numerical error at the interface is 0 for velocity,
and less than 2 105 for pressure and density. The rarefaction in water and the shock
in air are also properly located.
5.2.3.2
x 10
3
Analytical
Numerical
Pressure (Pa)
2.5
2
1.5
1
0.5
0
15
10
5
x (m)
3.5
x 10
1.06
Analytical
Numerical
x 10
Analytical
Numerical
1.05
3
Pressure (Pa)
Pressure (Pa)
1.04
2.5
1.03
1.02
1.01
1
1.5
0.99
1
12
11
10
x (m)
0.98
1
3
x (m)
Figure 5.6: A one-dimensional stiffened gas thin shell perfect gas problem:
analytical and numerical solutions for the pressure field at time t = 0.007 s. Top: the
entire computational domain; bottom-left: the vicinity of the rarefaction; bottomright: the vicinity of the shock.
109 Pa, k2 = 7.15, 0 = 1, 000 kg/m3 , and P0 = 3.0 107 Pa. The analytical solution
is still available, and a detailed solution procedure is described in Section 2.4.3 and
Appendix A. It is noted that although the Tait EOS assumes an isentropic flow, it
still retains some validity for flows with weak shocks, and more generally does not
prevent the development of shocks in the solution of Euler Equations. The analytical
solution for air is identical to the one in the previous example. As for water, the
solution is characterized by a rarefaction fan expanding in water with left and right
Analytical
Numerical
Density (kg/m3)
1000
800
600
400
200
0
15
10
5
x (m)
1001
1.35
Analytical
Numerical
1000
Analytical
Numerical
1.34
Density (kg/m3)
Density (kg/m3)
999
998
997
996
1.33
1.32
1.31
995
1.3
994
993
12
11
10
x (m)
1.29
1
3
x (m)
Figure 5.7: A one-dimensional barotropic liquid thin shell perfect gas problem:
analytical and numerical solutions for the density field at time t = 0.007 s. Top: the
entire computational domain; bottom-left: the vicinity of the rarefaction; bottomright: the vicinity of the shock.
Analytical
Numerical
12
Velocity (m/s)
10
8
6
4
2
0
15
10
10
Velocity (m/s)
10
6
4
4
2
0
11
10
x (m)
Analytical
Numerical
12
12
Analytical
Numerical
12
Velocity (m/s)
5
x (m)
3
x (m)
Figure 5.8: A one-dimensional barotropic liquid thin shell perfect gas problem:
analytical and numerical solutions for the velocity field at time t = 0.007 s. Top: the
entire computational domain; bottom-left: the vicinity of the rarefaction; bottomright: the vicinity of the shock.
5.2.3.3
x 10
3
Analytical
Numerical
Pressure (Pa)
2.5
2
1.5
1
0.5
0
15
10
5
x (m)
3.5
x 10
1.06
Analytical
Numerical
x 10
Analytical
Numerical
1.05
3
Pressure (Pa)
Pressure (Pa)
1.04
2.5
1.03
1.02
1.01
1
1.5
0.99
1
12
11
10
x (m)
0.98
1
3
x (m)
Figure 5.9: A one-dimensional barotropic liquid thin shell perfect gas problem:
analytical and numerical solutions for the pressure field at time t = 0.007 s. Top: the
entire computational domain; bottom-left: the vicinity of the rarefaction; bottomright: the vicinity of the shock.
structure is prescribed by
x(t) = r (1 cos(2t)) ,
= 100 Hz
in this example.
The extended computational domain, including the region occupied by the solid
body, is a square tube with dimensions set to 4.0 m x 0.5 m, 0.225 m
y 0.225 m, and 0.225 m z 0.225 m. Five CFD grids are created with 40,
80, 160, 320 and 640 grid points in the length direction. The first two grids are
displayed in Figure 5.10. The fluid-structure interface, located initially at x = 0 m,
is a square plate perpendicular to the length direction (x-axis). It is discretized with
2 triangular elements. The fraction of the extended computational domain to the
right(+x direction) of this interface is considered inactive in the fluid computation
as it is occupied by the structure body.
Figure 5.10: A one-dimensional perfect gas solid body problem: two non bodyconforming CFD grids with 40 and 80 grid points in length-direction.
The second-order semi-discretization method presented in Section 3.3 is applied to
the active part of the fluid computational domain except the direct neighborhood of
the interface where the numerical fluxes are calculated using the numerical algorithm
presented in Section 5.2.1 (Algorithm 5.1), which is locally first-order accurate.
The explicit second-order Runge-Kutta scheme presented in Section 3.3 is employed
to integrate the fluid system from t0 = 0 s to final time t1 = 0.01 s.
The predicted fluid pressure field at time t1 = 0.01 s is reported in Figure 5.11
for all five non body-conforming CFD grids. It can be observed that the solution
converges as the CFD grid is refined. An accuracy analysis is also performed using
the same pressure field, while the solution from the finest grid (640 grid points in
length direction) is considered as the reference. The relative errors in vector 1, 2, and
infinity norms are reported in Figure 5.12. The order of accuracy is 1.10 in infinity
norm, 1.40 in 2-norm, and 1.46 in 1-norm. This is reasonable given that Algorithm
5.1 introduces a first-order geometric error at the fluid-structure interface.
1.2E+05
40 nodes
pressure (Pa)
1.1E+05
80 nodes
160 nodes
1.0E+05
320 nodes
640 nodes
9.0E+04
8.0E+04
-4 -3.5 -3 -2.5 -2 -1.5 -1 -0.5
x (m)
0.5
Figure 5.11: A one-dimensional perfect gas solid body problem: pressure field in
at time t = 0.01s, computed on five non body-conforming CFD grids.
1.0E+00
slope= 1.10
relative error
1.0E-01
1-norm
slope = 1.40
2-norm
1.0E-02
inf-norm
slope = 1.46
1.0E-03
1.0E-04
1.0E-02
1.0E-01
dx (m)
1.0E+00
Figure 5.12: A one-dimensional perfect gas solid body problem: relative error in
the pressure field at time t = 0.01s.
5.3
5.3.1
Introduction
fiF
X
j
Z
cij
(ph )~nF Dj d
Fj
or
fiF
X
j
Z
cij
ph~nS Dj d,
Sj
(5.9)
where fiF is the local fluid-induced load at node Vi of the FE structural model of
interest, cij are some scalar coefficients, and Dj are some shape functions with support
on Ft j or Sj (see Figure 5.13). The two formulas in Eq. 5.9 correspond to evaluating
the load using the discretization of the interface either in the CFD grid or in the CSD
grid.
Dh
Vj
Cj
nF
Ci
Vi
WS
SE = S F
Figure 5.13: Spatial discretization of a two-dimensional fluid computational domain
using a body-conforming CFD grid with triangular elements.
However, as illustrated in Figure 4.1, for a non body-conforming CFD grid, Dh
does not contain a surface discretization corresponding to the wall boundary in a
fluid problem or the fluid-structure interface in a fluid-structure interaction problem.
Consequently, the CFD solver based on an embedded boundary method does not
have explicit representations of the surfaces where the flow-induced load needs to
be computed. On the other hand, in a fluid-structure interaction problem, the FE
structural model contains a discrete representation S of flexible wall boundaries but
the structural solver does not have direct access to the pressure field computed by
the flow solver. To resolve this issue, two consistent and conservative approaches are
proposed in this section for computing in an embedded boundary method the localized
flow-induced force loads. The first approach reconstructs the interface within the
CFD solver. The second one represents it as a zero level set, and works instead with
a surrogate fluid-structure interfaces obtained simply by joining contiguous segments
of the boundary surfaces of the fluid control volumes that are the closest to the this
zero level set.
5.3.2
N
[
q .
(5.10)
q=1
For a flexible wall boundary, E can be in practice S or another discrete representation of this wall boundary characterized by a different (for example, finer) mesh
resolution. A key observation, particularly in the context of this work, is that the
evaluation of the fluid state vector at an intersection of the embedded discrete interface and the CFD grid is simpler than its evaluation at a vertex of that surface.
Indeed, at the intersections of E and Dh , W can be and in this work, is already
determined from the solution of local, one-dimensional, one-sided Riemann problems
as explained in Section 5.2. On the other hand, the evaluation of W at the vertices
of E requires relatively complex interpolations, particularly in three dimensions.
This observation is behind the specific approach described next for reconstructing an
embedded discrete interface within a given CFD grid.
5.3.2.1
After the points Ik defined by the intersections of E and the edges (both in two and
three dimensions) of Dh have been computed for example, using either Algorithm
4.1 or Algorithm 4.2 presented in Chapter 4 Algorithm 5.2 outlined below
e E of a collection of
reconstructs the embedded discrete interface E as the union
triangles q obtained by connecting the points Ik that is,
eE =
N
[
q .
(5.11)
q=1
The choice of a triangle for eq is based on the fact that a tetrahedron, prism, pyramid,
or hexahedron all of which can be encountered in a CFD grid usually intersects a
surface in at least three non colinear points, and the simplest approach for connecting
more than three points is to use triangles.
Algorithm 5.2
For each element Dhe in Dh :
1. Inspect the status of the vertices of Dhe .
2. If all status are positive or all of them are negative, Dhe does not cross E .
E that
3. Otherwise, Dhe crosses E and contributes one or more elements eq to
are constructed as follows.
(a) Initialize a list Le of intersection points.
(b) Loop on the edges of Dhe : for each edge intersecting E , identify the intersection point Ik and add it to Le .
(c) Connect all points of Le to form one or more elements q .
As illustrated in Figure 5.14 which focuses on the two-dimensional case for the
e E is in general a geometrically
sake of clarity, the reconstructed discrete interface
accurate approximation of the embedded discrete interface E , provided that the
CFD grid has at least the same mesh resolution as E , which is usually the case.
nEq
xk
Ik
SE
tq SE
SE
tq SE
5.3.2.2
For simplicity, but without any loss of generality, assume next that the embedded discrete interface E is the discretization of the wet surface of the structure of interest,
S . For each computed intersection point Ik , the id of the intersecting element of E ,
q , and the natural coordinates of Ik in q , kq (see Figure 5.14), are collected. This information can be exploited to effectively interpolate at each point Ik the displacement
e E as follows
and velocity of the reconstructed discrete interface
u(Ik ) =
and
k) =
u(I
Njq (kq )u j ,
(5.12)
where Njq is the FE shape function associated with node Vjq of the element (or face)
q containing the point Ik , the summation in (5.12) is carried over all relevant shape
functions of q , these shape functions satisfy the partition of unity property. uj
and u j are the displacement and velocity of E at node Vjq of element (or face) q ,
respectively.
Let pk and vk denote the computed values of the fluid pressure and velocity vector
at the interface points Ik . In this work, these values are computed by solving local,
one-dimensional, fluid-structure Riemann problems as explained in Section 5.2.1. On
e E , the fluid pressure and velocity fields can be
the reconstructed discrete interface
approximated at each point ~x q as follows
ph (~x) =
3
X
q (~x) pk ,
N
k
and
~v (~x) =
k=1
3
X
q (~x) vk ,
N
k
(5.13)
k=1
(5.14)
E connected to Ik .
where ~n
qE is the normal to an element q of
Let NE denote the total number of vertices in E . Following the methodology
developed in [47] for computing the generalized flow-induced load vector in the context of the ALE framework where the CFD grid remains body-conforming at all
time-instances, the generalized flow-induced load vector at node Vi of the FE structural model, fiF , is computed here by applying the virtual power principle at the
fluid/structure interface.
More specifically, in this case the interface is approximated by the reconstructed
e E . Using the symbol to denote a virtual quantity, the virtual
discrete interface
e E can be written as
power of the fluid pressure force acting on
P
XZ
q
" 3
XZ X
q
q
E ~vh (~x) d
ph (~x) ~n
#
3
X
q
q
(~x) vk d
(~x) pk ~n
E
N
k
k=1
" 3
XZ X
q
q
N
k
k=1
q
N
k
#
3
hX
i
X
q (~x)
Njq (kq ) u j d.(5.15)
(~x) pk ~n
qE
N
k
k=1
k=1
The second equation of Eq. 5.15 is obtained via Eq. 5.13, while the last one is obtained
using Eq. 5.12 and the discretized transmission condition Eq. 5.14.
On the other hand, the virtual power of the finite element structure forces and
moments acting on S = E can be written as
P =
NE
X
i=1
fiF u i .
(5.16)
(5.17)
Eqs. (5.15), (5.16), (5.17) above and the partition of unity property of the shape
q lead to the following algorithm for computing the generalized
functions Niq and N
k
flow-induced load vector at node Vi of the given FE structural model and its resultant.
Algorithm 5.3
fiF
Z X
3
q /Ik
q
and
Ik (q 3Vi )
3
X
q
q (~x) pk ~n
q (~x) N q ( q ) d,
N
N
i
E
k
k
k
q k=1
(5.18)
k=1
and
NE
X
fiF
i=1
q k=1
3
X Z X
eE
q
i=1
3
X Z X
eE
q
NE
X
q k=1
3
X Z X
eE
q
q /Ik
q
and
Ik (q 3Vi )
Z X
3
3
X
q
q (~x) pk~n
q (~x) N q ( q ) d
N
N
i
E
k
k
k
q k=1
k=1
NE
3
X
X
q
q
q
~
Nk (~x) pk n
E
Nk (~x)
Niq (kq ) d
k=1
i=1
3
X
q
q (~x) pk ~n
q (~x) d
N
N
E
k
k
k=1
q (~x) pk ~n
N
qE d
k
(as expected).
(5.19)
q k=1
Remark: If the embedded discrete interface E does not coincide with the discretization of the wet surface of the structure of interest, S , the derivation of Algorithm 5.3 presented above is extended as follows. First, the projection of E onto
S is computed only once in a pre-processing step, and the natural coordinates of
(23)
fAF = A + A
(5.20)
p1
I1
Dh
12
p2
p3
I2
23 I 3
f AF
0
VA
E S
AB
f BF
VB
(23)
fBF = B + B ,
(5.21)
(ij)
where scalars fAF and fBF denote the y-component of fAF and fBF , respectively. S
Z
=
(12)
1(12) (x) p1 + N
2(12) (x) p2 N
1
2(12) (x) NA (I2 )
N
(x) NA (I1 ) + N
12
(5.22)
(23)
A
Z
=
(23)
(23)
(23)
(23)
(x) p2 + N3
(x) p3 N2
(x) NA (I2 ) + N3
(x) NA (I3 )
N2
23
(5.23)
(12)
B
Z
=
(12)
1(12) (x) p1 + N
2(12) (x) p2 N
1
2(12) (x) NB (I2 )
N
(x) NB (I1 ) + N
12
(5.24)
(23)
B
Z
=
2(23) ((x))p2 + N
3(23) ((x))p3
N
2(23) ((x))NB (I2 ) + N
3(23) ((x))NB (I3 )
N
23
(5.25)
In practice, fiF , i = 1, ..., N S can be evaluated by looping through the elements of
e E . For each q
e E , its contribution to the pressure load on each relevant structure
(12)
(23)
and B
(12)
and B
are
and (5.25).
The equivalence of Eqs. (5.17) and Eq. (5.18) is demonstrated as follows. The
virtual power of the finite element structure forces and moments acting on E can be
written as
(12)
A u A
(23)
(12)
B u B
(23)
A u A + B u B
((12) )
((23) )
(5.26)
be written as
(12)
(23) =
Z12
1(12) ((x))p1 + N
2(12) ((x))p2
N
1(12) ((x))~v1 + N
2(12) ((x))~v2 dx,
N
2(23) ((x))p2 + N
3(23) ((x))p3
N
2(23) ((x))~v2 + N
3(23) ((x))~v3 dx.
N
23
(12)
Z
=
(23)
Z
=
fBF u B
Z
=
Z
(ph (x)) vh (x)dx +
12
(ph (x)) vh (x)dx .
(5.27)
23
Algorithm 5.3 proposed above for computing the generalized flow-induced load
vector is conservative in the sense that the total work produced by the action of the
e E is equal and
fluid on the structure at the reconstructed fluid/structure interface
opposite to the total work produced by the reaction of the structure on the fluid at
eE.
Furthermore, for a constant pressure field p = p? , it follows from Eq. (5.18) and
q that the resultant
the partition of unity property of the shape functions Niq and N
k
fiF
= p
i=1
NE
X
i=1
= p
q /Ik
q
and
Ik (q 3Vi )
X Z
eE
q
= p
~n
qE
X Z
eE
q
~n
qE
X Z
eE
q
= p?
~n
qE
3
X
q (~x) N q ( q ) d
N
i
k
k
k=1
3
X
NE
X
q
Nk (~x)
Niq (kq ) d
k=1
i=1
3
X
q (~x) d
N
k
k=1
~n
qE d.
(5.28)
Hence, for a constant pressure field p = p? and a closed embedded discrete interface
E ,
NE
X
i=1
fiF
=p
X Z
eE
q
~n
qE d = 0
(5.29)
as it should be. Therefore, Algorithm 5.3 for computing the generalized flowinduced load vector is also consistent in the sense that it preserves the vanishing
property of the exact integration of a constant pressure field over a closed surface.
5.3.2.4
Accuracy analysis
first-order geometric error in both the computation of the numerical flux function
across an embedded interface and the pressure field on this interface, Algorithm
5.3 delivers a load computation scheme that is locally second-order space-accurate
and globally first-order space-accurate. Hence, improving the global accuracy of this
algorithm requires developing first a higher-order extension of Algorithm 5.1 proposed in Section 5.2 for treating a fluid-structure interface in an embedded boundary
method. Work on this extension is in progress (see [77]) but considered outside the
scope of this thesis.
5.3.2.5
Practical implementation
e E needs be reconstructed only locally, and Algorithm 5.2 and AlgoIn practice,
rithm 5.3 can be combined and implemented as follows for computing the generalized
load vector (5.18).
For each element Dhe in Dh :
1. Inspect the status of the vertices of Dhe .
2. If all status are positive or all of them are negative, Dhe does not cross E .
E that
3. Otherwise, Dhe crosses E and contributes one or more elements eq to
are constructed as follows.
(a) Initialize a list Le of intersection points.
(b) Loop on the edges of Dhe : for each edge intersecting E , identify the intersection point Ik and add it to Le .
(c) Connect all points of Le to form one or more elements q .
(d) Compute the contributions of the elements q to the generalized flowinduced load vectors fiF (5.18), for all Vi (q 3 (Ikq q ))
5.3.3
The first approach for load computation on embedded interfaces presented in Section 5.3.2 is practical even when the embedding CFD grid is unstructured, or made
of other elements besides hexahedra, thanks to the widely available computational
geometry tools, particularly from the computational graphics community. As demonstrated in Section 6.2, it even delivers a good CPU performance, despite the fact that
it relies on an explicit but local reconstruction of the embedded discrete interface.
Nevertheless, a second approach for load computation on embedded discrete interfaces is presented here. This alternative approach bypasses any reconstruction and
some of the complex computational geometry operations it entails.
5.3.3.1
points or cell centers of the computational domain. Hence, the idea here is to ease the
computation of a generalized flow-induced load vector or its resultant in an embedded
b E where all
boundary method by: (1) choosing a convenient surrogate interface
data forming the integrands of the integrals to be evaluated for load computation
are either readily available at this stage of a flow analysis or easy to obtain, and (2)
convert any integral over the real discrete interface E into an equivalent integral
b E or any other representation
e E of
b E where load computation
over its surrogate
is even easier. These two components of the idea developed here are discussed below,
in reverse order.
Let ~x = (x1 , x2 , x3 ) F denote a point in the computational fluid domain, and
let (~x) denote the distance from ~x to the embedded discrete interface E . Hence,
E is characterized by = 0 for ~x E . Moreover, using the level set function the
unit outward normal to E at a point ~x can be expressed as:
where
kx (~x)k2 = 1.
(5.30)
b E ~y = (y1 , y2 , y3 ) = (~x) E .
~x = (x1 , x2 , x3 )
(5.31)
In an ideal setting such as that graphically depicted in case (a) of Figure 5.16, is a
one-to-one mapping that can be expressed as
(5.32)
~
e E x = (x1 , x2 , x3 ) = ()
bE.
~ = (1 , 2 )
(5.33)
Let also
i
G = =
j
and
i=1, ,3;j=1,2
i
F = x =
xj
.
(5.34)
i,j=1, ,3
The computation of the matrix G is standard in any FE code. On the other hand,
the level set function can be used to compute the matrix F , assuming that is twice
differentiable. Indeed,
2
i
(~x) = ij
(~x)
(~x) (~x)
(~x),
xj
xj
xi
xj xi
(5.35)
where ij denotes the Kronecker delta, and using Eq. (5.30) and the notation
2
H(~x) =
(~x)
xj xi
,
(5.36)
i,j=1, ,3
F can be written as
(5.37)
p
b
p (~x) det[F T F ] d,
(5.38)
p
~
det[GT G] d1 d2 ,
p ()
(5.39)
!
p
~
p ()
det[GT F T F G] d1 d2 .
(5.40)
p(~y ) d =
E
bE
Z
p(~x) d =
bE
eE
and therefore
Z
Z
p(~y ) d =
eE
b E is conThe result (5.38) above reveals that a surrogate embedded discrete interface
venient if: (a) p (~x) is either readily available on this interface or can be obtained
there by simple computational means, (b) the associated mapping and corresponding gradient matrix F can be efficiently evaluated on this interface, and of course,
(c) E is explicitly contained in the discretization Dh . Even when such a surrogate
interface is identified, it may be even simpler to perform the load computation on a
e E for
b E , in which case the result (5.40) is used instead of
reference configuration
its counterpart (5.38).
The following academic example illustrates the application of the result (5.40)
b E is chosen as the projection
derived above. In this example, E is a triangle and
b E is also
of E onto a plane that does not contain (Figure 5.17. Hence in this case,
a triangle whose vertices are denoted here by Vi , i = 1, , 3, (~x) is a one-to-one
SE
SE
SE
SE
SE
SE
(a)
(b)
missed
region
(c)
Figure 5.16: Three typical situations arising from the choice of a surrogate embedded
b E : (a) an ideal situation, (b) a situation where is not a one-to-one
interface
b E is non smooth
mapping, and (c) a situation where the variation of the normal to
and leads to loss of accuracy.
mapping and more specifically a linear function, and therefore H(~x) = 0 and
F = F = I nE nE = det[GT F T F G] = det[GT F T F G] = det[GT F G],
(5.41)
where
GT F G = GT (I ~nE ~nE )G = GT G (GT ~nE ) (GT ~nE ).
(5.42)
Let ~g1 = V1 V2 and ~g2 = V1 V3 , let denote the reference right triangle with two edges
~ = 11 2 , 2 ()
~ = 1 ,
of unit length and a hypotenuse of length = 2, and let 1 ()
~ = 2 be the usual linear shape functions associated with a triangle such as
and 3 ()
. In this case,
~ =
()
3
X
~ i,
i ()A
= gj , j = 1, 2,
j
i=1
and
G = (g1 g2 ).
(5.43)
If |
| denotes the area of the triangle and ~n the normal to , then
|
| =
kg1 g2 k
2
(5.44)
and
q
det(GT F T F G) = 2|
||~n ~nE |.
(5.45)
y (x )
V1
V3
x ( )
V2
2
Z
~
~
f (~y ) d = 2|
| f () |~n ~nE | d1 d2 .
(5.46)
Z
d = 2|
| |~n ~nE | d1 d2 ,
(5.47)
| | = |
||~n ~nE |,
or equivalently,
|
| =
| |
,
|~n ~nE |
(5.48)
which is the well-known formula for the exact evaluation of the area of the projection
of a triangle onto a plane. Hence, the academic example presented above highlights
the potential of the results (5.38)(5.40) for load computation on a surrogate discrete
interface.
However, given a discretization Dh of the fluid domain of interest, finding a surb E that is convenient in the sense defined above and for which a
rogate interface
one-to-one mapping exists is not always possible. For example, Figure 5.16, case
(b), graphically depicts a situation where a convenient surrogate embedded discrete
interface does not lead to a projector that is a one-to-one mapping. Furthermore,
the computation of the gradient matrix F (5.37) can require computational and implementational efforts that are comparable to those of the reconstruction of E discussed
in Section 5.3.2. For these reasons, the practical exploitation of the results (5.38)
(5.40) calls for introducing some approximations in their computation. These are
suggested by the spatial and temporal characteristics of the underlying CFD method,
among others.
In the context of this thesis, the FV method presented in Section 3.3 for the semidiscretization of the fluid governing equations, and the flux-based method proposed
in Section 5.2.1 for enforcing the no-interpenetration transmission condition (2.15) on
an embedded discrete interface and recovering there the value of the fluid pressure,
suggest choosing as a surrogate fluid/structure interface
bE =
[
(l,m)/Ik
Clm ,
(5.49)
E 6=
where Ik is the edge associated with (or traversing) the boundary facet Clm (see
T
Figure 5.18), and Ik E 6= indicates that edge Ik intersects the embedded discrete
interface E .
Indeed, this interface is practical as it is made of control volume boundary facets
Clm that are explicitly built in a FV code operating on dual control volumes. Two
numerical algorithms are presented in Chap 4 for tracking the embedded discrete
interface, which includes the detection of intersecting edges. After Algorithm 1
has been used for enforcing the no-interpenetration transmission condition (2.15) on
b E , the value of the fluid pressure pk at the intersection of each facet Clm and its
Z
p(~y ) d
b
p(~x) d.
(5.50)
bE
It is notable that the level-set function (~x) is not involved in the evaluation of
the approximate form of the generalized flow-induced load shown in Eq. (5.50).
SE
Vi
xk
q
@Clm
Vl
tq
Ik V
m
SE
tq SE
SE
SE
Following the same conservative approach based on virtual work adopted in Algob E given
rithm 3 for computing the generalized flow-induced load vector, choosing
in (5.49) as a surrogate fluid/structure interface, choosing also to approximate (~x)
by ~x and therefore F by I to obtain the approximation (5.50), and recalling the partition of unity property of the shape functions Niq leads to the following conservative
algorithm for computing the distribution and resultant of the generalized flow-induced
load.
Algorithm 5.4
fiF
(l,m)/
q 3Vi
T
(Ik E 6= & Ik q )
(5.51)
Clm
and
NE
X
fiF
(l,m)/Ik
(Ik
pk~nlm
Clm
(l,m)/
E 6= & Ik q )
Niq (kq ) d
Vi q
X
T
q E
i=1
pk
~nlm d.
(5.52)
Clm
E 6=
From Eq. (5.52) above and the partition of unity property of the shape functions
Niq , it follows that for a constant pressure p = p? ,
NE
X
fiF
=p
i=1
X
(l,m)/l m <0
Z
~nlm d.
(5.53)
Clm
b E is also closed.
From the definition (5.49), it follows that if E is closed, then
Hence, for a constant pressure and a closed embedded surface, the resultant (5.52) is
equal to zero. This proves that Algorithm 4 for computing the generalized flowinduced load vector is also consistent in the sense of preserving the vanishing property
of the exact integration of a constant pressure field over a closed surface.
5.3.3.3
Accuracy properties
p(~x) is a first-order spatial approximation of the pressure on the embedded inb E (or equivalently on
e E ).
terface
The resulting approximation (5.50) is locally second-order space-accurate, and
therefore as accurate, if not more accurate, than the numerical solution delivered
at an embedded interface by a typical embedded boundary method.
Attempting to improve the spatial accuracy of the approximation (5.50) by
incorporating in it a correction factor of the form |~n ~nE | in order to account for
b E onto E would be counterproductive. To begin,
the effect of the projection of
as shown in Eq. (5.45) pertaining to the academic example of Section 5.3.3.1,
such a correction factor can be traced to the exact expression (5.32) of the
mapping and therefore is inconsistent with the approximation (~x) ~x. Furb E is not sufficiently smooth
thermore, as shown in Figure 5.16, case (c), when
as in the case of the surrogate interface (5.49) illustrated in Figure 5.18, prob E onto E results in missing some regions of E during the integration
jecting
process. Not only this degrades spatial accuracy, but it also leads to a load
computation scheme that does not preserve the vanishing property of the exact
integration of a constant pressure field over a closed surface (see Section 5.3.3.2).
It follows that in general, Algorithm 5.4 proposed in this paper for load computation in an embedded boundary method is locally second-order space-accurate,
and globally first-order space-accurate. In other words, it delivers a generalized flowinduced force distribution (5.51) that is second-order accurate in space, and a generalized flow-induced resultant (5.52) that is first-order accurate in space. In the
particular case where is the identity mapping that is, when the proposed surrogate interface is identical to the real interface Algorithm 5.4 becomes locally
5.3.4
In Section 5.3.2 and 5.3.3, it was concluded that in the ideal situation where Assumption 5.1 is satisfied, Algorithm 5.3 for load computation equipped with
Algorithm 5.2 for surface reconstruction is locally third-order space-accurate and
globally second-order space-accurate; in the realistic situation where Assumption
5.1 is not satisfied, Algorithm 5.3 is locally second-order space-accurate and globally first-order space-accurate. On the other hand, it was concluded in Section 5.3.3.3
that Algorithm 5.4 is in principle second-order space-accurate locally and firstorder space-accurate globally; however, it becomes locally third-order space-accurate
and globally second-order space-accurate when the surrogate interface (5.49) becomes
identical to real interface. In this section, these theoretical accuracy results are numerically verified for the two-dimensional manufactured problem described below.
Let = [2, 2] [2, 2] denote a square domain where the pressure field is
assumed to vary as follows
p(x, y) = sin(x + y),
(5.54)
where x and y are measured in an orthonormal frame whose origin is at the center
of . (See Figure 5.19.) Assume that a square ribbon playing here the role of a
closed interface with a side-length l = 2 is embedded in so that the center of
the area it covers coincides with the center of . The total pressure-induced force on
this interface is given by the integral of p(x, y) on the perimeter of this interface and
therefore is equal to
fF =
2(1 cos 2)
2(1 cos 2)
(5.55)
Figure 5.19: A prescribed pressure field in computational domain [0, 2] [0, 2].
To verify the global convergence properties of Algorithm 5.3 (equipped with
Algorithm 5.2) and Algorithm 5.4 applied to the approximation of the above
total force, two different sets of discretizations of are constructed. The first set, S1,
consists of a series of increasingly refined uniform quadrilateral meshes constructed
so that Assumption 5.1 is always satisfied and the surrogate interface (5.49) is
identical to the real interface (the ideal case) that is,the perimeter of the embedded
interface always coincides everywhere with control volume facets of the dual mesh.
For example, Figure 5.20 shows the coarsest mesh in the set S1, together with the
reconstructed interface according to Algorithm 5.2, and the surrogate interface
according to (5.49). The second set of meshes, S2, is a set of arbitrary triangular
meshes for which Assumption 5.1 is not satisfied and the surrogate interface (5.49)
is different from the real one (the realistic case). The coarsest of these meshes is shown
in Figure 5.21 together with the reconstructed and surrogate surfaces according to
Algorithm 5.2 and definition (5.49), respectively.
2
!1
!1
!2
!2
!1
!2
!2
!1
Figure 5.20: Ideal case: coarsest mesh in set S1 (uniform, quadrilateral, satisfying
Assumption 5.1) and reconstructed (left) and surrogate (right) interfaces.
Figure 5.22 and Figure 5.23 report in log-log format the variations with the element
mesh size of the absolute value of the relative error between the approximate and exact
evaluations of the total force f F for the ideal and realistic cases, respectively.
For the ideal case, the results shown in Figure 5.22 reveal that:
Algorithm 5.3 (equipped with Algorithm 5.2 for interface reconstruction)
and Algorithm 5.4 deliver the expected second-order accuracy.
Algorithm 5.4 delivers a significantly better accuracy than Algorithm 5.3.
This is because in this ideal case, the proposed surrogate interface (5.49) coincides everywhere with the real interface (see Figure 5.20).
For the realistic case, the variations of the relative error shown in Figure 5.23 are
!1
!1
!2
!2
!1
!2
!2
!1
Figure 5.21: Realistic case: coarsest mesh in set S2 (arbitrary, triangular, not satisfying Assumption 5.1) and reconstructed (left) and surrogate (right) interfaces.
irregular because of the irregular nature of the considered set of increasingly refined
meshes S2. Nevertheless, they reveal that:
Algorithm 5.3 (equipped with Algorithm 5.2 for interface reconstruction)
delivers on average a convergence rate equal to 1.98, which is significantly higher
than the theoretically determined rate of 1.
Algorithm 5.4 delivers on average 1.33, which is slightly higher than the
theoretically determined rate of 1.
Algorithm 5.3 delivers a significantly better accuracy than Algorithm 5.4.
This is because in the realistic case, the reconstruction of the interface performed
by Algorithm 2 can be expected to be in general a more accurate representation of this interface than the surrogate (5.49) (see Figure 5.21). However, it
should be emphasized that even in this case, both Algorithm 5.3 and Algorithm 5.4 deliver an excellent accuracy given a reasonable mesh resolution.
In summary, all numerical results obtained for the academic problem described
above verify the theoretical results, except in one case where a higher-order of accuracy
is obtained numerically.
10
!1
10
slope = 2 (reference)
Relative error
!2
10
!3
10
!4
10
!5
10
ALGORITHM 4
ALGORITHMs 2 & 3
!6
10
!3
10
!2
!1
10
10
10
Element size
Figure 5.22: Ideal case: performance of Algorithm 5.3 and Algorithm 5.4 for
load computation.
!1
10
!2
Relative error
10
slope = 1 (reference)
!3
10
!4
10
slope = 2 (reference)
!5
10
ALGORITHM 4
ALGORITHMs 2 & 3
!6
10
!2
10
!1
10
10
10
Element size
Figure 5.23: Realistic case: performance of Algorithm 5.3 and Algorithm 5.4
for load computation.
Chapter 6
Applications
In previous chapters, a computational framework based on an embedded boundary method for the simulation of highly nonlinear fluid-structure interactions is introduced, with focus laid on computational algorithms for tracking the embedded
fluid-structure interface (Chapter 4) and enforcing the fluid-structure transmission
conditions (Chapter 5). In this chapter, the main components of this computational
framework are assessed in the applications to fluid-structure interaction problems in
the context of aeronautics, underwater implosions and explosions, and pipeline explosions. These applications involve strong shock waves, complex structural geometry,
large and complex structural deformations, dynamic cracking, and dynamic fluid-fluid
material interfaces.
139
CHAPTER 6. APPLICATIONS
6.1
140
In this section, the problem of computing the unsteady inviscid airflow past a rigid
wing in heaving motion is considered. The wing in question is Wing 445.6, the
AGARD standard aeroelastic configuration for dynamic response. It has a root chord
length Lc = 22.0 in, a semi-span Ls = 30.0 in, a tip chordlength Lt = 14.5 in, and
a quarter-chord sweep angle of 45o . Its panel aspect ratio is equal to 1.65 and its
taper ratio is equal to 0.66. Its airfoil section is the NACA 65A004 with a maximum
thickness of approximately 0.9 in (4% of root chord length). Hence, this wing is quite
thin. Consequently, it is difficult to construct a non body-conforming CFD grid for
this inviscid flow problem which resolves the volume enclosed by the wet surface of
the wing without being unnecessarily too fine for the external flow computations.
The wet surface of the wing is discretized with 20, 721 grid points and 41, 438
triangles (Figure 6.1Left). This discrete representation DhE is embedded in a non
body-conforming CFD grid DhN BF with 105, 030 grid points and 609, 576 tetrahedra
(Figure 6.1Right).
This embedded discrete interface has been considered in Section 4.5 for assessing
the performance of fluid-structure interface tracking algorithms presented in Chapter 4. More specifically, it is shown there that interface tracking Algorithm 4.1
misses large swaths of the structure as it fails to detect double intersections, which
are the intersections along fluid edges that penetrates the interface twice. On the
other hand, Algorithm 4.2 successfully captures all intersection points. (See Figure 4.9.) This difference, illustrated as Case II of Figure 4.5, is essentially due to the
fact that the CFD grid does not resolve the structure everywhere.
CHAPTER 6. APPLICATIONS
141
Figure 6.1: The AGARD 445.6 wing: embedded discrete interface (left) and a cutview
at z = 0 of the non body-conforming CFD grid Dh (right).
The objectives of this section are (1) to investigate the influence of the missed
intersections to flow solutions, and (2) to assess and compare the performance of two
approaches for computing the fluid-induced force load acting on the embedded surface
(see Section 5.3). To this end, the wing is set in a prescribed harmonic heaving motion
characterized by the amplitude ha = 0.05 in and the frequency hf = 500 Hz. The
air flow is modeled by perfect gas with = 1.4. The free-stream conditions (Mach
number, angle of attack, density, and pressure) are set to M = 0.3, = 0deg ,
= 9.357255 108 (lb/in4 ).sec2 , and p = 14.5 psi, respectively. In order to
verify the CFD results generated by the embedded boundary method with different
algorithms for interface tracking and treatment, a body-conforming grid DhBF with
896, 167 grid points and 4, 664, 720 tetrahedra is also generated for the computation
of an ALE reference solution of this problem.
Using the three-point backward difference time-integrator presented in Section 3.3,
four implicit numerical simulations are performed:
1.1. using the CFD grid DhN BF and the embedded boundary method equipped with
Algorithm 4.1 for interface tracking, and the combination of Algorithm
CHAPTER 6. APPLICATIONS
142
5.2 and Algorithm 5.3 for calculating the flow-induced pressure load on the
wing surface.
1.2. using the CFD grid DhN BF and the embedded boundary method equipped with
Algorithm 4.2 for interface tracking, and the combination of Algorithm
5.2 and Algorithm 5.3 for calculating the flow-induced pressure load on the
wing surface.
1.3. using the CFD grid DhN BF and the embedded boundary method equipped with
Algorithm 4.2 for interface tracking, and Algorithm 5.4 for calculating
the flow-induced pressure load on the wing surface.
1.4. using the CFD grid DhBF and the ALE computational framework of AERO-F.
All the simulations outlined above are initialized with a uniform flow at the freestream condition specified above. The obtained time-histories of the total lift are
reported in Figure 6.2 and 6.3 for the first five periods of oscillation (0 sec t
0.01 sec). The lift predicted by simulation 1.4 can be considered as a reference
solution for two reasons: (1) it is computed on a much finer CFD grid (896, 167
grid points in DhBF versus 105, 030 in DhN BF ), and (2) as stated earlier, AERO-Fs
ALE computational framework has been successfully verified and validated for many
aerodynamic and aeroelastic applications in the past. In Figure 6.2, the reader can
observe that the lift predicted by Simulation 1.2 is almost identical to the reference
lift, whereas the one predicted by Simulation 1.1 is less accurate. This is not surprising
given the inaccurate intersection results obtained by Algorithm 1 (see Figure 4.9).
On the other hand, Figure 6.3 shows that Simulation 1.2 and 1.3 predict almost the
same lift history. Hence, at least for this problem, Algorithm 5.3 and Algorithm
5.4 deliver, as expected, the same accuracy.
143
CHAPTER 6. APPLICATIONS
Total lift
400
300
200
200
100
100
100
100
200
200
300
0
0.002
0.004
0.006
Time (sec)
0.008
0.01
300
Lift (lb.f)
Lift (lb.f)
Total lift
400
300
0
0.5
1.5
Time (sec)
2.5
3
3
x 10
Figure 6.2: Thin wing in heaving motion: comparison of the lift time-histories predicted by Simulation 1.1, 1.2, and 1.4.
Table 6.1 reports the CPU performance obtained on 32 processors for simulations
1.1, 1.2, and 1.3. Information provided in the first two columns indicates that Algorithm 4.2 for interface tracking is nearly three times slower than Algorithm
4.1. However, Algorithm 4.1 is not accurate in this case. Furthermore, interface
tracking with Algorithm 4.2 is also found to consume only 12% of the total CPU
time. Indeed, the computational complexity of an implicit flow solver is in general
of the order of O(N log N ), and that of both interface tracking algorithms is of the
order of O(N log NE ), where N and NE denote the number of points in the CFD
grid and number of elements in the embedded discrete interface. In most applications, NE is much small than N because Dh is a three-dimensional entity but DhE
is a two-dimensional one. In this example, the ratio between NE and N is roughly
NE
e =
= 0.39. It is shown in Section 6.3 that for a lower e, the relative cost of
N
interface tracking, compared to the total CPU time, is essentially negligible.
On the other hand, a comparison between the second and third columns of Table 6.1 indicates that the combination of Algorithm 5.2 and Algorithm 5.3
144
CHAPTER 6. APPLICATIONS
Total lift
Total lift
400
300
200
200
100
100
100
100
200
200
300
0
0.002
0.004
0.006
Time (sec)
0.008
0.01
300
Lift (lb.f)
Lift (lb.f)
400
300
0
0.5
1.5
Time (sec)
2.5
3
3
x 10
Figure 6.3: Thin wing in heaving motion: comparison of the lift time-histories predicted by Simulation 1.2, 1.3, and 1.4.
Table 6.1: AGARD wing in heaving motion: CPU performance results on 32 processors.
Sim. 1.1 Sim. 1.2 Sim. 1.3
Flow computations
720 sec
731 sec
730 sec
Finite volume fluxes and Jacobians 135 sec
139 sec
139 sec
Linear system solution (GMRES)
425 sec
427 sec
428 sec
Various other computations
160 sec
165 sec
163 sec
Interface tracking
41 sec
114 sec
114 sec
Load computations
18 sec
17 sec
11 sec
Total simulation time
825 sec
911 sec
905 sec
for force computation is only slightly more computationally expensive than the alternative Algorithm 5.4, despite the fact that Algorithm 5.2 reconstructs the
embedded surface. In either case, evaluating the load accounts for less than 2% of
the total CPU time.
CHAPTER 6. APPLICATIONS
6.2
145
In this section, a verification of the proposed embedded boundary method is performed in the context of a steady, subsonic flow around an aircraft. The complex
geometry of this aircraft indicates that the embedded boundary method can be favorable for studying its aerodynamic of properties, because it operates on non bodyconforming CFD grids. Indeed, generating a high-quality body-conforming CFD grid
around a complex geometry is usually both time-consuming and tedious.
The aircraft in question is Helios Prototype vehicle HP03, a remotely piloted
aircraft designed and developed by NASA for high altitude, long endurance (HALE)
flights [29] (see also Section 1.1.3). It features a wing with high aspect ratio (30.9),
which makes it possible to achieve high efficiency during long flights. A primary fuel
cell pod, four avionics side pods, and two fuel tanks are attached to the wing (see
Figure 6.4 and 1.4). The main dimensions of this aircraft are provided in [29] and used
to develop the numerical model used in the present simulation [30]. In the numerical
model, half of the aircraft (span-wise) is represented. The wing has a chordlength of
c = 8 ft, a semi-span of L = 123.5 ft, and a tip dihedral of 10 degrees starting at 2/3
of the midspan. Its airfoil section is assumed to be Selig S8055 with a 12% thickness
to chordlength ratio. Half of the primary pod, two side pods, and one fuel tank are
also modeled. The wet surface of this numerical model is discretized by a surface grid
with 853 grid points and 1, 687 triangular elements (Figure 6.5).
This surface grid is embedded in the fluid computational domain defined by
190.9 ft x 331.1 ft, 0 ft y 283.5 ft, and 160.0 ft z 160.0 ft.
Its relative location inside the computational domain is shown in Figure 6.6. The
146
CHAPTER 6. APPLICATIONS
flow is air modeled as perfect gas with = 1.4. The free-stream flow state is speciAt 10:34am, the helicopter observer advised again that the aircraft was approaching the northern
in the ocean, an indication a wind shear line was being approached. The helicopter
fied to account for awhitecaps
high-speed
(158
mph),
observer
suggested a left turn
of 20 degrees.
The SPhigh-altitude
responded with a 17-degree(60,
left turn.700 ft) flight condition.
More specifically, the density is set to 0 = 2.1826 104 slug/ft3 (or 0.1125 kg/m3 );
34
the pressure is set to p0 = 146.09 lbf/ft (or 6.995 103 Pa); the velocity vector
corresponds to a Mach number of M0 = 0.2394 and a 0.2 deg angle of attack.
To discretize the fluid computational domain, five unstructured, non body-conforming
CFD grids are created such that the convergence of the numerical solution in space
domain can be demonstrated. In each grid, three levels of refinements are conducted
(Figure 6.7). The characteristic element size in the most refined region of each grid
is set to 0.2 ft/0.25 ft/0.40 ft/0.56 ft/0.64 ft in the five grids. The finest grid has
9, 518, 381 grid points and 56, 777, 950 tetrahedron elements. A detailed statistics for
these grids is provided in Table 6.2.
CHAPTER 6. APPLICATIONS
147
Figure 6.5: Two different views of a surface grid with 853 grid points and 1, 687
triangular elements for an aircraft with high aspect ratio.
Steady state simulations are performed on each and all of the five grids. A tolerance = 1.07 is set for the relative residual at convergence. A cut-view of the
fluid pressure variation with respect to the free-stream pressure, predicted by the
simulation using Grid 2, is shown in Figure 6.8. Streamlines near the aircraft, predicted by the same simulation, are shown in Figure 6.9, together with the magnitude
of velocity along them. The solutions shown in both figures are smooth and possess
expected features. In particular, the deceleration of the flow near the leading and
trailing edges of the wing, as well as the acceleration of the flow above and below the
wing, are clearly evident.
The total lift computed on control volume boundaries (using Algorithm 5.4)
and the reconstructed surface (using Algorithm 5.2 and Algorithm 5.3) are
CHAPTER 6. APPLICATIONS
148
Figure 6.6: The fluid computational domain and the embedded discrete surface (colored in magenta).
shown in Figure 6.10. As the CFD grid being refined, the lift value computed using
different approaches converge in a very narrow range, namely 1, 142 8 lb.f.
Table 6.2: Statistics of five non body-conforming CFD grids created for the simulation
of a steady flow around an aircraft.
Characteristic element size (ft.)
Grid id.
# grid points # elements
Level 1 Level 2 Level 3 Far-field
1
0.2
0.4
0.8
20.0
9.52M
56.7M
2
0.25
0.5
1.0
25.0
5.01M
29.8M
3
0.4
0.8
1.6
40.0
1.34M
7.93M
4
0.56
1.12
2.24
60.0
520K
3.08M
5
0.64
1.28
2.56
70.0
361K
2.13M
CHAPTER 6. APPLICATIONS
149
Figure 6.7: A cut-view (at x = 0 in) of Grid 4 showing three levels of local mesh
refinement.
Figure 6.8: A cut-view (at x = 0 ft) of the fluid pressure variation with respect to
the free-stream pressure, predicted by the simulation on Grid 1
CHAPTER 6. APPLICATIONS
150
Figure 6.9: Streamlines near the aircraft, shown as ribbons. The color on the ribbons
represent the magnitude of velocity.
151
CHAPTER 6. APPLICATIONS
1600
Reconstructed surface
Control volume boundary
Lift (lb.f)
1500
1400
1300
1200
1100
0.2
0.3
0.4
0.5
Element size (ft)
0.6
0.7
CHAPTER 6. APPLICATIONS
6.3
152
153
CHAPTER 6. APPLICATIONS
Because they are extremely thin, the wings considered here are extremely flexible.
This makes the simulation of their structural dynamics particularly challenging as
the finite element modeling of such thin structures is typically sensitive to numerical
errors.
"
reinforcement
via thickening
81 mm
(0,0,0)
140 mm
reinforcement
by stiff fibers
140 mm
Figure 6.11: Ultra-thin flexible flapping wings: description and structural modeling.
In this example, the air flow is assumed to be inviscid. From the physics viewpoint, this is a gross assumption. However, it has no effect on the illustration and
evaluation of the interface tracking and treatment algorithms presented in this paper.
To this effect, a non body-fitted CFD grid with 2, 086, 997 grid points and 12, 462, 983
tetrahedra is generated (see Figure 6.12 and Figure 6.13). The 1, 152 surfacic triangles of the structural model are collected into a discrete fluid-structure interface that
is embedded in the CFD grid. This embedded discrete interface is an open surface.
Hence, its dynamics can be tracked by Algorithm 4.2 but not by Algorithm 4.1.
More specifically, the simulation considered herein illustrates Case III of Figure 4.5,
which is discussed in Chapter 4.
CHAPTER 6. APPLICATIONS
154
Figure 6.12: A pair of ultra-thin triangular wings: computational fluid domain and
the embedded discrete interface (magenta).
Figure 6.13: A pair of ultra-thin triangular wings: non body-conforming CFD grid
cutview at y = 5 mm (left) and cutview at z = 0 mm (right).
CHAPTER 6. APPLICATIONS
155
156
CHAPTER 6. APPLICATIONS
Displacement (mm)
100
50
50
100
0
0.05
0.1
0.15
Time (sec)
0.2
0.25
0.3
CHAPTER 6. APPLICATIONS
157
Figure 6.15: Ultra-thin flexible flapping wings: snapshots of the fluid pressure
(cutview at y = 20 mm) and structural deformation at six different time-instances.
CHAPTER 6. APPLICATIONS
6.4
158
Validation for the implosive collapse of an airbacked aluminum cylinder submerged in water
6.4.1
Experiment
CHAPTER 6. APPLICATIONS
159
pressure, under which the cylinder collapses, is pco = 690.5 psi. This pressure is equal
to the hydrostatic pressure at approximately 470 m under sea. Two photographs
of the collapsed cylinder are shown in Figure 6.17. The time at which the cylinder
starts to collapse is chosen to designate t = 0. The recorded pressure time-history
by a sensor (see Figure 6.18) reveals first a gradual pressure drop of 118.0 psi before
the cylinder gets into self-contact, followed by a sharp pressure pulse with a peak at
1.16 103 psi. This pressure pulse demonstrates the strong shock waves caused by
self-contact of the structure. Figure 6.18 reveals that the duration of collapse is about
1 ms. Given that the process of increasing the hydrostatic pressure from p0w to pco
takes about 1 minute, it is clear that the dynamic collapse of the cylinder essentially
happens under a constant pressure load pco = 690.5 psi.
This experiment provides a research model for analyzing the impact of an imploding external volume, such as an unmanned undersea vehicle, on a nearby submarine
(see Section 1.1.1 for more details). More specifically, the specimen may represent
an air-filled implodable volume attached to a submarine. Following this scenario,
the results of this experiment indicate that: (1) at a diving depth of 470 m, the
implodable volume will collapse; and more importantly (2) when it collapses, the
submarine hull will receive a strong pressure pulse characterized by a peak pressure
of 1.16 103 psi that is, about 70% higher than the hydrostatic pressure at the
diving depth (690.5 psi at 470 m under sea) or even higher. Moreover, this experiment also suggests that, due to the high overpressure resulting from implosion, it
may be necessary to take into account the dynamic response of external, implodable
volumes when investigating the survivability of the submarine hull.
160
CHAPTER 6. APPLICATIONS
t = 0.0280 in
D = 1.5007 in
r = 7/64 in
1.5 in
L = 2D = 3.0014 in
Figure 6.16: Schematic drawing of a cylindrical implodable with end caps designated
by stripes.
6.4.2
The aforementioned experiment is simulated using AERO-F as the fluid solver and
AERO-S as the structure solver. The geometric center of the cylinder is chosen as
the origin of the Cartesian coordinate system, and its axial direction is chosen as the
x-axis. In the structural dynamics sub-problem, half of the cylinder (length-wise) is
modeled. Two different finite element CSD models are created and compared. In both
models, the aluminum material of the specimen in the experiment is represented as a
nonlinear elasto-plastic medium with a Young modulus E = 1.014 107 psi, Poisson
ratio = 0.3, and density S = 2.599 104 (lb/in4 ).sec2 . The yield stress is set to
CHAPTER 6. APPLICATIONS
161
162
CHAPTER 6. APPLICATIONS
1200
Experiment
1100
Pressure (psi)
1000
900
800
700
600
500
0.2
0.4
0.6
0.8
1
Time (sec)
1.2
1.4
1.6
3
x 10
CHAPTER 6. APPLICATIONS
163
Figure 6.19: CSD grid for Model 1.1 and 1.2. The elasto-plastic shell elements for
the cylinder and the rigid shell elements for the plug are distinguished by color.
during the simulation.
In both models, a Mode 4 sinusoidal imperfection is imposed on the geometry
of the cylinder to trigger its collapse. More specifically, the circular cross section of
the cylinder verifies the following parametric equation in the (r, ) polar coordinate
system:
r = r0 (1 cos 4).
r0 = 0.73635 in is the radius of the true circular cross section measured to the midsurface of the cylinder. The imperfection magnitude is set to be equal to the
maximum ovalization of the specimen in the experiment namely = o = 0.083%.
Both models provide for the self-contact of the aluminum cylinder. More specifically, geometric constraints are added to the CSD system in order to avoid interpenetration [98, 99], and Lagrange multipliers are introduced to enforce the these
inequality constraints. In Model 2, the contact between the aluminum cylinder and
the steel plug is also modeled, following the same approach.
In the CFD sub-problem, air inside the cylinder is modeled as perfect gas with a
specific heat ratio a = 1.4. Because of the ultrahigh compressions involved in this
164
CHAPTER 6. APPLICATIONS
water
1.0 in
1.5007 in
air
(0, 0, 0)
steel plug
fixed
ux u y uz x y z 0
aluminum cylinder
symmetry
ux y z 0
Figure 6.20: Model 1.1 for an aluminum cylinder and a steel plug.
problem, water is modeled as stiffened gas with w = 7.15 and p0 = 41, 990 psi. The
fluid computational domain is a rectangular box: = {(x, y, z) R3 : 0 in x
12 in, 10 in y 10 in, 10 in z 10 in}. (Figure 6.24) It is discretized by
an unstructured non body-conforming CFD grid Dh with 1, 689, 089 grid points and
10, 064, 277 tetrahedra (Figure 6.25).
Symmetry boundary conditions are applied to the fluid model at the transversal
plane passing through the middle cross section of the cylinder, i.e. the plane determined by x = 0. Non-reflecting boundary conditions are applied at the remaining
boundaries of the fluid domain. At t = 0, the initial state of air inside the cylinder
is set to 0a , va0 , and p0a . The initial density and velocity for water are set to 0w and
vw0 . To imitate the pressure ramp-up process in the experiment, water pressure is
165
CHAPTER 6. APPLICATIONS
water
1.0 in
1.5007 in
air
(0, 0, 0)
steel plug
aluminum cylinder
u y uz x y z 0
symmetry
ux y z 0
Figure 6.21: Model 1.2 for an aluminum cylinder and a steel plug.
initially set to pa = pco 10 psi and increased statically and linearly to pco , while
the air pressure inside the cylinder is maintained at p0a . Time-integration of the fluid
system starts as soon as the collapse pressure pco is reached. The duration of pressure
increase, denoted by tinc , is chosen such that when pco is reached, the cylinder has
just started collapsing. Here, for CSD Model 1.1, 1.2, and 2, tinc is 5.6 104 sec,
1.5 104 sec, and 2.5 104 sec, respectively.
Three fluid-structure coupled simulations are conducted:
Simulation 1: using CSD Model 1.1;
Simulation 2: using CSD Model 1.2;
Simulation 3: using CSD Model 2.
CHAPTER 6. APPLICATIONS
166
In all simulations, the embedded fluid-structure interface is tracked using Algorithm 4.1; the no-interpenetration transmission condition is enforced using Algorithm 5.1; and the equilibrium transmission condition is enforced using Algorithm 5.4. The second-order implicit three-point backward difference (3PBDF)
scheme (see Section 3.3) is employed to integrate the semi-discretized fluid system,
whereas the second-order explicit central difference (CD) scheme (see Section 3.4) is
employed to integrate the semi-discretized structural system. A constant time step
t = 6.5108 sec is applied to both. The water pressure time-histories are recorded
at the sensor locations for the first T = 1.5 103 sec.
The predicted pressure time-histories are reported at a sensor location in Figure 6.26, 6.27, 6.28, together with the corresponding experimental data. Time is
shifted such that the highest pressure peaks are aligned for ease of comparison. In
general, the main features of the experimental signal are reproduced in all three simulations. These include first the gradual pressure drop, then a sudden pressure jump
with two peaks, followed next by a sharp pressure drop. The magnitude of the first
pressure drop and the width of the pressure jump are accurately captured in all simulations. As for the highest pressure peak, the predictions of Simulation 2 and 3
match the experimental data perfectly with relative errors less than 3%, whereas the
prediction of Simulation 1 differs by approximately 15%. By comparing results of
Simulation 1 (Figure 6.26) and Simulation 2 (Figure 6.27), it is clear that the axial shrinking motion of the cylinder and the plugs has a significant influence to the
magnitude of the highest pressure peak. As expected, the prediction of Simulation
2 agrees better with the experimental data compared to the one of Simulation 1, as
the sliding boundary condition imposed in Model 1.2 on the plug and the bonded
region of the cylinder represents more accurately the experimental setup than the
CHAPTER 6. APPLICATIONS
167
CHAPTER 6. APPLICATIONS
168
169
CHAPTER 6. APPLICATIONS
water
1.2 in
0.3 in
steel plug
1.5007 in
air
(0, 0, 0)
u y uz 0
aluminum cylinder
symmetry
ux y z 0
CHAPTER 6. APPLICATIONS
170
Figure 6.24: Underwater implosion problem: the CFD domain and the embedded
discrete interface.
CHAPTER 6. APPLICATIONS
171
172
CHAPTER 6. APPLICATIONS
1200
Experiment
CSD Model 1.1
1100
Pressure (psi)
1000
900
800
700
600
500
0.2
0.4
0.6
0.8
1
Time (sec)
1.2
1.4
1.6
3
x 10
173
CHAPTER 6. APPLICATIONS
1200
Experiment
CSD Model 1.2
1100
Pressure (psi)
1000
900
800
700
600
500
0.2
0.4
0.6
0.8
1
Time (sec)
1.2
1.4
1.6
3
x 10
174
CHAPTER 6. APPLICATIONS
1200
Experiment
CSD Model 2
1100
Pressure (psi)
1000
900
800
700
600
500
0.2
0.4
0.6
0.8
1
Time (sec)
1.2
1.4
1.6
3
x 10
175
CHAPTER 6. APPLICATIONS
800
Experiment
CSD Model 1.2
CSD Model 2
Pressure (psi)
750
700
650
600
550
1.05
1.1
Time (sec)
1.15
1.2
3
x 10
Figure 6.29: Pressure at a sensor location for 103 sec t 1.2 103 sec, predicted
by Simulation 2 and 3, together with the experimental data.
Figure 6.30: Transversal and axial views of the deformed structure predicted by
Simulation 1.
CHAPTER 6. APPLICATIONS
176
Figure 6.31: Transversal and axial views of the deformed structure predicted by
Simulation 2.
Figure 6.32: Transversal and axial views of the deformed structure predicted by
Simulation 3.
CHAPTER 6. APPLICATIONS
177
Figure 6.33: Snapshots of the fluid pressure (two cut-views at x = 0 in and z = 0 in)
and structural deformation at eight different time-instances. The structure is shown
in wireframes such that the air inside the structure is visible.
CHAPTER 6. APPLICATIONS
6.5
178
6.5.1
Experiment
The specimen used in this series of tests is an aluminum 6061-T6 pipe with a length
of L = 0.916 m, a thickness of h = 8.9 104 m, and a circular cross section with
outer diameter d = 4.128102 m. It is sealed at one end by a Mylar diaphragm, and
connected by a flange at the other end to an aluminum tube, called the detonation
tube. Two fixtures are built to clamp the specimen, leaving an effective test length
179
CHAPTER 6. APPLICATIONS
detonation pipe
detonation wave
41.28 mm
457 mm
916 mm
clamp
180
CHAPTER 6. APPLICATIONS
pressure gauge
aluminum pipe
(specimen)
detonation pipe
I beam
CHAPTER 6. APPLICATIONS
181
The overpressure, defined as the pressure difference relative to the value in the
ambient state, is recorded by a sensor located at 2.4 102 m above the notch
(see Figure 6.35). The signal obtained from the test with a notch length of 3.81
102 m
a peak overpressure of 130 MPa and a duration of 0.3 ms. More generally, the peak
overpressure obtained from all the tests are reported in Figure 6.43, which suggests
a linear relationship with respect to the initial notch length.
6.5.2
CHAPTER 6. APPLICATIONS
182
CHAPTER 6. APPLICATIONS
183
Pin is 20% lower than the peak pressure at the detonation front (6.2 106 Pa), yet
significantly higher than the equilibrium state of the detonation product (a mixture
of carbon dioxide and steam), which is approximated as 2.24 106 Pa using the
Chapman-Jouguet theory.
A non body-conforming CFD grid is generated to discretize the aforementioned
fluid computational domain. It consists of of 1, 434, 600 grid points and 8, 544, 527
tetrahedron elements. Two cut-views of the grid taken at y = 0 m and x = 0.457 m
are shown in Figure 6.40.
Five fluid-structure coupled simulations are performed using the CSD models with
different notch lengths. In all simulations, the embedded fluid-structure interface,
which undergoes topological change due to cracking, is tracked with respect to the
embedding CFD grid using Algorithm 4.2. The no-interpenetration transmission
condition is enforced using Algorithm 5.1. The equilibrium transmission condition
is enforced using Algorithm 5.4. The second-order explicit Runge-Kutta scheme
(see Section 3.3) is employed to integrate the fluid system, while the second-order
explicit Central Differencing scheme (see Section 3.4) is employed to integrate structural system. A constant time step t = 4.0 108 sec is used in both systems.
The termination time of the simulation is set to T = 1.6 103 sec. Each simulation
requires approximately 9 hours of CPU time on 238 MPI processors.
Six snapshots of the structural deformation and the fluid pressure, taken at different time instances, are shown in Figure 6.41. They are obtained from the simulation
characterized by a notch length of 3.81 102 m. The burst of high pressure fluid
through the crack opening is clearly evident, while the crack propagation path resembles the experimental result shown in Figure 6.36. The predicted time-history
CHAPTER 6. APPLICATIONS
184
of overpressure at the sensor location (0.24 m above the notch) is reported in Figure 6.42right. Compared with the corresponding experimental data shown in Figure 6.42left, it can be observed that the pressure pulse is accurately reproduced in
the simulation. In particular, both the magnitude of the peak and the duration of
the pulse are accurately captured.
The peak overpressure at the sensor location, predicted by all five simulations
characterized by different notch lengths, are reported in Figure 6.43, together with
corresponding experimental data. Two observations are notable. First, for each given
notch length, the predicted peak overpressure is close to the value obtained in the
experiment, with an error less than 15%. Second, and more importantly, the linear
relationship between peak overpressure and notch length, reported in [28] basing on
the experimental data, is also captured in the simulations.
6.5.3
The reader is reminded that in the aforementioned simulations, the burning of the
explosive gas inside the pipe and the resulting detonation wave are not modeled.
Instead, the fluid inside the pipe is assumed to be air and initialized by a uniform
flow. As a result, when the simulation starts, the structure receives a uniform hydrostatic loading, rather than the highly nonlinear detonation loading that occurs
in the experiment. It has been shown that the present simulation model allows the
reproduction of several important features observed in the experiment, such as the
linear relationship between peak overpressure and initial notch size. However, it can
not be used for studying the direct influences of the detonation loading. For example, at a fixed sensor location, the blast arrival time, measured to the ignition time,
can not be accurately predicted (see Figure 6.42). In order to model the detonation
CHAPTER 6. APPLICATIONS
185
inside the pipe, the unburned gas (a mixture of ethylene and oxygen) and the detonation product (a mixture of carbon dioxide and water vapor) must be modeled as
different fluid media. In particular, in order to reproduce the thermodynamic behavior of the detonation product, it might be necessary to apply the highly nonlinear
Jones-Wilkins-Lee (JWL) EOS [105].
alibration
alibration
1172 s
s
1172
1439 s
186
CHAPTER 6. APPLICATIONS
187
187
187
994
Calibration
994 s
s
1083
994
1083
994 s
s
s
108
10
1261
1172 s
s
1261
1172
s
s
1350
1261
s
1350
1261s
s
s
135
13
Figure 6.36: Photographs showing crack propagation and the burst of detonation
product through the crack opening. Time zero corresponds to the ignition spark in
the detonation tube.
1528
1439 s
s
1617
1528 s
s
161
CHAPTER 6. APPLICATIONS
187
Figure 6.38: The initial crack prescribed on the CSD model for a pre-flawed aluminum
pipe.
CHAPTER 6. APPLICATIONS
188
Figure 6.39: The embedding CFD domain and the embedded CSD model.
Figure 6.40: Two cut-views of the non body-conforming CFD grid for a fluid-structure
coupled pipe explosion simulation. Left: cut-view at y = 0 m; right: cut-view at
x = 0.457 m
CHAPTER 6. APPLICATIONS
189
190
CHAPTER 6. APPLICATIONS
&%"
&"#
'()**+(),-.'/0
&$"
&##
%"
"#
$"
#
!$"
!"#
#,,,,,,,,,,,,,,#45,,,,,,,,,,,,#46,,,,,,,,,,,,&4$,,,,,,,,,,,,&47,,,,,,,,,,,,$4#
123),-3*0
180
160
140
120
100
80
60
experiment
40
simulation
20
0
0
20
40
60
80
100
Figure 6.43: The peak overpressure at a sensor location (0.24 m above the notch) as
a function of initial notch length: experimental and simulation results.
CHAPTER 6. APPLICATIONS
6.6
191
Finally, a fluid-structure coupled simulation is performed to demonstrate the capability of the computational framework presented in this thesis for underwater explosion
problems involving cracking structures and two immiscible fluid media directly in
contact. The fluid sub-problem is solved using AERO-F, while the structural subproblem is solved using the in-house version of DYNA3D equipped with XFEM.
The CSD model presented in Section 6.5 with an initial notch length l = 2.54
102 m is used here.
CHAPTER 6. APPLICATIONS
192
Figure 6.44: The CFD domain and the embedded CSD model for an underwater
explosion simulation.
In this simulation, all the major components of the computational framework,
summarized in Section 3, are involved. More specifically, the finite volume based
multi-phase compressible flow solver is employed to solve the fluid sub-problem in
which water and air are directly in contact. The immiscible interface between water
and air is located using the level-set method summarized in Section 3.3. The finite
element based structural solver (Section 3.4), equipped with the XFEM technology
(Section 3.5), is employed to solve the dynamics of the cracking aluminum pipe.
Finally, the embedded boundary method ,summarized in Section 3.6 and detailed in
Chapter 4 and 5, is employed to track the fluid-structure interface and enforce the
transmission conditions.
The second-order implicit Three Point Backward Difference (3PBDF) scheme (see
CHAPTER 6. APPLICATIONS
193
CHAPTER 6. APPLICATIONS
194
Figure 6.46: The structural deformation and effective plastic strain at T = 1.0
103 sec.
CHAPTER 6. APPLICATIONS
195
Figure 6.47: The fluid pressure field (left) and the domains of water and air (right)
at 1.78 104 sec (top), 6.25 104 sec (middle), and 1.0 103 sec (bottom). A
cut-view for y = 0 is shown. The structure geometry is shown in wireframe.
Chapter 7
Conclusions and Perspectives for
Future Work
7.1
This thesis presents a fluid-structure coupled computational framework for the solution of nonlinear multi-phase fluid-structure interaction problems involving high
compressions and shock waves, large structural displacements and deformations, selfcontact, and possibly the initiation and propagation of cracks in the structure.
Basing on an embedded boundary method, this computational framework solves
on non body-conforming CFD grids multi-phase compressible inviscid flows interacting with structures undergoing large displacements and deformations, and possibly
cracking.
In this context, two robust and efficient computational algorithms are proposed
for tracking the discrete embedded interface with respect to arbitrary (i.e. structured
and unstructured) non body-conforming CFD grids. The first one is based on a
196
projection approach, whereas the second one is based on a collision approach. The
first algorithm is faster. However, it is restricted to closed interfaces and resolved
enclosed volumes. The second algorithm is more versatile as it can handle open
surfaces and underresolved enclosed volumes. Both computational algorithms have
been equipped with a parallel distributed bounding box hierarchy to efficiently store
and retrieve the elements of the discretized embedded interface. As a result, both
interface tracking algorithms deliver a fast parallel performance. For several threedimensional examples arising from aeroelastic and underwater implosion applications,
they were found to consume a small to tiny percentage of the total simulation CPU
time.
As part of a finite volume multi-phase compressible flow solver, the embedded
boundary method relies on the exact solution of local, one-dimensional, fluid-structure
Riemann problems for enforcing the no-interpenetration transmission condition at the
fluid-structure interface and at the same time recovering at the same interface the
fluid state. This approach has a clear physical background, and is fully consistent
with the flow solver which enforces the interface conditions at fluid-fluid interfaces
by solving two-phase fluid Riemann problems. In addition, it does not need any
interpolation or extrapolation of fluid states near or across the interface, thus greatly
simplifies implementation for unstructured grids. Using the recovered fluid state,
two consistent and conservative approaches are proposed for computing the flowinduced forces and moments on rigid and flexible embedded interfaces. The first
one reconstructs locally the fluid/structure interfaces. The second one bypasses this
reconstruction and chooses instead to work with surrogate fluid/structure interfaces.
Both algorithms are equally applicable to structured and unstructured CFD grids.
A fluid-structure coupled computational framework is then developed by coupling
the multi-phase compressible flow solver equipped with the aforementioned embedded
boundary method with an extended finite element method (XFEM) based structure
solver using a partitioned procedure and provably second-order explicit-explicit and
implicit-explicit time-integrators. In particular, the interface tracking algorithms in
the embedded boundary method are adapted to tracking embedded discrete interfaces
with phantom elements and carrying implicitly represented cracks.
The computational framework is demonstrated, and its performance is assessed,
with the solution of several challenging fluid-structure interaction problems in the
fields of aeronautics, underwater implosions and explosions, and pipeline explosions.
In particular, two laboratory experiments are considered for validation purpose. The
first experiment concerns the implosive collapse of an air-filled aluminum cylinder,
while the second one studies the dynamic fracture of pre-flawed aluminum pipes driven
by detonation waves. In both cases, the numerical simulation is able to reproduce in
a quantitative sense the important features in the experiment.
7.2
The embedded boundary method is currently developed for Euler equations which
govern compressible inviscid flows. Indeed, in many engineering applications, such as
underwater implosions and pipeline explosions, viscous effects in the flow can be safely
neglected. However, strong viscous effects can be found in other fluid-structure interaction problems, such as the flight of flapping wing MAVs. In order to study these
problems, the Navier-Stokes equations needs to be solved, while turbulence models
may need to be incorporated for modeling turbulent flows. The extension of the
current embedded boundary method to compressible viscous (laminar or turbulent)
flows is also in progress in Prof. Charbel Farhats group [107].
Appendix A
Exact solution of the
one-dimensional fluid-structure
Riemann problem
The present appendix provides a detailed solution procedure for the one-dimensional
fluid-structure Riemann problem (P3) presented in Section 2.4.1. The mathematical
formulation of this problem is first stated as follows.
P3: Let F (t) = (, vwall t ] be the one-dimensional time-dependent space computational domain. Find the solution w(x, t) to the one-dimensional Euler equations:
w
+
F(w) = 0,
t
x
in F
(A.1)
where
w = v ,
F(w) = v 2 + p
v(E + p)
200
(A.2)
v(vwall t, t) = vwall
= x vwall t
,
=t
(A.3)
is introduced. Using the chain rule, The partial derivatives in Eq. A.1 can be expressed
as
F(w) F(w)
F(w)
F(w)
=
=
x
w
w w
w w
=
= vwall
+
t
t
t
(A.4)
w
+ (F(w) vwall w) = 0,
(A.5)
v(0, ) = 0.
v (
v 2 + p)
+
=0
E (E + p)
v
+
=0
(A.6)
(A.7)
(A.8)
1
in which E is defined as E = (e + v2 ). This is in fact the Euler equations for a
2
w =
v .
(A.9)
In terms of w,
P3.1 can be re-formulated as:
P3.2: For (, 0] and (0, + ), find the solution w(,
) to
+ F(w)
= 0,
(A.10)
+ F() = 0,
(A.11)
(A.12)
moving wall
t
moving wall
rarefaction
x vwall t
shock
x vwall t
vwall
p
vwall
p
L
vL
p
L
L
vL
p
L
not involved
not involved
(0, 0)
(0, 0)
=
L
pL + pc
2a
2aL
= vwall +
vL +
1
1
(A.13)
(p + pc )
.
2a
do not
1
change across the rarefaction. and p can be expressed in closed-form as
Eqs. A.13 are derived from the fact that Riemann invariants s and u +
p = (pL + pc )
1
2
1
(vwall vL ) + 1 1 pc
2aL
(A.14)
and
= L
p + pc 1
.
pL + pc
(A.15)
At a given time t > 0, the rarefaction expands over [x1 (t), x2 (t)], with
x2 (t) = (v a )t.
The structure of the rarefaction is also determined from the Riemann invariants.
More specifically, with defined as = x/t, the state vector within the rarefaction
fan is determined by
( 1)vL + 2(aL + )
+1
1
L (v() )2 1
() =
(pL + pc )
()
p() = (pL + pc )
pc
L
v() =
(A.16)
(A.17)
(A.18)
In the case of a shock wave, density and pressure components of the intermediate
state ( , vwall , p ), as well as the shock speed vs , satisfy the following RankineHugoniot jump conditions:
L (vL vs ) = (vwall vs )
(A.19)
L (vL vs )2 + pL = (vwall vs )2 + p
(A.20)
1
1
(vL vs )2 + hL =
(v vs )2 + h
2
2
(A.21)
1+
4(pL + pc + ) + 1
2
p +pc
+ 1
pL +pc
+1
L p +pc 1
(
)+
pL +pc +1
L vL vwall
,
L
vs =
(A.22)
(A.23)
(A.24)
where
=
,
(v vL )2
2
,
( + 1)L
and = (pL + pc )
1
.
+1
Tait barotropic liquid: The EOS for a Tait barotropic liquid is provided in Eq. 2.8.
In the case of a rarefaction, the density component of the intermediate state ( , vwall , p )
can be derived using one Riemann invariant only:
vwall +
2aL
2a
= vL +
,
1
1
(A.25)
r
1
where a =
is the speed of sound. a can be obtained from Eq. A.25 as
0
a =
1
( 1)(vL vwall ) + 2aL .
2
1
1
(( 1)(vL vwall ) + 2aL )2 1 .
4
(A.26)
At last, p can be derived using and the EOS. At a given time t, the rarefaction
The structure of the rarefaction fan can be expressed in terms of q = x/t, given by
( 1)vL + 2(aL + q)
+1
a(q) = v(q) q
1
a(q)2 1
(q) = 0
,
v(q) =
(A.27)
(A.28)
(A.29)
(A.30)
L (vL vs )2 + pL = (vwall vs )2 + p
(A.31)
Together with the EOS, there are three nonlinear equations and three unknowns: ,
p , and vs . The exact solutions are not available in closed-form expression. However,
approximate solutions can be computed using numerical nonlinear equation solvers
such as the Newtons method.
Bibliography
[1] Stokes, G. On the effect of the internal friction of fluids on the motion of pendulums. Transactions of the Cambridge Philosophical Society 1851; 9:8106.
[2] Sane S. and Dickinson M., The control of flight force by a flapping wing: lift
and drag production. Journal of Experimental Biology 2001; 204:26072626.
[3] Au-Yang M.K. Flow-induced vibration of power and process plant components:
a practical workbook. ASME Press, New York, 2001.
[4] Garrick, L.E. Aeroelasticity frontiers and beyond. Journal of Aircraft 1976;
13:641657.
[5] Dugundji J. Personal perspective of aeroelasticity during the years 19531993.
Journal of Aircraft 2003; 40:809812.
[6] Livne E. Future of airplane aeroelasticity. Journal of Aircraft 2003; 40:1066
1092.
[7] Naitoh M., Kasahara F., Mitsuhashi T., and Ohshima I. Analysis on pipe
rupture of steam condensing line at Hamaoka-1, (I) Accumulation of noncondensible gas in a pipe. Journal of Nuclear Science and Technology 2003;
40:10321040.
209
BIBLIOGRAPHY
210
[8] National Transportation Safety Board. Pacific gas and electric company natural
gas transmission pipeline rupture and fire, San Bruno, California, September 9,
2010. Pipeline Accident Report NTSB/PAR-11/01 2011. Washington DC.
[9] Shyy W., Aono H., Chimakurthi S.K., Trizila P., Kang C.-K., Cesnik C.E.S.,
and Liu H. Recent progress in flapping wing aerodynamics and aeroelasticity.
Progress in Aerospace Sciences 2010; 46:284327.
[10] Kerbl K., Rehman J., Landman J., Lee D., Sundaram C., and Clayman V. Current management of urolithiasis: progress or regress? Journal of Endourology
2002; 16:181288.
[11] Zhou Y. and Zhong P. The effect of reflector geometry on the acoustic field
and bubble dynamics produced by an electrohydraulic shock wave lithotripter.
Journal of Acoustic Society of America 2006; 119:36253636.
[12] Klaseboer E., Fong S.W., Turangan C.K., Khoo B.C., Szeri A.J., Calvisi M.L.,
Sankin G.N., and Zhong P. Interaction of lithotripter shockwaves with single
inertial cavitation bubbles. Journal of Fluid Mechanics 2007; 593:3356.
[13] Iloreta J.I., Fung N.M., and Szeri A.J. Dynamics of bubbles near a rigid surface
subjected to a lithotripter shock wave. Part 1. Consequences of interference
between incident and reflected waves. Journal of Fluid Mechanics 2008; 616:43
61.
[14] Calvisi M.L., Iloreta J.I., and Szeri A.J. Dynamics of bubbles near a rigid surface
subjected to a lithotripter shock wave. Part 2. Reflected shock intensifies nonspherical cavitation collapse. Journal of Fluid Mechanics 2008; 616:6397.
BIBLIOGRAPHY
211
[15] Cleveland R.O. and Sapozhnikov O.A. Modeling elastic wave propagation in
kidney stones with application to shock wave lithotripsy. Journal of Acoustic
Society of America 2005; 118:26672676.
[16] Wijerathne M.L.L., Hori M., Sakaguchi H., and Oguni K. 3D dynamic simulation of crack propagation in extracorporeal shock wave lithotripsy. IOP Conference Series: Material Science and Engineering 2010; 10:012120.
[17] Udaykumar H, Mittal R, Shyy W. Computation of solid-liquid phase fronts in
the sharp interface limit on fixed grids. Journal of Computational Physics 1999;
153:535574.
[18] Udaykumar H, Mittal R, Rampunggoon P, Khanna A. A sharp interface Cartesian grid method for simulating flows with complex moving boundaries. Journal
of Computational Physics 2001; 174:345380.
[19] Glowinski R, Pan TW, Kearsley AJ, Periaux J. Numerical simulation and optimal shape for viscous flow by a fictitious domain method. International Journal
for Numerical Methods in Fluids 2005; 20:695711.
[20] Gilmanov A, Sotiropoulos F. A hybrid Cartesian/immersed boundary method
for simulating flows with 3D, geometrically complex, moving bodies. Journal of
Computational Physics 2005; 207:457492.
[21] Mittal R, Dong H, Bozkurttas M, Najjar FM, Vargas A, von Loebbecke A. A
versatile sharp interface immersed boundary method for incompressible flows
with complex boundaries. Journal of Computational Physics 2008; 227:4825
4852.
BIBLIOGRAPHY
212
BIBLIOGRAPHY
213
BIBLIOGRAPHY
214
[39] Ivings M.J., Causon D.M., and Toro E.F. On Riemann solvers for compressible
liquids. International Journal for Numerical Methods in Fluids 1998; 28:395
418
[40] Cantwell B.J. Fundamentals of compressible flow. Lecture notes for AA210A.
Stanford University, Stanford, 2009.
[41] Ishihara D., Yoshimura S. A monolithic approach for interaction of incompressible viscous fluid and an elastic body based on fluid pressure Poisson equation.
International Journal for Numerical Methods in Engineering 2005; 64:167203.
[42] Hubner B. and Dinkler D. A simultaneous solution procedure for strong interactions of generalized Newtonian fluids and viscoelastic solids at large strains.
International Journal for Numerical Methods in Engineering 2005; 64:920939.
[43] Bazilevs Y., Calo V.M., Hughes T.J.R., and Zhang Y. Isogeometric fluidstructure interaction: theory, algorithms, and computations. Computational
Mechanics 2008; 43:337
[44] Piperno S., Farhat C., and Larrouturou B. Partitined procedures for the transient solution of coupled aeroelastic problems. Part I: Model problem, theory
and two-dimensional application. Computational Methods in Applied Mechanics
and Engineering 1995; 124:79112.
[45] Piperno S. and Farhat C. Partitined procedures for the transient solution
of coupled aeroelastic problems. Part II: Energy transfer analysis and threedimensional applications. Computational Methods in Applied Mechanics and
Engineering 2001; 190:31473170.
BIBLIOGRAPHY
215
[46] Farhat C, Lesoinne M. Two efficient staggered procedures for the serial and
parallel solution of three-dimensional nonlinear transient aeroelastic problems.
Computer Methods in Applied Mechanics and Engineering 2000; 182: 499516.
[47] Farhat C, Lesoinne M, LeTallec P. Load and motion transfer algorithms for
fluid/structure interaction problems with nonmatching discrete interfaces: momentum and energy conservation, optimal discretization and application to
aeroelasticity. Computer Methods in Applied Mechanics 1998; 157: 95114.
[48] Farhat C, van der Zee KG, Geuzaine P. Provably second-order time-accurate
loosely-coupled solution algorithms for transient nonlinear computational aeroelasticity. Computer Methods in Applied Mechanics and Engineering 2006;
195:19732001.
[49] Wang K, Rallu A, Gerbeau J-F, Farhat C. Algorithms for interface treatment
and load computation in embedded boundary methods for fluid and fluidstructure interaction problems. International Journal for Numerical Methods
in Fluids 2011; 67:11751206.
[50] Wang K, Gretarsson J, Main A, and Farhat C. Computational Algorithms for
Tracking Dynamic Fluid-Structure Interfaces in Embedded Boundary Methods.
International Journal for Numerical Methods in Fluids. (to appear)
[51] Michler C., van Brummelen E.H., Hulshoff S.J., and de Borst R. The relevance of conservation for stability and accuracy of numerical methods for fluidstructure interaction. Computer Methods in Applied Mechanics and Engineering
2003; 192:41954215.
BIBLIOGRAPHY
216
[52] Seo J.H. and Mittal R. A sharp-interface immersed boundary method with
improved mass conservation and reduced spurious pressure oscillations. Journal
of Computational Physics 2011 230:73477363
[53] Guendelman E, Selle A, Losasso F, Fedkiw R. Coupling water and smoke to thin
deformable and rigid shells. SIGGRAPH 2005, ACM TOG 24 2005; 973981.
[54] Gretarsson J, Kwatra N, Fedkiw R. Numerically Stable Fluid-Structure Interactions Between Compressible Flow and Solid Structures. Journal of Computational Physics 2011; 230:30623084.
[55] Fadlun EA, Verzicco R, Orlandi P, Mohd-Yusof J. Combined-immersed boundary finite difference methods for three-dimensional complex flow simulations.
Journal of Computational Physics 2000; 161:3560.
[56] Kim J, Kim D, Choi H. An immersed-boundary finite volume method for simulations of flow in complex geometries. Journal of Computational Physics 2001;
171:132150.
[57] Choi JI, Oberoi RC, Edwards JR, Rosati JA. An immersed boundary method for
complex incompressible flows. Journal of Computational Physics 2007; 224:757
784.
[58] Tseng YH, Ferziger JH. A ghost-cell immersed boundary method for flow in
complex geometry. Journal of Computational Physics 2003; 192:593623.
[59] Berthelsen PA, Faltinsen OM. A local directional ghost cell approach for incompressible viscous flow problems with irregular boundaries. Journal of Computational Physics 2008; 227:43544397.
BIBLIOGRAPHY
217
[60] Lohner R, Baum JD, Mestreau EL, Sharov D, Charman C, Pelessone D. Adaptive embedded unstructured grid methods. AIAA-03-1116, 2003.
[61] Mark A, van Wachem BGM. Derivation and validation of a novel implicit
second-order accurate immersed boundary method. Journal of Computational
Physics 2008; 227:66606680.
[62] Fedkiw R. Coupling an Eulerian fluid calculation to a Lagrangian solid calculation with the ghost fluid method. Journal of Computational Physics 2002;
175:200224.
[63] Geuzaine P, Brown G, Harris C, Farhat C. Aeroelastic dynamic analysis of a full
F-16 configuration for various flight conditions. AIAA Journal 2003; 41:363
371.
[64] Farhat C, Geuzaine P, Brown G. Application of a three-field nonlinear fluidstructure formulation to the prediction of the aeroelastic parameters of an F-16
fighter. Computers & Fluids 2003; 32:329.
[65] Methods Development Group. DYNA3D: A nonlinear, explicit, threedimensional finite element code for solid and structural mechanics. User manual,
Lawrence Livermore National Laboratory (LLNL) 2005.
[66] Stephens C.H., Arena A.S., and Gupta K.K. Application of the transpiration
method aeroservoelastic prediction using CFD. AIAA Journal 1998; 982071.
[67] Farhat C, Geuzaine P, Grandmont C. The discrete geometric conservation law
and the nonlinear stability of ALE schemes for the solution of flow problems on
moving grids. Journal of Computational Physics 2001; 174: 669694.
BIBLIOGRAPHY
218
BIBLIOGRAPHY
219
BIBLIOGRAPHY
220
[83] Sethian J.A. Level Set Methods and Fast Marching Methods. Cambridge University Press, Cambridge, United Kingdom, 1999. Springer, Berlin, Germany,
2003.
[84] Van Leer B. Towards the ultimate conservative difference scheme V: a secondorder sequel to Goudonovs method. Journal of Computational Physics 1979;
32:361370.
[85] LeVeque R.J. and Li Z. The immersed interface method for elliptic equations
with discontinuous coefficients and singular sources. SIAM Journal on Numerical Analysis 1994; 31:10191044.
[86] Lee L. and LeVeque R.J. An immersed interface method for incompressible
Navier-Stokes equations. SIAM Journal on Scientific Computing 2003; 25:832
856.
[87] Geradin M. and Rixen D. Mechancial Vibrations: Theory and Application to
Structural Dynamics. Wiley, New York, 1997.
[88] Belytschko T., Liu W.K., and Moran B. Nonlinear finite elements for continua
and structures. Wiley, New York, 2000.
[89] Moes N., Dolbow J., and Belytschko T. A finite element method for crack
growth without remeshing. International Journal for Numerical Methods in Engineering 1999; 46:131150.
[90] Belytschko T., Chen H., Xu J., and Zi G. Dynamic crack propagation based
on loss of hyperbolicity with a new discontinuous enrichment. International
Journal for Numerical Methods in Engineering 2003; 58:18731905.
BIBLIOGRAPHY
221
[91] Song J.H., Areias P.M.A., and Belytschko T. A method for dynamic crack
and shear band propagation with phantom nodes. International Journal for
Numerical Methods in Engineering 2006; 67:868893.
[92] Belytschko T., Moes N., Usui S., and Parimi C. Arbitrary dicontinuities in
finite elements. International Journal for Numerical Methods in Engineering
2001; 50:9931013.
[93] Lohner R. Applied computational fluid dynamics techniques: an introduction
based on finite element methods (second ed). Wiley, New York, 2008.
[94] Robinson-Mosher A, Shinar T, Gretarsson J, Su J, Fedkiw R. Two-way coupling
of fluids to rigid and deformable solids and shells. SIGGRAPH 2008, ACM TOG
27 2008; 46.146.9.
[95] Farhat C. CFD on moving grids: from theory to realistic flutter, maneuvering, and multidisciplinary optimization International Journal of Computational
Fluid Dynamics 2005; 19:595603.
[96] Hartmann D., Meinke M., and Schroder W., A strictly conservative Cartesian
cut-cell method for compressible viscous flows on adaptive grids. Computer
Methods in Applied Mechanics and Engineering 2011; 200:10381052.
[97] Quartapelle L., Castelletti L., Guardone A., and Quaranta G., Solution of the
Riemann problem of classical gasdynamics. Journal of Computational Physics
2003; 190:118140.
[98] Puso M.A. and Laursen T.A. A mortar segment-to-segment contact method for
large deformation solid mechanics. Computer Methods in Applied Mechanics
and Engineering 2004; 193 601629
BIBLIOGRAPHY
222
[99] Glass M.W., Crane N.K., Gullerud A.S., Jones R.E., Voth T.E. and Brown
K.H., ACME: algorithms for contact in a multiphysics environment. User manual, 2004.
[100] Whaley R. and Wilkins J.R. III. U.S. Navy general specification for the design, construction, and reparing of diving and hyperbaric equipment. Technical
Report TS500-AU-SPN-010, NAVSEA, 2006.
[101] Avery P., Rebel G., Lesoinne M., and Farhat C. A numerically scalable dualprime substructuring method for the solution of contact problems part I:
the frictionless case. Computer Methods in Applied Mechanics and Engineering
2004; 193 24032426.
[102] Farhat C., Avery P., Tezaur R., and Li J. FETI-DPH: A dual-primal domain
decomposition method for acoustic scattering. Journal of Computational Acoustics 2005; 13 499524.
[103] Avery P., Farhat C., and Hetmaniuk U. A Pade-based factorization-free algorithm for identifying the eigenvalues missed by a generalized symmetric eigensolver. International Journal for Numerical Methods in Engineering 2009; 79
239252.
[104] Chapman D.L. On the rate of explosion in gases. Philosophical Magazine 1899;
47 90104.
[105] Dobratz B.M. and Crawford P.C. LLNL explosives handbook. properties
of chemical explosives and explosive simulants. Technical Report A572272,
Lawrence Livermore National Laboratories, 1985.
BIBLIOGRAPHY
223