Bz5xn Multiphysics Anfd Multiscale Modeling Techniques and Applications

MULTIPHYSICS
and MULTISCALE
MODELING
TECHNIQUES AND APPLICATIONS
Young W. Kwon
Boca Raton London New York
CRC Press is an imprint of the

Taylor & Francis Group, an informa business
2016 by Taylor & Francis Group, LLC
viii
ix
xi
xii
Preface
Multiscale and multiphysics analyses have become popular for better understanding of
complex physical behaviors. However, to the best of my knowledge, books on these topics
are limited. Most are focused on specific topics. As a result, this book is expected to serve
a wider spectrum of engineers and scientists who work on multiscale and multiphysics
analyses. This book is a collection of research in the subject areas of multiscale and multiphysics analysis that coworkers and I conducted during the past decade.
The first several chapters present various computational techniques that are useful for
multiscale and multiphysics analyses. Those include the finite element method, lattice
Boltzmann method, cellular automata, molecular dynamics, and their coupling techniques.
Some of those techniques are useful for multiscale analysis, while others are beneficial
for multiphysics analysis. Then, the next three chapters present examples for multiscale
analysis, which is used for composite materials, metallic materials, and biomaterials. The
final three chapters present examples for multiphysics analysis. Those examples are fluidstructure analysis of composite structures, electromechanical rail guns, and blood vessel
aneurysms.
Young W. Kwon
Monterey, California
xiii
Acknowledgments
I am indebted to my many coworkers, especially former graduate students because they
have made significant contributions to research programs. Especially, this book contains large portions of research undertaken with the following individuals (in alphabetical order): Ahmet Altekin, Jamal AlRowaijeh, Jermaine Bailey, Joe Berner, Stuart Blair,
Brandon Clumpner, Ryan Conner, Linda Craugh, Jarema Didoszak, Anthony Harrell,
Selcuk Hosoglu, Sunghoon Jung, Daniel Kidd, Scott Knutton, Chaitanya Manthena, Ryan
McCrillis, Jung Joo Oh, Angela Owens, Moon Shik Park, Spyridon Plessas, Nikolaos
Pratikakis, Eric Priest, Kevin Roach, Matthew Shellock, Fatma Gulden Simsek, Michael
Violette, and Kangjie Yang. If there is anyone omitted here, I apologize for my mistake.
Without their contribution, this book would not be possible. I also appreciate the guidance
and help from the CRC staff, especially that of Jonathan Plant.
xv
Author
Dr. Young W. Kwon is a distinguished professor in the Mechanical and Aerospace
Engineering Department of the Naval Postgraduate School in Monterey, California. He
was past chair of the department. He was also professor and chair of the Department of
Mechanical Engineering and Energy Processes of Southern Illinois University Carbondale.
He received his PhD degree from Rice University and a BS degree from Seoul National
University, both in mechanical engineering. Before joining the Naval Postgraduate School,
he was an assistant professor at the University of MissouriRolla.
His research interests include multiscale and multiphysics computational techniques
for material behaviors bridging the nanoscale to macroscale, fluid-structure interaction
problems, ship shock modeling and simulation; composite materials; fracture and damage
mechanics; nanotechnology; and biomechanics. He has authored or coauthored extensive
technical publications, many of which appeared in archived, refereed publications. He
wrote the textbook Finite Element Method Using MATLAB (CRC Press), which was translated
into Greek; and contributed book chapters, including Nanomechanics in Nanoengineering
of Structural, Functional and Smart Materials (CRC Press), Multi-Scale Computational
Modeling and Simulation in Progress in Engineering Computational Technology (Saxe-Coburg
Publishing), Computational and Experimental Study of Composite Scarf Bonded Joints
in Structural Integrity and Durability of Advanced Composites: Innovative Modeling Methods and
Intelligent Design (Woodhead Publishing), Dynamic Loading on Composite Structures
with Fluid-Structure Interaction in Dynamic Deformation, Damage and Fracture in Composite
Materials and Structures (Woodhead Publishing), and others. He edited a book, Multiscale
Modeling and Simulation of Composite Materials and Structures, p
ublished by Springer.
Dr. Kwon has received various awards, including the Cedric K. Ferguson Medal from
the Society of Petroleum Engineers, Menneken Faculty Award, Excellent Research Award
from the American Orthopedic Society of Sports Medicines, Outstanding Instruction and
Research Awards, American Society of Mechanical Engineers PVPD (Pressure Vessels and
Piping Division) Outstanding Service Award, National Deans List, Whos Who in Science
and Engineering, American Society of Mechanical Engineers Dedicated Service Award,
American Society of Mechanical Engineers Board of Governors Award, and more.
Dr. Kwon is a fellow of the American Society of Mechanical Engineers. He is the technical
editor of the American Society of Mechanical Engineers Journal of Pressure Vessel Technology
as well as the journal Materials Sciences and Applications, and he serves on the editorial boards
for multiple journals. He is vice president for the International Society of Multiphysics.
Dr. Kwon was the American Society of Mechanical Engineers PVPD chair and its senate
president. He has served as a member of the executive committee of American Society
of Mechanical Engineers PVPD, technical program chair for the 2009 American Society
of Mechanical Engineers PVP Conference, and conference chair for the 2010 American
Society of Mechanical Engineers PVP Conference. He is also an honorary theme editor of
Pressure Vessels and Piping Systems of the Encyclopedia of Life Support Systems under the
auspices of the United Nations Educational, Scientific, and Cultural Organization.
Dr. Kwon has conducted extensive research projects sponsored by government agencies
and private sectors and supervised more than 100 graduate students. Dr. Kwon provided
numerous talks as an invited or keynote speaker at professional meetings and institutions
worldwide.
xvii
1
Introduction
1.1Overview
There are systems that exist in nature and there are systems developed by human beings.
For example, the bodies of humans and animals are living systems in nature, and power
plants are man-made systems. To understand or develop such systems, it is necessary to
conduct multiphysics and multiscale analyses.
Multiphysics analysis is the study of multiple physical behaviors as they interact with
one another. For example, the human body is a good example of something that requires
multiphysics analyses. Chemistry and biology are basic knowledge necessary to understand living cells, tissues, organs, and the like. However, those subjects are not considered
in this book because they are outside of the scope of the study. Only the physical aspects
are considered in multiphysics analysis.
Considering the blood circulation system in the human body, blood vessels require fluid
mechanics analysis to study blood flow and blood pressure; the system also requires structural analysis to investigate contraction and dilation as well as aneurysms and potential
ruptures of the blood vessels. In this case, both analyses should be coupled because blood
and vessels interact. As a result, multiphysics analysis is required. As an example for a
man-made system, a power plant contains many heat pipes that carry hot fluid, which also
requires fluid mechanics analysis and structural analysis, leading to multiphysics analysis. When there is an exchange of heat between the fluids inside and outside the heat pipes,
heat transfer analysis should also be considered.
In many cases, a single analysis is conducted, neglecting the interactive aspect of multiphysics because the single analysis is much simpler. However, if the interaction is strong
and influencing each participant, a single analysis may not provide reliable results. In
that case, multiphysics analysis must be undertaken even though it is more complex and
requires more effort. With advances in computing power as well as computational techniques, multiphysics analysis has been more common in engineering design and analysis.
On the other hand, almost every material and living organism has a complex hierarchical structure in different length scales. Human bones are good examples. They consist
of simple elements at the nanoscale from a mechanics point of view. However, through
the complex hierarchical structures in different length scales, bones are optimized and
provide necessary strength and stiffness for the human skeleton. Likewise, man-made
composite materials consist of fibers and matrix materials. Through aligning or weaving
the fibers, the fiber composite structures become strong, stiff, and light for use for new
technology. Metals also have many different characteristics at the different length scales.
To understand and predict their behaviors as well as to develop new materials, it is
necessary to undertake multiscale analysis. This analysis links the main characteristics
1
Multiscale and Multiphysics Modeling
at different length scales so that we can have more fundamental understanding of their
behavior as well as knowledge of what are the most important hierarchical structures
influencing the macroscale behavior.
1.2Computational Methods
Advances in computing capabilities as well as computational techniques contribute to the
study of multiphysics and multiscale analysis. In this section, we discuss some computational methods used for multiscale and multiphysics analyses. There is no intention to list
and discuss all the available computational methods here. Some computational methods
are presented that are relevant to the contents of the book.
The finite element method is arguably the most popular and powerful computational
technique to analyze continuous media [13]. This technique can solve virtually any
differential equation with proper boundary and initial conditions for complex shapes
of domains. The finite element method was developed initially for structural analysis,
and the technique has spread to other types of applications, including problems in fluid
mechanics, electromagnetic waves, and so on. As a result, the finite element method is useful for multiphysics problems.
The finite element technique is useful for solving continuous domain problems; the
molecular dynamics technique is applicable to discrete domain problems such as atomistic
and molecular modeling [4,5]. As a result, the length scale of the molecular dynamics problems is much smaller than the length scale of the finite element analysis problems. In order
to combine the two techniques, a coupling technique is necessary for multiscale analyses.
The lattice Boltzmann method as well as cellular automata can be applied to both continuous and discrete domain problems [69]. These are the advantages of the techniques.
However, those methods are based on rules, and those rules are not straightforward for
development for any differential equation. As a result, the lattice Boltzmann method and
the cellular automata have been applied to a limited number of problems. Because those
methods are based on rules, they are easy to program and are computationally efficient.
1.3Organization of This Book

This book is organized in the following manner: first, multiple computational techniques
are presented in Chapters 2 through 5, respectively: the finite element method, lattice
Boltzmann method, cellular automata, and molecular dynamics technique. Then, Chapter
6 presents coupling techniques to take advantage of these methods. Finally, several
example problems are discussed for multiphysics and multiscale problems in Chapters 7
through 12. The first three of these chapters are for multiscale analyses, and the last three
chapters are for multiphysics problems.
Multiscale analysis for composite materials and structures is provided in Chapter 7;
fibrous composite, particulate composite, and woven fabric composite materials and structures are presented as examples. Multiscale analysis for metals is given in Chapter 8. In
this chapter, a simplified analysis is provided for crystalline structures using the finite
Introduction
element method and molecular dynamics. Then, Chapter 9 discusses multiscale analysis
of biomaterials using human bones as an example.
Multiphysics problems are discussed for composite structures in Chapter 10. Fluid
structure interaction is presented for composite structures surrounded by water. Ship
structures are examples of these structures. Chapter 11 provides an example of the multiphysics analysis of electromechanical problems. A rail gun is selected as an example problem; this couples electromagnetic waves, heat transfer, and rigid body dynamics. Finally,
Chapter 12 presents a multiphysics analysis of biomechanics using the example of blood
vessels, for which there is fluidstructure interaction.
2
Finite Element Method
2.1Introduction
The finite element method (FEM) is one of the most widely used numerical solution techniques [13]. Especially, the FEM has been used almost exclusively for analyses of solids
and structures. This chapter introduces the basic finite element concepts and presents formulations for various applications, in particular for solid and structural analyses.
The FEM can solve differential equations whether they are boundary value problems,
initial value problems, or eigenvalue problems. The method of weighted residual (MWR)
is a good starting point for the finite element formulation [1]. Therefore, the next section
presents the MWR to introduce the basic finite element concept and formulation. Then, the
finite element technique is applied to various physical problems in the sections that follow.
2.2Method of Weighted Residual

To illustrate the MWR, let us begin with a simple example problem that is expressed as an
ordinary differential equation:
d2u
u = 0, 0 x 1 (2.1)
dx 2
and the differential equation should satisfy the following boundary conditions:

u(0) = 0 and u(1) = 1
(2.2)
This is a simple two-point boundary value problem. The exact solution is

u( x) =
ex e x
(2.3)
e e 1
which satisfies both the governing equation, Equation 2.1, as well as the boundary conditions, Equation 2.2, simultaneously. Let us assume that we do not have the exact solution for
the sake of introduction of the MWR. The purpose of the MWR is to find an approximate
5
solution to the given problem. We want to make the approximate solution represent the
exact solution as closely as possible even though we do not know the exact solution.
To begin with the MWR, we assume a possible set of solutions. Because we do not know
the exact solution, what would be the proper forms for the approximate solutions? One
fact is that any complex function can be expressed in terms of a series of polynomials with
an infinite number of terms. In a practical sense, we cannot assume an infinite order of
polynomial function. Instead, a finite order of polynomial would be a reasonable choice.
Therefore, let us assume a polynomial function. For the present example, it is easy to find
an approximate solution that satisfies the boundary conditions. Because there are two
boundary conditions, the lowest order of polynomial is a linear function. However, both
coefficients of the linear function can be determined uniquely from the two boundary
conditions. As a result, there is no room to optimize the approximate solution as closely as
possible to the exact solution that also satisfies the differential equation. To this end, the
lowest order of polynomial function is a quadratic function as given next:

(x) = a0 + a1x + a2x2 (2.4)
where ai (i = 0, 1, 2) is the coefficients to be determined, and denotes an approximate

solution. Such an assumed function is called the trial function. First, we want to satisfy the
boundary conditions. Substitution of the boundary conditions as given in Equation 2.2
into the quadratic function of Equation 2.4 yields

(0) = a0 = 0
(2.5)
(1) = a0 + a1 + a2 = 1
(2.6)
Because there are two equations for three unknowns, we can express any two unknowns
in terms of the third unknown. In other words, we can obtain a0 = 0 and a1 = 1 a2. Then,
the approximate solution can be rewritten as

(x) = (1 a2)x + a2x2 (2.7)
The mathematical expression in Equation 2.7 satisfies the boundary conditions regardless of the choice of the third coefficient a2. To check how the approximate solution of
Equation 2.7 satisfies the differential equation, Equation 2.1, Equation 2.7 is plugged into
Equation 2.1, leading to

R(x) = 2a2 (1 a2)x a2x2 (2.8)
in which R(x) is called the residual. The residual will be zero for any value of x within the
problem domain if the approximate solution happens to coincide with the exact solution.
However, because the approximate solution is different from the exact solution, the residual does not vanish for all x values within the problem domain.
From the concept of the residual, it is reasonable to state that a smaller residual means
the approximate solution is closer to the exact solution. As a result, we want to find the
third coefficient a2 to minimize the residual. In particular, we want to minimize the residual throughout the whole problem domain. In that aspect, we plan to minimize the sum
of the residual over the whole domain. However, we do not want to sum the residual as
it is because the residual may be positive or negative from point to point in the domain.
In that case, those residuals can cancel out one another, resulting in a negligible sum of
residuals. In other words, the residuals, may be large at every point, but their sum may be
small because of such cancellation. To avoid such a problem, the residual is multiplied by
another function, which is called the test function or weighting function. Then, such a residual is called a weighted residual. Now, our goal is to make the sum of the weighted residual
over the problem domain vanish. That is,
1
I=
w(x)R(x) dx = w(x){2a (1 a )x a x } dx = 0 (2.9)

2
The next question is what would be the best choice as the test function. There are some
popular choices [1], and there is a specific name depending on the selection of the test function. In this study, the Galerkin method is presented. In that technique, the test function is
selected based on the trial function, and their relationship to the present trial function is
w( x) =
( x)
du
= x + x 2 (2.10)
da2
Inserting Equation 2.10 into Equation 2.9 gives

1
I=
( x + x ){2a (1 a )x a x }dx = 0 (2.11)

2
5
Solving Equation 2.11 for the unknown a2 yields a2 =
, and the approximate solution
22
becomes

( x) =
u
17 x + 5 x 2
(2.12)
22
The approximate solution can be compared to the exact solution as shown in Figure 2.1.
Both solutions agree well.
To improve the approximate solution, we need to increase the order of the polynomials
in the trial function. For example, we use a cubic function such as

(x) = b0 + b1x + b2x2 + b3x3 (2.13)
Applying the boundary conditions to this expression yields

(0) = b0 = 0
(2.14)
(1) = b0 + b1 + b2 + b3 = 1
(2.15)
1
Exact
Approx.
FEM
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
FIGURE 2.1
Comparison among the exact solution, approximate solution using quadratic polynomials, and finite element
solution using two linear shape functions.
Elimination of two variables from the equations gives

b0 = 0
(2.16)
b1 = 1 b2 b3 (2.17)
Then, the trial function is expressed as

(x) = (1 b2 b3)x + b2x2 + b3x3 (2.18)
Substitution of the trial function into the differential equation results in the residual

R(x) = 2b2 + (1 + b2 + 7b3)x b2x2 b3x3 (2.19)
To apply the MWR, we need two test functions because there are two unknowns in
Equation 2.19. Those are determined such that

w1 ( x ) =
( x)
du
= x + x 2 (2.20)
db2
w2 ( x) =
( x)
du
= x + x 3 (2.21)
db3
The resulting two weighted residual expressions are

1
I1 =
( x + x ){2b + (1 + b + 7b )x b x
2
b3 x 3 dx = 0 (2.22)
I2 =
( x + x ){2b + (1 + b + 7b )x b x
3
b3 x 3 dx = 0 (2.23)
Solving the two equations for b1 and b2 gives the approximate solution.
2.3Galerkin Finite Element Formulation

If a problem domain has a complex shape, it is not easy to assume a trial function over the
domain while satisfying prescribed boundary conditions throughout the domain. In addition, the solution process should be conducted systematically using a computer algorithm.
To that end, the problem domain is divided into a number of simple shapes of subdomains. For the one-dimensional (1-D) problem discussed previously, the problem domain
is divided into two equal size subdomains as shown in Figure 2.2. Each subdomain is
called a finite element. As a result, there are two finite elements in Figure 2.2, and each
element has two nodes. However, because one node is shared between the first and the
second elements, there are three nodes in the problem. The finite element formulation is
going to determine solutions at the nodes, which is called the nodal variable. Let the nodal
variable at the ith node be called ui. Then, the first element has nodal variables u1 and u2,
and the second element has nodal variables u2 and u3.
In the finite element formulation, a trial function is expressed in terms of so-called shape
functions and nodal variables. The shape functions are used to interpolate the solution
within an element. Because each element has two nodes, a linear interpolation would be
suitable to express the solution inside the element with two nodal variables. For example,
let an element have nodal variables ui and uj. The linear interpolation function within the
element is expressed as
(x) = a + bx (2.24)
We want to replace a and b by the nodal variables ui and uj. To achieve this, we evaluate
the linear interpolation function at the nodal points as follows:

(xi) = a + bxi = ui (2.25)
(xj) = a + bxj = uj (2.26)
Element #1
x=1
Node #1
Element #2
x = 1.5
Node #2
FIGURE 2.2
Finite element mesh with two linear elements and three nodes.
x=2
Node #3
10
Solving these two equations gives
a=
ui x j u j xi
(2.27)
x j xi
b=
u j ui
(2.28)
x j xi
Substitution of Equations 2.27 and 2.28 into Equation 2.24 yields
( x) =
u
ui x j u j xi u j ui
x (2.29)
+
x j xi
x j xi
Rearranging the expression with collecting terms for ui and uj, respectively, gives
( x) =
u
xj x
x j xi
ui +
x x
x xi
j
uj =
x j xi
x j xi
x xi ui

x j xi u j (2.30)

The shape functions are expressed as
H i ( x) =
xj x
(2.31)
x j xi
H j ( x) =
x xi
(2.32)
x j xi
and
(x) = Hi(x)ui + Hj(x)uj (2.33)
The shape functions are called linear shape functions, and they are plotted in Figure 2.3.
The functions have the following properties:

Hi(xi) = 1, Hi(xj) = 0, Hj(xi) = 0, Hj(xj) = 1,
(2.34)
Hi(x) + Hj(x) = 1
(2.35)
and

11
1
Hi(x)
Hj(x)
0
xi
xj
FIGURE 2.3
Plot of linear shape functions.
Now, the weighted residual formulation becomes

1
I=
0.5
wR dx =
wR dx +
wR dx =
wR dx (2.36)
e = 1 e
0.5
where e denotes an element domain. For the present example,

d2u
d =
wR d = w 2 u
dx
e
e
dw du
du
dx dx wu d + w dx d (2.37)
in which e denotes an element boundary. The first expression of Equation 2.37 is called the
strong formulation, and the second expression is called the weak formulation. The Galerkin
method results in the test function such that
w1 =
du
du
= H i ( x) and w2 =
= H j ( x) (2.38)
du j
dui
In addition,
dH j
dH i
du
=
ui +
u j (2.39)
dx
dx
dx
Substitution of Equations 2.332.39 into Equation 2.37 gives
dw du
=
+
wu
d
dx dx
e
e

dH i
dx
dx
dH i
dx
dH j
dw1
dx
dw2
dx

du w1
+
dx w2
dH j H i
+
[H i
dx H j

d
u
(2.40)
u
i
H j ]
d
u j
12
Let us plug the shape functions and evaluate the integral for an element whose left nodal
coordinate is xi and its right nodal coordinate is xj. Then, the resultant matrix expression is
1 h
+
h 3
1 h
+
h 6
1 h
+
h 6
1 h
+
h 3
u
i
u
j
(2.41)
here, h = xj xi.
Let us apply this to each finite element as shown in Figure 2.2. Because each element has
the same value of h = 0.5, the corresponding matrices become
13
6
23
12
23
12
13
6
u
1
u2
(2.42)
23
12
13
6
u
2
u3
(2.43)
for the first element and
13
6
23
12
for the second element. The two matrices must be added as shown in Equation 2.36.
However, we cannot add 2 2 matrices on top of each other because the column vectors are
different. To add them, each matrix expression is expanded into a 3 3 matrix as follows:
13
6
23
12
0
23
12
13
6
0
u
1
u2
u
3
(2.44)
u
1
u2
u
3
(2.45)
for the first element and
0
13
6
23
12
0
23
12
13
6
13
for the second element. Then, the two matrices can be summed together as
13
6
23
12
23
12
13
3
23
12
23
12
13
6
u
1
u2
u
3
F
1
0
F
3
(2.46)
The right-hand side column vector comes from the boundary integral in Equation 2.37,
which is explained further in the next section. From the boundary conditions, u(0) = u1 = 0
23
. The finite element solution is compared to the exact soluand u(1) = u3 = 1. Then, u2 =
52
tion in Figure 2.1.
2.4Axial Bar and Beam

An axial bar carries a load along its longitudinal axis, while a beam carries a load through
its bending. In this section, a 1-D bar and a beam are considered. Finite element formulations are presented for an axial member and then for a beam.
2.4.1Axial Member
The axial member is a 1-D structure deforming longitudinally. Let us consider an infinitesimal element as a free-body diagram, as indicated in Figure 2.4. For simplicity, let us
assume a uniform cross-sectional area. The summation of force along the longitudinal
direction is

2 u
F = +
dx A A = Adx 2 (2.47)
x
t
where A is the cross-sectional area, is the axial stress, is the mass density, u is the axial
displacement, and t denotes time. Simplifying, the equation becomes
+ dx
x
dx
FIGURE 2.4
Free-body diagram of an infinitesimal axial member.
14
2 u
= 2 (2.48)
x
t

Hookes law states
= E (2.49)
in which E is the elastic modulus, and denotes the normal strain. The strain-displacement
relationship is
=
u
(2.50)
x
Plugging Equation 2.50 into Equation 2.49 and the resultant expression into Equation
2.48 yields
u
2 u
E = 2 (2.51)
x
x
t
If the elastic modulus is constant along the bar, it is simplified as

c2
2 u 2 u
=
(2.52)
x 2 t 2
where the speed of sound in the material is

c=
E
(2.53)
Let us apply the Galerkin finite element formulation to this equation. The weighted
residual expression is
2 u
2 u
w 2 c 2 2 d =
t
x
n
2 u 2 w u
u
2
w 2 + c
d c w d = 0
x
x
n
e =1
2 u 2 w u
u
2
d c w d
w 2 + c
n
x x
t
(2.54)
For a typical element, let us use the linear shape function as follows:
u( x , t) = H i ( x)ui (t) + H j ( x)u j (t) = H i ( x)
u (t)
i
H j ( x)
u j (t) (2.55)
15
This expression shows that the shape functions are used to interpolate the solution in the
spatial domain and the nodal variables vary as a function of time. As a result, their derivatives are computed as
H j ui
x u j
u H i
=
x x
(2.56)
and
u
= Hi
t
H j
(2.57)
ui
t
u j
Substitution of Equations 2.56 and 2.57 into Equation 2.54 for each element level gives
2 u
2 w u
d =
w 2 + c
x
x
t
e
+ c2
xi
xj
xj
xi
H
i
H i
H
j
H j dx
2 ui
t 2
2 u j
t 2
(2.58)
H i
x
x
H i
x
H j
H j ui
dx
x u j
Evaluation of the matrices in Equation 2.58 results in the following expression:
} + [ k ]{u} =
[m] {u
i
x j xi 2 1 u
c 2 1 1 ui

6 1 2 u j x j xi 1 1 u j (2.59)

in which the superimposed dot denotes a partial derivative in terms of time. Summing
element-level matrices yields
[M]{} + [K]{U} = {F} (2.60)
Let us discuss the column vector in Equation 2.60, which comes from the boundary
integral in Equation 2.54. For the 1-D problem, the whole domain is from 0 to L, and the
boundary integral is simplified as

x= L
u
u
u(L, t) 2 u(0, t)
c w
d = c 2 w
= c2w
c w
(2.61)
n
x
x x= 0
x
2
16
The first term is for the last element, which can be written as
c2w
H (L) u(L, t)
i
= c2
H
L
(
)
x
j
u(L, t)
=> c 2
x
0 u(L, t)

(2.62)
1 x
Therefore, the second component in Equation 2.62 goes to the last component of the column vector {F}. Likewise, the second term in Equation 2.61 is written as
c2 w
u(0, t)
=> c 2
x
H (0) u(0, t)
i
= c2
0
H
(
)
x
j
1 u(0, t)

(2.63)
0 x
The first term of Equation 2.63 contributes to the first term of {F}.
As an example, let us consider an axial bar that is 3 m long. The bar has an elastic modulus of 12 GPa and mass density of 2000 kg/m. The bar is fixed at one end, and a constant
force 2000 N is applied to the other end suddenly while the bar is at rest initially. The bar
is divided into five equal-length elements. Each element has
} + [ k ]{u} = 0.2
[m]{u
0.1

1
i
0.1 u
7
+ 10
0.2 u j
1

u
i
u
j
(2.64)
Summing all the matrices yields
0.2
0.1
0
0
0
0
0.1
0.4
0.1
0
0
0
+ 107 0
0
0
0
0
0.1
0.4
0.1
0
0
0
0
0.1
0.4
0.1
0
0
0
0
0.1
0.4
0.1
1
2
1
0
0
0
0
1
2
1
0
0
0
0
1
2
1
0
0
0
0
0.1
0.2
0
0
0
1
2
1
1
2
u
u
3
4
u
u
5
6
u
0
0
0
0
1
1
u
1
u2
u
3
u4
u5
u6
f
1
0
0
0
0
2000
(2.65)
The boundary condition states u1 = 0, and the initial conditions are

u1(0) = u2(0) = u3(0) = u4(0) = u5(0) = u6(0) (2.66)
17
and
u 1 (0) = u 2 (0) = u 3 (0) = u 4 (0) = u 5 (0) = u 6 (0) (2.67)
The technique to solve Equation 2.65 is discussed further in Section 2.4.3.

2.4.2Beam
A beam member supports transverse loading through bending like a bridge structure. An
infinitesimal length of a beam is shown in Figure 2.5 for its free-body diagram. The force
equilibrium in the transverse direction is

(V + dV ) + qdx V = Adx
2 v (2.68)
t 2
Here, v is the transverse displacement, is the mass density, and A is the cross-sectional
area. This equation is simplified to
dV
2 v
+ q = A 2 (2.69)
dx
t

The moment equilibrium gives

M (M + dM) Vdx/2 (V + dV)dx/2 = 0
(2.70)
This also simplifies to

dM
+ V = 0 (2.71)
dx
Combining Equations 2.69 and 2.71 yields

2 w 2 M
+
= q (2.72)
t 2
x 2
q
M
V
FIGURE 2.5
Free-body diagram of an infinitesimal beam.
M + dM
dx
V + dV
18
Now, the bending moment is related to the curvature of the deformation, such as
M = EI
2 v
(2.73)
x 2
where EI is the beam rigidity. Substitution of Equation 2.73 into Equation 2.72 results in
2 v
4v
+ EI 4 = q (2.74)
2
t
x
if EI is constant. This is the governing equation for classical beam theory. Applying the
weighted residual formulation to the differential equation gives
L
I=
2 v
4 v
w A 2 + EI 4 q dx = 0 (2.75)
t
x
This is the strong formulation. To develop the weak formulation, we take the integrations
by parts twice for the fourth-order term as follows:
L
I=
2 v
wA 2 dx +
t
2 w 2 v
3 v w 2 v
EI 2 = 0 (2.76)
EI
dx
wq
dx
+
wEI

x 2
x 2
x 3 0 x
x 0
Discretizing the domain gives

n
I=
i=1
2 v
wA 2 dx +
t
e
i=1
2 w 2 v
EI 2 dx +
x 2
x
e
i=1
L
w
wq dx wV
M = 0 (2.77)
0
x 0
e
in which n is the number of elements.

For the beam element, we need shape functions. However, the shape functions given
in Equations 2.31 and 2.32 cannot be used for the beam bending. The shape functions
in Equations 2.31 and 2.32 provide a continuous solution from one element to the next
element, but their derivative is not continuous between two neighboring elements. As a
result, these shape functions are called C0-type shape functions, that is, continuity up to
the zeroth order of the shape functions.
For beam-bending solutions, not only deflections but also slopes must be continuous
from one element to the next element. Therefore, the beam element has both deflection and
slope as nodal degrees of freedom, as illustrated in Figure 2.6. In addition, for the classical
beam-bending theory, the deflection and slope are related to each other as follows:

v
(2.78)
x
19
v2
v1
FIGURE 2.6
Beam-bending element.
Therefore, the shape function should be the C1 type, that is, the functions and first-order
derivatives must be continuous. To develop such shape functions, let us start with a cubic
polynomial function such as the following:

v = a0 + a1x + a1x2 + a1x3 (2.79)
The reason we use the cubic function is that it has four coefficients, and the number of
the nodal variables for the beam element shown in Figure 2.6 is also four, so the cubic function can uniquely interpolate the solution within the element.
Now, let us replace the coefficients ai by the nodal variables. To do that, we evaluate the
cubic function and its derivative at the nodal points as follows:
v
1
1
v2
2
0
1
0
1
l
1
0
0
l2
2l
0
0
l3
3l 2
a
0
a1
a2
a3
(2.80)
or, in a short notation,

{v} = [C]{a} (2.81)
Solving for {a} and substituting it into Equation 2.79 yields

v = {X}T [C]1 {v} (2.82)
where
{X}T = {1 x x2 x3} (2.83)
The shape functions are
{H}T = {X}T [C]1 (2.84)
Explicit expressions for the shape functions are

H 1 ( x) = 1
3x 2 2 x 3
+ 3 (2.85)
l2
l
20
2x2 x3
+ 2 (2.86)
l
l
H 2 ( x) = x
3x 2 2 x 3
3 (2.87)
l2
l
H 3 ( x) =
H 4 ( x) =
x2 x3
+ 2 (2.88)
l
l
Substitution of the shape functions into Equation 2.77 gives the following matrix expression, where the first term in the equation yields the elemental mass matrix:
[m]{ v} =
H
1
H
A 2
H3
H4
H1
156
Al 22l
=
420 54
13l
H2
22l
4l 2
13l
3l 2
v
1
1
H3
H 4 dx
v
2
2

13l v1
3l 2 1
22l v2

4l 2
2
54
13l
156
22l
(2.89)
The second term in the equation yields the elemental stiffness matrix:
[ k ]{ v} =
H ,
1 xx
H ,
EI 2 xx
H 3 , xx
H 4 , xx
12
EI 6l
= 3
l 12
6l
H 1 , xx
6l
4l 2
6l
2l 2
H 2 , xx
12
6l
12
6l
6l
2l 2
6l
4l 2
H 3 , xx
v
1
1
v2
2
H 4 , xx
v
1
dx 1
v2
2
(2.90)
The third term yields the elemental force vector:

l
{f} =
H
1
H2
H3
H4
q( x) dx
(2.91)
21
If the applied load intensity function q(x) is constant within the element and its value is
qo, then the elemental force vector becomes
ql
o
2
qo l 2
{ f } = 12
qo l
2
2
qo l
12
(2.92)
If the load intensity function is linear over the element, such as q(x) = cx, then, the elemental load vector becomes
3cl
20
cl 2
{ f } = 30
7 cl
20
2
cl
20
(2.93)
Other loading cases can be found in Reference 1.

Assembly of all the elemental matrices and vectors yields the following matrix equation:
} + [ K ]{V } = { F } (2.94)
[ M]{V
As an example, let us consider a cantilever beam with a tip load P as shown in Figure 2.6.
We will use one element to represent the beam. Then, the system mass and stiffness matrices are equal to the elemental mass and stiffness matrices as given in Equations 2.89 and
2.90, respectively. The force vector becomes
{F}T = {f1 m1 P 0} (2.95)
where f1 and m1 are the unknown reaction force and moment at the clamped node 1 of
Figure 2.7. The boundary condition states v1 = 0 and 1 = 0. The initial conditions state zero
0)} = {0}.
nodal variables {V(0)} = {0} and zero nodal velocities {V(
2.4.3Solution Techniques
Let us consider the equations of motion as follows:

} + [C] {U } + [ K ]{U } = { F } (2.96)

[ M]{U
22
v2
v1
1
I
P
FIGURE 2.7
Cantilever beam with a tip load.
We discuss some techniques to solve the equations of motion. One way to solve the second-
order differential equation is to break it into two first-order differential equations. To this
end, let us say
{ V} = {U } (2.97)

Then, Equation 2.96 can be written as
[ M]{V } + [C]{V } + [ K ]{U } = { F } (2.98)
Putting the two equations together gives

[ M]
[0]
[0]
[ I ]
{V }
+
{U }
[C]
[ I ]
[K ]
[0]
{V }
=
{U }
{ F }
(2.99)
{0}
For convenience, we rewrite it in a simpler form at time t:

t
[ A]{X } + [B]{X }t = { P}t (2.100)
Equation 2.100 can be solved using the finite difference method, such as using backward,
forward, or Crank-Nicholson techniques [1].
If the forward difference technique is used, the temporal derivative is expressed as

{X }t =
{X }t+t {X }t
(2.101)
t
in which t is the time step size. Substitution of Equation 2.101 into Equation 2.100 and
rearrangement of the resulting equation yields
[A]{X}t+t = t{P}t t[B]{X}t + [A]{X}t (2.102)
The equation is solved as time progresses. For example, we set t = 0. Because {X}0 is given
from the initial condition, {X}t is solved after applying the boundary condition. Then, we
obtain {X}2t from the previous solution {X}t. This process continues until the time reaches
23
the termination time set by the user. If the matrix [A] is a diagonal matrix, Equation 2.102
does not require inverting the matrix. It is called the explicit solver, and it is conditionally
stable. In other words, if the time step size t is not so small, the solution diverges as time
increases.
To make [A] a diagonal matrix, the matrix [M] should be a diagonal matrix. The matrix
[M] in Equation 2.59 is transformed into a diagonal matrix, such as
[m] =
x j xi 1
2 0
0
(2.103)
1
Likewise, the diagonal mass matrix for the beam element is expressed as
0
Al
[m] =
2 0
0
l2
78
0
(2.104)
0
l2
78
0
0
These are called diagonal mass matrices.

On the other hand, the backward difference technique is an implicit solver and requires
inversion of a matrix. The technique is unconditionally stable. That is, the solution is stable
regardless of the time step size. The backward difference technique uses
{X }t =
{X }t {X }tt
(2.105)
t
Plugging Equation 2.105 into Equation 2.100 and rearranging yields

([A] + t[B]{X}t = t{P}t + [A]{X}tt) (2.106)
This is solved in the same way as for the forward difference technique.
The other way is to solve the second-order differential equation as it is. We use the central difference technique as follows:
t
{U }t+ 2 = {U }t 2
}t (2.107)
+ t {U
{U }t+t = {U }t + t {U }
t+
t
2
(2.108)
To solve the problem, we first compute the initial acceleration from the initial displacement and velocity, such as the following:

} = { F } [C]{U } [ K ]{U }0 (2.109)

[ M]{U
24
Then, the velocity is computed from Equation 2.107 and the displacement is computed
from Equation 2.108. This process repeats itself. One thing to note is that the velocity is
t
computed at time t + . To compute the velocity at t, we use the average; for instance,
2
{U }t = {U }
t+
t
2
+ {U }
2
t
2
(2.110)
The central difference technique is also an explicit solver and a conditionally stable
scheme. The time step size should be smaller than the critical step size, which can be
computed as
tc =
x
(2.111)
cs
where x is the element length and cs is the wave speed in the material.
An unconditionally stable technique is the Newmark method [1]. This technique can use
any time step size without worrying about the instability of the solution. Of course, the
solution accuracy is proportional to the time step size regardless of the stability.
2.5Truss and Frame

Truss structures consist of axial members, which are two-force members; the structures
carry loads along the structural members axes. Every member of a truss structure is in
either tension or compression. On the contrary, frame structures can carry loads transverse to structural members orientations. In this section, the finite element formulation is
presented for both structures.
2.5.1Truss
A truss structure has axial members connected by pin joints as shown in Figure 2.8. Let us
us consider a truss element in two dimensions (2-D) as shown in Figure 2.9. It shows two
different coordinate systems; one is the global coordinate system, and the other is the local
coordinate system, which rotates along with the truss member. All the variables associated
with the former are denoted using capital letters; those for the local axis are symbolized
using lowercase letters. Now, let us examine the relationship between the two coordinate
systems. The displacements in those coordinates are related to each other as
u
1
v1
u2
v2
cos
sin
0
sin
cos
cos
sin
0
sin
cos
0
U
1
V1
U2
V2
or {d} = {T }{D}
(2.112)
25
FIGURE 2.8
Truss structure.
y, v
x, u
Y, V
X, U
FIGURE 2.9
A 2-D truss element.
where the angle is the measure of the rotation of the local system from the global system.
It is considered positive along the counter-clockwise direction.
The stiffness matrix of the truss member is expressed as
AE 0
[ klocal ]{d} =
l 1
0
0
0
0
0
1
0
1
0
0
0
0
u
1
v1
u2
v2
(2.113)
Because the strain energy should be the same no matter how it is expressed in terms of any
coordinate system, we can write

1 T
1
{d} klocal {d} = {D}T k global {D} (2.114)
2
2
26
Substitution of Equation 2.113 into Equation 2.114 results in

c2
AE cs
T
[ k global ] = [T ] [ klocal ][T ] =
l c2
cs
c2
cs
c2
cs
cs
s2
cs
s2
cs
s2
(2.115)
cs
s2
Here, c = cos and s = sin .

Every truss element must be transformed as shown in Equation 2.115 and assembled together
to make up the system stiffness matrix. On the other hand, the diagonal mass matrix is
c2
Al 0
[m] =
2 0
0
s2
0
0
0
0
c2
0
0
0
0
s2
(2.116)
A similar derivation can be conducted for a three-dimensional (3-D) truss, whose details
are given in Reference 1.
2.5.2Frame
Let us us consider a 2-D frame element as shown in Figure 2.10. In terms of the local coordinate, the element stiffness matrix is a combination of those from the 1-D axial and beam
members as expressed as follows:
Al 2
0
E 0
[ klocal ]{d} = 3
l Al 2
0
0
12 I
6 Il
0
12 I
6 Il
0
6 Il
4 Il 2
0
6Il
2 Il 2
Al 2
0
0
Al 2
0
0
0
12 I
6 Il
0
12 I
6 Il
6 Il
2 Il 2
0
6 Il
4 Il 2
u
1
v1
u2
v2
(2.117)
The transformation matrix between the local and global axes is
u
1
v1
u2
v2
s
0
0
0
0
s
c
0
0
0
0
0
0
1
0
0
0
0
0
0
c
s
0
0
0
0
s
c
0
0
0
0
0
1
U
1
V1
U2
V2
or {d} = [T ]{D} (2.118)
27
y, v
x, u
Y, V
X, U
FIGURE 2.10
A 2-D frame element.
After the coordinate transformation, the element matrix in the global coordinate system
becomes
a
11
a21
a
[ k global ] = 31
a41
a51
a
61
a12
a13
a14
a15
a22
a23
a24
a25
a32
a33
a34
a35
a42
a43
a44
a45
a52
a53
a 54
a55
a62
a63
a64
a65
a16
a16
a36
a46 (2.119)
a56
a66
where
AE 2 12EI 2
a11 = a14 = a44 =
c + 3 s (2.120)
l
l
AE
12EI
a12 = a15 = a24 = a45 =
cs 3 cs (2.121)
l
l
6EI
a13 = a16 = a34 = a46 = 2 s (2.122)
l
AE 2 12EI 2
a22 = a25 = a55 =
s + 3 c (2.123)
l
l
6EI
a23 = a26 = a35 = a56 = 2 c (2.124)
l
28
a33 = a66 =
a36 =
4EI
(2.125)
l
2EI
(2.126)
l
aij = aji (2.127)

The diagonal mass matrix for the 2-D frame is given as
1
0
Al
[m] =
2 0
0
0
0
0
0
0
0
0
0
l2
78
0
0
1
0
0
1
0
1
0
0
(2.128)
0
0
l2
78
The 3-D frame element matrices can be developed similarly (as given in [1]).
2.6Solid Element
In this section on the solid element, the theory of elasticity is briefly presented, and the
finite element formulation is followed. Even though the 2-D theory and its formulation are
given here, it is straightforward to extend them to 3-D problems [1].
2.6.1Theory of Elasticity
Let us consider a 2-D infinitesimal element for a free-body diagram as shown in Figure2.11.
There are normal stresses, denoted by , and shear stresses, designated as . The equations
of equilibrium can be developed as shown next. The summation of force in the x axis is
x
dx dy x dy
yx
2 u
+ yx +
dy dx yx dy + f x dxdy = dxdy 2
y
t
F =
(2.129)
29
y +
dy
yx +
xy
yx
y
dy
dy
x +
dx
x +
yx
x
x
x
x
dx
dx
y
FIGURE 2.11
Free-body diagram for 2-D infinitesimal element.
in which fx is the body force per unit volume along the x axis and u is the displacement in
the x axis. Simplifying the expression yields
x yx
2 u
+
+ f x = 2 (2.130)
x
y
t
Similarly, the force equilibrium in the y axis gives

xy
y
y
+ fy =
2 v
(2.131)
t 2
The moment equilibrium about the z axis, which is in the direction perpendicular to the
xy plane, states the stress tensor is symmetric, that is,
xy = yx (2.132)
The next set of equations presents the relationship between stresses and strains. These
are called constitutive equations. For an isotropic material, their relationship is written as
E
x

y =
E
xy
0
E
1
E
1
G
x
y
xy
or {} = [C]{ }
(2.133)
30
The inverse relationship is written as
x
y
xy
2
1 v
E
2
1 v
0
E
1 v2
E
1 v2
0

x
y
xy
} = [D]{}
or {

(2.134)
The last set of equations is the kinematic equations, which relate strains to displacements as follows:
u
x
x v
y = y

xy u v
+
y x
(2.135)
Finally, there are two types of boundary conditions. One type is the given displacement,
and the other is the prescribed traction, which can be expressed as
p
x
py
n + n
xy y
x x
xy nx + y ny
(2.136)
where px and py are the given traction, and nx and ny are the outward normal unit vector
components at the given boundary.
2.6.2Finite Element Formulation
Let us apply the finite element formulation for the governing equations and the boundary
conditions developed previously. To do that, we begin with the equations of motion and
apply the weighted residual formulation to those equations:
2
w1 u
2
2 v

w2 t 2

x xy
fx
x
y

d = 0

xy y
fy
y
x

(2.137)
where w1 and w2 are two test functions. Now, we apply integrations by parts to this expression and substitute Equation 2.136 into the resultant expression to yield the following
equations:
31
w1
w
1
2 u
t 2
2 v
t 2
d +
w f
1 x
w2 f y
d +
w
1
w1
y
y d
w2

y xy
(2.138)
0
w2
y
w p
1 x
w2 py
The stress terms are replaced by the strain terms using Equation 2.134, and then the
strain terms are replaced by the displacement terms using Equation 2.135. The resultant
mathematical expression becomes
w1
w
1
w
1
u
t 2
2 v
t 2
d +
w f
1 x
w f
2 y
d +
d =
w1
y
w2
y
w2
y
w p
1 x
w p
2 y
[D] y
u v (2.139)
y x
Now, a problem domain is discretized into a number of finite element domains as

sketched in Figure 2.12, and each finite element has a triangular shape with nodes at the
vertices as seen in Figure 2.13. We will develop shape functions for the triangular element.
The element has three nodes (nodes i, j, k), and each node has a nodal coordinate value as
shown in Figure 2.13. Let node i have nodal variables (ui, vi), and the other two nodes have
similar nodal displacements. Then, to interpolate the displacements inside the element in a
unique way, the lowest order of polynomial function is a linear function in terms of x and
y, such as the following:
FIGURE 2.12
Finite element mesh.
32
Node j (xj, yj)

Node k (xk, yk)
Node i (xi, yi)

FIGURE 2.13
Triangular element.
{1
u = a0 + a1 x + a2 y =
a
0
a1
a
2
(2.140)
As we evaluate the displacement at the three nodes, we obtain

u
i
uj
uk
yi
yj
xk
xi
xj
xk
a
0
a1
a
2
(2.141)
Inverting the matrix and substituting it into Equation 2.140 results in
u = {H i
Hj
u
i
H k } uj
uk
(2.142)
where
) (
) (
) (
) (
) (
) (
H1(x, y) =
1
x j y k xk y j + y j y k x + xk x j y (2.143)
2A
H 2 (x, y) =
1
xk y i xi y kj + y k y i x + xi xk y (2.144)
2A
H 3 (x, y) =
1
xi y j x j y i + y i y j x + x j xi y (2.145)
2A
33
and
1
1
A = det 1
2
yi
y j (2.146)
xk
xi
xj
xk
Because the two displacements u and v are independent, the same shape functions can
be used for both displacements. That is,
v = {H i
v
i
Hk } vj
vk
Hj
(2.147)
Substitution of the displacements into the kinematic equation gives

{} = [B]{d} (2.148)
where

{}T = { x
[B] =
xy } (2.149)
H 1
x
H 2
x
H 3
x
H 1
y
H 2
y
H 1
y
H 1
x
H 2
y
H 2
x
H 3
y
{d}T = {u1
v1
u2
v2
u3
H 3
(2.150)
y
H 3
v3 } (2.151)
In addition, the same shape functions are used for the test functions w1 and w2. Then, we
obtain the following expression:
w
1
0
H
0
= 1
w2
0
H2
H3
H1
H2
0
= [N ]
H 3
(2.152)
34
Elemental mass matrix can then be computed as follows:

2 u
2
t
2
v
t 2
w1
[m] =
w
e
1
0
A 1
=
12 0
1
0
2 u
t 2
2 v
t 2
= [ N ]{d}
1
d =
0
w2
0
2
0
1
0
1
1
0
2
0
1
0
0
1
0
2
0
1
(2.153)
2 u
t 2
2 v
t 2
T
d = [ N ] [ N ] d
1
0
1
0
2
1
0
1
0
2
0
(2.154)
The diagonal mass matrix is

[m] =
A
[ I ] (2.155)
3
where [I] is a 6 6 identity matrix.

The elemental stiffness matrix becomes
[k ] =
w
1
w1
y
w2
y
w2
y
T
B] d (2.156)
[D] y d = [B] [D][B
u v
y x
Evaluation of matrix [B] using the shape functions gives
(y y )
j
k
1
0
[B] =
2A
( xk x j )
(y k yi )
(yi y j )
( xk x j )
( xi x k )
(y j y k )
( xi x k )
(y k yi )
( x j xi )
( x j xi ) (2.157)
(yi y j )
35
Because matrix [B] is constant (i.e., not a function of x and y), the element stiffness matrix
can be evaluated easily over the element domain, and the resultant matrix becomes
[k] = [B]T[D][B]A (2.158)
Finally, the force column vector consists of two parts: body force and boundary traction.
The column vector due to the body force is computed as
w f
1 x
w2 f y
e
= [ N ]T
{ fb } =

d =
e
f x
d
fy
w
1
0
0 f x
d
w2 f y

(2.159)
and the traction force vector is

w p
w
1 x
1
d =
w
P
0
2 y
e
e
p
x
= [ N ]T d
py
e
{ ft } =

0 px
w2 py
(2.160)
One of the main differences between the two force vectors is that the body force vector is an
integral over the domain while the traction force vector is an integral over the boundary. As a
result, the shape functions over the domain are simplified when they are evaluated along the
boundary. In other words, 2-D shape functions for the 2-D domain become 1-D shape functions
along the boundary. Let us say the element boundary ij is located along the boundary of the
problem domain. Then, the shape function Hk disappears along the element boundary, and
shape functions Hi and Hj become 1-D functions that have the axis along the boundary.
For example, if s denotes the axis along the boundary, the shape functions used for the
boundary integral are

H i ( s) =
sj s
(2.161)
s j si
H j ( s) =
s si
(2.162)
s j si
Then, the traction force vector becomes
sj
{ ft } =
si
H
i
0
Hj
0
Hi
0
H j
p
x
p
y
ds (2.163)
36
If the traction is constant over an element boundary that is located on the problem domain
boundary, the corresponding column vector becomes
p
x
s j si py
{ ft } =
2 px
p
y
(2.164)
Assembly of element matrices and vectors results in the system matrix equation as
follows:

} + [ K ]{D} = { F } (2.165)
[ M]{D
in which

[ M] =
[m] (2.166)
[K ] =
[k] (2.167)
{F} =
{ f } + { f } (2.168)
b
The set of equations can be solved using one of the techniques discussed previously
using the finite difference method for the temporal derivative.
2.7Isoparametric Formulation
The linear triangular shape functions discussed previously are simple to use and do not
require integration over the domain because the matrix [B] is a constant matrix. However,
to obtain an accurate approximation, we have to use a large number of elements. Therefore,
other types of finite elements and the corresponding shape functions were developed.
Among them, isoparametric shape functions have been the most popular.
For the isoparametric formulation, we consider mathematical mapping between two different coordinate systems. One coordinate system is called the natural coordinate domain,
and the other coordinate system is called the physical coordinate domain. The problem
domain is defined in terms of the physical coordinate system. As a result, the finite element mesh is made in the physical coordinate domain, and every finite element may have
different shapes. However, shape functions are defined in terms of the natural coordinate
system. As long as each finite element has the same kind of polygon, mostly the quadrilateral shape, with the same number of nodes per element, the same shape functions defined
37
in the natural coordinate domain can be used for all those elements with proper mapping.
The 2-D formulation is presented in the following material.
The two coordinate systems are shown in Figure 2.14 along with their quadrilateral elements. In the natural coordinate denoted by (, ), the square shape of the element remains
the same all the time. On the other hand, the quadrilateral element in the physical coordinate system denoted by (x, y) can be in any location and of any shape depending on what
element is chosen in the mesh of the physical problem. Each node is numbered sequentially in the counterclockwise direction in both coordinate systems.
Shape functions are defined in terms of the natural coordinate system. They can be
determined using the 1-D shape functions in both axes. In other words, the shape functions along the axis are
1 ( ) =
1
(2.169)
2
1.0
1.0
1.0
(a)
1
1.0
y
3 (x3, y3)
2 (x2, y2)
4 (x4, y4)
1 (x1, y1)
x
(b)
FIGURE 2.14
Isoparametric formulation: (a) natural coordinate and (b) physical coordinate.
38
2 ( ) =
1+
(2.170)
2
1 ( ) =
1
(2.171)
2
2 ( ) =
1+
(2.172)
2

Similarly, along the axis are
Then, the shape functions for the element in the natural coordinate system are

H 1 ( , ) = 1 () 1 ( ) =
1
(1 )(1 ) (2.173)
4
H 2 ( , ) = 2 () 1 ( ) =
1
(1 + )(1 ) (2.174)
4
H 3 ( , ) = 2 () 2 ( ) =
1
(1 + )(1 + ) (2.175)
4
H 4 ( , ) = 1 () 2 ( ) =
1
(1 )(1 + ) (2.176)
4
To interpolate the physical variable in the physical domain, we use shape functions such
as
4
u( x , y ) =
H (, )u (2.177)
i
i=1
where ui is the nodal variable. The differential equation to be solved in terms of the physical coordinate system requires derivatives in terms of x and y. However, the shape functions are expressed in terms of and . As a result, we need to relate the two coordinate
systems using mathematical mapping. To do that, we use the following geometrical mapping from the natural to physical coordinate systems:
4
x=
H (, )x (2.178)
i
i=1
y=
H (, )y (2.179)
i
i=1
39
These expressions state that the ith node in the natural coordinate is mapped to the ith
node in the physical coordinate shown in Figure 2.14. Likewise, the inside and boundary
of the square element in the natural coordinate are mapped to the inside and boundary of
the quadrilateral in the physical domain.
Let us compute the Jacobian matrix associated with the mapping. It is defined as
[ J ] =

(2.180)
y
The components of the Jacobian matrix are computed as follows:

J11 =
J12 =
J 21 =
x
=
y
=
x
=
y
J 22 =
=
H (, ) x (2.181)
i
i=1
4
H (, ) y (2.182)
i
i=1
H (, ) x (2.183)
i
i=1
H (, ) y (2.184)
i
i=1
Now, let us discuss how to take derivatives of the shape functions Hi(, ) in terms of
x and y. To this end, we consider functions (x, y) and (x, y) are functions of (, ). Taking
derivatives of (x, y) with respect to (, ) results in the following expressions using the
chain rule:
(2.185)
Thus, premultiplying the inverse of the Jacobian matrix to the equation yields

(2.186)

40
These expressions are used for the shape functions. In other words, we have
H
i
H i
y
H i

(2.187)
H i

As an example, we apply the isoparametric formulation to the 2-D solid elements. Then,
we need to compute the matrix [B] for a quadrilateral shape of finite element in the physical
domain. The matrix is expressed as
[B] =
H 1
x
H 2
x
H 3
x
H 4
x
H 1
y
H 2
y
H 3
y
H 1
y
H 1
x
H 2
y
H 2
x
H 3
y
H 3
x
H 4
y
H 4
(2.188)
y
H 4
and it is a function of (, ). The element stiffness matrix is written as

1 1
[k ] =
[B] [D][B] d = [B] [D][B] J d d (2.189)

T
1 1
where |J| is the determinant of the Jacobian matrix and is equal to J11J22 J12J21.
Evaluation of the double integrals in Equation 2.189 is undertaken numerically. The
Gauss-Legendre quadrature rule is used for the numerical integration. Any numerical
integration can be expressed in terms of sampling points and weight factors. For example,
a function f() can be integrated numerically, and its result is written as
1
f ( ) d =
W f ( ) (2.190)
i
i=1
in which Wi is the weight factor for the sampling point i. In addition, n is the number of
sampling points for numerical integration. Some sampling points and their corresponding weight factors are listed in Table 2.1 for the Gauss-Legendre quadrature, which is the
most commonly used numerical integration technique for isoparametric finite elements.
To extend this to double integrals, we apply the procedure one by one as follows:
1 1
f ( , ) d d =
W f ( , )d
i
1 i= 1
1 1
Wj
j=1
m
i=1
Wi f ( i , j ) =
(2.191)
j=1 i=1
WjWi f ( i , j )
41
TABLE 2.1
Gauss-Legendre Quadrature Sampling Points and Weight
Factors
No. of Sampling
Points
Sampling Point
Coordinate Value
n=1
n=2
Weight Factor
0.0
1
2.0
1.0
1.0
3
n=3
15
5
0
5
9
8
9
15
5
5
9
Therefore, the same sampling points and weight factors as in Table 2.1 can be used in both
integrals. Because the shape functions have balanced expressions in terms of (, ), the
number of sampling points is equal in both directions.
The Gauss-Legendre quadrature states that any polynomial function of order 2n 1 can
be numerically integrated exactly. In other words, the following integral
1
(a + a x + a x
0
+ + a2 n1x 2 n1 dx (2.192)
requires n number of sampling points for exact integration.
2.8Plate and Shell Structures

Plates and shells are common structural components because they are efficient for carrying loading in the transverse direction. The plate is the 2-D extension of the beam. While
the plate has a flat surface, the shell has a curved surface. As a result, the beam is a subset
of the plate, and the plate is a subset of the shell. This section presents a shell formulation
that can be easily degenerated into a plate or a beam element.
The shell element developed here is like a 3-D solid element. It has four nodes at the
bottom and top surfaces, respectively, leading to eight nodes total. In addition, the shell
element has three displacements as nodal degrees of freedom but no rotational degrees
of freedom. In that aspect, the shell element looks the same as the 3-D solid element.
However, the constitutive equation is different between the two elements. Furthermore,
42
the shell element can model a geometry with a large aspect ratio (i.e., the ratio of the side
length to the thickness).
The advantage of the present shell element is that the shell element can be easily stacked
on one another without special consideration at the interface between two neighboring
shell elements. The shell element can be connected to the 3-D solid elements directly as it
is, and it defines clear interfaces when in contact with other media, such as fluid.
Let us consider three displacements in terms of the local coordinate system of the shell
element, which is attached to the element. In-plane displacements u and v are parallel to
the shell surface, and the transverse displacement w is normal to the shell surface as seen
in Figure 2.15. The displacements are

u = u(x, y, z) (2.193)
v = v(x, y, z) (2.194)
w = w(x, y, z) (2.195)
where (x, y, z) is the local coordinate system. These displacements are interpolated using
the following shape functions:
u( x , y , z) =
v( x , y , z) =
n1
n2
i=1
j=1
n1
n2
i=1
j=1
N (, )H ()u (2.196)
i
ij
N (, )H ()v
i
y, v
z, w
Y, V
Global coordinate
X, U
FIGURE 2.15
Shell element in terms of its local and global coordinate systems.
(2.197)
Local coordinate
x, u
Z, W
ij
43
w( x , y , z) =
n1
n2
i=1
j=1
N (, )H ()w
i
ij
(2.198)
in which Ni(, ) is the 2-D shape functions for the in-plane interpolation, and Hj() is the
1-D shape function for the transverse interpolation. Both shape functions are expressed in
terms of the natural coordinate system (, , ) because the isoparametric formulation is
used. Because we use four nodes on the in-plane and two nodes along the thickness direction, n1 = 4 and n2 = 2. Of course, different numbers of nodes may be used for the in-plane
and thickness directions, respectively. However, the total eight nodes are selected because
they resemble the popular eight-node brick element.
For the shell formulation, we consider in-plane strains, transverse shear strains, and the
transverse normal strain independently. For the classical formulation, the transverse shear
and normal strains are neglected. However, the present formulation includes those as a
higher-order formulation. The in-plane strains are expressed as follows:

x
{ inpl } = y =

xy
u

v (2.199)
w

The transverse shear strains are also expressed in the following manner:

xz
{trans } =
yz
u

v (2.200)
w

Finally, the transverse normal strain is given as

z =
w
(2.201)
z
The constitutive equations are

{inpl} = [Dinpl]{inpl} (2.201)
{trans} = [Dtrans]{trans} (2.202)

z = Dtransnz (2.203)
44
is
If the material is orthotropic, the material property matrix for the in-plane components
D
11
[Db ] = D12
0
D12
D22
0
0 (2.204)
D33
where

D11 =
E1
(2.205)
1 v12 v21
D12 =
E2 v21
(2.206)
1 v12 v21
D22 =
E2
(2.207)
1 v12 v21
D33 = G12 (2.208)
Here, Ei and ij are the elastic modulus and Poissons ratio, respectively, and Gij is the shear
modulus of the in-plane directions.
The material property matrix for the transverse shear components is expressed as
G
[Dtrans ] = 13
0
0
(2.209)
G23
and the transverse normal component has

Dtransn = E3 (2.210)
As the shape functions are plugged into the kinematic equation, we obtain
{inpl} = [Binpl]{d} (2.211)
where
[Binpl] = [Binpl1 Binpl2] (2.212)
45
[Binpli ] =
H i
H i

N 1
x
0 Hi
N 2
x
0 Hi
N 3
x
0 Hi
N 4
x
Hi
N 1
y
Hi
N 2
y
Hi
N 3
y
N 1
y
Hi
N 1
x
0 Hi
N 2
y
Hi
N 2
x
0 Hi
N 3
y
Hi
N 3
x
0 Hi
N 4
y
N 4
0
Hi
y
N 4
0
Hi
x

(2.213)
0
{d}T = {d11 d21 d31 d41 d12 d22 d32 d42} (2.214)
{dij} = {uij vij wij} (2.215)
Furthermore, the transverse shear strains are expressed as follows:
{trans} = [Btrans]{d} (2.216)
where
[Btrans ] = [Btrans1
Btrans 2 ] (2.217)
[Btransi ] =
N 3
H 1
N 1
H 1
N 2
H 1
H 1
N 4
N1
H1
N2
H1
N3
0
0
0
H1
N4
0
H1
z
x
z
x
z
z
x
N 3
N 2
H 1
H 1
N 4
H 1
N 1
H 1
N1
H1
N2
0
0
H1
0
N3
H1
0
N4
H1
y
z
y
z
y
z
y
z
(2.218)
The transverse normal strain is

z = {Btrann}T{d} (2.219)
in which
{Btrann }T =
{B
T
trann1
T
Btrann
2
} (2.220)
46
N i
{Btranni }T = 0 0 H 1
z
0 0 H1
N i
z
0 0 H1
N i
z
0 0 H1
N i
z (2.221)
The element stiffness matrix in terms of the local coordinate can be written as
[ Klocal ] =
[B
inpl
]T [Dinpl ][Binpl ]d +
[B
trans
]T [Dtrans ][Btrans ]d
(2.222)
{Btrann }T Dtrann {Btrann }d
The first term represents the bending energy, the second term represents the transverse
shear energy, and the last term comes from the transverse normal energy.
Finally, we need to transform the local displacements into the global displacements. To
this end, we use the direction cosines between the x axis and the global axis denoted by
(l1,m1, n1), the direction cosines between the y axis and the global axis denoted by (l2, m2,n2),
and the direction cosines between the z axis and the global axis denoted by (l3, m3,n3).
The compatibility of displacements between the local and global coordinate systems can
be written as
ui 1 + ui 2
v i1 + v i 2
w i1 + w i 2
l
1
l2
l
3
m1
m2
m3
n1
n2
n3
U i1 + U i 2
V i1 + V i 2
W i 1 + W i2
(2.223)
and
ui 2 ui 1
i2
v v i1
i2
i1
w w
U i 2 U i1
i2
V V i1
i2
i1
W W
(2.224)
where the superscript i denotes the node from 1 to 4. Solving the two sets of equations
results in
ui 1
i1
v
w i 1
i2
u
vi2
i2
w
l +1
l2
1 l3
=
2 l1 1
l2
m1
n1
l1 1
m1
m2 + 1
n2
l2
m2 1
m3
n3 + 1
l3
m3
m1
n1
l1 + 1
m1
m2 1
n2
l2
m2 + 1
m3
n3 1
l3
m3
n2
n3 1
n1
n2
n3 + 1
n1
U i1
i1
V
W i 1
i2
U
V i2
i2
W
(2.225)
47
Applying this transformation expression to all sets of nodal displacements results in the
following equation:
{dlocal} = [T]{dglobal} (2.226)
Then, the element stiffness matrix in terms of the global coordinate system can be written as
[Kglobal] = [T]T[Klocal][T] (2.227)
This element stiffness matrix can be assembled into the system stiffness matrix.
2.9Acoustic Wave Equation

Sound wave propagates in a fluid medium. To develop the acoustic wave equation, fluid
flow and its viscosity are neglected in the fluid medium. We will develop the velocity
potential formulation for the acoustic wave equation. To do that, let us begin with the
continuity equation:

+ v = 0 (2.228)
t
in which is the fluid density, v is the velocity vector, is the gradient operator, and t
denotes time. The change in density can be written as
= o(1 + s) (2.229)
where s is called condensation, and o is an ambient fluid density. Plugging Equation 2.229
into Equation 2.228 with an assumption of s being so mall (i.e., s 1) results in the following
equation:
s
+ v = 0 (2.230)
t
Applying Newtons 2nd law yields

vo
v
= p (2.231)
t
Here, p is the pressure. The velocity potential is defined to express the velocity as follows:

v = (2.232)
48
Now, we substitute Equation 2.232 into Equation 2.231, which yields

o
+ p = 0 (2.233)
t
Pressure is proportional to condensation, and the proportional constant is called the

bulk modulus; it is written as

p = Bs (2.234)
where B is the bulk modulus. Plugging Equations 2.232 and Equation 2.234 into Equation
2.230 results in

1 p
2 = 0 (2.235)
B t
Elimination of pressure from Equations 2.233 and 2.235 produces the final wave equation in terms of the velocity potential:
2 2 2
c = 0 (2.236)
t 2
where c is the speed of sound, and c2 = B/o. The boundary conditions to the acoustic wave
equation are either prescribed velocity or pressure. The velocity is computed from the
velocity potential as defined in Equation 2.232. On the other hand, the pressure boundary
condition is expressed from Equation 2.233, such as

p = o
(2.237)
t
The acoustic wave equation can also be expressed in terms of pressure instead of velocity potential:

2 p 2 2
c p = 0 (2.238)
t 2
Either equation can be used depending on which expression is easier for a given application. For the fluid-structure interaction, the velocity potential formulation is selected, as
will be discussed below. For a 1-D case, the acoustic wave equation is simplified to

2 p 2 2 p
c
= 0 (2.239)
t 2
x 2
This is the same kind of equation as Equation 2.52.
49
The weighted residual expression for Equation 2.236 yields

2
w 2 c 2 2 d =
t
d + w ud w d = 0 (2.240)
2
n
t
Discretizing the domain with a finite element mesh, this equation is rewritten as
ne
e = 1 e
2
d +
t 2
ne
w ud =
e = 1 e
nb
e =1
d (2.241)
n
If we use the linear triangular element for a 2-D domain problem, the first term in
Equation 2.241 becomes
2
w 2 d =
t
e
A 2
[ H ] [ H ] d{} =
1
12
1
e
1
2
1
1
2
(2.242)
in which A is the area of the triangular element.

The second term in Equation 2.241 becomes
w d =
a a a
11 12 13
w w
x x + y y d = a211 a22 a23
e
a31 a32 a33

1
2 (2.243)
where
a11 =
2
2
1
x3 x2 ) + ( y 2 y 3 ) (2.244)
(
4A
a12 = a21 =
1
( x3 x2 ) ( x1 x3 ) + ( y2 y3 ) ( y3 y1 ) (2.245)
4A
a13 = a31 =
1
( x3 x2 ) ( x2 x1 ) + ( y2 y3 ) ( y1 y2 ) (2.246)
4A
a22 =
a23 = a32 =
1
( x1 x3 )2 + ( y3 y1 )2 (2.247)
4A
1
( x1 x3 ) ( x2 x1 ) + ( y3 y1 ) ( y1 y2 ) (2.248)
4A
50
a33 =
2
2
1
x2 x1 ) + ( y1 y 2 ) (2.249)
(
4A
The term on the right-hand side becomes

w n d = wv d (2.250)
Here, vn is the normal velocity at the boundary. The outward direction is considered positive. Similarly, isoparametric quadralateral elements can be used for the finite element
formulation.
2.10Interaction of Structure with Acoustic Domain

This section presents how to model coupling between a structure and an acoustic fluid
medium. First, a 1-D problem is discussed for an axial member with an acoustic domain.
Then, a 2-D problem is presented using a beam in a 2-D acoustic fluid medium.
2.10.1One-Dimensional Case
To model the effect of an acoustic fluid medium on the longitudinal vibration of a rod, the
wave equations that follow are considered for the rod and the fluid, respectively, as presented previously. For the rod, the wave equation is expressed as

2 u
2 u
= cr2 2 (2.251)
2
t
x
where u, x, and t are the longitudinal displacement of the rod and the spatial and time variables, respectively. Furthermore, cr is the speed of sound in the rod; it is equal to Er /r ,
where Er and r are the elastic modulus and density of the rod material, respectively. The
wave equation for the acoustic fluid is

2 p
2 p
= c 2f 2 (2.252)
2
t
x
where p is the acoustic pressure, and cf is the speed of sound in the fluid and is equal to
B f / f , where Bf and f are the bulk modulus and density of the fluid, respectively.
The weighted residual finite element formulation for Equation 2.251 was developed previously, and the final matrix expression is
[Mr]{} + [Kr]{u} = {Fr}+{Ff} = {Fr} + Arpint (2.253)
51
in which the superimposed dots denote the temporal derivative, and [Mr] and [Kr] are the
mass and stiffness matrices, respectively, as given Equation 2.59. In addition, {Fr} and {Ff}
are the force vectors applied to the rod by mechanical and fluid forces, respectively. The
fluid force results from the pressure at the interface (i.e., pint), and Ar is the cross-sectional
area of the rod. For simplicity, the cross section is assumed to be unity.
The same formulation is applied to Equation 2.252, leading to
Lf
2 p
2 p
w 2 c 2f 2 dx =
x
t
Lf
2 p
w 2 dx +
t
Lf
Lf
c 2f
p
w p
dx c 2f w = 0 (2.254)
n 0
x x
The first and second terms on the right-hand side of Equation 2.254 yield [Mf] and [Kf],
which are the same as those in Equation 2.59, like the rod, but the third term is for the
boundary condition. For the fluid-solid interface, the boundary condition becomes
p
2 u
= f 2 (2.255)
n
t
The resulting matrix equation becomes

int = 0 (2.256)
[ M f ]{ p} + [ K f ]{ p} c 2f f u
assuming the interface is located between the right end of the rod and the left end of the
fluid domain.
For the eigenvalue analysis of the rod in contact with fluid, the two equations, Equations
2.253 and 2.256, are solved together with an assumption of harmonic solutions for the rod
displacement and the acoustic pressure. In the example provided in the following, the rod
has an elastic modulus of 20 GPa and a density of 2000 kg/m3, resulting in the speed of
sound, 3162 m/s. For the fluid, its speed of sound is varied from 1500 to 2500 and 4000 m/s.
Some of this speed may not be realistic. However, as a demonstration of the effect of the
fluid speed of sound, its value is varied.
Figure 2.16 shows the first mode shape for each different case of the fluid speed of sound.
The rod is fixed at the left end, while it has no constraint on the right end except for its contact with fluid. No fluid means there is no fluid on the right end of the rod. In this case, the
first mode is a half-sine curve, as predicted by the vibration theory of the rod. When the
fluid speed of sound is 1500 m/s, the mode shape is very close to the no-fluid case. When
the fluid speed of sound is increased to 4000 m/s, the bulk modulus of the fluid becomes
higher with a constant fluid density. This means the fluid becomes very incompressible.
Then, the beam behaves like the case when both ends are constrained, resulting in a fullsine curve as shown in the figure. When cf equals 2500 m/s, the mode shape is between
the two cases.
This study also showed that the natural frequency and the mode shape of the wet structure did not change as the fluid density was varied while the ratio of the speeds of sound
of the solid and fluid media remained constant. Thus, the ratio of the speeds of sound is
the main parameter that can affect the frequency as well as the mode shape.
52
Normalized longitudinal displacement
0
0.1
No f luid
Fluid speed = 1500
Fluid speed = 2500
Fluid speed = 4000
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized distance
0.8
0.9
FIGURE 2.16
Mode shape of axial member coupled with 1-D acoustic domain.
2.10.2Two-Dimensional Case
To present a 2-D coupling example, a beam is considered. The dynamic equation for beam
bending is written as previously:
s As
2 u 2 2 u
+
EI
= qbs (2.257)
t 2 x 2 x 2
As discussed previously, the finite element matrix equation becomes

[Ms]{s} + [Ks]{us} = {Fs} + bs = [Q fsi]{pfsi} (2.258)
where the mass and stiffness matrices are given in Equations 2.89 and 2.90.
The Helmholtz wave equation results in the following matrix equation as given
previously:
fsi } (2.259)
[ M f ]{ p f } + [ K f ]{ p f } = f [Q fsi ]T {u
with the following boundary condition:

p
s (2.260)
= f u
n
Putting together the equations for the beam and acoustic problems and neglecting any
external load gives
[M ]
s
f [Q fsi ]T
[0] u
s
[ M f ] p f
[K ]
s
[0]
bs [Q fsi ]
[K f ]
u
s
p f

0
= (2.261)
0
53
0.60
Frequency ratio
0.50
0.40
0.30
0.20
0.10
0.00
1000
2000
3000
4000
Beam density
5000
6000
FIGURE 2.17
Plot of the frequency ratio with and without acoustic domain versus beam density (kg/m3).
where [Q fsi] is only associated with the structural and fluid nodes at the fluid-structure
interaction (FSI) interface. Therefore, the submatrix [Q fsi] contains many zero rows and
columns for the nodal variables, which are not associated with the FSI interface.
As an example, the fluid domain has a density of 1000 kg/m3, and the speed of sound
is 1500 m/s. The beam has an elastic modulus of 20 GPa, and the density is 2000 kg/m3.
Then, the first natural frequency is computed for the beam in the acoustic domain as well
as without considering the acoustic domain. The ratio of the frequency with and without
the acoustic domain interaction is 0.4. In other words, the acoustic domain reduces the frequency by 60%. As a parametric study, the ratio of the frequency is plotted as a function of
the density of the beam as shown in Figure 2.17. The figure shows that the lighter the beam
density is, the greater the effect of the acoustic domain is.
3
Lattice Boltzmann Method
3.1Introduction
The Navier-Stokes equation has been solved using computational fluid dynamics techniques. The control volume technique is one of the most popular. Another common
numerical method is the finite element method. Those formulations dealt with macroscopic variables, such as velocities and pressure, directly. As a result, those techniques are
suitable for macroscale analysis of fluid flow. Another alternative computational technique
is the lattice Boltzmann method (LBM). This technique is a bottom-up approach and is
presented here.
The LBM has been popular for modeling and simulating fluid flow [1,2]. In contrast to
other conventional methods described previously, the LBM does not begin with the macroscale parameters as used in a continuous medium. Instead, the technique uses a collection of fictitious particles. The flow domain consists of regular shapes of lattices. Then, the
particles move along the lattices to have local interactions with other particles in accordance with simple rules. The macroscale conservation laws, such as conservation of mass
and linear momentum as described by the Navier-Stokes equations, are satisfied from the
local collision of particles and their redistribution. Because of the simplicity of the concept
using a collection of particles, the LBM has been applied to various flow problems, including those of porous media [35], two-phase flow [68], and magneto-hydrodynamics [911],
among others.
The LBM technique can also produce other partial differential equations of interest by altering the formulation. Those equations include the Burgers equation [12], the
Korteweg-de Vries equation [13], the Brinkman equation [14], and the Schrodinger equation [15]. A review of the current state of the art in LBM can be found in Reference 1, with
a more recent update in Reference 16. An analysis of LBM theory with a critique and comparison with traditional computational fluid dynamics techniques can be found [17]. This
chapter presents various LBM formulations for the solution of the Navier-Stokes equations
for single-component fluid flows as well as a limited number of multicomponent fluid
flows. In addition, the LBM formulation is discussed for the wave equation.
3.2Standard Lattice Boltzmann Method

The LBM technique can be derived from the concepts of the cellular automaton (CA) [18,19].
The CA model underlying a fluid dynamics model incorporates movement of particles
55
56
from one lattice site to another along discrete lattice directions. The rule for the lattice site
update is applied to all particles arriving at a given lattice site in a given time step and is
represented formally in Equation 3.1:
f ( x + x e , t + t ) = f ( x , t) + ( f ) (3.1)
where f is the distribution function of particle velocity along the direction, t is time, e
is the discrete velocity vector along the direction, and denotes the collision operator.
The discrete velocity vector e is provided for the specified directions of the lattice model
used in the analysis. When the lattice space and the time increment are normalized as
unity, Equation 3.1 is expressed as

f ( x + e , t + 1) = f ( x , t) + ( f ) (3.2)
The expression given in Equation 3.1 can also be recast into the following differential
equation:
f
+ e f = (3.3)
t
in which indicates the gradient vector. Equation 3.1 can be derived by applying the finite
difference concept to Equation 3.3.
Equations 3.1 through 3.3 are applied to a regular lattice of the same size in all directions. The lattice is defined by the problem dimension d and the number of lattice vectors
q and is denoted as DdQq. One of the most common lattice structures is D2Q9, which suggests that the lattice is a two-dimensional (2-D) discretization and there are nine lattice
velocity sets e as shown in Figure 3.1.
(0,1)
(1,1)
e6
(1,1)
e2
e5
e0
(1,0)
e3
e7
(1,1)
FIGURE 3.1
D2Q9 lattice with nine velocity vectors.
e1 (1,0)
e4
(0,1)
e8
(1,1)
57
The discrete velocity vectors are given next for the D2Q9 lattice structure of unit spacing
as shown in Figure 3.1:
e = 0
0
1
0
0
1
1
0
0
1
1
1
1
1
1
1
1
(3.4)
1
This
means that the nine
velocity
vectors for D2Q9 are

e0 = (0, 0), e1 = (1, 0), e2 = (0, 1),
e3 = (1, 0), e4 = (0, 1), e5 = (1, 1), e6 = (1, 1), e7 = (1, 1), e8 = (1, 1). For three-dimensional
(3-D) analysis, the D3Q15 lattice model as illustrated in Figure 3.2 has the discrete velocity
vectors expressed as
0
e = 0
1
0
0
1
0
0
0
1
0
0
1
0
0
0
1
0
0
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1 (3.5)
1
Other common 3-D lattice structures are D3Q19 and D3Q27. Figure 3.3 compares the
three lattice structures. The discrete velocity vectors for D3Q19 and D3Q27 are
0 1 1 0 0 0 0 1 1 1 1 1 1 1 1 0 0 0 0
e = 0 0 0 1 1 0 0 1 1 1 1 0 0 0 0 1 1 1 1 (3.6)
0 0 0 0 0 1 1 0 0 0 0 1 1 1 1 1 1 1 1
(1,1,1)
e7 (1,1,1)
e8
e5
e3
(1,1,1)
z
e10 (1, 1,1)
e9
e0
e2
y
(1,1,1) e12
x
(1,1,1)
e13
FIGURE 3.2
D3Q15 lattice with 15 velocity vectors.
e4
e1
e11 (1,1,1)
e6
e14 (1,1,1)
58
D3Q15 lattice
D3Q19 lattice
(b)
(a)
D3Q27 lattice
(c)
FIGURE 3.3
Comparison of (a) D3Q15, (b) D3Q19, and (c) D3Q27 lattice structures.
for D3Q19 and
0

e = 0
0
1
0
0
0
1
0
0
0
1
0
1
0
0
0
1
1
1
0
1
1
0
0
1
0
1
0
1
1
1
0
1
0
1
1
0
1
0
1
1
1
0
1
0
1
1
0
1
1
1
1
1
1
1
1
0
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
0
0
(3.7)

for D3Q27. In general, more lattice vectors improve the accuracy of the solution at the
expense of the computational cost.
59
The collision operator in Equations 3.1 through 3.3 can be expressed in different ways.
When a single relaxation time technique is used for the collision operator, such as the BGK
technique [20], the collision operator is expressed as
1
= ( f f ) (3.8)
where f denotes the local equilibrium distribution of f. The relaxation parameter can
be related to the fluid kinematic viscosity using the following expression:
=
1
+ (3.9)
c s2 2
where cs is the lattice speed and is equal to 1/ 3 for the unit spacing lattice structures.
The local equilibrium distribution is written as

2

f = 1 + 3 e + 9( e ) 3 (3.10)
c s2
2 c s4
2 c s2
in which is the macroscopic fluid density, and is the fluid velocity. The pressure and
velocity vector can be expressed, respectively, as

(3.11)
and

1
=
f e
(3.12)
In addition, is the weighting parameter for each velocity direction as follows:
4/9
= 1/9
1/36
=0
= 1, 2 , 3, 4 (3.13)
= 5, 6, 7 , 8
2/9
= 1/9
1/72
=0
= 1 to 6 (3.14)
= 7 to 14
for D2Q9;
60
for D3Q15;
1/3
= 1/18
1/36
=0
= 1 to 6 (3.15)
= 7 to 18
for D3Q19; and
= 1 3, 14 16
= 10 13, 23 26 (3.16)
= 4 9, 17 22
=0
8 27
2 27
1 54
1 216
for D3Q27.
The computational procedure for the LBM consists of repeats of the two processes called
streaming and collision (or called redistribution) processes. For the streaming process,
particles at each lattice point move to its neighboring lattice points as specified by DdQq.
The distribution of the particles along those directions is determined by as expressed
in Equations 3.13 through 3.16, depending on the lattice structure used in the analysis.
This streaming process is completed at all the lattice points. Then, Equation 3.8 is used
for the collision process, after which the particles at every lattice point are updated using
Equation 3.1 or 3.2. Then, the two processes repeat themselves.
3.3Multiple Relaxation Lattice Boltzmann Formulation

The single relaxation collision operator is simple and computationally efficient because the
update is based on a single parameter. However, the collision operator has a problem of
stability in some applications. To improve stability, the multiple relaxation collision operator, also referred to as the generalized lattice Boltzmann equation, was presented [21]. The
objectives of the multiple relaxation collision operators were to resolve the fixed Prandtl
number associated with the single-parameter collision operator. The technique allows
for varying kinematic and bulk viscosities and introduces a mechanism for improving
simulation stability. The multiple-relation collision technique utilizes the projection of
the density distribution functions f onto an orthogonal vector space of momenta using
the operator M. The particular momenta depend on the lattice structure chosen, but all
include a combination of the mass density, kinetic energy, momentum flux, energy flux,
and viscous stress tensor. They are expressed in the vector R. Relaxation occurs over the
momentum space using the relaxation times given in S, and the result is transformed back
to the density space f using the inverse of M.
For the D2Q9 lattice, the momentum space and transformation matrix are given in
Equations 3.17 and 3.18, respectively.

RD 2Q 9 =
jx
qx
jy
qy
pxx
qxy (3.17)
61
M D 2Q 9
4
4
0
= 0
0
0
0
0
1
1
2
1
2
0
0
1
0
1
1
2
0
0
1
2
1
0
1
1
2
1
2
0
0
1
0
1
1
2
0
0
1
2
1
0
1
2
1
1
1
1
1
0
1
1
2
1
1
1
1
1
0
1
1
2
1
1
1
1
1
0
1
2
1
1
(3.18)
1
1
1
0
1
where R = Mf, and is fluid density, e is the energy, is related to the square of the
energy, jx and jy are mass fluxes, qx and qy correspond to energy flux, and pxx and pxy correspond to the diagonal and off-diagonal components of the viscous stress tensor. The
coefficients for relaxation over this momentum space are given in a diagonal matrix as
shown in Equation3.19:

S = diag(0, s2, s3, 0, s5, s7, s8, s9) (3.19)
In Reference 22, it was shown that the same fluid viscosity is given in the fluid flow when
s8 = s9 = 1/. The other parameters in Equation 3.19 can be set as desired to promote solution stability or as required to further tailor fluid behavior. If all nonzero coefficients of S
are set to 1/, the single-relaxation technique is recovered. Once the coefficients of S are
provided, the LBM collision operator is given as shown in Equation 3.20:

MRT = M1SM(f feq) (3.20)
where all values of f are relaxed with a single matrix collision operator. While the multiple
relaxation technique requires more computations per time step, it has been found that
simulations are able to be conducted with much lower lattice density with improved stability. Furthermore, fewer time steps are typically required for flows to overcome the noise of
nonequilibrium initial conditions and to reach accurate flow configurations.
3.4Finite ElementBased Lattice Boltzmann Method

One of the requirements of the standard LBM is the use of regular lattice structures of
square or cubic shapes of equal spacing. If there is a curved boundary, such a regular lattice structure requires a very fine mesh to represent the curved boundary accurately. This
is directly related to computational cost. One way to overcome this limitation is to use the
finite elementbased LBM, called FELBM [23]. Because more general shapes of finite elements, such as triangular or quadrilateral shapes for a two-dimensional (2-D) domain, can
be used with FELBM, any complex boundary shape can be modeled using a reasonable
size of lattice structure of a nonregular shape like a finite element mesh.
62
The formulation of the FELBM starts with Equations 3.3 and 3.8 for the single-relaxation
time operator. The derivation shown next is also applicable to the multiple-relaxation time
operator with a minor change. The weighted residual technique is applied to the equation:
f
w t + e
1
f + ( f f ) d = 0 (3.21)
where w is a test or weighting function. To develop the FELBM, the problem domain is discretized into a number of finite elements called a finite element mesh. Then, the unknown
variable f is expressed in terms of the interpolation (or called the shape) functions and
nodal variables for every finite element as follows:
n
f =
H f
i
= [ H ]{ f } (3.22)
i=1
where Hi is the spatial shape function associated with the ith node of the finite element, fi
is the ith nodal variable of the finite element, and n is the number of nodes per element. In
addition, [H] is a row vector consisting of the shape function Hi, and {f} is a column vector
containing unknown solutions at the nodes. Substitution of Equation 3.22 into Equation
3.21 yields the weighted residual equation:

1
{w} [ H ]{ f } + e [H ]{ f } + [ H ] { f } { f } dS = 0 (3.23)
where the integration is conducted over each finite element domain e, and the summation is performed over the total number of elements. Furthermore, {w} is a column vector consisting of weighting functions. Calculation of each element integral from Equation
3.23 and summation of the resulting matrices and vectors results in the following matrix
equation:

[ M]{ F } + [ K ]{ F } + [C]{ F } [C]{ F } = 0 (3.24)
where the matrices and vectors in Equation 3.24 are expressed as follows:

[ M] =
[K ] =
Se
[k ] =
[C] =
[m] =
Se
{w}[ H ] dS (3.25)
{w} ( e [H ]) dS (3.26)
[c ] =
Se
1
{w}[ H ] dS (3.27)
63
{ F } =
{ f } (3.28)
{ F } =
{ f } (3.29)
Depending on the choice of the weighting function, there are many choices in the
weighted residual method. Some of them are the Galerkin method, collocation method,
method of moments, least-square method, or subdomain method. In this section, the
Galerkin method is presented because it is one of the most popular choices. For the
Galerkin method, the weighting function is selected to be the same as the shape functions
used in Equation 3.22, that is, {w} = [H]T. Then, Equations 3.25 through 3.27 are rewritten,
respectively, as

[ M] =
[K ] =
[m] =
Se
[k ] =
[C] =
Se
[ H ]T [ H ] dS (3.30)
[ H ]T ( e [H ]) dS (3.31)
[c ] =
Se
1
[ H ]T [ H ] dS (3.32)
For the collocation method, the weighting functions are selected to be Dirac delta functions. The Dirac delta functions may be defined at the nodal points of each element for convenience. Therefore, for a 2-D case, w(x, y) = (x xi) (y yj), where xi and yj are the nodal
coordinate values. For the 3-D case, w(x, y, z) = (x xi) (y yj) (z zk). On the other hand,
for the method of moment the weighting functions are chosen to be monomial terms, such
as xpyqzr (where p, q, and r are nonnegative integers), starting from the lowest order.
Once the matrix equation of the first-order temporal derivative, as given in Equation
3.24, is developed from the weighted residual finite element formulation, a numerical time
integration scheme is applied to the equation. There are many different numerical techniques for time integration. Those include, but are not limited to, the forward difference
technique, backward difference technique, Crank-Nicolson technique, Runge-Kutta technique, and predictor-corrector technique. The forward difference technique is the simplest
and easiest. If the matrix [M] becomes a diagonal matrix, the forward difference technique
becomes the explicit method. As a result, even though a small time step size t is used
because of the conditional stability, the overall computation is efficient. If an unconditionally stable method is preferred, the Crank-Nicolson technique may be selected.
Using the forward difference scheme for time integration to Equation 3.24 yields the following equation:

{ F }t+ t = { F }t + t[ M]1 [C]{ F }t [C]{ F }t [ K ]{ F }t (3.33)
Equation 3.33 is solved for the given initial and boundary conditions. How to apply the
boundary conditions using the LBM is described in Section 3.8.
64
3.5Element-Free-Based Lattice Boltzmann Method

Consider a lattice point x and its neighborhood as a subdomain x. Inside the subdomain,
there are n numbers of randomly distributed lattice points. To represent the solution u(x)
inside the subdomain, a polynomial function is assumed:

u(x) = {p(x)}T{a(x)} (3.34)
where {p(x)} is a vector containing a complete monomial basis of order m as expressed as

follows:
{p(x)T} = {1 x y} for 2-D with m = 3
(3.35)
{p(x)T} = {1 x y x2 xy y2} for 2-D with m = 6
(3.36)
{p(x)T} = {1 x y z} for 3-D with m = 4
(3.37)
and {a(x)} is a vector consisting of coefficients of the monomial terms. One thing to be mentioned here is that the coefficient vector {a(x)} is not a constant vector but a function of x,
which will be determined subsequently.
The coefficient vector is determined to best fit the solutions at the lattice points inside the
subdomain. To achieve the goal, the weighted least square technique is utilized. The sum
of the weighted square is expressed as
n
J( x) =
w (x)[{p(x )} {a(x)} u ]
k
(3.38)
k =1
where wk(x) is the weighting function associated with the lattice point k, and k is the solution at the same node. Minimization of this equation with respect to {a(x)} results in
[A]{a} = [B]{} (3.39)
in which
n
[ A( x )] =
w (x){p(x )}{p(x )}
k
(3.40)
k =1
[B(x)] = [w1(x){p(x 1)} w2(x){p(x 2)} wn(x){p(x n)}] (3.41)

Solving for the coefficient vector and substituting the resulting expression into Equation
3.34 yields

u(x) = {}T{} (3.42)
where the interpolation function vector {} is given as

{(x)}T = {p(x)}T [A(x)]1[B(x)] (3.43)
65
The interpolation function, Equation 3.43, is applied to the weighted residual formulation as described in the previous section to develop the element-free-based LBM, called
EFLBM[23].
This interpolation function vector is different from that used in the finite element
method. The interpolation function vector for the finite element method satisfies the following relationship:

u(x k) = k (3.44)
However, the interpolation vector developed for the element-free technique does not satisfy the same relationship.
For the matrix [A] to be invertible, n m should be satisfied, and the weighting function
is selected to be a nonnegative function, such as a spline function, defined as follows:
2
3
4
1 6 dk + 8 dk 3 dk 0 d r
k
k
r
r
r
wk ( x) =
k
k
k
(3.45)
0
dk rk
where dk is the distance between the lattice points x and x k, and rk is the size of the support
for the weighting function wk.
The derivative of the interpolation function requires the derivatives of {p(x)}, [A(x)]1, and
[B(x)], respectively. For the derivative of [A(x)]1, the following expression is used:
([A]1)l = [A]1([A])l [A]1 (3.46)
3.6Hybrid Lattice Boltzmann Formulation

To mitigate the computational demands of the FELBM while retaining the ability to model
a domain with complex and irregular shapes without an unnecessarily dense lattice, the
hybrid LBM (HLBM) was developed. The HLBM couples classical LBM (CLBM) described
in Section 3.2 or 3.3 and FELBM [24]. The CLBM requires a regular lattice structure, but
FELBM does not. Therefore, any domain with curved boundaries can be constructed as
a sum of two types of subdomains. One subdomain has a regular lattice structure; the
other subdomain consists of a random lattice structure. It is also acceptable to have multiple regular lattice subdomains and irregular lattice subdomains. The CLBM is applied to
the subdomains made of regular lattices, and FELBM is applied to the other subdomains
made of irregular lattices. The two techniques are properly coupled at the interfaces of
the two different types of subdomains. Either the single-relaxation or multiple-relaxation
technique can be used for CLBM as well as FELBM.
All of the theoretical development from the CLBM and FELBM formulations is preserved, and the logical sequence of computations is maintained on each individual sub
domain. A typical HLBM time step is portrayed schematically in Figure 3.4.
66
Collision
CLBM subdomain:
FELBM subdomain:
classical streaming
advective time step
FIGURE 3.4
Schematic of the hybrid LBM time step. Methodology differs only in implementation of the particle streaming
phase.
To couple two neighboring subdomains, an interface layer is provided. Computationally,

the streaming process of the CLBM domain and the advection process of the FELBM
domain can be executed concurrently, with each subdomain retaining a halo of depth 1
into the adjoining subdomain [24]. Within each subdomain, the outermost layer of lattice
points represents the halo as shown in Figures 3.5 and 3.6.
While this coupling scheme is conceptually simple, great care must be taken with the
time and space integration methods used for advecting particle density data on the FELBM
subdomain. First-order time integration schemes tend to have too much dissipation error,
while second-order schemes suffer from dispersion errors. Both effects propagate onto the
CLBM subdomain and have an impact on solution quality and stability everywhere. For
the results presented here, simple bilinear elements are used for the spatial discretization,
and a four-stage third-order explicit time integration method shown in Equation 3.47 is
used for temporal integration [24].
(a)
(b)
FIGURE 3.5
Schematic hybrid lattice on regular domain. Assignment following streaming in the CLBM domain and advection in the FELBM domain is only made to the interior of each respective subdomain. Data drawn from the lattice points on the halo facilitate communication between each subdomain. (a) Solid circles: CLBM interior; open
circles: CLBM halo. (b) Solid squares: FELBM interior; open squares: FELBM halo.
67
Hybrid test patch

32
30
Lattice Y position
28
26
24
22
20
18
35
40
45
50
Lattice X position
FIGURE 3.6
Interface region for CLBM and FELBM domains on a uniform mesh. Lattice points with both an asterisk and
circle belong to the interface.
F(U ) = FEM advection operator on FELBM sub-domain

n elements
U n fOUT n
U (1) = U n +

1
t F(U n )
2
U ( 2 ) = U (1) +
U ( 3) =
FELBM sub-domain
1
t F(U (1) )
2
(3.47)
2 n 1 (2) 1
U + U + t F(U ( 2 ) )
6
3
3
U n+ 1 = U ( 3 ) +
1
t F(U ( 3) )
2
U n+1 fIN n+1
FELBM interior
This method effectively controls both dissipation and dispersion errors during the advection process, and it allows a single FELBM advection time step for every CLBM streaming
step. The CLBM subdomain undergoes the classical LBM streaming to adjacent neighbors,
68
restricted to only destination lattice points that lay within the CLBM interior as indicated
in Equation 3.48.

fOUT n
CLBM sub-domain
streaming
fIN n+1
CLBM interior
(3.48)
Geometrically simple portions of the domain are discretized with a uniform, regular
lattice for the CLBM. Subdomains containing complex or irregular shapes are identified
and discretized with the FELBM. For example, in simulations such as those requiring
fluid-structure interaction (FSI), in the case of flow past a heat exchanger tube bundle, it
would be sufficient to employ the FELBM only in a region around the actual tubes. This
area could employ a mesh with isoparametric elements to efficiently describe the shape
of the tube and be used with the FELBM, while the remainder of the domain could use a
uniform, regular lattice and the CLBM. The resulting HLBM could accurately capture the
flow properties while reducing the total number of lattice points required significantly. To
make the most of the computational benefits, the size of the FELBM subdomains should be
much smaller than the size of the CLBM subdomains.
3.7Multicomponent Flow
One of the advantages of the LBM lies in its natural amenability to multicomponent fluids. There are four main multicomponent fluid models using the LBM theory: the color
fluid model [25], the interparticle-potential model [26], the free-energy model [27], and
the mean-field theory model [28]. The different methods are classified based on the way
in which the surface tension of the component interface is taken into account in the evolution of the particle distribution functions as well as how the location of this interface is
determined. References 2 and 16 provide good surveys of multicomponent fluid flows. In
this section, the interparticle potential model is discussed in greater detail and the other
methods are briefly reviewed.
3.7.1Color Fluid Model
The color fluid model [25] allows the simulation of immiscible binary fluids in two dimensions. The method is based on the two-component CA model introduced in Reference 29
and is modified for use with LBM. In this technique, each component of the binary fluid
mixture is denoted by a color. That is the reason the method is referred to as the color fluid
model. For example, one component is called red particles and the other blue particles.
The LBM is carried out for each fluid species, and the effect of surface tension on particle
distributions is considered with an additional perturbation term appended to the collision
operator. Furthermore, a recoloring step is applied to make a correction based on the
local color gradient that forces a shift to a direction leading to other like-colored particles.
This method is frequently criticized in the literature for the artificial recoloring process
[30], although each of the multicomponent models has some heuristic process that can
be subjected to the same criticism. Because of that, spurious resultant velocities are commonly exhibited in the vicinity of fluid interfaces. More importantly, the recoloring step
69
is executed based on a costly and time-consuming calculation of local color gradients that
requires considerably more information sharing between neighboring particles compared
to other methods.
3.7.2Free-Energy Model
The free-energy model proposed in Reference 27 takes a different approach. Instead of
maintaining density distributions for each phase, a single-density function is used along
with a density difference . Despite the terminology, which leads one to believe that
the method is intended mainly for single-component multiphase flow, the method was
originally introduced to model phase separation in nonideal one- and two-component fluids. The free-energy model gets its name through the use of the so-called Cahn-Hilliards
approach for nonequilibrium thermodynamics [31]. In this approach, the form of the pressure tensor is defined based on a nonlocal pressure and a parameterized van der Waals
equation of state. This pressure tensor is added to an expanded equilibrium distribution
function that produces the desired interfacial effect. This approach has been used to simulate Rayleigh-Taylor instability [32], bubble motion [33], and simulation of spontaneous
emulsification of liquid droplets in oil-water-surfactant systems [34].
3.7.3Mean Field Theory Model
The mean-field theory was introduced in Reference 28 for nonideal gas flow. In this
method, two distribution functions are used. The first distribution function is used to calculate the pressure and velocity fields of an incompressible liquid. The second is an index
function that is used to locate the interface. The model is so named because the interparticle interactions are treated using a mean-field approximation in the same way that the
Coulomb interaction among charged particles of a plasma is treated in the Vlasov equation [20,35]. As several authors have pointed out, this approach is similar to the traditional
computational fluid dynamics methods for interface capturing and is the LBM analogy to
the level set [36] and volume of fluid methods [37]. This model has been successfully used
to model Rayleigh-Taylor and Kelvin-Helmholtz instabilities [38,39] with nonideal dense
fluids, among other applications.
3.7.4Interparticle Potential Model
The interparticle potential model [26] simulates multiphase and multicomponent fluids as
a simple means. The fundamental idea is that the surface tension effect is microscopic, and
the same effect could be incorporated into the LBM via these same interparticle potential
forces. In this model, only the nearest-neighbor particle densities are considered, and they
are introduced as follows:
q1
w (x + e t, t)e
F ( x , t) = G ( x , t)
(3.49)
=1
where x is particle position, G is a parameter indicating interaction strength, (x, t) is a

function describing the interaction potential, w is the lattice weight for direction , and
e is the corresponding lattice velocity. The form of the potential function can be varied
70
to obtain the desired interparticle potential behavior. For multiphase flow, is commonly
expressed as in Equation 3.50:

() = o exp o (3.50)

where o and o are arbitrary parameters selected to achieve appropriate dynamics for
a selected fluid system. The multicomponent fluid systems used in this work are only
qualitatively correlated with real fluid systems in that only density and viscosity are set
and scaled consistently with the LBM. The parameter G is set to generate a desired level
of immiscibility while maintaining simulation stability. The potential function is set as
(x,t) = (x, t). Notice that Equation 3.49 specifies a weighted summation of the value of
for each lattice position in the neighborhood of a given lattice particle.
3.7.5Immiscible Multicomponent Lattice Boltzmann Procedures
The basic LBM process with multiple components is similar in most ways to that used
for single-component systems. The obvious difference is that there are now two sets of
distribution functions: as before f and for a second component that will conventionally be
referred to as g. A second difference is that, as discussed in the preceding paragraphs, the
interparticle force must be calculated in accordance with Equation 3.49 and incorporated
into the calculation of macroscopic velocity used for computing feq and geq. The biggest and
most fundamental difference is the need to structure the computations to account for the
fact that, due to the desired macroscopic dynamic evolution of the system, some lattice
sites will have zero density for one or the other component and only the interfaces will
have a significant mixture.
To be as explicit as possible, the immiscible multicomponent LBM time step for fluid lattice points carried out for this work is implemented as follows:

1. Compute the macroscopic density of each fluid:

q1
f =
(3.51)
=1
q1
g =
(3.52)
=1
2. Compute the macroscopic momentum of each fluid:

q1
f uf =
f (3.53)
=1
q1
g ug =
e g
=1
(3.54)
71
3. Compute a weighted combined macroscopic density and velocity:

= ff + gg (3.55)
u=
f u f f + g ug g
(3.56)
1
where we recall = .
4. Compute the interparticle force using Equation 3.49, setting G to a constant and
(x, t) = :
q1
F f = G g
w (x + e t, t) (3.57)
=1
q1
Fg = G f
w (x + e t, t) (3.58)
=1
5. Apply these interparticle forces as momentum inputs to each respective lattice

population:

ueqf = u f f Ff (3.59)
ueq
g = u g g Fg (3.60)
6. Complete the usual Lattice Bhatnagar-Gross-Krook (LBGK) collision and streameq

ing steps using u f , ueq
g , and the corresponding macroscopic density for computaeq
eq
tion of f and g accordingly.
3.8Boundary Condition
Boundary conditions must be applied to the boundary of the fluid domain. Because the
LBM has different variables from conventional fluid mechanics, applying the boundary
condition to the LBM model is not straightforward like the conventional computational
fluid dynamics application. The LBM has the particle velocity distributions as the nodal
variable, while conventional fluid mechanics has fluid velocity and pressure as the nodal
variables. As a result, fluid velocity and pressure boundary conditions must be translated
into the particle density distribution conditions. Some of common boundary conditions
are discussed in this section.
72
Before streaming
After streaming with

periodic boundary
FIGURE 3.7
Streaming from north to south as a periodic boundary.
3.8.1Periodic Boundary
Periodic boundary conditions are common and easy to implement into the LBM. For
nodes along a periodic boundary, the nodes along the counterpart periodic boundary are
assigned as the nearest neighbor for streaming purposes. For example, the density distribution along the north direction at the north boundary node is streamed into the north
direction at the south boundary node, which is supposed to be overlapped with the north
boundary node with the periodic boundary condition, as shown in Figure 3.7.
3.8.2Fixed Rigid Boundary
Fixed rigid boundaries appear in a wide variety of applications. A flat rigid boundary is
modeled using the so-called bounce-back technique in which all unknown values of f
are replaced with the values that are known, but from the opposite direction. In addition,
directions parallel to the solid boundary are also reversed, resulting in the exchange of
density distribution values for all opposing directions. This is illustrated in Figure 3.8.
Before bounceback
Boundary
FIGURE 3.8
Application of on-grid bounce-back boundary condition.
After bounceback
Boundary
73
Fluid lattice point
Solid
boundary
Solid lattice point
Bounced-back
directions
FIGURE 3.9
Halfway bounce-back solid boundary condition schematic.
Solid boundaries implemented in this fashion are often referred to as dry nodes because
they do not undergo the collision process. This simplifies implementation and computational efficiency considerably because macroscopic values need not be computed at these
nodes and the equilibrium distribution does not need to be evaluated. This so-called ongrid version of the bounce-back boundary conditions has been shown to be first-order
accurate.
An alternate scheme was introduced where the lattice points are arranged so that the
physical wall is actually located exactly halfway between the first fluid point inside the
domain and the corresponding solid node representing the wall. This scheme is illustrated
in Figure 3.9 and has been shown to exhibit second-order convergence.
3.8.3Pressure Boundary Condition
The pressure boundary condition at the inlet is described first for the D2Q9 lattice. The
inlet is assumed to be located at the left side (i.e., west boundary) of the domain, which is
normal to the x axis direction. First, the inlet pressure pin is converted into the density as
in. Then, the distribution density function is written as

f1 + f5 + f8 = in (f0 + f2 + f3 + f4 + f6 + f 7) (3.61)
f1 + f5 + f8 = in (uin)x + (f3 +f6 + f 7) (3.62)
f5 f8 = f2 + f4 f6 + f 7 (3.63)
Because f1, f2, and f8 contribute to the flow domain, they are obtained from these equations along with the inlet velocity (uin)x.
74
(uin)x = 1 ( f0 + f2 + f4 + 2f3 + 2f6 + 2f 7)/in (3.64)

f1 = f 3 +
2
in (uin )x (3.65)
3
f 5 = f7
1
1
( f2 f 4 ) + in (uin )x (3.66)
2
6
f 8 = f6 +
1
1
( f2 f 4 ) + in (uin )x (3.67)
2
6
The pressure boundary condition at the outlet is described as follows: The outlet is
located at the right side (i.e., east boundary) of the domain, which is normal to the x axis
direction. First, the outlet pressure pout is converted into the density as out. Then, the distribution density function is written as

f3 + f6 + f 7 = out (f0 + f1 + f2 + f4 + f5 + f8) (3.68)
f3 + f6 + f 7 = out (uout)x + (f1 + f5 + f8) (3.69)
f6 f 7 = f2 + f4 f5 + f8 (3.70)
Because f3, f6, and f 7 contribute to the flow domain, they are obtained from these equations along with the outlet velocity (uout)x.
(uout)x = [(f0 + f2 + f4 + 2f1 + 2f5 + 2f8)/out] 1
f 3 = f1
(3.71)
2
out (uout )x (3.72)
3
f6 = f 8
1
1
( f2 f 4 ) out (uout )x (3.73)
2
6
f7 = f 5 +
1
1
( f2 f 4 ) out (uout )x (3.74)
2
6
3.8.4Velocity Boundary Condition

The velocity inlet for the west side of the domain of the D2Q9 lattice, which is normal to
the x axis, has the following density distribution function:

f1 + f5 + f8 = ( f0 + f2 + f3 + f4 + f6 + f 7) (3.75)
75
f1 + f5 + f8 = (uin)x + ( f3 + f6 + f 7) (3.76)
f5 f8 = (uin)y f2 + f4 f6 + f 7 (3.77)
From these equations,

1
( f0 + f2 + f 4 + 2 f3 + 2 f6 + 2 f7 ) (3.78)
1 (uin )x
f1 = f 3 +
2
(uin )x (3.79)
3
f 5 = f7
1
1
1
( f2 f 4 ) + (uin )x + (uin )y (3.80)
2
6
2
f 8 = f6 +
1
1
1
( f2 f 4 ) + (uin )x (uin )y (3.81)
2
6
2
The velocity outlet for the east side of the domain, which is normal to the x axis, has the
following density distribution function:

f3 + f6 + f 7 = (f0 + f1 + f2 + f4 + f5 + f8) (3.82)
f3 + f6 + f 7 = (uout)x + (f1 + f5 + f8) (3.83)
f6 f 7 = (uout)y f2 + f4 f5 + f8 (3.84)
1
( f0 + f2 + f 4 + 2 f1 + 2 f5 + 2 f8 ) (3.85)
1 + (uin )x
f 3 = f1
2
(uout )x (3.86)
3
f6 = f 8
1
1
1
( f2 f 4 ) (uout )x + (uout )y (3.87)
2
6
2
f7 = f 5 +
1
1
1
( f2 f 4 ) (uout )x (uout )y (3.88)
2
6
2
76
For the D3Q15 model, the velocity boundary conditions may be described as follows: If
the flow goes into the negative x plane, the density distribution functions satisfy the following equations:

f1 + f 7 + f10 + f11 + f14 = (f0 + f2 + f3 + f4 + f5 + f6 + f8 + f9 + f12 + f13) (3.89)
f1 + f 7 + f10 + f11 + f14 = ux + f2 + f8 + f9 + f12 + f13 (3.90)
f 7 f10 + f11 f14 = uy f3 + f4 f8 + f9 f12 + f13 (3.91)
f 7 + f10 f11 f14 = uz f5 + f6 f8 f9 + f12 + f13 (3.92)
From these equations, we obtain the following equations:

1
( f0 + 2 f2 + f3 + f 4 + f5 + f6 + 2 f8 + 2 f9 + 2 f12 + 2 f13 ) (3.93)
1 ux
f1 = f 2 +
2
ux (3.94)
3
f7 = f13
1
1
1
1
( f 3 f 4 + f 5 f6 ) +
ux + uy + uz (3.95)
4
12
4
4
f10 = f12
1
1
1
1
( f 3 + f 4 + f 5 f6 ) +
ux uy + uz (3.96)
4
12
4
4
f11 = f9
1
1
1
1
( f 3 f 4 f 5 + f6 ) +
ux + uy uz (3.97)
4
12
4
4
f14 = f8
1
1
1
1
( f 3 + f 4 f 5 + f6 ) +
ux uy uz (3.98)
4
12
4
4
If there is a flow out of the positive x plane, the density distribution functions satisfy the
following expressions:

f2 + f8 + f9 + f12 + f13 = (f0 + f1 + f3 + f4 + f5 + f6 + f 7 + f10 + f11 + f14) (3.99)
f2 + f8 + f9 + f12 + f13 = ux + f1 + f 7 + f10 + f11 + f14 (3.100)
f8 f9 + f12 f13 = uy f3 + f4 f 7 + f10 f11 + f14 (3.101)
f8 + f9 f12 f13 = uz f5 + f6 f 7 f10 + f11 + f14 (3.102)
77
From the equations, we obtain the following expressions for the density and distributions:

1
( f0 + 2 f1 + f3 + f 4 + f5 + f6 + 2 f7 + 2 f10 + 2 f11 + 2 f14 ) (3.103)
1 + ux
f 2 = f1
2
ux (3.104)
3
f13 = f7 +
1
1
1
1
( f 3 f 4 + f 5 f6 )
ux uy uz (3.105)
4
12
4
4
f10 = f12 +
1
1
1
1
( f 3 + f 4 + f 5 f6 )
ux + uy uz (3.106)
4
12
4
4
f11 = f9 +
1
1
1
1
( f 3 f 4 f 5 + f6 )
ux uy + uz (3.107)
4
12
4
4
f14 = f8 +
1
1
1
1
( f 3 + f 4 f 5 + f6 )
ux + uy + uz (3.108)
4
12
4
4
When there is a flow into the negative y plane, the density distributions have the following relationships:

f3 + f 7 + f8 + f11 + f12 = (f0 + f1 + f2 + f4 + f5 + f6 + f9 + f10 + f13 + f14) (3.109)
f 7 f8 + f11 f12 = ux f1 + f2 + f9 f10 + f13 f14 (3.110)
f3 + f 7 + f8 + f11 + f12 = uy + f4 + f9 + f10 + f13 + f14 (3.111)
f 7 + f8 f11 f12 = uz f5 + f6 f9 f10 + f13 + f14 (3.112)
1
( f0 + f1 + f2 + 2 f 4 + f5 + f6 + 2 f9 + 2 f10 + 2 f13 + 2 f14 ) (3.113)
1 + uy
f3 = f 4 +
f7 = f13
2
uy (3.114)
3
1
1
1
1
( f 3 f 4 + f 5 f6 ) +
ux + uy + uz (3.115)
4
12
4
4
78
f10 = f12
1
1
1
1
( f 3 + f 4 + f 5 f6 ) +
ux uy + uz (3.116)
4
12
4
4
f11 = f9
1
1
1
1
( f 3 f 4 f 5 + f6 ) +
ux + uy uz (3.117)
4
12
4
4
f14 = f8
1
1
1
1
( f 3 + f 4 f 5 + f6 ) +
ux uy uz (3.118)
4
12
4
4
A similar expression can also be derived for the positive y plane as well as z planes.
Those expressions derived previously require intensive bookkeeping in terms of computer
programming because the identity of each boundary plane must be provided for each
boundary lattice point. For instance, it is necessary to know the normal vector to each
boundary node as well as the direction of flow, such as inflow or outflow. To avoid such
cumbersome data, another approach was proposed to describe the velocity boundary conditions. This approach does not require bookkeeping because it can be applied to any
lattice point regardless of its location and the flow direction. In this approach, the present
velocities are computed from the present density distributions at the lattice point where
velocities are prescribed as follows for D3Q15:

ux = ( f1 + f7 + f10 + f11 + f14 f2 f8 f9 + f12 f13 )/ (3.119)
uy = ( f3 + f7 + f8 + f11 + f12 f 4 f9 f10 + f13 f14 )/ (3.120)
uz = ( f5 + f7 + f8 + f9 + f10 f6 f11 f12 + f13 f14 )/ (3.121)

Then, the present velocities are subtracted from the described velocities as follows:
ux = uxp ux (3.122)
uy = uyp uy (3.123)
uz = uzp uz (3.124)
where the superscript p denotes the prescribed value as a boundary condition. The final
density distributions are given as

f1 = f1 + ux/3 (3.125)
f2 = f2 ux/3 (3.126)
f3 = f3 + uy/3 (3.127)
79
f4 = f4 uy/3 (3.128)
f5 = f5 + uz/3 (3.129)
f6 = f6 zz/3 (3.130)
f 7 = f 7 + ux/24 +uy/24 + uz/24 (3.131)
f8 = f8 ux/24 + uy/24 + uz/24 (3.132)
f9 = f9 ux/24 uy/24 + uz/24 (3.133)
f10 = f10 + ux/24 uy/24 + uz/24 (3.134)

f11 = f11 + ux/24 + uy/24 uz/24 (3.135)

f12 = f12 ux/24 + uy/24 uz/24 (3.136)
f13 = f13 ux/24 uy/24 uz/24 (3.137)
f14 = f14 + ux/24 uy/24 uz/24 (3.138)
As stated, these expressions can be used for any boundary lattice point with a prescribed
velocity condition regardless of whether the point is located on an inlet or outlet or x plane,
y plane, or z plane.
3.9Turbulent Flow
A turbulent model is incorporated into the LBM by modifying the viscosity. The total
viscosity of a fluid is the sum of the molecular viscosity and the eddy viscosity resulting
from turbulence [40]:

t = o + e (3.139)
where t is the total viscosity, and o and e are the molecular and eddy viscosities, respectively. The viscosity is related to the relaxation constant as follows:

1
= (3.140)
2
As a result, the relaxation constant can be written as

t = o + e (3.141)
80
The eddy viscosity is computed using the Prandtl mixing length approach [37] as follows:
e = (lmix)2 || (3.142)
where is the von Karman constant and is equal to 0.41, and || is the magnitude of the
strain rate tensor, which can be computed from
1
2p
e e ( f
ji ji
ji
f ji ) (3.143)
ji
in which p is the pressure.

The mixing length is determined based on the two-layer model. If the flow is close to a
boundary wall, the distance from the wall is assumed equal to the mixing length. On the
other hand, if the flow is far from the boundary wall, the mixing length is considered a
constant value.
3.10Wave Equation
The previous sections presented various LBM formulations to solve fluid flows. This section discusses the LBM formulation to analyze the wave equation. The LBM equation
based on the single parameter is used as before and expressed as

1 eq
fi (r + ei t , t + t) fi (r ,t) =
fi (r ,t) fi (r ,t) (3.144)
where fi (r , t) denotes the probability of finding a particle at lattice site r and

timet that
moves along the ith lattice direction with the local discrete particle velocity ei . fieq (r , t) is
the local equilibrium solution. Furthermore, t is the time increment, and is the relaxation time.
For the wave equation using the 2-D square lattice, the D2Q5 lattice structure is used
[41]. The five lattice points include the points at the center, east, west, north, and south. The
local equilibrium of particle distribution is expressed as
fieq =
2 c s2
if i = 0
e2
(3.145)

c s2
1 ei
+
if i 0
2 e2
2e 2
where

f (3.146)
i
81
and
f e (3.147)
i i
and cs is the wave propagation speed.

The local velocity vectors are given as follows:
e = 0
0
1
0
0
1
1
0
0
(3.148)
1
With the choice of = 1/2, Equation 3.144 is rewritten as
fi (r + tei , t + t) =
n2 1
fo (r , t) if i = 0
2
n

fi+ 2 (r , t) if i 0
2
2n
(3.149)
and the wave equation is recovered. For these equations, 0 1 is the wave attenuation
factor. If = 0, it denotes perfect reflection of the wave. On the other hand, = 1 indicates
perfect transmission. Any in-between value for represents that the wave is partially
absorbed. Furthermore, n 1 is the refraction index, which is defined as the ratio of the
maximum propagation speed of the model to cs [41].
3.11Scaling
Most of the LBM literature is cast in lattice units, with the lattice spacing and the time
increment unity. To solve a problem in a physical domain, it is necessary to convert the
physical units into the lattice units. As an intermediate step, it is sometimes customary to
rescale physical units to nondimensional units. This is particularly useful if some knowledge of the system state in terms of some nondimensional parameters such as the Reynolds
number is needed. Figure 3.10 illustrates the process to change from the physical units to
the lattice units by way of the nondimensional units.
TP, LP
TD
TP
LP
LD
T0,P
L0,P
Physical
units
T0, L0
TLBM
Dimensionless
units
FIGURE 3.10
Scaling from physical units to dimensionless units to LBM units.
T0
Nt
LLBM
L0
Nx
TLBM, LLBM
Lattice
units
82
To illustrate the unit conversion, an example is presented as shown in Figure 3.11. The
problem involves flow within a 2-D channel around a circular object. Flow enters from
the left boundary with a prescribed parabolic velocity and exits the right boundary with
a prescribed constant pressure. The top and bottom boundary are modeled as no-slip
walls.
The process of scaling for this example problem is completed in two steps as described
previously. First, a characteristic time scale T0 and length scale L0 are identified. For this
problem of a circular obstruction in 2-D channel flow, the natural choice is to use the conventions for Reynolds scaling where the characteristic length is the diameter of the circle.
Therefore, the characteristic length in physical units L0,p = 0.2 m. The characteristic time is
assigned to be the time required for an average fluid particle to traverse the diameter of
the cylinder. For the assigned inlet boundary condition, the average fluid velocity is twothirds of the maximum inlet velocity of 0.5 m/s for the parabolic profile. Consequently,
T0,p = L0,p/U0,p = 0.2/0.5 = 0.4 s. All of this corresponds to a Reynolds number of 100, which
is convenient to know when comparing the output of the LBM simulation against experimental data or benchmark values.
The second step is to decide how finely the reference time and space scales are to be
subdivided. For this example, the reference length L 0 will be represented with 25 lattice
points. In terms of dimensionless units, L 0 = 1. The conversion between dimensionless
units and the LBM units is LLBM = x = 1/(25 1) = 0.0417. To convert between a distance
in terms of lattice units and a distance in physical units, one would multiply by both
the conversion factors. Therefore, the physical spacing of the lattice points is x L0,p =
0.0083m. Similarly, the time domain is discretized by deciding how many time steps will
be used to traverse a single unit of the reference time T0. For this problem, the reference
time will be divided by 250 time steps, so t = T0/Nt = 1/250 = 0.004 s. As with the spatial
scaling, to convert a single lattice time step to physical elapsed time, one must multiply
by the scaling parameters between the dimensionless units and lattice units in addition
to the conversion between physical and dimensionless units. For this problem, those conversions are TP = TLBM t T0 = 0.0016. Once the spatial and temporal scaling factors are
determined, the properly scaled LBM parameters must be determined from the given
physical data.
X, Y = (0,1)
1
u (0, y)= 1
2
y 0.5
0.5
No-slip boundaries top/bottom

L=1m
P = 0 Pa
D = 0.2 m
X, Y = (0,0)
v = le
m2
FIGURE 3.11
Schematic diagram of the channel flow example problem.
= 1000
kg
m3
X, Y = (4,0)
83
3.11.1Viscosity Scaling
To obtain dynamic LBM behavior that corresponds to the desired physical fluid under the
prescribed conditions, the temporal and spatial scaling factors need to be applied to
the specified fluid kinematic viscosity . Because has units of square meters per second,
thenecessary conversion can be accomplished using Equation 3.150.
LBM = physical
T0 t
(3.150)
(L0 x )2
Carrying out this conversion for the specified fluid with the chosen discretization results
in LBM = 0.023. This is the value that is applied to Equation 3.9 to determine the LBM relaxation parameter, = 0.57.
3.11.2Velocity Boundary Condition Scaling
The problem shown in Figure 3.11 has a prescribed inlet velocity profile. The velocity is
expressed in terms of meters per second, which is not compatible with the unit system
assumed when the LBM boundary conditions were developed. The velocity is simply
scaled in accordance with Equation 3.151.
uLBM = uphysical
T0 t
(3.151)
L0 x
For this problem, this conversion reads:

u(0, y ) =
( y 0.5)
( y 0.5) 0.0016
3
= 0.144 1
1
(3.152)
0.5
4
0.5 0.0083
This is the velocity that will be passed to the LBM time-stepping routine to set the prescribed velocity at the inlet lattice points.
3.11.3Pressure Boundary Condition Scaling
For the problem shown in Figure 3.11, the prescribed outlet pressure is set to 0 Pa. This is a
relative pressure, of course. Otherwise, the density for lattice points at the outlet would be
set to zero. The pressure in the physical units is expressed as
2
L
c x 0 (3.153)
tT0
2
LBM s
Pphysical =
The solution procedure is provided in Figure 3.12; the solutions after 50,000 time steps are
illustrated in Figure 3.13 [42].
84
Start
Increment
time step
Apply bounce-back
boundary condition
Solid
Type of lattice point
BC
Apply microscopic
and macroscopic
boundary conditions
Fluid
Compute equilibrium
density distribution
No
Collide
Stream
Last time step?
Yes
End
FIGURE 3.12
LBM time step flowchart.
85
FIGURE 3.13
Velocity magnitude (top), pressure (middle), and vorticity magnitude (bottom) for example flow case after
50,000 time steps.
3.12Example Problems
3.12.1Poiseuille Flow
The first test case is the Poiseuille flow which is a 2-D flow between parallel plates. The flow
condition is depicted in Figure 3.14. In the figure, the channel width 2b is 1 m wide, fluid
density is 1000 kg/m3, and fluid viscosity is 1 N s/m2. The maximum inlet velocity Umax
is 0.015 m/s. The solution to this problem is known to be a function of y only and is given as
u( y ) =

where
1 dp 2
(b y 2 ) (3.154)
2 dx
dp
is given as a function of maximum velocity and fluid viscosity as follows:
dx
dp
2
= 2 U max (3.155)
dx
b
No slip wall
b
X
u(0, y, t) = Umax 1
y b
b
FIGURE 3.14
Poiseuille flow configuration.
No slip wall
p(L, y, t) = pout = constant
86
For the LBM model of this problem, the D2Q9 lattice with the single-parameter collision
operator is used along with the on-grid boundary conditions for the prescribed velocity
on the west boundary and constant prescribed pressure on the east boundary. The initial
lattice discretization is set so that 30 lattice points can span the channel entrance. The
time step is set to achieve a relaxation parameter of 1.30. While refining the grid to test
for convergence, the time step is adjusted to maintain a constant relaxation parameter for
all tests. The results are shown in Figure 3.15 [42]. As expected, first-order convergence is
obtained for this boundary condition.
The halfway bounce-back boundary condition is also implemented and used in an identical set of tests. The goal of this step, in addition to showing the convergence properties
of the boundary condition, is to illustrate the second-order convergence properties of the
LBM. Results for double precision are shown in Figure 3.16 [42].
In this study, it may seem arbitrary to have selected a constant relaxation parameter
= 1/ = 1.25. This conclusion is partially correct insofar as there is considerable flexibility
On-grid double precision
Relative error
1 slope
2 slope
Relative error
0.1
0.01
0.001
0.0001
0.00001
1.48
1.78
2.08
2.38
Log grid resolution
2.68
FIGURE 3.15
Poiseuille flow convergence with on-grid bounce-back boundary conditions.
Halfway double precision
Relative error
1 slope
2 slope
Relative error
0.1
0.01
0.001
0.0001
0.00001
0.000001
1.48
1.78
2.08
2.38
Grid resolution
FIGURE 3.16
Poiseuille flow convergence with halfway bounce-back boundary conditions.
2.68
87
Horizontal velocity fluctuation x/Lx = 0.75 vs. time step, Re = 10

0.06
0.05
Umax (m/sec)
0.04
0.03
0.02
0.01
0
0.01
(a)
4
5
6
Time step
10
104
Horizontal velocity fluctuation x/Lx = 0.75 vs. time step, Re = 10

0.06
0.05
Umax (m/sec)
0.04
0.03
0.02
0.01
0
0.01
(b)
4
5
Time step
8
105
FIGURE 3.17
1
Stabilization time for Poiseuille flow, Re = 10, = = 1.3: (a) with Ny = 30 and (b) with Ny = 480.
regarding how this value is picked. Recall that the fluid viscosity in lattice units is scaled
t
by 2 in accordance with Equation 3.150. Consequently, if x is reduced by a factor of 2, t
x
must be reduced by a factor of 4. With this refined time step, the number of time steps is
increased by a factor of 4 for the fluid simulation, including the time required for the LBM
simulation to arrive at the equilibrium from nonequilibrium initial conditions. For a given
value of , this initial instability can last for many time steps. Even for this simple problem
geometry, the LBM system does not reach a stable answer for many time steps. An illustration of this is given in Figure 3.17. This figure shows variation in the horizontal velocity at
the center of the channel geometry for lattice density of Ny = 30 and 480, respectively. Note
the change in time scales for the time step axis.
3.12.2Backward-Facing Step
The occurrence of flow separation of internal flows by sudden geometric changes is well
known and is important to engineering applications. While the Poiseuille flow test case is
88
convenient for code validation insofar as analytic solutions are known, it does not fully test
the ability of the LBM to reproduce correct fluid behavior. It will be shown that for modest Reynolds numbers, the LBM captures flow separation typified by the backward-facing
step problem accurately.
For this benchmark study, the experimental results presented in Reference 43 are used.
The problem setup is as depicted in Figure 3.18. For these computations, the inlet boundary
conditions are prescribed velocity and for the outlet is prescribed pressure. The multiple
relaxation scheme for D2Q9 is used for bulk dynamics. Measurements were taken based
on streamlines computed from the velocity data. An example for a Reynolds number of
100 is provided in Figure 3.19 [42].
To conveniently compare with measured benchmark values, representative data points
are pulled from Reference 43 and plotted separately along with the values taken from the
computed data. For each successively increased Reynolds number, the lattice density and
time step were both adjusted to maintain a constant relaxation factor = 1.25. Results are
given in Figure 3.20 [42]. Good agreement can be seen in this case with experimental results.
3.12.3Lid-Driven Cavity
Another validation example of the LBM implementation is the lid-driven flow in two
dimensions. A commonly used benchmark for this flow condition is given in Reference 44.
A schematic illustration of the 2-D lid-driven cavity problem is given in Figure 3.21.
No-slip wall
b
b
p(L, y, t) = pout = constant
X
No-slip wall
FIGURE 3.18
Schematic of domain and boundary conditions for backward-step benchmark in two dimensions.
Backward step, Re = 100
0.731
FIGURE 3.19
Backward-step simulation. Step height = 0.25 m, outlet width = 0.5 m, Re = 100.
89
Primary vortex length vs. Reynolds number

10
9
Vortex length (1/s)
8
7
6
1 Armaly et al. 1983
1 LBM
4
3
2
1
0
100
400
200
300
Reynolds number
500
FIGURE 3.20
Comparison of primary vortex reattachment length normalized by step height with results reported in
Reference 45.
u(x,t) = constant = ulid

Moving boundary
Y
No-slip boundary
X
FIGURE 3.21
Schematic of the 2-D lid-driven cavity problem.
The on-grid bounce-back technique is applied to model no-slip stationary walls. The
moving-wall boundary condition is used to model the lid. The lid velocity is set to a constant value in the x direction with the desired speed. The standard benchmark stipulates a Reynolds number of 1000, although other authors have published results at higher
Reynolds numbers. In two dimensions, the D2Q9 lattice is used with either the single- or
multiple-relaxation LBM techniques.
90
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.2
0.4
0.6
0.8
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.2
0.4
0.6
0.8
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.2
0.4
0.6
0.8
FIGURE 3.22
Lid-driven cavity in two dimensions with 1600 1600 lattice showing, from left to right, streamlines, vorticity
contours, and pressure contours for Re = 1000.
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.2
0.4
0.6
0.8
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.2
0.4
0.6
0.8
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.2
0.4
0.6
0.8
FIGURE 3.23
Lid-driven cavity in two dimensions with 1600 1600 lattice showing, from left to right, streamlines, vorticity
contours, and pressure contours for Re = 5000.
X-velocity comparison to benchmark
LBM simulation
Ghia et al.
0.9
0.8
Y-position
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
1
0.8
0.6
0.4
0.2
X-velocity
FIGURE 3.24
Comparison of X velocity for lid-driven cavity flow for Re = 1000.
0.2
0.4
91
Figures 3.22 and 3.23 show the streamlines, pressure contours, and vorticity contours
for Reynolds numbers of 1000 and 5000, respectively [42]. The results compare well with
the results in References 45 and 46. In Figures 3.24 through 3.29, two samples of velocity
are taken in the computed problem domain. The first is the x velocity component sampled
along the vertical centerline. As can be seen, good agreement with benchmark values is
obtained for all macroscopic fluid parameters.
Y-velocity comparison to benchmark
0.4
0.3
0.2
LBM simulation
Ghia et al.
Velocity
0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.1
0.2
0.3
0.4 0.5 0.6

X-position
0.7
0.8
0.9
0.9
FIGURE 3.25
Comparison of Y velocity for lid-driven cavity flow for Re = 1000.
Pressure profile horizontal slice
0.1
LBM simulation
Ghia et al.
0.08
Pressure (Pa)
0.06
0.04
0.02
0
0.02
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
X-position
FIGURE 3.26
Comparison of pressure along the central horizontal line for lid-driven cavity flow for Re = 1000.
92
Pressure profile vertical slice
LBM simulation
Ghia et al.
0.9
0.8
Pressure (Pa)
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.02
0.02
0.04
0.06
Y-position
0.08
0.12
0.1
FIGURE 3.27
Comparison of pressure along the central vertical line for lid-driven cavity flow for Re = 1000.
Vorticity comparison horizontal slice
LBM simulation
Ghia et al.
Vorticity
0
2
4
6
8
10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
X-position
FIGURE 3.28
Comparison of vorticity along the central horizontal line for lid-driven cavity flow for Re = 1000.
3.12.4Channel Flow over Cylinder

The flow configuration for the example of channel flow over a cylinder is illustrated in
Figure 3.30. Constant velocity is specified on the inlet, and constant pressure is specified
on the outlet. The top and bottom of the domain are simulated as periodic boundaries with
the result that the effective domain is a linear array of cylinders in uniform flow. There
is similar work for the benchmarks [4753]. The trailing vortex region for this benchmark
was measured visually following flow simulation. Numeric results are given in Table 3.1
93
Vorticity comparison vertical slice
1
0.9
LBM simulation
Ghia et al.
0.8
Y-position
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
5
5
10
Vorticity
15
20
FIGURE 3.29
Comparison of vorticity along the central vertical line for lid-driven cavity flow for Re = 1000.
X, Y = (0,1)
Periodic top/bottom
L=1m
u(0, y) = U0
P = 0 Pa
D = 0.1 meter
X, Y = (0,0)
m2
v = le 3 sec
= 1000
Kg
m3
FIGURE 3.30
Channel with cylindrical obstacle for the 2-D problem.
TABLE 3.1
Comparison of Trailing Vortex Length to
Benchmark Values
Reference 49
Reference 51
Reference 53
Reference 55
Present solution
Re = 20
Re = 40
0.92
0.91
0.94
1.04
0.95
2.20
2.18
2.35
2.55
2.05
X, Y = (4,0)
94
and graphically in Figure 3.31 [42]. It can be seen from the table that the computed values
using LBM are comparable to those reported in the literature.
Above a Reynolds number of approximately 45, the trailing vortex detaches in an alternating pattern referred to as a Von Karman street. As a last measure, the rate of vortex
shedding was measured and the nondimensional Strouhal number was evaluated, with
the result compared to the literature. A visualization of the vorticity in the wake of a circular cylinder in channel flow at a Reynolds number of 100 is given in Figure 3.32. Notice the
alternating regions of positive and negative vorticity resulting from the vortex shedding
alternately from the top and bottom of the cylinder. The equation for the Strouhal number
is given in Equation 3.156:
St =
f L0
(3.156)
U0
where St is the nondimensional Strouhal number, f is the frequency of vortex shedding, L0 is

the characteristic length, and U0 is the characteristic velocity. For this problem, the characteristic length is the diameter of the cylinder, and the characteristic velocity is the average flow
velocity. During flow simulation, the drag and lift forces were computed; results are presented
in Figure 3.33. The Strouhal number for this simulation is determined by taking the discrete
Fourier transform of the computed coefficient of lift data and applying this along with U0
and L0 in Equation 3.156. The resulting spectrum is presented in Figure 3.34 [42]. The result
was compared with others reported in the literature and showed good agreement (Table 3.2).
0.950
(a)
2.05
(b)
FIGURE 3.31
Streamline visualization of trailing vortex at (a) Re = 20 and (b) Re = 40.
95
Vorticity Z
0
100
151.979
100
146.2506
FIGURE 3.32
Vorticity plot for cylinder in 2-D flow at Re = 100.
Drag/lift coef f icient versus time, Re = 100
CL
CD
Drag/lift coef f icient
1.5
1
0.5
0
0.5
0
3
4
5
6
Simulation time (sec)
FIGURE 3.33
Drag and lift coefficient for cylinder in uniform flow, Re = 100.
Strouhal number from CL energy spectrum
Energy magnitude
0.12
0.1
0.08
0.06
0.04
0.02
0
0.1
0.2
0.3
0.4 0.5 0.6 0.7

Strouhal number
0.8
0.9
FIGURE 3.34
Strouhal number computed from the energy spectra of the lift coefficient at Re = 100.
96
TABLE 3.2
Comparison of Strouhal Number
to Benchmark Values
Strouhal Number
Reference 55
Reference 52
Reference 49
Present study
0.16
0.171
0.172
0.169
3.13Lattice Boltzmann Method Implementation

on Graphics Processing Units
With recent advances in modern graphics processing units (GPUs), the interest in using
these devices for scientific calculations has been growing. In particular, the memory
bandwidth and computing capability of GPUs compared to contemporary central processing units (CPUs) has made their use for LBM applications particularly appealing.
Implementations of the LBM using CUDA (NVIDIA) and the C programming language
have been published and the viability of the GPU as an effective platform for executing the
LBM has been demonstrated extensively [5456].
In this section, the computational requirements for the LBM are reviewed. Next, the
NVIDIA CUDA GPU computing platform is introduced, and its use for the LBM simulations presented in this work is outlined. Comparisons are made to recently published performance benchmarks for systems with a single GPU. Last, multi-GPU implementations
of the LBM in hybrid parallel schemes employing CUDA with Open Multi-Processing
(OpenMP) as well as CUDA with Message Passing Interface (MPI) are presented. Perfor
mance of these codes is compared with a more conventional parallel implementation with
MPI.
3.13.1Computational Requirements for the Lattice Boltzmann Method
Although the operations to be executed for each lattice point during each time step are
conceptually straightforward and easy to implement on a computer, it has been well recognized that the LBM is particularly computationally intensive and memory demanding
[57]. Considerations for precision and stability combine to dictate a requirement for a large
number of lattice points to effectively discretize a problem domain. For a large number
of lattice points, each lattice point in the domain requires storage for each value of f: 9
values for the popular D2Q9 lattice and 1327 values for most commonly used 3-D lattices.
This is roughly double the requirement for a more traditional solver for the incompressible
Navier-Stokes equation.
In addition to ample memory and computational capability, the LBM requires great
memory bandwidth so that data can be streamed into the computing cores. For each time
step, each value of f must be loaded from memory and stored again at least once. The
number of floating point operations needed depends on the choice of boundary conditions
and collision operator, but as a rule of thumb, roughly 20 floating point operations are
required for every value of f. This implies that to achieve a computational performance of
97
a high-end CPU, say 500 billion floating point operations per second (500 GFLOPS), with
single-precision arithmetic, the computing device would require a memory bandwidth of
at least 200 gigabytes (GB) per second.
3.13.2Basic Implementation on Graphics Processing Units
The basic implementation of the LBM on the GPU is discussed here. The discussion has
two parts. The essential calculations required for the LBM routine are considered followed
by the question of how to best arrange the main LBM variablesf for all of the lattice
pointsin memory.
3.13.2.1LBM Routine
Every LBM routine must provide for certain identifiable milestones; briefly, these are the
following:
Problem initialization
Computation of macroscopic flow properties such as and u
Enforcement of boundary conditions to force the proper flow and solve the correct
problem
Collision to relax toward equilibrium
Streaming to propagate information across the LBM grid
Exportation of data to allow postprocessing
Several methods for initializing the values of f at each lattice point have been analyzed
in the literature [56]. For this work, all lattice points are initialized by setting u = 0 and
equal to the nominal density of the fluid to be used in the simulation. Then, feq is computed using Equation 3.10. These tasks either can be done with the CPU prior to copying
the lattice data to the GPU or can be implemented in a separate kernel prior to commencing time stepping.
Computation of macroscopic properties and enforcement of boundary conditions are
frequently done in conjunction with the collision step. This is done because the macroscopic properties are often required, within the context of the LBM simulation, for calculation of feq, which is required for collision and, for some schemes, boundary condition
enforcement. To compute macroscopic properties separately from either boundary condition enforcement or collision would require storing the values in global GPU memory. For
this reason, in light of the CUDA programming guidance to minimize global memory
transactions, the steps of computing macroscopic flow properties, boundary condition
enforcement, and collision are always done in the same kernel.
The streaming step for the classical LBM is simply a data copy operation. While this is
simple to implement, it is the subject of much research regarding how to best execute the
streaming step in a way that memory accesses are coalesced.
Last, any simulation is pointless if there is no way to evaluate the results. For this work,
intermediate values for the fluid velocity and pressure field were periodically transferred
from the GPU to the CPU and written to disk using Visualization Toolkit (VTK) file formats. For FSI computations, displacement, velocity, and acceleration data were similarly
stored for later postprocessing.
98
3.13.2.2Data Layout
The two principal alternative data layouts for the LBM computations are the so-called
array of structures (AoS) or structure of arrays (SoA). The two alternatives are illustrated
schematically in Figure 3.35 [42]. In AoS, the density distribution values f for a given lattice point are assigned in consecutive memory locations. In SoA, the density distribution
for all of the lattice points for a given lattice speed are assigned consecutive memory locations; these are followed by the density distribution function for the next velocity and so
on until all of the data have a location.
For the LBM calculations conducted on the CPU, it is most appealing to use the AoS
because this will allow the CPU to access sequential memory locations while accessing
the data for a particular lattice point. This allows for efficient memory transfers as well as
effective use of the memory cache hierarchy. In contrast, most LBM implementations on
the GPU use the SoA approach. With the SoA, when data are loaded from memory within
a kernel, each thread in a given warp reads from consecutive memory locations, as illustrated in Figure 3.36. When loads are coalesced in this fashion, the data are transferred
Consecutive memory locations

1
f0
f1
f2
i
f0
i
f1
i
f2
N2
N2
f 3 f2 f1
(a)
1
f0
f0
f0
fi
fi
f 1 f1 f1
fi
(b)
d4
d3
Th
rea
d2
rea
Th
rea
d1
Th
rea
Th
f0
f0
f0
f0
f0
Th
rea
d0
Th
rea
d1
Th
rea
d2
Th
rea
d3
Th
rea
d4
Float f0 = fEven(0*nnodes+tid);
Th
rea
d0
FIGURE 3.35
Schematic of data layout schemes: (a) the AoS and (b) the SoA. Superscripts indicate lattice node number; subscripts indicate the lattice velocity.
F loat f1 = fEven(1*nnodes+tid);
f1
f1
f1
f1
f1
Consecutive memory locations

FIGURE 3.36
When using SoA, load instructions executed by consecutive threads read from consecutive locations in memory.
99
from memory in a single transaction. A similar condition exists during store operations as
well. This is in conformance with the guidance to ensure coalesced memory access. As a
rule of thumb, using the AoS approach on the GPU penalizes achievable performance by
a factor of approximately two.
3.13.3Performance Benchmark
To compare the effectiveness of the implementation strategies adopted for this work, a
3-D lid-driven cavity problem was selected as a benchmark [42]. Three comparable works
recently published in the literature were used as comparisons [5456]. To make a fairer
comparison between all of the results, the reported performance figures were normalized
for memory bandwidth capability for the GPU device on which each comparable result
was computed. The relevant characteristics of these devices are listed in Table 3.3. The
normalized performance is shown in Figure 3.37 [42].
It can be seen from Figure 3.37 that the result of this work is comparable to or better than
other recently reported implementations. To achieve the best performance for each problem size, the number of threads per block must be adjusted accordingly. The dependence
of execution performance on the thread block size is illustrated in Figure 3.38. Using 96
threads per block performs well for all problem sizes, while the use of 256 threads in a
block performs very poorly for all problem sizes.
TABLE 3.3
Properties of GPU Devices Used in Benchmark Computations in Figure 3.38
Device
Million of lattice point updates per second
Number of CUDA cores

Global memory (MB)
Memory bandwidth (GB/s)
Estimated peak performance (GFLOPS)
GTX-260
GTX-295
GTX-480
GTX-580
192
896
111.9
805
240 2
896 2
111.9 2
805 2
480
1536
177.4
1345
512
1536
192.2
1581
Lid driven cavity D3Q19

(scaled for memory bandwidth)
800
700
600
500
This work
Rinaldi et al. (2012)
Obrecht et al. (2010)
Astorino et al. (2011)
400
300
200
100
0
32 3
64 3
96 3 128 3 144 3 160 3

N
FIGURE 3.37
Performance benchmark for LBM on a 3-D lid-driven cavity scaled for device memory bandwidth.
100
Millions of lattice point updates per second
Lid driven cavity D3Q19
800
700
600
500
64 threads per block

400
300
200
100
0
32 3
64 3
96 3
128 3 144 3 160 3
FIGURE 3.38
LBM on a 3-D lid-driven cavity with various numbers of threads per block.
4
Cellular Automata
4.1Introduction
Instead of providing a formal definition of a cellular automaton (plural: cellular automata, CA) directly, a simple example is used to explain CA. A square board is divided into
equal spacing in both directions, and each small square is called a cell. Then, a finite
number of colors are selected, and each cell is assigned to one of the colors. However,
each cell cannot have the transition of any colors. For example, one cell can be red,
another may be blue, but no cell can be half red and half blue. In this board, time is also
as discrete as the colors of the cells are. In every time step, the colors of the cells can
change according to a specified rule. As an example, the rule dictates the change in the
color of any cell depending on the colors of neighboring cells and the cells own color.
This rule applies to every cell at each time step, and the process repeats itself with every
time step. The whole board with the given rule is called the CA. The objective of the CA
technique is to represent or explain a complex physical or social phenomenon using a
simple, automated rule.
These cells do not have to be colored, but they are assigned to one of a finite number of
states at any given time step. These states may be represented by colors or may be represented by integer numbers (0, 1, 2, ). Furthermore, a finite number of alphabets will do.
Usually, the number of states is small, but in principle any finite number is acceptable.
The way that the neighboring cells are defined may be different. One can only use the
four cells on the east, west, north, and south (von Neumann neighborhood), but another
can use eight cells, such as east, west, north, south, northeast, southeast, southwest, and
northwest (Moore neighborhood). One can even use a hexagonal lattice instead of a square
lattice. These are for the two-dimensional (2-D) cases. The concept can be extended to
three-dimensional (3-D) problems. For example, the equivalents to von Neumann neighbors in 3-D are east, west, north, south, front, and back.
In the beginning, it may be assumed that every cell is in the same state, except some
finite number of cells that are in a different state. This is called the configuration. Let us
give an example that was initially proposed by Edward Fredkin [1]. The rule is defined
on a 2-D plane that consists of regular lattices (i.e., cells). Each cell is labeled by its position (i, j), where i and j are the row and column indices, respectively. Furthermore, each
cell positioned at (i, j) has the state defined at iteration t. The state can be either 1 or 0. The
CA rule determines the state of each cell at iteration t + 1 by using the states at iteration
t. The rule starts from an initial condition at time t = 0 with a given configuration of the
101
102
values 0 (i, j) on the lattice. The state at time t = 1 at a cell is determined based on the
following rule:

1. For each cell, compute the sum of states (1 or 0) on the four nearest neighboring
cells (north, south, east, and west). The domain has periodic boundaries on both i
and j directions so that this calculation applies to all cells in the domain.
2. If the sum of the four neighbors is an even number, the new state 1 (i, j) becomes
0 (or is called white) and 1 (or is called black) otherwise.
3. Repeat the process for the next time step.
This CA rule can be expressed by the following mathematical expression:
i+1 (i, j) = i (i + 1, j) + i (i 1, j) + i (i, j + 1) + i (i, j 1)
(4.1)
i+1 (i, j) = remainder of (i+1 (i, j)/2) (4.2)
Applying the CA rule starting with a small black square at the center of the domain develops a complex shape as the number of iterations continues. Figure 4.1 shows the shape
after 119 iterations. This example shows that, despite the simplicity of the local rule, the
behavior of a CA model can be complex [1].
In the previous example, the rule is applied to every cell homogeneously, leading to a
synchronous dynamics. One way to introduce spatial inhomogeneity is using nonperiodic
boundary conditions. The cells at the boundary do not have as many neighboring cells as
those inside the domain. Therefore, if the boundary condition is not periodic, a new rule
must be applied to the boundary cells. In other words, the boundary cells do not follow
the same rule as the cells inside the domain. Furthermore, multiple rules can be applied
selectively in tandem. For example, one rule is applied to one set of cells, while another
rule is applied to the other set of cells. This application can also be in tandem in terms of
the temporal domain.
t = 119
50
100
150
200
250
50
100
150
200
250
FIGURE 4.1
The rule expressed in Equations 4.1 and 4.2 on a 256 256 periodic lattice after 119 iterations.
Cellular Automata
103
The history of CA goes back to the 1940s when von Neumann introduced the concept of
extracting the abstract mechanisms leading to self-reproduction of biological organisms
[2]. In 1970, Conway developed the game of life, the well-known example of a CA [3]. The
game of life has attracted much interest because the example showed that simple rules
could generate complex patterns. In the 1980s, Wolfram showed that a CA can be applied
to many behaviors of a continuous system, as summarized in References 46. Other areas
that received attention regarding the CA technique include, but are not limited to, fluid
dynamics problems such as porous media, granular flows, spreading of a liquid droplet
and wetting phenomena, microemulsion and physical situations such as pattern formation, reaction-diffusion processes, nucleation-aggregation growth phenomena, traffic process, and so on.
4.2Strengths and Weaknesses of Cellular Automata

The power of the CA approach comes from its simplicity. In searching for a way to model
a physical system, the traditional methodology has been to solve a set of equations (e.g.,
differential equations) that satisfies the complex behavior of the system and whose solution gives the desired answers to the system. With the increasing processing power of
computers, a new way became feasible. Instead of trying to model the system as a whole,
modeling it as a sum of parts becomes possible. By using the CA approach, one can model
the system by means of simple local rules governing the behavior of the whole system. The
CA model must use some simple (and intuitive or experimental at some level) local rules
at the microscopic level but at the same time it must reflect the macroscopic behavior of the
physical system under consideration.
Numerically, an advantage of the CA approach is its simplicity and its ease of implementation on computers and parallel machines. In addition, working with Boolean quantities
prevents the problem of instability.
The weaknesses of the CA approach mostly result from its discrete nature. Some of them
are the statistical noise requiring a systematic averaging process and a less flexible rule to
describe a wider range of physical situations [1]. According to its definition and its nature,
the CA approach seems not to be suitable for modeling large-scale moving objects because
in the CA approach the only things changing are the states of the cells, not the positions of
the cells. However, there are some suggested models [7] to overcome this problem. In the
1980s, some researchers [8,9] showed that directly working with real numbers representing
the state of the cells has some advantages instead of working with Boolean cellular states.
This approach is called the lattice Boltzmann method and is numerically more efficient
than Boolean dynamics.
4.3Modeling Moving Objects Using Cellular Automata

The interest here is to model vibration of a string using the CA approach. In its current
form, the main elements that constitute a CA domain are cells, which can be in a finite
number of states, namely, 0, 1, 2, ; white/black or dead/alive; and so on. The example in
104
Section 4.1 demonstrated a CA model that consists of cells with the state of either white or
black (1 or 0). On an automaton, the particles do not move from one site to another. Instead,
the states of the cells change during iterations without any transport of matter.
To model moving objects, Chopard proposed a simple model [7] with features to deal
with large-scale objects that can move and interact with their surroundings and to allow
these objects to have adjustable mass, energy, and momentum. In addition, in the mode,
these objects maintain their sizes and their integrity during the evolution but the particles
composing them do not spread out in the entire space.
In the proposed model, the CA space is composed of particles on the lattice and the
springs that connect and hold the particles together on the lattice. Only one particle can be
linked to each end of a spring. The particles can be of two kinds: white and black. The end
points of a spring can have either color, and the consecutive particles should have different
colors. That is, both neighbors of a particle have the same color (e.g., both of the neighbors
of a white particle must be black). A spring has an orientation and length. According to
this definition, a particle initially on the arbitrary positive side of a spring should not pass
to the other side during the evolution process. The particles are allowed to alternate in a
3-D cubic lattice, but no two particles are allowed to occupy the same lattice position in the
same time step (this means no spring can be zero length or fold onto itself). An example of
a one-dimensional (1-D) lattice (called a string) is illustrated in Figure 4.2. In this example,
the positive x direction is arbitrarily selected to the right.
The rule for time evolution of the internal particles (particles with two neighbors) is
given as [7]

u(t + 1) = u+ + u u(t) (4.3)
where u(t) represents the position of a particle at time t, and u+ and u represent the positions of two neighboring particles. This equation implies that the position of a particle
at the next time step is a reflection of the particle with respect to the center of mass of
its two neighbors. In Figure 4.2, u(t) represents the position of the second black particle,
u(t) = 5. Therefore, u and u+ denote the positions of the first and second white particles,
respectively.
It is apparent that Equation 4.3 is valid only for particles with two neighbors. For the
particles at both end points, because they only have one neighbor, the reflection is performed with respect to u a, where a is a constant that represents the unstretched length
and orientation of the spring that links the particles. For convenience with the sign convention, a should be positive. Furthermore, to prevent the particles from moving off the
lattice points, a should be an integer or half integer (assuming that the lattice coordinates
are given in integers). We discuss a in detail further in the chapter.
1.5
1
0.5
0
0.5
FIGURE 4.2
One-dimensional CA lattice.
6
t=0
10
11
105
Cellular Automata
(a)
(b)
FIGURE 4.3
Time evolution of one-dimensional string particles: (a) internal particle and (b) right-end particle.
The evolution rule for a particle at the left end of a string, whose current position of the
particle is u(t), is given by

u(t + 1) = 2(u+ a) u(t) (4.4)
Similarly, the new position of a particle at the right end becomes

u(t + 1) = 2(u + a) u(t) (4.5)
The time evolution has two phases. At each time step, only one kind of particle can
move. For instance, in the first time step, all the white particles are held fixed and the equations described previously are applied to black particles only. In the first time step, only
black particles change their positions, and they move simultaneously. In the second time
step, black particles are held fixed, and only white particles move on the lattice simulta
neously, and so on [7].
Figure 4.3 demonstrates the time evolution of an internal particle and the right-end particle on a 1-D lattice with a = 3/2. The new position of the internal particle is calculated
by reflecting the particle with respect to the center of mass of its two neighboring black
particles using Equation 4.3. The new position of the particle at the right end of the string
is calculated by reflecting the particle with respect to (u + a) using Equation 4.5.
4.3.1Spring Constant
As mentioned previously, there are some constraints for defining the spring constant a.
The constant should be positive and be an integer or half integer to prevent the particles
from moving off the lattice points when assuming the lattice coordinates are given as
integers.
In addition, there are two other requirements. A spring should not fold over itself. For
instance, a particle at the left end of the string should not pass to the right of its right neighbor in any of the time evolution steps. Similarly, a particle at the right end of the string
should not pass to the left of its left neighbor in any of the time evolution steps. Here, right
and left are defined assuming the positive x coordinate is increasing to the right. Last, the
end particles should not go out of the lattice (not the same as moving off the lattice points).
This last requirement can be relaxed according to the physical system modeled. If this is
the case, one can ignore the constraint on a that comes from the last requirement. These
additional two requirements are studied separately for the left-end and right-end particles
next.
4.3.1.1Left-End Particles
The third requirement is that a spring should not fold over itself, which implies r(t + 1) < r+.
Applying Equation 4.4 to this expression yields
106
u(t + 1) < u+
2(u+ a) u(t) < u+

2 u+ 2 a u(t) < u+ (4.6)

a>
u+ u(t)
2
The last requirement implies u(t + 1) xo, where xo represents the left-end coordinate of
the lattice. Again, by substituting Equation 4.4 into the left-hand side of this equation, we
obtain
u(t + 1) xo

2(u+ a) u(t) xo (4.7)

a
xo + u(t) 2 u+
2
By combining these two expressions, the allowed interval of a for the left-end particle on
the lattice becomes

u+ u(t)
x + u(t) 2 u+
<a o
(4.8)
2
2
4.3.1.2Right-End Particles
Following the same formulation, the third requirement states u(t + 1) > r. Substitution of
Equation 4.5 into this expression yields
u(t + 1) > u
2(u + a) u(t) > u

2 u + 2 a u(t) > u (4.9)

a>
u(t) u
2
The last requirement implies u(t + 1) xL, where xL represents the right-end coordinate of
the lattice. By substituting Equation 4.5 into the left-hand side of this equation, we obtain
u(t + 1) xL

2(u + a) u(t) xL (4.10)

a
xL + u(t) 2 u
2
107
Cellular Automata
By combining these two, the allowed interval of a for the right-end particle on the lattice is

r(t) r
x + r(t) 2 r
<a L
(4.11)
2
2
The important part of these results is that if the distance between an end particle and
its neighbor is more than 2a, this rule overshoots the end particle and causes it to move to
the other side of its neighbor and fold onto itself. This also implies that the spring changes
orientation, which is a violation because we defined a as a constant. If it is exactly 2a, then
the particle collides with its neighbor and tries to occupy the same lattice point with its
neighbor. If this rule is to be used, one should bear in mind that a spring linking an end
particle and its neighbor cannot stretch more than twice its original length, whereas there
is no limit for the springs linking internal particles with exactly two neighbors.
4.3.2Velocity, Mass, Momentum, and Energy
Because the goal of the proposed time evolution model is to simulate moving objects, it is
necessary to define the velocity of a model. There are two kinds of velocities to be defined.
The first one is the velocity of the string V. The second is the speed of each particle vi,
which is the distance it will travel in the next time step. Please note that every time step is
1 unit time in this formulation, t = 1.
We define the total number of particles as N, and Ni is the ith particle numbered from the
left to the right in the string, that is, increasing toward the positive x direction. The total
mass of the string is N 1 because each end particle has a mass of and internal particles
have a mass of 1. According to this, the momentum of a string is
P=
1
v1 +
2
N 1
v + 21 v
i
(4.12)
i= 2
Reference 4 showed that the total energy of the string is given by
E=
1
2
N 1
(x
i+1
xi ax )2 (4.13)
i=1
where x stands for the coordinate of the particles, and ax is the reference value for zero strain.
It can be proved that the mass, momentum, and energy of a string are conserved during
the time evolutions by using a discrete Hamiltonian formalism [7]. According to Equations
4.12 and 4.13, one can adjust the mass, momentum, and energy of a string by adjusting the
number of particles and the initial configuration of the particles on the lattice.
To this point, our examples were for a 1-D lattice. Of course, this can be extended to
a multidimensional lattice. For that case, the time evolution formulation (i.e., Equations
4.3 through 4.5) can be applied to all directions separately, or one can define all particle
locations as vectors and apply the equations as vector algebra. The motion of a string
along the x direction is called longitudinal and along y direction is transverse. Figure 4.4
demonstrates the time evolution of a string on a 2-D lattice where a = (3/2, 0). Note that
108
(a)
(b)
FIGURE 4.4
Time evolution of particles on a two dimensional lattice: (a) t = 0 and (b) t = 1.
the internal particles are reflected with respect to the center of mass of their neighbors,
and the end particles are reflected with respect to the point that is the vector sum of their
neighbors and the a vector.
This example already deviates from the classical CA definition by using coordinates of
particles instead of the states of cells, which allowed us to model a moving object, but in
philosophy and methodology, we still use the CA approach in the sense that the coordinates of particles in the next time step are calculated according to a rule that depends on
the coordinates of neighboring particles. This rule is spatially and temporally inhomogeneous because the rules for internal and end particles are different, and they are applied
to only one kind of a particle (black/white) at each time step.
Let us go one step further and relax the rule by saying that the particles are no longer
bounded to lattice points and can go off the lattice points. This will allow us to work with
real numbers and bring us a step closer to modeling real physical situations. The practical consequence of this relaxation is that we can ignore the second rule defined for spring
vector a previously in this section; thus, this constant vector can be any real number but it
should still agree with other limitations of the spring vector.
4.4Physical Examples Using Cellular Automata

Some 1-D physical problems are solved using the CA technique. Those examples include
longitudinal vibration of a rod and transverse vibration of a string [10].
109
Cellular Automata
4.4.1Longitudinal Vibration of a Long Uniform Rod

The physical system to be modeled using the CA approach is the longitudinal vibration of
a long uniform rod fixed at one end. Initially, a force F is applied to the rod at its free end
and released at t = 0. The physical system is shown in Figure 4.5. The analytical solution
was given in [9] as
u( x , t) =
8 FL
2 AE
(2(n+1)1)
n
n= 0
sin
(2 n + 1)x
(2 n + 1)ct
(4.14)
cos
2L
2L
Let us try to model a steel rod with L = 1 m, A = 10 4 m2, E = 200 GPa, = 7860 kg/m3,
F = 200 N using the CA rule defined previously. We will use 31 particles (16 black and 15
white). The spring constant a is calculated as
a=
L
1
=
= 0.05 m (4.15)
N 1 21 1
To match the temporal part of the analytical solution, we propose a method that relates the
discrete time model of CA to real-time units, namely, seconds. In the CA model, the time
steps are originally defined as iterations, so t = 1 and it is unitless. We approach the problem by using the speed of sound for solids and introduce a time-scaling factor. The speed
of sound in the rod and speed of sound in the CA model are given as
cr =

cCA
E
; speed of sound in rod
(4.16)
L
a a
; speed of sound in CA model
=
= =
t 1 N 1
Because the distance a sound wave travels in t seconds should be the same in the rod
and the CA model, we obtain the following relationship:
cr tr = cCA tCA

E
L
L
tr =
tCA tr = tCA
N 1
( N 1) E
(4.17)
L
Time Scaling Factor =
tr = tCA TSF
( N 1)cr
x
u
A, E,
L
FIGURE 4.5
Longitudinal rod.
110
When we study the unit of the time-scaling factor TSF, it should be in the time unit because
tCA is unitless and equals 1. The unit analysis of the TSF is as follows:

[L]
[TSF ] =
[ N 1]
[E]
[]
[L]
[ M][L]
[T 2 ][L2 ]
[ M]
[L3 ]
[L]
2
[L ]
[T 2 ]
[T ][L]
= [T ] (4.18)
[L]
This analysis shows that the TSF is consistent in time units, as expected. Therefore, one
iteration in the CA model is equal to 1 TSF unit in real time. Now, let us turn back to our
example. First, we should calculate the initial displacement at the tip of the rod caused by
the force F so that we can calculate the initial configuration of the particles on the lattice.
The initial displacement is calculated as

u( x , 0) =
Fx
FL
200 1
u(L, 0) =
=
= 105 m (4.19)
AE
AE 104 200 109
so that at t = 0 the total length of the rod is 1.00001 m. When we compare the analytical and
CA solutions at the rod tip, we can see that the CA solution is in excellent agreement with
the analytical solution, as shown in Figure 4.6.
The longitudinal vibration of the rod is governed by the 1-D wave equation given in
Equation 4.20. The CA approach and local time evolution rules defined previously successfully model the macroscopic behavior of a physical system governed by Equation 4.20.
In the next examples, we also try to model another system governed by the 1-D wave
equation.
2 u
2 u
= c 2 2 (4.20)
2
t
x
10
8
10
Displacement
6
4
2
CA sol.
Analytical sol.
2
4
6
8
0
0.5
1.5
2
Time (s)
2.5
3.5
10
FIGURE 4.6
Comparison of CA and analytical solutions of the rod problem at the free tip (x = L).
111
Cellular Automata
4.4.2Transverse Vibration of String

4.4.2.1String Plucked at the Midpoint
The next example is to model a vibrating string whose both ends are fixed. It is an example
of stringed musical instruments. The string is initially plucked at the middle and released
at t = 0. The physical system and the corresponding CA model are shown in Figure 4.7.
There are some points to be mentioned. The first one is the boundary condition. In the
previous example, only the left end of the rod was fixed. However, in the string problem,
both ends are fixed so that the first and last particles in the CA model must be defined as
constrained particles. The second thing is the TSF, as defined in Equation 4.17. The TSF
includes the speed of sound cr term in the denominator. In the string problem, the speed
of sound and the TSF are defined as
To
L
TSF =
=
( N 1)cr
cr =
L
T
( N 1) o
(4.21)
where To represents the tension of the string, which is assumed to be constant along the
string, and represents the mass of the string per unit length.
The string is initially plucked at the midpoint and released with zero initial velocity. The
initial displacement at the midpoint is H = 0.1 m, the length of the string is L = 1 m, and the
speed of sound in the string is c = 1 m/s. The total number of particles is N = 101 (51 black
and 50 white). The analytical solution was given in [9] as

u( x , t) =
8H
2
n= 0
(2 n + 1)ct
(1)n
(2 n + 1)x
(4.22)
sin
cos
2
L
L
(2 n + 1)
u
H
L
x
(a)
0.1
0.05
0
0
0.1
(b)
0.2
0.3
0.4
0.5
t=0
0.6
0.7
0.8
0.9
FIGURE 4.7
String plucked at the center and corresponding CA model: (a) string plucked at the midpoint and (b) CA model
of the string problem.
112
0.1
0.08
0.06
Displacement
0.04
0.02
0
0.02
0.04
0.06
0.08
0.1
3
Time (s)
FIGURE 4.8
CA and analytical solutions of the string problem at the center.
The comparison of the analytical and CA solutions at the center of the string is shown in
Figure 4.8. The CA solution again matches the analytical solution for the string problem.
This proves that the proposed CA rule and the CA modeling approach can be used to
model physical systems governed by the 1-D wave equation.
4.4.2.2String with a Force Applied on the Middle
To this point, we have only considered displacement boundary conditions, but we need to
find a method to implement the applied force. Our next example is the vibration of a string
with a force applied to it. We will use Newtons second law to apply the force. The first
step is to discretize the total mass of the string as lumped masses connected with springs.
The white and black particles in the CA model represent the lumped masses. According to
the CA definition, the particles at the two ends of the string have half the mass of internal
particles. That is,
mi =

M
, for internal particles
( N 1)
M
, for end particles
mi =
2( N 1)
(4.23)
where mi denotes the mass of each particle, M represents the total mass of the string, and
N is the number of particles. To apply an external force to the CA model, the following
algorithm is developed:
ai (t + 1) =

Fi (t + 1) Fi (t)
mi
vi (t + 1) = vi (t) + ai (t)t
pi (t + 1) = pi (t) + vi (t + 1)t
(4.24)
113
Cellular Automata
where Fi is the force applied to the ith particle, and mi is the lumped mass of the particle.
Once the displacements of the CA particles to which the forces are applied are computed
using these equations, the local rules given in Equations 4.3 through 4.5 are applied to all
particles. Then, the computational cycle repeats the process.
The physical properties of a string are given as M = 10 2 kg, L = 0.5 m, and c = 0.5 m/s,
and a sinusoidal force is applied at the midpoint of the string as shown in Figure 4.9. The
forcing function is given as F(t) = 0.05*sin(10t). The CA solution is compared to the analytical solution at x = L/2 as seen in Figure 4.10.
There is an interesting point when the mass of the string is involved in the problem. The
TSF is no longer valid (in the form as defined earlier) in the problems where external forces
are involved. The new form of the TSF turns out to be
TSF =
( N 1)
To
M
L
=
( N 1)cr
L
M
LM
=
(4.25)
( N 1)cr
L
This formula comes from numerical experimentation and lacks a rigorous mathematical foundation. It works for the string problems when external forces are applied, so it
u
F(t)
L
FIGURE 4.9
String with force at the center.
0.05
CA sol.
FEM sol.
0.04
Displacement (m)
0.03
0.02
0.01
0
0.01
0.02
0.03
0.04
0.05
0.5
1.5
Time (s)
FIGURE 4.10
Solution for the string with a forcing function problem.
2.5
3.5
114
0.02
CA sol.
FEM sol.
0.015
Displacement (m)
0.01
0.005
0
0.005
0.01
0.015
0.02
0.5
1.5
2
2.5
Time (s)
3.5
FIGURE 4.11
Solution for the string with two forcing functions.
should be used as it is. To show that TSF works for string problems with forcing functions, here is another example. This time, there are two forces acting on the string, F1(t)=
0.05*sin(10t) and F2(t) = 0.03*sin(12t), acting at x = L/4 and x = 3L/4, respectively. The
physical quantities are given as L = 2.5 m, M = 0.01 kg, and c = 1.5 m/s. Figure 4.11 shows
the comparison of the CA and finite element method (FEM) solutions at the midpoint of
the string.
4.4.3One-Dimensional Wave Equation
The classic illustration of DAlemberts solution to the 1-D wave equation is the perturbation of a string subject to tension. For the moment, we shall apply the fixed boundary conditions to the ends of the string and focus our attention on an interior point far away from
those ends. Let us consider a string subject to a Gaussian perturbation at its midpoint as
shown in Figure 4.12 and with zero initial velocity expressed as
2
2 u
2 u
, < x < , 0 < t <
=
c
t 2
x 2
u( x , 0) = f ( x) = e
x2
2
(4.26)
u
( x , 0) = g( x) = 0
t
The DAlembert solution is

u( x , t) =
1
1
[ f ( x ct) + f ( x + ct)] +
2
2c
x + ct
x ct
g() d (4.27)
115
Cellular Automata
Initial condition
f (x)
0.8
0.6
0.4
0.2
0
20
0
< x<
10
10
20
FIGURE 4.12
Initial perturbation of an infinite string.
Substituting functions f(x) and g(x) from Equation 4.26 into Equation 4.27 yields
u( x , t) =
1
e
2
( x ct )2
2
+e
( x ct )2
2
(4.28)
The CA solution is obtained from the following cellular automata rule, which is applied
to black and white nodes in tandem. The rule can be rewritten for the 1-D wave equation
as follows:
uc(t + t) = uw(t) + ue(t) uc(t t) (4.29)
for the wave equation, Equation 4.20. Subscripts c, w, and e denote the center, west, and east
nodes, respectively. The comparison between the DAlembert solution and the CA solution
is provided in Figure 4.13 [11]. The two solutions agree well.
1
u(x, t = 4.6377)
0.8
DAlembert solution
CA solution
0.6
0.4
0.2
0
20
10
0
<x <
10
20
FIGURE 4.13
Comparison of the CA solution to DAlemberts solution to the one-dimensional wave equation in an infinite
string.
116
4.5Boundary Conditions
Fixed boundary conditions are easy to implement, but they have limited utility in modeling a potentially infinite wave domain. One possible approach is to model a much larger
domain than that of interest so that the area of interest is a distance away from the boundary. Then, the solutions are unpolluted by the imposed boundary condition. This requires
many computations that will be ignored. This is a waste of computational time. Another
approach is to apply a nonreflecting boundary, such that the wave in question is unaffected by its proximity. In the CA arena, the nonreflective boundary condition is implemented by generating virtual cells adjacent to the boundary cells as shown in Figure 4.14a
[12]. Fixed or specified boundary conditions do not require a virtual neighbor.
Figure 4.15 shows the application of several different boundary conditions to the positive
side of the spatial domain of our 1-D wave equation discussed previously. For reference,
the f(x ct) portion of the DAlembert solution is shown as the exact solution. The nonreflecting boundary condition corresponds well with the analytical solution. All four waves
peak initially in unison; the reflecting wave returns with equal amplitude, while the free
wave returns with negative amplitude. These boundary conditions can be easily extended
to the 2-D and 3-D domains.
u1
(a)
u1
(b)
u61
un1
un
un
un
un1
u1
(c)
un1
un 2unun1
u1
(d)
un1
un
u1
(e)
un1 un (1 )un + un1
u1
FIGURE 4.14
Various boundary conditions at the right end. The broken line indicates the virtual node for the boundary condition. (a) Nonreflecting boundary at the right end; (b) reflecting boundary at the right end; (c) free boundary
at the right end; (d) periodic boundary at the right end; and (e) partially reflecting boundary at the right end ( = 0:
no reflection; = 1: full reflection).
117
Cellular Automata
0.5
0.4
Wave amplitude
0.3
0.2
0.1
0
0.1
0.2
DAlemberts solution
Nonreflecting BC
Free BC
Ref lecting BC
0.3
0.4
0.5
0
4
5
Time (s)
FIGURE 4.15
Application of various boundary conditions to CA calculation of one-dimensional wave propagation.
4.6Discretization and Model Fidelity

The waves modeled by CA rules display some coarseness in the results because every
point is updated at alternating time steps. The initial perturbation displayed previously
was a Gaussian wave of medium width that proved to be smooth enough to demonstrate
the desired characteristics. Attempts to model a point source along the lines such that f(x) =
0 if x 0 and f(x) = 1 if x = 0 were unsuccessful. The question is how smooth a function
or discretization is necessary to use the CA technique to model the wave equation. To
determine a sufficiently fine discretization relative to the sharpness of the initial perturba
x2
2 2
, a series of progressively narrower Gaussians, is considered as shown in

tion, f ( x) = e
Figure 4.16. For each initial condition, the CA solution to the wave equation is calculated
for the same kind of discretization but with progressively smaller uniform lattice spacing.
Then, error norms relative to the DAlembert solution are calculated at the same arbitrary
time. Nonreflecting boundary conditions are applied in all cases. The resultant convergence is computed as a function of the lattice spacing dx. An error norm of 1% is chosen
as the reference value. The largest dx value required to achieve that level of convergence is
plotted against the half width for each initial condition as shown in Figure 4.17. A linear
2
estimate of that data is dxcrit = . Applying this estimate to the specific f(x) used previ3
ously suggests that 4.5 nodes are required to represent the peak value within an error of
1%. Therefore, a discretization that uses 11 or more nodes to represent both sides of a peak
should be sufficient.
118
1
0.8
Exp(x2/2)
0.6
Exp(x2)
0.4
Exp(4x2)
Exp(2x2)
Exp(8x2)
0.2
0
10
10
FIGURE 4.16
Initial perturbations of varying widths.
0.8
dxcrit
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
FIGURE 4.17
Critical discretization versus initial perturbation width.
4.7Convergence
To examine convergence of the CA rule for wave propagation with respect to mesh or
lattice spacing, comparison to the steady-state plane wave as presented by Reference 13
is used. Let us consider a semi-infinite fluid-filled space with a given uniform vibration
u (t) = Ue it at the z = 0 boundary and a nonreflecting boundary as z . The steady-state
pressure in the waveguide is p(z,t) = cUe(ikzit), where k = /c. The given function is applied
to the z = 0 nodes in a CA domain with the initial values everywhere else uniformly zero.
The CA rule is applied for a number of iterations corresponding to over three periods of
the steady-state solution, and the point-by-point error is calculated relative to the analytic
2
solution over the range z 0,
and plotted in Figure 4.18. The rate of convergence is
k
linear, as expected for a first-order method.
119
Cellular Automata
Relative error
100
101
Linear convergence
102
0
0.005
0.01
0.015
0.02
dx (m) [dx = dy = dz]
0.025
FIGURE 4.18
Convergence of CA wave equation rule to an analytic solution as a function of dx.
4.8Two-Dimensional Wave Problem

4.8.1Cellular Automata in Two Dimensions
In a 2-D system, each particle has four neighboring particles except for the boundary particles [7,12,14]. These particles are named the east, west, north, and south neighbors and
constitute a von Neumann neighborhood. The subscripts c, n, s, e, and w represent center,
north, south, east, and west, respectively, as shown in Figure 4.19.
For any particle surrounded by exactly four neighboring particles, the evolution in time
is expressed as

uc (t + t) =
ue (t) + uw (t) + un (t) + us (t) 2 uc (t)

(4.30)
2
2
1.5
1
0.5
0
0.5
1
1.5
2
2
1.5
0.5
FIGURE 4.19
Lattice and particles in a two-dimensional CA model.
0.5
1.5
120
Equation 4.30 is an expansion of the local CA rule that we defined for 1-D grids. On the
other hand, the rule used for particles at the east-side boundary with three neighbors is

uc (t + t) =
2(uw (t) + awe ) + un (t) + us (t) 2 uc (t)

(4.31)
2
Similarly, the rules for west, north, and south boundary particles are, respectively,

uc (t + t) =
2(ue (t) awe ) + un (t) + us (t) 2 uc (t)

(4.32)
2
uc (t + t) =
2(us (t) + asn ) + ue (t) + uw (t) 2 uc (t)

(4.33)
2
uc (t + t) =
2(un (t) asn ) + ue (t) + uw (t) 2 uc (t)

(4.34)
2
Finally, there is a different rule for a corner particle. For example, for the northeast boundary of the lattice, the rule is

uc (t + t) =
2(us (t) + asn ) + 2(uw (t) + awe ) 2 uc (t)

(4.35)
2
Similar rules can be developed for other corners of the grid. In the equations just presented, ai (i = we or sn) is the spring vector constant that gives the orientation and equilibrium length of the springs at each direction. Vector ui is the position vector of each particle
with respect to a given origin point.
These equations define the spatial time evolution of the particles, but for the temporal
part to match the real-time scale, we must define a TSF for 2D domain problems. The initial
central displacement influences all the neighboring lattice nodes after two iterations. The
maximum distance between the center node and the neighboring corner node is 2x, where
x is the uniform spacing between two neighboring lattice points. Because the speed of
sound is c, the TSF is

TSF =
2x
(4.36)
c
Because the equations and methodology have been defined, we give an example of a 2-D
membrane with an initial displacement at the center.
4.8.2Example of Membrane Vibration
The membrane has a square shape and is 1.7 m in each dimension; it has a mass of 0.2 kg.
The membrane domain is divided into 40 equal spacings in both directions, resulting in a
total of 41 41 = 1681 lattice nodes. The central lattice node is initially displaced with 0.01 m
as shown in Figure 4.20.
121
Cellular Automata
Displacement
0.015
0.01
0.005
0
20
15
10
y-axis
5
0
20
15
10
x-axis
FIGURE 4.20
Initial configuration of the membrane.
The membrane is clamped at all four sides, and it is released from its initial configuration with the zero initial velocity. The problem is analyzed using the CA rules described
previously. Figure 4.21 shows the comparison of the membrane displacements at the center
for both CA and FEM solutions [14]. Both solutions agree well each other.
The CA technique is simple and easy to implement. If all of the boundary nodes are
fixed, Equation 4.30 is the only equation used for all internal nodes with white and black
nodes in tandem. The operation does not deal with any large system matrix. Therefore,
the CA calculations are much faster than those for FEM, especially as the number of nodes
increases in the system. Table 4.1, for example, compares the computing times for the CA
and FE methods for the membrane example using different numbers of nodes in the model.
0.01
CA sol.
FEM sol.
0.008
Displacement (m)
0.006
0.004
0.002
0
0.002
0.004
0.006
0.008
0.01
0
3
Time (s)
FIGURE 4.21
Comparison of CA and FEM solutions for the membrane problem.
122
TABLE 4.1
Normalized Computing Times for CA and FEM
Number of Nodes
CA
FEM
FEM/CA
11 11 = 121
21 21 = 441
31 31 = 961
41 41 = 1681
51 51 = 2601
1
7.08
24.4
59.6
111
1.58
20.4
237
1042
3231
1.58
2.88
9.71
17.5
29.1
TABLE 4.2
Comparison of Number of Required Algebraic
Operations for CA and FEM
Addition and subtraction
Multiplication and division
CA
FEM
4s
2s
2s2 + 6s
2s2 + 8s
An approximate number of algebraic operations (i.e., addition, subtraction, multiplication, and division) required during each time step in CA and FEM is given in Table 4.2,
where s is the system degrees of freedom, which are equal to the degrees of freedom times
the number of nodes. As shown in Table 4.2, the number of algebraic operations for FEM
increases in a quadratic function of the total degrees of freedom, while the number of
operations increases linearly for CA. When modeling a large system, the CA technique is
much quicker than FEM, as shown in Table 4.1.
4.9Three-Dimensional Wave Problem

It follows from the 1-D and 2-D CA expressions that the 3-D wave equation can be modeled as

uc (t + t) =
1
(uW (t) + uE (t) + uS (t) + uN (t) + uF (t) + uB (t) 3uC (t t) (4.37)
3
with the subscripts F and B on the new terms standing for front and back, respectively. To
test this supposition, consider a point source located in the center of a domain of interest. As discussed previously, CA is not expected to faithfully model a true point source, so
the source under consideration is a smooth radial function that has a maximum value of
one at the origin of the domain and is zero valued outside a radius of five nodes from the
origin. The domain comprises 73 equally spaced nodes in each direction, and nonreflecting boundary conditions are applied on all six sides of the domain. Three points in space
are examined: the origin, an arbitrary point inside the initial perturbation (r < 5dx), and an
arbitrary point outside the initial perturbation (r > 5dx).
The analytical solution for the 3-D wave problem is given as

2 u
= c 2 2 u, ( x , y , z) R 3 (4.38)
2
t
123
Cellular Automata
u(x, y, z, 0) = (x, y, z) (4.39)
u
( x , y , z, 0) = ( x , y , z) (4.40)
t
and is expressed as
u( x , y , z, t) = t +
t (4.41)
t
( )
where is the gradient operator, and are the averages of their respective initial conditions (Equations 4.39 and 4.40) over the sphere of radius ct centered at (x, y, z). In other words,
( x , y , z) =
1
4c 2 t 2
( x + ct sin cos , y + ct sin sin ,
(4.42)
z + ct cos )(ct) sin dd
( x , y , z) =
1
4c 2t 2
( x + ct sin cos , y + ct sin sin ,

z + ct cos )(ct)2 sin dd
(4.43)
The integrals in Equations 4.42 and 4.43 are difficult to evaluate analytically unless the
functions and have a simple mathematical expression.
Using the notations presented, the initial conditions for the point source considered are
expressed as
r
1
( x , y , z) =
a
if r < a
(x, y, z) = 0
(4.44)
if r a
(4.45)
where r = x 2 + y 2 + z 2 is the Cartesian radius from the center of the source. The analytical solution of the integrals in Equation 4.41 is not easily calculated. Therefore, a numerical
integration using Simpsons rule yields a suitable comparison [11].
Figures 4.22 through 4.24 show the analytical solutions at the respective points plotted as a
function of time directly over their CA counterparts versus the number of iterations through
which the rule is applied. These plots illustrate that we need to match the discrete iterations of a
cellular automaton to the continuous time domain or calculate an appropriate time-scalingfactor.
Consider a true point source located at the origin of an otherwise-zero-valued CA
domain. Following the current CA rule, the earliest a node at (dx, dy, dz) can reach a value
other than zero is after the third iteration. Thus, 3dt = dx 2 + dy 2 + dz 2 . For the equal-size
spacing, the time-scaling factor for one iteration in CA is equivalent to

dtCA =
dx
c ndim
(4.46)
124
Analytical solution at (0,0,0)
1
0.5
0
0.5
1
0
Time
5
104
At CA solution (0,0,0)
1
0.5
0
0.5
1
0
10
20
30
Iterations
50
40
FIGURE 4.22
Comparison of a three-dimensional wave model at domain origin: Time versus iterations.
Analytical solution at (dx, 3dx, 2dx) (r = 3.75dx)

0.2
0.1
0
0.1
0.2
Time
CA solution at (dx, 3dx, 2dx)
5
104
0.2
0.1
0
0.1
0.2
10
20
30
Iterations
40
50
FIGURE 4.23
Comparison of a three-dimensional wave model at a point inside the initial perturbation: Time versus iterations.
125
Cellular Automata
Analytical solution at (7dx, 8dx, 6dx) (r = 12.2dx)

0.04
0.02
0
0.02
0.04
0.06
Time
5
104
CA solution at (7dx, 8dx, 6dx)

0.04
0.02
0
0.02
0.04
0.06
10
20
30
Iterations
40
50
FIGURE 4.24
Comparison of a three-dimensional wave model at a point outside the initial perturbation: Time versus iterations.
in which ndim is the dimension of the problem, which is three for the present example, and
dx is the lattice spacing. The points of interest for this exercise were chosen arbitrarily, but
in such a way that both the inside and outside points are displaced from the origin in all
three directions. Comparisons between the analytical and CA solutions plotted with this
time equivalency are shown in Figures 4.25 through 4.27 [11]. Both solutions agree well in
all the locations.
0.8
Analytical
0.6
CA
0.4
0.2
0
0.2
0.4
0.6
0.8
FIGURE 4.25
Three-dimensional wave model at a domain origin.
Time
5
104
126
Analytical
0.2
CA
0.1
0
0.1
0.2
Time
4
104
FIGURE 4.26
Three-dimensional wave model at a point inside the initial perturbation.
0.06
Analytical
CA
0.04
0.02
0
0.02
0.04
0.06
Time
5
104
FIGURE 4.27
Three-dimensional wave model at a point outside the initial perturbation.
4.10Application to Underwater Acoustics

An acoustic model of wave propagation is applied to sound travel in water. The basis of
underwater wave propagation models falls into one of the following five categories, which
are based on different variances of the wave equation [15]:

1. Ray theory
2. Normal mode
3. Multipath expansion
4. Fast field (wave number integration)
5. Parabolic equation
127
Cellular Automata
Current selections of propagation models are limited in their application due to constraints presented in the domain [15]. Many propagation models circumvent this problem
by establishing a hybrid modeling scheme [16]. The prior modeling approach links several
different models into one to solve complex problems with various domain features. This
section illustrates transmission losses due to wave propagation as a function of frequency
and domain constraints.
4.10.1Transmission Loss Development
Sound propagation in water undergoes a reduction in intensity as it travels between points.
Spreading and attenuation are mechanisms used to quantify the reduction in sound as it
travels over a distance. By definition, propagation losses are calculated by Equation 4.47:
Intensity 1
PropagationLosses (PL) = 10log
(dB) (4.47)
Intensity r
where
Intensity =
(PressureAmplitude)2
(4.48)
density ofseawater *speedofsound inseawater
Here, the subscript 1 refers to the intensity at 1 m, and r refers to an arbitrary length
greater than 1 m [17]. For simplicity, the domain is homogeneous, and the speed of sound
is constant throughout the domain. As a result, Equation 4.47 reduces to the form shown
in Equation 4.49:
PressureAmplitude1
PropagationLosses(PL) = 20log
(dB) (4.49)
PressureAmplituder
As a way of illustrating propagation losses, a 2-D domain is established with the following parameters: The number of grids in both horizontal and vertical directions is 101. The
acoustic source is located at the grid (35, 65) as a sinusoidal function with a frequency of
800 Hz, and the radius of the sound source is four grid spacings. All boundaries are nonreflective, and the speed of the sound of water is 1500 m/s.
Because the source peak value is known, this value is implemented for the pressure
amplitude at 1 m in Equation 4.49. Propagation losses are displayed in Figure 4.28. The
losses are based on Equation 4.49. Figure 4.29 extracts a point at the grid point (35, 65) from
Figure 4.28 and illustrates how the pressure changes as a function of time. From looking at
Figure 4.29, the peak value of nodal pressures changes significantly around time 0.015 s. As
a result, calculating propagation losses solely on peak values do not give an accurate depiction of the predominant values seen at that point throughout all the iterations. To correct
this issue, let us take the root mean square (RMS) pressure from all iterations. Calculation
of the RMS pressure at point (35, 65) gives a value of 0.2830 Pa. This has a significant effect
in calculating propagation losses seeing that a peak of 2.709 Pa is originally observed in
128
Rows
Propagation losses
50
10
45
20
40
30
35
40
30
50
25
60
20
70
15
80
10
90
100
20
40
60
Columns
80
100
FIGURE 4.28
Propagation losses based off peak-amplitude pressure values. Nonreflecting boundary conditions were applied
to all sides.
Location: row = 35 and column = 65
Nodal pressure values (Pa)
2.5
2
1.5
1
0.5
0
0.5
1
0
0.02 0.04
0.06 0.08 0.1

Time (s)
0.12 0.14 0.16
FIGURE 4.29
Depiction of nodal pressure changes inside a 101 101 domain at 800 Hz.
Figure 4.29. As a result of making this change, Figure 4.30 plots how propagation losses
are affected for every location in the domain [18]. In turn, modification to Equation 4.49
results in the RMS values being used instead of peak values in the calculation of propagation losses, as shown in Equation 4.50:

RMSPressureofSource
PropagationLosses(PL) = 20log
(dB) (4.50)
RMSPressurer
129
Cellular Automata
Propagation losses (dB)
60
10
50
20
30
40
Rows
40
50
30
60
20
70
80
10
90
100
20
40
60
Columns
80
100
FIGURE 4.30
Propagation losses based on RMS pressure values at 800 Hz with nonreflecting boundary conditions on all
sides.
4.10.2Various Boundary Conditions

The beauty behind the CA method is the ease in which various boundary conditions can
be added to the domain. The intent of this section is to illustrate wave propagation in a
domain not easily modeled by ordinary modeling methods. Parameters that define the
domain in this section are as follows: The acoustic domain is 980 980 m with a uniform
nodal spacing of 10 m. The acoustic source is now at a depth of 10 m below the water line
and 30 m from the left boundary of the domain. The sound source has a frequency of
25Hz. Details of its location are described in the sections that follow.
4.10.3Wave Propagation across a Flat-Bottom Ocean Floor
For the first set of simulations, there is no restriction on the left and right boundaries. This
in turn requires nonreflecting boundary conditions on the left and right walls. The ocean
floor behaves as a perfect reflector, and the air-to-waterline boundary is a free boundary.
The receiving point is 630 m below the waterline and also 630 m away from the source on
the horizontal axis. Figure 4.31 shows the results of how the pressure varies at the receiver
as a function of time [18]. The effects of modeling the floor as a perfect reflector become
apparent at this particular point around 1.056 s, as shown in Figure 4.31. The initial wave
front peaks are approximately 1.2 Pa. As time progresses, the wave front changes the
peaks to approximately 2.1 Pa around time 1.0559 s at our receiver location because of the
reflection from the perfect bottom reflector. In viewing Figure 4.32, the distortion propagates its way throughout the entire domain. Because the acoustic source is still active in the
domain, interference patterns develop and result in a more chaotic profile after 2.2156 s.
Nevertheless, Figure 4.33 provides an effective measure of propagation losses over a range
of reflection from the bottom boundary.
130
Location: row = 65 and column = 65

2
1
0
1
2
3
0
0.5
1.5
Time (s)
Time = 0.58454 s
10
0
10
10
20
30
40
Rows
50
60
70
80
90
100
10
20
30
40
50
60
70
80
100
90
Nodal pressure (Pa)
Nodal pressure (Pa)
FIGURE 4.31
Receiver reception of changes in acoustic pressure from a 25-Hz source with a flat ocean bottom with 100%
reflection.
15
10
5
0
5
10
15
10
Columns
Nodal pressure (Pa)
10
20
30
40
50
Rows
60
70
80
90
100
20
30
40
50
60
Rows
Time = 0.84856 s
5
0
5
Time = 0.70239 s
10
20
30
40
50
60
70
80
90
80
70
80
90
100
10
20
30
90
100
70
60
50
40 Columns
100
Columns
FIGURE 4.32
Three-dimensional depiction of wave propagation with nonreflecting boundary conditions on the left and
right, free conditions on the top, and 100% reflection on the bottom surface. Snapshots of wave propagation, as
a function of time, were captured at a source frequency of 25 Hz.
(Continued)
131
Cellular Automata
Nodal pressure (Pa)
Time = 1.0559 s
10
0
10
10
20
30
40
50
Rows 60
80
90
100
10
20
30
40
10
0
10
10
20
30
40
50
60
Rows 70
80
90
100
60
70
100
90
Columns
Time = 2.2156 s
Time = 1.3058 s
Nodal pressure (Pa)
Nodal pressure (Pa)
70
50
80
15
10
5
0
5
10
10
20
30
40
50
60
70
80
90
100
Columns
15
10
20
30
40
50
Rows 6070
80
90
100
20
40
60
80
100
Columns
FIGURE 4.32 (CONTINUED)

Three-dimensional depiction of wave propagation with nonreflecting boundary conditions on the left and
right, free conditions on the top, and 100% reflection on the bottom surface. Snapshots of wave propagation, as
a function of time, were captured at a source frequency of 25 Hz.
4.10.4Wave Propagation over a Curved Hill

This example illustrates the nature of wave propagation with an ocean floor that exhibits a
curved shape. The simulated domain is the same as that in the previous section. The only
exception is that reflective boundary conditions are applied to the curved hill. Changes in
the bottom boundary yield significant changes in the acoustic pressure profile. Changes
in the receiver pressure begin increasing around 0.59 s after the initial front of the wave
begins reflecting back from the curved hill, as shown in Figure 4.34 [18]. This is a result of
reflective waves reinforcing the propagation of the incoming source. As time progresses,
there is attenuation in the magnitude as a result of the reflective phenomenon occurring
out of phase with the incoming source while returning at the same incident angle.
132

Free BC
10
250
20
20
200
150
Nonreflecting BC
Nonreflecting BC
60
Rows
Rows
50
100
50
10
20
30
40
50
60
70
80
90
100
100
20% Reflection BC
10
20
30
Free BC
40
50
60
Columns
70
80
90
100
250
100
200
20
30
250
Free BC
10
20
200
30
150
Nonreflecting BC
Nonreflecting BC
60
100
70
Rows
Rows
Nonreflecting BC
50
90
Columns
10
40
50
150
Nonreflecting BC
Nonreflecting BC
60
100
70
50
80
90
100
150
Nonreflecting BC
80
100% Reflection BC
90
50
50
70
80
40
40
60
70
100
200
30
30
40
250
Free BC
10
90
65% Reflection BC
10
20
30
40
50
60
Columns
70
50
80
80
90
100
100
10
20
30
2% Reflection BC
40
50
60
Columns
70
80
90
100
FIGURE 4.33
Propagation losses of a 980 980 m domain at a frequency of 25 Hz with various ranges of reflection from the
bottom boundary.
When comparing Figure 4.34 through Figure 4.32 at time 1.0559 s, the addition of the
curved boundary has allowed for a reflected wave to essentially propagate throughout
the entire domain in a faster time frame. This essentially shows an accurate depicted of
wave propagation when the curvature of the domain is changed and the model is accurately accounting for the constraints along the boundary. Nevertheless, the addition of the
curved boundary causes significant changes in propagation losses along the lower left and
right corners of the domain, as shown in Figure 4.35. The decrease in magnitude can be
attributed to the fact that the perfect reflection off the curved hill increases the acoustic
pressure. The reflection is at an oblique angle to the source, so destructive interference
does not occur in a manner as shown on the top left and right portions of the hill.
4.10.5Wave Propagation over a Sloping Bottom
The scenario in this example describes modeling wave propagation across a linearly sloping bottom. The bases of this domain formulation are from the wedge shape domain
depiction in Reference 19. Here, a similar size domain emulates the 3-D shape depicted in
Figure 4.36. The first part discusses propagation in an open ocean environment, that is, a
domain with infinite lateral limits. The latter part of this scenario models wave propagation in a confined waterway channel.
133
Cellular Automata
Nodal pressure (Pa)
Time = 1.0559 s
10
0
10
10
20 30
Nodal pressure (Pa)
Time = 0.59397 s
40
50
Rows
60
70 80
90
100
10
20
30
40
50
60
70
80
90
100
Columns
10
0
10
20
30
40
Rows
80
50
60
70
80
90
100
10
20
30
40
70
60
50 Columns
90
100
Time = 2.2156 s
Nodal pressure (Pa)
10
10
0
10
10
20
30
40
50
Rows
60
70
80
90
100
10
20
30
40
50
60
70
80
90
100
Columns
FIGURE 4.34
Snapshot at three instances in time of wave propagation in a two-dimensional domain with a curved floor bottom boundary.
To simulate wave propagation in a confined water channel, the front and back walls of
the domain must exhibit reflective boundary constraints. The floor of the domain, on the
other hand, is maintained at 65% partial reflection. Last, the water-to-air interface is maintained as a free boundary.
Simulation results of propagation losses are plotted in Figures 4.36 and 4.37 [18].
Accordingly, the effects of reflection are really noticeable along the sides of the domain
at higher percentages of reflection. Consequently, development of this modeling scheme
causes changes in the pressure profile and emulates changes in the loss profile. Significant
changes can be on the order of 16 dB, significant when comparing Figure 4.37a and d.
Nevertheless, refinement of this domain can be adapted to meet the needs of a modeling
scenario that may require these restraints.
4.11Multigrid Technique of Cellular Automata

The intent of this section is to develop a refinement scheme to increase the accuracy of a
solution in certain regions of the cellular domain. The problem that has not been addressed
is what happens if the number of columns and rows is expanded to be large in magnitude,
134

Free BC
10
20
Nonreflecting BC
100
100% Reflection BC
80
50
10
20
30
40
50
60
Columns
70
80
90
100
100
20% Reflecting BC
80
100
50
60
100
65% Reflection BC
50
90
30
40
50
60
Columns
70
80
90
100
60
70
80
90
100
40
250
Free BC
200
Nonreflecting BC
Nonreflecting BC
150
50
60
100
2% Reflecting BC
70
80
50
Columns
30
150
Nonreflecting BC
40
20
200
20
30
Nonreflecting BC
30
10
20
10
20
70
10
250
Free BC
10
Rows
60
150
90
100
100
50
Nonreflecting BC
Nonreflecting BC
70
90
50
Rows
70
40
150
Rows
Rows
Nonreflecting BC
200
30
60
40
20
200
40
250
Free BC
10
30
50
250
80
50
90
100
10
20
30
40
50
60
70
80
90
100
Columns
FIGURE 4.35
Propagation losses in a 980 980 m domain with a curved ocean floor for various ranges of bottom reflection.
The acoustic source had a frequency of 25 Hz.
for example, 104 nodes in all three dimensions. In all likelihood, a user utilizing a standard personal computer is going to experience many problems and receive a simulation
error due to insufficient memory. To circumvent this issue requires the user to perform the
simulation on a supercomputer or in an iterative save-and-repeat manner.
To circumvent issues with memory allocation as well as to achieve computational efficiency, this section presents a refinement process to capture the essence of the pressure
profile in a uniform grid with development of a localized finer grid. A uniform grid correlates to a domain with one kind of global grid inside the CA domain. A multigrid domain
refers to a domain that has global grid points with a local grid system internal to some
parts of the global domain. An analogy to this method in computational mechanics is that
it is similar to creating smaller mesh sizes inside a large mesh size.
4.11.1Global-Local Technique
To illustrate the concept of this technique, let us examine the cellular automaton grid points
in Figure 4.38. Here, the multigrid domain has 11 11 global grid points. Of interest is
the local grid domain-associated solid circles. The solid circular corners and the bounded
zone by them in Figure 4.38 are further refined to 9 9 local grid points as shown in
135
Cellular Automata
10
20
30
40
50
60
70
80
Propagation losses (dB) Nonreflective BC
100% Reflective BC
Free BC
20
Nonreflective BC
30
30 40
Z-nodes
10
50 60
20
50
10
30
30
20
50
40
60
80
Nonreflective BC
20% Reflective BC
40
50
60
70
80
30
10
20
10
20
40
50
60
40
Columns
60
60
70
80
2% Reflective BC
Free BC
50
65% Reflective BC
30
50
2% Reflective BC
40
50
20
30
40
Columns
20
Nonreflective BC
30
40
10
60
10
65% Reflective BC
30 40
50 60
40
Rows
Rows
20
Nonreflective BC
30
Z-nodes
Columns
Free BC
Z-nodes
70
65% Reflective BC
60
10
10 20
60
Nonreflective BC
20
65% Reflective BC
60
50
40
65% Reflective BC
10 20
40
30
100% Reflective BC
Rows
60
30
10
40
50
20
20% Reflective BC
Rows
10
10
50
60
65% Reflective BC
10
20 30
40 50
Z-nodes
60
10
20
30
40
50
60
Columns
FIGURE 4.36
Propagation losses of a wedge-shape bottom domain at 110 Hz for various ranges of front and rear deflection.
The bottom boundary will reflect 65% of the incoming wave.
Figure 4.39, where the solid circular corners are also shown. Such a local refinement can be
undertaken at selective global grids. If all the global grids are refined in the same manner,
it is equivalent to the 81 81 grid points. In that case, there is no reason to distinguish the
global and local grids because all the local grids can be solved in the same way as handled
previously. Instead, the local refinement is conducted only in necessary global grid zones.
The inclusion of the surrounding gray area in Figure 4.39 accounts for necessary boundary
regions to be performed for the CA technique, which is discussed further in this chapter.
To this point, all simulations were created for one uniform grid. The remaining parts of
this section illustrate how to exchange information between the local and global grids for
the multigrid analysis. For example, analysis is conducted to relate the pressure change at
the point (5, 5) in Figure 4.38 to its corresponding point (2, 2) in Figure 4.39. Throughout all
simulations, the acoustic source is placed in the local grid for all multigrid scenarios. First,
the CA analysis is performed for the local domain grids, including the acoustic source.
We apply nonreflecting boundary conditions to the boundary of the local grid domain.
The number of analyses of the local CA grid system before the next global CA analysis
depends on the number of grids in the local grid system. This is decided to match the time
increments between the local and global mesh sizes. Considering Figures 4.38 and 4.39, the
mesh size in the local mesh is one-eighth of the global mesh size. The time increment is
136
55
50
45
40
35
30
10
20
30
40
50
60
60
55
50
45
40
35
10
20
30
65
60
55
50
45
40
35
(c)
Z-nodes
(a)
(b)
40
50
60
20
30
Z-nodes
40
50
60
10
20
30
40
Z-nodes
50
60
65
60
55
50
45
40
35
(d)
Z-nodes
10
FIGURE 4.37
Propagation losses of a wedge-shape bottom domain along column 45 and row 31 from Figure 4.36 at 110 Hz
for (a) 100%, (b) 65%, (c) 20%, and (d) 2% front and rear reflection. The bottom boundary exhibited 65% partial
reflective boundary conditions.
1
2
3
Rows
4
5
6
7
8
9
10
11
7
6
Columns
10 11
FIGURE 4.38
Cellular automaton multigrid domain including the boundary nodes.
proportional to the mesh size. In other words, the time increment in the local mesh is oneeighth that in the global mesh. This means that the local CA analysis is conducted eight
times, and then the global CA analysis is conducted once. After the global CA analysis is
conducted for all the grid nodes, the local CA analysis is undertaken eight times again.
This process repeats continuously. For the global analysis, the values at the global nodes,
which were determined from the local grid analysis, are used.
137
Cellular Automata
1
2
3
4
Rows
5
6
7
8
9
10
11
5
6
7
Columns
10
11
FIGURE 4.39
Local grid domain of a cellular automaton domain including the associated boundary nodes.
4.11.2Example Problems
Throughout numerical experimentation to assess this method, a dependence on the frequency and the level of refinement of the global and local grids is considered [18]. The
first example case has a uniform grid of 401 401. An equivalent multigrid scheme has its
global grid of 71 71. The acoustic source is properly aligned between the multigrid and
uniform grid cases. On execution of both models, the results of both simulations for the
zero-input frequency case are shown in Figure 4.40. A direct comparison of the associated
pressure rise, at near-identical locations, shows similar profiles between the two methods.
In particular, the starting time for the associated pressure rise is nearly identical. On the
other hand, the multigrid method appears to approach the peak magnitude at a faster rate.
As the next problem, the acoustic input frequency has changed to 10 Hz. The correlations between the two domains show consistently higher peaks with the multigrid propagation scheme as shown in Figure 4.41. On the contrary, there is consistency in the onset
time for wave propagation. Nevertheless, the effects of the local and global grid mesh size
contribute to prevent the multigrid solution from matching the uniform grid solution. As
such, the following simulation set evaluates how changes in their associated values affect
the overall performance.
The first sets of revisions are made to the global grid. The grid spacing in the global
model is reduced to one-quarter of the original global model. The comparison of the
refined global grid in reference to the original uniform grid is shown in Figure 4.42. The
effects of refining the global grid produce results that closely match the original uniform
grid, with its computational cost approximately one-third that for the uniform grid.
The next set of evaluations considers the effects of changing the mesh size inside the
local grid. The local grid is increased by a factor of approximately four. The reason for
choosing this factor is for direct comparison with the global refinement process previously
discussed. The simulation results are shown in Figure 4.43. The performance of the model
shows that refining only the local grid does not produce results that are more accurate in
the global domain by that associated method alone.
138
Multigrid nodal points
Uniform grid nodal points
Location: row = 2, column = 2, Z = 1


8
X: 0.182
Y: 0
0
0
0.5
1
1.5
Time (s)
0
0
X: 0.1254
Y: 0.01015
0.5
1
1.5
Time (s)
0.5
1
1.5
Time (s)
X: 0.1212
Y: 0.006057
X: 0.05798
Y: 0.002103
0.5
1
1.5
Time (s)
0
0
X: 0.05893
Y: 0.003903
0.5
1
1.5
Time (s)
0
0
0
0
0
0
0
0
1
1.5
Time (s)
0.5
X: 0.189
Y: 0
X: 0.08485
Y: 0.0
0.5
1
1.5
Time (s)
0
0
X: 0.08155
Y: 0
0.5
1
1.5
Time (s)
FIGURE 4.40
Comparison of equivalent nodal points, inside a uniform grid and a multigrid, from a 0-Hz acoustic source.
139
Cellular Automata
2
1.5
1
0.5
0
0.5
1
1.5
2


2
1.5
1
0.5
0
0.5
1
1.5
2
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)

2
1.5
1
0.5
0
0.5
1
1.5
2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time (s)
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
2
1.5
1
0.5
0
0.5
1
1.5
2
0
2
1.5
1
0.5
0
0.5
1
1.5
2
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
2
1.5
1
0.5
0
0.5
1
1.5
2
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
2
1.5
1
0.5
0
0.5
1
1.5
2
1.5
1
0.5
0
0.5
1
1.5
2
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
FIGURE 4.41
Comparison of a uniform grid and a multigrid, from a 10-Hz acoustic source, at near-identical nodal distances
from the source.
(Continued)
140

2.5
2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
0
2
1.5
1
0.5
0
0.5
1
1.5
2
0
2.5
2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)

Comparison of a uniform grid and a multigrid, from a 10-Hz acoustic source, at near-identical nodal distances
from the source.
141
Cellular Automata
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
2
1.5
1
0.5
0
0.5
1
1.5
2
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
2
1.5
1
0.5
0
0.5
1
1.5
2
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
2
1.5
1
0.5
0
0.5
1
1.5
2
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
2
1.5
1
0.5
0
0.5
1
1.5
2
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
2
1.5
1
0.5
0
0.5
1
1.5
2

2
1.5
1
0.5
0
0.5
1
1.5
2
2
1.5
1
0.5
0
0.5
1
1.5
2
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
2
1.5
1
0.5
0
0.5
1
1.5
2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Time (s)
FIGURE 4.42
Comparison of a uniform grid and a multigrid where the global grid points of the multigrid were refined by a
factor of four. The nodal points from both domains are at approximately equal nodal distances from the 10-Hz
acoustic source.
(Continued)
142


2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)

2
1.5
1
0.5
0
0.5
1
1.5
2

2
1.5
1
0.5
0
0.5
1
1.5
2
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)

Comparison of a uniform grid and a multigrid where the global grid points of the multigrid were refined by a
factor of four. The nodal points from both domains are at approximately equal nodal distances from the 10-Hz
acoustic source.
143
Cellular Automata
2
1.5
1
0.5
0
0.5
1
1.5
2
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
2
1.5
1
0.5
0
0.5
1
1.5
2
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Time (s)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
2
1.5
1
0.5
0
0.5
1
1.5
2
0
2
1.5
1
0.5
0
0.5
1
1.5
2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Time (s)
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
2
1.5
1
0.5
0
0.5
1
1.5
2

2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Time (s)
3
2
1
0
1
2
3
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
FIGURE 4.43
Comparison of a uniform grid and a multigrid where the local grid points of the multigrid were refined by a
factor of four. The nodal points from both domains are at approximately identical distances from the 10-Hz
acoustic source.
(Continued)
144

2.5
2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
2
1.5
1
0.5
0
0.5
1
1.5
2
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
2
1.5
1
0.5
0
0.5
1
1.5
2
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)
2.5
2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time (s)

Comparison of a uniform grid and a multigrid where the local grid points of the multigrid were refined by a
factor of four. The nodal points from both domains are at approximately identical distances from the 10-Hz
acoustic source.
5
Molecular Dynamics
5.1Introduction
Molecular dynamics (MD) studies the interaction among neighboring atoms or molecules
to trace the time history of their motions [1,2]. Then, the macroscopic material properties
are computed from the motions of the atoms or molecules. This chapter presents the basic
formulation of classical MD and the mathematical expressions for commonly used interatomic potential energy functions. MD is time consuming and computationally expensive
because the time step size used in the analysis is small and the number of atoms or molecules is usually large. As a result, a version of molecular mechanics is developed to save
computational cost [3]. Then, some application cases are discussed. The examples are the
analysis of carbon nanotubes (CNTs), polymers, metals, and heat transfer.
5.2Classical Molecular Dynamics Formulation

Classical MD is based on Newtonian mechanics. At any given time, the resultant force
applied to each molecule or atom by neighboring molecules or atoms is computed using
the appropriate intermolecular or interatomic potential energy function. Then, the acceleration of the molecule or atom is determined using Newtons second law:
n
i=1
d2 r
Fik = mk 2k (5.1)
dt
where Fik is the

force vector applied to the kth molecule or atom by the ith molecule or
atom; mk and rk are the mass and the position vector of the kth molecule or atom, n is the
number of neighboring molecules or atoms that influence the kth molecule or atom; and t
denotes time. From this point, the term atom is used, which can simply be replaced by the
term molecule.
The interactive forces among atoms are computed from the interatomic potential energy
functions as discussed in the material that follows. For an isolated system, the total energy
E is conserved. The total energy can be expressed as the Hamiltonian H. For a system with
N atoms, H takes the form
n
H=E=
i=1
pi2
+ U (5.2)
2 mi
145
146
where pi is the momentum of an atom and the potential energy U results from the inter
atomic interactions. Taking the derivative of this equation with respect to the position ri ,
the explicit relationship between the Hamiltonian and the potential energy becomes
H U
= (5.3)
ri
ri
The resultant Hamiltons equations of motion are

H pi ri
(5.4)
=
=
pi mi
t
and

p
H
2 r
= i = mi 2i (5.5)
ri
t
t
From Equations 5.3 and 5.5, we obtain
U
2 ri
= m 2 (5.6)
ri
t
Using Newtons second law, Equation 5.6 yields

H
U
Fi = = (5.7)
ri
ri
From Equation 5.7, the interatomic force can be computed from the negative gradient of the
interatomic potential energy function U.
The acceleration of each atom is computed using the interaction forces obtained using
Equation 5.7 and Newtons second law, Equation 5.1. The acceleration is integrated with
respect to time once and twice to find the velocity and displacement, respectively, using
the prescribed time step size. This process is applied to all molecules. These calculations
are repeated for the next time until the specified final time arrives.
There are many different mathematical expressions for the interatomic potential energy
function depending on the bonding nature of the molecules. The details of the interatomic
potential energy function are discussed in the next section.
5.3Time Integration Technique

As the interatomic potential energy function is expressed in terms of the relative positions among atoms, it is necessary to update the new positions of atoms continuously. As
147
Molecular Dynamics
a result, once the acceleration of each atom is computed from the interatomic forces of the
neighboring atoms, they should be integrated once and twice, respectively, to determine
the velocity and position. In that aspect, a reliable and accurate time integration scheme is
necessary.
For numerical time integration, the finite difference method is used. One of the simplest
finite difference schemes is the forward difference:
dr(t + t) r(t + t) r(t)
(5.8)
dt
t
This is computationally efficient but lacks accuracy. The error is linearly proportional to
t. This means that if the t is reduced by half, the error is also reduced by half. For
Equation 5.8 as well as subsequent equations, the superimposed arrow symbol for a vector
notation is omitted for simplicity when obvious.
Another finite difference method is the central difference scheme, which is given by the
following equations:
a(t) =
F(t)
(5.9)
m
t
t
v t + = v t + a(t)t (5.10)
2
2
t
r(t + t) = tv t + + r(t) (5.11)
The velocity at time t is computed from the average value as follows:
v(t) =
t
1 t
v t + + v t (5.12)
2
2
2
The advantage of the central difference technique is computational efficiency, and it has
second-order accuracy; the forward difference method has first-order accuracy.
However, in MD, the motion of a particle is computed over a large number of time steps,
and it turns out that the numerical errors associated with these two methods (central and
forward difference) are too big to tolerate. Therefore, it is necessary to use a more complicated scheme for solving the differential equations arising from Newtons second law
and Hamiltonian dynamics. The scheme adopted here is known as the Gear predictorcorrector method [2].
The Gear method is composed of three steps: prediction, evaluation, and correction.
This method is presented because it is important to understand the process of error accumulation with the simulation time step in the MD simulation runs. The equations that follow are given for the ith atom to calculate the trajectories, such as the position ri, velocity
ri , and acceleration ri at time t + t.
148
5.3.1Prediction Step
An atoms position at time t + t is predicted using a fifth-order Taylor series based on the
positions and their derivatives at time t.

riP (t + t) = ri (t) + ri (t)t + ri (t)
(t)5
(t)2
(t)3
(t)4
+
ri (t)
+ ri( 4) (t)
+ ri( 5) (t)
(5.13)
2!
3!
4!
5!
riP (t + t) = ri (t) + ri (t)t +

ri (t)
(t)4
(t)2
(t)3
+ ri( 4) (t)
+ ri( 5) (t)
(5.14)
2!
3!
4!
riP (t + t) = ri (t) +
ri (t)t + ri( 4) (t)
(t)3
(t)2
(5.15)
+ ri( 5) (t)
2!
3!
ri P (t + t) =
ri (t) + ri( 4) (t)t + ri( 5)
(t)2
(5.16)
2!
ri(P4) (t + t) = ri( 4) (t) + ri( 5) (t)t (5.17)
ri(P5) (t + t) = ri( 5) (t) (5.18)
where the superimposed dot or superscript (n) denotes temporal derivatives of the corresponding order. Furthermore, the superscript P stands for the predicted value.
5.3.2Evaluation
The interatomic force Fi on each atom is calculated at time t + t using the predicted positions. The evaluation of the forces is the most time-consuming process. The force on each
atom is given by the potential energy function as

Fi =
j i
U (rij )
rij
(5.19)
where U(rij) is a continuous potential energy function that acts between atoms i and j.
5.3.3Correction
The predicted positions and their derivatives are corrected using the discrepancy ri
between the predicted acceleration and that given by the evaluated force Fi as follows:

ri = ri (t + t) riP (t + t) (5.20)
where
ri (t + t) is the acceleration using Equation 5.19 and Newtons second law.
149
Molecular Dynamics
In Gears algorithms for a second-order differential equation of motion, this difference

term is used to correct all predicted positions and their derivatives as follows:

riC (t + t) = riP (t + t) +
3
ri (t)2 (5.21)
32
riC (t + t) = riP (t + t) +
251
ri (t) (5.22)
720
riC (t + t) = riP (t + t) + ri (5.23)
11 ri
(5.24)
6 t
ri C (t + t) =
ri P (t + t) +
ri(C4) (t + t) = ri(P4) (t + t) +
2 ri
(5.25)
(t)2
ri(C5) (t + t) = ri(P5) (t + t) +
ri
(5.26)
(t)3
5.4Interatomic Potential Energy Function

The interatomic potential energy function is a semiempirical expression. Depending on
the bonding nature among atoms, many different potential functions have been developed. In this section, some of the most commonly used potential functions are discussed.
5.4.1Lennard-Jones Potential
The most common energy potential function is the Lennard-Jones (LJ) model [4,5] for gas:

n m
U (r ) = (5.27)
r
r
where
m
n n n m
=
(5.28)
n m m
In addition, and are constants depending on the atoms or molecules, and m and n are
positive integers.
150
5.4.2Morse Potential
The Morse function describes the potential energy (rij) of two atoms i and j with a distance rij as follows:

2 ( rij ro )
(r r )
(rij ) = D e
2 e ij o (5.29)
in which D and are constants with dimensions of energy and reciprocal distance, respectively. The ro is the equilibrium distance of the two atoms with (ro) = D as shown in
Equation 5.29. The constants in Equation 5.29 are provided in Reference 6 for various pure
metals. The force between the two atoms is computed from

F(rij ) =
nr (5.30)
rij
in which nr is the unit position vector between the two atoms. The forces among N atoms
in a given system must be in equilibrium using Equation 5.30.
5.4.3Abell-Tersoff-Brenner Potential
A key feature of the Abell-Tersoff-Brenner (A-T-B) type potential is that short-range bonding interactions tend to participate readily in reactions so that chemical bonds can form
and break during the MD simulations. Tersoff developed, for the first time, an analytical potential energy function based on Abells functional forms that realistically describe
bonding in silicon for a large number of solid-state structures [79]. After that, Brenner
presented a practical empirical many-body potential energy function for the hydrocarbon
molecules as well as graphite and diamond lattices [10].
However, even though the original Brenner potential has been effective for reacting
hydrocarbon molecules, it is not efficient for purely carbon atom systems because its
potential function has a term associated with hydrogen atoms. For that reason, when the
Brenner potential was coupled with the Tersoff potential, the terms associated with hydrogen atoms in the original Brenner potential were removed. Thus, the resultant modified
A-T-B type potential function was developed. The A-T-B type potential energy as a function of the atomic coordinates is given by
U=
ui =
1
2
u (5.31)
i
[V (r ) B V (r )] (5.32)
R
ij
ij
ij
j i
where the repulsive and attractive pair terms VR and VA are given by

VR (rij ) = fij (rij )Dij( e ) /(Sij 1) exp 2Sij ij r Rij( e )
)) (5.33)
151
Molecular Dynamics
))
VA (rij ) = fij (rij )Dij( e )Sij/(Sij 1) exp 2/Sij ij r Rij( e ) (5.34)
where D, S, , and R are empirical parameters, and the function fij(r), which restricts the
pair potential to nearest-neighbor atoms, is given by
fij (r ) =
r > Rij(1)
1,
(r R(1) )
1 + cos ( 2 ) ij(1) /2 ,
Rij Rij
0,
((
Rij(1) < r < Rij( 2 )
(5.35)
r > Rij( 2 )
))
) (
Bij = 1 +
Gi (ijk ) fik (rik ) exp ijk rij Rij( e ) rik Rik( e )
k (i , j)
c2
c02
G(ijk ) = a0 1 + 02 2
d0
d0 + (1 + cos )2
(5.36)
(5.37)
where is the angle between atoms i j and i k bonds, r is the distance between atoms,
and other constant variables are constant. These equations require a total of 11 fitting
parameters for carbon atoms. The values for these parameters are given in Table 5.1, where
the subscripts cc means that those parameters are used just for the carbon-carbon bonding
network without other elements, such as hydrogen.
The A-T-B potential energy function for two carbon atoms is plotted as a continuous
function in Figure 5.1 using Equations 5.31 through 5.37. In the equilibrium state, two
carbon atoms are separated by the equilibrium separation distance r0 ~ 1.42 , which is
called the effective carbon bond length. At equilibrium, the potential energy of the system
becomes minimal. Consequently, any equilibrated carbon nanotube (CNT) structure will
be formed only if the potential energy per each carbon atom can attain a minimum-energy
value as they approach each other. This minimum energy is defined as the bond energy.
The bond energy for the CNT is known to be 2.5 eV, which is consistent with the carboncarbon tight bonding overlap energy accepted from numerous prior experiments [11,12].
TABLE 5.1
Optimized Parameters for the A-T-B Type Potential
Parameter
Value
Parameter
Value
Parameter
Value
Rcc( e )
1.42
ccc
0.0
d02
3.52
6.0 eV
( 1)
cc
1.7
(2)
cc
2.0
0.00020813
3302
(e)
cc
cc
Scc
cc
2.1
1.22
0.5
R
a0
c02
Source: From D. W. Brenner, Physical Review B, vol. 42, 1990, pp. 94589471.
152
Potential energy (eV/atom)
3
2.5
2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
3
Bond length ~ 1.42
Bond energy ~ 2.5 eV
1.25
1.5
1.75
2.25
2.5
Interatomic separation, r ()
FIGURE 5.1
Potential energy versus interatomic separation of A-T-B potential from MD simulation.
The simulated A-T-B potential energy as a function of interatomic separation distance can
be seen in Figure 5.1.
Although these arguments are considered for only two carbon atoms, similar arguments
can be applied to the entire bonding networks consisting of many carbon atoms, as in
CNTs. Because any system tends toward the minimum potential energy, carbon atoms
exert forces on one another so that they can adjust themselves to the bottom of the potential energy as shown in Figure 5.1.
Figure 5.1 shows that the A-T-B potential curve is asymmetric. The A-T-B potential curve
is broader in the region r > r0. Thus, when two paired atoms vibrate about their equilibrium positions at a given temperature (i.e., stretching and compressing the bond according
to kinetic molecular theory); the atoms will spend more time in the region r > r0, that is,
more time in stretching the bond than in compressing the bond with respect to the equilibrium bond length r0. However, when the temperature is approximately room temperature
3
(~300 K), the mean vibration kinetic energy is 0.039 eV, which is computed from kT, and
2
the atoms will vibrate near the bottom of the potential well. In this case, the A-T-B curve
can be taken as approximately symmetric; consequently, the thermal vibrational motion of
two atoms can be approximated as a simple harmonic oscillator motion [13,14].
It is interesting to discuss the attractive force region because it gives us a hint for studying the deformation behaviors and Youngs modulus of CNTs under an external tensile
load. For the range of 1.42 < r < 2.0 , the paired carbon atoms experience a significant
force change when their separation reaches around 1.72 ; namely, the interaction force
derived from the potential energy function suddenly changes from the gentle positive
slope to the steep negative slope at around 1.72 when it goes far away from the equilibrium position, as shown in Figure 5.2.
Figure 5.2 shows the magnified parts of the attractive force region from Figure 5.1, and
U
.
the force function is evaluated as the negative gradient of the potential energy, F =
r
In the figure, the positive y axis represents the attractive force region, while negative y
axis represents the repulsive force region. At the equilibrium position r0, the interaction
force between two carbon atoms becomes zero, and then the attractive force increases
153
Molecular Dynamics
Potential energy and force
Force function
3
2
dF
1
0
dr
1
2
Potential function
3
4
5
1.3
1.4
1.5
1.6
1.7
1.8
1.9
2.1
Separation distance, r ()
FIGURE 5.2
Idealized interaction force and potential function of one pair of carbon atoms.
until the separation distance reaches around 1.72 . However, as it goes away from that
point, the attractive force decreases until the separation distance reaches 2.0 . For separations larger than about 2.0 , the interaction force between two atoms becomes essentially
zero; this is a reason that the A-T-B potential is called a short-range interaction potential,
and the interacting force becomes a strong repulsive force for r 1.42 . Thus, from this
perspective, it can be anticipated that CNTs will show strong resistance against external
compressive loads.
If a CNT is subjected to an external tensile load, the system is in a nonequilibrium state.
The displacement of neighboring atoms due to the tensile force results in a net attractive
force; subsequently, this net force will be balanced by a portion of the applied force acting
on these atoms. The elastic modulus can be evaluated by the measurement of the balance
mentioned [14]. Thus, the elastic modulus E depends on the gradient of the interaction
force F versus separation distance r at the equilibrium position r0 as shown in Figure 5.2
and can be written

E=
1
r0
F
(5.38)
r r = r0
Based on Equation 5.38, the computer-generated elastic modulus is 16.74 TPa when r0 =
1.42 . This elastic modulus is unrealistically high due to the consideration of the idealized one pair of carbon atoms. That is, there is no special effect of many-body interaction
forces. Therefore, the realistic elastic modulus of a CNT system can be found only when
the many-particle interaction forces are considered, including the bonding angle effect
represented in Equation 5.37 within the previously mentioned A-T-B potential function.
Meanwhile, it is geometrically impossible for every pair of carbon atoms in a CNT system to be separated corresponding to the minimum energy. Moreover, the thermal energy,
which is manifested as interactions among atoms, constantly knocks atom pairs away
from the minimum-energy separation, and the overall behavior of the collection of atoms,
therefore, may be complex. Thus, it is expected that the actual elastic modulus of CNTs
154
will converge to a realistically accepted value, the order of approximately 1 TPa, when all
the facts mentioned are considered.
5.4.4Embedded Atom Potential
The embedded atom method describes the total energy of a system of metallic atoms:

Etot =
E (
ee
h ,i
)+
1
2
(r ) (5.39)
ij
ij
i, j
Equation 5.39 states that the total energy of the system of atoms is the sum of the embedding energy and the core-core repulsions.
Specifically, the embedding energy term treats each atom i as an impurity and is a calculation of the amount of energy required to embed atom i into the electron density of its
actual location. The electron density of its actual location is determined by

h , i =
(r ) (5.40)
a
j
ij
j i
Equation 5.40 states that the electron density at the position of atom i is a sum of the densities contributed by all other atoms in the sample within a given distance rij. Atoms beyond
rij have no effect on atom i.
The core-core repulsion term in Equation 5.39 is a calculation of the core-core repulsions
ij of each atom i paired with all other atoms in the sample within a given distance rij.
Similar to the electron density, atoms beyond rij have no core-core repulsion with atom i.
The use of one-half in Equation 5.39 is necessary to avoid counting the same repulsion two
times during summation over the entire sample.
5.5Molecular Mechanics Formulation

The molecular mechanics used for discrete atoms are described next: The force acting on
an atom is computed from

F(rij ) =
nr (5.41)
rij
in which nr is the unit position vector between two atoms. All resultant forces among N
atoms along with any external load in a given system must be in equilibrium.
Consider any two atoms located at positions i and j, called atom i and atom j, as shown
in Figure 5.3. In that figure, solid circles indicate present positions of the two atoms under
equilibrium before applying external loads. As external loads are applied to the system,
the two atoms move to the positions denoted by circles, called atoms i* and j*, as seen in
Figure 5.3. Then, displacement vectors from initial positions to final positions of the atoms
155
Molecular Dynamics
Atom #j*
uj
Rij
Atom #i*
ui
Atom #j
rij
y
Atom #i
x
FIGURE 5.3
Relative positions of two atoms before and after movement.
of the two atoms at the initial

are expressed as ui and u j , respectively. The position vectors
and final positions, respectively, are denoted by rij and Rij , as shown in the figure. The displacement vectors and position vectors are related as

Rij = rij + u j ui = rij + uij (5.42)

where uij = u j ui is the relative displacement vector of the two atoms.
The force between the two atoms i and j at their new equilibrium positions is expressed
as Equation 5.43:

Fij (Rij ) = Fij (Rij )nR (5.43)
in which Rij is the distance

between the two displaced atoms. nR is the directional unit vec
tor along the vector Rij, and it is expressed as follows:
rij + uij
nR =
(5.44)
Rij
Equation 5.44 is substituted into Equation 5.43, resulting in the following expression:

uij
rij
Fij (Rij ) = Fij (Rij )
+ Fij (Rij )
(5.45)
Rij
Rij
Applying Equation 5.45 to atoms i and j yields a matrix expression as shown in Equation
5.46:
156
k
0
0
k
0
0
0
k
0
0
k
0
0
0
k
0
0
k
k
0
0
k
0
0
0
k
0
0
k
0
0
0
k
0
0
k
uix Fix

uiy Fiy

uiz Fiz
=
u jx Fjx (5.46)

u jy Fjy

u jz Fjz
where

k=
Fix = Fjx =
Fiy = Fjy =
Fiz = Fjz =
Fij (Rij )
Rij
Fij (Rij )
Rij
Fij (Rij )
Rij
Fij (Rij )
Rij
(5.47)
( xi x j ) (5.48)
( y i y j ) (5.49)
( zi z j ) (5.50)
Here, (xi, yi, zi) and (xj, yj, zj) are the coordinates of the atoms in Figure 5.3.
The matrix expression, Equation 5.46, is computed for all atoms that interact with one
another and is assembled into the system matrix [K] consisting of all atoms displacements
{u}. The resultant system matrix equation is
[M]{} + [K]{u} = {F} (5.51)
where matrix [M] is the diagonal system mass matrix constructed from atomic masses,
vector {F} is the system force vector resulting from atomic interacting forces and additional
external loads, and the superimposed dot indicates a temporal derivative. In the
of
system
equations, the stiffness matrix [K] and force vector {F} are nonlinear because Rij and Fij (Rij )
are not known a priori.
5.6Application to Carbon Nanotubes: Elastic Modulus

Understanding the basic geometrical structures of CNTs is a preliminary essential step
to start MD simulations with the initial geometrical configurations of the given CNT systems. For example, it should be a structure that is properly energy minimized. This section describes the basic structural properties of CNTs, explains the synthesis possibility of
the heterojunction-type nanotubes such as the bamboo shape nanotube (BSNT), and then
157
Molecular Dynamics
presents the resultant initial geometrical models of single wall nanotubes (SWNTs) and
bamboo shape single wall nanotubes (BSWNTs) [15,16].
5.6.1Basic Structures of Carbon Nanotubes
Many experiments have confirmed that CNTs are cylindrical tubular shell structures
based on a benzene-type hexagonal lattice of carbon atoms. The cylindrical tube can be
tens of microns long, and the ends of nanotubes are capped with half-dome-shaped
fullerene molecules.
The ideal single-wall carbon nanotube (SWCNT) can be thought of as the fundamental
structure that forms the building blocks of both multiwall carbon nanotubes (MWCNTs)
and an ordered array of SWCNTs called nanoropes. It is known that Youngs modulus of
MWCNTs is not significantly different from SWCNTs because the high modulus mainly
results from the carbon-carbon interaction bonds within the individual layers. In addition,
there is only a weak van der Waals interaction force between two graphite layers because
the average interlayer spacing distance of 3.4 is larger than the maximum carbon-carbon
bond length of about 2.0 . Thus, once understanding of the SWCNT structures is obtained,
the consideration of other CNT structures is much easier. Consequently, the outline for the
basic structural characteristics of CNTs focuses on the SWCNT.
It has been found that three types of CNTs are possible: armchair, zigzag, and chiral SWNTs. These depend on how the two-dimensional graphene layer is rolled up. To
understand the structural features of SWCNTs, Figure 5.4 shows the two-dimensional
graphene layer with atoms labeled using (n, m) notation, where n and m are integers and
typically
In Figure 5.4, the
n m; this notation was suggested by Dresselhaus et al. [17].
vector C is called the chiral vector of the nanotube and is defined as C = (m, n) in the twodimensional hexagonal graphene sheet. The magnitude of these unit vectors is known
to be 0.246 nm.
The physical properties of CNTs are significantly determined by their diameter and chiral angle, both of which depend on n and m. When the graphene layer is rolled up to form
the cylindrical part of the nanotubes, the ends of the chiral vector meet each other. Thus,
the chiral vector forms the circumference of the nanotubes circular cross section, and different values of n and m therefore lead to different nanotube structures. The diameter of
nanotubes dt is simply the length of the chiral vector divided by ; thus, it can be expressed
as a function of n and m:
(0,2)
(0,1)
(0,0)
(1,2)
(1,1)
(1,0)
(2,2)
(2,1)
(2,0)
(3,2)
(3,1)
(3,0)
FIGURE 5.4
Graphene layer with carbon atoms identified by the notation C = (m, n).
C = (3,2)
158
dt =
0.246 2
(n + n m + m2 )1/2 (nm) (5.52)
The chiral angle can be calculated by
31/2 m
= sin 1
2
2 1/2 (5.53)
2(n + n m + m )
According to prior research, armchair nanotubes are formed when n = m and the chiral
angle is 30. Zigzag nanotubes are formed when either n or m is zero with the chiral angle
0. All other nanotubes are known as chiral nanotubes, with the chiral angle between 0
and 30 [18]. In particular, the chiral vector (n, m) can determine whether a SWCNT can be
either a metal or a semiconductor. For example, a metallic nanotube can be obtained when
the difference n m is a multiple of three. If the difference is not a multiple of three, a semiconductor nanotube can be obtained. These characteristics are important for studying the
electronic properties of SWCNTs.
5.6.2Simulation Time Step
The simulation is propagated through time at intervals of t by iteration using Gears
method at every time step. Based on the energy conservation principles, typically Gears
algorithm is about one order higher in accuracy than Verlets. A reasonable simulation
time step t is estimated as 0.01 ps using the relation between the bond length r0 and the
bond energy . The method to find the reasonable time step size is presented next.
For the isolated system, the kinetic energy of an atom should be about the same as the
potential energy of the atom. Let us assume that one atom can move within the bond
length r0 during the time step t. Then, the kinetic energy Ek and the velocity v of the atom
are written, respectively, by

Ek =
1
mv 2 = (5.54)
2
v=
r0
(5.55)
t
From Equations 5.54 and 5.55, the maximum possible t value is computed as

t = r0 m/2 0.02 ps
(5.56)
where m is the mass of the carbon atom, and r0 and are the carbon bond length and CC
tight bonding overlap energy, respectively. Based on Equation 5.56, the time step 0.01 ps is
considered to have a reasonable value.
159
Molecular Dynamics
5.6.3Freestanding Thermal Vibration Method

This section describes the method to calculate Youngs modulus of CNTs by introducing the
freestanding room temperature vibration method presented in Reference 19. According to
the kinetic molecular theory, carbon atoms vibrate about their equilibrium positions with a
3
mean vibration kinetic energy that increases with the temperature as kT. From this fun2
damental principle, Youngs modulus of SWCNTs is estimated as Y = 1.25 0.35 / + 0.45
(TPa) by observing SWCNTs freestanding room temperature vibrations in a transmission
electron microscope (TEM).
The formula to calculate Youngs modulus was derived from the relationship between
the motion of a vibrating clamped cylindrical cantilever rod governed by the classical
fourth-order wave equation and the quantum mechanicsstatistical probability theory
given by the Boltzmann factor. The resultant relationship between Youngs modulus Y;
length L; inner and outer tube radii b and a, respectively, that form a cylindrical CNT wall;
and the root mean square (rms) displacement , which is represented by the vibration
amplitude at the tip of a CNT at a temperature T, can be expressed by

2 = 0.8486
L3 kT
(5.57)
YWG(W 2 + G 2 )
where W is the SWCNT width (diameter), G = (a b) is the graphite interlayer spacing of

0.34 nm, and k is the Boltzmann constant.
Equation 5.57 is used to predict the Youngs modulus as follows:

Y = 0.8486
L3 kT
(5.58)
2WG(W 2 + G 2 )
From this equation, notice that Youngs modulus depends strongly on the rms displacement at the tip of the SWNT for a given length and width W of the CNT and the given
ambient temperature T.
5.6.4Equilibrium and Vibration Motion of CNTs
The system with atoms randomly assigned initial velocities taken from a Gaussian distribution must be allowed to reach equilibrium before any reliable measurements can
be taken. Equilibrium is monitored using the Boltzmann H function that represents the
instantaneous velocity distribution: the distribution at one moment during the simulation.
The H function for the x component can be written as
H x (t) =
f (v ) ln f (v ) dv
x
(5.59)
where f(vx)dvx is Maxwells velocity distribution, which is the fraction of atoms having
velocities between vx and vx + dvx, and can be evaluated analytically, giving
160
f ( vx )dvx =
mvx2
m
exp
dv (5.60)
2 kT x
2 kT
Based on Boltzmanns H theorem, the H function decreases from an initial value to its
value given by the Maxwell velocity distribution Equation 5.60 until the system reaches
equilibrium. Once equilibrium is attained, the H function shows steady-stable fluctuations with respect to an average value. Figure 5.5a and b show the simulated H function
for the SWCNT and bamboo-structured single-wall carbon nanotube (BSCNT) models
under study from the initial to the equilibrium states, respectively [15]. Both H functions
converge to some values, and they stay close to the values. It is an indication that the equilibrium of both systems is achieved. In Figure 5.5, the H function of the BSCNT model
displays more fluctuations than that of the SWCNT model within the same range of time
steps between 200 and 400. This indicates that the fluctuation of carbon atom velocities
in the BSCNT model is larger than in the SWCNT model. As a result, it may be expected
that Youngs modulus of BSCNT should be lower than for the SWCNT under the same
temperatures and constraints because the atomic rms displacement of the BSCNT model
is expected to be larger than SWCNT by the relation between Y and in Equation 5.58.
Figure 5.6 shows the equilibrated top configuration of the armchair (4, 4) SWCNT model
made of 185 carbon atoms after the equilibrium MD simulation. The configuration is captured when the potential energy per carbon atom of the SWCNT model reaches the minimum value and atoms are vibrating about the equilibrium state. In the center of Figure 5.6,
the local topology of the cap region of the SWCNT model shows a pentagon-shape defects
ring pole. This pentagon topology becomes a seed for the stable cap formation mechanism
0
H function
200
400
600
800
1000
800
1000
2
3
4
5
6
Time steps
(a)
0
1
200
400
600
H function
2
3
4
5
6
7
(b)
Time steps
FIGURE 5.5
H function of SWNT and BSNT model from MD simulation: (a) SWCNT models H function with time steps and
(b) BSCNT models H function with time steps.
161
Molecular Dynamics
Observed atomic vibration
FIGURE 5.6
Top view of SWCNT model under equilibrium.
Knots
FIGURE 5.7
The top view of BSCNT model under equilibrium.
of the SWCNT model [20]. We calculate the Youngs modulus of the SWCNT model using
the thermal vibration method because it consists of the tip of the SWNT model. A few
atomic vibration motions along the circumference of the SWCNT model can be seen in
Figure 5.6.
The equilibrated configurations for the BSCNT, which consists of two armchair (5, 5)
SWCNTs and a total of 210 carbon atoms, is shown in Figure 5.7. The equilibrium process
for the BSCNT model is similar to the previous SWCNT model. Figure 5.7 shows the equilibrated top view. The equilibrated BSCNT model shows somewhat different features with
respect to the SWCNT model. The remarkable local topology changes (e.g., knot-shaped
defects that look like protuberant lumps at a point from which a stem or branch grows)
are created at the heterojunction region along the circumference of the BSCNT. Atomic
162
vibrational motions are also observed, but the magnitudes of those are more conspicuous
than for the SWCNT model. In both cases, the central pentagon ring of the end tip of both
models shows visible vibration motions because it is difficult to catch the rms displacements of those pentagon rings atoms by eye on the atomic scale. Thus, to find the atomic
scales displacements of end tips of both models, computer algorithms were developed for
calculating Youngs modulus.
5.6.5Elastic Modulus of CNTs under Equilibrium
The two models considered previously were tested under equilibrium MD simulation at
room temperature (300 K), and then the elastic moduli of SWCNT and BSCNT were evaluated based on Equation 5.58. The numerical values were compared with the theoretical
and experimental results presented from other studies.
Figure 5.8a and b show histograms that demonstrate the spread in the evaluated Youngs
modulus values for the SWCNT and BSCNT by running 100 simulations each, and measured objects were the carbon atoms of the pentagon ring pole at the tip in both models.
The MD code was designed to calculate the rms displacements of the tip carbon atoms
3
when the instantaneous kinetic energy per atom of both models reached 0.039 eV ~ kT
2
at the given room temperature of 300 K. Then, Youngs modulus was calculated using
Equation 5.58. In both cases, the use of 100 simulations was considered sufficient to reduce
statistical errors due to the random initial velocities for both models.
30
25
Occurence %
20
15
10
1.
0
0.
75
1
.0
1
.2
5
1.
15
1
.5
1.
5
1.
75
1.
75
2
.0
2.
0
2.
25
2.
25
2
.5
2.
5
2.
75
2.
75
3
.0
3.
0
3.
25
(a)
Elastic modulus (GPa)

30
25
Occurence %
20
15
10
0.
0.
4
0
.
45
45
0
.5
0.
5
0.
55
0.
55
0
0. .6
6
0.
65
0.
65
0. 07
7
0.
75
0.
75
0
0. .8
8
0.
85
0.
85
0
.9
(b)
Elastic modulus (GPa)
FIGURE 5.8
Histograms of Youngs modulus from MD for (a) SWCNT and (b) BSCNT.
163
Molecular Dynamics
The observed distributions of Figure 5.8a and b are fairly asymmetrical about the mean
values, displaying a tail extending to higher values. These asymmetrical distributions may
affect the ability to interpret the reliability of a mean elastic modulus without a normal
distribution assumption. Thus, under the equilibrium MD simulation, there are essential
embedded random errors during the simulation process due to the time evolution method
and different initial conditions on each independent simulation, such as random initial
velocities. Although embedded time evolution errors in MD code cannot be completely
removed, a normal distribution assumption about the randomly distributed elastic moduli
evaluating from MD results is feasible due to the fact that the assigned initial velocities are
normally distributed by the Maxwell-Boltzman velocity distribution.
According to the assumption that the measured Youngs modulus shows a normal distribution about a mean value, normal probability density functions for the SWCNT and
BSCNT models can be computed. The normal probability density function for a random
variable x, which represents the Youngs modulus value, with mean and standard deviation is given by

( x )2
1
f ( x) =
exp
, where x + (5.61)
2
2
(2 )
The parameters used for this normal probability density function for the computed
Youngs moduli are shown in Table 5.2. The data were collected from the equilibrium MD
simulation results of the SWCNT and BSCNT models.
Figure 5.9 shows the normal probability density functions of both SWCNT and BSCNT
models plotted using Equation 5.61 with the parameters of Table 5.2. The mean value of
Youngs modulus for the SWCNT model was 1.42 TPa with a standard deviation of 0.34
TPa. Based on these data, the expected mean Youngs modulus for the real population of
SWCNTs can be predicted as 1.42 0.23 TPa with 50% confidence intervals. These values
are in good agreement with the theoretical and experimental Youngs modulus values
reported by others, as shown in Tables 5.3 and 5.4 [19,2127].
In particular, considering about 48% of the elastic values for the SWCNT model, which is
distributed within the range 1.0 to 1.5 TPa as shown in Figure 5.8a, and comparing with the
previously mentioned mean Youngs modulus of SWCNT 1.25 TPa as evaluated by the same
thermal vibration method presented by Krishnan et al. [19] and obtained from 27 SWCNTs,
the consistency of both results are comparable. In their method, the tubes length and tip
vibration amplitudes were estimated directly from the digital micrographs (TEM images).
From these consistent results, the MD simulation is reasonable for use for evaluating Youngs
modulus of the SWCNT model, and it may be feasible to simulate the BSCNT model.
TABLE 5.2
Statistical Data Measured from MDSS Simulations of SWCNT/
BSCNT Models
Parameter
Number of simulations
Mean Youngs modulus (TPa)
Standard deviation (TPa)
Maximum Y value (TPa)
Minimum Y value (TPa)
SWCNT
BSCNT
100
1.4243
0.342
3.4202
0.7402
100
0.6041
0.1002
0.9241
0.4046
164
1012
Normal probability density
3.5
BSNT
3
2.5
2
1.5
SWNT
1
0.5
0
0
0.5
1.5
(Pa)
2.5
3
1012
FIGURE 5.9
Normal probability density functions of elastic moduli (Y value) for SWCNT and BSCNT models.
TABLE 5.3
Various Theoretical Calculations for Youngs Modulus
Method
MD Using TersoffEmpirical Force

Nonorthogonal
Brenner Potential [21] Constant Method [22] Tight Binding [23]
Wall thickness
Elastic modulus
0.06 nm
SWCNT: 5.5 TPa
0.34 nm
SWCNT: 0.97 TPa
0.34 nm
SWCNT: 1.2 TPa
Ab Initio DFT [24]

0.34 nm
SWCNT rope: 0.8 TPa
MWCNT: 0.95 TPa
TABLE 5.4
Various Experimental Studies for Youngs Modulus
Method
Wall thickness
Elastic modulus
Thermal
Vibration [25]
Restoring Force
ofBending [26]
Thermal
Vibration [19]
0.06 nm
SWCNT:
1.81.4 TPa
0.34 nm
SWCNT:
1.280.59 TPa
0.34 nm
SWCNT: 1.3 to
0.4/+0.6 TPa
Deflection Forces [27]

0.34 nm
SWCNT: 1.0 TPa
The equilibrium MD simulation results of the SWCNT and BSCNT models show two
important features. One is that although the SWCNT MD results and prior studies were
consistent with each other within the mean error, the elastic modulus of the mean value
1.42 TPa of the SWCNT model calculated from MD simulation shows a somewhat higher
value than the elastic modulus of the mean 1.25 TPa used with the same thermal vibration
method reported by Krishnan et al. [19]. A second is that the calculated elastic modulus of the BSCNT model shows an obvious difference with respect to the SWCNT model.
The mean elastic modulus of the BSCNT model is calculated as 0.604 TPa with a standard deviation of 0.1 TPa. Consequently, if BSCNTs can be synthesized in practice, the
expected mean elastic modulus for a real BSCNT population would be predicted to be
0.604 0.07 TPa with a 50% confidence level. Contrary to the general belief that expects
Molecular Dynamics
165
high elastic modulus values due to the fact that the BSCNT has an internal support structure, the results show a significantly lower elastic modulus than that of the SWCNT model,
but still show a strong elastic modulus compared with other materials, such as hardened
steel, on the order of 210 GPa. Unfortunately, so far there are no available data for the heterogeneous CNTs elastic moduli to be compared with that of the BSCNT model.
One reason for a somewhat higher mean elastic value of the SWCNT model calculated
from MD simulation can be understood from the parameter relationships of Equation
5.58. In this equation, the elastic modulus is inversely proportional to the width of the
CNTs W when the other parameters are held the same. This means that the elastic modulus of CNTs has dependence on the diameter of the tube. Krishnan et al. used nanotube
widths as a range of 1.01.5 nm (1015 ) to calculate Youngs modulus using Equation
5.58, while the present SWCNT and BSCNT model diameter used in MD simulation was
just 6.78. Therefore, it is natural to expect that the measured Youngs modulus of the
SWCNT model is higher than Krishnan et al.s Youngs modulus from the fundamental
relations mentioned.
There is another possible factor in explaining the difference in the elastic moduli
between our MD results and the prior experimental results reported by others. In most
of the prior experiments, the CNTs not only were made of carbon atoms but also had
impurities due to the limit of the SWCNT synthesis process in practice. Typically, it is
known that impurities contained in CNTs lower their mechanical strength. Thus, it is a
reasonable inference that the Youngs modulus evaluated from the MD simulation using
the SWCNT model made of pure carbon atoms may have higher strength than real CNTs.
On the other hand, why is the elastic modulus of the SWCNT model roughly twice as high
than BSCNT model? Earlier studies reported that the Youngs modulus of CNTs probably
depends on the presence of structural imperfections, such as the nesting of tubular cylinders,
which can create a joint or knuckle, thereby weakening the tube, and structural defects
with the pentagon (5)heptagon (7) rings. The defects that appeared along the circumference
of the BSCNT model can prove the arguments presented in Reference 28: When heterogeneous CNTs are formed, the 57 defect rings are induced, and the induced defects absorb
partial energy from the total energy of the system because it requires defect formation energy.
Thus, the defects at the junction regions of the BSCNT model may absorb energy, with
the other regions energy per atom increasing for balance so that the total energy of an
isolated system is constant. As a result, the energy needed for the atomic vibration in the
BSCNT tips would increase, and then the elastic modulus of the BSCNT model would
be lower than that of the SWCNT model based on the relationship of Equation 5.58. This
analytical interpretation is also consistent with the results of Figure 5.5b, as mentioned
previously. However, it is not clear which energy (i.e., kinetic or potential energy) has an
important role for that mechanism.
5.6.6Comparative Results of Equilibrium and Nonequilibrium Simulations
In this section, a relative Youngs modulus ratio concept for the SWCNT and BSCNT models is introduced with the force-strain diagram that is extracted from the nonequilibrium
MD simulation, which is then compared with the Youngs modulus values calculated from
the previous equilibrium MD simulation.
By conventional continuum mechanics, the ultimate tensile strength and Youngs modulus for a bulk material can be determined from the force-displacement data under the external tensile loading test. The ultimate tensile strength is measured as the maximum stress
prior to fracture. To obtain the Youngs modulus from the collected data, a second-degree
166
polynomial function can be fitted to the applied force-displacement diagram, with the
modulus measured as the slope within the small strain limit range [29].
However, because it is difficult to estimate a definite cross-sectional area of CNTs due to
the nanoscale model configuration, the previous conventional method is not appropriate
for calculating a realistic Youngs modulus directly from the applied force-strain diagram
of nonequilibrium MD simulation. For that reason, a relative Youngs modulus ratio, in
which two materials are assumed to have the same cross-sectional area, is introduced to
remove the effect of cross-sectional area in computing the Youngs modulus.
To collect data for the analysis of atomic-scale tensile deformations of our SWCNT and
BSCNT models, the tensile forces were applied to both models to move the end atoms with
constant strain rates (~10 2/ps), and the strain per atom at every force was measured. Once
the force-strain per atom data are collected from the simulation, a relationship between the
applied force and strain can be plotted as the force-strain diagram.
The tensile force-strain diagram for the SWCNT and BSCNT models obtained from the
nonequilibrium MD simulation is plotted in Figure 5.10. The applied tensile forces per
atom of the SWCNT and BSCNT models were recorded on the y axis, and the average
strains of carbon atoms in the location right below the forced atoms were recorded on
the x axis, respectively. Overall, the observed force-strain diagrams showed a significant
nonlinear relationship. In particular, observe that the strain of the BSCNT model was more
linearly increased than that of the SWCNT model as the applied force increased. From this
observation, it is possible to expect that the Youngs modulus of the BSCNT model would
be less than the modulus of the SWCNT model.
Force per atom (nM)
60
50
40
30
20
10
0
(a)
6
Milli-strain
10
12
Force per atom (nM)
50
40
30
20
10
0
(b)
8
10
Milli-strain
12
14
16
FIGURE 5.10
Tensile force-strain per atom for the SWCNT and BSCNT models: (a) SWCNT and (b) BSCNT.
167
Molecular Dynamics
To calculate the relative Youngs modulus ratio value between two materials and to satisfy an elastic limit criteria, an applied tensile force range of 1530 nN/atom was selected,
and the linear regression analysis was applied to this selected data as shown Figure 5.11.
Based on the linear regression analysis under the assumption that there is an elastic limit
on a force-strain diagram, the ratio of force to strain of a specific material can be represented as the slope of the regression line.
In Figure 5.11, although the data set of the SWCNT model shows a characteristic that was
more nonlinear than those of the BSCNT model within the same applied force range, the
linear regression analysis could be used for both data sets due to a good-fitting statistical
coefficient R 2, which quantifies goodness of fit and is a fraction between 0.0 and 1.0, with
no units. Higher R2 values indicate that the fitting line or curve comes closer to the data. In
this case, the computed R 2 using the statistical software package was 0.91 for the SWCNT
and 0.99 for the BSCNT models. These numerical values indicate that the linear assumption to calculate the slope of a regression line for SWCNTs and BSCNTs is appropriate.
The resultant slope of the regression line was 7416.3 for the SWNT and 3691.3 for the
BSNT models, as seen in Figure 5.11. Consequently, the ratio value for the BSCNT model
with respect to the SWCNT model was calculated as 0.498. This means that the average
Youngs modulus of the BSCNT model was 49.8% of that of SWCNT.
From the results of the previous equilibrium MD simulations, the evaluated mean
Youngs modulus of the SWCNT and BSCNT models were 1.424 TPa and 0.604 TPa, respectively. Based on these two values, the calculated ratio is 0.424 (42.4%). This is consistent
Force per atom (nM)
40
30
20
10
0
0.5
1.5
2.5
3.5
2.5
3.5
Milli-strain
(a)
Force per atom (nM)
40
30
20
10
0
0.5
(b)
1.5
Milli-strain
FIGURE 5.11
The linear regression fitting for the SWCNT/BSCNT within the elastic limit: (a) SWCNT and (b) BSCNT.
168
with the ratio value of the nonequilibrium MD simulation results, 0.498 (49.8%), within an
error of 15%. Consequently, although we could not find an exact numerical value for the
Youngs modulus of the SWCNT and BSCNT models from the nonequilibrium MD simulations, the Youngs modulus of the BSCNT model was observed to be lower than that of
the SWNT model; that is, a heterogeneous CNT might be found to have a lower Youngs
modulus than pure CNTs.
5.7Application to Carbon Nanotubes: Vibrational Mode Shapes

In this section, radial breathing modes of armchair nanotubes of different diameters are
first discussed, and a comparison is provided for vibrational modes of the same tubes computed using different methods and interatomic interactions based on tight-binding and ab
initio discrete Fourier transform (DFT) approaches. The other low-frequency modes are
described in material that follows; these modes include axial breathing, nonaxial bending,
and torsional shear for single-wall, multiwall, and bamboo-like CNTs.
5.7.1Comparison of Radial Breathing Modes of Armchair SWCNTs
The natural frequencies of radial breathing modes in the low-frequency region were calculated for armchair-type nanotubes with different diameters (Table 5.5). The results obtained
with the present technique were summarized along with the results obtained using other
techniques [3033]. Overall, the results obtained using the present technique were comparable to results using other methods. In particular, these results agreed well with the ab
initio results. On the other hand, these were slightly higher than the experimental data and
the constant force results and slightly lower than the results using the tight-binding methods. The computational results from the constant force model agreed well with experiments
because the constant force models were generally fitted directly to the experimental data.
5.7.2
Natural Frequencies and Mode Shapes for SWCNTs
Natural frequencies and mode shapes were computed for an (8, 8) armchair-type SWCNT
with both ends fully constrained to their equilibrium position. There were various vibrational mode shapes for the SWCNT, such as the axial vibrational mode, lateral bending
mode, radial breathing mode, twisting mode, axial shear mode, and so on. Each vibration type had different mode shapes associated with natural frequencies, which ranged
from the lowest to the highest. However, only the first few lowest-mode shapes for some
TABLE 5.5
Natural Frequencies (in THz) Associated with the Radial Breathing Mode in the
Low-Frequency Region
Chirality
(7, 7)
(8, 8)
(9, 9)
(10, 10)
Present
Technique
Constant
Force [30]
Tight
Binding [31]
Ab Initio [32]
Experiments [33]
7.39
6.86
5.83
5.57
6.18
5.49
4.95
8.04
7.14
6.42
5.85
6.57
5.85
5.25
6.90
6.18
5.46
4.86
169
Molecular Dynamics
vibration types are discussed here. This is because the lowest natural frequencies and
mode shapes are more important in most thermomechanical and vibrational resonatortype engineering applications.
In the following discussion, all the mode shapes are presented in the ascending order
of the natural frequency, starting from the lowest. The first two accordion-like axial vibrational mode shapes, with natural frequencies 4.683 and 9.328 THz, are shown in Figure5.12.
The first mode is the movement of all atoms in one direction along the tube axis, and the
second mode shows the atoms moving away from the center line toward the top and bottom ends of the nanotubes longitudinal direction.
A set of lateral bending vibrational modes is shown in Figure 5.13. The first mode is a
half-sine shape, and the second mode is a full-sine shape. The exact frequency and shape
16
16
14
14
12
12
10
10
2
10
10
10
2
10
10
10
FIGURE 5.12
First two axial vibration modes of the SWCNT.
16
16
14
14
12
12
10
10
2
10
10
10
FIGURE 5.13
First two bending vibration modes of the SWCNT.
2
10
10 0 10
170
of the modes depend on the length of the nanotube under consideration. The longer the
tube is, the lower the natural frequency is for lateral bending vibration.
Figure 5.14 shows the first two radial breathing modes with frequencies of 8.690 and
9.664 THz. Because of the fixed end boundary constraints, the atoms at the center of the
tube have the largest displacement for the first radial mode. On the other hand, the second
mode has a vase shape. Figure 5.15 illustrates a twisting (or torsional) mode shape with a
frequency of 5.331 THz. In other words, the atoms move along the circumferential direction except for the constrained top and bottom atoms. Another vibrational mode was the
axial shear mode with a frequency of 5.331 THz, as seen in Figure 5.16. In this case, atoms
at one side of the tube diameter move in one axial direction while rest of the atoms move
in the other direction of the axis.
16
16
14
14
12
12
10
10
2
10
10
10
10
FIGURE 5.14
First two radial vibration modes of the SWCNT.
6
4
20
10
10
20
10
FIGURE 5.15
Twisting mode of the SWCNT.
6
10
0 10
171
Molecular Dynamics
16
14
12
10
8
6
4
2
0
2
10
10
10
FIGURE 5.16
Axial shear mode of the SWCNT.
5.7.3Natural Frequencies and Mode Shapes for MWCNTs

For MWCNTs, the interaction among atoms in the same nanotube was modeled using
the Tersoff-Brenner potential; the interaction of atoms located in two different concentric
nanotubes was modeled using the van der Waals interactions of the LJ 6-12 type. The
parameters for the LJ potential were selected such that = 0.0778 kcal/mol and = 3.4 for
Equation 5.27. The MWCNTs consisting of two concentric SWCNTs can have symmetric
and asymmetric modes. For the symmetric modes, both the inner and outer tubes have
deformations in the same directions for the same mode, while for the asymmetric modes
the inner and outer tubes have deformations in different directions for the same mode
shape. The analysis of MWCNTs did not show any vibrational mode, which consisted of
two different types of modes for the inner and outer nanotubes, such as a bending mode
for one tube and a radial mode for the other. This is probably because van der Waals interactions are strong enough to disallow any partial or full mixing of different vibrational
modes of the inner and outer tubes. This may also be because both the inner and outer
tubes in the multiwall tube model are of same chirality (i.e., both are armchair nanotubes).
The mode mixing may be possible in tubes of different chiralities, and it needs to be further investigated. The results also showed that, for the same chirality, the symmetric and
asymmetric mixing of modes in the nanotubes was possible through weak van der Waals
forces in a multiwall nanotube.
The natural frequencies of the MWCNTs were compared to those of the two individual
(5, 5) and (10, 10) SWCNTs of the same length under similar conditions. The computed
natural frequencies of the MWCNTs were sensitive to the strength of the van der Waals
forces acting between two concentric tubes. The present results were obtained using the
LJ parameters given previously. As far as the axial accordion and lateral bending modes
are concerned, all three cases had almost similar natural frequencies. Similar natural frequencies in the (10, 10) and (5, 5) SWCNTs for these modes were probably because bending
and axial modes do not vary with the radii of the nanotubes. This behavior was similar to
the one seen in continuum models of axial vibration of a rod and bending vibration of a
hollow beam. The natural frequency depends only on length and not radius in the case of
axial vibration. In the case of bending of a hollow beam, length has a more dominant effect
on the natural frequency than the cross-sectional area.
172
TABLE 5.6
Comparison of Natural Frequency of the First Radial Mode between MWCNT
and SWCNT
MWCNT
Frequency (THz)
Symmetric
Asymmetric
(5, 5) SWCNT
(10, 10) SWCNT
4.684
5.667
5.665
4.682
For the symmetric and asymmetric mixing, however, the MWCNT has two natural
frequencies, for the first accordion and bending modes, respectively. The frequencies
for the asymmetric modes are slightly higher than those for symmetric modes. For
the radial mode of vibration, the (10, 10) SWCNT had a 17% lower frequency than the
(5, 5) SWCNT. The natural frequency associated with the symmetric radial mode of
MWCNT was slightly higher than the (10, 10) SWCNT, while the frequency related to
the asymmetric radial mode was slightly higher than the (5, 5) SWCNT, as shown in
Table 5.6.
5.7.4Natural Frequencies and Mode Shapes for BSCNTs
As one of the possible heterogeneous CNTs, the BSCNT was studied for its natural frequencies and mode shapes and compared to a conventional SWCNT. The simulated stable
structure was investigated for low-frequency vibrational characteristics. Radial breathing
modes of a BSCNT are also shown in Figure 5.17. Those mode shapes for the SWCNT
and BSCNT look similar in a global sense. However, the local behaviors were different
between the two CNTs depending on the mode shape.
Natural frequencies of SWCNT and BSCNT are compared in Table 5.7 for different vibrational modes for the CNTs with the same diameter. The table shows the ratios of the BSCNT
natural frequencies in relation to the SWCNT natural frequencies. As shown in Table 5.7,
16
16
14
14
12
12
10
10
2
5
FIGURE 5.17
First two radial vibration modes of the BSCNT.
2
5
173
Molecular Dynamics
TABLE 5.7
First Natural Frequency Ratios between the SWCNT and BSCNT
Vibrational Mode
Accordion Mode
Bending Mode
Radial Mode
BSCNT frequency/
SWCNT frequency
0.917
0.917
1.160
the natural frequency of BSCNT associated with the first axial accordion or lateral bending
mode shape was lower than that of the SWCNT. On the other hand, the first natural frequency of the radial mode of the BSCNT was higher than that of the SWCNT. These results
can be explained as follows: The additional number of the atoms in the cross-sectional
layer at the junction in a BSCNT can contribute to both stiffness and mass ofthe overall
CNT structure. For the first accordion-like axial and lateral bending modes, theeffect of
the additional mass was more critical than the change of stiffness because the internal
atomic membrane structure of a BSCNT does not contribute much to the axial and bending stiffness. However, the case was opposite for the radial vibration. The cross-sectional
layer structure at the junction increased the radial stiffness significantly so that the radial
frequency of the BSCNT increased compared to the SWCNT.
5.8Application to Polymers
5.8.1Cross-Linking of Polymers
Molecular dynamics simulations are in heavy demand for the study of randomly crosslinked polymers. Barsky and Plischke previously reported on simulations that involved
short chains and small system sizes [34]. These simulations showed that there was a universal function that stated the distribution of localization lengths for a wide range of crosslink density. The relationship between the shear modulus and density were never taken
into account. During later studies, Barsky and Plischke extended their simulations to longer polymer chains and systems [34]. Their focus was on the shear modulus as a function
of the density of cross-links.
The purpose of the cross-links is to tie all the polymer molecules together. This prevents molecules from flowing past or around each other once the temperature increases
and increases their resistance to melting. The advantage of being tied together is that
the molecules are not easily broken apart from each other. The difference between
uncross-linked polymer chains and a cross-linked network is described in Figure 5.18.
However, in this study the temperature was held constant as was the cross-linking
density.
This section describes the construction of a molecular model. The goal is to randomly
generate polymer chains within a cube and determine how many molecules per given
chain are necessary to produce homogeneous behavior. Computer simulations were used
to focus on randomly dispersed particles in a three-dimensional (3-D) space. These simulations contained different volumes as well as polymers with different lengths, N = 5, ,
45, with unit multiples of five molecules in a chain. The 3-D models of polymers were
used to create structures that properly represented real molecules, thus providing various
174
When polymers become cross-linked, this becomes this

FIGURE 5.18
How polymers appear once they are cross-linked.
35
30
25
20
15
10
5
0
5
10
150
100
50
50 20
20
40
60
80
100
FIGURE 5.19
Composite material with particles randomly dispersed in a 90 90 15 space volume.
construction and analysis strategies. Figure 5.19 shows a simulation with 12 polymer
chains extending in both the x and y directions and 2 chains in the z direction.
Using cross-linking potential systems, MD simulations were conducted in the 3-D polymeric matrix state. The simulations emphasized the relation between the correlation volume of a space and the concentration of the random particles.
5.8.2Allocation of Polymers
Molecules formed in this study were first confined to a cube with a set of dimensions in
the x, y, and z directions. For the construction of the polymer chain, an allowable distance
between molecules within the polymer was first established. All measurements during the
construction were dimensionless. To ensure atoms were not too close or too far apart, an
allowable tolerance input was set. After creating the foundation, the actual construction of
the polymer chain began. For optimum results that are not too computationally expensive,
the number of polymer chains in the x and y directions were the same, and the z direction
was one-third of the x directions value. The number of molecules contained in each chain
was chosen to be constant.
175
Molecular Dynamics
To determine the spacing in each direction, the lengths of the cube were divided by the
number of polymer chains in the same direction. The total number of atoms contained in
the cube was the product of the number of molecules in a chain and the number of polymer chains in the x, y, and z directions.
As far as the atomic structure was concerned, the bead-spring model was used. In
essence, the bead-spring model simulates the hydrodynamic properties of a chain macromolecule. These chains consist of a sequence of beads, each of which offers hydrodynamic
resistance to the flow of the surrounding medium. The beads are connected to each other
by a spring, which does not contribute to the frictional interaction. However, the spring is
responsible for the elastic and deformational properties of the chain. The mutual orientation of the springs is random.
For this project, each bead, or mapping point, represented a specific monomer. This
required a fixed distance of the beads along the backbone, bond angles around the backbone, and torsion states around the bead-bead connection to be held constant.
A loop was generated to account for the individual atoms and their physical location.
To initiate the chain creation, the first atoms position was generated. The remainder of
the atoms along the chain were randomly generated. During the random generation, the
atoms positions were checked to verify whether they were within the tolerance.
5.8.3Potential Energy for Polymers
The interactions between the polymers and atoms led to prediction of the large-scale
bulk properties of material. This MD program used the LJ potential method to evaluate
molecule interactions between the polymers and atomic atoms. The LJ potential is an
effective potential that describes the interaction between two uncharged molecules or
atoms.
The LJ potential was mildly attractive as two uncharged molecules or atoms approached
each other from a distance, but strongly repulsive when they approached too closely. At
equilibrium, the pair of atoms or molecules tended to go toward a separation corresponding to the minimum of the LJ potential. The strong close-in repulsion between atoms or
molecules is understandable, resulting from mutual deformation of their structures (one
atom cannot diffuse through another). The mild attraction at larger distances was due to
the induced dipole-dipole moment interaction of the particles.
As stated, the LJ potential is used to ensure self-avoidance of polymers. The following
formula was used to compensate for the added attractive potential between neighboring
atoms:
1 2 rij
kR
ln
U nn (rij ) = 2 o Ro
r <R
o
ij
,
rij Ro
(5.62)
with Ro = 1.5 and k = 30 /2. By incorporating this method, polymers were prevented
from passing through each other [34].
To consider coupling of metallic and polymeric materials, the embedded atomic potential was used for the metallic atoms. The foundation of the embedded atom method, based
on the quasi-atom concept, is the realization that the cohesive energy of a metallic system
176
140
120
100
80
60
40
20
0
100
50
0 0
20
40
60
80
100
120
140
FIGURE 5.20
Randomly generated polymers with aluminum at the center contained in a 141 141 141 volume. The number
of polymers in each chain in the x, y, and z directions is six with six beads in each chain. Total number of atoms
in the structure is 8160.
can be expressed in terms of embedding energies. Figure 5.20 shows the creation of a polymer with an aluminum core [35].
For the present MD simulation, each system was equilibrated to an average dimensionless temperature of 300 and a time step equal to 0.01. A molecule was selected at random
and then connected to another random molecule within the same polymer chain. A standard distance was used for the molecules along with a combination of the potentials for
polymers described previously. Molecules contained in the same chain were only linked
once, and no cross-linking of neighboring atoms was permitted.
5.8.4Autocorrelation
All of the viscometric functions are derived from elements of the stress tensor. The stress tensor can be determined from the pressure tensor, so only one of these needs to be calculated by
the program. In an equilibrium MD simulation, Newtons equations describing the motion of
the atoms in the model system are solved as a function of time. The viscosity is then given
as an integral of the stress-stress autocorrelation functions determined during the simulation:
V
=
lim dt Pxy (t)Pxy (0) (5.63)
kBT
where V is the volume of the system, kB is Boltzmanns constant, T is temperature, and t is

time. The quantity Pxy(t) is the value of the xy component of the traceless symmetric stress
tensor at time t, so Pxy(t)Pxy(0) is the stress-stress autocorrelation function measured during
the course of the simulation.
5.8.5Examples
Dimensionless units were used to compute the stress tensor in the xy, yz, and xz directions.
Several variables remained constant throughout each simulation to observe the polymer
177
Normalized shear modulus
Molecular Dynamics
800
700
600
500
400
300
200
100
0
5000
10,000
Number of atoms
15,000
FIGURE 5.21
Plot of the normalized shear modulus as a result of proportionally varying the density with the volume of the
cube.
behavior. The temperature remained constant at 300 with a time step of 0.01. The number
of atoms was initially computed by the number of beads or molecules in a chain multiplied
by the number of polymer chains in each direction. The density was then calculated by
dividing the number of atoms by the cubes volume. All simulations were run a minimum
of three times. The average of the results was recorded and plotted. In all of the simulations completed, there was a noticeable homogeneous trend.
The first simulation required proportionally varying the x and y dimensions of the cube
along with the number of atoms in the same directions. The cube length in the z direction
was set to a constant, and the number of polymer chains extending in this direction was
two. The number of monomers in a chain remained constant at 20. Note that the normalized shear modulus approached a constant value, as seen in Figure 5.21, when the number
of atoms became more than 8000.
During the next simulation, the volume was held constant, and the number of monomers per chain was held at 20. The number of polymer chains that extended in the z direction remained constant at Nz = 2; however, the x and y directions were increased by one
unit starting from Nx = Ny = 3 after a series of runs were completed. After the total number
of atoms became about 9000, the shear modulus remained roughly constant, as seen in
Figure 5.22.
Shear modulus
3.0000E+01
1.5000E+01
0.0000E+00
2000
4000
6000
8000
10,000 12,000 14,000
Number of atoms
FIGURE 5.22
Plot of the normalized shear modulus determined from a cube of constant volume (112.5 112.5 15) as a function of varying density.
178
The following simulation required holding both the volume and density constant. The
one varying parameter was the number of monomers per given chain (N). The increments
were in units of five beginning with N = 5, , 45. The results showed that the normalized
shear modulus began to approach a constant value as the number of atoms became more
than 8000, as seen in Figure 5.23.
The final example considers a face-centered cubic (FCC) structure inserted in the center of a cube with a constant volume. The FCC was used to simulate the characteristics
of aluminum. There were six polymer chains extending in each direction with variance in the number of monomers per given chain for the first simulation. The varying
parameters for this simulation included the size of the FCC and the number of polymers
interacting with this metallic structure. Thus, the number of monomers per given chain
decreased proportionally with the expansion of the metallic structure to maintain a constant density.
Ideally, this simulation was created to represent a composite material. As the metallic
core size increased, the polymers space became more confined due to the constant volume of the cube. This means the volume fraction of the metallic particle increased in a
polymer composite. As shown in Figure 5.24, the normalized shear modulus continuously
increased along with the increasing size of the metallic core.
Shear modulus
1.2
1.0
0.8
0.6
0.4
0.2
0.0
2000
4000
6000
8000
10,000
12,000 14,000
Number of atoms
Normalized shear modulus
FIGURE 5.23
Plot of the normalized shear modulus determined from a cube with constant volume and polymer density as a
function of molecule variance.
5
4
3
2
1
0
50
100
150
200
250
Number of FCC cubes in metal core

FIGURE 5.24
Plot of the average normalized shear modulus for a cube with a constant volume and increasing metallic core
size under constant polymer density.
179
Molecular Dynamics
5.9Application to Nanoscale Fluid Flow

A 3-D microchannel used in the simulation was represented by a lattice atomic configuration. The simulation model consisted of fluid atoms and solid atoms. The most outside
atoms parallel to the x axis were considered a wall. The wall, solid, and liquid atom masses
were assigned in the following way: mw = ms = 3ml, respectively. The LJ potential was used
for all atoms, for example, with the following parameters: = 2; w = 0.5, s = 0.5, and f =
0.1 for the wall, solid, and liquid atoms, respectively; and ws = 0.5, wl = 0.2, and sl = 0.2 for
the energy value between the wall, solid, and liquid atoms, respectively.
For the current study, the periodic boundary conditions were applied to the computational domain such that the number of atoms leaving the computational domain was equal
to the number of atoms entering the inlet domain. Due to the fact that the simulated system of molecule geometry was very small, the surface effect had a great influence on the
molecular dynamic program performance and calculations; therefore, the periodic boundary conditions were used to reduce the atoms surface effect interaction with the walls [36].
The geometry used for performing the required calculations consisted of a system of
molecules in which there were 800 atoms. In all cases, the y-axis velocity value was equal
to zero; this allowed us to observe the system of molecules in the x z plane only.
5.9.1Fluid Flow Simulation
The fluid flow simulation was performed to obtain the velocity profile of the flow at different locations along the x axis. In this section, the liquid atoms only had an assigned x-axis
velocity by applying the force fxf = 0.1 along the x axis. Figures 5.25 through 5.27 show the
velocity profile at x = 4, 20, and 37, respectively.
Figure 5.25 indicates that the velocity profile was somewhat similar to most of the velocity profiles obtained at the inlet of the computational domain. This figure shows that the
solid atoms had a small amount of motion compared to the fluid atoms.
However, in this case it was expected to have fluctuation in the profile due to the atoms
vibration when simulation was performed at an atomic level. Figures 5.26 and 5.27 show
that the flow profile was gradually developed until the velocity profile somewhat similar
to the expected parabolic shape was obtained at x = 37.
0.012
Velocity
0.008
0.004
0.000
0.00 16.14 25.45 35.42 26.60 60.50
Distance
FIGURE 5.25
The x-axis velocity profile at x = 4.
74.76
180
0.012
Velocity
0.008
0.004
0.000
0.00
13.07 24.47 34.64 48.75 60.85 72.78

Distance
FIGURE 5.26
0.0012
0.01
Velocity
0.008
0.006
0.004
0.002
0
0.00
11.67
24.73
33.64
47.47
61.11
72.69
Distance
FIGURE 5.27
5.9.2Particle-Fluid Interaction
Different simulation cases were performed for the particle-fluid interaction. Figure 5.28
shows the arrangement of solid atoms inside fluid atoms. The solid atoms are denoted by
the boxes in the figures. The solid atoms were arranged horizontally. To trace the atomic
motions, the initial and final atomic locations are plotted in Figures 5.29 and 5.30 for the
solid and fluid atoms, respectively [37]. The fluid atoms had larger motion than the solid
atoms.
Both the solid and liquid atoms were assumed to have an equal x-axis force. The velocity
profiles at different locations along the x axis in the x-z plane were reviewed. By observing
these behaviors, it can be concluded that due to the effect of the particles (i.e., solid atoms)
dragging at the center of the computational domain, the solid atoms had a great effect on
the velocity profile, as shown in Figure 5.31. This effect was associated with a significant
reduction in the x-axis velocity in all profiles, as expected.
In the next study, the particles were assigned at random locations throughout the
domain, as indicated in Figure 5.32. All atoms except for the boundary solid atoms had an
181
Molecular Dynamics
80
70
60
50
z
40
30
20
10
0
0
10
15
25
20
x
30
35
40
FIGURE 5.28
Solid atom locations inside fluid atoms. Atoms in the boxes are the solid atoms.
43.00
42.00
41.00
z
40.00
39.00
38.00
37.00
36.00
0.00
4.00
8.00
x
12.00
16.00
FIGURE 5.29
Original and final configurations of solid atoms. Black circles indicate the initial locations, and the gray circles
denote the final locations.
43.00
42.00
41.00
40.00
39.00
38.00
37.00
0.00
4.00
8.00
12.00
16.00
20.00
FIGURE 5.30
Original and final configurations of liquid atoms. Black circles indicate the initial locations, and the gray circles
denote the final locations.
182
1.40E-02
1.20E-02
Velocity
1.00E-02
8.00E-03
6.00E-03
4.00E-03
2.00E-03
0.00E+00
0.00
14.27
22.36
35.03 48.76
Distance
58.16
73.49
FIGURE 5.31
Liquid velocity profile at x = 37 with solid atoms at the center.
80
70
60
50
z
40
30
20
10
0
10
15
20
x
25
30
35
FIGURE 5.32
Solid atom locations inside fluid atoms. Atoms in the boxes are the solid atoms.
0.020
Velocity
0.016
0.012
0.008
0.004
0.000
0.00
10.15
18.67
38.06 45.20
Distance
FIGURE 5.33
Random particle-liquid interaction velocity profile at x = 37.
59.90
72.08
40
183
Molecular Dynamics
x-axis force fxf = 0.1. The results obtained had a significant amount of atomic motion. The
concept of the drag forces of particles still exists, as shown in Figure 5.33.
5.10Application to Fatigue of Metals

5.10.1 M odeling
For modeling, the position of each atom was assigned an identification number (1 to 500
in this study) and the Cartesian coordinate had angstrom units. The final array of atoms
was in the shape of a cube or a box. A random 3-D velocity in angstroms/picosecond was
also assigned to each atom. The potential function used was the embedded atom method,
which includes the atomic number, atomic mass, lattice spacing, potential cutoff range
for each atom, grid of interatomic force, grid of atom electron density, and the spherically
averaged atomic density. In this study, periodic boundary conditions were selected so that
atoms moving outside the box boundary were remapped to the other side of the box.
Because the system was initialized with random velocities assigned to the atoms, it took
time for the system to reach equilibrium. Analysis of the output data showed that values
such as total energy fluctuated as a result of the random velocities before reaching a constant value. Equilibrium occurred consistently around 1400 time steps. This began the
production phase, and data produced during this period were used in the analysis.
The simulation used, for example, the following units: energy in electron-volts, temperature in kelvin, pressure in bars, and volume in cubic angstroms. A model of 500 copper
atoms arranged in a FCC array was implemented to investigate the behavior of copper
under conditions of cyclic loading. From the original 500-atom model, configurations
were created that contained defects in the form of vacancies or impurities. At least two
entire rows of atoms were removed to create the configuration containing vacancies. This
assisted in visualization of the movement of the vacancies relative to one another. To create the configuration with impurities, the vacancies of the previous configuration were
replaced with nickel atoms and assigned corresponding potentials [38].
The simulation was performed on these configurations of atom models to simulate cyclic
loading by applying a pressure of 10 bars for a time step of 0.005 ps with a duration of
1ps followed by 0 bar applied for the same length of time. This gave a load cycle of 2 ps,
as shown in Figure 5.34. The positions of the defects and the energy data were tabulated
at the end of each period of loading and unloading. The positions of the impurities were
tracked by recording the coordinates given in the output position file. The positions of the
vacancies were tracked using the average position value of the nearest neighbors in the z
direction.
Force
Time
FIGURE 5.34
Cyclic loads.
184
5.10.2Examples
Figure 5.35 shows an initial arrangement of metallic atoms with initial vacancies. Figure
5.36 shows the relative distance between defects versus the number of cycles. The ones
labeled as loaded state represent the positions for each load period. In other words, it
shows the position every 2 ps starting with the first load. The figures also compare a parallel load versus a normal load. A parallel load indicates a load applied along the z axis of
Figure 5.35. It is parallel to the line connecting the vacant positions. The figures show a
fluctuation in distances with increasing numbers of cycles. The general trend was a separation of the vacancies. The separation of the vacant positions was larger for the parallel
load than for the normal load as the number of cycles increased.
8
6
4
z (1/10 nm)
2
0
2
4
6
8
10
10
y (1/10 nm)
FIGURE 5.35
Two-dimensional view of 3-D copper atoms with two rows of vacancies.
2.05
Norm. relative distance
2.04
2.03
Parallel load
Normal load
2.02
2.01
2
1.99
1.98
1.97
1.96
0
15
10
No. of load cycles
20
25
FIGURE 5.36
Plot of the normalized distance between the two vacancies versus the number of cyclic loads at the unloaded
state.
185
Molecular Dynamics
Figure 5.37 compares the change in average potential energy of the two loading types.
There was an increase in potential energy in both cases as the number of cycles increased.
This indicated that as the atoms moved apart, the progressive strength decreased along
with increased cycles. Thus, the potential energy can be an indicator of fatigue damage accumulation. At the same time, there was an increase in the average kinetic energy
with increasing cycles, as shown in Figure 5.38. There was only a minor difference in the
energy plots for the two loading orientations, indicating that whether parallel or normal
was irrelevant. Figures 5.39 through 5.41 also plot the relative distance, average potential
energy, and kinetic energy with impurity, respectively. The general trends were between
the vacancy and impurity models.
3.5655
Parallel load
Normal load
Potential energy per atom
3.566
3.5665
3.567
3.5675
3.568
3.5685
3.569
10
15
No. of load cycles
20
25
FIGURE 5.37
Plot of average potential energy per atom in the system with vacancies versus the number of cyclic loads at the
unloaded state.
0.022
Parallel load
Normal load
Kinetic energy per atom
0.0215
0.021
0.0205
0.02
0.0195
0.019
0.0185
10
15
No. of load cycles
20
25
FIGURE 5.38
Plot of average kinetic energy per atom in the system with vacancies versus the number of cyclic loads at the
unloaded state.
186
2.05
Parallel load
Normal load
Norm. related distance
2.04
2.03
2.02
2.01
2
1.99
1.98
1.97
1.96
0
10
15
No. of load cycles
20
25
FIGURE 5.39
Plot of the normalized distance between the two impurities versus the number of cyclic loads at the unloaded
state.
Potential energy per atom
3.5355
3.536
Parallel load
Normal load
3.5365
3.537
3.5375
3.538
3.5385
3.539
10
15
No. of load cycles
20
25
FIGURE 5.40
Plot of average potential energy per atom in the system with impurities versus the number of cyclic loads at the
unloaded state.
Kinetic energy per atom
0.0205
0.02
0.0195
0.019
0.0185
Parallel load
Normal load
0.018
0.0175
0
10
15
No. of load cycles
20
25
FIGURE 5.41
Plot of average kinetic energy per atom in the system with impurities versus the number of cyclic loads at the
unloaded state.
6
Coupling Techniques
6.1Introduction
Various computational techniques have their own merits and limitations. For example,
the finite element model (FEM) is useful for a continuous medium level, while molecular
dynamics (MD) is applicable to the molecular or atomic level. As a result, if the FEM is
applied to the continuum level and the MD is applied to the atomic level, it is necessary
to develop a coupling scheme in a multiscale analysis that bridges both levels of analysis.
This chapter describes coupling schemes among various analysis techniques so that multi
scale or multiphysics problems can be modeled and simulated.
6.2Coupling Technique between Finite Element and Atomic Models

To combine discrete atoms and a continuous medium, the atomic model and the FEM of
the continuous medium are coupled [16]. In this section, a two-dimensional (2-D) case
is presented to show the coupling scheme; the developed technique can be applicable
to three-dimensional (3-D) cases. Figure 6.1 shows atoms surrounded by a continuous
medium that is discretized for a finite element mesh. In other words, the inner domain
is the atomic domain, and the outer domain is the finite element domain of a continu
ous medium. To exchange information between the atomic domain and the continuous
domain, an overlapped interface domain is defined as shown in the figure; the domain is
bounded in the figure by inner and outer bold line rectangles. The interface domain con
tains both finite elements and atoms.
In the coupling scheme presented here, the atomic and FEMs are solved independently
one after the other in a staggered manner. As shown in Figure 6.1, the external loading
and boundary conditions are applied to the continuous medium. The first analysis is con
ducted for the continuous domain using finite element analysis. The step-by-step solution
procedures for the coupled analysis are described next for static problems [1,2].
1. Develop the finite element matrix equation for the continuum domain, such as
[Kf]{uf} = {Ff}. From this point, subscripts f and a denote finite element and atomic
solutions, respectively. Then, solve the matrix equation with applied loading and
boundary conditions. For the first iteration, the nodal forces of the finite elements
in the interface domain are set to zero.
187
188
Applied force
Finite element
analysis domain for
continuous medium
Discrete
atomic
domain
Overlapped
interface domain
between the
continuous and
discrete domains
FIGURE 6.1
Problem with partial continuum and partial atomistic domains.
2. Obtain the nodal displacement of the finite elements in the interface domain.
Compute the embedded atoms displacements from the finite element nodal dis
placements using the finite element shape functions, such as {uae } = [ N ]{uef }, where
superscript e indicates the embedded atoms or finite elements containing such
atoms in the overlapped interface domain, and [N] is the shape function matrix of
finite elements.
{ }
3. Compute the new positions xae of the embedded atoms in the interface domain
by adding the displacements computed to the previous position vectors.
4. Solve for the rest of the atoms new positions with the fixed embedded atoms
positions using atomic analysis.

{ }
5. Compute the resultant atomic forces Fae on the embedded atoms exerted by all
neighboring atoms.
6. Compute the equivalent nodal forces of the finite elements containing the atoms in
the interface domain using Ffe =
[ N ]T Fae , where the summation is conducted
{ }
{ }
for all the atomic forces included in the finite elements in the interface domain.
7. With the finite element nodal forces computed in the interface domain, a new finite
element solution is obtained by applying the external and boundary condition as
well as the nodal forces at the interface boundary. Then, continue and repeat the
process from step 2 until the most recent solutions are within the acceptable toler
ance level.
6.3Example Problems of Coupled Finite Element and Atomic Models

Some examples are provided in this section to demonstrate the coupling technique
between the discreet atomic and continuum finite element domains.
Coupling Techniques
189
6.3.1 Hexagonal Array of Atoms with Dislocation

Finite element domains are applied around atomic domains. This kind of domain setting
makes it convenient to apply physical loading and boundary conditions because it is easy
to apply those conditions in the finite element domains. Therefore, as coupled examples,
uniform shear or axial force is applied at the top of the FEM. Figure 6.2 shows the hex
agonally arranged atoms surrounded by the finite element mesh. To minimize the com
putational cost, the overlapped interface zone does not contain many atoms. In addition,
the atomic model has multiple atoms removed in the middle to represent dislocation, as
shown in Figure 6.2. To represent dislocation properly (i.e., to represent the atomic rear
rangement after removal of the middle atoms), the initial equilibrium of atoms is com
puted. In this calculation, no load is applied to the FEM model. Only a minimum number
of constraints are applied to the surrounding FEM to avoid rigid body motions because
the static equilibrium is considered in this study. For a 2-D domain, two linear and one
rotational motions must be suppressed.
The iteration procedure described in the preceding section is applied to the atomic
and finite element domains until the new atomic positions are determined after removal
of the atoms. Figure 6.3 shows the equilibrium positions of atoms after removal of
atoms in the center. The figure shows atomic dislocation; the finite element domain is not
shown in the figure. After the initial equilibrium of atoms is achieved, an external load is
applied to the finite element domain. The first load is the shear load on the top boundary
of the domain; the bottom boundary is constrained from the vertical motion, and the left
bottom corner is constrained from both directional linear motions. Figure 6.4 shows the
movement of atoms under shear loading. The direction and magnitude of the motion are
denoted by the arrow and the length. The longer arrow means a larger displacement along
the arrowhead direction. Because of the dislocation introduced in the middle, there are
noticeable vertical components in the atomic motions around the dislocation zone. The
displacements of atoms are magnified in the arrow plots. Therefore, the final positions of
atoms are not at the tip of the arrows. Figure 6.5 shows the final atomic positions with the
applied shear load. Figure 6.6 shows the displacements of atoms with dislocation under
a tensile load. The atoms around the dislocation zone show large displacements under
tensile loading.
FIGURE 6.2
Initial atomic domain with removed atoms in the middle surrounded by finite element domain.
190
FIGURE 6.3
Equilibrated positions of atoms after removal of atoms in the middle.
FIGURE 6.4
Atomic displacements under shear loading on the boundary of the finite element domain.
FIGURE 6.5
Equilibrium position of atoms under shear loading.
Coupling Techniques
191
FIGURE 6.6
Atomic displacements under tensile loading on the boundary of the finite element domain.
6.3.2 Atomic Array Embedded in the Finite Element Mesh with a Crack
This example uses an atomic model for a crack tip, and the atomic model is surrounded
by a continuum FEM as sketched in Figure 6.7. To apply loading and boundary conditions
a distance away from the crack tip, it is necessary to introduce the continuum finite ele
ment domain. Otherwise, the number of atoms to be included in the study would be tre
mendously huge. Two different kinds of atomic arrangement are considered: square and
hexagonal arrays. The finite element domain has zero displacements at the bottom edge
and a uniform tensile displacement at the top edge. In other words, the domain is under a
uniaxial tensile load.
This example shows how the atoms near the crack tip behave under the external tensile
loading. Displacements of atoms at the notch tip zone are plotted in Figures 6.8 and 6.9 for
the square and hexagonal arrays, respectively. In those figures, the atomic displacements
are plotted relative to those of the atom at the crack tip, which is located in the middle in
FIGURE 6.7
Atomic array embedded in the finite element mesh with a crack.
192
FIGURE 6.8
Atomic displacements in a square array near the crack tip.
FIGURE 6.9
Atomic displacements in a hexagonal array near the crack tip.
the vertical direction. Therefore, the crack tip displacement is zero. There is a negligible
difference between the two different kinds of atomic arrangement. As a result, the atomic
arrangement does not seem to affect the atomic displacements in any notable magnitude.
In addition, the atomic domain in the crack tip zone is replaced by a continuum finite
element domain for comparison. In other words, the whole domain is the finite element
domain.
With application of the same loading and boundary condition, the nodal displacements
are plotted in Figure 6.10. The same number of finite element nodes was used as the num
ber of atoms. Of course, this is not physically acceptable because each atom cannot be rep
resented by a continuous medium. However, this comparison was conducted for the sake
of curiosity. Comparison of the displacement between the atomic and continuous models
near the crack tip shows that those are quite different even though there is some qualita
tive similarity between the two models.
Coupling Techniques
193
FIGURE 6.10
Finite element nodal displacement plot near the crack tip in a whole continuous model.
6.3.3 Atomic Behavior at the Crack Tip

The response in atoms was investigated when there were some vacant atoms at the crack
tip. This example is similar to the previous one. However, the crack tip extended into the
atomic model as shown in Figure 6.11, which shows the embedded atoms in a finite ele
ment mesh. An edge crack is shown in the figure, and the crack tip extends into the atomic
model. The two left bottom nodes were constrained, while the left top node was subject to
tensile force.
The equilibrated atomic positions before applying the external load are shown in Figure
6.12. Then, a tensile load was applied at the left side of the model like a compact specimen
as described previously. Figure 6.13 is the plot for atomic movement with the tensile load
and the final atomic positions in equilibrium with the external load. The figure shows
some vertical separation of atoms at the bottom side of the notch tip because of the stress
FIGURE 6.11
Atomic and finite element model with a crack.
194
FIGURE 6.12
Initial equilibrium positions of atoms.
FIGURE 6.13
Movement of atoms with tensile opening force.
concentration at that point. To see if there was stress concentration at that point in a con
tinuous medium, a finite element analysis was conducted for the whole domain. In other
words, the atomic domain was replaced by a finite element domain. The FEM showed that
the highest stress point of the continuum model exactly corresponded to the location of
the atomic model, which means that the atomic behavior in nanoscale was comparable to
that in the continuum scale.
6.4Homogenization of an Atomic Model into a Continuum Model

Because the typical distance between any two neighboring atoms is less than a nanometer,
even a small size atomic domain contains a huge number of atoms. The number of atoms
195
Coupling Techniques
is directly proportional to the computational cost. As a result, it would be wise to limit the
total number of atoms in the atomic domain. On the other hand, a more realistic atomic
model requires a reasonable size atomic domain. As a compromise for the two conflicts,
a technique was developed so that a group of atomic arrangements can be represented by
an equivalent homogenized continuum model. By doing so, a large atomic domain with
fewer individual atoms can be modeled. The homogenized continuum model is solved
using the FEM. Then, the coupling of the individual atomic domain and the homogenized
atomic domain was analyzed as described in Section 6.2.
To develop a homogenized continuum model for discrete atoms, a representative contin
uum domain was selected out of the discrete atoms [7]. Then, the matrix equation as devel
oped in Chapter 5 (i.e., Equation 5.46) was developed for all possible interacting atoms
inside the chosen representative continuum domain to determine the stiffness matrix of
the atomic model. The stiffness matrix of the atomic model can be written as [Ka], where
the superscript a denotes the atomic model. As the representative continuum domain is
represented using the finite element technique, the displacement vector in the finite ele
ment {u} can be expressed in terms of the shape functions [H] and nodal displacements of
the finite element {uf}, such as the following:
{u} = [H]{uf} (6.1)
As the coordinate values of each atom, which is included in the representative finite
element, are substituted into Equation 6.1, the atomic displacements at the position can
be related to the finite element nodal displacement. Repetition of the process for all atoms
inside as well as on the boundary of the finite element results in
{ua} = [N(xi)]{uf} (6.2)
where the matrix [N(xi)] consists of shape functions evaluated at the atomic position xi
included in the representative finite element. The size of the atomic displacement vec
tor {ua} is equal to nad. Here, na is the number of atoms in the finite element and d is the
problem dimension, that is, 1 for one dimension, 2 for two dimensions, and 3 for three
dimensions. Using the relation given in Equation 6.2, the atomic stiffness matrix [Ka] can
be converted into the finite element matrix equation [Kf] as follows:
[Kf] = [N]T[Ka][N] (6.3)
The finite element stiffness matrix is computed from

[K f ] =
[B] [D][B] dV (6.4)

T
where [B] is the matrix resulting from the kinematic relation (i.e., strain-displacement rela
tion), [D] is the constant material property matrix, and V is the volume of the finite element.
For the simplex finite element family, matrix [B] is also a constant matrix. Then, the inte
gration of Equation 6.4 results in
[Kf] = [B]T[D][B]V (6.5)
196
From Equations 6.3 and 6.5, the material property matrix [D] is
[D] = ([B][B]T)1 [B][N]T [Ka][N][B]T ([B][B]T)1/V (6.6)
The matrix [D] is the smeared material property matrix of the atoms for an equivalent
continuum. As long as the atoms in the finite element are a representative collection of all
atoms in the problem domain, the material property matrix can be used for all finite ele
ment formulations of the problem domain. By doing so, the smeared continuum domain
can replace discrete atomic regions. If necessary, the material property values, such as
elastic moduli, can be extracted from the material property matrix [D].
6.5Example Problems for a Smeared Atomic Model

To verify the smeared model presented in the previous section, example problems are
solved using the discrete atomic model as well as the smeared continuum model, and then
the two solutions are compared.
6.5.1 Vibration of Atoms in a One-Dimensional Arrangement
Let us consider a vibration problem of atoms in a one-dimensional (1-D) arrangement. The
total number of atoms is 101, and they are located linearly with an even spacing d. For sim
plicity, two neighboring atoms are used to compute the equivalent elastic modulus of the
smeared continuum medium even if any number of atoms may be selected for homogeni
zation. Then, the natural frequencies of the discrete atomic domain as well as the smeared
continuum domain are computed, respectively, to compare the two solutions.
Regardless of any interatomic potential energy used for this example, the stiffness matrix
of the discrete atomic model with two neighboring atoms is expressed as

[K a ] = k
k
k
(6.7)
k
where the k value depends on the specific interatomic potential and can be any value for
the present example. The two atoms are smeared into a continuum finite element with two
nodal points whose coordinate locations coincide with the locations of the two discrete
atoms. As a result, the matrix [N] in Equation 6.2 becomes the identity matrix of 2 by 2. The
stiffness matrix of the continuum finite element is

[K f ] =
EA 1
d 1
1
(6.8)
1
where E is the equivalent elastic modulus of the 1-D continuum domain, and A is the crosssectional area of the continuum. Because E and A will be used together as a product for finite
element analysis, there is no need to compute them separately. Comparison of Equation 6.7
to Equation 6.8 yields the equivalent EA value for the continuum model, which is equal to kd.
Using this value, the whole domain can be analyzed using continuum finite elements.
197
Coupling Techniques
When three atoms in equal spacing d are chosen for the homogenization, the atomic
stiffness matrix becomes
k
[ K a ] = k
0
k
2k
k
k (6.9)
k
and the matrix [N] relating atomic and continuum displacement vectors is
1
[ N ] = 0.5
0
0.5 (6.10)
1
The matrix [B] for a two-node linear element is expressed as

[B] = [1/2d 1/2d] (6.11)
The matrix operation as given in Equation 6.6 results in the equivalent EA = kd as pre
viously found. This means that the material property of the smeared continuum is not
dependent on the number of atoms selected for the homogenization process as long as the
atoms are arranged in an equal length space.
Now, the atomic domain with 101 atoms is modeled using one continuum finite ele
ment. Then, the number of finite elements is increased to 11 gradually with uniform
finite element meshes. The mass matrix of the continuum model is also computed from
the atomic masses with an assumption of uniform mass distribution within each finite
element. Natural frequencies are computed using the atomic and smeared continuum
models, respectively. Figure 6.14 shows the plot of normalized natural frequencies as a
function of the number of finite elements in the smeared continuum domain. The natural
Frequency ratio
1
0.8
0.6
0.4
1st mode
2nd mode
3rd mode
0.2
0
0
5
10
No. of finite elements
15
FIGURE 6.14
Plot of the frequency ratio of the continuum model to the atomic model versus the number of continuum finite
elements for the first three low natural frequencies of a one-dimensional system.
198
frequencies of the continuum model are normalized with respect to those of the atomic
model. Natural frequencies of both models agree well as the number of finite elements
increased. As expected, the lower natural frequency requires a smaller number of finite
elements to match the frequency of the continuum domain to that of the atomic domain.
This example shows that using the smeared continuum model can reduce the degrees of
freedom with good accuracy compared to that needed in the atomic model.
6.5.2 Vibration of Atoms in 2-D Arrangement
This example considers atoms in a 2-D arrangement. The embedded atomic potential is
selected to represent the interatomic forces [8]. The parameters used for the present analy
sis were obtained from the reference. The atomic arrangement is assumed to be the hex
agonal arrangement as shown in Figure 6.15, which also shows four triangular shapes of
finite elements that represent the smeared continuum domain. Each triangular element
has three nodes at its vertices. The first case considers a uniform tensile static load. The
number of atoms varies from 400 to 3000, while the number of finite elements is held at
four. The difference in the displacements between the atomic model and FEM was less
than 3% for all the atoms considered.
The next case studies a vibrational analysis using the same atomic model and FEM as
shown in Figure 6.15. The same interatomic potential as before is used again to deter
mine the natural frequencies and mode shapes. Three typical mode shapes of the 2-D
model without constraints are plotted in Figures 6.16 through 6.18, which also compare
the mode shapes determined using the discrete atomic and smeared continuum models.
The first three vibrational mode shapes are tensile, shear, and bending modes. The figures
show that the three mode shapes determined by both models agree well with each other
between the atomic and smeared continuum models. The frequency associated to each
mode was also compared between the two models. The difference in the frequencies is
presented in Table 6.1. The difference lies between 5% and 8% for the three mode shapes.
Considering only four linear triangular elements were used for the present analysis, such
accuracy is acceptable for the smearing technique.
FIGURE 6.15
Two-dimensional atomic domain represented by a smeared domain using four linear triangular continuum
finite elements.
199
Coupling Techniques
Initial shape
Atomistic model
Deformed shape
Smeared continuum model
FIGURE 6.16
Axial mode vibration shape.
Initial shape
Atomic model
Deformed shape
FIGURE 6.17
Shear mode vibration shape.
Initial shape
Atomic model
FIGURE 6.18
Bending mode vibration shape.
Deformed shape
200
TABLE 6.1
Comparison of Frequencies between Atomic and Smeared Models
Frequency of Smeared Model/Frequency
ofAtomic Model
Vibrational Mode Shape

Axial mode
Shear mode
Bending mode
0.95
0.96
1.08
6.6Coupling Technique between Finite Element

and Lattice Boltzmann Models
Two different kinds of problems are discussed here. First, a wave propagation problem is
presented by coupling the FEM and the lattice Boltzmann method (LBM). Then, the cou
pling of fluid and structural domains is discussed.
6.6.1 Coupling Acoustic Domains
An acoustic domain consists of two parts. One part was modeled using the FEM, and
the other part was modeled using the LBM. One of the reasons to use multiple modeling
techniques for a single problem is to take advantage of each technique to solve the prob
lem. For example, the LBM technique can represent radiated or partially radiated bound
ary conditions elegantly. However, the technique in its classical formulation requires a
regular lattice structure. On the other hand, the FEM technique is useful for a complex
domain shape. As a result, the combination of the two techniques can take care of both
requirements.
The coupled domain problem is solved in a staggered manner with an exchange of data
at the interface of the two subdomains, which are modeled by two different methods. To
couple the two techniques, overlapping grid points are necessary at the FEM and LBM
interface. For the following 2-D wave equation,

2
2
2
2
=
c
+
s
(6.12)
2
t 2
y 2
x
the FEM uses the variable as the unknown at the FEM nodes, while the LBM uses fi as
the unknowns at the LBM grid points. The two variables are related to each other in the
following manner:

f (6.13)
i
When the transient solution data are transferred from the LBM to the FEM at the overlap
ping grid points, Equation 6.13 is utilized. However, transferring the transient solution
data from the FEM to the LBM is not unique. For this directional coupling, the following
technique is utilized: First, fi is computed as follows:
201
Coupling Techniques
fem i 0 (6.14)
2 n2
fi =
f o = 2
n2 1
fem (6.15)
n2
where the subscript fem denotes the FEM solution at the interface grid points. Then, for fi
that is normal to the interface boundary,
a modification is made to incorporate the flow of fi
from the neighbor. That is, ()ei is added to fi, in which is the difference between two
neighboring grids, one in the FEM subdomain and the other in the LBM subdomain, and
a constant is a function of the grid size, time step, and the relaxation constant. Finally, fi
is corrected so that their sum is equal to the FEM solution at the grid.
6.6.2 Coupling Fluid and Structural Domains
Structural analysis is commonly performed using the FEM, while the LBM is useful for
analysis of fluid flow. As a result, a multiphysics problem of fluid-structure interaction
(FSI) can be analyzed using the FEM for the structural behavior and the LBM for fluid flow.
Such a problem requires coupling between the FEM and the LBM. Because the FEM and
LBM have totally different kinds of unknown variables, the structural and fluid domains
are solved in a staggered manner instead of simultaneously. In other words, one domain
is solved first followed by the other domain. For a fully coupled problem, or a so-called
two-way coupling problem, the staggering procedure repeats until the end of the analysis.
If the same time step size is used for both the FEM and LBM, each analysis is conducted
once followed by the other analysis. However, the time step size for one analysis is n times
greater than that for the other analysis, and the latter analysis is performed n times while
the former analysis is conducted once to make sure there is consistency in time marching
in the computation.
The general scheme of coupling the structural motion to the fluid flow is sketched in
Figure 6.19. The structural analysis provides velocities to the fluid model at the FSI, while
the fluid analysis provides surface tractions to the structural model. Velocities and surface
tractions are explicit variables for the FEM. However, they are implicit variables for the
LBM. Surface tractions can be computed from the particle distribution functions, the fluid
density, and viscosity. However, the velocity obtained from finite element analysis should
Velocities
Tractions
Structural
Fluid
model
model
FIGURE 6.19
Exchange of information between the structural model and the fluid model.
202
be properly related to the particle distribution functions used for the LBM. More details
are provided next.
To couple the fluid domain and the structural domain, the following compatibility con
ditions are enforced at the fluid/structure interface: First, both models should have the
same tractions at the interface as follows:

klf nl = kls nl (6.16)
f
where kl is the stress tensor of the fluid, skl is the stress of the structure at the interface,
and nl is the normal unit vector at the interface. The fluid stress tensor is computed as
ij = pij + (vi,j + vj,i) (6.17)
in which p is pressure, is the viscosity, ij is the Kronecker delta, and vi,j is the fluid veloc
ity gradient. Equations 6.16 and 6.17 are used to calculate the surface tractions on the struc
ture from the fluid solution. After the interface surface traction is computed, it is applied to
the structural FEM through the nodal forces; for example:

{F} =
{} d (6.18)
in which

i = ijnj (6.19)
where [B] is the finite element matrix relating strains to nodal displacements, the matrix
depends on the element shape functions used in the FEM, and is the interface surface
or line depending on 3-D or 2-D analysis. If the fluid viscosity is very small or neglected,
there is only normal traction and no tangential traction.
The velocity compatibility condition is applied to the interface assuming no slip bound
ary condition at the interface as follows:

vi =
ui
(6.20)
t
where ui is the structural displacement at the interface, and vi is the fluid velocity. Equation
6.20 is used to provide the fluid velocity from the structural velocity solutions. However,
the LBM does not have fluid velocities as the primary unknown variables. Instead, the
density distribution function f k is the primary unknown variable for the LBM. As a result,
there is no unique way to apply the velocity compatibility condition to the LBM. For a 3-D
domain problem, the following equations can be used if ux, uy, and uz are the velocity
difference components between the structural velocity and the previous fluid velocity:

f1 = f1 + vx/3 (6.21)
f2 = f2 vx/3 (6.22)
f3 = f3 + vy/3 (6.23)
203
Coupling Techniques
f4 = f4 vy/3 (6.24)
f5 = f5 + vz/3 (6.25)
f6 = f6 vz/3 (6.26)
f 7 = f 7 + vx/24 + vy/24 + vz/24 (6.27)
f8 = f8 vx/24 + vy/24 + vz/24 (6.28)
f9 = f9 vx/24 vy/24 + vz/24 (6.29)
f10 = f10 + vx/24 vy/24 + vz/24 (6.30)
f11 = f11 + vx/24 + vy/24 vz/24 (6.31)
f12 = f12 vx/24 + vy/24 vz/24 (6.32)
f13 = f13 vx/24 vy/24 vz/24 (6.33)
f14 = f14 + vx/24 vy/24 vz/24 (6.34)
These equations satisfy the conservation of mass as well as the interface velocity continu
ity at the fluid domain.
6.7Coupling Technique between the Finite Element

and Cellular Automata Models
The coupling between the FEM and cellular automata (CA) model is shown for both the
acoustic problem and the acoustic-structure interaction problem.
6.7.1 Coupling Acoustic Domains
Both the CA and the FEM techniques have their own merits and limitations for modeling
wave propagation. For example, the CA technique is computationally efficient and can
model various boundary conditions easily, but it requires a regular grid structure, such
as a square or cubic shape. On the other hand, the FEM technique can model a complex
domain shape easily but is computationally expensive. As a result, it is useful to combine
both techniques for solving the wave propagation problem, and the coupling technique
is presented here. One part of the domain is modeled using the FEM and the other part
is modeled using the CA technique. The coupling between the FEM and CA technique is
undertaken as discussed next [9]. The acoustic solution s at the interface and its neighbors
are passed between the two techniques. The FEM analysis utilizes the CA solution at the
interface grid point directly. On the other hand, the CA calculation uses the FEM solution
204
at the grid point next to the interface to compute the CA solution at the interface grid
point using the local rule as shown for the 2-D case:

C(t + t) = (E(t) + W(t) + N(t) + S(t) 2C(t))/2 (6.35)

The procedure for the coupled CA and FEM analyses is as follows:
1. First, divide the physical domain into CA and FEM parts. Define the CA and FEM
nodes and make the nodes at the interface shared nodes.
2. Apply forces to the CA nodes, if any.
3. Apply CA rules to the CA subdomain. For calculating the next values of the nodes
at the interface (i.e., nodes shared between the CA and FEM subdomains), the
current values of two neighboring particles for each interface node are necessary.
Because these are the last CA node, there are no CA nodes on the FEM side. As a
result, the values from the nodes next to the interface of the finite element model
are borrowed.
[r+ (t)] = [r(t t)]FEM node next to the FEM interfface
[r (t)] = the CA node next to the interface CA node
(6.36)
[r(t + )]interface node of CA = [r+ (t)] + [r (t)] [r(t)]interface node of CA

4. At t = 0 (i.e., initially), the FEM data may require the solution at t = t for the cen
tral difference time integration scheme. It is assumed that the data at t = t are
equal to the data at t = 0.
5. Use the data of the CA interface nodes as the boundary conditions for the FEM
analysis.
6. Return to step 2.
During the procedure, the same time step size is used for both the FEM calculations and
the CA calculations so that FEM and CA calculations can be synchronized. The time step
size for the CA model was discussed in Chapter 4; the time step size for the FEM model
is more flexible as long as the stability condition is met. As a result, the time step size is
mostly controlled by the CA analysis. If the time step size for the CA analysis is greater
than the critical time step size for the FEM analysis in terms of the numerical stability
perspective, the CA time step size is divided by an integer such that the divided time
step size becomes smaller than the critical time step size for the FEM analysis. Then, the
divided time step size is used for the FEM analysis, which repeats in the same number as
the integer utilized previously so that the calculations in the CA and FEM analyses are
synchronized.
6.7.2 Coupling the Acoustic and Structural Domains
The structural domain is modeled using the FEM; the acoustic domain is modeled using
the CA method. The main variables for the structural FEM are nodal displacements; the
acoustic domain has velocity potential as the nodal variables. The CA model for the wave
205
Coupling Techniques
CA acoustic domain
FEM structural domain
FEM acoustic domain

FIGURE 6.20
Coupling of the acoustic and structural domains.
equation is simple to implement and can be easily adapted for a variety of nontrivial
boundary conditions.
To couple the acoustic and structural domains, the velocity and pressure are exchanged
between the two domains such as for the FSI problem. The acoustic domain passes the
pressure to the structure, which acts normal to the interface boundary, while the structural
domain passes to the normal velocity to the acoustic domain. When the CA technique is
used for the acoustic domain, the velocity is computed from the velocity potential using
the finite difference technique. However, the CA technique applies CA rules every other
time to each node as discussed in Chapter 4. As a result, it is not easy to couple the struc
tural FEM domain to the CA acoustic domain directly. Therefore, a small FEM acoustic
domain is introduced between the structural FEM domain and the CA acoustic domain
[10], as presented in Figure 6.20.
The FEM is computationally more expensive than the CA method for acoustic analy
sis, but the former can model a complex domain shape and be easily coupled with the
structural analysis without difficulty. As a result, the intermediate FEM acoustic domain
is useful for the present coupling problem. To maximize the computational efficiency of
the total analyses, the size of the FEM acoustic domain is minimized. In this scheme, the
coupling between the FEM structural and acoustic domains was discussed in Chapter 2;
the coupling between the FEM and CA acoustic domains was discussed in Section 6.7.1.
Therefore, the three analyses are conducted one after another. For example, the FEM struc
tural analysis is followed by the FEM acoustic analysis, which is also followed by the CA
analysis. Likewise, the data also flow in the reverse direction for the two-way coupling.
7
Multiscale Analysis of Composite Structures
7.1Introduction
Composite materials are made of two different materials that are mostly bonded physically
rather than chemically. Thus, any chemically compounded material generally is not called
a composite material. Most composite materials have strong and stiff materials bonded by
weak and soft materials. The former is called the reinforcing material; the latter is called
the binding material. There are many different shapes and forms of reinforcement. For
example, the reinforcement may be particles, short fibers, long fibers, or woven fibers.
Depending on the type of the reinforcement, those composites are called particulate composite, short-fiber composite, fibrous composite, or woven fabric composite, respectively.
Common fibers are glass fibers, carbon fibers, boron fibers, and others. Common binding
materials, called matrix materials, are resins, metals, and so on. Because of light weights
and low costs, resins are most commonly used. In addition, many composite structures are
fabricated with multilayers of the composites, such as laminated composites or sandwich
composites. Laminated composites have multiple layers, each of which is mostly made of
the same materials but has a different fiber orientation relative to one another. Sandwich
composites have top and bottom skin layers and a middle core layer. Skin materials are
strong and stiff, while the core material is soft and light.
Composite structures are good examples of structures consisting of multiple length
scales from the fiber scale to a structural scale. The former is mostly measured in microns,
and the latter is mostly measured in meters. In this chapter, computational techniques
are presented for multiscale analyses of composite structures. The emphasis of the techniques is exchange of critical information from one length scale to another so that the
overall analysis can use all the critical information to reliably predict the behavior of the
composite structures. The next section presents the overall strategy and main concepts for
multiscale analysis modeling of general composite structures, which is followed by necessary modeling schemes for composite materials. Then, more specific composite structures
are discussed.
7.2Multiscale Modeling of Composite Structures

The basic materials for composite structures are fibers (or particles) and a matrix material.
Failures in composite structures can be attributed to matrix cracking, fiber breakage, and
207
208
fiber/matrix interface debonding. For simplicity, the interface debonding may be included
as matrix cracking in the failure mode.
Matrix cracking can occur in various locations depending on the loading, boundary
condition, and so on. For example, matrix failure between two adjacent composite layers is
called interlayer delamination in the macroscale perspective. In addition, the matrix may
crack between fiber bundles within the same layer. If the crack is parallel to the fiber orientation, it is called fiber splitting. On the other hand, if the crack is perpendicular to the fiber
orientation, the failure is called transverse matrix cracking. All of these are nothing but
matrix failure. However, they are classified as different failure modes at the macroscale,
and different failure criteria are usually developed for those failure modes.
To understand and predict the failure in composite structures using a physics-based
approach rather than a phenomenological approach, the multiscale approach is important so that all the failures can be described in terms of the constituent materials, such
as the fibers and the matrix. In that aspect, the multiscale technique has been developed
[19].
The multiscale analysis of composite structures consists of two loops that complete
one cycle, and the cycle repeats itself. The loops are the preloop and the postloop
before the composite structural level analysis. The former may be called the stiffness
loop and the latter called the strength loop. The objective of the stiffness loop is to
compute the effective material properties of the composite structure so that the composite structure can be analyzed using the effective material properties. On the other
hand, the objective of the strength loop is to decompose effective stresses and strains at
the composite structure level into stresses and strains at the constituent material level
(i.e., fibers, particles, and matrix materials). Then, failures and damage are assessed
using the stresses and strains at the constituent material level to determine damage
and failure locally in the constituent materials. If there is damage or failure, degraded
material properties of the damaged or failed constituents are determined, and those
material properties are used for the next cycle of the stiffness loop. Those loops repeat
in cycle until the multiscale analysis is terminated. In general, the multiscale analysis
stops when the composite structure fails completely or the maximum applied load is
reached in an incremental manner.
Figure 7.1 illustrates the multiscale analysis described for the laminated woven fabric
composite structure. In the figure, the starting point is the mechanical properties of the
constituent materials. The multiscale technique consists of five different models, which
are represented by hexagons in the figure. Each model has two objectives. The fiber model
computes the effective material properties of the unidirectional composite layer from the
material properties of the fiber and matrix as well as their volume fractions. In addition,
the model can decompose the effective stresses and strains at the unidirectional composite
layer into those at the constituent material level.
Likewise, the fabric model has two similar objectives. The model can compute the effective material properties of the woven fabric composite layer from the material properties of
the unidirectional layer as well as decompose the stresses and strains at the latter level into
those at the former level. The lamination model computes the effective material properties
of the laminate from each individual layers properties as well as decomposes the stresses
and strains at the laminate level into those at each lamina level. To make the multiscale
analysis computationally efficient, all those models are analytical models, so they require
minimum computational time.
The major computational time involves the finite element model applied to the laminated
composite structure. This is a typical finite element analysis. As a result, the accuracy and
209
Preloop or stiffness loop
Mechanical
properties of
constituent
materials
Unidirectional
composite
properties
Woven-fabric
composite
properties
Laminated
composite
structural
properties
Finite
element
model
Damage
failure
model
Mechanical
properties
of fiber and
matrix
Unidirectional
stresses and
strains
Woven-fabric
stresses and
strains
Laminated
structural
stresses
and strains
Post loop or strength loop

FIGURE 7.1
Schematics of multiscale analysis of composite structures.
computational time depend on the mesh density and the number of iterations necessary to
reach a converged solution. For the transient problem, the number of time increments also
influences the computational time.
Depending on the nature of the composite structure, some models in Figure 7.1 may be
omitted. For example, if a composite structure is made of unidirectional layers with different layer orientations, the fabric model is skipped. In addition, if every layer is modeled
discretely, the lamination model is also skipped. In the following sections, each model is
discussed in detail.
7.3Fiber Model
The fiber model considers straight fiber bundles embedded in a matrix material. To make
the model applicable to cases that are more general, the fiber model was developed for
long-fiber, short-fiber, and particle-reinforced composites. In other words, one general
model can deal with each case [10,11].
To simplify mathematics, the cross-sectional shapes of the fibers or particles are assumed
to be rectangular in the fiber model. The representative unit cell for the model is shown in
Figure 7.2. The unit cell consists of eight subcells. The subcells are numbered from 1 to 8.
The dimensions of the unit cell are a by b by c. Without losing generality, let us assume the
fiber is oriented along the 1 axis in Figure 7.2. Dimensions of the ith subcell are denoted
by ai, bi, and ci. For example, the first subcell has dimensions of a1, b1, and c1, while the fifth
subcell has dimensions of a5, b5, and c5. Then, some subcells have the same dimension
along one axis, such as a1 = a3 = a5 = a7, a2 = a4 = a6 = a8, b1 = b2 = b5 = b6, b3 = b4 = b7 = b8, c1 =
210
6
6
2
7
5
5
5
1
a
1
b
FIGURE 7.2
Unit cell for fiber model.
c2 = c3 = c4, and c5 = c6 = c7 = c8. If particles are assumed to be a cubic shape, a1 = b1 = c1. On

the other hand, if fibers are assumed to have a square cross section, b1 = b2 = c1 = c2. For
long fibers, a1 and a2 are arbitrary; for short fibers or whiskers, a1 = c1, where is the aspect
ratio of the short fiber or whisker geometry. For convenience, the unit cell is assumed to
have a unit volume, and the specific values of the dimensions are determined based on the
inclusion volume fraction.
For mathematical simplicity, each subcell is assumed to have a constant strain and stress
state. In other words, the stress and strain at the nth subcell are denoted by tensors nij and
nij , respectively, where i and j vary 1 to 3, indicating the coordinate axes. Subcell stresses
must satisfy equilibrium at the subcell interfaces. To apply the formulation to nonelastic
problems, an incremental formulation is used in the development. The incremental stress
and strain tensors are denoted by nij and nij, respectively, for the nth subcell. There are 12
subcell interfaces total, with 4 interfaces in each axis. Then, the normal force equilibrium
at the 12 subcell interfaces can be expressed as

2
3
4
5
8
111 = 11
11
= 11
11
= 611 711 = 11
(7.1)
4
122 = 322 222 = 22
522 = 722 622 = 822 (7.2)
4
133 = 533 233 = 633 333 = 733 33
= 833 (7.3)
Likewise, tangential force equilibriums must be satisfied at the interfaces, and there are
two orthogonal tangential forces at each interface. For instance, if the interface is normal
to the 1 axis, there are two tangential forces along the 2 and 3 axes. Such tangential force
equilibrium yields the following expressions in terms of shear stresses:
211
2
3
4
5
8
112 = 12
= 12
= 12
12
= 612 = 712 = 12
(7.4)
2
5
3
4
8
113 = 13
= 13
= 613 13
= 13
= 713 = 13
(7.5)
4
123 = 323 = 523 = 723 223 = 23
= 623 = 823 (7.6)
The total number of equilibrium equations is 30, of which 12 are for the normal stress
components and 18 are for the shear stress components.
In addition to equilibriums in terms of stresses, the subcells need to deform compatibly.
For example, the length change of the subcells along each axis should be equal. In other
words, the total length change in subcells 1 and 2 should be equal to that in subcells 3 and
4 along the 1 axis. Such compatibility can be written in terms of the subcell normal strains
as

2
3
4
5
8
a1111 + a2 11
= a311
+ a411
= a511
+ a6611 = a7 711 + a811
(7.7)
4
a1122 + a3 322 = a2 222 + a4 22
= a5 522 + a7 722 = a6622 + a8 822 (7.8)
3
4
a1133 + a5 533 = a2 233 + a6633 = a3 33
+ a7 733 = a4 33
+ a8 833 (7.9)
Similarly, there is compatibility among subcell shear strains. The average shear strain of
subcells 1, 2, 3, and 4 is assumed to be equal to that of subcells 5, 6, 7, and 8 for the 12 component. These relations can be expressed in the same manner as the following:

2
3
4
8
a1b1112 + a2b2 12
+ a3b312
+ a4b412
= a5b5152 + a6b6612 + a7 b7 712 + a8b812
(7.10)
2
5
4
8
a1c1113 + a2 c2 13
+ a5c513
+ a6c6613 = a3c3 133 + a4c 413
+ a7 c7 713 + a8c813
(7.11)
4
b1c1123 + b3c3 323 + b5c5 523 + b7 c7 723 = b2 c2 223 + b4c4 23
+ b6c6 623 + b8 c8 823 (7.12)
The number of compatibility equations is 12, with 9 equations for the subcell normal
strains and 3 for the subcell shear strains. The sum of the equilibrium and compatibility
equations is 42. The normal components have 21, and the shear components have 21.
Each subcell has its own constitutive equation as follows:

n
nij = Dijkl
nkl nkl (7.13)
n
where nkl is the thermal expansion coefficients, is the temperature change, and Dijkl
is
the material stiffness matrix. In general, every subcell can have different material properties. For elastic deformation of fibrous and particulate composites, there are two different material properties: the fiber and particle property or the binding matrix properties.
Therefore, each subcell will use either of the material properties. However, for elastoplastic
deformation, the tangent modulus of every subcell may be different.
212
Let us consider a composite made of elastic whiskers and an elastoplastic matrix material, such as a metal matrix composite with carbon whiskers. Then, subcells representing
the matrix material, for instance, subcells 2 through 8, may have different states of elastoplastic deformation at the subcell levels. Then, the tangent moduli representing different elastoplastic states will be different for each subcell representing the matrix material.
To simplify it, the average state of stresses in the matrix material subcells can be used to
determine elastoplastic deformation of the unit cell level, and the corresponding tangent
modulus can be used for every matrix subcell.
There are six constitutive equations for each subcell, resulting in 48 equations for the
eight subcells. So far, the sum of all the previous equations is 90 (i.e., sum of 30 equilibrium, 12 compatibility, and 48 constitutive equations). The number of unknowns is 96, half
of which are stresses and the other half are strains at the eight subcells. That means we
need more equations. The next set of equations comes from the relation between the unit
cell and subcell-level stresses and strains. The stresses and strains at the unit cell level are
assumed to be the volume averages of those at the subcell level. That is,
8
ij =
v
n
n
ij
(7.14)
n
ij
(7.15)
n= 1
8
ij =
v
n
n= 1
in which the superimposed bar denotes stresses and strains at the unit cell level, and vn is
the volume fraction of the nth subcell, which is computed as

vn =
anbn cn
(7.16)
abc
Furthermore, there are constitutive equations at the unit cell level, which can be
expressed as

ij = Dijkl (kl kl ) (7.17)
Equations 7.14, 7.15, and 7.17 result in 18 additional equations, and there are 12 additional
unknowns, which are the unit cell stresses and strains, ij and ij , respectively. Finally,
the total number of equations is equal to the total number of unknowns, which is 108.
Those equations can be solved together to find the material properties at the unit cell level,
that is, the smeared effective composite material properties of Dijkl and kl as shown in
Equation7.17. If every subcell has either an isotropic or an orthotropic material property,
the normal and shear components are decoupled. As a result, they can be solved independently. Most composites have either isotropic or orthotropic fiber and matrix materials.
The sets of equations presented provide two important mathematical expressions. The
first is that the smeared effective composite material properties of Dijkl and kl can be
n
determined from the subcell-level data, such as Dijkl
, nkl, and the geometric information (an,
bn, and cn). Furthermore, the subcell-level stresses and strains nij and nij, respectively, can
be computed from the stresses and strains at the unit cell level, ij and ij , respectively.
213
These two functions are necessary to complete the preloop (i.e., stiffness loop) and the
postloop (i.e., strength loop) as sketched in Figure 7.1. The detailed mathematical expressions for the two final expressions are given in Reference 9.
For orthotropic constituent materials (i.e., for the fiber and matrix materials), some more
detailed mathematical expressions are provided for the normal stress and strain components. Similar expressions can be developed for shear stress and strain components. First,
the subcell constitutive equations as shown in Equation 7.13 are substituted into the 12
interface equilibrium equations, Equations 7.1 through Equation 7.3, so that the latter 12
equations can be expressed in terms of subcell strain components. In addition, nine compatibility equations from Equations 7.7 through 7.9 as well as Equation 7.15 yield the following matrix expression:
[T]{}N = {f} (7.18)
n
where [T] is the 24 24 matrix consisting of Dijkl
and the geometric dimensions of an, bn,
and cn. Furthermore, {}N is the column vector of 24 1 containing three normal strains of
the eight subcells, and {f} is also a column vector, which can be expressed as
{ f }T = {}T {0}T { }TN (7.19)
Here, {} is the 12 1 column vector containing the components of thermal expansion

coefficients, {0} is the null vector, and { } N is the 3 1 vector consisting of three normal
strains at the unit cell level.
Solving Equation 7.18 yields
{} N = [R]{ f } = [R1 ]{} + [R2 ]{ } N (7.20)

where
[R] = [T]1 = [[R1] [0] [R2]] (7.21)

and [R1] is a 24 12 matrix, and [R2] is a 24 3 matrix.
The constitutive equations at the subcell level as shown in Equation 7.13 can be expressed
in terms of the matrix form as follows:
{}N = [E]N{}N {}N (7.22)
Equation 7.14 can also be written as
{ } N = [V ]N { } N (7.23)
in which the subscript N also denotes only the normal components.

Substitution of Equation 7.20 into Equation 7.22 and the resulting expression into
Equation 7.23 yields

{ } N = [V ]N [E]N [R2 ]{ } + [V ]N [E]N [R1 ]{} [V ]N {} N (7.24)
214
Let us rewrite Equation 7.17 as

{ } N = [E ]N ({ } N {} N ) (7.25)
Comparing Equation 7.2 to Equation 7.25 shows that the effective material property
matrix at the unit cell level [E ]N can be expressed as
[E ]N = [V ]N [E]N [R2 ] (7.26)
and the effective thermal property matrix becomes

{} N = [E ]1[V ]N ([E]N [R1 ]{} {} N )/ (7.27)
These equations suggest that the effective mechanical property expressed in Equation 7.26
is independent of the thermal property, but the effective thermal property shown in Equation
7.27 depends on the mechanical properties at the subcell level. Two equations, Equations 7.26
and 7.27 are used for the preloop analysis; Equations 7.20 and 7.22 are used for the postloop
to decompose the strains at the unit cell level into subcell-level strains and stresses.
A composite can have thermal stresses at the subcell level while it deforms freely at the
unit cell level. For example, let us consider a uniform temperature change in the unit cell
level. Then, the thermal strain at the unit cell level is
{ } N = {} N (7.28)
Substitution of Equation 7.20 into Equation 7.22 gives

{ } N = [E]N ([R1 ]{} + [R2 ]{ } N ) {} N (7.29)
Plugging Equation 7.28 into Equation 7.29 yields the thermal stresses at the subcell level:

{ } N = [E]N ([R1 ]{} + [R2 ]{} N ) {} N (7.30)
7.4Fabric Model
The fiber model is general so that it can be applicable for different types of reinforcement. However, different fabric models should be developed for different types of fabric
structure. For example, plain weave and 2/2 twill composites are shown in Figure 7.3. The
fabric model relates material properties of a unidirectional strand to the effective material
properties of a plain weave composite. Like the fiber model, this model has two functions:
computation of the effective properties of the plain weave composite using the strand
materials and their weaving information for the preloop and decomposition of the woven
strains and stresses into the strand strains and stresses for the postloop.
215
Warp
Warp
Fill
Fill
(a)
(b)
FIGURE 7.3
Fabric patterns for (a) plain weave and (b) 2/2 twill composites.
7.4.1Fabric Model for Plain Weave Composite

To develop the fabric model, it is necessary to observe any representative unit cell of the
fabric. For the plain weave fabric composite as shown in Figure 7.3a, the unit cell model is
shown in Figure 7.4. The unit cell for the plain weave consists of 13 subcells. Most of those
subcells in Figure 7.4 have the fiber orientations along the 1 or 2 axis [5,7]. In other words,
subcells 1, 9, 11, and 12 have fibers along the 1 axis, while subcells 3, 7, 13, and 10 have fibers
along the 2 axis. On the other hand, four subcells (2, 4, 6, and 8) have fibers in inclined orientations, and subcell 5 is filled with the matrix material.
First, normal stress/strain components are discussed. There are three normal strains for
n
each subcell, for a total of 39 strains str

, n = 1, 2 , , 13 . There are three normal strains
kl
at the unit cell level wf
kl . Those strains represent the average values of each subcell or the
unit cell, respectively, as assumed in the fiber model. To relate those strains, equilibrium
and compatibility conditions are applied to the 13 subcells. At the interface between any
( )
( )
t2
13
12
t1
5
1
10
a2
11
b2
1
a1
b1
FIGURE 7.4
Unit cell geometry for plain weave composite.
a1
a2
216
two neighboring subcells, the normal subcell stresses to the interface plane should be in
equilibrium as follows:

( ) = ( ) , ( ) = ( ) , ( ) = ( ) , ( ) = ( )
( ) = ( ) , ( ) = ( ) , ( ) = ( ) , ( ) = ( )
( ) = ( ) , ( ) = ( ) , ( ) = ( ) , ( ) = ( )
( ) = ( ) , ( ) = ( ) , ( ) = ( ) , ( ) = ( )
( ) = ( ) , ( ) = ( ) , ( ) = ( ) , ( ) = ( )
str
11
str
11
str
22
str
11
str
22
str
22
str
33
str
11
str
22
str
33
str
11
str
11
12
str
22
10
str
33
str
22
str
22
11
str
33
str
33
str
22
str
11
str
22
str
22
13
str
22
12
str
33
str
11
str
11
str
22
10
str
11
str
11
10
str
22
str
11
13
str
11
str
22
11
str
11
str
33
str
11
str
11
12
str
22
str
22
str
33
11
(7.31)
(7.32)
(7.33)
(7.34)
13
(7.35)
where the superscript numbers denote subcell numbers in Figure 7.4, and the subscript
numbers indicate the stress components. The superscript str is used to differentiate the
subcells in the fabric model from the subcells in the fiber model.
The strain compatibility conditions among subcells are as follows:
( ) + ( )
1
str
11
( )
str
a1 11
str
a1 11
10
( )
( )
str
+ b1 11
( )
str
+ b1 11
11
str
11
( )
6
str
+ a1 11
str
= 11
( ) + ( )
str
= 22
( )
str
a2 22
str
a2 22
str
22
10
( )
( )
str
+ b2 22
( )
str
+ b2 22
str
22
12
( )
13
str
11
( )
str
+ a2 22
( ) + ( )
str
22
( )
str
t1 33
str
22
13
str
+ b1 11
( )
str
+ b1 11
str
+ a1 11
( )
str
+ a1 11
13
(7.37)
(7.38)
( ) + ( )
(7.40)
( )
12
str
11
str
= a2 22
str
22
10
( )
str
+ b2 22
( )
str
+ b2 22
( ) + ( )
str
= 22
10
( )
(7.39)
str
= a2 22
( )
str
+ t2 33
(7.36)
( ) + ( )
( )
str
+ a2 22
10
( )
( )
str
= a1 11
( ) + ( )
7
str
11
str
= a1 11
str
11
( )
str
+ a1 11
( )
( ) + ( )
str
= 11
str
22
11
( )
str
+ a2 22
( )
str
+ a2 22
13
(7.41)
(7.42)
(7.43)
( )
str
= (t1 + t2 ) 33
(7.44)
217
( )
( )
str
t1 33
str
t1 33
( )
str
= (t1 + t2 ) 33
( )
str
= (t1 + t2 ) 33
11
str
+ t2 33
11
str
+ t2 33
( )
( )
(7.45)
(7.46)
( ) = ( ) (7.47)
( ) = ( )
str
33
str
33
str
33
( )
str
+ t2 33
( )
str
+ t2 33
str
t1 33
str
t1 33
str
t1 33
( )
str
33
(7.48)
( )
str
= (t1 + t2 ) 33
( )
( )
str
= (t1 + t2 ) 33
( )
str
= (t1 + t2 ) 33
12
( )
12
( )
13
str
+ t2 33
(7.49)
(7.50)
(7.51)
Here, ai, bi, and ti are geometric dimensions of the plain weave unit cell as shown in
Figure 7.4. Finally, the unit cell strain and stress are assumed to be equal to the volume
averages of the subcell strains and stresses, respectively. That is,
13
wf
ij
v ( )
n
str
ij
(7.52)
n= 1
13
wf
ij =
v ( )
n
str
ij
(7.53)
n= 1
Here, vn is the volume fraction of the nth subcell. The constitutive equation at each subcell level can be written as
( ) = (E ) ( )
str
ij
str
ijkl
str
kl
(i, j, k , l = 1, 2 , 3) (7.54)
where the summation sign convention is applied only to the subscripts k and l. For each
n
str
should be determined based on the orientation of the strand inside the subsubcell, Eijkl
cell. Algebraic manipulation of the previous equations, as was done for the fiber model,
yields the following relationships:
( )
((
wf
str
Eijkl
= f Eijkl
( )
str
ij
, ai , bi , ti (7.55)
(7.56)
= g wf
ij , ai , bi , ti
218
Equations 7.55 and 7.56 are used for the preloop and postloop of the multiscale analysis
as the fabric model. Equation 7.55 computes the effective material properties of the plain
weave fabric composite based on the material properties of the strands and weaving geometry, such as Equation 7.26. Equation 7.56 calculates the strains at the subcell level from the
unit cell level similarly as for Equation 7.20. Once the subcell strains are computed, subcell
stresses are computed from the constituent equations at the subcell level.
A similar derivation can be made for shear components. For example, the shear component parallel to the 12 plane can be derived as presented next. The shear stress equilibrium at the subcell interfaces is written as
( )
str
t1 12
str
t1 12
str
t1 12
str
t1 12
str
t1 12
str
t1 12
str
t1 12
( )
( )
str
+ t2 12
str
= (t1 + t2 ) 12
( )
str
= (t1 + t2 ) 12
( )
str
= (t1 + t2 ) 12
( )
str
= (t1 + t2 ) 12
str
+ t2 12
( )
str
+ t2 12
( )
str
+ t2 12
( )
( )
( )
str
= (t1 + t2 ) 12
( )
11
str
+ t2 12
( )
10
10
13
12
( )
( )
( )
( )
( )
str
= (t1 + t2 ) 12
( )
str
= (t1 + t2 ) 12
str
+ t2 12
str
+ t2 12
13
13
(7.57)
(7.58)
(7.59)
(7.60)
(7.61)
( ) (7.62)
8
( ) (7.63)
6
( ) = ( ) . (7.64)
str
12
str
12
The assumed strain compatibility expressions are

( ) = ( )
( ) = ( )
( ) = ( )
( ) = ( ) . (7.68)
str
12
str
12
str
12
str
12
str
12
str
12
str
12
str
12
10
11
12
(7.65)
(7.66)
(7.67)
219
Combining Equations 7.57 through 7.68 and Equations 7.52 through 7.54 for the shear
components yields the expressions for the shear components, which are similar to
Equations 7.55 and 7.56.
7.4.2Fabric Model for 2/2 Twill Composite
The fabric pattern of a 2/2 twill composite is shown along with its representative unit cell
[6] in Figure 7.5. The unit cell for the 2/2 twill composite has 77 subcells; some are considered to have the same average stresses. To confirm the assumed states of stress, a finite
element model was run on several materials under a uniform displacement to validate this
assumption. Based on observation of the numerical results, the 2/2 twill composite unit
cell is assumed to have 17 different states of stresses as shown in Figure 7.6. As a result,
17 independent subcells are used in the unit cell model. Fifty-one linearly independent
equations were developed that relate the normal stresses and normal strains of the 17
independent stressed subcells. The average stresses and strains of each subcell are used in
all the subsequent equations. The first set of equations represents the stress equilibrium at
the subcell interfaces. Applying equilibrium to any two neighboring subcells results in the
normal stress equations that follow, for which the subscripts indicate stress components
and the superscripts designate the subcell numbers. Equations 7.69 to 7.71 are a set of normal stress equilibriums in the 1, 2, and 3 directions respectively.
( ) = ( ) ; ( ) = ( ) ; ( ) = ( ) ; ( ) = ( )
( ) = ( ) ; ( ) = ( ) ; ( ) = ( ) ; ( ) = ( )
str
11
str
11
str
11
str
11
15
str
11
str
11
str
11
15
14
str
11
( ) = ( ) ; ( ) = ( )
str
11
str
11
16
str
11
str
11
str
11
17
str
11
str
11
str
11
10
str
11
11
str
11
str
11
str
11
str
11
( ) = ( ) + ( ) ; ( ) = ( )
str
11
14
str
11
12
( )
str
+ 11
13
Unit cell
FIGURE 7.5
Unit cell for 2/2 twill composite.
str
11
str
11
(7.69)
220
a
(a)
2
15
10
14
13
(b)
13
16
11
15
10
14
13
15
17
12
11
17
11
12
16
10
17
12
16
14
Vertical strands
Horizontal strands
FIGURE 7.6
Unit cell model for 2/2 twill woven fabric composite: (a) unit cell sketch and (b) subcells inside unit cell.
( ) = ( ) ; ( ) = ( ) ; ( ) = ( ) ; ( ) = ( )
( ) = ( ) ; ( ) = ( ) ; ( ) = ( ) ; ( ) = ( )
str
22
str
22
str
22
str
22
16
17
str
22
str
11
17
str
22
11
str
11
( ) = ( ) ; ( ) = ( )
str
22
str
22
14
str
22
15
str
22
str
22
str
22
str
22
str
22
str
22
13
str
22
( ) = ( ) + ( ) ; ( ) = ( ) + ( )
str
22
12
str
22
str
22
str
22
str
22
str
22
16
str
22
str
22
10
(7.70)
221
( ) = ( ) ; ( ) = ( ) ; ( ) = ( )
2
str
33
str
33
str
33
( ) = ( ) ; ( )
8
str
33
str
33
str
33
11
str
= 33
str
33
str
33
( ) ; ( )
10
str
33
str
33
( )
12
str
= 33
13
(7.71)
;
The following equations represent the directional strain compatibility assuming uniform deformation of the unit cell, where a and h are the dimensions of the fill and warp
strands, respectively, of the composite material shown in Figure 7.6:
((
str
2 a 11
) + ( ) ) + h (( ) + ( ) )
2
((
str
= 2 a 11
((
2 a ((
str
2 a 11
str
11
((
2 a ((
str
11
str
11
((
((
str
11
10
0
12
str
11
13
str
11
14
str
11
str
11
16
str
22
11
str
11
((
str
22
12
str
22
str
22
13
str
22
10
str
22
str
22
str
22
((
12
str
22
str
22
str
22
str
22
13
(7.72)
str
22
( ) + ( )
str
+ 22
16
str
22
) + ( ) + ( ) + ( )
6
str
11
) + ( ) ) + h ( )
str
= 2 a 22
) + ( ) ) + h (( )
str
22
) + 3h (
) + ( ) ) + h (( ) + ( )
str
22
17
str
11
str
11
13
11
str
22
) + ( ) ) + h ( )
2
15
str
11
( ) + ( )
str
= 11
( ) + ( ) = ( ) + ( )
6
) + ( ) ) + h ( )
str
= 2 a 11
) + ( ) ) + h (( ) + ( )
= a 2str2
str
22
str
11
) + ( ) + ( ) + ( )
str
11
((
str
22
str
11
) + ( ) ) + h (( ) + ( ) + ( )
2
str
= 2 a 22
str
2 a 22
str
11
str
11
10
((
str
11
) + ( ) ) + h (( ) + ( ) )
( ) + ( )
str
2 a 22
str
11
) + ( ) ) + h ( )
str
= a 11
str
11
str
11
14
str
22
15
17
str
22
) + 3h (
str
22
12
(7.73)
The constitutive equation for every subcell is expressed as shown in Equation 7.74.

( ) = (E ) ( ) (7.74)
str
ij
str
ijkl
str
kl
222
( )
( ) is the
subcell material property matrix in terms of the unit cell axes. The stiffness matrix ( E ) is
where ijstr
( )
and str
kl
str
are the nth subcell stresses and strains, respectively, and Eijkl
str
ijkl
determined from stress and strain equations in conjunction with the proper transformation
matrices depending on the orientation of the strand direction, as discussed in the next section.
The final equations describe the relationship between the woven fabric unit cell stresses
(strains) and the strand subcell stresses (strains). The woven fabric unit cell stresses and
strains are computed as the volumetric average of subcell stresses and strains.
17
wf
ij
V ( )
str
ij
(7.75)
n= 1
17
wf
ij
V ( )
n
str
ij
(7.76)
n= 1
where Vn is the volume fraction of the nth subcell. Finally, the following relationships are
developed:
((
wf
str
Eijkl
= f Eijkl
( )
str
ij
, a, h, t (7.77)
(7.78)
= g wf
ij , a , h, t
These two equations provide the bidirectional passage of the fabric model. Equation 7.77
wf
is used to compute the effective woven fabric stiffness Eijkl
of a 2/2 twill composite from the
n
str
stiffness Eijkl
of the unidirectional strand and the weave geometric dimensions a,h, andt,
which is the thickness of the strand thickness. In addition, Equation 7.78 decomposes the
2/2 twill composite strains into the subcell strains. The subcell stresses can then be calculated using Equation 7.74.
( )
7.5Coordinate Transformation
Because of weaving, fiber strands must be undulated. Typical undulation angles are 10
to 15. The undulation angle is measured from the global coordinate axes. The undulated
portions of the fiber strands must have their material properties adjusted to that of the
global coordinate system. A brief explanation of stiffness transformation follows.
Let the coordinate system X, Y, and Z be the global coordinate system and x1, y1, and z1
be the local coordinate system where the fiber strands of interest are aligned with the local
axis as shown in Figure 7.7. The stress and strain transformations from the local coordinate
system to the global coordinate system are
{ } x1y1z1 = [T ]{ }XYZ
{} x1y1z1 = [T ]{}XYZ
(7.79)
223
y1
x1
X
FIGURE 7.7
Global and local coordinate axes.
The constitutive equations are

{ }XYZ = [C]XYZ {}XYZ
{ } x1y1z1 = [C]x1y1z1 {} x1y1z1
(7.80)
Mathematical manipulation results in the following stiffness transformation equations:

{ } x1y1z1 = [C]x1y1z1 [T ]{}XYZ
[T ]{ }XYZ = [C]x1y1z1 [T ]{}XYZ
{ }
XYZ
x1y1z1
= [T ] [C]
[T ]{}
(7.81)
XYZ
{ }XYZ = [C]XYZ {}XYZ

where
[C]XYZ = [T ]1 [C]x1y1z1 [T ] (7.82)
Finally, the transformation matrices are as follows:
m2
2
n
[T ] = 0
0
0
2 mn
n2
m2
0
0
0
2 mn
0
0
1
0
0
0
0
0
0
m
n
0
0
0
0
n
m
0
mn
mn
0
0
0
m 2 n2
(7.83)
224
m2
2
n
[T ] = 0
0
0
mn
n2
m2
0
0
0
mn
0
0
1
0
0
0
0
0
0
m
n
0
0
0
0
n
m
0
2 mn
2 mn
0
0
0
2
m n2
(7.84)
The matrices [T] and [T] are the stress and strain transformation matrices, respectively;
m and n are the cosine and sine of the angle between the global and local axis, respectively.
When the local axis rotates in the counterclockwise direction from the global axis, it is
considered positive.
7.6Lamination Model
The lamination model may or may not be needed depending on the finite element analysis
model used in the study. If each layer is modeled discretely in the finite element analysis,
the lamination model is not necessary. For such an analysis, the finite element model contains a large number of elements. If generic three-dimensional (3-D) solid elements are
used to model individual layers, the total number of elements will be extremely large for
a general size composite structure. Instead, solid-like plate/shell elements can be used to
represent each layer as needed [10]. The element was presented previously in Chapter 2, so
it is not repeated here. The element can model each layer with a reasonable number of total
elements in the finite element analysis.
Another way to model a multilayer composite is by applying different material properties in each layer (e.g., due to different orientations of fibers in layers) at the numerical integration points of a plate/shell element. In that case, the number of numerical integration
points through the thickness of the plate/shell element should be at least equal to the number of layers, and the integration points should be properly spaced to represent the layers.
If every layer has the same thickness, the integration points should be spaced equally.
If the plate/shell element requires bending stiffness properties directly instead of basic
properties such as elastic moduli and Poisson ratios, the lamination plate/shell formulation must be used to compute the equivalent stiffness. For instance, the laminated plate
theory has the following constitutive equation:

N A
= T
M B

(7.85)

where N and M are in-plane force and bending moment vectors, respectively; and and
are the in-plane strain and curvature vectors, respectively. Furthermore, the material
property matrix in this equation is computed from the summation of each layers properties. Let the ith layer have the following stress-strain relation:
{}i = [E]i {}i (7.86)
225
Then, we obtain
n
[ A] =
[E] (z z
i
i1
) (7.87)
i=1
[B] =
1
2
[D] =
1
3
[E] ( z
i
2
i
zi21 (7.88)
i=1
[E] ( z
i
3
i
zi31 (7.89)
i=1
in which zi is the coordinate value along the plate thickness of the ith layer, and z0 and zn
are the bottom and top coordinate values of the laminated plate, respectively.
These equations are used for the preloop. For the postloop, if the in-plane strains and
curvatures are computed, the strain at the ith layer is computed as
{}i = [E]i ({} + z{}) (7.90)
This completes the full cycle of the lamination model. In the lamination model, the coordinate transformation equation, Equation 7.82, is used to find the material property matrix
of the layers whose fibers are not oriented in the global coordinate axes.
7.7Examples of the Fiber Model

In this section, both particulate and fibrous composites are considered for their effective
stiffness and strength and their examples are presented [11,12].
7.7.1Particulate Composite Made of Al2O3 Particle/Aluminum Matrix
A particulate composite is made of Al2O3 particles and the aluminum matrix. The particles
have a Youngs modulus of 350 GPa and a Poisson ratio of 0.3; the matrix material has a
Youngs modulus of 70 GPa and a Poisson ratio of 0.3. The particle volume fraction (PVF)
is 0.3. Results obtained using the fiber model are shown in Table 7.1, which also compares
TABLE 7.1
Comparison for Particulate Composite Made of Al2O3 Particle/Aluminum
Matrix with a Particle Volume Fraction of 0.3
Elastic Modulus
Poisson Ratio
Shear Modulus
110 GPa
105 GPa
116 GPa
0.27
0.29
0.27
37 GPa
41 GPa
Present
Reference 13
Reference 14
226
the results to other results. Those include the Mori-Tanaka method [13] based on Eshelbys
inclusion theory for a spherical inclusion as well as Yu and Tangs variational asymptotic
method for a cubic inclusion [14]. The fiber model provides results comparable to both
models.
7.7.2Particulate Composite Made of SiC Particle/Aluminum Matrix
For the example of a particulate composite made of SiC particle and aluminum matrix,
Youngs moduli of 450 and 70 GPa and Poisson ratios of 0.22 and 0.3 were used for the
SiC particle and the aluminum matrix, respectively. Elastic moduli are plotted for various
PVFs in Figure 7.8, where the values of elastic moduli from the data published by commercial manufacturers are compared [15]. The data are available only up to a PVF of 0.25.
Figure 7.9 compares the present predicted results to experimental values [16] as well as the
Mori-Tanaka results [13] for the whole range of PVFs. The present method predicted effective elastic moduli with acceptable accuracy.
7.7.3Fibrous Composite Made of Boron Fibers and Aluminum Matrix
Boron fibers have an elastic modulus of 379.3 GPa and a Poisson ratio of 0.1; the aluminum
matrix has an elastic modulus of 68.3 GPa and a Poisson ratio of 0.3. The unidirectional
composite has a fiber volume fraction of 0.47. Results of the fiber model were compared to
other results as shown in Table 7.2. Two results were compared from Reference 14, with one
the finite element solution and the other the variational asymptotic solution. In addition,
the comparison includes an asymptotic homogenization solution developed by Oliveira et
al. [17] as well as the experimental data obtained by Kenaga et al. [18]. The present solution
agrees well with these other solutions.
120
Ref. 15
Ref. 13
Present
Elongational modulus (GPa)
100
80
60
40
20
0
0.05
0.1
0.15
Particle volume fraction
FIGURE 7.8
Comparison of elastic modulus for SiC particle and aluminum matrix.
0.2
0.25
227
600
Ref. 16
Ref. 13
500
Present
400
300
200
100
0
0.0
0.1
0.2
0.3
0.4
0.5
0.6 0.7
Particle volume fraction
0.8
0.9
1.0
FIGURE 7.9
Comparison of elastic modulus for SiC particle and aluminum matrix for the whole range of PVF.
TABLE 7.2
Comparison of Effective Moduli for Boron Fiber/Aluminum Matrix Composite
Models
Present
Reference 14 (FEM)
Reference 14 (analytical)
Reference 17
Reference 18
E11 (GPa)
E 22 (GPa)
G 12 (GPa)
G 23 (GPa)
12
23
215.1
215
215.3
214.6
216
142.6
144
144.1
144.5
140
51.3
57.2
54.4
54.7
52
43.7
45.9
45.3
46.2
0.20
0.19
0.20
0.19
0.29
0.25
0.29
0.26
0.25
Note: FEM, finite element method.
7.7.4Fibrous Composite Made of Graphite Fiber and Epoxy Matrix

Graphite fibers are transversely orthotropic material. The longitudinal and transverse
elastic moduli of the graphite fibers are EL = 235 GPa and ET = 14 GPa, respectively. The
shear modulus is GLT = 28 GPa, and the Poisson ratios are LT = 0.2 and TT = 0.25. Here,
the subscripts L and T denote the longitudinal and transverse directions, respectively.
Thematrix is an isotropic resin material with an elastic modulus of 2.48 GPa and a Poisson
ratio of 0.34. The fiber volume fraction is 0.6. Table 7.3 shows the results using the present
model and the finite element and analytical solutions from Reference 14. All the results are
in good agreement.
7.7.5Short-Fiber Composite Made of SiC Whiskers and Aluminum Matrix
This study examined whiskers, which are short fibers. Aspect ratios from 1.0 to 20 were
considered, with the same whisker volume fraction of 0.2. Figure 7.10 compares the present
228
TABLE 7.3
Comparison of Effective Moduli for Graphite Fiber and Aluminum Matrix Composite
Models
Present
Reference 14 (FEM)
Reference 14 (analytical)
E11 (GPa)
E 22 (GPa)
G 12 (GPa)
G 23 (GPa)
12
23
142.9
142.6
142.9
9.61
9.60
9.61
5.45
6.00
6.10
3.03
3.10
3.12
0.25
0.25
0.25
0.35
0.35
0.35
2.5
Composite modulus/matrix modulus
2.25
2
1.75
1.5
Ref. 19 (experimental)
Ref. 19 (analytical)
Present
Continuous fiber
1.25
1
0.75
0.5
0.25
0
11
Particle aspect ratio
16
FIGURE 7.10
Comparison of normalized elastic modulus for SiC whiskers and aluminum matrix with a whisker volume
fraction of 0.2.
solution against the Eshelby model and experimental data [19]. The present model is closer
to the experimental values, and the solution approaches the upper dashed horizontal line,
which represents the case of continuous SiC fiber and aluminum matrix composite as the
aspect ratio increases.
7.7.6Porous Material Made of Al2O3
A material with porosity can also be modeled using the fiber model. As an example, alumina with microvoids was investigated. The elastic modulus and Poisson ratio for the
99.9% alumina were 370 GPa and 0.22, respectively. Voids were treated as fictitious particles with a very small elastic modulus, such as 10 6 to 10 20, compared to that of the alumina. Figure 7.11 shows the present results, which were compared with the Mori-Tanaka
model [13] and experimental values [20]. In the present analysis, the aspect ratio of the
void can be modeled as discussed previously. A narrow and long void (an aspect ratio of
2) along the loading direction gives a higher modulus in the longitudinal direction than in
the transverse direction. The present results agree well with the experimental data until a
void volume fraction of 0.4.
229
400
Ref. 20
Ref. 13
Present, AR=1
300
Present, AR=2, long.
Present, AR=2, trans.

200
100
0
0.0
0.1
0.2
0.3
0.4 0.5 0.6 0.7

Void volume fraction
0.8
0.9
1.0
FIGURE 7.11
Comparison of the elastic modulus for alumina with the porosity. AR, aspect ratio.
7.7.7Thermal Stress of Glass Fiber/Epoxy Matrix Fibrous Composite

Because of the mismatch in the coefficients of thermal expansions (CTEs) between the fiber
and matrix materials, thermal stresses occur at the constituent materials. The composite is
made of glass fibers with an elastic modulus of 72.38 GPa, a Poisson ratio of 0.2, and a CTE
of 5.0 10 6 m/m/K. The matrix material had an elastic modulus of 2.75GPa, a Poisson
ratio of 0.35, and a CTE of 54.0 10 6 m/m/degree K. The CTEs of the fiber composite
obtained by the fiber model were compared with the analytical solutions in Reference
21, as shown in Figure 7.12. The longitudinal CTEs of the two results are almost identical,
but the transversal CTEs in Reference 21 deviated from the present results. The present
method gave exactly the same values to the pure matrix case and the pure fiber case, as
expected, when the fiber volume fraction became 0 and 1, respectively. One of the features
of the present fiber model is that microscale stresses (i.e., stresses at the subcell level) can
be computed. Figure 7.13 shows subcell-level thermal stress values in megapascals for the
composite plate undergoing a 100 K temperature rise by either the plane strain condition
or the plane stress condition. The fiber volume fraction was 0.5 for this study. The fibers
had longitudinal tensile stresses of 16.2 MPa and 25.3 MPa for the plane stress and plane
strain conditions, respectively, and the transverse stresses were 4.7 MPa in tension and
34.3 MPa in compression for the plane stress and plane strain, respectively.
7.7.8 Hierarchical Composite
Aluminum-based metal matrix composite in a trimode developed in References 22 and
23 was considered to estimate its elastic modulus of the composite. The composite consisted of B4C particles embedded into ultrafine grain (UFG) aluminum, and the set of
230
Coefficient of thermal expansion (/deg)
105
8
L. CTE, Ref. 21
T. CTE, Ref. 21
L. CTE, present
T. CTE, present
5
4
3
2
1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Fiber volume fraction

FIGURE 7.12
Comparison of the CTE of glass/epoxy fibrous composite. L, longitudinal; T, transverse.
Plane strain: 33 = 0
33 = 30.6 MPa
Plane stress: 33 = 0
33 = 3.43 103
11.3
24.1
3
20.9
32.77
21.8
21.8
16.2
4.7
34.2
16.5
34.3
24.1
4.7
4.7
15.3
28.4
4.7
11.3
11.3
3
15.3
11.3
25.3
10.0
(a)
(b)
FIGURE 7.13
Thermal stresses in megapascals at the glass fiber and epoxy matrix, respectively, for a temperature change of
100 K and fiber volume fraction of 0.5: (a) plane stress and (b) plane strain.
materials were included in coarse grain (CG) aluminum. The two-level composition
of each mode can be modeled as seen in Figure 7.14 from the morphological observation of the material. The level I microstructure is an idealization of the conglomeration
of the hard particle (B4C) embedded in an UFG aluminum matrix by the cryomilling
process. The level II microstructure has the final product with a CG aluminummatrix.
231
UFG (ultra fine grains)
Level-I
CG
(coarse grain)
B4C
Level-I
Level-II
FIGURE 7.14
Trimodal B4C/AlUFG/AlCG metal matrix composite consisting of two levels.
TABLE 7.4
Constituent and Composite Material Properties for Trimodel Composite
Constituent materials,
Reference 23
Level I
Level II
Monolithic
E (GPa)
(kg/m3)
B4C
UFG (Al 5083)
Bimodal
B4C-UFG
CG
Trimodal
B4C/UFG-CG
448
70
E (GPa)
97
70
E (GPa)
83
0.1
0.3
0.27
0.3
0.28
2510
2657
PVF (%)
10:40
50
E (GPa), ROM
108
Theelastic modulus can be calculated by applying two-level microstructures one after

the other using the fiber model. Table 7.4 shows the data for the constituent materials and the calculated modulus. The present model yielded an elastic modulus of
83GPa, while the rule of mixtures (ROM) gave 108 GPa. The elastic modulus reported
in Reference 22 was 95 GPa by the compression experiment. Neglecting the morphological aspect and hierarchical procedure in Figure 7.14, the single-level calculation of
10% B4C and 90% aluminum gave an elastic modulus of 80.9 GPa using the present fiber
model. The ROM calculation gave the same modulus value by either the single-level or
two-level calculations.
7.7.9Glass Fiber/Epoxy Matrix Fibrous Composite under Tensile Load
The fiber model was implemented into a finite element analysis program. A square composite plate was analyzed under uniform tensile loading in the y axis, and the plate had
the fiber orientation measured from the x axis in the counterclockwise direction. The
results are shown in Figures 7.15 and 7.16 by obtaining the nominal stress and strain from
the finite element solutions. The plate was loaded by the displacement control and exhibited skew symmetric behavior when the fiber orientation angle was neither 0 nor 90. The
elastic modulus and Poisson ratio with respect to the fiber orientation angle were identical
to the exact solutions [24].
232
40
Tensile modulus, Ey (GPa)
35
Present
30
Exact, Ref. 24
25
20
15
10
5
0
10
20
30
40
50
60
70
Fiber orientation angle (deg)
80
90
FIGURE 7.15
Tensile elastic modulus as a function of the fiber angle, which is the measure between the fiber orientation and
x axis while loaded in the y axis.
0.7
Present
Exact, Ref. 24
0.6
Poissons ratio
0.5
0.4
0.3
0.2
0.1
0
10
20
30
40
50
60
70
Fiber orientation angle (deg)
80
90
FIGURE 7.16
Poisson ratio as a function of the fiber angle, which is the measure between the fiber orientation and x axis while
loaded in the y axis.
233
7.7.10Filament Wound Cylinder

Cylinders were fabricated using the filament-winding technique shown in Figure 7.17 [25].
The cylinders were made of six layups of boron fibers and epoxy matrix. The cylindrical
walls were considered for various loading conditions, such as axial loading, pressure loading, and torsion. A boron/epoxy composite has an orthotropic pattern consisting of a 2-to-1
ratio of circumferential (or hoop) to axial (or longitudinal) windings. Actual cylinders have
small variations in their component compositions, winding angles, overall dimensions,
and the like because of uncontrollable variation during the fabrication process. The boron
fibers had an elastic modulus of 450 GPa and Poisson ratio of 0.2, while the epoxy had
an elastic modulus of 3.28 GPa and a Poisson ratio of 0.35. The fiber volume fraction was
0.45, and the matrix volume fraction was 0.39. Then, the void volume fraction was 0.16. A
square patch of the cylinder wall was modeled here instead of using the full cylindrical
finite element model because the square patch gave exactly the same elastic modulus as
the full cylinder.
Two-step calculations were conducted for this study. The first step was to include voids
and to compute the reduced constituents properties resulting from the voids as discussed
previously. The second step used a finite element model of the square composite patch
having six layers of boron/epoxy laminate with the reduced properties. Each layer had
a uniform thickness. The square patch was subjected to tension in the longitudinal (L)
and hoop (H) directions as well as in-plane shear. Elastic properties for the composite
wall were computed from the results of the finite element analysis. The present results are
listed in Table 7.5 and are compared to other results [25]. The present method gave reasonable values. In addition, the present multiscale analysis can compute the fiber and matrix
stresses as well as the effective laminate stress, as listed in Table 7.5.
H
L
Orthotropic layup (2H/2L/2H)

Thickness = 1.4 mm
FIGURE 7.17
Filament-wound cylinder with orthotropic layers.
TABLE 7.5
Comparison of Effective Properties of a Filament-Wound Cylinder
Models
Component
Volume Fraction
E H (GPa)
EL (GPa)
G (GPa)
Present
Vf:Vr:Vv = 45:39:16
160
85
3.6
160 (140190)
145
78 (6983)
76
6.2 (4.86.9)
3.9
Test, Reference 25 (range)

Theory, Reference 25
Maximum
Stress (MPa)
at = 0.002
Fiber: 621
Matrix: 19.0
Laminate: 335
234
7.7.11Fibrous Composite Plate with a Hole

A square unidirectional fibrous composite plate with a hole was considered for this
example. The plate was made of glass fibers and epoxy matrix with a circular hole at
the center. The width of the plate was 76.2 mm, and diameter of the hole was 12.7 mm.
The composite plate was subjected to a tensile load. Material properties of the constituents were the same as those in the example in Section 7.7.7. The fiber volume fraction
was 0.5. Figure 7.18 shows the plate with a finite element mesh with line segments AB
and AC. Stresses were computed along the line segments. The segment AC was used to
obtain the circumferential stress component at a radius of 6.5 mm. Figure 7.19 plots the
results for the ratio of circumferential stress to the applied stress 0 along the segment
AC. The fiber orientation angle was varied from 0 to 90 with increments of 30. For
comparison, the isotropic material solution is included. The stress concentration factor
at point A was approximately 3 when = 0, and the factor decreased when = 30 or
= 60 and then increased when = 90. The maximum stress concentration for = 60
occurred approximately at = 160 (or = 20). This asymmetric behavior was reported
in Reference 26.
Figure 7.20 compares stresses in the composite level as well as in the fiber and matrix
material level, = 30. Even though the von Mises stress does not have a special meaning
for the nonisotropic composite material, it was selected just for comparison of the stresses
at different material levels. All stress values are in megapascals and for the applied nominal strain of 0.01.
Von Mises stresses in the fiber, matrix, and composite levels are plotted along the line
segment AB in Figure 7.21 for four different fiber angles. The effect of the fiber volume
FIGURE 7.18
Square fibrous composite plate with a center hole subjected to tension.
235
Isotropic
= 90
= 60
= 0
30
60
= 30
90
120
150
180
(deg)
FIGURE 7.19
Circumferential stress variation along segment AC in Figure 7.18.
S, Mises
(Avg: 75%)
180.0
165.0
150.0
135.0
120.0
105.0
90.0
75.0
60.0
45.0
30.0
15.0
0.0
(a)
SDV2
(Avg: 75%)
225.4
220.0
188.3
156.7
125.0
93.3
61.7
30.0
1.7
33.3
65.0
96.7
128.3
160.0
168.7
(b)
SDV50
(Avg: 75%)
60.0
55.0
50.0
45.0
40.0
35.0
30.0
25.0
20.0
15.0
10.0
5.0
0.0
(c)
FIGURE 7.20
Stress contour plots for square fibrous composite with a hole with the fiber angle = 30 as shown in Figure
7.18 (stresses in MPa): (a) composite von Mises stress; (b) fiber longitudinal stress; and (c) matrix von Mises
stress.
fraction on the stress concentration behaviors can also be obtained by the present analysis.
Fiber volume fractions of 0.5 and 0.7 are compared in Figure 7.21. When the fiber volume
fraction increased from 0.5 to 0.7, the stress carried by the fibers was reduced significantly,
while the stress concentration of the composite remained unchanged regardless of the
fiber volume fraction.
236
6
= 0
= 30
Fiber volume
fraction = 0.5
Fiber volume
fraction = 0.7
vm
0 2
Fiber volume
fraction = 0.5
Fiber volume
fraction = 0.7
vm
0 2
Fiber stress
Fiber stress
Composite stress
Composite stress
0
(a)
2
6
Matrix average stress
16
x (mm)
26
36
(b)
2
6
16
vm
0 2
= 60
= 90
Fiber volume
fraction = 0.5
Fiber volume
fraction = 0.7
Fiber volume
fraction = 0.5
Fiber volume
fraction = 0.7
36
Fiber stress
vm
0 2
Fiber stress
Composite stress
Composite stress
2
6
26
(c)
x (mm)
16
x (mm)
26
36
(d)
16
x (mm)
26
36
FIGURE 7.21
Stress plots along the line AB as shown in Figure 7.18, with fiber angles of (a) 0; (b) 30; (c) 60; and (d) 90.
7.7.12Particulate Composite Plate with a Hole

A metal matrixbased particulate composite plate was considered. The plate had a
square shape with a circular cutout at the plate center as shown in Figure 7.18. The
plate had a PVF of 0.7 and was subjected to tensile loading. A small area around the
hole (i.e., 3 mm wide in the radial direction) had a decreasing (soft) or increasing (hard)
PVF. Figure 7.22 plots the stress along the y axis (i.e., along the loading direction) as
a function of the distance from the center of the hole. Both composite-level stresses
and constituent material-level stresses are compared in the figure. The softer tip zone
reduced the stress concentration under the axial loading. A 10% reduction of the PVF
at the tip zone resulted in an approximately 10% reduction in stress concentration; this
was because of the decreased particle stress, as shown in Figure 7.22b. A higher PVF at
the tip zone area resulted in a greater stress concentration.
237
4
PVF = 0.7 and 0.7
PVF = 0.7 and 0.7
PVF = 0.6 and 0.7
PVF = 0.6 and 0.7
PVF = 0.8 and 0.7
3
y
0
PVF = 0.8 and 0.7
3
y
0
Hard tip zone

2
Hard tip zone
Soft tip zone
Soft tip zone

1
(a)
10
x (mm)
12
14
(b)
10
x (mm)
12
14
3
PVF = 0.7 and 0.7
PVF = 0.6 and 0.7
PVF = 0.8 and 0.7
2
y
0
Soft tip zone

1
Hard tip zone
0
(c)
10
x (mm)
12
14
FIGURE 7.22
Stress plots for a particulate composite plate with a hole with different PVFs: (a) Stress at the composite level;
(b)stress in the particles; and (c) stress in the matrix.
7.7.13Fibrous Composite with Statistical Consideration

Effective material properties of unidirectional composites were computed using the fiber
model based on the constituent materials, such as the given fiber and matrix materials.
For the first example, the fibers were graphites with an orthotropic material property. The
fibers had an elastic modulus and a Poisson ratio of 232 GPa and 0.275 in the longitudinal direction and 15 GPa and 0.49 in the transverse direction, respectively. The in-plane
shear modulus was 25 GPa. The matrix material was an epoxy with an elastic modulus of
5.35GPa and a Poisson ratio of 0.354.
The elastic moduli of the unidirectional graphite/epoxy composite are shown in
Figures7.23 and 7.24, which plot the longitudinal and transverse moduli as a function of
the fiber volume fraction. As Gaussian statistical distribution was assumed for the materials with a 90% confidence level, the upper and lower bounds are plotted in the figures
238
Longitudinal elastic modulus (GPa)
300
250
Experimental data
200
150
100
50
0.1
0.2
0.3
0.4 0.5 0.6 0.7

0.8
0.9
FIGURE 7.23
Plot of longitudinal elastic modulus of unidirectional graphite/epoxy composite as a function of fiber volume
fraction.
18
Transverse elastic modulus (GPa)
16
Experimental data
14
12
10
8
6
4
0
0.1
0.2
0.3
0.4 0.5 0.6 0.7

0.8
0.9
FIGURE 7.24
Plot of transverse elastic modulus of unidirectional graphite/epoxy composite as a function of fiber volume
fraction.
239
1500
Experimental data
Mean value
Failure strength (MPa)
1000
Lower value
Upper value
500
10
20
30
40
50
60
Off-axis angle
70
80
90
FIGURE 7.25
Plot of off-axis strength of unidirectional boron/epoxy composite.
and show that the experimental data obtained in Reference 27 were well bounded by the
predicted results using the fiber model.
7.7.14Fibrous Composite for Strength
Strength of a unidirectional boron/epoxy composite in the off-axis direction was predicted and compared to the experimental data as a function of the off-axis angle measured from the fiber orientation, as seen in Figure 7.25. Both results agree well each
other.
Delamination of a (90/0)s fibrous composite with a notch was investigated as the next
example. Because interlayer delamination occurs in a very thin resin layer between two
neighboring composite layers, it is important to model the very thin resin layer explicitly.
Figure 7.26 shows the predicted interlayer zone using the present analysis model [8]. The
result agreed well with the experimental observation.
7.8Examples of Fabric Model

Thermal stress analyses were conducted for plain weave composites using the finite element method (FEM) and the present multiscale techniques [7]. The composites were fabricated using glass/epoxy and boron/epoxy. Tables 7.6 and 7.7 list the fiber and matrix
material properties. The fiber volume fractions considered in the study were 0.18, 0.33, and
240
1.5
Delamination zone 2a/w = 0.125

103
0.5
0
0
0.2
0.4
0.6
0.8
1
103
FIGURE 7.26
Sketch of the delamination zone in a symmetric cross-ply composite with a notch-to-width ratio of 0.125.
TABLE 7.6
Material Properties of Glass and Epoxy
Material
Glass fiber
Epoxy
Modulus (GPa)
Poisson Ratio
CTE (10 6 m/m/degree K)
72.38
2.75
0.20
0.35
5.0
54.0
Modulus (GPa)
Poisson Ratio
399.62
3.45
0.20
0.35
5.04
36.0
TABLE 7.7
Material Properties of Boron and Epoxy
Material
Boron fiber
Epoxy
0.50. Figure 7.27 shows finite element meshes of a plain weave composite, that is, meshes
for the composite as well as for the fiber strands. The finite element mesh did not model
individual fibers because such modeling requires too many elements, which was beyond
the present computational capacity. However, the present finite element mesh can capture
the general behaviors of the strands in the average sense.
For analyses, a constant temperature change was applied to the plain weave composites
while they were free to deform. The effective CTEs of the plain weave composites and
241
y
x
(a)
(b)
FIGURE 7.27
Finite element model for a plain weave composite: (a) model for both woven strands and filling matrix and
(b)model for woven strands only.
the thermal stresses at the constituent level were computed from the analyses. Thermal
stresses in the fiber and matrix materials occurred because of the mismatch of CTEs of the
fiber and matrix materials.
Tables 7.8 and 7.9 show the CTE results for glass/epoxy and boron/epoxy plain weave
composites, respectively, and compare the present multiscale model results to the finite
element analysis results for three different fiber volume fractions. The multiscale predictions agreed well with the finite element results for both in-plane and out-of-plane effective
CTEs. Because the matrix material had a greater influence on the out-of-plane directionthan
the in-plane direction, the effective CTE was higher in the out-of-planedirection, resulting
from the higher CTE values of the matrix materials. Furthermore, the much stiffer boron
fibers yielded lower CTE values than the glass fibers.
TABLE 7.8
Effective CTE Comparison of Plain Weave Composite Made of Glass/Epoxy
Fiber Volume
Fraction
0.18
0.33
0.50
Present Result
FEM Solution
In-plane CTE
Out-of-plane CTE
In-plane CTE
Out-of-plane CTE
In-plane CTE
Out-of-plane CTE
0.2323
0.6707
0.1687
0.5588
0.1307
0.3964
0.2291
0.5922
0.1632
0.5095
0.1271
0.4012
TABLE 7.9
Effective CTE Comparison of Plain Weave Composite Made of Boron/Epoxy
Fiber Volume
Fraction
0.18
0.33
0.50
CTE (10 4m/m/degree K)
Present Result
FEM Solution
In-plane CTE
Out-of-plane CTE
In-plane CTE
Out-of-plane CTE
In-plane CTE
Out-of-plane CTE
0.0866
0.5204
0.0713
0.4180
0.0654
0.2920
0.0892
0.4584
0.0772
0.4023
0.0684
0.2754
242
TABLE 7.10
Stress Comparison of Plain Weave Composite Made of Glass/Epoxy
Fiber Volume
Fraction
0.18
0.33
0.50
Stress (MPa)
Present
Result
FEM Solution
Longitudinal fiber stress

Transverse matrix stress
1.7339
0.0695
1.159
0.130
0.7795
0.1915
1.4535 (avg), 1.8702 (max)

0.0851 (avg), 0.1275 (max)
1.052 (avg), 1.252 (max)
0.149 (avg), 0.171 (max)
0.7522 (avg), 0.9271 (max)
0.1935 (avg), 0.2188 (max)
Note: avg, average; max, maximum.
TABLE 7.11
Stress Comparison of Plain Weave Composite Made of Boron/Epoxy
Fiber Volume
Fraction
0.18
0.33
0.50
Stress (MPa)
Present
Result
FEM Solution

2.3539
0.1026
1.570
0.139
1.1481
0.1817
1.9877 (avg), 2.896 (max)

0.1382 (avg), 0.228 (max)
1.481 (avg), 1.831 (max)
0.143 (avg), 0.163 (max)
1.1252 (avg), 1.5199 (max)
0.1870 (avg), 0.2131 (max)
Note: avg, average; max, maximum.
Tables 7.10 and 7.11 compare thermal stresses at the fiber and matrix material level for a
unit temperature rise. The finite element model calculates stress variations from element
to element. As a result, both volume average stresses and maximum stresses were computed for each constituent material and compared to the multiscale model predictions.
The fiber stresses from the multiscale model were well bounded by the average and maximum stresses calculated from the finite element analysis. On the other hand, the matrix
stresses were slightly smaller than the average stresses obtained using the finite element
technique. Overall, the results from two different analyses compared well for different
ranges of fiber volume fractions.
The next example is a 2/2 twill composite subjected to the strain of xx = 0.016 and yy=
0.0009. The finite element results were compared to the multiscale results, as seen in
Table7.12, for some selected subcell numbers, which are shown in Figure 7.6. Both strain
results were comparable.
TABLE 7.12
Comparison of Strains at Selected Location of 2/2
Twill Composite
Subcell Number
(Figure 7.6)
1
4
11
Present Result
Finite Element
Results
0.0221
0.0165
0.0164
0.0277
0.0142
0.0150
243
7.9Elastoplastic Analysis of Composite Materials

Because plastic deformation is dependent on the loading history, the incremental stress
and strain formulation is used. The basic forms of equations are the same as developed
for the previous sections. For example, the equations for the fiber model are rewritten in
terms of incremental stresses and strains. The constitutive equations are written in the
incremental form as well. For a given strain increment nij and a temperature increment
of the nth subcell, the elastoplastic constitutive equation is
n( ep )
nij = Dijkl
nkl nkl (7.91)
n( ep )
where Dijkl
is the tangent modulus for each subcell. Equation 7.91 is almost the same as
n( ep )
Equation 7.13 except that the material property tensor Dijkl
is generally the tangent modulus for an elastoplastic material instead of an elastic material. Normally, the inclusion
material like fibers or particles is an elastic material with isotropy or anisotropy, while the
matrix material may have plastic deformation. For better convergence with the Newton
method, the consistent tangent modulus in Equation 7.91 should be used from the return
mapping of the yield function [2729]. The present model used the radial return mapping algorithm (RMA) for the J2 plasticity with isotropic hardening. The consistent tangent
modulus can be given as follows:
h
( ep )
( e )*
Dijkl
= Dijkl
+
3* ij kl (7.92)
1 + h / 3
( e )*
denotes
Here, superscript n is omitted for notational simplicity. In Equation 7.92, Dijkl
( e )*
the elastic part of the consistent modulus, Dijkl = * ij kl + *( ik jl + il jk ), and the other
term represents the consistent modulus due to plastic flow. Other quantities in Equation
7.92 are * = (/(tr)), * = (2/3)*, ij = f/ij, and h = d/dp. In addition, f is the yield
function; (tr) is the equivalent trial stress; h is the hardening slope of the uniaxial stressstrain curve; , , are Lame constants of elasticity; and ij is the Kronecker delta. Typical
isotropic hardening yield stresses for the matrix material are in the form of the Ludwik
equation, = 0 + K np, or modified Hollomon equation, = K(0 + p)n. Here, 0, 0, K, and n
are material constants from the uniaxial tensile test. For a multiaxial stress state, and p
denote the von Mises stress and equivalent plastic strain, respectively.
As in Section 7.3, substituting Equation 7.91 into other equations finally yields the following matrix equation, similar to Equation 7.18:
[T ]{} = { f (ij , )} (7.93)
where [T] is a 48 by 48 square matrix, and {f} is the vector on the right-hand side. Equation
T
7.93 can be solved for {} = 111 122 133 112 113 123 , which consists of the subcell
strain increment of Equation 7.91. This solution is plugged into Equation 7.91 to compute
the subcell stress increment. Consequently, composite level stresses and strains can be
written in the vector-matrix form:
{ } = [D]({ } {}) (7.94)
244
where { } and { } denote the vector form of unit cell (composite) stress and strain increment. A consistent tangent modulus matrix [D] and CTE vector {} for the unit cell (i.e.,
composite) can be derived as follows:

[D] = [V ][Dep ][R2 ] (7.95)
{} = [D]1[V ]({} [Dep ][R1 ]{}/) (7.96)
where [R1] and [R2] are submatrices of [T]1 = [R] = [R1 : R0 : R 2], which corresponds to the
segregation of { f } = {T : 0T : T }T. The unit cell stresses and strains are related to subcell
stresses and strains in the matrix form, such as { } = [V ]{ } and { } = [V ]{}. In addition, [Dep] and {} can be constructed by [Dep] = [Diag(Dn(ep))] and {} = {1 1 1 0 0 0 8 8
8 0 0 0}T, respectively, from Equation 7.91. Detailed descriptions for deriving the matrices
and vectors can be found in Reference 9.
As a result, the subcell strain increments nij, subcell stress increments nij, and unit
cell stress increments nij can be calculated from the given ij and . Furthermore, the
consistent tangent modulus [D] and CTE {} are derived using Newtons iteration method.
Using Equations 7.93 and 7.94, the present method ensures complete consistency between the
averaging and recovery processes. In addition, no subcell-level iteration is required for the
RMA for J2 plasticity. Usage of its consistent tangent modulus as expressed in Equation7.92
for a component of the structural modulus, Equation 7.94, ensures fast convergence for
Newtons method. Therefore, the present method needs iterations not for the subcell level
but for the structural level only such that it makes the computation efficient. The algorithms
in the Newtons iteration for the present method are summarized as follows:
Call from the incremental Newtons step

1. For a given unit-cell { } and , construct the state matrix [T] and right-hand-side
vector {f}:
n( e )
[T ( e ) ] T Dijkl

[T] [T(e)], if elastic for all subcells
[T] [T(ep)], if plastic for any subcell
{ f } { f (ij , )}

2. Solve for the subcell strain increment {} from the unit cell state equation: recovery
process:
{} = [T]1{f}
[T]1 = [R1 : R0 : R 2]
245
{ f } = {T : 0T : T }T

for n = 1, 2, , 8.
nij collect from {} {}

3. For the given subcell strain increment nij , perform the RMA for n = 1, 2, , 8:
( )
np f np = 0 [RMA]
np np + np [RMA]
1
ij nkk(tr ) [RMA]
3
nij nij n +
nij( p ) nij( p ) +
3 n n
ij p [RMA]
2
nij( e ) nij( e ) + nij n(p)

[RMA]
ij
hn
n( ep )
n( e )*
Dijkl
Dijkl
+
3 n* nij nkl [RMA]
n
n
1 + h /3
4. Update the state matrix [T]:
n( ep )
[T ( ep ) ] T Dijkl
5. Construct and update the unit cell variables from the subcell quantities: averaging
process:
8
ij
v
n
n
ij
n= 1
8
ij( e )
(e)
n ij
n= 1
8
ij( p )
( p)
n ij
n= 1
246
[D] = [V ][Dep ][R2 ]
{} = [D]1[V ]({} [Dep ][R1 ]{}/)
6. Return.
2 n( p ) n( p )
ij ij
is the equivalent plastic strain for the nth subcell. For postprocess3
ing purposes, other variables can be calculated accordingly.
where np =
3
sij sij , composite von Mises stress
2
vm =
2 ( p) ( p)
ij ij , composite plastic (inelastic) strain
3
p =
ij
1
VM
(7.97)
(7.98)
V
n
n
ij
, average matrix stress components
(7.99)
n=k
1
VM
V , average matrix plastic strain

n
n p
(7.100)
n=k
where sij is the composite stress deviator, k = 2 for particulate composite, k = 3 for fibrous
composite, and VM is the total volume of the matrix material. Strain energy and plastic dissipation energy can also be calculated appropriately. The present method keeps track of all
the subcell stresses and strains as well as composite stresses and strains. As a result, the
method is applicable to path-dependent loading-unloading simulation or residual stress
analysis by mechanical and thermal loading.
7.10Examples of Elastoplastic Analysis of Composite Materials

7.10.1Particulate Metal Matrix Composite Made of SiC and Aluminum
A typical particulate metal matrix composite (PMMC) was selected to apply the present
model. The composite consisted of 15% SiC particles in an A365 T6 aluminum extruded bar.
The particle diameter was 16 m, and the experimental results are reported in Reference
15. The material properties used for the study were as follows:
SiC:

E = 427 GPa, = 0.22
247
A356 T6:

E = 76 GPa, = 0.33, 0 = 199.7 MPa, K = 4.848 GPa, n = 0.1493
The stress-strain curve for the aluminum matrix was fitted by the modified Hollomon
equation = K(0 + p)n from the experimental curve in Reference 15. Figure 7.28 compares
the present results to the experimental and other theoretical results. The pure aluminums
stress-strain curve is also included as a reference. The present model was applied for both
the matrix-only problem and the composite problem. The matrix-only problem was solved
by setting the number 1 subcell of the micromechanics model to be matrix material like the
surrounding subcells 2 through 8. The curve in Reference 13 represents the Mori-Tanaka
result, which is a mean field theory based on Eshelbys method. The FEM curve represents
the finite element solution.
The result of the present model is in reasonable agreement with the experimental and
other results. In detail, the present model predicted the good elastic range, slightly lower
initial yielding stress ( < 0.01), and gradually overestimated postyielding stress ( > 0.02).
Yielding started at subcell 5 in Figure 7.2 if the loading was applied along the 3 direction.
The composite strain at the start of yielding was about 0.0015 by both the present and
the FEM models. At the composite strain of 0.01, the von Mises stress and the equivalent
plastic strain of subcell 5 were 267 MPa and 0.01565, respectively, for the micromechanics
model. Using the FEM, those values were 264 MPa and 0.01562, which were calculated by
the average over the volume equivalent to subcell 5.
Several strengthening mechanisms, including the size effect [15,30], may contribute at
the initial yielding range such that the strength of the composite may have higher strength
than the classical theory. Several failure mechanisms, including particle-matrix interface
decohesion [30,31], may occur at the postyielding range so that hardening of the composite
400
PMMC
15% SiCp/A356 T6
d = 16 m
Stress (MPa)
300
Matrix only
A356 T6
200
Experimental
Ref. 15
Ref. 13
FEM
Present
100
0.01
Strain
0.02
0.03
FIGURE 7.28
Stress-strain curves of 15% SiCA356 T6 particulate composite (particle diameter: 16 m).
248
TABLE 7.13
Comparison of Computational Statistics between the Present Micromechanics Method and
the Conventional Finite Element Method
100-Increment Scheme
Maximum Strain Level: 0.05
Total number of elements
Total number of increments and iterations
Average iterations per increment
Total CPU time(s)
CPU time per iteration(s)
20-Increment Scheme
Present
FEM
Present
FEM
1
101
1.01
9.2
0.091
8000
111
1.11
1947
17.54
1
38
1.9
4.7
0.124
8000
44
2.2
675.9
15.36
may be lower than for the present model. The experiments in Reference 15 showed the size
effect on strengthening; that is, the smaller particle diameter yielded higher composite
strength. Because the present model does not consider such strengthening and failure
mechanisms, the stress-strain curve in Figure 7.28 is reasonable in terms of both strength
and deformation.
Computational aspects of the present method along with the FEM are summarized in
Table 7.13. The FEM used 8000 (20 20 20) hexahedral elements, but the present method
used only 1 element. The analysis used 100 increments or 20 increments up to the global
maximum strain level of 0.05, respectively. A laptop computer with an Intel Core 2 Duo at
1.86 GHz was used for comparison. Not only a small number of iterations per increment
but also dramatically less central processing unit (CPU) time per iteration were observed
by the present model.
7.10.2Particulate Composite Made of SiC and Cu Alloy
The next example considers a particulate composite consisting of 20% SiC and an
Al-3.5wt% Cu alloy as studied in Reference 32, which shows the stress-strain curves for
the composite and the matrix alloy. Material properties are given in the following, and the
stress-strain equation of the matrix alloy was fitted using the modified Hollomon equation, = K(0+p)n:
SiC:
E = 427 GPa, = 0.22

Al-3.5 wt% Cu:

E = 72 GPa, = 0.33, 0 = 172.4 MPa, K = 618.5 MPa, n = 0.2117
The particle diameter varied from 1 to 7 m, and its median value was between 3 and
4 m. As a result, 3.5 m was selected for the present model. Because of the extensive
aging process after the fabrication for the composite [32], the size effect by quenching was
considered negligible. Figure 7.29 compares the present result to other experimental, FEM,
and analytical results. The result by the present model was close to the experimental
result up to a strain level of 0.015. The FEM overestimated and the method in Reference
13 underestimated the experimental result.
249
400
PMMC
20% SiCp/Al-Cu
d = 17 m
Stress (MPa)
300
200
Matrix only
Al-3.5 pct Cu
100
Experimental
Ref. 32
Ref. 13
FEM
Present
0.01
0.02
0.03
Strain
FIGURE 7.29
Stress-strain curves of 20% SiCAl-3.5 wt% Cu particulate composite (particle diameter: 17 m).
7.10.3Thermal Residual Stresses for a Whisker Composite

Most composite materials have a cooling period during the fabrication process, which
results in residual stresses in the composite because of mismatch in CTE between the
reinforcing inclusions and the soft matrix materials. Such thermal residual stresses in the
metal matrix composite were studied in Reference 33 using the Eshelby method. The composite was the whisker-shaped SiC with the 6061 aluminum matrix. The aspect ratio of the
short fiber was 1.8. Material properties were as follows:
SiC:
E = 427 GPa, = 0.17, CTE = 4.3 106/K

Al 6061:

E = 72 GPa, = 0.33, CTE = 2.36 105/K, 0 = 47.5 MPa, ET = 2.3 GPa
The aluminum matrix was assumed to be an isotropic, linear-hardening material of initial yield stress 0 and plastic modulus ET. The cooling temperature was T = 200 K.
The theoretical value of the residual stress in the matrix material was derived in a matrix
volume average form [33]. Figure 7.30 shows the longitudinal stress zzM and tangential
stress M obtained from Reference 33. The tangential and radial stresses were in tension and compression, respectively, with an identical magnitude, while the longitudinal
stress was in compression. As a result, the absolute stress values are plotted in Figure 7.30.
The effective von Mises stress vmM can be calculated from component stresses, and the
result is also indicated in Figure 7.30. The effective von Mises stress was also calculated by
the present model as shown by the solid line straightforwardly in a similar fashion by the
250
70
SiCAl 6061
Residual stress (MPa)
60
T = 200K
vm M
50
40
zz M
30
20
Ref. 33
10
0
Present
0
0.05
0.15
0.1
SiC volume fraction
0.2
0.25
FIGURE 7.30
Predicted residual stresses for SiCAl 6061 whisker composite by cooling.
matrix average. Both predictions were close for a dilute system (less than 5% volume fraction). For a larger volume fraction, the result given by Reference 33 begins to overestimate
the residual stress. Physically, the von Mises residual stress must be bounded by a certain
value because of the yield criterion. It is evident that the present method predicted the
bounded value as seen in Figure 7.30.
7.10.4Fibrous Metal Matrix Composite
Studies of fibrous metal matrix composites (FMMCs), including those for matrix material
plasticity, can be found in References 34 and 35. Although extensive experimental and
analytical investigations were presented in Reference 34, basic constituents material properties were not explicitly provided. Therefore, the composite material used in Reference
35 was selected for this study. The composite was made of 34% boron fibers with 2024
aluminum matrix. The stress-strain curve for the 2024 aluminum was fitted by the Ludwik
equation, = 0 + K np. Constituents material properties were as follows:
Boron:
E = 379 GPa, = 0.20

Al 2024:

E = 60.5 GPa, = 0.33, 0 = 83.5 MPa, K = 562 MPa, n = 0.447
Figure 7.31 shows the stress-strain predictions for transverse and longitudinal loading
by the present model and FEM. Both results agreed well with each other. The figure also
251
250
Long.
Stress (MPa)
200
FMMC
34% BAl 2024
150
Trans.
100
Matrix only
Al 2024
50
Ref. 35
FEM
Present
0.001
0.002
Strain
0.003
0.004
0.005
FIGURE 7.31
Stress-strain curves of 34% BAl 2024 unidirectional FMMC.
compares the experimental results for the transverse loading. Three experiments for the
composite and a uniaxial test of the 2024 aluminum are included in the figure. The experimental data by the transverse loading were a little bit higher than the present analysis for
the elastic range of the stress-strain behavior. The elastic modulus could be affected by
several factors, such as the packing arrangement of the fibers and boundary conditions
such as plane strain or plane stress. The present model considered the plane stress condition, while the experiments may be between plane stress and plane strain conditions.
Therefore, the experimental modulus was higher than for the present model. The plane
strain condition caused higher hydrostatic stress, which led to higher yield stress by the
von Mises yield criterion. Postyielding behaviors ( > 0.2%) by the two analyses deviated
from the experiments. The experimental results showed severe failure or damage after
initial yielding, while the analyses exhibited apparent hardening after initial yielding.
This can be explained by the damage observed during the experiments. According to the
detailed description in Reference 35 through photomicrographic observations, the test
specimen failure was associated with boron filament failure along its diametrical plane
normal to the direction of the applied stress. It was investigated that boron fibers may fail
at an applied stress of about 170 MPa. In addition, other damages from microflaws, such as
interfacial defects and matrix voids, may lower the strength. This is possible because the
diameter of the boron filament is normally large (over 100 m). Because the present model
does not account for any failure other than matrix plasticity, the prediction resulted in
higher hardening, as seen in Figure 7.31.
7.10.5Fibrous Composite Plate with Preexisting Crack
A composite plate with an initial crack was studied as a structural example using the
present model. The plate was 100 100 mm, and its thickness was 1 mm. The crack was
252
15 mm long. The material for the plate was a unidirectional composite made of 34% boron
fibers and 2024 aluminum, as in the previous example. Adopting ideas from the study
in Reference 34, the crack orientation as well as fiber orientation were the parameters for
this study. Figure 7.32a shows a crack perpendicular to the loading direction with an arbitrary fiber orientation angle . Figure 7.32b shows a slanted crack at an angle , with fibers
always parallel to the crack.
This example focused on a local-global effect due to matrix yielding and plastic deformation. The present study can predict many physical aspects around a crack tip, such as
the crack propagation direction, matrix failure area, and residual stresses for the composite material from a micromechanics viewpoint. Here, an extensive qualitative study
is presented, and results are compared with experimental observations described in
Reference34.
Figure 7.33 compares propagation of the plastic yielding zone in the matrix material
around the perpendicular crack tip of Figure 7.32a for the composites with fiber angles =
0, = 45, and = 90. The applied stress level is indicated as = /max (max = 100 MPa) in
(a)
(b)
FIGURE 7.32
A 34% BAl 2024 fibrous composite cracked plate problem: (a) perpendicular crack and (b) oblique crack.
= 3/4
= 3/4
= 3/4
=1
=1
=1
= 1/2
= 1/2
= 1/4
= 1/2
= 1/4
= 1/4
1 mm
(a)
FIGURE 7.33
1 mm
1 mm
(b)
(c)
Perpendicular crack: propagation of plastic failure region p > 0 : (a) fiber angle = 0; (b) fiber angle = 45
and (c) fiber angle = 90.
253
the plots. Contour lines in the figure represent frontal lines of plastic failure p > 0 . The
plastic yielding propagates along the fibers longitudinal direction. This is clear for fiber
orientation of = 45 or 90 as shown in Figure 7.33b and c, where the yielding emanates
from the notch tip and eventually propagates toward the fiber orientation. This prediction
coincides with the experimental results for the fatigue crack propagation with = 90 for
up to 20,000 cycles [34].
The slanted crack results are shown in Figure 7.34 for various crack orientation angles
in Figure 7.32b. The plot was reoriented to make the slanted cracks at different angles be
aligned to one another for easy comparison. At applied stress level of = 1/2, the plastic
yielding zone is shown in Figure 7.34. It agrees well with the results from Reference 34.
Finally, residual stress analysis was undertaken. A complete cycle of loading up to
100MPa followed by unloading was applied to the perpendicular crack plate with = 0
in Figure 7.32a, and the results are shown in Figure 7.35. At the maximum applied stress,
the entire plate is in plastic deformation. After complete unloading, at a remote material
= 1/2
= 0
= 30
= 60
1 mm
FIGURE 7.34
Oblique crack: Plastic failure region p > 0 .
S, Mises
(Avg: 75%)
B
A
210
190
170
150
130
110
90
70
50
30
10
(a)
(b)
(c)
FIGURE 7.35
Perpendicular crack: Residual von Mises stress: (a) composite stress; (b) fiber stress and (c) matrix average stress.
254
point, the fibers remain in a slight compression (4 MPa) longitudinally, while the matrix
remains in slight tension (1 MPa) along the fiber direction. The net stress for the composite
remains stress free.
Near the crack tip, the stress field is severe and complicated, as shown in Figure 7.35.
Those residual stresses occurred at two distinct levels. One was the global structural level
due to nonuniform near- and far-field stresses, and the other was the micromechanics
level resulting from the CTE mismatch in the fiber and matrix materials. Fibers were in
biaxial compression with about 200 MPa. Along the fiber direction, however, the matrix
was in compression (35 MPa) at point A while in tension (35 MPa) at point B in Figure 7.35c
near the crack tip. The matrix was in compression in the loading direction (50190 MPa).
Therefore, the maximum residual von Mises stress was approximately 230 MPa, which is
a high stress. Note that the initial yield stress of the matrix material was about 80 MPa.
These detailed aspects, including both structural and material behaviors, can be predicted
directly by the present model.
7.10.6Laminated Composite Plate
The last example demonstrates the capability of the present model for a laminated composite structure. The same 34% boronaluminum fibrous composite was used for each ply
with various fiber orientations. The plate had 10 plies of [45/90/45/90/0]s with equal
thickness (0.1 mm). At the structural level, the plate was quasi-isotropic because of the
layer orientations. However, each ply should have behaved differently. Figure 7.36 shows
the results of plastic deformation and stresses at the respective layers when applied stress
was 100 MPa. Figure 7.36 shows larger plastic deformation in plies 1 and 3 (i.e., 45 layers)
and smaller plastic deformation in ply 2 and 4 (i.e., 90 layer). This result was consistent
with analyses carried out previously for the single ply. Therefore, it can be predicted that
the structural plastic deformation may initiate in ply 1 and ply 3.
Ply 1 (45)
3
Ply 3 (45)
Ply 5 (0)
Ply 2 and 4 (90)
Ply 2 and 4 (90)
Ply 5 (0)
Ply 3 (45)
Ply 1 (45)
1 mm
FIGURE 7.36
Plastic deformation contours in each ply of [45/90/45/90/0]s laminated composite when = 1 (applied
stress: 100 MPa).
8
Multiscale Analysis of Metallic Materials
8.1Introduction
Most metallic materials have grain structures in the microscale. With the advancements
of nanotechnology, new materials are constructed with nanoscale grain sizes because
the material strength is inversely proportional to the grain size. Strength of the material
greatly depends on the grain size and characteristics. This chapter presents a simple multiscale analysis technique for a polycrystalline material. The nature of such a material is still
too complex to fully understand in all length scales. More refined multiscale techniques
will be developed as more knowledge is available to bridge different length scales.
8.2Polycrystalline Materials
Polycrystalline metallic materials are commonly used for engineering applications, especially for load-carrying structural members. The stiffness and strength of those materials are influenced by various factors associated with various length scales. For instance,
atomic defects such as vacancies, impurities, and dislocations play important roles at the
atomic or molecular level, which is at the nanoscale. On the other hand, grain boundaries
at the microscale also affect material strength and deformation. Furthermore, macroscale
defects such as notches, holes, and cracks can also influence the strength of the bulk material. Most of them in various length scales cannot be avoided, but some of them may be
altered through some engineering processes. Therefore, it is necessary to include all those
characteristics in different length scales to understand and predict the mechanical behaviors of engineering structural materials.
8.3Previous Study of the Multiscale Analysis of Metals

Multiscale models and analyses have been developed for metallic materials [110] in which
microstructural characteristics, including voids, dislocations, and the like, were examined
through multiscale analysis. The multiscale models considered two neighboring length
scales; some researchers undertook coupling of a discrete model, such as an atomistic
255
256
model and a continuum model [11,12]. Some of the previous multiscale modeling works on
metallic materials are discussed next.
Chen and Mehraeen [3] used the principle of virtual power for an asymptotic expansion
of field variables into multiscale Euler equations at different scales. The approach was
applied to grain deformation and evolution under stress. Takano et al. [4] applied both
the enhanced mesh superposition method and the asymptotic homogenization theory for
multiscale finite element analyses to study porous materials. Cuitino et al. [5] identified
controlling unit processes at the microscale (e.g., dislocation mobility, interactions and evolution for plastic deformation). Then, atomistic modeling was conducted for the controlling
unit process assuming independency of those processes. The atomistic parameters were
used to correlate the macroscopic driving force to the macroscopic response. Tadmor et al.
[9] developed the quasi-continuum method to link the atomic and continuum models. Zhu
et al. [10] incorporated an interatomic potential into a continuum finite element analysis
using the Cauchy-Born hypothesis [13]. They studied dislocation nucleation induced by
nanoindentation. The finite element result incorporating the interatomic potential energy
compared well with the pure molecular dynamics result.
One of the important aspects of multiscale modeling is linking one scale to the next
scale. For example, it is necessary to link an atomic model to a continuum model. Liu et al.
[14] provided a recent survey on multiscale modeling and simulation. They summarized
and compared various coupling techniques between the atomistic model and the continuum model. Some of those techniques were hierarchical [10], concurrent [9], and bridging
scale methods [15].
This chapter presents a systematic approach for multiscale modeling to implement characteristics of different length scales of a polycrystalline material into the structural behavior, from the nanoscale (i.e., the atomic dimension) to the macroscale (i.e., the engineering
structure dimension).
8.4Procedure for Multiscale Analysis of Polycrystalline Metals

The proposed multiscale model consists of four different length scales: macroscale, mesoscale, microscale, and nanoscale. The macroscale is an engineering structural level whose
dimension can range generally from centimeters to meters. This scale is a continuum level,
so finite element analysis is conducted for the macroscale analysis. At this scale, the structure may have holes, notches, cracks, and the like, and it is subjected to external loading,
possibly with geometric constraints.
The mesoscale consists of a multigrain whose size is usually on the order of millimeters
or less. This level is also a continuum level so that another finite element analysis is conducted at this scale. Intergrain boundary characteristics can be employed at this level.
Therefore, special care is provided for the grain boundaries, as discussed further in the
chapter.
The microscale model is for a single-grain boundary, which is generally in the dimension
of micrometers. The single-grain domain is divided into finite elements whose material
properties are obtained from smeared atomic behaviors. In this model, material defects
inside grains, such as dislocation, vacancies, impurities, and the like, can be considered
collectively.
257
Macroscale:
Mesoscale:
Microscale:
Nanoscale:
Test coupon
level
Multigrain
level
Single grain
level
Atomic level
FIGURE 8.1
Schematic diagram of the multiscale analysis of polycrystalline metal.
Finally, the nanoscale is for a local section inside a single grain where dislocations,
vacancies, impurities, and so on at the atomic level can be considered explicitly. For example, interaction among different atomic defects can be examined at this level. Molecular
mechanics is suitable for this length scale.
Coupling of different length scale models is depicted in Figure 8.1. The figure illustrates
how each length scale analysis is connected the others. The following sections describe
each scale analysis in more detail and how the neighboring length scales are connected to
each other.
8.5Macroscale Analysis
Macroscale analysis is undertaken for load-carrying engineering structures such as test
coupons or structural components subjected to loading. Many different numerical analysis techniques can be applied at this level. For simple problems, analytical solutions may
be available, although those are limited cases. Among all different solution techniques,
the finite element method is the technique used most often because of its flexibility and
generality. The structure to be investigated is discretized into a finite element mesh as
illustrated in Figure 8.2. To obtain a more accurate solution, an adaptive mesh technique
may be utilized for a local zone where multiscale analysis will be undertaken. At the
selected local points, the following set of information is obtained from the finite element
analysis:
{u/x u/y u/z v/x v/y v/z w/x w/y w/z} (8.1)
258
FIGURE 8.2
Macroscale finite element mesh of a structure with a concentrated force and supports at both ends.
for a three-dimensional (3-D) domain analysis, and

{u/x u/y v/x v/y} (8.2)
is used for a two-dimensional (2-D) domain analysis, where u, v, and w are the displacements along the x, y, and z coordinate axes, respectively. The solutions given in Equation
8.1 or 8.2 are utilized for the mesoscale analysis. Because 2-D and 3-D multiscale analyses
use the same techniques, 2-D analysis is presented in the descriptions that follow.
8.6Mesoscale Analysis
A representative 2-D unit cell of a rectangular shape was selected for the mesoscale analysis. The size of the unit cell is generally in millimeters or micrometers, depending on
the average grain sizes in the metallic material under study. The unit cell consists of
a number of grains. In most cases, because grain sizes and shapes vary from location
to location inside the material, it may be necessary to construct generic grain sizes and
shapes properly representing the actual grains that can be measured using a scanning
electron microscope. In 2-D analysis, the Voronoi diagram may be utilized to generate
grain boundaries inside the unit cell as shown in Figure 8.3. The average size of Voronoi
FIGURE 8.3
Finite element mesh of Voronoi polygons in the mesoscale unit cell (solid bold lines indicate the boundary of
Voronoi polygons, and the broken lines are the finite element boundaries).
259
polygons is determined based on the average size of the measured grains. If the average
grain size is longer in one axis due to a prior unidirectional deformation process, Voronoi
polygons are constructed to have a longer dimension in the same axis. The seed points to
generate Voronoi polygons are determined using a random number generation process.
The number of grains is determined by dividing the unit cell length by the average grain
size.
Each Voronoi cell, which represents a grain, is further divided into a finite element mesh
for the mesolevel finite element analysis (Figure 8.3). Then, the deformation gradient solution obtained from the macroscale analysis as given in Equation 8.2 is used as boundary
conditions of the mesoscale model as indicated in Figure 8.4. In the figure, line segments
AB and DC are parallel to each other, while line segments BC and AD are also equal and
parallel, too.
Because grain boundaries have different crystallographic orientations with dislocations, the interface stiffness of a grain boundary is different from the internal stiffness
of the mostly regularly positioned grains. As a result, the grain boundary is modeled
using different finite elements. A simple way is to use spring elements as illustrated in
Figure8.5. Two neighboring finite elements at a grain boundary have independent nodes
that are overlapped at the grain boundary as shown in the figure. Then, the two independent nodes are coupled using spring elements in both directions. The spring constant
should be properly determined to represent the stiffness at the interface. A low value of
the spring constant means low stiffness of the interface along the spring orientation. A
molecular dynamics simulation may be undertaken to determine the proper stiffness of
the interface for various conditions, such as different crystallographic orientations and
dislocations. Depending on the arrangement of the Voronoi polygons, a point at multi
grain boundaries may have multiple nodes, so each grain sharing the point can have
its own node. The finite element solution of the mesoscale unit cell provides the nodal
displacements of a grain, which is analyzed in the subsequent microscale analysis. The
selected grain with known nodal displacements is further analyzed as described in the
next section.
u
L
y y
C
D
v
L
y y
Ly
v
L
x x
A
Lx
u
L
x x
FIGURE 8.4
Sketch of deformation of the mesoscale unit cell.
260
Grain boundary
Elem. A
Elem. A
Double
Elem. B
Nodes
Elem. B
Spring element
FIGURE 8.5
Two finite elements meeting at a grain boundary and connected by spring elements.
8.7Microscale Analysis
Microscale analysis is typically at the length scale of micrometers or less. The unit cell is a
single grain selected from the previous mesoscale analysis. Independent multiple grains
can be analyzed if necessary. Each single grain is also analyzed using the finite element
method because there are still too many atoms in a single grain so a continuum model is
still applicable. As a result, finite element analysis is undertaken for the single-grain unit
cell with the boundary displacements obtained from the mesoscale analysis.
The grain may have defects, such as dislocations or vacancies. The stiffness of each finite
element in the grain may be adjusted based on the respective state of defects. If there is no
specific information available for the locations and states of defects inside the grain, a random assignment of defects for each element may be assumed in terms of statistical data.
From the microscale analysis, the deformation gradients as in Equation 8.2 are obtained at
a selected location for the next-level nanoscale analysis. If desirable, the microscale analysis and nanoscale analysis can be conducted simultaneously without two subsequent
analyses.
8.8Nanoscale Analysis
Eventually, nanoscale analysis is performed at the atomic scale, which can represent all the
defects as they are. The dimension of the nanoscale model is nanometers. Because there are
still too many atoms in the nanoscale model in general, we cannot model all atoms in the
nanoscale model. Instead, the nanoscale model consists of three subdomains. Individual
atoms are modeled in the inner subdomain, while a smeared continuum model is used for
the outer subdomain. Then, two subdomains are coupled through the interface subdomain.
Figure 8.6 shows a nanoscale model. The inner subdomain contains a detailed atomic
arrangement, such as atomic vacancies, dislocations, impurities, and so on. Molecular
mechanics is applied to the inner subdomain. The outer subdomain is modeled as a continuum with smeared material properties of the atoms, and it is modeled using the finite
261
Continuum region with

smeared atoms
Discrete atomic
region
Interface region
FIGURE 8.6
Nanoscale analysis model containing three subdomains. The broken lines indicate a finite element mesh.
element technique [11]. Eventually, the interface subdomain contains both finite elements
and discrete atoms as denoted by bold lines in Figure 8.6. This subdomain provides coupling between the discrete atoms and the smeared continuous subdomain as discussed in
Chapter 6 [16]. Boundary conditions are applied to the outer subdomain using the deformation gradients obtained from the microscale analysis as expressed in Equation 8.2.
8.9Example Problems
As an example for multiscale analysis, a large plate with a hole at the center is considered
[17]. The plate is subjected to a uniform applied stress o as shown in Figure 8.7. At the edge
of the hole, a nanoscale crack is located, denoted by A in the figure. Multiscale analysis
is applied to point A. Because the nanoscale crack is too small to be included in the macroscale analysis, the macroscale model neglects the nanoscale crack. In this example case,
an analytical solution is available so it is used for the macroscale level instead of finite element analysis. The deformation state at point A in Figure 8.7 is given as

u
3 o
=
,
x
E
u
= 0,
y
v
= 0,
x
v 3 o
=
(8.3)
E
y
where u and v are the displacements along the x and y axes, respectively, and E and are
the elastic modulus and Poisson ratio of the material, respectively. Pure aluminum is chosen as the material, so the elastic modulus is 70 MPa and the Poisson ratio is 0.3.
Figure 8.8 shows the mesoscale analysis model at point A in Figure 8.7. The average grain
size is assumed to be 10 m. The multigrains are generated using the Voronoi diagram with
random point generation. Each grain in Figure 8.8 is discretized into a number of finite
262
o
Plate
Nanoscale
crack
x
A
Hole
o
FIGURE 8.7
A very large perforated plate under remote uniform stress with a nanoscale crack.
Selected grain for microscale analysis

FIGURE 8.8
Multigrains for mesoscale analysis in its local coordinate system with the unit in micrometers. The grain in the
bold-line boundaries is for the next-level microscale analysis.
elements. Approximately 1000 finite elements are used in the mesh. If the grain boundary information (e.g., grain orientations and defects in grain boundaries) is known at the
specific location, it should be used in the mesoscale model. Otherwise, statistical data for
the grain boundary can be utilized. For example, the stiffness of grain boundaries (i.e.,
the interface spring constants in Figure 8.5) of the mesoscale model is assigned randomly
based on the selected statistical model to represent the random grain orientations within
grains and grain boundary defects. For the present example, a uniform random distribution is considered for the grain boundary stiffness, which is assumed to vary between the
263
prescribed ranges. Four different cases are considered in this study. It is assumed that the
grain boundary stiffness varies randomly between (a) 90% and 100%, (b) 50% and 100%,
(c) 10% and 100%, and (d)1% and 100% of the value that represents no defect in the grain
boundary. The major stress component that occurs in the y axis of Figure 8.7 is calculated
for the grain, denoted by bold boundary lines in Figure 8.8 and normalized with respect
to the applied tensile stress. The stresses are calculated 200 times for each case of the four
scenarios for statistical analysis. Then, the probability of events at different stress states is
plotted in Figure 8.9 for the four cases. As expected, the actual state of stress depends on
the grain boundary state.
Microscale analysis is conducted for the bottom left grain as selected in Figure 8.8 using
the finite element method. The selected grain is shown in Figure 8.10 with a finite element mesh. Approximately 2000 elements are used in the microscale model. However,
for visual clarity, a much coarser mesh is shown in Figure 8.10. The boundary conditions
of the microscale model are obtained from the deformation of the single-grain boundary
from the mesoscale analysis. The elastic modulus of each finite element inside the grain
is randomly varied to represent possible defects inside the grain, such as dislocations
or vacancies. The first case assumes a uniform random variation of the elastic modulus
between 90% and 100% of the modulus of the no-defect case. A total of 100 computer simulations are performed with the randomly generated modulus. The second case assumes a
Gaussian distribution of the elastic modulus between 90% and 100% of the modulus of the
no-defect case, with 95% for the mean elastic modulus. Figure 8.11 plots the histogram of
0.2
Probability of event
0.2
0.15
0.1
0.05
0
(a)
2
3
Normalized stress
0.15
0.1
0.05
0
(b)
0.15
0.1
0.05
0
2
3
Normalized stress
2
3
Normalized stress
0.2
0.2
(c)
2
3
Normalized stress
0.15
0.1
0.05
0
(d)
FIGURE 8.9
Probability of stress states at the selected grain location in Figure 8.8. The grain boundary stiffness varies randomly within (a) 90% to 100%; (b) 50% to 100%; (c) 10% to 100%; and (d) 1% to 100% of the stiffness of the grain
boundary without any defect.
264
Selected element for nanoscale model
0.16
0.16
0.14
0.14
0.12
0.12
0.1
FIGURE 8.10
Finite element mesh for the microscale model of the selected grain in the mesoscale model.
0.08
0.06
0.1
0.08
0.06
0.04
0.04
0.02
0.02
(a)
0
2.7
2.8
2.9
3
3.1
Normalized stress
3.2
(b)
0
2.7
2.8
2.9
3
3.1
Normalized stress
3.2
FIGURE 8.11
Probability of stress states at the selected location in the grain as shown in Figure 8.10. The material defects in
the finite elements of the grain vary in (a) uniform random distribution or (b) Gaussian random distribution
between 90% and 100% of the stiffness of the grain boundary without any defect.
the probability of events at the stress state at the bottom left corner of the selected grain in
Figure8.10.
Finally, Figure 8.12 shows the nanoscale analysis model. The discrete atomic region is
surrounded by the smeared continuum region. Figure 8.12 shows the coupled atoms and
the smeared continuum. Even though more than 1500 atoms are used, only a small number of atoms are shown in the figure for visual clarity.
265
Discrete atomic subdomain
Smeared continuum subdomain

FIGURE 8.12
Nanoscale model with the discrete atomic subdomain surrounded by a smeared continuum subdomain that
was divided into a finite element mesh.
In the nanoscale analysis, the nanoscale crack is modeled discretely. The nanoscale
model is subjected to the stress field at the boundary as determined in the microscale
analysis. Because the finite element model is used for the smeared continuum region, the
boundary conditions are applied to the finite element mesh. The displaced atomic positions near the crack tip due to the load are shown in Figure 8.13. The figure illustrates the
atomic separation around the crack tip indicating crack propagation at the bottom side of
the crack tip.
Crack growth
FIGURE 8.13
Crack growth shown with separation of atoms.
9
Multiscale Analysis of Biomaterials
9.1Introduction
Biomaterials are complex living tissues that have evolved for many years. Those materials
have distinct hierarchical structures constructed of simpler materials [1]. The human body
has a limited number of basic materials to make the desired biomaterials in the macroscale
[2]. Bones, tendons, ligaments, muscles, and skin are examples of such biomaterials. This
chapter studies the bone because it is the main part of the body skeleton that can support
the load in both tension and compression [35].
Complex multiscale hierarchies of the bones can carry various large loading with an
optimal weight and serve as the storage site for bone marrow, calcium, and phosphate.
To understand and predict the structural properties of the bones, a multiscale model was
developed and its analysis was conducted. The multiscale model began with the nanoscale
model and continued up the hierarchies to the macroscale level. Figure 9.1 shows a sketch
of multiscale hierarchical structures of the bones. In the following sections, some biological descriptions in different length scales are provided. Then, multiscale biomechanics
models and analyses are presented [6].
9.2Nanoscale Model
The major nanoscale components of the bone are hydroxyapatite (HA) and tropocollagen.
These two components become the building blocks for the next scale model: the microscale
model. The bone contains microscopic inclusions of noncollagenous proteins and proteoglycans other than hydroxyapatite and tropocollagen. These two materials act as a binding matrix, while noncollagenous proteins also act for metabolic functions [7]. Because
the noncollagenous proteins and proteoglycans do not serve for mechanical functions
and their volume fractions are low, they are not considered in the nanoscale model discussed here. As a result, only hydroxyapatite and tropocollagen are further discussed and
modeled.
9.2.1Hydroxyapatite
Hydroxyapatite is a hexagonal, close-packed crystal lattice crystalline solid material consisting of calcium phosphate, and it is the only ceramic-type material created naturally
inside the human body [8,9]. Hydroxyapatite is sometimes called bone mineral, and it is
267
268
Nanoscale
(a)
(b)
(c)
(d)
Microscale
(e)
Macroscale
(f )
(g)
FIGURE 9.1
Diagram of bone hierarchies from nanoscale to macroscale: (a) Tropocollagen; (b) Hydroxopatite; (c) Fibril;
(d)Fiber; (e) Lamellar bone; (f) Trabecular bone; and (g) Cortical bone.
the major component providing bone stiffness. The material has a polycrystalline form as
small thin plates within the body [4,10].
Hydroxyapatite has a chemical formula of Ca10(PO4)6(OH)2. The growth of hydroxyapatite is regulated by heterogeneous nucleation factors that promote the growth of the mineral phase and contribute to the highly ordered structure of the microlevel hierarchy of
the bone. The growth is considered concentrated within the gap zones. Because hydroxyapatite exhibits a hexagonal close-packed crystal lattice, a single crystal can be tested to
determine material properties [10].
269
TABLE 9.1
Properties of Hydroxyapatite Materials
Source
Elastic Modulus, E (GPa)
Poisson Ratio,
Measuring Technique
150
144
114
165
0.27
Nanoindentation along [0001]

Nanoindentation along [1010]
Ultrasonic
[10]
[11]
[8]
Nanoindentation testing on single hydroxyapatite crystals showed that the stiffness in

the [0001] direction of the hexagonal close-packed crystal lattice was greater than that in
the [1010] direction [10]. Some of the previous test results are listed in Table 9.1. The data
showed some variations. Because the nanoindentation test data were the median, those
values were used in this study as the reference values. Although the size of hydroxyapatite
crystals varies depending on locations, average crystal size is considered 50 25 3 nm
[1,3,4,8,10,12], and the crystals are organized into small plates.
Hydroxyapatites also store ions, and they are responsible for 99%, 90%, 90%, and50%
of the bodys store of calcium, phosphorus, sodium, and magnesium, respectively. Hydroxy
apatites have a large ratio of surface area to volume, which allows for rapid absorption and
dissolution of ions as needed [4,10]. This is important in determining effects of bone diseases and disorders that affect bone density and calcium absorption.
9.2.2Tropocollagen
Collagen is constructed of tropocollagen molecules as the basic unit. Collagen has 28 different variations to serve different physiological purposes as needed in the body [13].
Collagens in human bones consist mostly of collagen I type, about 95% of the total collagen [14].
Collagen I is composed of three helical protein strands. Two are alpha-1 type 1 (COL1A1)
strands, and the third one is the alpha-1 type 2 (COL1A2) strand [13,1519]. Each strand is
different in terms of its amino acid sequence, but each strand maintains a similar helical
structure. Both COL1A1 and COL1A2 are left-handed polyproline II (PPII) helices, and
they are 300 nm long. They also contain approximately 1000 amino acids [1925]. Those
helices have a common repeating subunit, called Gly-X-Y [13,1618,20,21,26]. Gly is the
amino acid glycine, which is the smallest of the amino acids. For each repeated strand,
an amino acid glycine is along the central axis of the tropocollagen molecule [27,28]. This
arrangement makes the three strands orient themselves to form stable hydrogen bonds.
There is a pattern between the X and Y positions of the three-amino-acid subunit. Two
commonly repeated patterns are Gly-Pro-Y and Gly-X-Hyp, where Pro and Hyp represent
the amino acids proline and hydroxyproline, respectively [1618,23]. A common model for
the tropocollagen molecule utilizes the repeating sequence Gly-Pro-Hyp.
9.2.3Helical Spring
A helical spring model was selected to estimate the stiffness of a single molecule of the tropocollagen. This model was selected because each polyproline helix of the tropocollagen
molecule can be independently represented as a spring. The stabilizing hydrogen bonds
as well as the van der Waals force among the three helices keep the tropocollagen from
buckling and make sure that all three helices deform symmetrically. The close packing
270
of the tropocollagen molecules within collagen fibrils also prevents buckling, which is
discussed further in this chapter. The stiffness of the tropocollagen molecule is the sum of
the stiffness of the COL1A1 and COL1A2 strands. Equation 9.1 was used to determine the
spring stiffness of each helix.

k=
Gd 4
(9.1)
8D3 N A
where k represents the spring constant, G represents the shear modulus, d is the wire
diameter, D denotes the mean coil diameter, and NA represents the number of coils.
To compute the quantitative values for each protein helix, we assumed that each
repeating subunit is consists of the Gly-Pro-Hyp sequence. This representative subunit
provides a relatively stable position for each amino acid. It also allows for proper crosslinking and hydrogen bonds to form between strands [18,20,23,29]. The presence of the
amino acids proline and hydroxyproline greatly influences the stiffness of collagen [30].
Modeling every subunit with this sequence provides good agreement with the correct
bond angles and axial repeat, 10/3, for the majority of the polyproline helices [18]. Some
recent studies discussed the predominance of an axial repeat of 7/2 over portions of the
PPII. However, this variation is more important for molecular dynamics simulations in
which strain energies and steric effects of the amino acid interactions are computed and
analyzed [27,28].
The wire diameter for the helical spring model as expressed in Equation 9.1 was taken
tobe the average spacing between residues, 0.286 nm [19,27,28]. The coil diameter of the
helix was determined from the diameter of a tropocollagen molecule. A tropocollagen molecule has a diameter of 1.5 nm. Hence, a single PPII helix was assumed to have a mean coil
diameter of 0.5 nm [18,20,21]. The number of coils was determined from a constant axial
repeat of 10/3 [16,23]. In other words, the number of coils was 300 for 1000 aminoacids.
The shear modulus was the most difficult parameter to derive because no data are available
for the shear modulus of a single-chain amino acid helix. The bond energy of the backbone
of a single subunit was computed to estimate the shear modulus. There are six CN bonds
and three CC bonds along the backbone inside the Gly-Pro-Hyptriplet. Bondenergies were
5.06 10 19 J/bond and 5.75 10 19 J/bond for CN andCCbonds,respectively;therefore, the
total bond energy of the backbone was then 4.76 1018 J/subunit. Asasubunithadavolume
of 7.02 1029 m3, the shear stress of the backbone was computed as energy over volume,
which was equal to 67.9 GPa. This is close to the typical values of aluminum (70 GPa) and
thus not unreasonable. This calculation did not include all strengthening effects of crosslinking, nonbackbone atoms, and other atomistic considerations. However, this value provides a basic starting point to calculate the overall stiffness of collagen.
9.2.4Results of the Nanoscale Model
Table 9.2 is a summary of the parameters and their values used for Equation 9.1. The
springconstant k for a helix is 0.0015 N/m. The elastic modulus of a helix was computed
from the spring constant using the following equation:

L
E = k (9.2)
A
271
TABLE 9.2
Parameter Values Used for Equation 9.1
Parameter
Value
Units
G
d
D
NA
k
67.9
0.286
0.5
300
0.0015
GPa
nm
nm
N
m
where L is the length and A is the cross-sectional area of the helix. The length and diameter of a PPII helix were 300 nm and 0.5 nm, respectively. However, due to the helical twist,
the cross-sectional diameter of the helix was defined as 0.7 nm. Using the helix length of
300 nm and cross-sectional area of 3.85 10 19 m2, Equation 9.2 yielded the elastic modulus
of 1.18 GPa for each protein helix. Because each tropocollagen molecule had three helices,
the elastic modulus of a tropocollagen molecule was 3.54 GPa.
The present estimated modulus is compared to other experimental data and theoretical
analysis results in Table 9.3. All results showed quite a variation in the elastic modulus of
the tropocollagen, from less than 1 GPa to more than 10 GPa. When compared, the present
nanoscale model using the equivalent helical spring presented a reasonable elastic modulus of the tropocollagen, confirming the validity of this model.
TABLE 9.3
Comparison of Elastic Modulus for Tropocollagen
Elastic Modulus (GPa)
3.54
1.2
2.8
616 (average 6)
7
2.4
3
9
5.1
3
0.3512
4.8 1
1.4
(PHG)a 11.37
(PPG)b 13.43
a
b
Method
Source
Spring model
Property used in finite element model
Property used in finite element model
Molecular dynamics
Atomistic modeling
Atomistic modeling
X-ray diffraction
Brillouin light scattering
Brillouin light scattering
Estimate from persistence length
Estimate from persistence length
Molecular dynamics
Debye-Waller factor
Molecular dynamics
Molecular dynamics
Present study
[1]
[31]
[29]
[32]
[33]
[34]
[35]
[36]
[37]
[33,38]
[39]
[23]
[29]
[29]
PHG is a Gly-Pro-Hyp model.

PPG is a Gly-Pro-Pro model.
272
9.3Microscale Model
The previous section considered the material properties of the elemental composite materials of bone at the nanoscale, that is, hydroxyapatite and tropocollagen. These two simple constituents produce a more complex hierarchy at the next length scale, called the
microscale. The structures at the microscale have organized and regular arrangements of
collagen and hydroxyapatite. A micromechanics model can be developed and used at this
length scale.
9.3.1Two-Dimensional Fibril Structure
The fibril is the smallest size of the microstructures at the microscale. It consists of a staggered two-dimensional (2-D) array of hydroxyapatite and tropocollagen molecules. Those
arrays depend on the binding and cross-linking tendencies of tropocollagen molecules.
The regular fibrils are organized during a process called fibrillogenesis, which involves
the expansion and regrowth of fibrils. A collagen molecule has a C terminal and an N
terminal, so named for the peptides cleaved during the transition from procollagen to
collagen [19,25]. As each end of a tropocollagen exhibits different cross-linking tendencies,
the amino acid sequence at each terminal of the tropocollagen defines a polar reference for
each molecule. When the collagen molecules are aligned near each other, the C and N ends
join to produce an organized structure. Not only do the ends of the tropocollagen molecules have preferential alignment, but also the entire length of the tropocollagen molecule
contains segments of amino acids preferential to cross-linking with adjacent tropocollagen
molecules [26]. These preferential segments form a collagen network with a 40-nm gap
and a regular periodicity of 67 nm [22,25,26,4044]. Hydroxyapatites grow preferentially
in these gap regions, allowing for the inclusion of the reinforcing particles within the fibril
composite. Figure 9.2 shows the staggered alignment of hydroxyapatite in the fibril composite. This figure is not to scale.
The height of the crystals and tropocollagen matrix is 3 nm, and the hydroxyapatite
crystals are 50 nm in length. Figure 9.2 is used to portray the periodic arrangement known
to occur. This shows a regular structure driven by the molecular sequence of the tropocollagen molecules. Furthermore, such cross-linking results in stability on the tropocollagen phase of the fibril [22,30]. The cross-linking keeps the individual molecules from
shearing. This in turn increases the strength and brittleness of the structure [22,31]. The
67 nm
67 nm
67 nm
67 nm
67 nm
67 nm
1
3
~N
C~
~N
C~
~N
C~
FIGURE 9.2
Visualization of 67-nm periodicity in bone fibril. Gray regions denote hydroxyapatite crystals, and white regions
depict tropocollagen.
273
presence of cross-linking also allows the collagen phase to be represented as a single,

continuous, material in a simplified model. An important observation of the tropocollagen
and hydroxyapatite microstructures shows that the hydroxyapatite crystals are aligned in
the parallel direction to the microfibril orientation. This results from the growth behavior
of the mineral phase within the gap zones.
The initial deposition of hydroxyapatite at a molecular level is influenced by heterogeneous nucleation factors. One of the factors is bone sialoprotein, which binds directly to the
ends of collagen I molecules. In addition, bone sialoprotein is found within the gap zones of
microfibrils in a high concentration and has been found to be a nucleator of hydroxyapatite
formation, creating a hydroxyapatite crystal within a supersaturated medium [45].
Hydroxyapatite crystals grow in a preferential direction, depending on the substrate
chemistry on which they are grown [46]. As a result, mineral components supersaturate
the gap zone, followed by crystal growth on the ends of the collagen I molecules initiated
by the bone sialoproteins. Subsequently, crystals grow in an ordered manner that aligns
all crystals parallel with the microfibril axis [45]. Because all crystals have grown in the
same orientation, they can be considered as a transversely isotropic material using the
material values given in Table 9.1.
There is one conflict in the physical models based on tropocollagen cross-linking. The
collagen network has 40-nm gaps, which are smaller than the 50 25 3 nm hydroxyapatite crystals. It is believed that the ends of the tropocollagen molecules must be surrounded
by hydroxyapatites, which grow between adjacent molecules [44,47]. This helps to prevent
shearing of the hydroxyapatitetropocollagen-bonded surfaces. The width of the crystals
represents another challenge in modeling and determining the three-dimensional (3-D)
structure of collagen fibrils.
9.3.2Three-Dimensional Fibril Structure
Even though the 2-D regular staggered array of a collagen fibril can be easily visualized,
3-D fibrillogenesis in vivo is still a debated topic [25,26]. When fibrils are examined using
electron microscopy, they have shown different 3-D structures. One is a linear fibril structure, and the other is a twisting fibril structure.
A linear fibril model has been accepted for many years [25,27,47,48]. The linear model
was considered as a long, thin filament with alternating bands of mineral-rich phase and
mineral-deficient phase. The former is sometimes referred to as the gap, while the latter is
referred to as the overlap. The mineral content in rich phases is approximately twice that
of the deficient phase [48]. Figure 9.3 shows the 3-D packing of hydroxyapatite crystals in
the linear fibril model.
2
1
3
FIGURE 9.3
Simplified linear fibril model.
274
67 nm
67 nm
67 nm
1
3
FIGURE 9.4
Twisting fibril model.
Even though the linear fibril structure is a widely accepted model for the proposed theory of 3-D fibrillogenesis, there has not been any evidence that the fibril grows purely linearly in the lateral direction. A recent study with the in vitro organization of fibril growth
showed that fibrils grow laterally in 4-nm steps [41]. Each lateral step was defined by electrostatic attraction during the formation of the microscale fibril, but the lateral growth was
not related to longitudinal growth [41].
More recently, a twisting crystalline structure was considered [7]. A spiral orientation
resulted in a more even distribution of hydroxyapatite crystals. In the twisting model,
periodicity was assumed in the lateral direction with 67 nm, and the 2-D stacking also
directed the 3-D pattern. A simplified twisting fibril model is shown in Figure 9.4.
Figures 9.3 and 9.4 show the two proposed fibril models and illustrate a simplified
visualization of hydroxyapatite crystal distributions. The organization of bone fibril in
vivo is much more complex because it includes noncollagenous proteins; errors in collagen stacking, amino acid sequencing, and continuous absorption and regrowth of fibrils.
Furthermore, biological materials are never made of pure substances. The collagen used
in bone contains only 95% collagen I [14]. It is postulated that the other collagen species
assist in regulating the fibrillogenesis process, but they may also introduce irregularities
into the structure [19].
Fibrils are generally considered to have a circular cross section whose diameter varies
from 50 to 300 nm [49,50]. It is reasonable to state that only one crystal width is present
across the structure in those small-diameter fibrils. This would validate the use of the
linear fibril model. However, either linear or twisting structures may be possible for fibrils
with larger diameters. To explore the two models discussed previously, a micromechanics
model was modified to characterize both structures, and their results were compared to
other experimental and theoretical data to determine their validity.
9.3.3Micromechanics Fibril Model
Kwon and his coworkers [5159] developed a micromechanics model for the analysis of
composite structures. This model is presented in Chapter 7, and the representative unit
cell is shown in Figure 7.2. The micromechanics model was used for the present analysis.
For the present work, the collagen network was assumed to have a staggered array with
at least a 25-nm width to allow for the generally accepted dimensions of the hydroxyapatite crystals. Then, the micromechanics model would provide relevant results for both the
linear packing model and the twisted packing model. Table 9.4 lists the material properties
used to model the fibril with the micromechanics model.
275
TABLE 9.4
Material Properties for Fibril Components
Material
Hydroxyapatite
Tropocollagen (compression)
Tropocollagen (tension)
E1 (GPa)
E 2 (GPa)
12
G 12 (GPa)
150.38
3.428
3.256
143.56
3.428
3.256
0.23
0.35
0.35
59.744
1.270
1.206
Water is present inside fibrils. The presence of water makes the fibril a bimodulus composite material because water can only support compressive load. Bone is known to contain
approximately 10% to 25% water [31], and some of this is thought to be contained inside
the nanoscale tropocollagen and hydroxyapatite. Water serves as a binding and stabilizing agent within the tropocollagen [26,30], and small amounts of water are tightly bound
within the hydroxyapatite crystal [31]. In addition, the rest of the water is assumed to be
held within the various hierarchies. For the present fibril model, each unit cell was postulated to have 8% water by volume. In other words, the stiffness of collagen included 8%
water by volume in compression and an 8% void space volume in tension. Tropocollagen
in compression and tension exhibits the different properties shown in Table 9.4.
9.3.3.1Linear Fibril Subunit
The linear packing model was a lateral repeat of the 2-D fibril array shown in Figure 9.2.
Each crystal had a length of 50 nm, width of 25 nm, and height of 3 nm. Figure 9.3 shows
the cross section used to create a repeating subunit of the array. This subunit section
repeats throughout the array in 67-nm increments. The dimensions used for the micromechanics unit cell are listed in Table 9.5. Subcell 1 was assigned hydroxyapatite properties;
the remaining subcells were assigned tropocollagen properties.
9.3.3.2Twisting Fibril Subunit
The twisting fibril takes into account periodicity in the 2 direction as well as the 3 direction, as shown in Figure 9.5. Table 9.6 lists the dimensions used for the twisting model.
The dimensions for the a1, a2, b1, and c1 lengths were determined with the 67-nm periodicity in the 1 direction and the known size of the hydroxyapatite crystals. For the twisting
model, subcell 1 was assigned hydroxyapatite properties, and the remaining subcells were
assigned tropocollagen properties.
TABLE 9.5
Unit Cell Dimension for Linear
Fibril Model
Dimension
a1
a2
b1
b2
c1
c2
Value (nm)
50
17
25
3
3
9
276
2
1
3
FIGURE 9.5
Subunit employed in the twisting fibril model.
TABLE 9.6
Unit Cell Dimension for Twisting
Fibril Model
Dimension
Value (nm)
50
17
25
75
3
6
a1
a2
b1
b2
c1
c2
9.3.4Fibril Results
The micromechanics model calculated the effective properties of the fibril unit cell. Fibrils
are expected to have transverse isotropic properties because of the random dispersion
in the radial orientations. The 1 axis in the micromechanics model lies along the fiber
axis, but the 23-plane orientation will be different for each fibril. To consider the random
dispersion in the radial direction to compute transverse isotropic material properties, the
elastic and shear moduli and Poisson ratios were averaged in the 2 and 3 axis. The resulting values are shown in Table 9.7, and the results are compared to other experimental and
theoretical values in Table 9.8. All the results except for the experimental nanoindentation
and dynamic mechanical analysis evaluated the longitudinal elastic modulus in tension.
The others determined the transverse elastic modulus of a fibril in compression.
The micromechanics model resulted in the elastic moduli, which compared well with
other data for bone fibrils. The micromechanics model provided a simple solution to a
complex problem. The relatively elementary approach yielded realistic results for both the
TABLE 9.7
Transverse Isotropic Fibril Results
Model
Linear
Twisting
Compression
Tension
Compression
Tension
E1 (GPa)
E 23 (GPa)
G 23 (GPa)
G 12 (GPa)
12
32
6.050
5.760
4.264
4.054
6.601
6.298
3.812
3.621
1.534
1.458
1.354
1.286
1.677
1.593
1.368
1.299
0.299
0.299
0.309
0.309
0.295
0.295
0.366
0.366
277
TABLE 9.8
Comparison of Values for Fibril Results
7.65 3.85
1.00 0.75
7.5 5.5
4.96 0.57
3.07 0.23
4.36
38
10 2.0
5
4.81
2.05 0.75
4.75 3.06
a
Method
Source
Nanoindentation of rat tail tendon (transverse)

Finite element model, tension (longitudinal)
Finite element model, compression (longitudinal)
Dynamic mechanical analysis using AFM, peak (transverse)a
Dynamic mechanical analysis using AFM, trough (transverse)a
Molecular multiscale modeling (small strain)
Molecular multiscale modeling (large strain)
3-D molecular dynamics (longitudinal)
Molecular modeling (longitudinal)
Molecular dynamics and finite element modeling (longitudinal)
Full atomistic model (longitudinal)
Finite element model (longitudinal)
[60]
[31]
[31]
[50]
[50]
[22]
[22]
[42]
[44]
[47]
[61]
[48]
Peak and trough refer to mineralized and unmineralized sections, respectively. AFM, atomic force
microscopy.
TABLE 9.9
Fibril Model Mineral Volume Fraction
Fibril Model
Linear
Twisting
Mineral Volume Fraction (%)

16.66
6.22
linear fibril model and the twisting fibril model. In addition, the micromechanics model
calculated the Poisson ratios for the fibril rather than assuming a value, as did all results in
Table 9.8. The micromechanics model resulted in accurate results for the elastic modulus
of bone fibrils in both tension and compression. The complete data set for the fibril calculated using the micromechanics model can be used to increase the accuracy of subsequent
hierarchies, as the micromechanics model was used for the bone fiber.
The bone mineral content was used as a metric for comparing the validity of theoretical models. However, the mineral content at the microstructure level is not well known.
Therefore, the mineral content was tracked at each length scale and was compared at the
macrolevel in further study. Table 9.9 shows the mineral content of the two fibril models.
9.3.5Bone Fiber
In the hierarchical structure, the bone fibril makes up the bone fiber. Fibers consist of
fibrils and hydroxyapatite. The fiber contains hydroxyapatite in the form of extrafibrillar
mineral, deposited between densely packed fibrils [2,7,62,63]. Bone fibers are sometimes
called fibril bundles. Minerals are deposited on the outside of fibrils, which are arranged
in parallel. The extrafibrillar minerals provide a stiffener for the bone fiber. These mineralized fibrils are tightly packed in a uniform direction [7].
Each closely packed fibril is surrounded by a crust of hydroxyapatite minerals, which
grow as small plates with an approximate thickness of 2030 nm to coat the fibril [7,63]. As
a result, there is an ordered arrangement of minerals such that the 1 axis of minerals and
278
3
2
FIGURE 9.6
Three-dimensional fiber array. Dark rectangles represent hydroxyapatite crystals; the rod structures represent
fibrils.
the longitudinal axis of fibrils are parallel to each other [63]. This allows each fiber to be
modeled as a transversely isotropic material. The thickness of the crust was considered to
be 26 nm, and the diameter of a single uniform fibril was chosen to be 150 nm [7]. Figure
9.6 illustrates the 3-D structure with the surface mineral removed for visual clarity.
The fibrils in Figure 9.6 show a parallel arrangement, and both linear and twisting mineral fibril models were analyzed. There was little information regarding the degree of
mineralization along the lengths of the fibrils [7]. Different variations of mineralization
were considered to analyze the effects of extrafibrillar mineralization (EFM).
Mineral packing at the fiber level is not a close-packed assembly because there are void
spaces among the minerals. Water is contained in those voids along with dissolved nonstructural proteins and macromolecules [7]. This solution is often called the extrafibrillar
matrix. This liquid phase at the fiber level represents the remaining component of bone
water, and these spaces are treated as water, resulting in a bimodulus property for fibers
as considered for the fibril model.
9.3.6Micromechanical Fiber Model
The fibers are made of single and multiple bundles of fibrils. The bundles have organized
and repeating arrangements of fibrils and hydroxyapatite. Therefore, the micromechanics
model discussed in Chapter 7 was also applicable to model the bone fiber. The micro
mechanics model relies on volume fractions of each material. To ensure a volume of unity,
Equation 9.3 was used:

a1 + a2 = b1 + b2 = c1 + c2 = 1
(9.3)
Furthermore, a circular cross section of the fibrils was modeled as an equivalent square
shape and minerals surrounding the fibrils.
In the micromechanics model, the 1-axis cross section remained constant like a continuous fiber composite model. This area percentage was based on a total crust thickness
of 26 nm and a fibril diameter of 150 nm. To apply this to the micromechanics model,
Table 9.10 shows the dimensions applied to the 1-axis face. This resulted in a fibril area
of 72.25% when converted into the micromechanics model. This is shown in Figure 9.7,
where subcell 1 represents the fibril. To test the effect of fibril mineralization, the degree
of mineralization was varied. Percentage EFM (%EFM) was defined as the total percentage
of fibril covered by mineral and was equal to the volume of subcells 3, 5, and 7, shown in
279
TABLE 9.10
Fiber Unit Cell 1-Axis Face Dimension
Dimension Parameter
Fraction
0.85
0.15
0.85
0.15
b1
b2
c1
c2
3
c2
c1
b2
b1
(a)
(b)
FIGURE 9.7
Longitudinal fiber cross section as compared to micromechanics model cross section. (a) Fiber cross section:
dark area represents mineral, light section represents fibril. (b) Micromechanics model longitudinal cross section.
Figure9.8. The %EFM was altered by changing the length of a1. It was postulated that the
fiber consisted of no more than 95% EFM. The different levels of mineralization analyzed
are listed in Table 9.11. The effect of water was also studied under compression. Subcells
4, 6, and 8 were assigned the properties of water in Figure 9.8. In tension, these subcells
wereassumed to be void spaces.
3
6
8
15
85
5
7
1
2
85
a2
FIGURE 9.8
Unit cell for fiber micromechanics model.
a1
15
280
TABLE 9.11
Fiber Unit Cell Extrafibrillar
Mineralization (EFM) Dimensions
%EFM
a1
a2
50
70
90
95
50
70
90
95
50
30
10
5
9.3.7Fiber Results
The material properties were computed using the fiber micromechanics model for compression and tension. As expected, the fiber exhibited transverse isotropic material properties because the radial arrangement of the fiber arrays had a random distribution. Tables
9.12 and 9.13 show the results for compression and tension, respectively.
The compressive results showed that the linear fibril model produced a stiffer fiber in
elastic and shear moduli than the twisting fibril model. For all %EFM, the fiber shear
modulus in the plane 12 direction was much less than for the plane 23 directions. As
%EFM increased from 50% to 95%, the plane 12 shear modulus increased by an order of
magnitude, while the plane 23 modulus approximately doubled. This was exhibited for
both the linear and the twisting models.
TABLE 9.12
Fiber Results in Compression
50% EFM
E1 (GPa)
E2 = E3 (GPa)
G23 (GPa)
G12 (GPa)
21
32
70% EFM
90% EFM
95% EFM
Linear
Twisting
Linear
Twisting
Linear
Twisting
Linear
Twisting
6.511
16.66
1.053
0.012
0.369
0.140
5.446
14.62
0.931
0.012
0.413
0.132
7.567
21.41
1.473
0.019
0.364
0.128
6.509
19.16
1.302
0.019
0.397
0.119
11.60
26.35
1.893
0.056
0.295
0.126
10.47
23.91
1.672
0.056
0.307
0.118
16.17
27.72
1.998
0.109
0.245
0.127
14.97
25.23
1.765
0.108
0.250
0.120
TABLE 9.13
Fiber Results in Tension
50% EFM
E1 (GPa)
E2 = E3 (GPa)
G23 (GPa)
G12 (GPa)
21
32
70% EFM
90% EFM
95% EFM
Linear
Twisting
Linear
Twisting
Linear
Twisting
Linear
Twisting
4.527
14.10
1.001
0.012
0.227
0.113
3.374
12.80
0.885
0.012
0.266
0.099
4.691
19.74
1.401
0.019
0.318
0.113
3.592
17.95
1.237
0.019
0.372
0.099
4.870
25.38
1.800
0.056
0.408
0.113
3.841
23.03
1.589
0.056
0.478
0.099
4.919
26.79
1.900
0.109
0.431
0.113
3.911
24.31
1.677
0.107
0.505
0.099
281
TABLE 9.14
Fiber Model Mineral Volume Fraction
%EFM
50
70
90
95
Linear
Twisting
25.91
31.46
37.01
38.40
18.37
23.92
29.47
30.86
The tensile data also showed similar results to those in compression. The linear model
exhibited greater moduli than the twisting model. In addition, the values of the transverse
elastic modulus were much greater than the longitudinal elastic modulus.
The shear moduli in the plane 12 direction were very low for a low %EFM. This was
due to a relatively large amount of void space present within the unit cell. However, this
is discussed further with the structure of lamellar bone. The inclusion of fibers within the
disordered fibrillar matrix can increase the shear modulus in the plane 12 direction.
Because fibers exist within the macrostructures of bone and are not easily isolated for
testing, there are no data available for the fiber model. Therefore, the present data cannot
be compared to assess the fiber model.
The mineral content completely surrounds the fibril. The mineral content of the different
fiber models was calculated. These calculations included both intrafibrillar and extrafibrillar hydroxyapatite. The resulting values are shown in Table 9.14.
9.4Macroscale Model
So far, nanoscale fibril and fiber models were developed from the nanoscale hydroxyapatite
and tropocollagen. These microscale components served as the base unit for the macro
scale biomaterials. To investigate the macroscale characteristics of bones, three different
models were examined. The first was a unit cell model of lamellar bone; the second was
a layered fiber-reinforced composite model of cortical and trabecular bone; and the third
was a finite element model of a tetrakaidecahedron modeling cancellous bone. Although
the lamellar layer was still measured in microns, it was quantified in the macroscale section because it is the main constituent of cortical and cancellous bone.
9.4.1Lamellar Bone
Lamellar bone is the next hierarchical structure of bones. It is equivalent to a fiber-
reinforced composite made of bone fibers and bone fibrils. More recently, the presence of
a disordered fibril matrix was discovered in the lamellar bone. As a result, the lamellar
bone comprises bone fibers surrounded by a matrix of disordered fibrils [7]. This is a new
finding that many previous models have not taken into account. Previous studies have
postulated that lamellar layers are constructed merely of an ordered array of fibers [3,64].
The discovery of the disordered fibril matrix required the previous micromechanics unit
cell model for the present hierarchical level. It was found that the lamellar bone exhibited
282
3
2
1
FIGURE 9.9
Single lamellar bone layer.
an ordered motif of 23 m, which represents the layer thickness, and a disordered matrix
thickness of 0.25 to 1 m, which was associated with the fibrillary matrix [7]. Small pores
with a diameter of 50 nm were also found along with the disordered matrix [7]. These
pores were represented as voids in the lamellar matrix. The voids could be included in the
model, but their volume seemed small. As a result, voids were neglected in the model. The
simplified view of a single lamellar layer is shown in Figure 9.9.
9.4.2Lamellar Model
The model used for lamellar bone to predict the material properties was similar to that utilized for bone fibers (i.e., the micromechanics unit cell model for continuous fiber composites). Because fibers within a lamellar layer are unidirectional and the disordered matrix
has a relatively random thickness, the lamellar properties are considered as transversely
isotropic material. The dimensions used for the lamellar unit cell model were as follows:
layer thickness of 2.5 m and matrix thickness of 0.375 m. These were the averages of
the ranges obtained in Reference 7. The fiber diameter was then computed to be 1.75 m.
Figure 9.10 shows the representation of the longitudinal cross section used for the lamellar model. The longitudinal cross section of the unit cell model had 38.4% fiber volume
fraction for the lamellar unit cell. The presence of the small pores was neglected as they
accounted for approximately 0.1% of the current models volume.
The unit cell model used the volume fractions of each constituent. To model the lamellar
layer as a continuous fiber, the dimensions of a1 and a2 were unimportant; other dimensions are listed in Table 9.15.
3
(a)
c2
c1
b2
b1
2
(b)
FIGURE 9.10
Longitudinal lamellar cross section as compared to micromechanics model cross section. (a) Lamellar cross section with shaded area representing disordered matrix and light section representing fiber. (b) Micromechanics
model longitudinal cross section.
283
TABLE 9.15
Unit Cell Dimensions for
Lamellar Model
Dimension
Value
0.50
0.50
0.62
0.38
0.62
0.38
a1
a2
b1
b2
c1
c2
The lamellar composite matrix contained a random distribution of fibrils, resulting in

isotropic material behavior. To determine the isotropic material properties, the averaging
technique was used to compute the equivalent elastic modulus and the Poisson ratio. The
values listed in Table 9.7 were used. In addition, fiber material properties of varying %EFM
were considered. Because the lamellar model contained the bimodulus materials, the
lamellar layer was expected to show transverse isotropic, bimodulus, material properties.
9.4.3Lamellar Model Results
Material properties were computed using the lamellar model for both tensile and compressive loading, respectively, as well as four different values of %EFM. The results are
shown in Tables 9.16 and 9.17.
TABLE 9.16
Lamellar Results in Compression
50% EFM
E1 (GPa)
E2 = E3 (GPa)
G23 (GPa)
G12 (GPa)
21
32
70% EFM
90% EFM
95% EFM
Linear
Twisting
Linear
Twisting
Linear
Twisting
Linear
Twisting
6.428
9.001
1.336
0.622
0.328
0.237
4.609
6.425
1.156
0.529
0.366
0.266
6.841
9.717
1.552
0.629
0.322
0.235
5.024
6.928
1.338
0.536
0.350
0.269
8.391
10.35
1.705
0.665
0.280
0.243
6.547
7.409
1.466
0.572
0.286
0.288
10.14
10.55
1.737
0.715
0.242
0.252
8.272
7.576
1.492
0.620
0.237
0.302
TABLE 9.17
Lamellar Results in Tension
50% EFM
E1 (GPa)
E2 = E3 (GPa)
G23 (GPa)
G12 (GPa)
21
32
70% EFM
90% EFM
95% EFM
Linear
Twisting
Linear
Twisting
Linear
Twisting
Linear
Twisting
5.514
8.305
1.270
0.592
0.312
0.235
3.718
5.992
1.098
0.503
0.370
0.262
5.578
9.174
1.475
0.599
0.338
0.230
3.801
6.538
1.271
0.510
0.394
0.257
5.646
9.781
1.621
0.635
0.355
0.226
3.895
6.902
1.393
0.546
0.407
0.254
5.665
9.905
1.651
0.684
0.359
0.225
3.922
6.976
1.418
0.593
0.410
0.253
284
TABLE 9.18
Lamellar Model Mineral Volume Fraction
%EFM
50
70
90
95
Linear
Twisting
20.22
22.35
24.48
25.02
10.89
13.02
15.16
15.69
Table 9.16 shows that the linear model produced a greater elastic modulus than the
t wisting model. Furthermore, the elastic modulus in the 1 direction was smaller than that
in the 2 or 3 direction for a low value of %EFM. As the %EFM value increased, the two
values became closer. For the twisting model, the elastic modulus in the 1 direction was
greater than that in other directions.
The tensile property also showed that the linear model yielded a higher elastic modulus
than the twisting model. In addition, the modulus in the 2 or 3 direction was larger than
that in the 1 direction for all considered values of %EFM.
The mineral content of the lamellar layers was analyzed as before, and it is tabulated in
Table 9.18. The mineral volume fraction within lamellar layers is a representative parameter for the macroscale bone mineral content. As a result, the predicted mineral content
could be compared to previous theoretical and experimental values. There was a wide
variation in previous data, which showed the mineral content from around 30% to 70%
[6568]. Among them, the data with better understanding of the hierarchical structure of
the bone gave approximately 30% to 40% mineral contents. Comparing the present bone
mineral content results to those data, the linear crystal pattern was considered the more
viable model. The mineral content of the twisting model was much lower than those values. Even though some perturbations were considered in the fibril and fiber models, the
mineral content of the twisting model would not match the current estimates. As a result,
the calculations that follow used the linear model.
9.4.4Cortical Bone
Cortical bone is the part of the bone with high density, and it is the outer shell of the bone
surrounding the spongy center of cancellous bone. Cortical bone makes up a large percentage of the weight fraction of the skeletal system and provides major stiffness of and
strength to the bone. Cortical bone has concentric layers of lamellar bone, called osteons,
which can be made of either primary or secondary bone. Primary bone is found where
bone has been grown de novo. Secondary bone is created when the primary bone is broken
down and regrown in place [7]. Secondary osteons are called Haversian systems, which
were modeled in this study.
The concentric layers composing Haversian systems form a cylindrical structure whose
diameter is approximately 200 m [7,64,69]. Haversian canals are located at the center of
these cylinders, and they contain the blood supply and nerve endings for the surrounding
bone. The diameter of the canals is approximately 3050 m [7,64,69].
Lamellar bone constitutes each concentric layer inside the osteon. As a result, each
layer has unidirectional fibers. The Haversian canal is represented as unbound water [7].
Multiple Haversian systems are packed together inside cortical bone. Due to their circular
shape, there are incomplete layers at the interface of each Haversian system where the
285
boundaries intersect. These boundaries are defined by a cement line, which is an identifiable region where osteon the growth direction has transitioned. The properties of cement
lines are similar to those of the surrounding bone, despite the misnomer cement [70].
The longitudinal axis of a Haversian system is preferentially aligned with the longitudinal axis of the bone for a majority of cortical bone [64]. Transversely oriented Haversian
canals do exist that interconnect each longitudinal Haversian canal, but they are short
compared to the longitudinal canals. This study assumed all Haversian systems were
aligned parallel to the long axis of the bone.
The fiber directions of the concentric lamellar layers vary for each layer like a filament
winding composite cylinder. Two distinct patterns of fiber orientation have been found
in lamellar bone. One is a periodic alternating pattern, and the other is a continuous fiber
twist. The distinct alternating pattern has high- and low-angle fibers. The high-angle
fibers are at 6580, and low-angle fibers are at 1530. Because of the imperfect nature
of biomaterials, there are some fibers with intermediate angles. The average orientations
of threesamples were calculated [71]. The results showed 45% high-angle fibers, 35% lowangle fibers, and 20% intermediate fibers [71]. The findings in Reference 64 also showed
a section of femur exhibiting a continuous twist of fiber angles from 0 to 180 between
lamellar layers. The continuous twist fibers were varied with an approximately 10 increment. Both fiber orientation patterns (i.e., the periodic alternating pattern and the continuous twist) can be modeled as a laminated fibrous composite.
9.4.5Cortical Bone Model
The lamellar layers of the cortical bone behave like a laminated composite. Each layer is
a fibrous composite with a transverse isotropic material. The material properties of each
layer can be determined from the fiber orientation. Figure 9.11 shows the respective layers
when the cylindrical shape of an osteon is cut and arranged flat on top of one another.
The material property of the cortical bone was computed using the volume average of
all the layers. The material properties of a single lamellar layer can be determined by
the coordinate transformation technique, which is discussed in Chapter 7 for a laminated
composite as given in Equations 7.82 through 7.84.
In addition to considering the fiber orientation of the individual lamellar layer, the
Haversian canal and canaliculi were addressed in the macroscale model. Taking into consideration the dimensions of the Haversian system, the Haversian canal, and the microscopic canaliculi, the macroscale cortical bone had 75% densely packed bone and 25% void
space. As done with models at a smaller length scale, the void space was addressed as a
liquid in compression and a void in tension.
FIGURE 9.11
Osteon cross-sectional view.
286
TABLE 9.19
Preferential Orientation Layered Composite
Model Parameters
Fiber Direction
72.5
22.5
0
40
55
90
Volume Percentage
0.45
0.35
0.05
0.05
0.05
0.05
TABLE 9.20
Smooth Orientation Layered Composite
Model Parameters
Fiber Direction
0
10
20
30
40
50
60
70
80
90
Volume Percentage
0.10
0.10
0.10
0.10
0.10
0.10
0.10
0.10
0.10
0.10
The periodic alternating pattern and the continuous fiber twist are referred to as preferential orientation and smooth orientation, respectively, from this point. Their effects were
examined. The two models relied on volume percentage of a defined fiber orientation.
Tables 9.19 and 9.20 list the volume percentages present for each fiber orientation of the
preferential model as well as for the smooth model.
The stiffness matrix of each layer was calculated using Equations 7.82 through 7.84 and the
fiber orientations provided in Tables 9.19 and 9.20. Because of the random radial distribution of
osteons, the material properties in the radial direction were assumed to be constant.
9.4.6Cortical Bone Model Results
The results of the layered cortical bone model were computed using only the linear fibril
model because it was more viable than the twisting fibril model. Both the preferential and
smooth orientation models were used for tension and compression. Table 9.21 shows the
compressive material properties; Table 9.22 shows the tensile properties.
Comparing the results for both tensile and compressive properties shows that the preferential fiber orientation model yielded greater stiffness than the smooth orientation model for all
%EFM values considered here. However, the difference between the two models was not large.
The elastic and shear moduli increased as %EFM increased, as expected.
The results of the present cortical model are compared to other experimental and theoretical data in Table 9.23. Comparing the results of the present model to the experimental
287
TABLE 9.21
Cortical Bone in Compression
50% EFM
E1 (GPa)
E2 = E3 (GPa)
G23 (GPa)
G12 (GPa)
21
32
70% EFM
90% EFM
95% EFM
Pref.
Smooth
Pref.
Smooth
Pref.
Smooth
Pref.
Smooth
5.074
5.836
0.700
0.944
0.319
0.232
4.642
5.679
0.735
0.957
0.357
0.226
5.390
6.217
0.773
1.037
0.316
0.227
4.914
6.045
0.818
1.050
0.355
0.221
5.944
6.818
0.839
1.144
0.311
0.222
5.453
6.586
0.889
1.157
0.350
0.219
6.424
7.282
0.870
1.213
0.306
0.218
5.923
6.990
0.920
1.226
0.344
0.218
TABLE 9.22
Cortical Bone in Tension
50% EFM
E1 (GPa)
E2 = E3 (GPa)
G23 (GPa)
G12 (GPa)
21
32
70% EFM
90% EFM
95% EFM
Pref.
Smooth
Pref.
Smooth
Pref.
Smooth
Pref.
Smooth
4.110
4.779
0.666
0.887
0.375
0.257
3.704
4.649
0.699
0.900
0.427
0.250
4.396
5.138
0.735
0.970
0.378
0.256
3.930
5.009
0.778
0.983
0.432
0.247
4.622
5.408
0.798
1.039
0.398
0.268
4.120
5.286
0.846
1.054
0.455
0.257
4.696
5.485
0.829
1.069
0.423
0.284
4.193
5.371
0.876
1.085
0.482
0.272
Note: Pref., preferential.
TABLE 9.23
Comparison of Values for Cortical Results
6
9
15
21
21.422.1
17.5 1.9
17.8 2.1
19.1 5.4
17.1 3.15
17.0
14.91 0.52
20.55 0.21
16.58 0.32
23.45 0.21
22.5 1.3
a
Method
Source
FEM, 20% mineral volume fraction

Uniaxial tension
Uniaxial tension
Uniaxial tension
Nanoindentation (transverse)
Uniaxial compression and tensiona
Uniaxial compression and tension
Acoustic microscopy (transverse)
Acoustic microscopy (longitudinal)
Nanoindentation (longitudinal)
Nanoindentation (longitudinal)
[1]
[1]
[1]
[1]
[66]
[72]
[73]
[74]
[75]
[76]
[77]
[77]
[77]
[77]
[78]
Only two femurs sampled had statistically significant differences in

compression and tension.
288
TABLE 9.24
Elastic Moduli of Compact Bone (GPa)
Location
Femur
Tibia
Femur
Femur
Method
Ultrasound
Mechanical Testing
Ultrasound
Ultrasound
Preferential Model 95%

EFM, Compression
[76]
11.17
11.17
17.21
3.3
3.3
3.6
0.595
0.298
0.298
[79]
6.94
8.56
18.45
4.91
3.56
2.41
0.495
0.142
0.119
[80]
13.18
14.56
21.67
6.56
5.85
4.74
0.377
0.237
0.225
[81]
18.57
19.43
28.29
8.71
8.71
8.71
0.323
0.209
0.126
This study
6.42
6.39
8.46
0.87
0.97
1.62
0.419
0.191
0.193
Source
E1
E2
E3
G23
G13
G12
12
13
23
data showed that the present model resulted in a lower longitudinal modulus. There were
many factors affecting the result. First, the mineral content assumed in our model was generally lower than those obtained in other studies. Increasing the mineral content would
increase the stiffness, which is shown for the solutions obtained using the finite element
method in Table 9.23. At higher mineral contents, the finite element results approached the
experimental data. The present study utilized a maximum of 25% mineral volume fraction,
which resulted in a longitudinal modulus of approximately 7 GPa. From the finite element
solutions in Table 9.23, the results would predict a longitudinal modulus of approximately
7.5 GPa. This shows that the present simplified multiscale models presented results that
were comparable to other solutions and the difference results from the assumptions made
at various hierarchical levels.
While most results showed either a longitudinal or a transverse modulus, some studies
determined complete material properties, which are also compared to the present data in
Table 9.24. The values depicted previously were the orthotropic material properties. The
present model exhibited similar trends as other experimental tests. The modulus in the 3
direction was greater than both those in the 1 and 2 directions. As discussed previously,
the present model underestimated the properties compared to others.
9.4.7Cancellous Bone
Cancellous bone is a porous macrostructure located inside the cortical bone layer, and it
is composed of small plates and beams. The porous nature of cancellous bone stores bone
marrow and reduces the weight of the skeletal system. The individual beams of the cancellous bone are called trabeculae.
While cortical bone provides more strength and stiffness, cancellous bone is more metabolically active, resulting in more frequent absorption and deposition of bone. The deposition of new bone is related to the applied stresses on the cancellous bone [82], and the
directed deposition makes the outermost layers of trabecular bone oriented parallel to
the longitudinal axes of the trabeculae [7]. The orientations of adjacent layers are slightly
different because they are formed under a different state of stress. The trabecula has a
resulting structure of lamellar layers with a relatively uniform fiber direction. Because the
layers have low fiber angles with respect to the longitudinal direction of the trabecula, a
289
single trabecula is considered to have laminated composite layers at angles of 10, 5, 0,

5, and 10.
While a laminated fibrous composite material model is a good representation of a single
trabecula, it cannot predict the properties of the cancellous bone. The structure of cancellous bone is truly 3-D, and it has much heterogeneity at different anatomical locations [83].
Early models for cancellous bone used beams or plates. The early models are useful as
simple models of cancellous bone, but these models result in asymmetrical properties for
cancellous bone and consider Euler buckling a failure mechanism [82]. Later experimental studies showed that the cancellous bone failed mostly due to microscopic cracking.
Therefore, buckling was not considered as a failure mode for trabeculae [84].
Three-dimensional finite element models were developed using microcomputed tomography to replicate small sections of bone, and two unit cells were proposed to accurately
model cancellous bone [85]. One was a prismatic unit cell, and the other was a tetrakaidecahedral unit cell. It was found that both unit cells accurately represented the mechanical
properties. Furthermore, a study in Reference 86 showed that a complex finite element
model of tetrakaidecahedral cells could accurately represent different levels of bone loss
resulting from aging.
This study used a tetrakaidecahedral unit cell to compute the macroscale properties of
cancellous bone. Several values of trabecular bone properties were used for the tetrakaidecahedral unit cell along with different bone densities. The results of the macroscale properties were then compared to previous experimental data.
9.4.8Trabecular Bone Model
Trabecular bone was modeled in the same way as the cortical bone, but the fiber orientations and volume percentage of each layer were different for the two. The parameters used
for trabecular bone are given in Table 9.25. Due to the lack of Haversian canals in trabecular bone, the void space results from microscopic canaliculi. This void space was assessed
to be 5%, and it was treated as a liquid only responsible for compressive properties.
9.4.9Result of Trabecular Bone
Table 9.26 shows both compressive and tensile material properties predicted using the
trabecular model. The modulus was higher in compression than in tension because of
the liquid contained in the void. The difference between tensile and compressive moduli
became larger as %EFM increased. The results of the present trabecular bone model were
also compared to previous experimental and theoretical results, as shown in Table 9.27.
Early methods tested macroscale cancellous bone to back calculate the material properties
TABLE 9.25
Trabecular Bone Layered Composite Model
Parameters
Fiber Direction
10
5
0
5
10
Volume Fraction
0.2
0.2
0.2
0.2
0.2
290
TABLE 9.26
Trabecular Bone in Compression (Comp.) and Tension (Tens.)
50% EFM
E1 (GPa)
E2 = E3 (GPa)
G23 (GPa)
G12 (GPa)
21
32
70% EFM
90% EFM
95% EFM
Comp.
Tens.
Comp.
Tens.
Comp.
Tens.
Comp.
Tens.
6.047
8.519
1.259
0.658
0.242
0.234
5.099
7.762
1.197
0.623
0.254
0.253
6.427
9.186
1.461
0.672
0.235
0.232
5.158
8.569
1.389
0.636
0.264
0.241
7.848
9.782
1.605
0.717
0.236
0.240
5.223
9.134
1.526
0.673
0.318
0.237
9.451
9.975
1.636
0.773
0.241
0.248
5.245
9.252
1.555
0.720
0.382
0.237
TABLE 9.27
Comparison of Values for Trabecular Results
11.38
12.7 2.0
15
3.81
5.35 1.36
10.4 3.5
14.8 1.4
13.4 2.0
18.0 2.8
11.4 5.6
17.5 1.12
18.14 1.7
Method
Source
Inelastic buckling
Ultrasound (isotropic)
Microhardness (transverse)
Three-point bending
Four-point bending
Tensile test
Ultrasound (isotropic)
Tension and compression tests
Acoustic microscopy (transverse)
[87]
[88]
[89]
[90]
[91]
[92]
[92]
[78]
[73]
[74]
[77]
[77]
of trabecular bone. On the other hand, current methods use microelectronic-mechanical

systems and nanoindentation to directly measure trabecular material properties.
The results for the trabecular bone model were similar to those for the cortical bone
model, resulting in underestimation of the stiffness values. However, the trabecular result
was closer to the mean values of accepted longitudinal and transverse results. The same
discussion for the cortical bone model also applies to the trabecular bone model.
9.4.10Cancellous Bone Model
A tetrakaidecahedral unit cell was used as the cancellous bone model. We developed the
unit cell to analyze cellular materials [93]. The tetrakaidecahedron is a regular truncated
octahedron and was developed by Lord Kelvin in 1887 to model the soap bubble formation in foam. It is a compact and optimal geometry for space filling by repeating itself. It
provides minimum surface area for a given volume.
The tetrakaidecahedron has 8 hexagonal faces, 6 square faces, 36 edges, and 24 vertices.
For the open-cell structures present in cancellous bone, the faces were treated as voids,
and the edges were treated as trabeculae. A simple finite element model was developed
by considering the symmetry of the unit cell [93]. Because the unit cell had open faces,
291
TABLE 9.28
Anatomical Cancellous Bone Indices
BV/TV
Tb.Th (mm)
0.2654
0.2067
0.27
0.27
0.105
0.111
0.34
0.41
0.27
0.172
0.2
0.19
0.36
0.33
0.47
0.186
0.174
Anatomical Location
Source
Femoral neck
Femoral head
Distal femur
Proximal femur
Greater trochanter
Tibia
Radius
Distal tibia
Distal tibia
Average (young age)a
Average (medium age)a
[94]
[95]
[96]
[96]
[94]
[94]
[97]
[98]
[98]
[99]
[99]
Young age was defined as 1639 years old, and medium age
was defined as 4059 years old.
the edges were modeled as beams with the properties of trabecular bone. As a result, the
tetrakaidecahedral unit cell was modeled as a 3-D frame structure consisting of beams.
The cross section of the beam elements depended on the volume fraction provided for the
unit cell [93]. Each beam element had one node at each end with six degrees of freedom per
node. Loads were applied to the model at vertices. No buckling was considered.
The finite element model for the tetrakaidecahedral unit cell yielded isotropic material
properties. In addition, because each beam element was treated as an isotropic material,
two variations were considered: one using compressive material properties and the other
using tensile material properties. This created a bound on the material properties of cancellous bone for each %EFM used. The beam elements of the tetrakaidecahedral model
were represented by the material properties of the trabecular long axis. The properties for
the different %EFM, in tension and in compression, were considered, and the effect of bone
marrow was neglected.
The radius of the beam elements and the volume density of the bone were determined
as follows: A universally accepted set of indices for trabecular structure are bone volume
per tissue volume (BV/TV), bone surface per tissue volume (BS/TV), and bone surface per
bone volume (BS/BV) [83]. These indices helped derive the structure of the trabeculae:
trabecular thickness (Tb.Th), trabecular separation (Tb.Sp), and trabecular number (Tb.N)
[83]. Table 9.28 shows the parameters discussed previously at various anatomical locations.
The diameter of the beams was set to the average thickness of the femoral trabeculae,
which was 0.19 mm.
9.4.11Cancellous Bone Results
The results of the present model were determined using the trabecular results obtained
previously, and their values are tabulated in Table 9.29. As expected, the modulus of the
cancellous bone was greater in compression than in tension. The distal and proximal femur
were stiffer than both the femoral neck and the head. This was expected due to its greater
density. The present predictions are compared to experimental and theoretical results in
Table 9.30. The comparison was good overall.
292
TABLE 9.29
Cancellous Model Elastic Modulus (MPa)
50% EFM
Femoral neck
Femoral head
Distal/proximal
70% EFM
90% EFM
95% EFM
Comp.
Tens.
Comp.
Tens.
Comp.
Tens.
Comp.
Tens.
292
196
300
246
165
253
311
209
319
249
167
256
379
255
390
252
170
259
457
307
469
253
170
260
Note: Comp., compression; Tens., tension.
TABLE 9.30
Comparison of Values for Cancellous Results
Elastic Modulus (MPa)
47 2
300 200
110 90
550 450
Method
Source
Finite element method

Microcompression testing
Compressive testing
Mechanical testing
[100]
[101]
[102]
[103]
9.5Further Adjustments in Models

In this section, we further discuss each hierarchical model for the multiscale analysis to
improve the results and capability. In addition, it determined which hierarchical model
had the greatest influence on macroscale properties. Minerals are the main element to
result in necessary stiffness and strength of the biomaterial. The fiber model showed that
there could be a high level of EFM within living bone. Comparing the linear and twisting
fibril models based on the assessed mineral volume contents, the linear model seemed
to be the proper one for fibril crystal packing. Therefore, the selected model for further
analysis comprised the linear fibril hydroxyapatite packing and 95% EFM. Moreover, the
model was studied for compressive properties. We used the preferential fiber model for
the macroscale hierarchical structure.
We examined further the macroscale longitudinal stiffness, transverse stiffness, and
mineral volume fraction. The results from the present modification were assessed for the
cortical and trabecular bones. Another independent analysis was conducted to assess bone
loss. To examine the effects of bone loss on the material properties of cancellous bone, different bone densities were considered for the macroscale cancellous bone.
9.5.1Modified Hierarchy
Hierarchies in both microscales and macroscales were altered to find out their effects on
the cortical and trabecular properties. The fibrillar subunit was changed at the microscale
level. The diameter of the fibrils and the width of the EFM were varied. On the other hand,
the dimensions of the fiber and the disordered matrix were changed at the macroscale
level. Finally, to define the upper and lower bounds of the composite stiffness, both the
Voigt and the Reuss averages of cortical and trabecular bone are presented [104].
The modification to the hierarchy was conducted one by one, and the results from each
modification are presented and compared. The first change was for the unit cell of the
293
TABLE 9.31
Unit Cell Alteration to Fibrillar Model
Value (nm)
Dimension
a1
a2
b1
b2
c1
c2
Model 1
Model 2
Model 3
50
17
25
3
3
6
50
17
25
1
3
6
50
17
25
1
3
3
fibrillar model. The hydroxyapatite crystal had the same dimensions as previously, but the
volume of the surrounding tropocollagen matrix was decreased as shown in Table9.31.
These changes increased the mineral volume percentage of the fibrillar model. Each
change was called model 1, model 2, and model 3. The next modification was made for
the bone fiber and depended on the volume percentage of both the bone fibril and the
extrafibrillar mineral. Two variations were tested by changing the volume percentage of
the fibril. The original model assumed the fibril volume percentage of 72.25%, which was
changed to either 62.25% or 82.25%. These variations are referred to as model 4 and model
5, respectively.
The only adjustment at the macroscale hierarchies was that of the lamellar bone. The
original model had a 38.44% fiber volume percentage. The fiber volume percentage was
increased such that model 6 had 50% fibers, model 7 had 70% fibers, and model 8 had 90%
fibers.
9.5.2Adjustment Results
Table 9.32 summarizes the results obtained using models 1 through 8. These models altered
only the hierarchy discussed without other changes. From the results, the transverse elastic modulus was greater than the longitudinal modulus for all models. As expected, the
TABLE 9.32
Elastic Moduli of Adjusted Hierarchies (GPa)
Cortical Bone
Voigt
Model
1
2
3
4
5
6
7
8
Original
Trabecular Bone
Reuss Model
Voigt
Reuss
E1
E2
E1
E2
E1
E2
E1
E2
7.05
8.27
10.26
6.95
5.84
7.21
8.93
11.25
6.42
8.03
9.53
11.87
7.87
6.61
8.30
10.49
13.39
7.28
3.75
3.92
4.33
3.58
3.61
3.35
2.75
1.81
3.58
4.48
4.72
5.15
4.33
4.23
4.21
3.81
2.84
4.29
10.28
11.63
14.18
10.79
8.15
10.49
12.29
14.08
9.45
11.19
13.87
17.73
10.64
9.07
11.84
16.07
22.04
9.98
8.04
8.76
10.20
8.15
6.84
7.77
7.68
6.38
7.52
9.00
10.48
12.49
8.60
7.72
9.12
10.34
10.05
8.24
29.97
31.50
42.18
28.03
22.00
27.53
31.89
36.22
25.01
294
increasing mineral volume fraction increased the macroscale stiffness. Reduction in the
tropocollagen matrix within the fibrillar subunit increased the macroscale properties. On
the other hand, an increase in the volume percentage of the fibril within the fiber model
showed an inverse effect on its stiffness because it was due to the extra fibrillar mineral. In
other words, because the fibril had less volume, the mineral occupied a larger volume. An
increase in the size of the bone fibers within the lamellar layers increased the macroscale
modulus.
As far as the Voigt and Reuss models are concerned, the Voigt model assumed isostrain
conditions, whereas the Reuss model assumed isostress conditions. Therefore, the Voigt
model resulted in the upper bound of the material properties, while the Reuss model
yielded the lower bound. This was derived from averaging the stiffness matrices or the
compliance matrices, respectively. The results of the cortical models showed a large difference between the Voigt and Reuss models. This was due to the large theta values of
the rotated layers. The trabecular models showed less variation because the layers were
constrained to small thetas.
Finally, the fibrillar model had the greatest influential hierarchy affecting the macro
scale properties. Small changes in this hierarchy were compounded throughout the upper
scales. Changes to the fibrillar model could be physically embodied by reducing the vertical spacing between hydroxyapatite crystals. The lamellar model was the second-most
influential hierarchy. The relatively small volume percentage of fibers present in this
model produced a large increase in the fiber diameter. This increase in the fiber diameter
decreased the relative proportion of the fibrillar matrix.
9.5.3Optimal Adjustment Results
The multiple hierarchies were adjusted simultaneously to optimize the macroscale results.
Both the fibrillar subunit and the fiber model were modified at the microscale. Although
the unit cell dimensions in model 2 were selected for the fibrillar subunit, the fiber model
was assumed to have a fibril volume percentage of 60%. The lamellar model had a fiber
volume percentage of 70%. The results are summarized in Table 9.33.
The optimized models of the multiscale analysis showed a mineral volume percentage
of 43.82% for the macroscale bone. The values of E1 and E2 for both models were within
the bounds of the tested results. An observation is that the accepted moduli of the cortical bone were greater than those of the trabecular bone. The previous tested data showed
that the cortical values average was greater than the trabecular results average. This study
showed that the trabecular material was stiffer than the cortical material. This difference
resulted from some variations at the hierarchies.
TABLE 9.33
Optimized Macroscale Bone Properties
Cortical
E1 (GPa)
E2 (GPa)
G23 (GPa)
G12 (GPa)
21
32
Trabecular
Voigt
Reuss
Voigt
Reuss
12.71
15.03
0.914
1.62
0.272
0.153
2.82
3.98
0.004
0.004
0.413
0.176
16.72
24.17
2.18
0.52
0.165
0.164
8.91
12.73
0.020
0.019
0.304
0.176
295
One of the reasons was the simplification of the void space within the macrostructures. While the cortical bone models assumed 25% void space due to the combination of
Haversian canals and canaliculi, the trabecular model assumed only 5% void space due to
the lack of Haversian canals. More information is needed for correct void space representation. Another factor relates to the lamellar layers, which should physically be portrayed
as layers that are similar in orientation to the surface layers. However, this assumption
will be unique for each individual strut based on its historical loading. It can be hypothesized that each trabecular strut exhibited different mechanical properties independent of
its dimensions.
9.5.4Bone Loss Results
Bone loss occurs with aging and disease. Some studies identified quantitative differences
between healthy individuals and individuals suffering from bone loss. The differences are
shown with respect to the cancellous indices of BV/TV and Tb.Th. The effects of bone loss
are shown in terms of the parameter in Table 9.34 that shows both a reduction in BV/TV
and thinning of trabeculae. To study the effects of reductions in bone density, both modes
of bone loss were modeled. The cancellous model assumed baseline values of 0.27 and
0.19 mm for BV/TV and Tb.Th, respectively. The models used the optimized material
properties of trabecular bone, as found through Voigt averaging. The baseline cancellous
results were computed for the optimized trabecular properties.
The finite element model for the cancellous bone used a ligament radius and length.
Either the ligament length or radius was varied to reflect the bone loss. In solving for
the cancellous modulus, the BV/TV index was the major parameter. Table 9.35 tabulates
the results, which showed a steady decrease in macroscale stiffness resulting from the
decreasing density. The effect of a 10% reduction in density yielded a 50% reduction in
modulus. Furthermore, the effects of the trabecular thinning may apply more to the failure
of trabecular bone than to the material stiffness.
TABLE 9.34
Cancellous Bone Indices for Individuals with Bone Loss
BV/TV
0.26
Tb.Th (mm)
Anatomical Location
Group Identifier
Source
0.33
0.157
Distal Radius
Proximal tibia
Iliac crest
Osteoporotic postmenopausal females

6079 years old
7090 years old
[98]
[99]
[105]
0.16
TABLE 9.35
Cancellous Bone Material Properties after
Bone Loss
BV/TV
0.27
0.24
0.20
0.16
0.27
0.27
0.27
Tb.Th (mm)
Stiffness (MPa)
0.19
0.19
0.19
0.19
0.17
0.15
0.13
830
689
514
357
830
830
830
10
Multiphysics Analysis of Composite Structures
10.1Introduction
Recently, composite materials have been used increasingly for maritime, aerospace, and
automotive structures. Early use of composites was limited to secondary structures; however, as knowledge and understanding of mechanical characteristics of composites have
grown, more primary load-bearing structures have been fabricated. In recent years, large
composite structures have been incorporated into ships and aircraft to increase operational performance while lowering ownership costs. For example, carbon fiber composite
material provides high strength and stiffness with low weight, which in turn translates to
increased fuel efficiency and increased payload. Further advantages of composites over
metals are lower maintenance costs and resistance to corrosion, which make composites
desirable for maritime applications. While composites provide advantages over metals,
they also come with complex and challenging engineering problems for analysts and
designers to overcome [1]. This chapter focuses specifically on the implications of utilizing
composite structures in maritime applications below the waterline because the structural
behavior is affected by fluid-structure interaction (FSI).
10.2Fluid-Structure Interaction Modeling

The FSI generally has two-way interactions between fluids and composite structures. Fluids
apply pressure loading to structures, while the structures affect the fluid velocity at the FSI
interface. If the effect on the fluid velocity is negligible, the FSI becomes a one-way interaction. The one-way interaction is easier to model because it does not require mutually interactive solutions of both media. In this case, the fluid medium is solved first, and then the
fluid pressure loading is applied to the structures to solve the structural responses. On the
other hand, two-way interactions require simultaneous solutions of both media because
the solutions are influenced by each other. If not solved simultaneously, one medium is
solved first, followed by the solution of the other medium, and the process iterates until
the solutions are converged.
We may consider two types of fluid media. One type of medium includes fluid flow, and
the other type neglects fluid flow. The former problem can be solved using the NavierStokes equation, while the latter is solved using the wave equation. The Navier-Stokes
equation can be modeled using the lattice Boltzmann method, while the wave equation
can be analyzed using the cellular automata (CA) as discussed in previous chapters.
297
298
In modeling FSI, the structure is modeled using the finite element method (FEM).
Especially, the three-dimensional (3-D) solidlike plate/shell elements are useful for modeling plate/shell structures with FSI because the FSI interfaces can be easily defined on
the bottom or the top of the thin structures. If 3-D solid elements are used for a thin shell
structure, the mesh requirement becomes prohibitive in terms of computation costs. On
the other hand, if the fluid is modeled using the FEM or control volume method, the compatibility conditions at the FSI interface can be easily enforced for those techniques use
physical variables such as fluid velocities and pressure as the unknown variables at the
nodal points.
The lattice Boltzmann method uses fictitious particle densities as the nodal variables,
which should be translated into fluid velocities and pressures. Such a translational process
may not be unique and requires special attention. The CA method has some difficulties
in taking derivatives of their own variables, which are needed to satisfy compatibility
conditions. However, both techniques are computationally efficient. Coupling between the
finite element structural model and the lattice Boltzmann or CA fluid model is discussed
in Chapter 6, so their details are omitted here.
10.3Low-Velocity Impact with FSI

This section discusses low-velocity impact on composite structures that are in contact with
water. The structures may be in contact with water on only one side or on both sides. For
the one-side contact case, water can touch the structures inside or outside. For a composite
plate considered in this study, Figure 10.1 illustrates four possible cases. The first case had
no water around the plate; the second case had water on the bottom side of the plate that
was impacted on the top side; the third case had water only on the top side of the plate; and
finally the fourth case had water on both sides of the plate.
Because the fluid motion during the dynamic response of the impacted plate may be
neglected as a simplified model, the fluid domain is considered an acoustic medium
represented by the wave equation. The impactor was modeled as a rigid material with
a given initial velocity. The impact/contact condition was applied between the impactor
and the plate. The plate was modeled using the 3-D solidlike plate/shell element, which
Impact side of
plate
(a)
Air
Impact side of
plate
(b)
Water
Air
Impact side of
plate
(c)
Air
Water
Impact side of
plate
(d)
Water
FIGURE 10.1
Four different cases as a composite plate is in contact with water: (a) no water contact; (b) water on the bottom
of the plate; (c) water on the top of the plate; and (d) water on both sides of the plate.
299
had displacement only as nodal degrees of freedom but did not have rotational degrees of
freedom as presented in Section 2.8.
The acoustic domain was modeled using the CA technique, while the composite plate
was analyzed using the FEM. However, as discussed in Chapter 6, a finite element structural model has some difficulty with direct coupling with the CA. As a result, the immediate neighbor of the structure was modeled using the finite element technique, which was
enclosed by the CA as sketched in Figure 6.20.
The overall analysis was conducted as follows: At a given time, the impact/contact analysis was conducted for the structure with the fluid pressure loading determined previously.
Then, the structural analysis yielded the velocity at the FSI interface, which was applied to
the fluid analysis. The pressure loading at the FSI interface was computed from the fluid
analysis. The computed pressure was compared to the previous pressure loading. If these
were within a given tolerance, the iteration stopped. Otherwise, the process iterated until
convergence.
10.3.1Single-Layer Plate Model
A clamped square plate was studied. The plate was 0.3048 0.3048 m and was 6.35 mm
thick. Because the plate was made of quasi-isotropic layers of composites, it was modeled
as a single layer of the equivalent isotropic material. Both continuous Galerkin (CG) and
discontinuous Galerkin (DG) techniques were used for modeling the composite plate. For
fluid, freshwater was considered. Figure 10.2 compares the transverse displacement of the
center of the plate subjected to a constant concentrated force applied at its center. In the
figure, dry means no FSI, while wet means inclusion of the FSI. The dry structure oscillated
about its predicted static deflection, while the wet structure showed an altered frequency
and varying magnitude as a result of FSI. As studied in Reference 2, the effect of the elastic
modulus and density of the plate was investigated along with FSI as shown in Figures 10.3
and 10.4.
The baseline model shown in Figure 10.2 had a structural density that was 2.7 times
larger than that of the fluid, and it had a frequency ratio of the dry-to-wet response of
1.84 and an amplitude ratio between the first peak and the first trough of the dry-to-wet
response of 2.49. As seen in Figure 10.3, doubling the elastic modulus of the base model
0
104
Displacement (m)
1
2
3
4
5
6
7
Dry
Wet
0
0.002
0.004
0.006
Time (s)
0.008
0.01
FIGURE 10.2
Displacement of a clamped plate with and without fluid-structure interaction.
300
104
Dry
Wet
Displacement (m)
0.5
1
1.5
2
2.5
3
0
0.002
0.004
0.006
Time (s)
0.008
0.01
FIGURE 10.3
Displacement of a clamped plate of doubled elastic modulus of the base model with and without fluid-structure
interaction.
10
Displacement (m)
1
2
3
4
5
6
7
Dry
Wet
0
0.002
0.004
0.006
Time (s)
0.008
0.01
FIGURE 10.4
Displacement of a clamped plate of doubled density of the base model with and without fluid-structure
interaction.
showed a slightly higher-frequency oscillation about a smaller static deflection for the dry
structure, as expected. In this case, the dry-to-wet frequency ratio was 1.86, and its amplitude ratio was 2.15, which was smaller than that of the base case. As seen in Figure 10.4,
doubling the density of the base model resulted in the expected lower-frequency oscillation for the dry structure. Its dry-to-wet frequency ratio was 1.49, and the dry-to-wet
amplitude ratio for the double density case was 2.00.
10.3.2Two-Layer Plate Model
The clamped plates discussed in this section were each 0.3048 0.3048 m and 3.5 mm
thick; they had two thickness layers and 16 elements in each planar direction. The material
301
was a quasi-isotropic E-glass. Initial conditions were zero displacement and zero velocity,
and a constant concentrated force of 1000 N was applied at the center of the plate at the first
time step. The damaged plates had debonding of four elements by four elements between
the two layers at the center of the plate. The debonded area was modeled using the contact
element without friction.
Figures 10.5 and 10.6 display time histories of the displacement and normal strain in the
plane on the bottom of each plate. The dry plate had larger displacement and strain than
the wet plate under this loading condition. Furthermore, the strains at the center of the
plate for both dry and wet cases had greater magnitudes than those at the edge of the damage zone. This is not realistic and suggests that an interface layer is necessary to properly
model debonding in laminated composites and other layered structures.
0
Displacement (m)
0.002
0.004
Dry
Wet
0.006
0.008
0.01
0.012
0.014
2
Time (s)
4
103
FIGURE 10.5
Displacement of a damaged clamped two-layer E-glass plate with and without fluid-structure interaction.
10
Strain
5
4
3
2
Center dry
Center wet
1
0
2
Time (s)
10
FIGURE 10.6
Strain at the center of a clamped two-layer E-glass plate with and without fluid-structure interaction.
302
10.3.3Three-Layer Plate Model

The two-layer model of the previous section did a poor job of reflecting debonding damage within a laminated plate. A three-layer model included a thin interface layer with
properties representing the resin used for the composite between two layers of E-glass,
and it was subjected to the same loading and boundary conditions as used previously.
Responses were calculated for 500 time steps. Figure 10.7 shows that the displacement of
the center of the plate did not reflect the presence or absence of a debonding zone but did
demonstrate the FSI effect in a fashion similar to that of the two-layer model. Figure 10.8
shows that the strain calculated in the E-glass element reflected the presence or absence
of damage mildly. Figure 10.9, on the other hand, shows clearly that the interface layer
was profoundly affected by the presence of a damage zone. To be clear, the strain values
at the centers of the interface layers of the damaged plates were not identically zero, but
they were four orders of magnitude lower than their undamaged counterparts. Figure
10.10 compares the strains in the interface layer at the center and the edge of the potential
debonding zone of the dry structure which was assumed to have no damage or damage.
Likewise, Figure 10.11 compares the strains in the interface layer at the center and the edge
0
Displacement (m)
0.005
0.01
Dry (damaged)
Wet (damaged)
Dry (undamaged)
Wet (undamaged)
0.015
0.02
0.5
1.5
Time (s)
2.5
3
103
FIGURE 10.7
Displacement of a clamped three-layer E-glass plate with and without fluid-structure interaction and with and
without damage.
0.01
Strain
0.008
0.006
Dry (damaged)
Wet (damaged)
Dry (undamaged)
Wet (undamaged)
0.004
0.002
0
0
0.5
1.5
Time (s)
2.5
103
FIGURE 10.8
Strain of a clamped three-layer E-glass plate with and without fluid-structure interaction and with and without
damage at center of lower E-glass layer.
303
104
Dry (damaged)
Wet (damaged)
Dry (undamaged)
Wet (undamaged)
Strain
4
3
2
1
0
0
0.5
1.5
Time (s)
2.5
3
103
FIGURE 10.9
Strain at the center of the interface layer of a clamped three-layer E-glass plate with and without fluid-structure
interaction and with and without damage.
10
10
Center
Zone edge
Center
Zone edge
Strain
Strain
105
6
4
2
2
0
(a)
Time (s)
103
0.5
(b)
1.5
Time (s)
2.5
3
103
FIGURE 10.10
Strains at the interface of a dry clamped three-layer E-glass plate (a) without and (b) with damage.
104
Center
Zone edge
3
2
1
(a)
Center
Zone edge
4
2
0
1
0
105
8
Strain
Strain
10
0
0.5
1.5
Time (s)
2.5
103
0.5
(b)
1.5
Time (s)
FIGURE 10.11
Strains at the interface of a wet clamped three-layer E-glass plate (a) without and (b) with damage.
2.5
103
304
of the debonding zone of the wet structure with and without damage. Once debonding
occurred at the center of the plates, the strain in the interface layer became nearly zero at
the plate center.
10.3.4Comparison to Experimental Data
Some experimental work was conducted to examine the response of composite plates to
low-velocity impact with and without FSI [35]. In general, it was found that, for a given
impact weight dropped from the same height, structures with FSI experienced higher
resultant forces and consequently greater damage than the same structure in dry conditions. The experimental data are compared to the numerical results here.
The vacuum-assisted resin transfer molding technique was used to construct a series of
0.3048 0.3048 m composite plates consisting of 16 layers of E-glass (approximately 3.5 mm
thick in total). The plates were subjected to low-velocity impact forces that resulted from
dropping a 10.8-kg weight from various heights to the center of the plates using the impact
equipment shown in Figure 10.12. The plates were instrumented with strain rosettes at
multiple positions.
Numerical comparison with these experimental data was conducted using a DG structural model consisting of a single layer of plate elements with a discretization of 12 elements in each planar direction. The overall structure had a length-to-thickness ratio of
87:1, and each element had a length-to-thickness ratio of 7.3:1. In this model, the material
properties used were those of E-glass, but the material was treated as quasi-isotropic.
Those nodes closest to the positions of the strain gauges in the experimental work were
compared to the experimental strains. Figures 10.13 through 10.15 are for the dry plate
and Figures 10.16 through 10.18 are for the wet plate, including FSI [6]. The locations for
strain gauges in the plate are shown in Figure 10.19. All plots show reasonable qualitative
agreement between experimental and numerical data. There were many reasons for discrepancies between the numerical and experimental results. For example, the experiment
had neither exact clamped boundary conditions nor homogeneous material properties. In
addition, for FSI, there was some flow motion, which was neglected in the computational
model.
FIGURE 10.12
Impact equipment.
305
0.2
Calculated
Measured
1000 strain
0.2
0.4
0.6
0.8
0
0.005
0.01
0.015
Time (s)
0.02
FIGURE 10.13
Measured versus calculated strain along the x axis in a dry plate at gauge 1.
1000 strain
0.5
Calculated
Measured
0.5
1
0
0.005
0.01
Time (s)
0.015
0.02
FIGURE 10.14
0.2
Calculated
Measured
1000 strain
0.1
0
0.1
0.2
0.005
0.01
0.015
Time (s)
0.02
FIGURE 10.15
1000 strain
0.5
Calculated
Measured
0.5
1
0.002 0.004 0.006 0.008 0.01 0.012

Time (s)
FIGURE 10.16
Measured versus calculated strain along the x axis in a wet plate at gauge 1.
306
1000 strain
0.5
Calculated
Measured
0
0.5
1
1.5
0
0.002
0.004
0.006 0.008
Time (s)
0.01
0.012
FIGURE 10.17
1000 strain
0.8
0.6
0.4
0.2
Calculated
Measured
0.2
0
0.002 0.004 0.006 0.008

Time (s)
0.01
0.012
FIGURE 10.18
Gauge #1
Gauge #2
x
Gauge #3
Gauge #4
FIGURE 10.19
Strain gauge locations.
10.4Vibration with FSI

Vibrational characteristics of composite structures were analyzed while they were submerged in water. The fluid domain was modeled as an acoustic domain by neglecting fluid
motions. The wave equation was solved using CA. On the other hand, the structures were
modeled using the beam and plate/shell elements, which had displacements only as nodal
degrees of freedom.
307
10.4.1Numerical Modal Analysis

Modal analysis extracts modal parameters like natural frequencies, mode shapes, and
damping ratios of a structure. In the experimental modal analysis, multiple sensors were
attached to a structure to be tested. Then, an impulse loading was applied to the structure.
The sensors mostly used were accelerometers because they are sensitive, small, and light
so that they would not alter the mass and stiffness of the original structure. Velocimeters
are bulkier and heavier than accelerometers, so the former sensors were not suitable for a
small structure used in a laboratory.
To measure a mode shape accurately, many sensors must be attached to a structure
with small spacing. On the other hand, the numerical modal analysis can obtain dynamic
responses at every nodal point of the finite element mesh. Therefore, as long as the mesh
is fine enough, an accurate mode shape can be determined from the numerical modal
analysis, for which any dynamic response (e.g., displacement, velocity, acceleration, etc.)
can also be selected.
Once the time domain responses of an excited dynamic structure were obtained, the
fast Fourier transform (FFT) was used to convert the time domain data into the frequency
domain data. Then, the modal parameters were obtained from the frequency domain data
at each nodal point. Reference 7 explained the modal analysis technique to extract the
mode shapes. Therefore, the explanation is omitted here.
10.4.2Verification Study
The results obtained using the numerical modal analysis were compared to some known
solutions. The example cases were vibrations of beams in air without FSI. The first example
case was a cantilever beam. The beam was 1.0 m long, 0.05 m thick, and 0.1 m wide. The
modulus of elasticity was 20 GPa, and its density was 2000 kg/m3. The analytical solutions
for natural frequencies and mode shapes were given in Reference 8. The mode shapes of a
cantilever beam are given in the following equation:

(x) = A(cos x cosh x) + (sin x sinh x) (10.1)
where

A=
sin l + sinh l
(10.2)
cos l + cosh l
m 2
(10.3)
EI
Here, l is the length of the beam, m is the mass per unit length, EI is the beam rigidity, and
is the frequency in radians/second. The coordinate x was measured from the fixed end of
thebeam. The first three mode shapes had l = 1.875104 for mode 1, l = 4.694091 for mode 2,
and l = 7.854757 for mode 3, from which the natural frequency could be determined.
The natural frequencies and mode shapes of the cantilever beam were computed using
numerical modal analysis. In addition, finite element eigenvalue analysis was also conducted to determine the natural frequencies and mode shapes using the mass and stiffness matrices obtained from the finite element models. Forty beam elements were used for
the analysis. Table 10.1 compares the first three natural frequencies among three different
308
TABLE 10.1
Comparison of Natural Frequencies of Cantilever Beam in Air
Analytical
Solution (Hz)
FEM Eigenvalue
Analysis (Hz)
Error of Eigenvalue
Analysis (%)
Numerical Modal
Analysis (Hz)
Error of Modal
Analysis (%)
25.54
160.06
448.19
25.54
160.06
448.19
0
0
0
26.20
160.00
460.00
2.58
0.03
2.63
Mode 1
Mode 2
Mode 3
solutions. Percentage errors are provided in the table compared to the analytical frequencies. The numerical modal analysis resulted in natural frequencies with an error of approximately 2%. The mode shapes obtained from the numerical modal analysis compared well
with the other results from the analytical solution as well as the FEM eigenvalue solution.
They were not distinguishable in the plots.
The second example was a clamped beam at both ends. The geometric data and material
properties were the same as previously measured. The mode shapes could be obtained
from Equation 10.1 except for the constant A, which was
A=
sin l sinh l
(10.4)
cos l cosh l
The first three mode shapes had l = 4.730041 for mode 1, l = 7.853205 for mode 2,and
l= 10.995608 for mode 3, from which the natural frequencies could be determined. The
first three natural frequencies are listed in Table 10.2. The results from the numerical
modal analysis agreed well with the other solutions.
As an example for a vibrational study including FSI, a composite beam clamped at both
ends was considered. The geometric and material properties of the beam were the same
as those used in the previous section. Water had a density of 1000 kg/m3, and the speed
of sound was 1500 m/s. The nonreflected boundary condition was applied to the fluid
domain. The numerical modal analysis was conducted using the multiphysics-based computational techniques described previously. As the first case, a short impulse load was
applied to the clamped composite beam. Then, the natural frequency and the mode shapes
were computed from the numerical modal analysis. To obtain the second rotationalsymmetric mode shape, the impulse load was applied to a finite element nodal point other
than the center for the center was the node of vibration for the second mode.
TABLE 10.2
Comparison of Natural Frequencies of Clamped Beam at Both Ends in Air
Mode 1
Mode 2
Mode 3
Analytical
Solution (Hz)
FEM Eigenvalues
Analysis (Hz)
Error of Eigenvalues
Analysis (%)
Numerical Modal
Analysis
Error of Modal
Analysis (%)
162.52
448.01
878.29
162.52
448.01
878.29
0
0
0
170.00
460.00
910.00
4.6
2.67
3.61
309
The first natural frequency obtained with FSI was 63.64 Hz, which was much smaller
than the first frequency without FSI, as shown in Table 10.2. The effect of FSI reduced
the frequency by more than 60%. Such reduction was comparable to what was measured
in the physical experiment for a cantilever beam. The experimental study in Reference 9
showed an approximately 70% reduction in the first natural frequency. Because the cantilever beam was less constrained than the beam clamped at both ends, the effect of FSI
was slightly larger. The first two mode shapes were compared in Figures 10.20 and 10.21
between the two solutions with and without FSI, respectively. The mode shapes look similar between the two cases. Then, the modal curvatures were compared; these were the second derivatives of the mode shapes and were proportional to bending strains. As shown
in Figures 10.22 and 10.23, there was greater difference in the modal curvatures than in the
mode shapes resulting from the FSI effect, especially for the second mode. Modal curvatures were directly proportional to the bending strains of the plates.
The next study investigated the free vibration of a beam clamped at both ends. As the
initial deformation, the static deflection of the beam subjected to a central force was used,
Normalized modal deflection
0
In air
In water
0.2
0.4
0.6
0.8
1
0.2
0.4
0.6
Normalized beam length
0.8
FIGURE 10.20
Comparison of the first mode shape of a beam clamped at both ends with and without FSI effect.
Normalized modal deflection
In air
In water
0.5
0.5
0.2
0.4
0.6
0.8
FIGURE 10.21
Comparison of the second mode shape of a beam clamped at both ends with and without FSI effect.
310
Modal curvature
In air
In water
0.2
0.4
0.6
0.8
FIGURE 10.22
Comparison of the first modal curvatures of a beam clamped at both ends with and without FSI effect.
80
60
In air
In water
Modal curvature
40
20
0
20
40
60
80
0.2
0.4
0.6
0.8
FIGURE 10.23
Comparison of the second modal curvatures of a beam clamped at both ends with and without FSI effect.
and the initial velocity was set to zero. The beam was freed to vibrate without any external
load. Figures 10.24 and 10.25 show snapshots of deformed beams during free vibration of
the beam at some selected times. Figure 10.24 is for vibration in air, while Figure 10.25 is for
vibration in water. The initial static deflection of the beam is close to its first mode shape.
As a result, the first mode shape was the major vibrational mode. The free vibration in air
looked like more or less the first mode shape of vibration. However, the free vibration in
water contained visible high-frequency modes on top of the first mode. In other words, the
FSI resulted in higher mode shapes to store the vibrational energy. This finding was also
observed in the previous experiment of a composite cantilever beam [9]. Figures 10.26 and
10.27 show the experimentally measured vibrations of a cantilever beam in air and water,
respectively, using the digital image correlation technique. The figures are also snapshots
at selected times. The experimental data also confirmed the high-frequency motion in
water resulting from FSI because the effect of FSI was not uniform over the beam.
311
103
0.2
Beam deflection
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
0.2
0.4
0.6
0.8
FIGURE 10.24
A snapshot of the deformed shape of a beam clamped at both ends with initial static deflection without FSI.
107
Beam deflection
1
0
1
2
3
4
5
6
0.2
0.4
0.6
0.8
FIGURE 10.25
A snapshot of the deformed shape of a beam clamped at both ends with initial static deflection with FSI.
The next set of studies was conducted for a composite plate whose material properties
were the same as the previous ones. The plate was square, 1 1 m, and plate thickness
was 0.02 m. A short-duration impulse force was applied to the composite plate while it was
clamped all around the boundary. The plate had zero initial displacement and velocity.
The deformed shapes of the vibrating composite plate in air and water, respectively, are
plotted in Figures 10.28 and 10.29 at arbitrary selected times.
The deformed shape of the vibrating composite plate in air resembled that of the static
deformation under the central load as shown in Figure 10.28. However, the vibrating composite plate in water showed a different transverse displacement, as shown in Figure10.29,
312
Amplitude (mm)
0
0.05
0.1
0.15
0.2
0
100
150
200
250
50
Distance from fixed end of beam
300
Amplitude (mm)
FIGURE 10.26
A snapshot of the deformed shape of a cantilever beam with initial static deflection without FSI, measured using
the digital image correlation technique.
t= 8 s
0.05
0.1
0
50
100
150
200
250
300
Distance from fixed end of beam

FIGURE 10.27
A snapshot of the deformed shape of a cantilever beam with initial static deflection with FSI, measured using
the digital image correlation technique.
106
Plate deflection
0
1
2
3
4
0
0.2
0.4
0.6
0.8
1 0
0.2
0.4
0.6
FIGURE 10.28
Vibrational shape of a clamped square plate without FSI under impulse load.
0.8
313
107
1
Plate deflection
0.5
0
0.5
1
1.5
2
0
0.2
0.4
0.6
0.8
0.2
0.4
0.6
0.8
FIGURE 10.29
Vibrational shape of a clamped square plate with FSI under impulse load.
where there are local peaks near the four corners of the clamped plate that were not
observed for the vibration in air. Furthermore, the excited mode shapes were compared
between the two vibrations in air and water, respectively. Because the force was applied
at the center, only symmetric modes were excited by the impulse load. The first two mode
shapes in air are shown in Figures 10.30 and 10.31. The first mode shapes in air and water
were close to each other. As a result, the first mode shape in water is not provided separately. However, the next symmetric modes excited by the central impulse force were very
different between the vibrations in air and water. Figure 10.31 shows the second symmetric mode excited in air, while Figure 10.32 shows the counterpart mode in water. The
mode shape shown in Figure 10.32 explains the local peak deformation near the corners as
observed in the experimental results [4,5]. When a clamped composite plate was impacted
Modal deflection
0
0.2
0.4
0.6
0.8
1
1
0.8
0.6
0.4
0.2
0 0
0.2
0.4
0.6
FIGURE 10.30
First mode shape of a clamped square plate under a central impulse force in air.
0.8
314
Modal deflection
1
0.5
0
0.5
1
1
0.8
0.6
0.4
0.2
0.2
0.4
0.6
0.8
FIGURE 10.31
Second symmetric mode shape of a clamped square plate under a central impulse force in air.
Modal deflection
1
0.5
0
0.5
1
0
0.2
0.4
0.6
0.8
0.8
0.6
0.4
0.2
FIGURE 10.32
Second symmetric mode shape of a clamped square plate under a central impulse force in water.
at the center while the plate was in air and water, respectively, the measured strains were
very different near the corners of the plate for the two situations compared to near the
center location [4,5]. That is because the excited mode shapes were different between the
vibrations in air and water.
The next examples examined a clamped plate with a given initial condition without any
external load. Two cases were considered. The first case used the static deformation of the
square plate under central force as the initial displacement along with zero initial velocity. The second case applied the static deformation under a concentrated bending moment
at the center as the initial displacement with zero initial velocity. Those initial displacements were obtained from the static bending finite element analysis of the same composite
plate clamped along the boundary. The same mesh was used for both static and dynamic
analyses. The first case corresponded to the initial deformation close to the first mode of
vibration, which was a symmetric mode. On the other hand, the second case corresponded
to the initial deformation close to the second mode of vibration, which was rotationally
symmetric about one axis.
315
Figures 10.33 and 10.34 compare the snapshots of deformed plates during free vibrations in air and water, respectively, at arbitrary times for the symmetric initial deflection. Likewise, Figures 10.35 and 10.36 compare the plate vibrations in air and water,
respectively, with the rotationally symmetric initial deflection. As shown in the beam
study, the free vibration of the composite plate in water stored the vibrational energy
into modes of higher frequencies, as shown in the figures. All deformed plots did not
consider slopes for smooth deformed shapes for simplicity. As a result, the deformed
plates showed facets.
105
Plate deflection
0
1
2
3
4
1
0.8
0.6
0.4
0.2
0.2
0.4
0.6
0.8
FIGURE 10.33
Snapshot of the deformed shape of the free vibration of a clamped square plate in air with symmetric initial
displacement.
108
Plate deflection
5
0
5
10
15
20
1
0.8
0.6
0.4
0.2
0 0
0.2
0.4
0.6
0.8
FIGURE 10.34
Snapshot of the deformed shape of the free vibration of a clamped square plate in water with symmetric initial
displacement.
316
105
3
2
Plate deflection
1
0
1
2
3
4
5
1
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
FIGURE 10.35
Snapshot of the deformed shape of the free vibration of a clamped square plate in air with antisymmetric initial
displacement.
1.5
105
Plate deflection
0.5
0
0.5
1
1.5
2
0
0.2
0.4
0.6
0.8
1
0.8
0.6
0.4
0.2
FIGURE 10.36
Snapshot of the deformed shape of the free vibration of a clamped square plate in water with antisymmetric
initial displacement.
317
The frequency spectra are plotted in Figures 10.37 and 10.38 for the symmetric initial
displacement, which was obtained from the time history of the central displacement of
the plate. The plot for vibration in air shows the first natural frequency as a sharp peak,
while the frequency spectrum for vibration in water shows the downward shift of the
first natural frequency with a much smoother peak because of contributions from various
frequency components. Furthermore, Figures 10.39 and 10.40 compare frequency spectra
of the plate motion with the rotationally symmetric initial displacement. Almost the same
kind of observation can be made for the two plots except the frequency in Figure 10.39
represents the second mode shape.
10
9
8
Magnitude
7
6
5
4
3
2
1
0
50
100
150
Frequency (Hz)
200
250
FIGURE 10.37
Frequency spectrum of the center displacement of a clamped square plate in air with symmetric initial
displacement.
0.3
Magnitude
0.25
0.2
0.15
0.1
0.05
0
50
100
150
Frequency (Hz)
200
250
FIGURE 10.38
Frequency spectrum of the center displacement of a clamped square plate in water with symmetric initial
displacement.
318
100
90
80
Magnitude
70
60
50
40
30
20
10
0
50
100
150
200
Frequency (Hz)
250
300
FIGURE 10.39
Frequency spectrum of the center displacement of a clamped square plate in air with rotationally symmetric
1
0.9
0.8
Magnitude
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
50
100
150
200
Frequency (Hz)
250
300
FIGURE 10.40
Frequency spectrum of the center displacement of a clamped square plate in water with rotationally symmetric
10.5Fatigue Loading with FSI

10.5.1Problem Description
Let us consider a composite structure that supports vibrating equipment on its top side and
is in contact with water on the bottom side, as shown in Figure 10.41. The water domain is
considered semi-infinite. The corresponding simplified engineering model is presented in
Figure 10.42, with detailed explanations presented next [10].
319
Vibrating equipment
Flexible composite structure
Support of structure
Support of structure
Water domain
Semi-infinite water boundary

FIGURE 10.41
Structure supporting vibrating equipment while in contact with water.
F = Fo sin(t)
Composite beam or plate
Water subdomain modeled using FEM

Water subdomain modeled using CA
FIGURE 10.42
Engineering model to represent a structure supporting vibrating equipment while in contact with water.
The vibrating equipment was simplified as a single mass and a linear spring under a
harmonic motion. Damping was neglected in the model. Neglecting damping would not
affect the objective of the present study because the structural response was compared
with and without the FSI effect under the same condition otherwise. The vibration of the
equipment was modeled as a harmonic force applied to the mass.
The structure was modeled as a beam or a plate/shell depending on the structural
dimensions. Both beam and plate structures are considered here. The FEM was used to
analyze the dynamic response of the beam or plate. Because the beam or plate must be
coupled with the fluid (i.e., water), the beam or plate finite elements were formulated to
have only displacement degrees of freedom at each node like a 3-D solid element, as discussed in Chapter 2. This makes the compatibility at the fluid-structural interface easy.
The boundary of the structure was either simply supported or clamped.
The water medium was assumed to be an acoustic medium represented by the wave
equation. Then, the water domain was broken into two subdomains as shown by broken
lines in Figure 10.42 to achieve the computational efficiency as well as easy compatibility at
the fluid-structure interface, as described in more detail in this chapter. Water subdomain
1 was analyzed using FEM, while water subdomain 2 was solved using the CA technique.
The top boundary of the water domain, represented by thick solid lines, was free while the
320
other boundary, denoted by double solid lines, were nonreflective, like an infinite boundary. Both boundary conditions could be easily represented by CA. In addition, both water
subdomains were properly coupled to have continuity of the solution.
As a harmonic force was applied to the single mass, the dynamic response of the structure
was analyzed. The structure may or may not have been in contact with water during the
harmonic excitation. The results from the two cases were compared to study the effect of FSI.
First, a beam structure was analyzed. The beam was made of a composite, aluminum,
or steel. Their material properties are given in Table 10.3. To avoid the resonance effect,
the vibration equipment had a frequency that was much lower than the natural frequencies of the structures. The beam was 1.0 m long, 0.1 m wide, and 0.02 m thick. Table 10.4
lists the first natural frequency of the beam in air for all three different materials. The
composite beam clamped at both ends had the lowest frequency of 65 Hz. This was the
frequency in air, that is, without the effect of water. In addition, the natural frequency of
the clamped composite beam, including the additional single mass and spring as shown in
Figure 10.42, was calculated using the FEM. The mesh sensitivity study showed that using
20 or more beam elements resulted in the first several natural frequencies being consistent.
The frequencies depended on the spring-to-mass ratio K/M . Figure 10.43 shows the first
natural frequency of the clamped composite beam with the centrally attached mass and
spring (called a beam system from this point) as a function of the spring-to-mass ratio K/M .
The graph is almost linear up to the ratio of K/M = 1000. This suggests the first natural
frequency of the beam system was almost linearly proportional to the ratio of K/M .
When the ratio K/M was approximately less than 500, the first natural frequency of
the clamped composite beam system was smaller than that of the beam only without the
spring and mass. In other words, the attached spring and mass decreased the first natural
frequency of the beam.
The vibrational frequency in water is much lower than that in air because of the added
mass effect. A previous experimental study [9] showed the FSI reduced the natural frequency of an E-glass composite cantilever beam by 70%. In other words, the frequency in
water was 30% of that in air. Metallic beams have smaller reductions in natural frequency
TABLE 10.3
Material Properties
E-glass composite
Aluminum
Steel
Density (kg/m3)
20
70
200
2000
2700
8000
TABLE 10.4
First Natural Frequencies of Beams in Air (Hz)
E-glass composite
Aluminum
Steel
Clamped
Simply Supported
65.09
104.8
102.9
28.68
47.18
45.35
321
150
Frequency (Hz)
100
50
200
400
600
2 (K/M) (1/s)
800
1000
FIGURE 10.43
Plot of first natural frequency of a clamped composite beam in air with centrally attached mass and spring as a
function of the spring-to-mass ratio.
because of a smaller FSI effect. A numerical modal analysis was conducted for the composite beam as well as the composite system as described in the following so as to determine
the natural frequencies.
A short impulse duration was applied to a location of the beam structure, and the dynamic
response was obtained at many locations as a function of time. For example, displacements, velocities, or accelerations can be computed from FEM. Among them, time histories
of nodal displacements were used for the present study. Then, the FFT was conducted to
convert the time domain response to the frequency domain data to determine the natural
frequencies.
For the present clamped composite beam without the spring and mass, an impulse was
applied to near the center of the beam, and the center displacement was obtained for FFT.
Figure 10.44 shows the FFT of the displacement. From the graph, the clamped composite
beam submerged in water had the first natural frequency of 16 Hz, which is approximately
one-fourth of the frequency in air.
The same numerical modal analysis was conducted for the clamped composite beam in
water with centrally attached mass and spring to find the natural frequency. Figure 10.45
shows the FFT plot with the first three natural frequencies of 25 Hz, 51 Hz, and 76Hz,
respectively, for K/M = 100. This ratio of K/M was selected because it resulted in the
first natural frequency of the dry composite beam system to be 16 Hz, like the wet composite beam. The first natural frequency of the wet composite beam increased with the
attached spring and mass, while the second natural frequency was reduced from 187 Hz.
When selecting values for the singles mass M and spring K, the force transmissibility was considered. Assuming the rigid base structure, the force transmissibility TR was
expressed as follows without damping:

TR =
1
(10.5)
1 r2
where r = /n, and is the applied frequency, while n = K/M . To consider the largest force transmission to the structure as the worst case, the frequency ratio r should be
322
0.1
0.09
Absolute value
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
10
20
30
40
50
Frequency (Hz)
60
70
80
60
70
80
FIGURE 10.44
The FFT plot of a clamped composite beam submerged in water.
0.1
0.09
0.08
Absolute value
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
10
20
30
40
50
Frequency (Hz)
FIGURE 10.45
The FFT plot of a clamped composite beam submerged in water with centrally attached mass and spring.
very small compared to unity. In other words, n should be much greater than . With n
chosen to be 100 Hz, the vibrational frequency was set to 10 Hz, which was lower than
the first natural frequencies for all the cases studied previously for the clamped composite
beam. All the plots that follow were normalized for the unit force, that is, Fo = 1N.
With the selected values, the composite beam was analyzed. The bending strain at the
center of the beam was computed for the composite beam with its time history as plotted
in Figure 10.46. Likewise, the bending strain at the boundary is also plotted in Figure 10.47.
The bending strains were calculated on the beam side in contact with water. The deformation of a clamped beam has opposite curvatures at the center and the boundary so that the
bending strains have opposite signs, as seen in Figures 10.46 and 10.47. The magnitudes of
the bending strains at both elements were similar to those without water. The FSI effect
323
106
w/o FSI
w/ FSI
3
2
Strain
1
0
1
2
3
4
0.05
0.1
0.15
0.2
0.25
Time (s)
0.3
0.35
0.4
FIGURE 10.46
Strain at the center element for a clamped composite beam subjected to a 10-Hz vibrating force.
106
4
3
Strain
2
1
0
1
2
3
w/o FSI
w/ FSI
4
5
0
0.05
0.1
0.15
0.2 0.25
Time (s)
0.3
0.35
0.4
FIGURE 10.47
Strain at the boundary element for a clamped composite beam subjected to a 10-Hz vibrating force.
significantly increased the bending strains in the composite beam. The bending strain was
more affected by FSI at the boundary than at the center. The maximum peak strain was
increased by approximately 50% at the center and 85% at the boundary because of FSI.
Also, the variation in peaks and valleys of the strain history was more profound with FSI.
The result clearly suggests that the FSI effect on the composite beam in contact with water
will significantly reduce the fatigue life of the structure.
To further understand the FSI effect, the force induced in the spring is plotted in Figure
10.48. The figure shows that the force in the spring was not affected by the FSI. The velocity
324
Spring force (N)
3
2
1
0
1
2
w/o FSI
w/ FSI
3
4
0
0.05
0.1
0.2 0.25
Time (s)
0.15
0.3
0.35
0.4
FIGURE 10.48
Spring force for a clamped composite beam subjected to a 10-Hz vibrating force.
at the center of the composite beam is plotted in Figure 10.49. The FSI influenced the velocity significantly. Interestingly, the plot of the central deflection of the beam was similar to
the strain plot at the center element as shown in Figure 10.46. Therefore, the deflection plot
is omitted here.
Both steel and aluminum beams were also tested under the same condition; their
responses are plotted in Figures 10.50 through 10.53. The steel beam had almost no FSI
effect on both the bending strain and velocity as seen in Figures 10.50 and 10.51, respectively. Likewise, the aluminum beam also had a negligible FSI effect on the bending strain,
as shown in Figure 10.52. The time history of the central velocity was close for the with
and without FSI conditions in terms of the phase and magnitude. However, FSI induced
Velocity (m/s)
103
w/o FSI
w/ FSI
0.05
0.1
0.15
0.2
0.25
Time (s)
0.3
0.35
0.4
FIGURE 10.49
Velocity at the center for a clamped composite beam subjected to a 10-Hz vibrating force.
325
107
Strain
w/o FSI
w/ FSI
5
0
0.05
0.1
0.15
0.2
0.25
Time (s)
0.3
0.35
0.4
FIGURE 10.50
Strain at the center for a clamped steel beam subjected to a 10-Hz vibrating force.
104
Velocity (m/s)
w/o FSI
w/ FSI
1
0
0.05
0.1
0.15
0.2
0.25
Time (s)
0.3
0.35
0.4
FIGURE 10.51
Velocity at the center for a clamped steel beam subjected to a 10-Hz vibrating force.
additional high-frequency responses in the velocity of the aluminum beam, like the composite beam. Such a high-frequency response was also observed during the free vibrational experiment of a composite cantilever beam [9].
Comparing the three different materials, the composite beam showed the largest FSI
effect because of its lowest density and stiffness. In the following study, some parameters
in the system were varied to determine which parameter yielded the most critical FSI
effect. As the first parameter, the equipment vibrating frequency was varied from 10 to
5 Hz. Comparison of Figures 10.46 and 10.54 suggests that the lower vibrating frequency
resulted in a little less FSI effect on the bending strain at the beam center because 10 Hz is
closer to the natural frequency of the composite beam in water.
326
106
1
w/o FSI
w/ FSI
Strain
0.5
0
0.5
1
0
0.05
0.1
0.15
0.2 0.25
Time (s)
0.3
0.35
0.4
FIGURE 10.52
Strain at the center for a clamped aluminum beam subjected to a 10-Hz vibrating force.
104
w/o FSI
w/ FSI
Velocity (m/s)
4
2
0
2
4
0
0.05
0.1
0.15
0.2
0.25
Time (s)
0.3
0.35
0.4
FIGURE 10.53
Velocity at the center for a clamped aluminum beam subjected to a 10-Hz vibrating force.
To further investigate the effect of the exciting vibrational frequency on the dynamic
response of the composite beam, the vibrational frequency was varied gradually, and
the ratio of the maximum dynamic deflection to the static deflection at the center of the
clamped composite beam was computed as plotted in Figure 10.55. The figure shows two
resonance frequencies for the beam submerged in water and one resonance frequency for
the dry beam up to 35 Hz. Those resonance frequencies agreed with the natural frequencies discussed previously. Comparing the deflection ratios between the conditions with
and without FSI, as shown in Figure 10.55, clearly suggests that the FSI effect significantly
increased the deflection of the composite beam (i.e., the greater bending strains).
The following study changed the boundary condition of the beam: The beam was simply
supported instead of clamped. The vibrating frequency was 5 Hz so it was much lower
than its natural frequency in water for the simply supported beam. Bending strains at
327
106
Strain
1
0
1
w/o FSI
w/ FSI
2
3
0.05
0.1
0.15
0.2
0.25
Time (s)
0.3
0.35
0.4
FIGURE 10.54
Strain at the center for a clamped composite beam subjected to a 5-Hz vibrating force.
Deflection ratio
15
10
w/o FSI
w/ FSI
10
15
20
Frequency (Hz)
25
30
35
FIGURE 10.55
Plot of the deflection ratio as a function of the exciting vibrational frequency.
the center and the boundary are plotted in Figures 10.56 and 10.57, respectively. The more
flexible boundary, such as the simply supported one, yielded a much larger difference
between the strains with and without the FSI effect. Such a difference was more profound
at the simply supported boundary, as seen in Figure 10.57. The FSI effect resulted in a
bending strain nearly three times greater at the boundary along with very-high-frequency
components in the strain-time history.
An experimental study confirmed the numerical results qualitatively. For example, threepoint bending tests were conducted with cyclic loading while beams were in air or in
water, respectively, under the same loading conditions. The numbers of cycles to failure
were compared between the dry and wet tests. Table 10.5 shows that there was a significant reduction in the number of cycles to failure in water due to FSI. The failure cycle was
about 50% greater in air than in water.
328
106
Strain
w/o FSI
w/ FSI
0.05
0.1
0.15
0.2 0.25
Time (s)
0.3
0.35
0.4
FIGURE 10.56
Strain at the center for a simply supported composite beam subjected to a 5-Hz vibrating force.
106
w/o FSI
w/ FSI
Strain
0.5
0
0.5
1
0
0.05
0.1
0.15
0.2 0.25
Time (s)
0.3
0.35
0.4
FIGURE 10.57
Strain at the boundary for a simply supported composite beam subjected to a 5-Hz vibrating force.
TABLE 10.5
Ratio of Number of Cycles to Failure in Air
versus in Water under Cyclic Loading
Average of ratios
Standard deviation
10 Hz
5 Hz
1.43
0.22
1.52
0.16
Finally, a clamped composite plate was also studied. The material was the same E-glass
composite as used for the beam, and its dimension was 1 1 m and it was 0.02 m thick. The
vibrating equipment was located at the center of the square plate with the same single mass
and spring ratio of K/M = 1000 as before. The first natural frequency of the dry composite plate without the single mass and spring was computed from the following formula:
329
Eh2
(10.6)
L4 (1 2 )
n = 1.655
where E and are the elastic modulus and Poisson ratio, is the volume density, and L and
h are the length and thickness of the square plate, respectively. The natural frequency of
the dry plate without the single mass and spring was 110 Hz (from Equation 10.6). When
the spring and mass were included, the natural frequency was computed from the eigenvalue analysis of the finite element model, and it became 16 Hz. The natural frequency of
the plate in contact with water was also computed using the numerical modal analysis,
which resulted in a frequency of 36 Hz. As a result, the exiting frequency of the equipment
was selected as either 10 or 20 Hz.
The transverse deflection was computed at multiple locations on the plate and compared between the dry and wet vibrations. The selected locations on the plate are shown
in Figure10.58. The ratio of the maximum transverse deflection of the wet plate to the
dry plate is provided in Table 10.6 for two different excitation frequencies. The results
showed that the largest difference between the wet and dry plate deflections occurred at
Loc. D
Loc. E
Loc. F
Loc. B
Loc. C
Loc. A
FIGURE 10.58
Locations on the clamped plate where displacements were compared between dry and wet vibrations.
TABLE 10.6
Ratio of Maximum Displacement between
the Wet and Dry Clamped Composite Plate
(Please see Figure 10.58 for the locations)
Excitation Frequency
Location
A
B
C
D
E
F
10 Hz
20 Hz
1.20
1.14
1.16
1.11
1.14
1.16
1.41
1.31
1.34
1.24
1.29
1.33
330
106
w/o FSI
w/ FSI
Displacement (m)
2
1
0
2
3
0
0.05
0.1
0.15
Time (s)
0.2
0.25
0.3
FIGURE 10.59
Displacement at the center of a clamped composite plate subjected to a 20-Hz vibrating force.
location A in Figure 10.58, while the least difference occurred at location D. These data
suggest that the FSI effect was larger at a location closer to the clamped boundary and
smaller at a location nearer the plate center. This statement can also be confirmed by
comparing locations B and C, as well as the locations D, E, and F, respectively. At location
A close to the boundary corner, the deflections of the wet plate were 20% and 41% greater
than that of the dry plate for the two different excitation frequencies, respectively. The
vibrating equipment with 20 Hz resulted in more than a 40% increase in the deflection
at location A. The transverse deflection at the center of the plate is plotted in Figure 10.59
to illustrate the vibrational motion of the plate.
10.6Hydrodynamic Loading
In this section, transient fluid flow around a flexible composite plate is studied [11]. Because
the fluid flows, the Navier-Stokes equation was used to model the fluid domain. The details
of the computer models and their results are discussed.
10.6.1Model Description
A box-shape structure 1 m3 was considered. The front side was a flexible E-glass composite plate that interacted with fluid flow, while the remaining five sides were assumed to
be rigid. The box-shape structure was varied further in terms of the shape and size for a
series of parametric studies. The structural element size remained relatively constant for
different cases.
The fluid domain was of a sufficient size to minimize interference from the sides of
the fluid domain. The fluid domain extended 1 m beyond the cube from its top, sides, and the
rear face and 2 m from the front face. The fluid domain had dimensions of 3 3 4 m. The
331
Velocity (m/s)
2.5
2
1.5
1
0.5
0
0.5
1
Time (s)
1.5
FIGURE 10.60
Inlet velocity profile of the base model.
structure was surrounded by water only on its outside. There was no fluid inside the box.
The water was assumed to have a density of 1000 kg/m3 and a kinematic viscosity of
1.0 10 6 m2/s. The Reynolds number for the base model of the box structure was 2.0 106.
The interfaces between the structure and fluid were defined as follows: The front interface had the FSI. The other five sides of the interface had no-slip boundary conditions
because the structural parts were held fixed in space. The inlet side of the fluid domain
had a prescribed velocity as a function of time. The outlet of the fluid domain had a prescribed constant pressure. Full-slip boundary conditions were also applied to all other
exterior surfaces of the fluid domain to mitigate boundary-layer effects from the fluid
domain onto the cubic structure.
The model described here was called the base model. The base model considered transient
fluid flow at the inlet under a constant acceleration for the first 0.5 s until a terminal velocity
of 2 m/s was reached, at which point the fluid velocity remained constant for the duration
of the simulation, as shown in Figure 10.60. A time step size of 0.01 s was used for plotting
the results even though the actual time step size used for computation was much smaller.
10.6.2Numerical Results with Constant Acceleration
As the first example, the flexible front plate of the base model was replaced by a rigid plate
so that no FSI existed at the front plate. Then, the two models with and without FSI were
compared to examine the effect of FSI on fluid flow. The numerical results showed that the
center node of the flexible base model had a maximum 45-mm inward displacement at
0.09 s and a maximum 15-mm outward displacement at 0.58 s followed by a steady-state
5-mm inward displacement at 2.0 s. These local maximum/minimum and steady-state
displacement should have the largest effect on fluid flow around the cube.
At a time of 0.09 s, the inward deflection of the deformable plate minimized overall resistance on the upstream flow field and allowed upstream fluid to progress at a faster average velocity than the rigid model. The maximum outward deflection of the flexible plate
occurred at 0.58 s and gave the cube a slight streamlined effect that resulted in an average
velocity around the cube of 2.32 m/s. The average fluid velocity of the rigid body around
the cube at 0.58 s was 2.22 m/s, slower than the flexible model. The velocity contours
332
between the two models at steady state (i.e., at 2.0 s) were virtually identical, with only
minor velocity differences of approximately 0.002 m/s.
Once the effect of FSI was investigated using the base model, the flexible base model was
further analyzed to investigate the effect of transient flow on the flexible front plate. From
this point in the discussion, the term base model refers to the box with the flexible front
plate. To determine the maximum stress/strain in the structural domain caused by transient flow, it was necessary to first verify the location of peak stress or strain. The numerical result showed that the maximum stress/strain occurred near the clamped boundaries.
One of the four maximum strain locations was selected, and from this point is called the
boundary point. Figure 10.61 compares the elastic equivalent strains at the boundary point
and the center as a function of time. The elastic equivalent strain is like the von Mises
stress in terms of strain components. This plot shows that the boundary point had a much
highest strain value than at the center.
The ratio of the maximum transient value to the steady-state value of the elastic equivalent
strain was a little more than 8 at both the center and the boundary nodes. This ratio remained
the same for the von Mises stress as well as the maximum displacement, as expected. Nodal
acceleration and displacement at the center are plotted in Figure 10.62. The plate displacement corresponded closely to the acceleration, with an expected minimal lag between the
two variables. There was an average 0.04-s delay of the local maximum and minimum
displacement values to the acceleration values. Peak values for structural stress and strain
occurred during times when displacement was at either a local maximum or minimum.
Last, at 0.5 s, the fluid acceleration suddenly changed to zero, as shown in Figure 10.60, and
there was an immediate response in structural nodal acceleration as well as in displacement.
Fluid pressure at the center of the plate was also evaluated from the base model and is
plotted in Figure 10.63. In addition, the average fluid pressure over the entire flexible plate
is also shown in the same plot. Multiplication of the average pressure by the plate surface
area resulted in the total fluid force on the plate. The average pressure plot closely matched
the pressure plot at the center node. The ratio of the maximum to the steady-state values
was also computed for the two kinds of fluid pressure (i.e., pressure at the plate center and
the average pressure). The ratio for the fluid pressure at the plate center was close to that
obtained for the stress and strain values (i.e., a little over 8). However, the ratio for the average pressure was 14% greater than that for the central pressure.
Elastic equivalent strain
0.004
Boundary
Center
0.003
0.002
0.001
0
0.5
1
Time (s)
1.5
FIGURE 10.61
Plot of the time history of elastic equivalent strain at two locations of the front plate of the base model.
333
Acceleration (m/s)
Displacement
0.06
200
0.04
100
0.02
0
100
0.02
200
0.04
300
0.2
0.4
0.6
0.8
1
1.2
Time (s)
1.4
1.6
1.8
Displacement (m)
Acceleration
300
0.06
FIGURE 10.62
Plot of the time history of acceleration and displacement at the center of the base model.
30,000
Center press.
Avg. press. interface
0.5
1
Time (s)
Pressure (Pa)
20,000
10,000
0
10,000
20,000
1.5
FIGURE 10.63
Plot of fluid pressure (press.) at the center and average (Avg.) fluid pressure over the entire plate of the base
model.
One more thing to note here is that the negative pressure in Figure 10.63 suggests potential local cavitation. However, such cavitation did not affect the initial maximum transient response because it occurred before the cavitation. Likewise, the final steady-state
response long after cavitation would not affected by the cavitation. As a result, potential
cavitation was neglected in this study.
10.6.3Numerical Results with Nonlinear Acceleration
In the next study, the inlet fluid velocity was varied as shown in Figure 10.64. The first case
modeled a monotonically increasing acceleration, and the second case was a monotonically decreasing acceleration. Together, they showed the effects of nonlinear acceleration
during the period of velocity variation. The monotonically increasing acceleration closely
resembled the profile of an accelerating ship, where acceleration increases with time. On
the other hand, the monotonically decreasing model was a near mirror image.
334
Base
Monotonic decreasing
Monotonic increasing
Velocity (m/s)
2.5
2
1.5
1
0.5
0
0.5
1.5
1
Time (s)
FIGURE 10.64
Inlet velocity profiles of monotonically varying accelerations.
As seen in Figure 10.64, the monotonically decreasing model had an almost-linear acceleration for the first 0.25 s, and the plate responded with oscillations similar to the base model,
with a linear acceleration as shown in Figure 10.65. However, the monotonically decreasing
model had a larger slope than the base model, which resulted in oscillations with a higher
amplitude. During the final 0.25 s of the velocity transient, acceleration slowly reduced to
zero in the monotonically decreasing model. This caused the oscillations to be reduced prior
to achieving the steady-state velocity so that the amplitude of stress became lower for the
monotonically decreasing model than for the base model. On the other hand, the monotonically increasing acceleration yielded an almost-linear increase in the center displacement.
The monotonically decreasing model had the largest ratio of peak to steady-state stresses,
which was nearly 14; the monotonically increasing model had a ratio of slightly larger than
10, and the linear model had a ratio of 8. This was expected based on the initial linearity of the fluid acceleration, with a greater slope than the base model. The monotonically
decreasing and linear models reached peak displacements in the negative direction at 0.08
and 0.09 s, respectively. The monotonically increasing model reached a maximum inward
Base
Displacement (mm)
50
Monotonic decreasing
Monotonic increasing
50
100
0.5
1
Time (s)
FIGURE 10.65
Plot of displacement at the center for monotonically varying accelerations.
1.5
335
displacement at 0.5 s, and the ratio was higher than for the base model because its slope
was higher during the second half of the transient.
When a monotonically increasing acceleration is followed by a monotonically decreasing acceleration, such a combined velocity profile portrays a transient velocity curve with
an inflection point and reduced acceleration slopes at both ends. This combination brings
the advantages of each curve and avoids the initial transient oscillation while encouraging
rapid oscillation decay prior to reaching steady-state velocity. This type of transient velocity curve is also most representative of a ship at sea and proves that the attributes of such
a transient acceleration are beneficial for the structure.
10.6.4Numerical Results with Free Surface
The boundary condition applied to the top surface of the fluid domain was changed in this
study. Previously, it was the free-slip boundary, but it was changed to the free boundary
with zero pressure. Then, the structure was placed inside the fluid with the varying depth
from the free surface. The n-meter model means the distance from the free surface of the
fluid domain to the top surface of the structural box is n meters.
Although the time for peak pressure on the 1-m model occurred at 0.07, which is slightly
earlier than that for the base model without the free fluid surface, the 1-m model resulted
in a far different response due to the significant effect of the free surface. Fluid along the
upper half moved freely and resulted in much lower pressures above the cube. This also
reduced the hydraulic forces applied to the fluid-structure interface.
The base model represented an infinite water depth, while the 1-m model represented
the shallow-water depth because the hydrostatic loading was neglected in the study. The
2-m and 3-m models showed increased fluid pressure in the region above the plate, indicating that the rise in depth increased fluid pressure and stress on the front plate. Figure 10.66
compares the central displacement of the plate with different water depths. The stress and
strain responded similar to the displacement for different water depths.
Comparing the maximum to the steady-state values was also conducted for the base, 1-m,
2-m, and 3-m models. The base model resulted in the highest ratio and was symmetric about
the center. The 1-m model had the smallest ratio at the upper boundary point, and the ratio
increased along the plate toward its bottom boundary. The ratios for the 2-m and 3-m models
Base
Displacement (mm)
30
1-meter
3-meter
2-meter
30
60
0.5
1
Time (s)
1.5
FIGURE 10.66
Comparison of the central displacement with different water depths from the free surface.
336
Infinite depth
Max. to steady state ratio
10
8
6
4
2
0
5
6
7
8
9
Distance from surface (m)
10
11
12
FIGURE 10.67
Plot of the ratios of the maximum to the steady-state stress/strain against the water depth from the free surface.
also tracked accordingly with depth. Additional models with different depths were also
analyzed to draw a curve as shown in Figure 10.67, which suggests that any water depth
greater than 9 m can be treated as the infinite depth in a practical sense when the water
depth effect is considered on the ratio of the maximum to steady-state values. The 1-m model
had a ratio of 2 approximately, and the ratio increased as a function of the water depth until it
reached a plateau at around 9 m. This result was similar with the data provided in Reference
12, which investigated energy scavenging from tidal forces and reported that normalized
velocity deficits in the water column decreased to near zero as depth exceeded 7 m.
10.6.5Numerical Results with Different Material Properties
This case considered different composite materials subjected to the same boundary conditions and acceleration profile as the base model. The base model had a density of 2000 kg/m3
and a Youngs modulus of 20 GPa; the first and second models changed Youngs modulus
to 50 GPa and 100 GPa, respectively. The third model changed the density to 3000 kg/m3
and the Youngs modulus to 50 GPa.
Peak stresses for the base model occurred at 0.08 s; all others occurred at 0.07 s. The three
models (basic model, model at 2000 kg/m3/50 GPa, model at 2000 kg/m3/100 GPa) had the
same mass density but different elastic moduli. Figure 10.68 compares the fluid pressure
at the center of the front plate. It shows almost the same magnitudes but small changes
in the frequency of the response. Even though the pressure had a small variation, the
plate displacement at the center was influenced by the plate material properties, as shown
in Figure 10.69. As expected, the lower modulus resulted in higher displacement if the
structure had the same density. However, the resultant stress was higher with the higher
modulus, as shown in Figure 10.70.
On the other hand, the two models (with 2000 kg/m3/50 GPa and with 3000 kg/m3/
50 GPa) could be compared to determine the effect of the mass density. The results suggest that the lower density resulted in greater strain and stress. The lower-density model
showed approximately 15% higher stress than the higher-density model. This indicates
that a lighter composite structure can have greater stress and strain under the same hydrodynamic loading, which should be considered in designs of marine composite structures.
337
Base
Pressure (Pa)
30,000
2000/50
2000/100
3000/50
20,000
10,000
0
10,000
0.5
1
Time (s)
1.5
FIGURE 10.68
Comparison of the pressure at the center of the plate with different material properties.
Base
Displacement (mm)
30
2000/50
3000/50
2000/100
30
60
0.5
1
Time (s)
1.5
FIGURE 10.69
Comparison of displacements at the center of the plate with different material properties.
Base (2000/20)
von Mises stress (MPa)
150
2000/50
3000/50
2000/100
120
90
60
30
0
0.5
1
Time (s)
1.5
FIGURE 10.70
Comparison of von Mises stresses at the center of the plate with different material properties.
338
When the ratios of the maximum to steady-state stresses were compared for all models,
the base model had a slightly higher ratio than the other models. However, the difference
was less than a couple percent, so that it was negligible in the practical sense because material properties influenced both the transient peak stress/strain and the steady-state stress/
strain at almost the same proportion.
10.7Hydrodynamic Ram
10.7.1What Is Hydrodynamic Ram?
Hydrodynamic ram (HRAM) refers to the damage process resulting from high pressures
generated when a projectile with a high velocity penetrates a compartment or vessel containing a fluid [13]. The large internal fluid pressure that acts on the walls of the fluid-filled
tank can result in severe structural damage, especially at the entrance and exit walls. The
study of HRAM effects on fuel tanks used on military aircraft is one good example.
In most nonexploding projectile impacts with penetration and traveling through a fluidfilled tank, the HRAM phenomenon can be described in four distinct phases [14]:
Shock phase: initial impact of a projectile into an entry wall of a fuel tank
Drag phase: movement of a projectile through fluid
Cavitation phase: development of a cavity behind the projectile as it moves through
the fluid and the subsequent cavity oscillation and collapse
Exit phase: projectile penetrates the exit wall and leaves the tank only when there
is sufficient energy remaining
Each phase contributes to structural damage of the tank walls via a different mechanism, and the extent of damage depends on numerous factors, such as projectile shape
and velocity, fluid level in the impacted tank, obliquity of impact, and fuel tank material.
The amount of structural damage can be significant, with large-scale peeling and tearing
of the entry and exit walls.
10.7.2Numerical Models
The HRAM model consisted of three parts: tank structure, projectile, and fluid inside the
structure. The simulation of HRAM required a very fine Euler mesh and small sampling
times to capture the propagation of shock waves in the fluid. For a simplified computational model, a generic cubic tank 200 200 200 mm, impacted by a 10-mm diameter spherical projectile was studied. Some parametric studies were also conducted of the
effect of the selected parameter on HRAM.
The projectile impacting at the center of the entry wall of the tank was a rigid sphere
with a mass of 4 g and 10-mm diameter. One reason for selecting a spherical projectile
was to prevent tumbling of the projectile during the drag phase, which would result in
significant pressure fluctuations in the fluid, causing an erratic response to the structure.
As shown in Figure 10.71, two models were constructed for this study [15]. The first
model was for the investigation of the shock phase of the HRAM with the projectile outside the tank, impacting the entry wall at a prescribed velocity. For this model, model 1, the
339
Model 1
Model 2
x
FIGURE 10.71
Schematic of models 1 and 2.
displacement of the tank walls due to projectile impact and the subsequent ram pressure
of the propagating hemispherical shock wave in the fluid from the impact point would be
of interest.
For the second model, model 2, the initial starting position of the projectile was at the
inner surface of the entry wall of the impact point to simulate the projectile movement
just after penetrating the entry wall. The initial velocity of the projectile was less due to
retardation of the projectile by the entry wall. Model 2 would be used to study the fluid
pressures and the tank wall response during the drag phase. By considering two models separately, the complex penetration process in the entry wall could be neglected. In
addition, we could study how each phase affected the dynamic responses of the system
individually even though the actual HRAM included both of them, one after the other
sequentially.
The structure was constrained at the bottom face of the box shape and stationary in
the beginning. The fluid was also stationary, while an initial velocity was assigned to the
projectile. For model 1, the initial velocity of the projectile was 300 m/s; the velocity was
250m/s for model 2. An important aspect of the FSI problem is the coupling of the interfaces between the structure and fluid mesh. In both models, FSI was defined at the interface of the structural tank and the internal fluid. In addition, the contact/impact condition
was applied to the sphere and the outside of the structure for model 1. On the other hand,
the projectile was coupled to the fluid for model 2.
10.7.3Numerical Results
The baseline model 1 simulation was set up for a 100%, 80%, or 60% water-filled tank
impacted without penetration at the center of the entry wall by a spherical rigid projectile with an initial velocity of 300 m/s [15]. Even though this was a hypothetical situation
because the projectile would likely penetrate the entry wall in an actual experiment, this
simulation provided some insight to the tank wall behavior during the initial shock phase
of the HRAM event. The event was simulated for 1 ms with a sampling rate of 20 s for
data collection to plot the event time history. For comparison, the following discussion
compares model 1 to an empty tank impacted under the same conditions.
The entry wall x displacement is plotted at the center of the plate in Figure 10.72. The
x direction corresponds to the major component of the entry wall because the direction
of projectile velocity impacting the entry wall was in the positive x direction. It can be
observed that the peak displacement of the entry wall for the 100% filled baseline model
1 was higher at around 9 mm as compared to 7 mm for the empty tank. An interesting
340
1.00E02
XDIS - empty
Displacement (m)
5.00E03
XDIS - 100%
0.00E+00
0.00E+00
2.00E04
4.00E04
6.00E04
8.00E04
1.00E03
5.00E03
1.00E02
Time (s)
FIGURE 10.72
Entry wall x displacement (XDIS) at the center for model 1.
phenomenon observed for model 1 is the entry wall displacing in the negative x direction
at around 0.06 ms after impact, which indicates the entry wall bulging outward. The x
component velocity indicates a much larger peak value of around 210 m/s in the negative x
direction right after projectile impact. This corresponded to the time when the entry wall
started to bulge. The effective entry wall stress (i.e., the von Mises stress) reached a higher
peak value for model 1 but over a shorter duration of time than for the empty tank.
The exit wall response to HRAM was of main interest in this study as it is an area where
main structural components and load-bearing members are likely to be located. Graphs
for exit wall response were plotted from data collected from the center node of the exit
wall panel. The x displacement plot in Figure 10.73 shows a peak displacement of around
2 mm experienced by the exit wall at the end of the simulation, a value much higher than
5.00E03
Displacement (m)
4.00E03
3.00E03
XDIS - empty
XDIS - 100%
2.00E03
1.00E03
0.00E+00
0.00E+00
1.00E03
2.00E04
2.00E03
4.00E04
Time (s)
FIGURE 10.73
Exit wall x displacement (XDIS) at the center for model 1.
6.00E04
8.00E04
1.00E03
341
that experienced by the empty tank. The exit wall for model 1 started deforming earlier, at
approximately 0.13 ms. This was approximately the time when the initial shock wave due
to projectile impact at the entry wall impinged onto the exit wall, causing it to displace.
The presence of fluid in the tank actually resulted in a much smaller velocity and effective stress at the exit wall. Peak stress at the center of the exit wall registered a much lower
value of approximately 100 MPa, as compared to 500 MPa for the empty tank.
Besides the propagation of a shock wave through the aluminum tank structure, a hemispherical shock wave was observed to propagate in the fluid toward the exit wall. This ram
pressure generated by the impact of the projectile in the shock phase was computed at the
center of the fluid. The peak pressure of 1.6 MPa occurred for the 100% filled tank in the
beginning. On the other hand, the 60% filled tank showed the peak pressure of 1.4 MPa
around 0.7 ms instead of the beginning.
Simulation for baseline model 2 was set up for a 100%, 80%, or 60% water-filled tank 2 mm
thick with an initial projectile velocity of 250 m/s. Model 2 was developed to assist in
understanding the structural response of the tank walls during the drag-and-cavitation
phase of HRAM. All displacement, velocity, and effective stress values plotted were
obtained from the center node or element of the tank walls. Because the collapse of the
cavity would most likely occur at a much later time, the cavitation collapse pressure and
its subsequent effect on the tank walls were omitted from this study. Instead, the effects
on tank walls due to drag-phase pressure and the formation of a cavity in the fluid would
be the main interest.
The exit wall response graphs are plotted up to 2 ms. However, the projectile impacted
the exit wall approximately at 1.5 ms. Thus, the results are only of interest before impacting the exit wall. The exit wall started to move and deform at approximately 0.13 ms into the
simulation due to the initial shock wave impinging onto the exit wall. At approximately
1 ms into the simulation, the rate of displacement of the exit wall registered an increase,
as can be observed from the steeper gradient of the displacement time history plot of the
exit wall as illustrated in Figure 10.74. This was due to the projectile approaching the exit
wall and the high-pressure region in front of the projectile during the drag phase exerting
a greater pressure and prestressing the exit wall before projectile impact. Likewise, the
1.00E02
Displacement (m)
8.00E03
Fluid level - 100%

Fluid level - 80%
Fluid level - 60%
6.00E03
4.00E03
2.00E03
0.00E+00
0.00E+00
5.00E04
2.00E03
1.00E03
Time (s)
FIGURE 10.74
Exit wall resultant displacement at the center for model 2.
1.50E03
2.00E03
2.50E03
342
6.00E+06
Fluid level - 100%
Fluid level - 80%
Fluid level - 60%
Pressure (Pa)
5.00E+06
4.00E+06
3.00E+06
2.00E+06
1.00E+06
0.00E+00
0.00E+00
5.00E04
1.00E+06
1.00E03
1.50E03
2.00E03
2.50E03
Time (s)
FIGURE 10.75
Drag-phase fluid pressure output from fluid location 2 for model 2.
effective stress had a peak value after 1 ms when the projectile approached the exit wall.
The exit wall reached a peak velocity of 7 m/s and a peak stress of around 94 MPa prior to
projectile impact.
Figure 10.75 shows the drag-phase fluid pressure recorded in the middle of the fluid.
A peak pressure of around 5 MPa was obtained as the projectile approached the fluid
location at around 0.5 ms. The drag-phase pressure rise was gradual and occurred over
a longer period of time as compared to the initial shock-phase pressure. As the projectile
moved past the fluid center, the pressure record went to zero, indicating the formation of
a cavity behind the projectile path. The cavitation phase of HRAM, which includes the
oscillation and the subsequent collapse of the cavity, was not a part of this study because
it would occur at a later time after the simulation ended. An interesting parameter that the
numerical simulation provided for the drag-phase analysis was the cavity evolution when
the projectile traversed the fluid toward the exit wall. The maximum cavity diameter measured from the fringe plot at 2 ms was approximately 60 mm. The bulging of the entry and
exit walls can also be observed.
11
Multiscale Analysis of Electromechanical System
11.1Introduction
One of the common engineering systems is an electromechanical system based on both
electrical and mechanical principles. An electric car is one example, and a rail gun launcher
is another example. This chapter introduces a multiphysics-based modeling technique for
a rail gun launcher using the finite element method. The multiphysics modeling was conducted for a rail gun launcher to predict the exit velocity of the launch object, temperature
distribution, and thermal contact stress distribution. For this modeling, electromagnetic
field analysis, heat transfer analysis, thermal stress analysis, and dynamic analysis were
conducted for a system consisting of two parallel rails and a moving armature. Especially,
a contact theory was used to estimate the electric as well as thermal conductivities at the
interface.
11.2Principle of Operation of a Rail Gun Launcher

The principle of operation of the simplest type of rail gun launchers is discussed. The rail
gun launcher is a type of projectile weapon [1]. The basic structure of the rail gun launcher
is shown in Figure 11.1. The simple rail gun launcher consists of two parallel conducting
rails and an armature between the rails, which accelerates a projectile between the rails to
a high speed in a short time using electromagnetic force. A large electric current flows to
one of the two parallel conducting rails, travels through the conducting armature between
the rails, and then goes back to the electric current source through the second rail. A projectile to be fired lies on the outer side of the armature and fits loosely between the rails.
The electric current in the rails produces a magnetic field between the parallel rails. The
magnetic fields are directed normal to the plane containing the two rails. The resultant
magnetic field exerts a force on the armature due to the electric current that flows through
it. This is called the Lorentz force [2]. The electromagnetic force points outward along the
rails and pushes the projectile, accelerates it, and launches it at a very high speed. There
are different types of rail guns to enhance the launch power. One is made using a solid
armature, and another one is constructed of a plasma armature. In addition, there are
series augmented rails or parallel augmented rails.
343
344
Parallel rails
(conductors)
Power supply
unit
Armature
(conductor)
Projectile
FIGURE 11.1
Basic structure of a rail gun launcher.
11.3Previous Study of Rail Gun Launchers

Rail gun launchers have been studied previously [37]. Because the electromagnetic field
is the major player of rail gun launchers, most studies examined electromagnetic fields.
Various formulations have been developed to solve the electromagnetic problems of moving conductors [815]. Most of them applied the finite element method; some others utilized the boundary element method [14] and the coupled finite and boundary element
methods [11]. Furthermore, a parallel algorithm was also investigated [15].
Some researchers investigated coupled problems, mostly electromagnetic and thermal
analyses together [3,10] because electric currents generate heat. Heat is an important aspect
inthe rail gun launcher. The contact interface condition between two conductors, such as
an armature and a rail, was studied [16] because those conditions affect the electromagnetic as well as thermal fields significantly.
The following sections present mathematical models for multiphysics analyses, a contact
theory for electric and thermal conductivities at the contact interfaces, a description of
analysis models, and example problems.
11.4Mathematical Models for Multiphysics Analysis

The overall schematic of the multiscale analysis of a rail gun launcher is given in Figure
11.2 and described next [16]. The whole analysis is a time-dependent transient problem.
With the initial location of the armature between two rails, electromagnetic wave analysis is first conducted. From the electromagnetic analysis, the Lorentz force, which is the
driving force of the armature and the projectile, is calculated. Then, Joules law is used
to determine the heat generation, which is utilized for the transient heat transfer analysis. Once the temperature distribution is computed, thermal stress analysis is undertaken
and the contact load is computed between the armature and the rails. Then, the acceleration of the armature is determined by applying Newtons second law to the armature and
345
Mathematical program of rail gun launcher
Is the armature inside

the rail?
No
Stop time
increment
Time increment loop
Yes
Increment of time step
Electromagnetic field analysis to
compute Lorentz force
Heat transfer analysis with electric
heat generation
Thermal stress
analysis to compute contact and
frictional forces
Rigid dynamic analysis to
determine acceleration, velocity,
and new position
FIGURE 11.2
Schematic of multiscale analysis of a rail gun launcher.
the projectile with Lorentz force and friction force between the armature and rails. The
acceleration is integrated twice over time to calculate the new position of the armature.
If the armature is still located inside the parallel rails, the whole analysis repeats itself.
Otherwise, the program is terminated, and the exit velocity is determined.
Electric and thermal conductivities at the interface between the armature and rails contribute significantly to the electromagnetic and temperature fields. As a result, it is necessary to estimate those properties accurately. To determine those interface properties, a
contact theory is also considered. Each analysis is described in more detail subsequently.
11.5Electromagnetic Theory
An electromagnetic system with moving conductors obeys five basic integral laws [10]:

1. Amperes circuital law

2. Faradays induction law
346
3. Law of source free magnetic flux

4. Gausss law
5. Law of charge conservation
Only four of them are independent. They are described in the following:

d
J c n da +
dt
( t )
H d l =
(t )

d
Ed l =
dt
(t )
( t )
( t )
( t )
( t )

B n da (11.2)

B n da = 0 (11.3)

G n da =

d
J n da +
dt
( t )

G n da (11.1)
( t )
( t )
e dv (11.4)
e dv = 0 (11.5)
intensity, B denotes magnetic flux, E denotes electric intensity,

where
H is the magnetic
G is electric flux, J c indicates conduction current density, e is the volume charge density,
t is time, (t) is the moving curve, (t) indicates the surface, and (t) denotes the volume.
Time-dependent configurations in integral Equations 11.1 through 11.5 imply that the conductor undergoes motion.
Governing differential equations of an electromagnetic system with moving conductors
can be deduced from the integral equations, and their forms depend on the chosen description of field variables. The field variables in the Lagrangian description are expressed as
a function of time and reference positions of the particles. Therefore, the integrations can
be performed over a conductor reference configuration that is fixed in space. Thus, convective terms involving velocity components drop out of the equations. Moreover, physical dimensions of electromagnetic systems in the applications are much shorter than the
wavelength of electromagnetic waves, so the displacement current can be neglected. By
virtue of the Gauss divergence theorem and Stokes theorem, quasi-static Maxwells equations in the Lagrangian form can be obtained from the integral equations and have the
following forms:

H = J (11.6)

B
E=
(11.7)
t
347
B = 0 (11.8)
J = 0 (11.9)
Elimination of the convective term greatly simplifies the numerical analysis as far as
storage requirements and numerical stability are concerned. A further consequence of
adopting the Lagrangian description is the position information on conductor boundaries available at all times during the motion. This is especially important in this analysis
where accurate data are needed at all times on the locations of rails and projectile boundaries. The materials of interest are assumed to be isotropic and nonferromagnetic but with
temperature-dependent electrical conductivity. The associated constitutive relations are
expressed as follows:

B = 0 H (11.10)
J = (T )E (11.11)
where T is temperature, 0 (= 4 10 7) is the permeability of free space, and (T) is the

temperature-dependent electrical conductivity.
By expressing magnetic flux as the curl of magnetic vector potential and electrical intensity as the negative sum of the time derivative of magnetic vector potential and the gradient of electric scalar potential, a set of magnetic diffusion equations can be deduced from
quasi-static Maxwells equations with constitutive relations as follows:

B = A (11.12)
A
E=
(11.13)
t
A
1
+
A + = J s (11.14)
t
0
A

= 0 (11.15)
where A is the magnetic vector potential, is the electrical scalar potential, and J s is the
impressed current density.
For nonconductive regions, the diffusion Equations 11.14 and 11.15 can be reduced to
one equation due to vanishing
electric conductivity and impressed current density. The
Coulomb gauge condition A = 0 is imposed to uniquely determine the magnetic vector
potential A.
The electromagnetic and temperature fields are coupled because the electrical conductivity is temperature dependent and the Joule heating is generated due to the electrical
348
resistivity. To obtain accurate magnetic fields, especially in high-current devices, it is necessary to include the thermal effect. From Fouriers law and energy balance, the thermal
diffusion equation from the Lagrangian viewpoint is expressed as

( kT ) + R = Cp
T
(11.16)
t

J J
R=
(11.17)
(T )
where R is the heat source, k is the temperature-dependent thermal conductivity, and Cp is

the temperature-dependent specific heat.
Moreover, changes in the magnetic field are assumed to only weakly depend on changes
in the instantaneous body configuration as a first approximation. Furthermore, the body
is assumed to be rigid so that the effect of deformations of the body is neglected. The
magnetic field is only affected by the rigid body motion of the conductor. The position
and velocity of the conductor are updated throughout the entire analysis. The equations of
motion are described in the following:

M =
( t )

J B dv Ff (11.18)
where M is the conductor mass, is the acceleration, and Ff is the frictional force, discussed further in the chapter.
Three sets of equationsmagnetic diffusion Equations 11.6 through 11.9, thermal diffusion Equations 11.16 and 11.17, and the equation of motion, Equation 11.18derived
previously with constitutive equations form the theoretical basis of the mathematical formulation modeling.
11.6Thermal Stress Analysis

From the electrical heat generation, thermal analysis is conducted using Equations 11.16
and 11.17. The heat transfer in the launcher is also affected by the contact interface condition. Therefore, the same contact theory used for electrical conductivity at the interface is
also utilized for thermal conductivity at the interface.
Using the temperature distribution in the launcher, thermal stress analysis is undertaken with the following set of equations. The stress equilibrium for two dimensions is
expressed as

xx xy
+
= 0 (11.19)
x
y
and

xy
x
yy
y
= 0 (11.20)
349
where ij is the stress tensor. Stresses and strains are related for an isotropic material using
the generalized Hookes law as follows:

xx =
1
( xx yy ) + T (11.21)
Y
yy =
1
( yy xx ) + T (11.22)
Y
xy =
1
xy (11.23)
2G
where ij is the strain tensor, Y is the elastic modulus, G is the shear modulus, is the
Poisson ratio, and is the coefficient of thermal expansion. Finally, the strain-displacement
relationship is written as

xx =
u
(11.24)
x
yy =
v
(11.25)
y
2 xy =
u v
+
(11.26)
y x
where u and v are the displacements in the x and y directions, respectively.

Equations 11.19 through 11.26 are solved using the finite element method with a known
temperature field [17]. From the stress analysis, the contact force between the armature
and the rails is computed. Then, using Coulombs frictional law, the frictional force at the
interface is calculated. Eventually, the net force along the parallel rails is determined by
subtracting the frictional force from the Lorentz force. Using the net force, rigid dynamic
analysis is performed to find the acceleration of the armature by applying Newtons second law by modifying Equation 11.18. In other words, the frictional force is subtracted on
the right side of the equation. The initial time integration of the acceleration yields velocity,
and the next time integration of the velocity results in the new position of the armature.
The Euler integration scheme is utilized for time integration. As long as the armature is
still in contact with the rails, the whole analysis starting from electromagnetic analysis to
dynamic analysis is repeated.
11.7Contact Theory
As the armature and the rails are in contact with each other, electricity and heat pass
through the contact surfaces. Therefore, both electric and thermal conductivities of the
350
contact interface play an important role in the electromagnetic and temperature fields,
which also influences the exit velocity of the projectile. To estimate conductivities (or sensitivities) at the contact interface, a statistics-based contact theory [18] is considered. A
nominally flat surface has a surface roughness (e.g., peaks and valleys) at the microscale.
When two nominally flat surfaces meet each other, those peaks and valleys encounter one
another as actual contacts. Therefore, the nominal contact interface consists of two parts:
the actual contact area and the void area. The void area usually contains air, which has
much less conductivity compared to metallic materials. Therefore, the overall conductivities of the interface depend on the amount of actual contact area.
A nominally flat surface is considered to have a large enough nominal area so that individual contacts of asperity are dispersed and the forces acting on neighboring contact
spots do not affect one another. The actual area of contact between two nominally flat
surfaces is determined by the elastic or plastic deformation of their highest asperities [18].
This leads to the result that the actual area of contact is directly proportional to the contact
load. Furthermore, the contact deformation depends on the topography of the surface.
To further simplify the problem, a nominally flat surface is assumed to have a large
number of asperities, which are spherical (or circular in a two-dimensional [2-D] case), at
least near their peaks. In addition, it is assumed that all the asperity peaks have the same
radius , and that their heights vary randomly. The probability that a particular asperity
has a height between z and z + dz above the given reference plane can be expressed as
(z)dz. If two surfaces become close together until their reference planes are separated by
a distance d, any set of asperities whose sum of their height was originally greater than d
will contact one another. Thus, the probability of having contact at any given asperity, of
height z, is expressed as
prob( z > d) =
(z) dz (11.27)
d
If there are N asperities total in the contact surface, the expected number of contacts
becomes
n = N ( z) dz (11.28)
From the Hertz contact solution, the actual contact area of two bodies with the same
radius is expressed as
A=
w (11.29)
2
The corresponding contact load is

P=
8
Ec 2 w 2 (11.30)
3
351
where w is the compliance (the distance at which points outside the deforming zone move
together during the deformation), which is equal to z d. In addition,
1 1 12 1 22
=
+
(11.31)
Ec
E1
E2
in which E1 and E2 are the elastic moduli of two bodies in contact, and 1 and 2 are their
Poisson ratios, respectively.
Then, the mean contact area is
(z d)(z) dz (11.32)
and the expected total area of contact can be written as
Atot = N ( z d)( z) dz (11.33)

2
The expected total load is

1
8
Ptot = NEc 2 ( z d) 2 ( z) dz (11.34)
3
A single asperity contact has electrical conductivity of 2a/c, where c is the average of
resistivity of two contact bodies. Furthermore, it is assumed that microcontacts are sufficiently separated so that the current flow through them can be also independent. Then,
the electrical contact over the total area can be expressed as

1
c
G = N
1
2
( z d) 2 ( z) dz (11.35)
The thermal conductivities can be computed from a similar expression.

For convenience, normalized variables are introduced. For example, the surface density
of asperities is introduced such that N = S, in which S is the nominal contact area. The
normalized separation distance h = d/s is also used, where s is the standard deviation of
the asperity height distribution. Then, Equations 11.33 through 11.35 can be rewritten as

Atot =
Ptot =
SsF1 ( h) (11.36)
2
1 3
8
SEc 2 s 2 F3 ( h) (11.37)
3
2
352
1 1
G = 2 Sc 1 2 s 2 F1 ( h) (11.38)
where
F =
(s d) *(s)ds (11.39)
Here, *(s) is the normalized height distribution, that is, the height distribution scaled to
make its standard deviation unity [18]. For the present study, the Gaussian distribution of
asperities is assumed.
11.8Analysis Model
A 2-D finite element analysis model of a rail gun launcher is shown in Figure 11.3. The
figure shows two rails and an armature between them, and the armatures shape was simplified as a rectangular shape. The rails were 0.5 m long and 9.5 mm wide; the armature
had a height of 19 mm and a width of 10 mm. The armature was assumed located initially
at 0.05 m from the left of the rails with zero initial velocity. An electric potential 6.5 kV was
applied to the rails at the initial room temperature of 300 K. As materials, aluminum, copper, and steel were selected for the present study. Their material properties are available in
Reference 19, so that they are omitted here.
0.05
0.04
0.03
0.02
Rails
Armature
0.01
0
0.01
0.1
0.1
0.2
0.3
0.4
0.5
0.6
FIGURE 11.3
Finite element mesh of the armature and rails (a crude mesh is shown for visual clarity even though more
refined meshes were used for actual computations).
353
TABLE 11.1
Statistical Values for Contact Surface Roughness
Surface
Parameters
Surface Condition A
Surface Condition B
Surface Condition C
1.37 m
13 m
0.065 m
0.24 mm
0.01 m
0.5 mm
Standard deviation
Mean radius of peaks
As far as the contact surface between the armature and rails is concerned, three different
surface conditions were considered [18]. Their statistical values for the Gaussian distribution are provided in Table 11.1. Furthermore, the rails are assumed to be constrained at
both outer edges from any movement and are insulated.
11.9Example Problems
The first model of the rail gun launcher was constructed using an aluminum armature
and copper rails. The contact surface condition A in Table 11.1 was considered. Because
the electric and thermal conductivities at the contact surface played an important role,
those properties were examined first. Considering different normalized separation distances, we calculated electric conductivity, thermal conductivity, and the contact force at
the armature/rail interface, which are listed in Table 11.2. Once we knew the normalized
separation distance or the contact force, we could estimate both electric and thermal conductivities at the interface.
After the interface properties were determined, the multiphysics-based computations as
outlined in Figure 11.2 were executed to determine the armature exit velocity, temperature
distribution, and thermal stresses in the rail gun launcher.
Table 11.3 shows the results of the rail gun constructed of an aluminum armature and
copper rails. The results show that as the distance between the reference planes became
smaller (i.e., larger contact load), the electric and thermal conductivities become larger,
which eventually resulted in a higher exit velocity with a larger Lorentz force and higher
temperature.
Figures 11.4 and 11.5 plot the distance between the reference planes of the contact surfaces as a function of the contact load and the electric conductivity, respectively. The figures show that as the contact load increased, the relative distance between the reference
planes decreased, as expected. Then, such a decrease in the relative distance increased
electric conductivity. Both plots show nonlinear relationships.
TABLE 11.2
Computed Values of Electric Conductivity, Thermal Conductivity, and Contact Load for
Different Normalized Separation Distances
Normalized Separation Distance
0.985
1.000
1.015
1.029
1.044
Electric conductivity (106 mho/m)

Thermal conductivity (W/m-degrees kelvin)
Contact load (N)
1.139
1.879
824
1.111
1.832
799
1.083
1.787
775
1.056
1.742
752
1.029
1.698
730
354
TABLE 11.3
Computer Simulation Results for Copper Rails and Aluminum Armature for Different
Distances between Reference Planes of Contact Surfaces
Distance between Reference Planes (m)
1.35
1.37
1.39
1.41
1.43
Exit velocity (m/s)

Total launch time (ms)
Lorentz force (kN)
Average temperature of contact elements
(degrees K)
4446
0.203
275
784
4332
0.212
262
769
4237
0.217
250
754
4132
0.222
238
745
4040
0.228
226
729
Distance between reference planes d (m)
2.4
106
Contact load P
2.2
2
1.8
1.6
1.4
1.2
1
0.8
200
400
600
800
Load P (N)
1200
1000
1400
FIGURE 11.4
Plot of distance between the reference planes of two contact surfaces versus contact load.
Distance between reference planes d (m)
2.4
106
Conductance G
2.2
2
1.8
1.6
1.4
1.2
1
0.8
0.4
0.6
0.8
1
1.2
1.4
Conductance G [(Ohmm)1]
1.6
1.8
106
FIGURE 11.5
Plot of electric conductivity versus distance between the reference planes of two contact surfaces.
355
Figure 11.6 compares the projectile velocities of three different armature materials (i.e.,
aluminum, mild steel, and polished steel) inside copper rails with the contact surface condition A in Table 11.1. Comparing the results, the aluminum armature yielded the highest
projectile velocity. Even if the analysis program already included frictional effect, the friction coefficient was assumed to be zero for the present study.
The next case switched the rail and armature materials. In other words, copper was used
for the armature and aluminum was selected for rails. Then, this case was compared to
the previous case before exchanging the materials, as shown in Figure 11.7. The aluminum
armature/copper rail case resulted in a higher velocity than for the opposite combination. From comparing the values of Lorentz forces that were exerted on the armature, no
significant change was observed between the two cases. However, the mass of the aluminum armature was smaller than that of copper because the volume of the armature was
kept constant in the analysis. The lighter mass of aluminum resulted in a higher velocity
according to Newtons second law.
However, heat generation and heat transfer were quite different between the two cases,
which gave very different average temperatures at the contact interfaces, as shown in
Figure 11.8. The average temperature at the interfaces was lower in the case of copper
armature/aluminum rail than that in the case of the opposite material combination. On
the other hand, the former case had a higher rate of temperature rise at the interface as
the armature moved along the rail compared to the latter case. To further investigate this
trend of temperature, the rail was extended to 1.0 m or 1.5 m long for the next study. The
results are discussed next.
First, the velocity was higher for the longer rail, as tabulated in Table 11.4. Figure 11.9
compares the velocity of the two cases with the rail length of 1.5 m. The increase of the
velocity followed a parabolic orbit. Hence, there must be an upper limit at the length of the
rail (also called a barrel) beyond which any further increase does not significantly affect
the value of the exit velocities and makes the design impractical.
Rail-copper
5000
Aluminum
Mild steel
Polished steel
4500
4000
Velocity (m/s)
3500
3000
2500
2000
1500
1000
500
0
0.05
0.1
0.15
0.2
0.25 0.3 0.35 0.4

Length of barrel (m)
0.45
0.5
0.55
FIGURE 11.6
Velocity as a function of the location in the copper rails for three different types of armature: aluminum, dotted
curve; mild steel, dashed curve; and polished steel, rigid curve.
356
Velocities
6000
Rail-Cu/proj-Al
Rail-Al/proj-Cu
5000
Velocity (m/s)
4500
3000
2000
1000
0
0.05
0.15
0.1
0.2
0.45
0.3
0.25
0.35
0.4
Length of the rail (m)
0.5
0.55
FIGURE 11.7
Velocity as a function of the distance along the rail for two different cases of materials.
Average temperatures
800
700
Temperature (K)
600
500
400
Rail-Cu/proj-Al
Rail-Al/proj-Cu
300
200
100
0
0.05
0.1
0.15
0.2
0.25
0.35
0.3
0.4
Length of the barrel (m)
0.45
0.5
0.55
FIGURE 11.8
Comparison of average temperatures at the interface between the rails and the armature for two different material combinations.
TABLE 11.4
Exit Velocities for Three Different Rail Lengths
Rail Length
Copper Rail/Aluminum
Armature
Aluminum Rail/Copper
Armature
4100 m/s
6200 m/s
7900 m/s
2500 m/s
3500 m/s
4500 m/s
0.5 m
1.0 m
1.5 m
357
Velocities
8000
Rail-Cu/proj-Al
Rail-Al/proj-Cu
7000
Velocity (m/s)
6000
5000
4000
3000
2000
1000
0
0.5
1
Length of the barrel (m)
1.5
FIGURE 11.9
Velocity as a function of the displacement in the 1.5-m rail.
Calculation of the kinetic energy from the exit velocity is given in Table 11.5. The kinetic
energy of the aluminum armature was lower than that of the copper one because of its
lighter mass. The total mass of the launch object was the mass of the armature and the
mass of the projectile. That means that the projectile mass can be determined based on
either the desired exit velocity or the kinetic energy. Some combination of the two criteria
may be also used for designing the weight of the projectile.
Furthermore, of greater interest is what happened at the distribution of the temperature
at the interfaces. The two temperature curves intersect as the rail length became about 0.8 m
beyond which the aluminum rail with a copper armature had a higher temperature than
the copper rail with an aluminum armature. The maximum values of average temperatures for the two cases are shown in Table 11.6. Two questions now arise. Why was the
initial average temperature of the aluminum armature case higher than that of the copper
armature case? In addition, why was the rate of increase of the former lower than that of
the latter?
Taking into consideration the specific design selected previously, the electric resistance
of the case of the aluminum armature/copper rail was lower than that of the opposite case.
This resulted in greater heat generation for the former case than the latter case. Therefore,
the temperature at the interface was higher for the aluminum armature/copper rail case
in the beginning.
TABLE 11.5
Kinetic Energy for Three Different Rail Lengths
Rail Length
Armature
Armature
82 kJ
188 kJ
304 kJ
100 kJ
198 kJ
327 kJ
0.5 m
1.0 m
1.5 m
358
TABLE 11.6
Maximum Average Temperatures (degrees kelvin) at the Interfaces
Rail Length
Armature
Armature
760
860
1010
660
920
1430
0.5 m
1m
1.5 m
As the armature moved down the rails with time, the heat generated from the electric
current was conducted along the rails and dissipated. Copper has higher thermal conductivity than aluminum. This caused faster heat removal from the rail gun launcher for the
copper rails. For this reason, the rate of increase of the temperature profile at the interfaces
was lower for the aluminum armature/copper rail case.
12
Multiphysics Analysis of Biomechanics
12.1Introduction
One of the examples for fluid-structure interaction (FSI) is blood flow in the heart and
blood vessels of the human body. Blood is a viscous fluid, and the heart and blood vessels
are generally viscoelastic solids. Pulsating motions in the heart and blood circulate blood
in the body. This chapter discusses the interaction between blood and a blood vessel. In
particular, the focus of the study is the aneurysm.
An aneurysm is a focal dilation of a blood vessel, which may rupture, leading to death
of the patient. There are two major aneurysms in the human body. One is a cerebral artery
aneurysm (CAA), and the other is the abdominal aortic aneurysm (AAA). The AAA occurs
in the infrarenal aorta and has a diameter greater than 3 cm; it can grow to 9 cm in length
[1,2]. Numerical modeling and simulation was conducted to investigate what effects influence an aneurysm.
12.2Review of Previous Work

Most of the previous studies on aneurysms considered either already existing realistic
aneurysms or idealized aneurysms with focuses on understanding and predicting ruptures of blood vessels. To predict ruptures, stresses in blood vessels must be determined.
As a result, blood vessel stresses were calculated as a function of the vessel diameter [3,4],
wall thickness [5], asymmetry [6], tortuosity [7], material property [8], calcification [9], and
blood flow [10]. Blood vessel strength was measured by ex vivo studies [11,12].
Researchers considered different constitutive models for investigating inception and
growth of aneurysms. Vena et al. presented an anisotropic model for early stages of the
aneurysm to study growth and remodeling [13]. Schmid and his colleagues investigated
the effects of differences in elastic properties, fiber orientations, and metabolic activities on
aneurysm formation and rupture in a layer-specific structural artery model [14].
A few computational studies considered the early stages of cerebral aneurysms. Inception
and growth of aneurysms in an idealized carotid artery were modeled in Reference 15.
They examined the initiation of the aneurysm by decreasing the modulus of elasticity
at a local region, where smooth muscle cell relaxation was assumed. A similar assumption was used in Reference 16 so that high wall shear stresses could cause wall weakening. The modulus of elasticity was decreased at the regions of high wall shear stresses in
curved and straight idealized intracranial arterial geometries. Another study [17] used a
359
360
hyperelastic (HE) material model with a large strain formulation at the bifurcation where
most saccular cerebral aneurysms exist; other regions of the vessel were modeled as linear
elastic (LE) material. They also assumed that smooth muscle cell relaxation could cause
aneurysm formation.
The thickness of the aorta consists of three layers: intima, media, and adventitia [18].
However, most studies assumed a single layer with a uniform material property for simplification even though in vivo studies showed that each layer had a different contribution
to the material property of the blood vessel [1921].
This chapter presents the results of a numerical study to determine the effects of material properties of a three-dimensional (3-D), idealized, three-layer abdominal aorta on
aneurysm initiation and fully developed aneurysm [22]. To this end, the numerical model
considered three individual layers in the blood vessel. Both LE and HE material properties
were considered for healthy and locally degenerated sections of the blood vessel. The FSI
between the blood and vessel was also considered in the study.
12.3Description of Numerical Models

Idealized 3-D models were generated for a three-layer abdominal aorta for aneurysm initiation as well as for an already-developed aneurysm [22]. The first model was for aneurysm
initiation; it had three concentric cylinders that were secured to one another. The actual
length of the blood vessel under study was 12 cm. However, the numerical model was 24cm
long so that the effect of the end boundary conditions could be neglected for the flow characterization. The in vivo wall thickness of the infrarenal aorta is between 0.14 and 0.15 cm
[21]. The ratio for the intima, media, and adventitia layers is 20:47:33. As a result, this study
considered a constant wall thickness of 0.15 cm with thicknesses of 0.03, 0.075, and 0.045 cm,
respectively, for the three layers. The lumen diameter was assumed to be 2 cm.
The vessel wall was modeled using two different materials. One material model was
nearly incompressible, isotropic, and LE. The elastic modulus was 1.2 MPa, and the
Poisson ratio was 0.49 [23]. The other material model used a HE wall and the coefficients
in Reference 24. It was a model of the Mooney-Rivlin type. The strain energy density function is given as

W = C10 ( I i 3) + C20 ( I1 3)2 +
( J 1)2
(12.1)
d
where W is the strain energy density of the material, C10 and C20 are the material constants,
I1 is the first deviatoric strain invariant, J is the ratio of the deformed elastic volume over
the undeformed volume of materials, and the parameter d is the material incompressibility
parameter. Some of the materials were related as follows:

K=
=
(12.2)
d 2(1 2 )
in which is the initial shear modulus, is the Poisson ratio, and K is the bulk modulus.
The material properties for each layer used in the study are listed in Table 12.1.
361
TABLE 12.1
Material Properties for Each Layer of a Healthy Blood Vessel
Intima
Media
Adventitia
Elastic Model
Hyperelastic Model
Thickness
Ratio
E (MPa)
C10 (MPa)
C20 (MPa)
1
3
2
0.522
1.566
1.044
0.490
0.490
0.490
0.034
0.101
0.067
0.363
1.090
0.727
d
1.190
0.397
0.595
The aneurysmal blood vessel section was assumed to have reduced material properties compared to a healthy blood vessel. The same reduction ratio was considered for
both LE and HE materials. The wall density was 1120 kg/m3. The blood was assumed
to have the characteristics of a Newtonian, laminar, and incompressible flow. The density of blood was 1050 kg/m3, and its dynamic viscosity was 0.0035 Pa s [25]. Youngs
modulus ratio for intima/media/adventitia was assumed to be a ratio of 1:3:2 as shown
in Table 12.1. The degenerated material properties were applied on the media layer by
decreasing its modulus of elasticity or the coefficients of the strain energy function
by 1/20 [26]. Four different sizes of degenerated regions were considered. The model
called degeneration in region A had local degeneration consisting of two medial arcs
(100 elements) located near the center of the longitudinal axis. The arcs were half circles
(i.e., 180). The other three models had degenerated circular rings. The second model,
degeneration in region B, had one circular medial ring (400 elements); the third model,
degeneration in region C, had two circular medial rings (800 elements); and the final
model, called degeneration in region D, had three consecutive medial rings (1200 elements). The four cases are called DR case A, DR case B, DR case C, and DR case D,
respectively, from this point.
As boundary conditions for blood flow, the inlet had time-dependent, fully developed
laminar flow velocity; the outlet had time-dependent pressure [27]. Peak systolic pressure
occurred at t = 0.53 s (15,594.5 kg/ms2), and peak systolic flow was obtained at t = 0.45 s
(0.437886 m/s). Figure 12.1 shows the pressure time history of two cycles. Because the initial condition of the blood flow was not known, it was assumed to be zero. To provide the
proper initial condition, the numerical computation was conducted for two cycles such
20
Pressure (kPa)
15
10
1st cycle
5
0
0.5
2nd cycle
1.5
Time (s)
FIGURE 12.1
Pressures applied to the blood.
2.5
362
that the solution after the first cycle could be naturally the initial condition for the second
cycle. Then, the solution during the second cycle was used for comparison. In the plot of
the result, the starting point of the second cycle was set to zero. For the solid domain, both
the inlet and outlet of the domain were fixed for all degrees of freedom. To minimize the
effect of the constrained blood vessel on the solution, the blood vessel was modeled much
longer than the actual length of the vessel under examination. At the interface between
fluid and solid, a no-slip boundary condition was applied.
12.4Comparison of Models with and without FSI

First, the result of the FSI model was compared to that of the model without FSI, which
did not consider FSI between the blood vessel and the blood. Then, the results of FSI were
further discussed [22]. Five different locations through the blood vessel thickness were
selected. Location 1 was at the interface of the intima and blood, location 2 was at the
interface of the media and intima, location 3 was at the middle of the media, location 4
was at the interface of the adventitia and media, and location 5 was at the upper side of
the adventitia. Figure 12.2 compares the applied pressure on the lumen sac for the analysis
without FSI (called noFSI) to the calculated pressure at the blood-intima interface for the
FSI analysis. The calculated pressure waveform for the FSI analysis (dashed line) was at
the middle of the blood vessel. The two pressure waveforms were similar to each other
except for a phase shift between them.
Figure 12.3 compares the von Mises stresses at the interfaces of the layers: the intimamedia interface and media-adventitia interface. The analysis results with FSI and without
FSI were significantly different from each other. The noFSI model overestimated the peak
von Mises stress by 52% at the intima-media interface and by 22% at the media-adventitia
interface. In addition, the FSI result showed a gradual continuous stress variation through
the thickness of the blood vessel, as shown in Figure 12.4; the noFSI result showed a sharp
change at the interface of each layer through the vessel thickness, as shown in Figure 12.5.
20
Pressure (kPa)
15
10
Lumen Sac-noFSI
Interface FSI-2nd cycle

0
0.5
Time (s)
FIGURE 12.2
Comparison of pressures between noFSI and FSI cases.
1.5
363
200
Intima/media noFSI
Intima/media FSI
Media/adventitia noFSI
von Mises stress (kPa)
150
Media/adventitia FSI
100
50
0.2
0.4
0.6
Time (s)
0.8
1.2
FIGURE 12.3
Comparison of von Mises stresses between noFSI and FSI results.
400
1st mesh
350
2nd mesh
300
3rd mesh
250
200
150
100
50
0
0.0004
0.0008
0.0012
0.0016
Radial distance (m)

FIGURE 12.4
von Mises stress distribution through the thickness of the blood vessel with FSI.
The figures also show the mesh sensitivity studies. All the meshes resulted in the same
solutions, so that the mesh size effect could be neglected.
In single-layer models, the von Mises stresses between FSI and noFSI models were
small. However, in the three-layer model, the two results were quite different in the von
Mises stresses at the interface of the layers [22]. Because blood flow does exist in human
arteries, FSI analysis is more realistic than noFSI analysis. Therefore, the blood flow
should be taken into consideration in studying mechanical properties of the arteries.
364
200
Mesh 1
Mesh 2
150
Mesh 3
100
50
0
0.0004
0.0008
Radial distance (m)
0.0012
0.0016
FIGURE 12.5
von Mises stress distribution through the thickness of the blood vessel without FSI.
12.5Results of the Aneurysm Initiation Studies

Four different models of aneurysm initiation were compared to the healthy abdominal
blood vessel without degeneration in terms of stresses and strains at the initiation site of
the three-layer abdominal aorta. The aneurysm initiation models may have had a combination of LE and HE materials in the healthy and degenerated sections of the blood vessel.
The combinations were the LE-vessel/LE-degeneration model, which used the LE material
for the blood vessel except for the degenerated area and the reduced LE material used for
the degenerated area; the LE-vessel/HE-degeneration model, which used the LE material
for the vessel except for the degenerated area and the reduced HE (i.e., Mooney-Rivlin)
material for the degenerated area; the HE-vessel/LE-degeneration model, which utilized
the HE material for the vessel except for the degenerated area and the reduced LE material for the degenerated area; and the HE-vessel/HE-degeneration model, which had the
HE material used for the vessel except for the degenerated area and the reduced HE (i.e.,
Mooney-Rivlin) material used for the degenerated area.
12.5.1Blood Vessel Wall Modeling
Hoop strains are compared among different material models. Figure 12.6 shows the
results of the blood vessel with the LE material for the healthy part and the reduced
LE material for the degenerated part acting as initiation of aneurysm. As expected, the
blood vessels with local degeneration had greater hoop strains compared to the healthy
blood vessel. A larger degeneration area resulted in a larger hoop strain. The strain variation along the vessel thickness was linear, but the slope of the line increased as the
degeneration zone became larger. A similar observation was made for the LE healthy
vessel with the reduced HE degenerated zones, as seen in Figure 12.7. When the healthy
vessel was modeled as the HE material while the degenerated zone was modeled as the
elastic material, the effect of the degeneration size on the increase in the hoop strain
was smaller compared to the previous cases. This can be shown when Figure 12.8 is
compared to Figures 12.6 and 12.7. The increase in the hoop strain in the LE vessel was
about 40% from one circular degenerated ring (DR case B) to two rings (DR case C) and
about 20% from two circular degenerated rings (DR case C) to three degenerated rings
365
Intact vessel
LE DR case B
LE DR case C
LE DR case D
0.4
0.35
Hoop strain
0.3
0.25
0.2
0.15
0.1
0.05
0
0.0004
0.0008
Radial distance (m)
0.0012
0.0016
FIGURE 12.6
Comparison of hoop strains for LE-vessel/LE-degeneration models with different degeneration sizes. DR case B:
one medial ring; DR case C: two medial rings; DR case D: three medial rings.
0.4
Intact vessel
MR DR case B
0.35
MR DR case C
Hoop strain
0.3
MR DR case D
0.25
0.2
0.15
0.1
0.05
0
0.0004
0.0008
0.0012
0.0016
Radial distance (m)

FIGURE 12.7
Comparison of hoop strains for LE-vessel/HE-degeneration models with different degeneration sizes. DR case B:
(DR case D). The HE-vessel/LE-degeneration model showed about a 15% increase in the
hoop strain from DR case B to DR case C and less than a 5% increase between DR case C
and DR case D. The hoop strains varied similarly at least qualitatively through the vessel thickness of different healthy and degenerated models. As a result, it was difficult to
detect any possible initiation of aneurysm in terms of the hoop strain behavior.
The von Mises stresses were compared among the healthy and degenerated models.
Figure 12.9 plots the von Mises stresses for the blood vessels whose healthy parts were
modeled using the LE material and the degenerated sections were modeled using the
reduced LE material. The figure shows an interesting result. Aneurysm initiation modeled as a reduced LE modulus decreased the von Mises stresses in the media layer and
increased them in the intima and adventitia layers. In other words, the media layer had a
366
0.4
Intact vessel
0.35
LE DR case B
LE DR case C
Hoop strain
0.3
LE DR case D
0.25
0.2
0.15
0.1
0.05
0
0.0004
0.0008
0.0012
0.0016
Radial distance (m)

FIGURE 12.8
Comparison of hoop strains for HE-vessel/LE-degeneration models with different degeneration sizes. DR case B:
400
Intact vessel
350
LE DR case B
LE DR case C
LE DR case D
300
250
200
150
100
50
0
0.0004
0.0008
0.0012
0.0016
Radial distance (m)

FIGURE 12.9
Comparison of von Mises stresses for LE-vessel/LE-degeneration models with different degeneration areas. DR
case B: one medial ring; DR case C: two medial rings; DR case D: three medial rings.
higher von Mises stress than the intima and adventitia layers in the healthy aorta, but the
initiation of an aneurysm, modeled as a loss in the medial layer, revised that behavior. The
medial stress decreased, and the intima and adventitia carried more loads. The collagenrich adventitia had the highest stress values. This was consistent with the previous finding in Reference 28, which stated that fibroblasts and smooth muscle cells increased the
synthesis of collagen due to the mechanical loading, which played a role in the growth and
remodeling of the AAA.
All different material models showed similar von Mises stress distributions resulting
from aneurysm initiation. One minor observation is that the von Mises stress at the medial
367
region increased as the area of the degeneration region increased for the models for the
LE-vessel/HE-degeneration model, as seen in Figure 12.10. Hoop and longitudinal stresses
were similar to von Mises stresses with a little change in magnitude.
The longitudinal strain increased along the blood vessel thickness when an aneurysm
initiated in both LE-vessel and HE-vessel models. However, the responses of the two
material properties were different from each other. Figure 12.11 compares the longitudinal
400
350
Intact vessel
300
MR DR case B
MR DR case C
250
MR DR case D
200
150
100
50
0
0.0004
0.0008
0.0012
0.0016
Radial distance (m)

FIGURE 12.10
Comparison of von Mises stresses for LE-vessel/HE-degeneration models with different degeneration sizes. DR
0.1
Intact vessel
LE DR case A
0.08
Longitudinal strain
LE DR case B
LE DR case C
0.06
LE DR case D
0.04
0.02
0
0.02
0.0004
0.0008
0.0012
0.0016
Radial distance (m)

FIGURE 12.11
Comparison of longitudinal strain for LE-vessel/LE-degeneration models with different degeneration sizes.
DR case A: one medial arc; DR case B: one medial ring; DR case C: two medial rings; DR case D: three medial
rings.
368
0.06
Intact vessel
Longitudinal strain
0.05
LE DR case B
LE DR case C
0.04
LE DR case D
0.03
0.02
0.01
0
0.01
0.0004
0.0008
0.0012
0.0016
Radial distance (m)

FIGURE 12.12
Comparison of longitudinal strain for HE-vessel/LE-degeneration models with different degeneration sizes.
DR case B: one medial ring; DR case C: two medial rings; DR case D: three medial rings.
strains for the healthy vessel made of the LE material and the degenerated parts made
of the reduced LE material. Likewise, Figure 12.12 plots the longitudinal strains for the
healthy vessel made of the HE material and the degenerated parts made of the reduced LE
material. Figure 12.13 is for the blood vessel models made of HE-vessel/LE-degeneration
material.
0.1
Intact vessel
MR DR case A
Longitudinal strain
0.08
MR DR case B
MR DR case C
0.06
MR DR case D
0.04
0.02
0
0.02
0.0004
0.0008
0.0012
0.0016
Radial distance (m)

FIGURE 12.13
Comparison of longitudinal strain for LE-vessel/HE-degeneration models with different degeneration sizes. DR
case A: one medial arc; DR case B: one medial ring; DR case C: two medial rings; DR case D: three medial rings.
369
Regardless of what materials were used for the healthy and degenerated regions of the
blood vessels, DR case C and DR case D showed similar longitudinal strain distributions,
but they were different from DR case A and DR case B, both of which also had similar
distributions. The medial longitudinal strain in the DR case B was much smaller than
those of DR case C and DR case D for the HE-vessel/LE-degeneration model, whereas
the magnitude of the medial strain in DR case B was close to the magnitudes in DR
case C and DR case D in the LE-vessel/LE-degeneration model. On the other hand, for
HE-degeneration models, the material property of the blood vessel also influenced the
longitudinal strain distribution through the vessel thickness but not as much as in the
case of LE-degeneration models. Furthermore, the distributions of longitudinal strains
of DR case A and DR case B were not the same, unlike those of the hoop strain and von
Mises stress.
Radial stresses in the two models of blood vessels like LE-vessel and HE-vessel were
also different from each other. As shown in Figure 12.14, the LE-vessel/LE-degeneration
model showed sharp changes in the radial stress at the interfaces of the layers for DR
case B. In DR case C, the change in radial stress at the interfaces was larger at the interface between media and adventitia than at the interface between intima and media. For
the largest area of degeneration, like the DR case D model, the radial stress had smaller
changes at the layer interfaces than those of less-degenerated areas. In the HE-vessel/
LE-degeneration model, as seen in Figure 12.15, the DR case B showed oscillations of the
radial stress value through the vessel thickness. At the HE-vessel/HE-degeneration model
with DR case B and DR case C, the radial stresses did not change at the interfaces of the
layers, unlike the LE-vessel/HE-degeneration model. Figures 12.16 and 12.17 support the
statement. The numerical results also indicated that the radial stresses were the largest in the adventitia for all models considered, and the adventitial radial stresses in the
40
Intact vessel
30
LE DR case B
Radial stress (kPa)
20
LE DR case C
LE DR case D
10
0
10
20
30
40
0.0004
0.0008
0.0012
0.0016
Radial distance (m)

FIGURE 12.14
Comparison of radial stresses for LE-vessel/LE-degeneration models with different degeneration sizes. DR case
B: one medial ring; DR case C: two medial rings; DR case D: three medial rings.
370
Radial stress (kPa)
Intact vessel
10
LE DR case B
LE DR case C
15
LE DR case D
20
0.0004
0.0008
0.0012
0.0016
Radial distance (m)

FIGURE 12.15
Comparison of radial stresses for HE-vessel/LE-degeneration models with different degeneration sizes. DR
40
Intact vessel
Radial stress (kPa)
30
MR DR case A
MR DR case B
20
MR DR case C
10
MR DR case D
10
20
30
40
0.0002
0.0004
0.0006
0.0008
0.001
0.0012
0.0014
0.0016
Radial distance (m)

FIGURE 12.16
Comparison of radial stresses for LE-vessel/HE-degeneration models with different degeneration sizes. DR
HE-vessel/LE-degeneration and HE-vessel/HE-degeneration models were significantly

larger than those in the intima and media layers.
Figures 12.18 through 12.21 show all the plots for the radial strains. The radial strain distribution resulting from aneurysm initiation was different for the HE-vessel and LE-vessel
models. The radial strains were almost constant in the LE-vessel/LE-degeneration
371
Radial stress (kPa)
0
5
10
Intact vessel
MR DR case B
15
20
MR DR case C
0
0.0004
0.0008
0.0012
0.0016
Radial distance (m)

FIGURE 12.17
Comparison of radial stresses for HE-vessel/HE-degeneration models with different degeneration sizes. DR
case B: one medial ring; DR case C: two medial rings.
0
0.05
0.1
Radial strain
0.15
0.2
0.25
0.3
Intact vessel
LE DR case A
LE DR case B
LE DR case C
LE DR case D
0.35
0.4
0.45
0.5
0.0004
0.0008
0.0012
0.0016
Radial distance (m)

FIGURE 12.18
Comparison of radial strains for LE-vessel/LE-degeneration models with different degeneration sizes. DR
case A: one medial arc; DR case B: one medial ring; DR case C: two medial rings; DR case D: three medial
rings.
model, while the magnitude depended on the degeneration size. On the other hand, the
HE-vessel/LE-degeneration model showed the reduced radial strain in the medial layer.
The LE-vessel/HE-degeneration and HE-vessel/HE-degeneration models also showed
different distributions of the radial strains. The LE-vessel/HE-degeneration model had
the increased radial strain in the media layer for DR case C and DR case D, while the
HE-vessel/HE-degeneration model had reduced radial strains for DR case B and DR
caseC.
372
0
0.05
Radial strain
0.1
0.15
0.2
0.25
Intact vessel
LE DR case B
LE DR case C
LE DR case D
0.3
0.35
0.4
0.0004
0.0008
0.0012
0.0016
Radial distance (m)

FIGURE 12.19
Comparison of radial strains for HE-vessel/LE-degeneration models with different degeneration sizes. DR case
B: one medial ring; DR case C: two medial rings; DR case D: three medial rings.
0
0.05
0.1
Radial strain
0.15
0.2
0.25
Intact vessel
0.3
MR DR case A
0.35
MR DR case B
0.4
MR DR case C
0.45
0.5
MR DR case D
0
0.0004
0.0008
0.0012
0.0016
Radial distance (m)

FIGURE 12.20
Comparison of radial strains for LE-vessel/HE-degeneration models with different degeneration sizes. DR case
A: one medial arc; DR case B: one medial ring; DR case C: two medial rings; DR case D: three medial rings.
The human abdominal aorta has nonlinear material characteristics, and it is believed
that the Mooney-Rivlin-type material is more realistic than the LE-type material. When
the initiation of an aneurysm was modeled as a reduction in the material properties, four
different material models resulted in significantly different distributions of stresses and
strains through the blood vessel. Therefore, it was considered proper that HE material
modeling of the blood vessel should be selected in modeling aneurysm initiation.
373
0
0.05
0.1
Radial strain
0.15
0.2
Intact vessel
0.25
MR DR case B
0.3
MR DR case C
0.35
0.4
0.0004
0.0008
0.0012
0.0016
Radial distance (m)

FIGURE 12.21
Comparison of radial strains for HE-vessel/HE-degeneration models with different degeneration sizes. DR
caseB: one medial ring; DR case C: two medial rings.
12.5.2Blood Viscosity Effect

It has been proposed that blood viscosity increases with age and diabetes. A study in
Reference 29 stated that the viscosity of blood increased with age in approximately 7%. The
effect of blood viscosity on the initiation of aneurysm was studied using the LE-vessel/
LE-degeneration model with the DR case C model. The dynamic viscosity of blood was
assumed to be 0.0038 Pa s. The numerical data suggested that the viscosity increase by
7% did not yield any noticeable change in the stress, strain, and deformation distribution
through the blood vessel. A previous study showed the effect of the viscosity by changing
the kinematic viscosity from 0.0027 Pa s, and 0.0097 Pa s made a small change in the blood
vessel diameter from 52.407 to 52.408 mm and increased the peak wall shear stress of the
aneurysmal wall [30]. However, the change in the blood viscosity was much greater in the
previous study, while it was small in the present study. Therefore, an increase in blood viscosity by 7% is not believed to influence aneurysm formation at the initial stages in terms
of the stress and strain perspectives.
12.6Already-Developed Abdominal Aortic Aneurysm

This section models the already-developed AAA. This numerical model was almost the
same as the previous initiation model except that a locally bulged section was included at
the center of the blood vessel. The maximum diameter of the bulged section was assumed
to be 6 cm. The material modeling of the healthy abdominal aorta without any degeneration or aneurysm was compared to the already-developed aneurysm model in terms of
stress-strain characteristics. The von Mises stresses were compared among different intact
374
and aneurysmal vessels. The von Mises stress was the highest in the media layer and the
lowest in the intima layer for all the models, as shown in Figure 12.22. These results were
consistent with the results from a previous study [20]. Although the magnitudes of stress
were not the same due to differences in types of vessels, the von Mises stress distribution
through layers was consistent with the present study. The range of stress from intima to
adventitia was from 40 to 100 kPa without aneurysm, while the present study showed the
stress ranged from 50 to 160 kPa.
The material property of the vessel did not change the qualitative distribution of the von
Mises stress in the intact or aneurysmal vessels as well as the magnitude of increase in
von Mises stresses from intact to aneurysmal vessels. However, Figure 12.23 shows that
an increase in hoop strain from intact to aneurysmal vessels was larger for the HE model
than the LE model. Likewise, the HE vessels yielded larger radial displacements than for
the LE vessels. The strain value was smaller for the aneurysmal vessel than the healthy
vessel because the aneurysmal vessel was stiffer than the healthy aorta.
400
LE-intact V.
350
LE-aneurysmal V.
300
MR-intact V.
250
MR-aneurysmal V.
200
150
100
50
0
0.0004
0.0008
0.0012
0.0016
Radial distance (m)

FIGURE 12.22
Comparison of von Mises stresses through the thickness of the intact and aneurysmal vessels using different
material models.
LE-intact V.
0.2
LE-aneurysmal V.
MR-intact V.
Hoop strain
0.15
MR-aneurysmal V.
0.1
0.05
0
0.0004
0.0008
0.0012
0.0016
Radial distance (m)

FIGURE 12.23
Comparison of hoop strains through the thickness of the intact and aneurysmal vessel.
References
Chapter 1
1. Y. W. Kwon and H.-C. Bang. The finite element method using MATLAB. 2nd ed. Boca Raton, FL:
CRC Press, 2000.
2. J. E. Akin. Finite element analysis for undergraduates. London: Academic Press, 1986.
3. O. C. Zienkiewicz and R. L. Taylor. The finite element method. 4th ed. London: McGraw-Hill,
1991.
4. D. A. Rapport. The art of molecular dynamics simulation. 2nd ed. Cambridge: Cambridge
University Press, 2004.
5. J. M. Haile. Molecular dynamics simulation: Elementary methods. New York: Wiley, 1997.
6. S. Wolfram. Cellular automata and complexity. Reading, MA: Addison-Wesley, 1994.
7. D. A. Wolf-Gladrow. Lattice-gas cellular automata and lattice Boltzmann models: Introduction.
Berlin: Springer-Verlag, 2000.
8. S. Succi. The lattice Boltzmann equation: For fluid dynamics and beyond. Oxford, UK: Clarendon
Press, 2001.
9. Z. Guo and C. Shu. Lattice Boltzmann method and its applications in engineering. Singapore: World
Scientific, 2013.
Chapter 2
1. Y. W. Kwon and H.-C. Bang. The finite element method using MATLAB. 2nd ed. Boca Raton, FL:
CRC Press, 2000.
2. J. E. Akin. Finite element analysis for undergraduates. London: Academic Press, 1986.
3. O. C. Zienkiewicz and R. L. Taylor. The finite element method. 4th ed. London: McGraw-Hill,
1991.
Chapter 3
1. H. Chen. Discrete Boltzmann systems and fluid flows. Computational Physics, vol. 7, 1993,
pp. 632637.
2. S. Chen and G. D. Doolen. Lattice Boltzmann method for fluid flow. Annual Review of Fluid
Mechanics, vol. 30, 1998, pp. 329364.
3. Z. Guo and T. S. Zhao. Lattice Boltzmann model for incompressible flows through porous
media. Physical Review E, vol. 66, September 2002, p. 036304.
4. G. H. Tang, W. Q. Tao, and Y. L. He. Gas slippage effect on microscale porous flow using the
lattice Boltzmann method. Physical Review E, vol. 72, November 2005, p. 056301.
375
376
References
5. M. Yoshino and T. Inamuro. Lattice Boltzmann simulations for flow and heat/mass transfer
problems in a three-dimensional porous structure. International Journal for Numerical Methods
in Fluids, vol. 43, no. 2, 2003, pp. 183198.
6. T. Inamuro, T. Ogata, S. Tajima, and N. Konishi. A lattice Boltzmann method for incompressible two-phase flows with large density differences. Journal of Computational Physics, vol. 198,
no. 2, 2004, pp. 628644.
7. G. Hzi, A. R. Imre, G. Mayer, and I. Farkas. Lattice Boltzmann methods for two-phase flow
modeling. Annals of Nuclear Energy, vol. 29, no. 12, 2002, pp. 14211453.
8. Y. Yan and Y. Zu. A lattice Boltzmann method for incompressible two-phase flows on partial
wetting surface with large density ratio. Journal of Computational Physics, vol. 227, no. 1, 2007,
pp. 763775.
9. G. Breyiannis and D. Valougeorgis. Lattice kinetic simulations in three-dimensional magnetohydrodynamics. Physics Review E, vol. 69, June 2004, p. 065702.
10. M. Pattison, K. Premnath, N. Morley, and M. Abdou. Progress in lattice Boltzmann methods
for magnetohydrodynamic flows relevant to fusion applications. Fusion Engineering and Design,
vol. 83, no. 4, 2008, pp. 557572.
11. J. Carter and L. Oliker. Performance evaluation of lattice-Boltzmann magnetohydrodynamics
simulations on modern parallel vector systems. In High performance computing on vector systems,
ed. M. Resch, T. Bnisch, K. Benkert, W. Bez, T. Furui, and Y. Seo. Berlin: Springer, 2006, pp. 4150.
12. B. H. Elton. A lattice Boltzmann method for a two-dimensional viscous Burgers equation:
Computational results. Supercomputing, 1991, pp. 242252.
13. G. Yan and J. Zhang. A higher-order moment method of the lattice Boltzmann model for the
Korteweg-de Vries equation. Mathematics and Computers in Simulation, vol. 79, January 2009,
pp. 15541565.
14. N. S. Martys. Improved approximation of the Brinkman equation using a lattice Boltzmann
method. Physics of Fluids, vol. 13, no. 6, 2001, pp. 18071810.
15. L. Zhong, S. Feng, P. Dong, and S. Gao. Lattice Boltzmann schemes for the non-linear
Schrdinger equation. Physical Review E, vol. 74, September 2006, p. 036704.
16. C. K. Aidun and J. R. Clausen. Lattice-Boltzmann method for complex flows. Annual Review of
Fluid Mechanics, vol. 42, no. 1, 2010, pp. 439472.
17. R. Nourgaliev, T. Dinh, T. Theofanous, and D. Joseph. The lattice Boltzmann equation method:
Theoretical interpretation, numerics, and implications. International Journal of Multiphase Flow,
vol. 29, January 2003, pp. 117169.
18. S. Wolfram. Cellular automaton fluids 1: Basic theory. Journal of Statistical Physics, vol. 45, 1986,
pp. 471526.
19. U. Frisch, B. Hasslacher, and Y. Pomeau. Lattice-gas automata for the Navier-Stokes equation.
Physical Review Letters, vol. 56, April 1986, pp. 15051508.
20. P. L. Bhatnagar, E. P. Gross, and M. Krook. A model for collision processes in gases. I. Small
amplitude processes in charged and neutral one-component systems. Physical Review, vol. 94,
1954, pp. 511525.
21. D. dHumeres. Generalized lattice-Boltzmann equations. Paper presented at the Rarefied gas
dynamicsTheory and simulations; Proceedings of the 18th International Symposium on
Rarefied Gas Dynamics, University of British Columbia, Vancouver, Canada, July 1992.
22. P. Lallemand and L.-S. Luo. Theory of the lattice Boltzmann method: Dispersion, dissipation, isotropy, Galilean invariance, and stability. Physical Review E, vol. 61, June 2000,
pp.65466562.
23. Y. W. Kwon and J. C. Jo. Development of weighted residual based lattice Boltzmann techniques
for fluid-structure interaction application. Journal of Pressure Vessel Technology, vol. 131, June
2009, p. 031304.
24. S. R. Blair and Y. W. Kwon. Modeling of fluid-structure interaction using lattice Boltzmann
and finite element methods. Journal of Pressure Vessel Technology, vol. 137, April 2015, p. 021302.
25. A. K. Gunstensen, D. H. Rothman, S. Zaleski, and G. Zanetti. Lattice Boltzmann model of
immiscible fluids. Physical Review A, vol. 43, April 1991, pp. 43204327.
References
377
26. X. Shan and G. Doolen. Multicomponent lattice-Boltzmann model with interparticle interaction. Journal of Statistical Physics, vol. 81, 1995, pp. 379393.
27. M. R. Swift, E. Orlandini, W. R. Osborn, and J. M. Yeomans. Lattice Boltzmann simulations of
liquid-gas and binary fluid systems. Physical Review E, vol. 54, November 1996, pp. 50415052.
28. X. He, X. Shan, and G. D. Doolen. Discrete Boltzmann equation model for non ideal gases.
Physical Review E, vol. 57, January 1998, pp. R13R16.
29. D. H. Rothman and J. M. Keller. Immiscible cellular-automaton fluids. Journal of Statistical
Physics, vol. 52, 1988, pp. 11191127.
30. J. I. Q. Chang and D. Alexander. Application of lattice Boltzmann method, thermal multiphase fluid
dynamics. Saarbrucken, Germany: Verlag Dr. Mller, 2000.
31. J. Kahn and J. E. Hilliard. Free energy of a nonuniform system. I. Interfacial free energy. Journal
of Chemical Physics, vol. 28, 1958, pp. 258267.
32. X. Nie, Y.-H. Qian, G. D. Doolen, and S. Chen. Lattice Boltzmann simulation of the twodimensional Rayleigh-Taylor instability. Physical Review E, vol. 58, November 1998, pp.68616864.
33. N. Takada, M. Misawa, A. Tomiyama, and S. Hosokawa. Simulation of bubble motion under
gravity by lattice Boltzmann method. Journal of Nuclear Science and Technology, vol. 38, 2001, p. 330.
34. A. Lamura, G. Gonnella, and J. M. Yeomans. A lattice Boltzmann model of ternary fluid mixtures. Europhysics Letters, vol. 45, 1999, pp. 314320.
35. J. Rowlinson and B. Widom. Molecular theory of capillarity. Dover Books on Chemistry. Mineola,
NY: Dover, 2003.
36. J. Sethian. Level set methods and fast marching methods: Evolving interfaces in computational geometry, fluid mechanics, computer vision and materials sciences. Cambridge: Cambridge University
Press, 1999, p. 136.
37. C. Hirt and B. Nichols. Volume of fluid (VOF) method for the dynamics of free boundaries.
Journal of Computational Physics, vol. 39, no. 1, 1981, pp. 201225.
38. X. He, S. Chen, and R. Zhang. A lattice Boltzmann scheme for incompressible multiphase flow
and its application in simulation of Rayleigh-Taylor instability. Journal of Computational Physics,
vol. 152, no. 2, 1999, pp. 642663.
39. X. He, R. Zhang, S. Chen, and G. D. Doolen. On the three-dimensional Rayleigh-Taylor instability. Physics of Fluids, vol. 11, no. 5, 1999, p. 1143.
40. C. M. Teixeira. Incorporating turbulence models into the lattice-Boltzmann method.
International Journal of Modern Physics C, vol. 9, no. 8, 1998, pp. 11591175.
41. B. Chopard, P. Luthi, and S. Marconi. A lattice Boltzmann model for wave and fracture phenomena. Condensed Matter, 1998, p. 98122201.
42. S. R. Blair. Lattice Boltzmann methods for fluid structure interaction. PhD dissertation, Naval
Postgraduate School, Monterey, CA, September 2012.
43. B. F. Armaly, F. Durst, and J. C. F. Pereira. Experimental and theoretical investigation of backwardfacing step flow. Journal of Fluid Mechanics, vol. 127, 1983, pp. 473496.
44. U. Ghia, K. Ghia, and C. Shin. High-resolutions for incompressible flow using the NavierStokes equations and a multigrid method. Journal of Computational Physics, vol. 48, no. 3, 1982,
pp. 387411.
45. O. Botella and R. Peyret. Benchmark spectral results on the lid-driven cavity flow. Computers
and Fluids, vol. 27, no. 4, 1998, pp. 421433.
46. C.-H. Bruneau and M. Saad. The 2D lid-driven cavity problem revisited. Computers and Fluids,
vol. 35, no. 3, 2006, pp. 326348.
47. H. Zhou, G. Mo, F. Wu, J. Zhao, M. Rui, and K. Cen. GPU implementation of lattice Boltzmann
method for flows with curved boundaries. Computer Methods in Applied Mechanics and
Engineering, vol. 225228, no. 0, 2012, pp. 6573.
48. W. Li, X. Wei, and A. Kaufman. Implementing lattice Boltzmann computation on graphics
hardware. The Visual Computer, vol. 19, 2003, pp. 444456.
49. D. Calhoun. A Cartesian grid method for solving the two-dimensional stream functionvorticity equations in irregular regions. Journal of Computational Physics, vol. 176, no. 2, 2002,
pp. 231275.
378
References
50. S. Xu and Z. J. Wang. An immersed interface method for simulating the interaction of a fluid
with moving boundaries. Journal of Computational Physics, vol. 216, no. 2, 2006, pp. 454493.
51. D. Russell and Z. J. Wang. A Cartesian grid method for modeling multiple moving objects in 2D
incompressible viscous flow. Journal of Computational Physics, vol. 191, no. 1, 2003, pp. 177205.
52. L. Ong and J. Wallace. The velocity field of the turbulent very near wake of a circular cylinder.
Experiments in Fluids, vol. 20, 1996, pp. 441453.
53. A. L. E. Silva, A. Silveira-Neto, and J. Damasceno. Numerical simulation of two-dimensional
flows over a circular cylinder using the immersed boundary method. Journal of Computational
Physics, vol. 189, no. 2, 2003, pp. 351370.
54. C. Obrecht, F. Kuznik, B. Tourancheau, and J.-J. Roux. A new approach to the lattice Boltzmann
method for graphics processing units. Computers and Mathematics with Applications, vol. 61,
no.12, 2011, pp. 36283638.
55. P. Rinaldi, E. Dari, M. Vnere, and A. Clausse. A lattice-Boltzmann solver for 3D fluid simulation on GPU. Simulation Modelling Practice and Theory, vol. 25, 2012, pp. 163171.
56. M. Astorino, J. B. Sagredo, and A. Quarteroni. A modular lattice Boltzmann solver for GPU
computing processors. Tech. Rep. MATHICSE-TR-06-2011. Lucerne, Switzerland: Mathematics
Institute of Computational Science and Engineering, July 2011.
57. K. Mattila, J. Hyvluoma, J. Timonen, and T. Rossi. Comparison of implementations of the latticeBoltzmann method. Computers and Mathematics with Applications, vol. 55, no. 7, 2008, pp. 15141524.
Chapter 4

1. B. Chopard, A. Dupuis, A. Masselot, and P. Luthi. Cellular automata and lattice Boltzmann techniques: An approach to model and simulate complex systems. Computer Science Department,
University of Geneva, 1998.
2. A. Burks, ed. Von Neumanns self-reproducing automata. In Essays on cellular automata.
Champaign: University of Illinois Press, 1970, pp. 364.
3. M. Gardner. The fantastic combinations of John Conways new solitaire game of life. Scientific
American, vol. 223, 1970, pp. 120123.
4. S. Wolfram. Theory and application of cellular automata. Singapore: World Scientific, 1986.
5. S. Wolfram. Cellular automata and complexity. Reading, MA: Addison-Wesley, 1994.
6. S. Wolfram. A new kind of science. Champaign, IL: Wolfram Media, 2002.
7. B. Chopard. A cellular automata model of large-scale moving objects. Journal of Physics A:
Mathematical and General, vol. 23, 1990, pp. 16711678.
8. W. T. Thomson, and M. D. Dahleh. Theory of vibration with applications. 5th ed. Englewood
Cliffs, NJ: Prentice Hall, 1998.
9. F. S. Tse, I. E. Morse, and R. T. Hinkle. Mechanical vibrations, theory and applications. 2nd ed.
Boston: Allyn and Bacon, 1978.
10. S. Hosoglu. Cellular automata: An approach to wave propagation and fracture mechanics
problems. MS thesis, Naval Postgraduate School, Monterey, CA, 2006.
11. L. E. Craugh. Coupled finite element and cellular automata methods for analysis of composite
structures in an acoustic domain. PhD dissertation, Naval Postgraduate School, Monterey, CA,
September 2012.
12. B. Chopard and M. Droz. Cellular automata modeling of physical systems. Cambridge: Cambridge
13. M. C. Junger and D. Feit. Sound, structures, and their interaction. Cambridge, MA: MIT Press, 1986.
14. Y. W. Kwon and S. Hosoglu, Application of lattice Boltzmann method, finite element method,
and cellular automata and their coupling to wave propagation problems. Computers and
Structures, vol. 86, no. 78, 2008, pp. 663670.
References
379
15. P. C. Etter. Underwater acoustic modeling. New York: Spon Press, 2003.
16. H. Schmidt and F. B. Jensen. Computational ocean acoustics: Advances in 3D ocean acoustic modeling. Beijing: American Institute of Physics, 2012.
17. A. D. White. SONAR for practising engineers. London: Wiley, 2002.
18. J. A. Bailey. Uniform and multi-grid modeling of acoustic wave propagation with cellular
automaton techniques. MS thesis, Naval Postgraduate School, Monterey, CA, 2013.
19. F. Sturm and J. A. Fawcett. On the use of higher-order azimuthal schemes in 3-D modeling.
Acoustical Society of America, vol. 113, no. 6, 2003, pp. 31343145.
Chapter 5
1. D. A. Rapport. The art of molecular dynamics simulation. 2nd ed. Cambridge: Cambridge
2. J. M. Haile. Molecular dynamics simulation: Elementary methods. New York: Wiley, 1997.
3. Y. W. Kwon and S. H. Jung. Atomic model and coupling with continuum model for static equilibrium problems. Computers and Structures, vol. 82, no. 2326, 2004, pp. 19932000.
4. J. E. Lennard-Jones. The determination of molecular fields. I. From the variation of the viscosity of a gas with temperature. Proceedings of the Royal Society (London), vol. 106A, 1924,
pp.441462.
5. J. E. Lennard-Jones. The determination of molecular fields. II. From the equation of state of a
gas. Proceedings of the Royal Society (London), vol. 106A, 1924, pp. 463477.
6. C. Hsieh and R. Thomson. Lattice theory of fracture and crack creep. Journal of Applied Physics,
vol. 44, 1973, pp. 20512063.
7. G. C. Abell. Empirical chemical pseudopotential theory of molecular and metallic bonding.
Physical Review B, vol. 31, 1984, pp. 61846196.
8. J. Tersoff. New empirical approach for the structure and energy of covalent systems. Physical
Review B, vol. 37, 1987, pp. 69917000.
9. J. Tersoff and R. S. Ruoff. Structural properties of a carbon-nanotube crystal. Physical Review
Letters, vol. 73, August 1994, pp. 676679.
10. D. W. Brenner. Empirical potential for hydrocarbons for use in simulating the chemical vapor
deposition of diamond films. Physical Review B, vol. 42, 1990, pp. 94589471.
11. W. G. Wilder, L. C. Venema, A. G. Rinzler, R. E. Smalley, and C. Dekker. Electronic structure
of atomically resolved carbon nanotubes. Nature, vol. 391, 1998, pp. 5962.
12. T. W. Odom, J.-L. Huang, P. Kim, and C. M. Lieber. Atomic structure and electronic properties
of single walled carbon nanotubes. Nature, vol. 391, 1998, pp. 6264.
13. S. T. Thornton and A. Rex. Modern physics for scientists and engineers. Orlando, FL: Sanders
College, 2000.
14. S. O. Kasap. Principles of electronic materials and devices. New York: McGraw-Hill, 2002.
15. J. J. Oh. Determination of Youngs modulus of carbon nanotubes using MD simulation. MS
thesis, Naval Postgraduate School, Monterey, CA, December 2003.
16. Y. W. Kwon, C. Manthena, J. J. Oh, and D. Srivastava. Vibrational characteristics of carbon
nanotubes as nanomechanical resonators. Journal of Nanoscience and Nanotechnology, vol. 5,
no.5, May 2005, pp. 703712.
17. M. S. Dresselhaus, G. Dresselhaus, and P. C. Eklund. Science of fullerenes and carbon nanotubes.
New York: Academic Press, 1996.
18. C. Schnenberger, A. Bachtold, C. Strunk, J.-P. Salvetat, and L. Forr. Interference and interaction in multiwall carbon nanotubes. Applied Physics A, vol. 69, 1999, pp. 283295.
19. A. Krishnan, E. Dujardin, T. W. Ebbesen, P. N. Yianilos, and M. M. J. Treacy. Youngs modulus
of single-walled nanotubes. Physical Review B, vol. 58, 1998, pp. 1401314019.
380
References
20. D. H. Oh and Y. H. Lee. Stability and cap formation mechanism of single-walled carbon nanotubes. Physical Review B, vol. 58, 1998, pp. 74077411.
21. B. I. Yakobson, C. J. Brabec, and J. Bernholc. Nanomechanics of carbon tubes: Instabilities
beyond linear response. Physical Review Letters, vol. 76, 1996, pp. 25112514.
22. J. P. Lu. Elastic properties of carbon nanotubes and nanoropes. Physical Review Letters, vol. 79,
1997, pp. 12971300.
23. E. Hernandez, C. Goze, P. Bernier, and A. Rubio. Elastic properties of C and BxCyNz composite
nanotubes. Physical Review Letters, vol. 80, 1998, pp. 45024505.
24. D. Sanchez-Portal, E. Artacho, J. M. Solar, A. Rubio, and P. Ordejon. Ab initio structural, elastic,
and vibrational properties of carbon nanotubes. Physical Review B, vol. 59, 1999, pp.1267812688.
25. P. Poncharal, Z. L. Wang, D. Ugarte, and W. A. DeHeer. Electrostatic deflections and electromechanical resonances of carbon nanotubes. Science, vol. 283, 1999, pp. 15131516.
26. S. Iijima, C. J. Brabec, A. Maiti, and J. Bernholc. Structural flexibility of carbon nanotubes.
Journal of Chemical Physics, vol. 104, 1996, pp. 20892092.
27. E. W. Wong, P. E. Sheehan, and C. M. Lieber. Nanobeam mechanics: Elasticity, strength, and
toughness of nanorods and nanotubes. Science, vol. 277, 1997, pp. 19711975.
28. D. Srivastava, Chenyu Wei, and Kyeongjae Cho. Nanomechanics of carbon nanotubes and
composites. Applied Mechanics Reviews, vol. 56, no. 2, 2003, pp. 215230.
29. P. Heino, H. Hkkinen, and K. Kaski. Molecular dynamics study of copper with defects under
strain. Physical Review B, vol. 58, 998, pp. 641652.
30. H. Kuzmany, B. Burger, A. Thess, and R. E. Smalley. Vibrational spectra of single wall carbon
nanotubes. Carbon, vol. 36, no. 56, 1998, pp. 709712.
31. A. M. Rao, E. Richter, S. Bandow, B. Chase, P. C. Eklund, K. A. Williams, S. Fang, K. R.
Subbaswamy, M. Menon, A. Thess, R. E. Smalley, G. Dresselhaus, and M. S. Dresselhaus.
Diameter selective Raman scattering from vibrational modes in carbon nanotubes. Science,
vol. 275, no. 10, 1997, pp. 187191.
32. L. Henrard, E. Hernandes, P. Bernier, and A. Rubio. Van der Waals interaction in nanotube bundles: Consequences on vibrational modes. Physical Review B, vol. 60, no. 12, 1999,
pp.R8521R8524.
33. N. Koratkar, B. Q. Wei, and P. M. Ajayan. Carbon nanotubes for damping applications. Advanced
Materials, vol. 14, 2002, pp. 9971000.
34. S. J. Barsky and M. Plischke. Elastic properties of randomly cross-linked polymers. Physical
Review E, vol. 54, 1996, pp. 53705376.
35. Y. W. Kwon and A. F. Harrell. How many monomer repeat units are necessary for reliable molecular dynamics simulation? Polymer and Polymer Composites, vol. 12, no. 6, 2004, pp. 483489.
36. O. C. Nwobi, L. N. Long, and M. M. Micci. Molecular dynamics studies of thermophysical properties
of supercritical ethylene. Reston, VA: American Institute of Aeronautics and Astronautics, 1998.
37. Y. W. Kwon, J. AlRowaijeh, and D. Kidd. Multiscale analysis of stationary and flowing media
with particle inclusion. Journal of Computational and Theoretical Nanoscience, vol. 7, no. 4, April
2010, pp. 700708.
38. Y. W. Kwon. Nanomechanics. In Nanoengineering of structural, functional and smart materials, ed.
M. J. Schulz, A. Kelkar, and M. J. Sundaresan. Boca Raton, FL, 2005, pp. 469500.
Chapter 6

1. Y. W. Kwon. Discrete atomic and smeared continuum modeling for static analysis. Engineering
Computations, vol. 20, no. 8, 2003, pp. 964978.
References
381
3. A. M. Cuitino, L. Stainier, G. Wang, A. Strachan, T. Cagin, W. A. Goddard III, and M. Ortiz.

A multi-scale approach for modeling crystalline solids. Journal of Computer-Aided Materials
Design, vol. 8, 2001, pp. 127149.
4. Y. W. Kwon. Multi-scale modeling of mechanical behavior of polycrystalline materials. Journal
of Computer-Aided Materials Design, vol. 11, no. 1, March 2005, pp. 4357.
5. G. J. Wagner and W. K. Liu. Coupling of atomistic and continuum simulations using a bridging
scale decomposition. Journal of Computational Physics, vol. 190, 2003, pp. 249274.
6. S. P. Xiao and T. Belytschko. A bridging domain method for coupling continua with
molecular dynamics. Computer Methods in Applied Mechanics and Engineering, vol. 193, 2004,
pp.16451669.
7. Y. W. Kwon and C. Manthena. Homogenization technique of discrete atoms into smeared continuum. Internal Journal of Mechanical Sciences, vol. 48, 2006, pp. 13521359.
8. R. G. Hoagland, S. M. Foiles, and M. I. Baskes. An atomic model crack tip deformation in aluminium using an embedded atom potential. Journal of Material Research, vol. 5, no. 2, February
1990, pp. 313324.
9. Y. W. Kwon and S. Hosoglu. Application of lattice Boltzmann method, finite element method,
and cellular automata and their coupling to wave propagation problems. Computers and
Structures, 86, 2008, pp. 663670.
10. L. E. Craugh and Y. W. Kwon. Coupled finite element and cellular automata methods for analysis of composite structures with fluidstructure interaction. Composite Structures, 102, August
2013, pp. 124137.
Chapter 7

1. Y. W. Kwon, D. H. Allen, and R. Talreja, eds. Multiscale modeling and simulation of composite materials and structures. New York: Springer, 2008.
2. Y. W. Kwon. Calculation of effective moduli of fibrous composites with micro-mechanical
damage. Composite Structures, vol. 25, 1993, pp. 187192.
3. Y. W. Kwon and J. M. Berner. Micromechanics model for damage and failure analyses of laminated fibrous composites. Engineering Fracture Mechanics, vol. 52, no. 2, 1995, pp. 231242.
4. Y. W. Kwon and J. M. Berner. Matrix damage analysis of fibrous composites: Effects of thermal residual stresses and layer sequences. Computers and Structures, vol. 64, no. 14, 1997,
pp.375382.
5. Y. W. Kwon and A. Altekin. Multi-level, micro-macro approach for analysis of woven fabric
composites. Journal of Composite Materials, vol. 36, no. 8, pp. 10051022.
6. Y. W. Kwon and K. Roach. Unit-cell model of 2/2-twill woven fabric composites for multi-scale
analysis. Computer Modeling in Engineering and Sciences, vol. 5, no. 1, 2004, pp. 6372.
7. Y. W. Kwon and W. M. Cho. Multi-scale thermal stress analysis of woven fabric composite.
Journal of Thermal Stresses, vol. 27, 2004, pp. 5973.
8. Y. W. Kwon and L. E. Craugh. Progressive failure modeling in notched cross-ply fibrous composites. Applied Composite Materials, vol. 8, no. 1, January 2001, pp. 6374.
9. Y. W. Kwon and C. Kim. Micromechanical model for thermal analysis of particulate and
fibrous composites. Journal of Thermal Stresses, vol. 21, 1998, pp. 2139.
10. Y. W. Kwon. Analysis of laminated and sandwich composite structures using solid-like shell
elements. Applied Composite Materials, vol. 20, no. 4, 2013, pp. 355373.
11. Y. W. Kwon and M. S. Park. Versatile micromechanics model for multiscale analysis of composite structures. Applied Composite Materials, vol. 20, no. 4, 2013, pp. 673692.
12. M. S. Park and Y. W. Kwon. Elastoplastic micromechanics model for multiscale analysis of
metal matrix composite structures. Computers and Structures, vol. 123, 2013, pp. 2838.
382
References
13. G. Weng. The overall elastoplastic stress-strain relations of dual-phase metals. Journal of the
Mechanics and Physics of Solids, vol. 38, no. 3, 1990, pp. 419441.
14. W. Yu and T. Tang. Variational asymptotic method for unit cell homogenization of periodically
heterogeneous materials. International Journal of Solids and Structures, vol. 44, no. 1112, 2007,
pp.37383755.
15. D. J. Lloyd. Particle reinforced aluminum and magnesium matrix composites. International
Materials Review, vol. 39, no. 1, 1994, pp. 123.
16. C. Hsieh and W. Tuan. Elastic properties of ceramicmetal particulate composites. Materials
Science and Engineering: A, vol. 393, no. 1, 2005, pp. 133139.
17. J. A. Oliveira, J. Pinho-da-Cruz, and F. Teixeira-Dias. Asymptotic homogenisation in linear elasticity. Part II: Finite element procedures and multiscale applications. Computational
Materials Science, vol. 45, no. 4, 2009, pp. 10811096.
18. D. Kenaga, J. F. Doyle, and C. T. Sun. Characterization of boron/aluminum composite in the
nonlinear range as an orthotropic elastic-plastic material. Journal of Composite Materials, vol. 21,
no. 6, 1987, pp. 516531.
19. R. J. Arsenault, S. Fishman, and M. Taya. Deformation and fracture behavior of metal-ceramic
matrix composite materials. Progress in Materials Science, vol. 38, no. 1, 1994, pp. 1157.
20. W. Pabst, E. Gregorova, G. Ticha, and E. Tynova. Effective elastic properties of alumina-zirconia
composite ceramicsPart 4. Tensile modulus of porous alumina and zirconia. Ceramics-Silikaty,
vol. 48, no. 4, 2004, pp. 165174.
21. Z. Hashin. Analysis of composite materials. Journal of Applied Mechanics, vol. 50, no. 2, 1983,
pp.481505.
22. J. Ye, B. Q. Han, F. Tang, and J. M. Schoenung. Mechanical behavior of a tri-modal Al matrix
composite. MRS Proceedings, vol. 880, 2004. http://dx.doi.org/10.1557/PROC-880-BB1.5.
23. Y. Li, Y. H. Zhao, V. Ortalan, W. Liu, Z. H. Zhang, R. G. Vogt, N. D. Browning, E. J. Lavernia, and
J. M. Schoenung. Investigation of aluminum-based nanocomposites with ultra-high strength.
Materials Science and Engineering: A, vol. 527, no. 12, 2009, pp. 305316.
24. R. M. Jones. Mechanics of composite materials. Washington, DC: Scripta Book, 1975.
25. H. W. Herring, R. M. Baucom, and R. A. Pride. Mechanical behaviors of boron-epoxy and
glass-epoxy filament-wound cylinders under various loads. NASA TN D-5050. Hampton, VA:
Langley Research Center, Langley Station, 1996.
26. L. B. Greszczuk. Stress concentrations and failure criteria for orthotropic and anisotropic plates
with circular openings. In Composite materials: Testing and design (second conference). Baltimore,
MD: American Society for Testing and Materials, 1972, pp. 363381.
27. S. W. Tsai and H. T. Hahn. Introduction to composite materials. Westport, CT: Technomic, 1980.
28. J. C. Simo and R. L. Taylor. Consistent tangent operators for rate-independent elastoplasticity.
Computer Methods in Applied Mechanics and Engineering, vol. 48, no. 1, 1985, pp. 101118.
29. E. A. de Souza Neto, D. Peric, and D. R. J. Owen. Computational methods for plasticity: Theory and
applications. New York: Wiley, 2011.
30. Y. S. Suh, S. P. Joshi, and K. T. Ramesh. An enhanced continuum model for size-dependent
strengthening and failure of particle-reinforced composites. Acta Materialia, vol. 57, no. 19,
2009, pp. 58485861.
31. S. Qu, T. Siegmund, Y. Huang, P. D. Wu, F. Zhang, and K. C. Hwang. A study of particle size
effect and interface fracture in aluminum alloy composite via an extended conventional theory of mechanism-based strain-gradient plasticity. Composites Science and Technology, vol. 65,
no. 7, 2005, pp. 12441253.
32. J. Llorca, S. Suresh, and A. Needleman. An experimental and numerical study of cyclic deformation in metal-matrix composites. Metallurgical and Materials Transactions A, vol. 23, no. 3,
1992, pp. 919934.
33. R. Arsenault and M. Taya. Thermal residual stress in metal matrix composite. Acta Metallurgica,
vol. 35, no. 3, 1987, pp. 651659.
References
383
34. C. T. Sun, J. F. Doyle, and D. Kenaga. The characterization of boron/aluminum composite as an

orthotropic elastic-plastic material. AFWAL-TR-85-3019. Wright-Patterson Air Force Base, OH:
Flight Dynamics Laboratory, Air Force Wright Aeronautical Laboratories, Air Force Systems
Command, 1985.
35. D. F. Adams. Inelastic analysis of a unidirectional composite subjected to transverse normal
loading. Journal of Composite Materials, vol. 4, no. 3, 1970, pp. 310328.
Chapter 8
1. J. Belak. Multi-scale applications to high strain-rate dynamic fracture. Journal of ComputerAided Materials Design, vol. 9, 2002, pp. 165172.
2. H. Rafii-Tabar and A. Chirazi. Multi-scale computational modeling of solidification phenomena. Physics ReportsReview Section of Physics Letters, vol. 365, No. 3, July 2002, pp. 145249.
3. J. S. Chen and S. Mehraeen. Variationally consistent multi-scale modeling and homogenization of stresses grain growth. Computer Methods in Applied Mechanics and Engineering, vol. 193,
no. 1720, 2004, pp. 18251848.
4. N. Takano, M. Zako, and Y. Okuno. Multi-scale finite element analysis of porous materials
and components by asymptotic homogenization theory and enhanced mesh superposition
method. Modelling and Simulation in Materials Science and Engineering, vol. 11, no. 2, March 2003,
pp. 137156.
5. A. M. Cuitino, L. Stainier, G. Wang, A. Strachan, T. Cagin, W. A. Goddard III, and M. Ortiz.
A multi-scale approach for modeling crystalline solids. Journal of Computer-Aided Materials
Design, vol. 8, 2001, pp. 127149.
6. S. Hao, W. K. Liu, B. Moran, F. Vernerey, and G. B. Olson. Multi-scale constitutive model
and computational framework for the design of ultra-high strength, high toughness steels.
Computer Methods in Applied Mechanical Engineering, vol. 193, 2004, pp. 18651908.
7. M. F. Horstemeyer and P. Wang. Cradle-to-grave simulation-based design incorporating
multiscale microstructure-property modeling: Reinvigorating design with science. Journal of
Computer-Aided Materials Design, vol. 10, 2003, pp. 1334.
8. G. Z. Voyiadjis and R. J. Dorgan. Bridging of length scales through gradient theory and diffusion equations of dislocations. Computer Methods in Applied Mechanical Engineering, vol. 193,
2004, pp. 16711692.
9. E. B. Tadmor, M. Ortiz, and R. Phillips. Quasicontinuum analysis of defects in solids.
Philosophical Magazine, vol. A73, 1996, pp. 15291563.
10. T. Zhu, J. Li, K. J. Van Vliet, S. Ogata, S. Yip, and S. Suresh. Predictive modeling of nanoindentationinduced homogeneous dislocation nucleation in copper. Journal of Mechanics and Physics of
Solids, vol. 52, 2004, pp. 691724.
11. Y. W. Kwon. Discrete atomic and smeared continuum modeling for static analysis. Engineering
Computations, vol. 20, no. 8, 2003, pp. 964978.
12. S. P. Xiao and T. Belytschko. A bridging domain method for coupling continua with molecular
dynamics. Computer Methods in Applied Mechanical Engineering, vol. 193, 2004, pp. 16451669.
13. J. L. Ericksen. The Cauchy and Born hypothesis for crystals. In Phase transformation and material
instabilities in solids, ed. M. E. Gurtin. New York: Academic Press, 1984, pp. 6178.
14. W. K. Liu, E. G. Karpov, S. Zhang, and H. S. Park. An introduction to computational nanomechanics and materials. Computer Methods in Applied Mechanical Engineering, vol. 193, 2004,
pp. 15291578.
15. D. Qian, G. J. Wagner, and W. K. Liu. A multiscale projection method for the analysis of carbon
nanotubes. Computer Methods in Applied Mechanical Engineering, vol. 193, 2004, pp. 16031632.
384
References
17. Y. W. Kwon. Multi-scale modeling of mechanical behavior of polycrystalline materials. Journal
of Computer-Aided Materials Design, vol. 11, no. 1, 2004, pp. 4357.
Chapter 9
1. T. J. Vaughan, C. T. McCarthy, and L. M. McNamara. A three-scale finite element investigation
into the effects of tissue mineralization and lamellar organization in human cortical and trabecular bone. Journal of the Mechanical Behavior of Biomedical Materials, vol. 12, 2012, pp. 5062.
2. F.-Z. Cui, Y. Li, and J. Ge. Self-assembly of mineralized collagen composites. Materials Science
and Engineering, vol. 57, 2007, pp. 127.
3. J.-Y. Rho, L. Kuhn-Spearing, and P. Zioupos. Mechanical properties and the hierarchical structure of bone. Medical Engineering and Physics, vol. 20, 1998, pp. 92102.
4. K. H. Buschow, J. Cahn, R. W. Flemings, M. C. Ilschner, B. Kramer, and E. J. Mahajan. Bone mineralization. In Encyclopedia of materialsScience and technology, eds. K. H. J. Buschow, R. Cahn,
M. Flemings, B. Ilschner, E. Kramer, S. Mahajan, and P. Veyssiere. Oxford, UK: Pergamon, 2001,
pp. 787794.
5. L. L. Hench. Clinical needs and concepts of repair. In Biomaterials, artificial organs and tissue
engineering, eds. L. L. Hench and J. R. Jones. Sawston, UK: Woodhead, 2005, pp. 7989.
6. B. R. Clumpner. Multiscale modeling of bone. MS thesis, Naval Postgraduate School,
Monterey,CA, December 2014.
7. N. Reznikov, R. Shahar, and S. Weiner. Bone hierarchical structure in three dimensions. Acta
Biomaterialia, vol. 10, 2014, pp. 38153826.
8. D. Taylor. Failure processes in hard and soft tissues. In Comprehensive structural integrity.
Volume 9, Bioengineering, eds. Y.-W. Mai and S.-H. Teoh. New York: Elsevier, 2003, pp. 3592.
9. Z.-C. Xing, K.-W. Chang, S. Chun, S. Kim, and I.-K. Kang. Immobilization of collagen onhydroxyapatite discs by covalent bonding and physical absorption and their interactionwithMC3T3-
E1osteoblasts. Tissue Engineering and Regenerative Medicine, vol. 11, no. 2, 2014,pp. 99105.
10. A. Zamiri and S. De. Mechanical properties of hydroxyapatite single crystals from nano
indentation data. Journal of the Mechanical Behavior of Biomedical Materials, vol. 2, no. 4, 2011,
pp.146152.
11. L. J. Katz, J. L. Katz, and K. Ukraincik. On the anisotropic elastic properties of hydroxyapatite.
Journal of Biomechanics, vol. 4, 1971, pp. 221227.
12. C. Rey, C. Combes, C. Drouet, and M. J. Glimcher. Bone mineral: Update on chemical composition and structure. Osteoporosis International, vol. 20, no. 6, 2009, pp. 10131021.
13. S. Ricard-Blum. The collagen family. Cold Spring Harbor Perspectives in Biology, vol. 3, 2010, a004978.
14. M. Tzaphlidou and P. Berillis. Collagen fibril diameter in relation to bone site. A quantitative
ultrastructural study. Micron, vol. 36, 2005, pp. 703705.
15. M. J. Buehler and S. Y. Wong. Entropic elasticity controls nanomechanics of single tropocollagen molecules. Biophysical Journal, vol. 93, 2007, pp. 3743.
16. M. D. Shoulders and R. T. Raines. Collagen structure and stability. Annual Review of Biochemistry,
vol. 78, 2009, pp. 929958.
17. B. Brodsky and A. V. Persikov. Molecular structure of the collagen triple helix. Advances in
Protein Chemistry, vol. 70, 2005, pp. 301339.
18. A. Bhattacharjee and M. Bansal. Collagen structure: The Madras triple helix and the current
scenario. Life, vol. 57, no. 3, 2005, pp. 161172.
19. K. A. Piez and A. Miller. The structure of collagen fibrils. Journal of Supramolecular Structure,
vol. 2, 1974, pp. 121137.
References
385
20. T. I. Nikolaeva, E. I. Tiktopulo, E. N. Ilyasova, and S. M. Kuznetsova. Collagen type I fibril

packing in vivo and in vitro. Biophysics, vol. 52, no. 5, 2007, pp. 489497.
21. M. J. Stevens and P. J. int Veld. Simulation of the mechanical strength of a single collagen molecule. Biophysical Journal, vol. 95, 2008, pp. 3339.
22. M. J. Buehler. Nanomechanics of collagen fibrils under varying cross-link densities: Atomistic
and continuum studies. Journal of the Mechanical Behavior of Biomedical Materials, vol. 1, 2008,
pp.5967.
23. N. Sasaki, H. Shirakawa, T. Nozoe, and K. Furusawa. Elastic properties of collagen in bone
determined by measuring the Debye-Waller factor. Journal of Biomechanics, vol. 46, 2013,
pp.28242830.
24. S. D. Bolboac and L. Jntschi. Amino acid sequence analysis on collagen. Bulletin of University
of Agricultural Sciences and Veterinary Medicine-Animal Sciences and Biotechnologies, vol. 63, 2007,
pp. 311316.
25. K. E. Kadler, D. F. Holmes, J. A. Trotter, and J. A. Chapman. Collagen fibril formation. Biochemical
Journal, vol. 316, 1996, pp. 111.
26. A. Steplewski, V. Hintze, and A. Fertala. Molecular basis of organization of collagen fibrils.
Journal of Structural Biology, vol. 157, 2007, pp. 297307.
27. J. P. R. O. Orgel, A. V. Persikov, and O. Antipova. Variation in the helical structure of native
collagen. PLoS One, vol. 9, no. 2, 2014, e89519.
28. J. Bella. A new method for describing the helical conformation of collagen: Dependence of the
triple helical twist on amino acid sequence. Journal of Structural Biology, vol. 170, 2010, pp. 377391.
29. S. Schweizer, A. Bick, L. Subramanian, and X. Krokidis. Influences on the stability of collagen
triple-helix. Fluid Phase Equilibria, vol. 362, 2014, pp. 113117.
30. T. I. Nikolaeva, S. M. Kuznetsova, and V. V. Rogachevsky. Collagen fibril formation in vitro at
nearly physiological temperatures. Biophysics, vol. 57, no. 6, 2012, pp. 757763.
31. R. Hambli and A. Barkaoui. Physically based 3D finite element model of a single mineralized
collagen fibril. Journal of Theoretical Biology, vol. 301, 2012, pp. 2841.
32. M. J. Buehler. Atomistic and continuum modeling of mechanical properties of c ollagen:Elasticity,
fracture, and self-assembly. Journal of Materials Research, vol. 21, no. 8, 2006, pp. 19471961.
33. S. Vesentini, C. F. C. Fiti, F. M. Montevecchi, and A. Redaelli. Molecular assessment of
the elastic properties of collagen-like homotrimer sequence. Biomechanics and Modeling in
Mechanobiology, vol. 2, 2005, pp. 224234.
34. N. Sasaki and S. Odajima. Stress-strain curve and Youngs modulus of a collagen molecule
as determined by the X-ray diffraction technique. Journal of Biomechanics, vol. 29, no. 5, 1996,
pp.655658.
35. R. Harley, D. James, A. Miller, and J. W. White. Phonons and the elastic moduli of collagen and
muscle. Nature, vol. 267, 1997, pp. 285287.
36. S. Cusack and A. Miller. Determination of elastic constants of collagen by Brillouin light scattering. Journal of Molecular Biology, vol. 135, no. 1, 1979, pp. 3951.
37. H. Hoffman, T. Voss, K. Khn, and J. Engel. Localization of flexible sites in thread-like molecules from electron micrographs: Comparison of interstitial, basement membrane and intima
collagens. Journal of Molecular Biology, vol. 172, 1984, pp. 325343.
38. Y.-L. Sun, Z.-P. Luo, A. Fertala, and K.-N. An. Stretching type II collagen with optical tweezers.
39. A. C. Lorenzo and E. R. Caffarena. Elastic properties, Youngs modulus determination and
structural stability of the tropocollagen molecule: A computational study by steered molecular dynamics. Journal of Biomechanics, vol. 38, 2005, pp. 15271533.
40. A. Gautieri, S. Vesentini, A. Redaelli, and R. Ballarini. Modeling and measuring visco-elastic
properties: From collagen molecules to collagen fibrils. International Journal of Non-Linear
Mechanics, vol. 56, 2013, pp. 2533.
41. D. L. Christiansen, E. K. Huang, and F. H. Silver. Assembly of type I collagen: Fusion of fibril
subunits and the influence of fibril diameter on mechanical properties. Matrix Biology, vol. 19,
2000, pp. 409420.
386
References
42. D. K. Dubey and V. Tomar. Understanding the influence of structural hierarchy and its coupling with chemical environment on the strength of idealized tropocollagen-hydroxyapatite
biomaterials. Journal of the Mechanics and Physics of Solids, vol. 57, 2009, pp. 17021717.
43. D. K. Dubey and V. Tomar. Role of the nanoscale interfacial arrangement in mechanical
strength of tropocollagen-hydroxyapatite-based hard biomaterials. Acta Biomaterialia, vol. 5,
2009, pp. 27042716.
44. M. J. Buehler. Nature designs tough collagen: Explaining the nanostructure of collagen fibrils.
Proceedings of the National Academy of Sciences of the United States of America, vol. 103, no. 33, 2006,
pp. 1228512290.
45. C. E. Tye, G. K. Hunter, and H. A. Goldberg. Glycobiology and extracellular matrices:Identification
of the type I collagen-binding domain of bone sialoprotein and characterization of the mechanism of interaction. Journal of Biological Chemistry, vol. 280, no. 14, 2005, pp. 1348713492.
46. K. Sato, T. Kogure, Y. Kumagai, and J. Tanaka. Crystal orientation of hydroxyapatite induced
by ordered carboxyl groups. Journal of Colloid and Interface Science, vol. 240, 2001, pp. 133138.
47. S. M. Pradhan, K. S. Katti, and D. R. Katti. Multiscale model of collagen fibril in bone: Elastic
response. Journal of Engineering Mechanics, vol. 140, 2014, pp. 454461.
48. I. Jger and P. Fratzl. Mineralized collagen fibrils: A mechanical model with a staggered
arrangement of mineral particles. Biophysical Journal, vol. 79, no. 4, 2000, pp. 17371746.
49. F. Yuan, S. R. Stock, D. R. Haeffner, J. D. Almer, D. C. Dunand, and L. C. Brinson. A new model
to simulate the elastic properties of mineralized collagen fibril. Biomechanics and Modeling in
50. C. A. Grant, M. A. Phillips, and N. H. Thomson. Dynamic mechanical analysis of collagen
fibrils at the nanoscale. Journal of the Mechanical Behavior of Biomedical Materials, vol. 5, 2012,
pp.165170.
51. Y. W. Kwon. Calculation of effective moduli of fibrous composites with micro-mechanical
damage. Composite Structures, vol. 25, 1993, pp. 187192.
52. Y. W. Kwon and J. M. Berner. Micromechanics model for damage and failure analyses of laminated fibrous composites. Engineering Fracture Mechanics, vol. 52, no. 2, 1995, pp. 23124.
53. Y. W. Kwon and J. M. Berner. Matrix damage analysis of fibrous composites: Effects of thermal residual stresses and layer sequences. Computers and Structures, vol. 64, no. 14, 1997,
pp.375382.
54. Y. W. Kwon and W. M. Cho. Multi-scale thermal stress analysis of woven fabric composite.
Journal of Thermal Stresses, vol. 27, 2004, pp. 5973.
55. Y. W. Kwon and L. E. Craugh. Progressive failure modeling in notched cross-ply fibrous composites. Applied Composite Materials, vol. 8, no. 1, January 2001, pp. 6374.
56. Y. W. Kwon and C. Kim. Micromechanical model for thermal analysis of particulate and
fibrous composites. Journal of Thermal Stresses, vol. 21, 1998, pp. 2139.
57. Y. W. Kwon. Analysis of laminated and sandwich composite structures using solid-like shell
elements. Applied Composite Materials, vol. 20, no. 4, 2013, pp. 355373.
58. Y. W. Kwon and M. S. Park. Versatile micromechanics model for multiscale analysis of composite structures. Applied Composite Materials, vol. 20, no. 4, 2013, pp. 673692.
59. M. S. Park and Y. W. Kwon. Elastoplastic micromechanics model for multiscale analysis of
metal matrix composite structures. Computers and Structures, vol. 123, 2013, pp. 2838.
60. M. P. E. Wenger, L. Bozec, M. A. Horton, and P. Mesquida. Mechanical properties of collagen
fibrils. Biophysical Journal, vol. 93, no. 4, 2007, pp. 12551263.
61. A. K. Nair, A. Gautieri, S.-W. Chang, and M. J. Buehler. Molecular mechanics of mineralized
collagen fibrils in bone. Nature Communications, vol. 4, 2013.
62. N. Sasaki, A. Tagami, T. Goto, M. Taniguchi, M. Nakata, and K. Hikichi. Atomic force microscopic studies on the structure of bovine femoral cortical bone at the collagen fibril-mineral
level. Journal of Materials Science: Materials in Medicine, vol. 12, 2002, pp. 333337.
63. E. A. McNally, H. P. Schwarcz, G. A. Botton, and A. L. Arsenault. A model for the ultrastructure of bone based on electron microscopy of ion-milled sections. PLoS One, vol. 7, no. 1, 2012,
e29258.
References
387
64. P. Varga, A. Pacurenu, M. Langer, H. Suhonen, B. Hesse, Q. Grimal, P. Cloetens, K. Raum, and
F. Peyrin. Investigation of the three-dimensional orientation of mineralized collagen fibrils
in human lamellar bone using synchrotron X-ray phase nano-tomography. Acta Biomaterialia,
vol.9, 2013, pp. 81188127.
65. E. P. Katz and S.-T. Li. Structure and function of bone collagen fibrils. Journal of Molecular
Biology, vol. 50, 1973, pp. 115.
66. M. B. Schaffler and D. B. Burr. Stiffness of compact bone: Effects of porosity and density. Journal
of Biomechanics, vol. 21, no. 1, 1988, pp. 1316.
67. S. P. Kotha and N. Guzelsu. Tensile behavior of cortical bone: Dependence of organic matrix
material properties on bone mineral content. Journal of Biomechanics, vol. 40, 2007, pp. 3645.
68. M. J. Olszta, X. Cheng, S. S. Jee, R. Kumar, Y.-Y. Kim, M. J. Kaufman, E. P. Douglas, and L. B.
Gower. Bone structure and formation: A new perspective. Materials Science and Engineering R
Reports, vol. 58, 2007, pp. 77116.
69. A. Faingold, S. R. Cohen, N. Reznikov, and H. D. Wagner. Osteonal lamellae elementary units:
Lamellar microstructure, curvature and mechanical properties. Acta Biomaterialia, vol. 9, 2013,
pp. 59565962.
70. J. G. Skedros, J. L. Holmes, E. G. Vajda, and R. D. Bloebaum. Cement lines of secondary osteons
in human bone are not mineral-deficient: New data in a historical perspective. The Anatomical
Record Part, vol. 286A, no. 1, 2005, pp. 781803.
71. N. Reznikov, R. Shahar, and S. Weiner. Three-dimensional structure of human lamellar bone:
The presence of two different materials and new insights into the hierarchical organization.
Bone, vol. 59, 2014, pp. 90104.
72. D. R. Carter and W. E. Caler. Uniaxial fatigue of human cortical bone. The influence of tissue
physical characteristics. Journal of Biomechanics, vol. 14, no. 7, 1981, pp. 461470.
73. H. H. Bayraktar, E. F. Morgan, G. L. Niebur, G. E. Morris, E. K. Wong, and T. M. Keaveny.
Comparison of the elastic and yield properties of human femoral trabecular and cortical bone
tissue. Journal of Biomechanics, vol. 37, no. 1, 2004, pp. 2735.
74. P. K. Zysset, X. E. Guo, C. E. Hoffler, K. E. Moore, and S. A. Goldstein. Elastic modulus and
hardness of cortical and trabecular bone lamellae measured by nanoindentation in the human
femur. Journal of Biomechanics, vol. 32, no. 10, 1999, pp. 10051012.
75. D. T. Reilly, A. H. Burstein, and V. H. Frankel. The elastic modulus for bone. Journal of
Biomechanics, vol. 7, no. 3, 1974, pp. 271275.
76. D. T. Reilly and A. H. Burstein. The elastic and ultimate properties of compact bone tissue.
Journal of Biomechanics, vol. 8, no. 6, 1975, pp. 393405.
77. C. H. Turner, J. Rho, Y. Takano, T. Y. Tsui, and G. M. Pharr. The elastic properties of trabecular
and cortical bone tissues are similar: Results from two microscopic measurement techniques.
78. J. Y. Rho, T. Y. Tsui, and G. M. Pharr. Elastic properties of human cortical and trabecular lamellar bone measured by nanoindentation. Biomaterials, vol. 18, 1997, pp. 13251330.
79. I. V. Knets. Mechanics of biological tissues. A review. Polymer Mechanics, vol. 13, no. 3, 1977,
pp.434441.
80. W. C. Van Buskirk and R. B. Ashman. The elastic moduli of bone. Paper presented at the
Joint ASME-ASCE Applied Mechanics, Fluids Engineering, and Bioengineering Conference,
Boulder, CO, vol. 36, 1981, pp. 131143.
81. H. S. Yoon and J. L. Katz. Ultrasonic wave propagation in human cortical boneII. Measure
ments of elastic properties and microhardness. Journal of Biomechanics, vol. 9, no. 7, 1976,
pp.459464.
82. L. J. Gibson. The mechanical behavior of cancellous bone. Journal of Biomechanics, vol. 18, no. 5,
1985, pp. 317328.
83. I. H. Parkinson and N. L. Fazzalari. Characterisation of trabecular bone. Studies in
Mechanobiology, Tissue Engineering, and Biomaterials, vol. 5, 2013, pp. 3151.
84. D. P. Fyhrie and M. B. Schaffler. Failure mechanisms in human vertebral cancellous bone.
Bone,vol. 15, no. 1, 1994, pp. 105109.
388
References
85. M. R. A. Kadir, A. Syahrom, and A. chsner. Finite element analysis of idealised unit cell
cancellous structure based on morphological indices of cancellous bone. Medical and Biological
Engineering and Computing, vol. 48, 2010, pp. 497505.
86. X. E. Guo and C. H. Kim. Mechanical consequence of trabecular bone loss and its treatment: A
three-dimensional model simulation. Bone, vol. 30, no. 2, 2002, pp. 404411.
87. P. R. Townsend, R. M. Rose, and E. L. Radin. Buckling studies of single human trabeculae.
88. R. B. Ashman and J. Y. Rho. Elastic modulus of trabecular bone material. Journal of Biomechanics,
vol. 21, no. 3, 1988, pp. 177181.
89. R. Hodgskinson, J. D. Currey, and G. P. Evans. Hardness, and indicator of the mechanical competence of cancellous bone. Journal of Orthopaedic Research, vol. 7, 1989, pp. 754758.
90. J. L. Kuhn, S. A. Goldstein, K. W. Choi, M. London, L. A. Feldkamp, and L. S. Matthews.
Comparison of the trabeclar and cortical tissue moduli from human iliac crests. Journal of
Orthopaedic Research, vol. 7, 1989, pp. 876884.
91. K. Choi, J. L. Kuhn, M. J. Ciarelli, and S. A. Goldstein. The elastic moduli of human subchondral, trabecular, and cortical bone tissue and the size-dependency of cortical bone modulus.
92. J. Y. Rho, R. B. Ashman, and C. H. Turner. Youngs modulus of trabecular and cortical bone
material: Ultrasonic and microtensile measurements. Journal of Biomechanics, vol. 26, no. 2,
1993, pp. 111119.
93. Y. W. Kwon, R. E. Cooke, and C. Park. Representative unit-cell models for open-cell metal
foams with or without elastic filler. Materials Science and Engineering, vol. 343, 2003, pp. 6370.
94. E. F. Morgan, H. H. Bayraktar, and T. M. Keaveny. Trabecular bone modulus-density relationships depend on anatomical site. Journal of Biomechanics, vol. 36, no. 7, 2003, pp. 897904.
95. D. Ulrich, B. van Rietbergen, A. Laib, and P. Regsegger. The ability of three-dimensional
structural indices to reflect mechanical aspects of trabecular bone. Bone, vol. 25, no. 1, 1999,
pp.5560.
96. S. Majumdar, M. Kothari, P. Augat, D. C. Newitt, T. M. Link, J. C. Lin, T. Lang, Y. Lu, and H.K.
Genant. High-resolution magnetic resonance imaging: Three-dimensional trabecular bone
architecture and biomechanical properties. Bone, vol. 22, no. 5, 1998, pp. 445454.
97. S. Majumdar, H. K. Genant, S. Grampp, D. C. Newitt, V.-H. Truong, J. C. Lin, and A. Mathur.
Correlation of trabecular bone structure with age, bone mineral density, and osteoporotic status: In vivo studies in the distal radius using high resolution magnetic resonance imaging.
Journal of Bone and Mineral Research, vol. 12, no. 1, 1997, pp. 111118.
98. R. Krug, J. Carballido-Gamio, A. J. Burhardt, G. Kazakia, B. H. Hyun, B. Jobke, S. Banerjee,
M. Huber, T. M. Link, and S. Majumdar. Assessment of trabecular bone structure comparing
magnetic resonance imaging at 3 tesla with high-resolution peripheral quantitative computed
tomography ex vivo and in vivo. Osteoporosis International, vol. 19, no. 5, 2008, pp. 653661.
99. M. Ding and I. Hvid. Quantification of age-related changes in the structure model type and
trabecular thickness of human tibial cancellous bone. Bone, vol. 26, no. 3, 2000, pp. 291295.
100. A. J. C. Ladd, J. H. Kinney, D. L. Haupt, and S. A. Goldstein. Finite-element modeling of trabecular bone: A comparison with mechanical testing and determination of tissue modulus.
Journal of Orthopaedic Research, vol. 16, no. 5, 1998, pp. 622628.
101. J. L. Williams and J. L. Lewis. Properties and an anisotropic model of cancellous bone from
theproximal tibial epiphysis. Journal of Biomechanical Engineering, vol. 104, no. 1, 1982, pp. 5056.
102. D. R. Carter and W. C. Hayes. The compressive behavior of bone as a two-phase porous structure. Journal of Bone and Joint Surgery, vol. 59, no. 7, 1977, pp. 954962.
103. R. Hodgskinson and J. D. Currey. Youngs modulus, density and material properties in cancellous bone over a large density range. Journal of Materials Science: Materials in Medicine, vol. 3,
no.5, 1992, pp. 377381.
104. B. Liu, X. Feng, and S.-M. Zhang. The effective Youngs modulus of composites beyond the
Voigt estimation due to the Poisson effect. Composite Science and Technology, vol. 69, no. 13, 2009,
pp. 21982204.
References
389
105. M. Hahn, M. Vogel, M. Pompesius-Kempa, and G. Delling. Trabecular bone pattern factorA new
parameter for simple quantification of bone microarchitecture. Bone, vol. 13, no. 4, 1992, pp.327330.
Chapter 10

1. J. N. Reddy and A. Miravite. Practical analysis of composite laminates. Boca Raton, FL: CRC Press,
1995.
2. Y. W. Kwon. Study of fluid effects on dynamics of composite structures. Journal of Pressure
Vessel Technology, vol. 133, no. 3, 2011, 031301.
3. Y. W. Kwon, A. C. Owens, A. S. Kwon, and J. M. Didoszak. Experimental study of impact on
composite plates with fluid-structure interaction. International Journal of Multiphysics, vol. 4,
no.3, 2010, pp. 259271.
4. Y. W. Kwon, M. A. Violette, R. D. McCrillis, and J. M. Didoszak. Transient dynamic response
and failure of sandwich composite structures under impact loading with fluid structure interaction. Applied Composite Materials, vol. 19, no. 6, 2012, pp. 921940.
5. Y. W. Kwon and R. P. Conner. Low velocity impact on polymer composite plate in contact with
water. International Journal of Multiphysics, vol. 6, no. 3, 2012, pp. 179197.
6. L. E. Craugh and Y. W. Kwon. Coupled finite element and cellular automata methods for
analysis of composite structures with fluid-structure interaction. Composite Structures, vol. 102,
August 2013, pp. 124137.
7. D. J. Inman. Engineering vibration. Englewood Cliffs, NJ: Prentice Hall, 1994.
8. F. S. Tse, I. E. Morse, and R. T. Hinkle. Mechanical vibrations, theory and applications. 2nd ed.
Boston: Allyn and Bacon, 1978.
9. Y. W. Kwon, E. M. Priest, and J. H. Gordis. Investigation of vibrational characteristics of composite beams with fluid-structure interaction. Composite Structures, vol. 105, November 2013,
pp. 269278.
10. Y. W. Kwon. Dynamic responses of composite structures in contact with water while subjected
to harmonic loads. Applied Composite Materials, vol. 21, 2014, pp. 227245.
11. Y. W. Kwon and S. C. Knutton. Computational study of effect of transient fluid force on composite structures submerged in water. International Journal of Multiphysics, vol. 8, no. 4, 2014,
pp.367395.
12. X. Sun. Numerical and experimental investigation on tidal and current energy extraction. PhD
dissertation, University of Edinburgh, 2008.
13. K. D. Kimsey. Numerical simulation of hydrodynamic ram. ARBRL-TR-02217. Adelphi, MD:
US Army Ballistic Research Laboratory, February 1980.
14. D. Varas, J. Lpez-Puente, and R. Zaera. Experimental analysis of fluid-filled aluminum tubes
subjected to high-velocity impact. International Journal of Impact Engineering, vol. 36, 2009,
pp.8191.
15. K. Yang, Y. W. Kwon, C. Adams, and D. Liu. Modeling and simulation of hydrodynamic ram
for aircraft survivability. Aircraft Survivability. (in print).
Chapter 11
1. H. D. Fair. Electromagnetic propulsion: A new initiative. IEEE Transactions on Magnetics, vol.
Mag-18, no. 1, January 1982, pp. 46.
2. D. Halliday, R. Resnick, and J. Walker. Fundamentals of physics. 9th ed. New York: Wiley, 2010.
390
References
3. J. D. Powell. Thermal-energy transfer from arc to rails in an arc-driven rail gun. IEEE
Transactions on Magnetics, vol. Mag-20, no. 2, March 1984, pp. 395398.
4. S. P. Atkinson. The use of finite element analysis techniques for solving rail gun problems.
IEEE Transactions on Magnetics, vol. 25, no. 1, January 1989, pp. 5256.
5. D. Rodger and P. J. Leonard. Modelling the electromagnetic performance of moving rail gun
launchers using finite elements. IEEE Transactions on Magnetics, vol. 29, no. 1, January 1993,
pp. 496498.
6. I. Kohlberg and W. O. Coburn. A solution for the three dimensional rail gun current distribution and electromagnetic fields of a rail launcher. IEEE Transactions on Magnetics, vol. 31, no. 1,
January 1995, pp. 628633.
7. S. Barmada, A. Musolino, M. Raugi, and R. Rizzo, Numerical simulation of a complete generatorrail launch system, IEEE Transactions on Magnetics, vol. 41, no. 1, January 2005, pp. 369374.
8. O. Biro and K. Preis. On the use of magnetic vector potential in the finite element analysis of
three-dimensional eddy currents. IEEE Transactions on Magnetics, vol. 25, no. 4, July 1989,
pp. 31453159.
9. K. Muramatsu, T. Nakata, T. Takahashi, and K. Fuhiwara. Comparison of coordinate systems
for eddy current analysis in moving conductors. IEEE Transactions on Magnetics, vol. 28, no. 2,
March 1992, pp. 11861189.
10. K.-T. Hsieh. A Lagrangian formulation for mechanically, thermally coupled electromagnetic
diffusive processes with moving conductors. IEEE Transactions on Magnetics, vol. 31, no. 1,
January 1995, pp. 604609.
11. A. Nysveen and R. Nilssen. Time domain simulation of magnetic systems with a general moving geometry. IEEE Transactions on Magnetics, vol. 33, no. 2, March 1997, pp. 13941397.
12. J.-F. Lee, R. Lee, and A. Cangellaris. Time-domain finite-element method. IEEE Transactions on
Antennas and Propagation, vol. 45, no. 3, March 1997, pp. 430442.
13. C. S. Biddlecombe, J. Simkin, A. P. Jay, J. K. Sykulski, and S. Lepaul. Electromagnetic analysis
coupled to electric circuits and motion. IEEE Transactions on Magnetics, vol. 34, no. 5, September
1998, pp. 31823185.
14. D.-H. Kim, S.-Y. Hahn, I.-H. Park, and G. Cha. Computation of three-dimensional electromagnetic field including moving media by indirect boundary integral equation method. IEEE
Transactions on Magnetics, vol. 35, no. 3, May 1999, pp. 19321938.
15. T. Iwashita and M. Shimasaki. Parallel processing of 3-D eddy current analysis with moving conductor using parallelized ICCG solver with renumbering process. IEEE Transactions on
Magnetics, vol. 36, no. 4, July 2000, pp. 15041509.
16. Y. W. Kwon, N. Pratikakis, and M. R. Shellock. Finite element based multiphysics modeling of
a rail gun launcher. Journal of Multiphysics, vol. 2, no. 4, December 2008, pp. 421436.
17. Y. W. Kwon and H. Bang. The finite element method using MATLAB. 2nd ed. Boca Raton, FL: CRC
Press, 2000.
18. J. A. Greenwood and J. B. P. Williamson. Contact of nominally flat surfaces, Proceedings of the
Royal Society of London, Series A, vol. A295, 1966, pp. 300319.
19. MatWeb website. http://www.matweb.com/search/GetProperty.asp.
Chapter 12
1. J. D. Humphrey and C. A. Taylor. Intracranial and abdominal aortic aneurysms: Similarities,
differences, and need for a new class of computational models. Annual Review of Biomedical
Engineering, vol. 10, 2008, pp. 221246.
References
391
2. S. S. Raut, S. Chandra, J. Shum, and E. Finol. The role of geometric and biomechanical factors
in abdominal aortic aneurysm rupture risk assessment. Annals of Biomedical Engineering, vol. 41,
2013, pp. 14591477.
3. M. F. Fillinger, S. P. Marra, M. L. Raghavan, and F. E. Kennedy. Prediction of rupture risk
in abdominal aortic aneurysm during observation: Wall stress versus diameter. Journal of
Vascular Surgery, vol. 37, 2003, pp. 724732.
4. D. A. Vorp, M. L. Raghavan, and M. W. Webster. Mechanical wall stress in abdominal aortic aneurysm: Influence of diameter and asymmetry. Journal of Vascular Surgery, vol. 27, 1998, pp. 632639.
5. C. M. Scotti, J. Jimenez, S. C. Muluk, and E. Finol. Wall stress and flow dynamics in abdominal
aortic aneurysms: Finite element analysis vs. fluid-structure interaction. Computer Methods in
Biomechanics and Biomedical Engineering, vol. 11, 2008, pp. 301322.
6. Z. Li and C., Kleinstreuer. A comparison between different asymmetric abdominal aortic aneurysm morphologies employing computational fluidstructure interaction analysis.
European Journal of Mechanics. B, Fluids, vol. 26, 2007, pp. 615631.
7. E. Georgakarakos, C. V. Ioannou, Y. Kamarianakis, Y. Papaharilaou, T. Kostas, E. Manousaki,
and N. Katsamouris. The role of geometric parameters in the prediction of abdominal aortic aneurysm wall stress. European Journal of Vascular and Endovascular Surgery, vol. 39, 2010,
pp.4248.
8. P. Rissland, Y. Alemu, S. Einav, J. Ricotta, and D. Bluestein. Abdominal aortic aneurysm risk
of rupture: Patient-specific FSI simulations using anisotropic model. Journal of Biomechanical
Engineering, vol. 131, 2009, 031001.
9. A. Maier, M. W. Gee, C. Reeps, H.-H. Eckstein, and W. Wall. Impact of calcifications on patientspecific wall stress analysis of abdominal aortic aneurysms. Biomechanics and Modeling in
10. C. M. Scotti and E. Finol. Compliant biomechanics of abdominal aortic aneurysms: A fluid
structure interaction study. Computers and Structures, vol. 85, 2007, pp. 10971113.
11. E. S. Di Martino, A. Bohra, J. P. Vande Geest, N. Gupta, M. S. Makaroun, and D. Vorp.
Biomechanical properties of ruptured versus electively repaired abdominal aortic aneurysm
wall tissue. Journal of Vascular Surgery, vol. 43, 2006, pp. 570576.
12. M. L. Raghavan, J. Kratzberg, E. M. Castro de Tolosa, M. M. Hanaoka, P. Walker, and E. S. da
Silva. Regional distribution of wall thickness and failure properties of human abdominal aortic aneurysm. Journal of Biomechanics, vol. 39, 2006, pp. 30103016.
13. P. Vena, D. Gastaldi, L. Socci, and G. Pennati. An anisotropic model for tissue growth and
remodeling during early development of cerebral aneurysms. Computational Materials Science,
vol. 43, 2008, pp. 565577.
14. H. Schmid, A. Grytsan, E. Poshtan, P. N. Watton, and M. Itskov. Influence of differing material
properties in media and adventitia on arterial adaptationApplication to aneurysm formation and rupture. Computer Methods in Biomechanics and Biomedical Engineering, vol. 16, 2013,
pp.3353.
15. I. Chatziprodromou, A. Tricoli, D. Poulikakos, and Y. Ventikos. Haemodynamics and wall
remodelling of a growing cerebral aneurysm: A computational model. Journal of Biomechanics,
vol. 40, 2007, pp. 412426.
16. Y. Feng, S. Wada, T. Ishikawa, K. Tsubota, and T. Yamaguchi. A rule-based computational
study on the early progression of intracranial aneurysms using fluid-structure interaction:
Comparison between straight model and curved model. Journal of Biomechanical Science and
Engineering, vol. 3, 2008, pp. 124137.
17. M. Nabaei and N. Fatouraee. Computational modeling of formation of a cerebral aneurysm
under the influence of smooth muscle cell relaxation. Journal of Mechanics in Medicine and
Biology, vol. 12, 2012, 1250006.
18. J. C. Lasheras. The biomechanics of arterial aneurysms. Annual Review of Fluid Mechanics,
vol. 39, 2007, pp. 293319.
392
References
19. F. Gao, Z. Guo, M. Watanabe, and T. Matsuzawa. Loosely coupled simulation for aortic arch
model under steady and pulsatile flow. Journal of Biomechanical Science and Engineering, vol. 1,
2006, pp. 327341.
20. K. Khanafer and R. Berguer. Fluid-structure interaction analysis of turbulent pulsatile flow
within a layered aortic wall as related to aortic dissection. Journal of Biomechanics, vol. 42, 2009,
pp. 26422648.
21. J. D. Humphrey and G. A. Holzapfel. Mechanics, mechanobiology, and modeling of human
abdominal aorta and aneurysms. Journal of Biomechanics, vol. 45, 2012, pp. 805814.
22. F. G. Simsek and Y. W. Kwon. Investigation of material modeling in fluid-structure interaction
analysis of an idealized three-layered abdominal aorta: Aneurysm initiation and fully developed aneurysm. Journal of Biological Physics, doi: 10.1007/s10867-014-9372-x.
23. Z. Li and C. Kleinstreuer. Blood flow and structure interactions in a stented abdominal aortic
aneurysm model. Medical Engineering and Physics, vol. 27, 2005, pp. 369382.
24. M. L. Raghavan and D. A. Vorp. Toward a biomechanical tool to evaluate rupture potential of
abdominal aortic aneurysm: Identification of a finite strain constitutive model and evaluation
of its applicability. Journal of Biomechanics, vol. 33, 2000, pp. 475482.
25. X. Wang and X. Li. A fluid-structure interaction-based numerical investigation on the evolution of stress, strength and rupture potential of an abdominal aortic aneurysm. Computer
Methods in Biomechanics and Biomedical Engineering, vol. 16, 2013, pp. 10321039.
26. G. N. Foutrakis, H. Yonas, and R. J. Sclabassi. Saccular aneurysm formation in curved and
bifurcating arteries. American Journal of Neuroradiology, vol. 20, 1999, pp. 13091317.
27. Y. Bazilevs, M.-C. Hsu, M. D. J. Benson, S. Sankaran, A. L. Marsden. Computational
fluidstructure interaction: Methods and application to a total cavopulmonary connection.
Computational Mechanics, vol. 45, 2009, pp. 7789.
28. J. S. Wilson, S. Baek, and J. D. Humphrey. Parametric study of effects of collagen turnover on
the natural history of abdominal aortic aneurysms. Proceedings of the Royal Society Series A,
vol. 469, 2013, 20120556.
29. C. Carallo, C. Irace, M. S. De Franceschi, F. Coppoletta, R. Tiriolo, C. Scicchitano, F. Scavelli,
and A. Gnasso. The effect of aging on blood and plasma viscosity. A 11.6 years follow-up study.
Clinical Hemorheology and Microcirculation, vol. 47, 2011, pp. 6774.
30. X. Wang and X. Li. Computational simulation of aortic aneurysm using FSI method: Influence
of blood viscosity on aneurismal dynamic behaviors. Computers in Biology and Medicine, vol. 41,
2011, pp. 812821.
Index
A
AAA, see Abdominal aortic aneurysm
Abdominal aortic aneurysm (AAA), 359
already-developed, 373374
growth and remodeling of, 366
Abell-Tersoff-Brenner (A-T-B) potential,
150154
Acoustic wave equation, 4750
ambient fluid density, 47
bulk modulus, 48
condensation, 47
fluid-structure interaction, 48
Newtons 2nd law, 47
triangular element, area of, 49
Aneurysm
abdominal aortic, 359, 373
cerebral artery, 359
initiation studies, results of, 364373
investigation of, 1
AoS, see Array of structures
Array of structures (AoS), 98
Artificial recoloring process, 68
Asymptotic homogenization theory, 256
A-T-B potential, see Abell-Tersoff-Brenner
potential
Atomic model, stiffness matrix of, 195
Axial bar and beam, 1324
axial member, 1317
backward difference technique, 23
beam, 1721
boundary condition, 16
Crank-Nicholson solution technique, 22
deflection and slope, 18
diagonal mass matrices, 23
elemental stiffness matrix, 20
explicit solver, 23
finite difference method, 22
free-body diagram, 13
Galerkin finite element formulation, 14
governing equation, 18
Hookes law, 14
load intensity function, 21
moment equilibrium, 17
Newmark solution method, 24
nodal variable, 15, 19
shape functions, 18, 19
solution accuracy, 24
solution techniques, 2124

unknown reaction force, 21
weak formulation, 18
B
Bamboo shape nanotube (BSNT), 156
Bamboo-structured single-wall carbon
nanotube (BSCNT), 160
models, Youngs modulus ratio concept for,
165
natural frequencies and mode shapes for,
172173
Benchmark values (LBM scaling), 82, 88
Biomaterials, multiscale analysis of, 267295
further adjustments in models, 292295
adjustment results, 293294
bone loss results, 295
hydroxyapatite crystal, dimensions of,
293
minerals, 292
modified hierarchy, 292293
optimal adjustment results, 294295
Reuss model, 292, 294
Voigt model, 292, 294
macroscale model, 281292
cancellous bone, 288289
cancellous bone model, 290291
cancellous bone results, 291
coordinate transformation technique, 285
cortical bone, 284285
cortical bone model, 285286
cortical bone model results, 286288
finite element method, 288
Haversian systems, 284, 285
lamellar bone, 281282, 284
lamellar model, 282283
lamellar model results, 283284
misnomer, 285
osteons, 284
preferential orientation, 286
primary bone, 284
result of trabecular bone, 289290
secondary bone, 284
smooth orientation, 286
soap bubble formation, modeling of, 290
stiffness matrix, 286
tetrakaidecahedral unit cell, 290
393
394
trabeculae, 288
trabecular bone model, 289
microscale model, 272281
amino acids, 272
bone fiber, 277278
collagen network, 272
extrafibrillar mineralization, 278
fiber results, 280281
fibril bundles, 277
fibrillogenesis, 272
fibril results, 276277
gap, 273
hydroxyapatite crystals, growth of, 273
linear fibril subunit, 275
micromechanical fiber model, 278279
micromechanics fibril model, 274276
overlap, 273
tensile data, 281
three-dimensional fibril structure, 273274
twisting crystalline structure, 274
twisting fibril subunit, 275
two-dimensional fibril structure, 272273
nanoscale model, 267271
bone mineral, 267
helical spring, 269270
helical twist, 271
hydroxyapatite, 267269
nanoindentation testing on, 269
results, 270271
tropocollagen, 269
Biomechanics, multiphysics analysis of, 359374
abdominal aortic aneurysm, 359
already-developed abdominal aortic
aneurysm, 373374
hoop strains, 374
strain value, 374
von Mises stress, 374
aneurysm initiation studies, results of,
364373
AAA, growth and remodeling of, 366
blood vessel wall modeling, 364373
blood viscosity effect, 373
Mooney-Rivlin-type material, 372
radial stresses, 369
von Mises stresses, 365
cerebral artery aneurysm, 359
comparison of models with and without FSI,
362364
blood-intima interface, calculated
pressure at, 362
single-layer models, 363
von Mises stresses, 363
hyperelastic material, 360
Index
linear elastic material, 360

numerical models, description of, 360362
aneurysm initiation, 360
cases, 361
degeneration model, 361
Mooney-Rivlin type model, 360
peak systolic pressure, 361
Poisson ratio, 360
vessel wall, modeling of, 360
review of previous work, 359360
cerebral aneurysms, 359
modulus of elasticity (aneurysm), 359
numerical study, 360
Blood circulation system, 1
Boltzmann constant, 159
Bone fiber, 277278
Boron/epoxy plain weave composites, 241
Brinkman equation, 55
BSCNT, see Bamboo-structured single-wall
carbon nanotube, 160
BSNT, see Bamboo shape nanotube
Burgers equation, 55
C
CA, see Cellular automata
CAA, see Cerebral artery aneurysm
Cancellous bone, 288289
Carbon nanotubes (CNTs), 145, 151
bond energy for, 151
elastic modulus, 156168
bamboo shape nanotube, 156
basic structures of carbon nanotubes,
157158
chiral nanotubes, 158
comparative results of equilibrium and
nonequilibrium simulations, 165168
elastic moduli, difference in, 165
elastic modulus of CNTs under
equilibrium, 162165
equilibrium and vibration motion of
CNTs, 159162
freestanding thermal vibration method,
159
Maxwell-Boltzman velocity distribution,
163
multiwall carbon nanotubes, 157
simulation time step, 158
single-wall carbon nanotube, 157
thermal vibration method, 161
time evolution method, 163
Youngs modulus ratio concept, 165
Index
equilibrated structure, 151

external tensile load, 153
multiwall, 157
physical properties of, 157
single-wall, 157
vibrational mode shapes, 168173
comparison of radial breathing modes of
armchair SWCNTs, 168
discrete Fourier transform approaches,
168
hollow beam, bending of, 171
natural frequencies and mode shapes for
BSCNTs, 172173
MWCNTs, 171172
SWCNTs, 168170
radial breathing modes, 172
Tersoff-Brenner potential, 171
Youngs modulus of, 152, 159
Cauchy-Born hypothesis, 256
Cellular automata (CA), 101144, 203
Boolean quantities, 103
boundary cells, 102
boundary conditions, 116117
DAlembert solution, 116
waste of computational time, 116
wave propagation, 117
configuration, 101
convergence, 118119
point-by-point error, 118
semi-infinite fluid-filled space, 118
coupling technique, 203205
discretization and model fidelity, 117118
critical discretization versus initial
perturbation width, 118
error norms, 117
Gaussian wave, 117
initial perturbations of varying widths,
118
nonreflecting boundary conditions, 117
game of life, 103
lattice Boltzmann method, 103
modeling moving objects using, 103108
evolution rule, 105
Hamiltonian formalism, 107
internal particles, rule for time evolution
of, 104
lattice coordinates, 105
left-end particles, 105106
particle position, 104
right-end particles, 106107
395
spring constant, 105107

string momentum, 107
string velocity, 107
velocity, mass, momentum, and energy,
107108
Moore neighborhood, 101
multigrid technique, 133144
example problems, 137144
global-local technique, 134137
local domain grids, 135
memory allocation issues, 134
multigrid method, 137
nodal points, 143
revisions, 137
uniform grid, 139, 140
wave propagation, onset time for, 137
physical examples using, 108115
constrained particles, 111
DAlembert solution, 115
finite element method solutions, 114
infinite string, 115
longitudinal vibration of long uniform
rod, 109110
Newtons second law, 112
one-dimensional wave equation, 114115
physical properties of string, 113
string with force applied on the middle,
112114
string plucked at midpoint, 111112
time-scaling factor, 110
time steps, 109
transverse vibration of string, 111114
prevention of instability, 103
rule, 101
standard lattice Boltzmann method and, 55
strengths and weaknesses, 103
synchronous dynamics, 102
three-dimensional wave problem, 122126
analytical solution, 122
Cartesian radius, 123
Simpsons rule, 123
true point source, 123
two-dimensional wave problem, 119122
cellular automata in two dimensions,
119120
corner particle, rule for, 120
example of membrane vibration, 120122
underwater acoustics, application to, 126133
air-to-waterline boundary, 129
confined water channel, wave
propagation in, 133
propagation losses, 127, 128
reflected wave, 132
396
root mean square pressure, 127

transmission loss development, 127129
various boundary conditions, 129
across flat-bottom ocean floor, 129131
over curved hill, 131132
over sloping bottom, 132133
von Neumann neighborhood, 101
Central processing units (CPUs), 96, 98
Cerebral artery aneurysm (CAA), 359
CG technique, see Continuous Galerkin
technique
Chiral nanotubes, 158
Classical LBM (CLBM), 65
CLBM, see Classical LBM
CNTs, see Carbon nanotubes
Coefficients of thermal expansions (CTEs), 229
Color fluid model, 6869
Composite structures, multiphysics analysis of,
297342
fatigue loading with FSI, 318330
beam system, 320
problem description, 318320
spring-to-mass ratio, 320
fluid-structure interaction modeling,
297298
fluid flow, 297
lattice Boltzmann method, 297, 298
Navier-Stokes equation, 297
nodal variables, 298
wave equation, 297
hydrodynamic loading, 330338
base model, 331
boundary point, 332
fluid pressure, 332
model description, 330331
numerical results with constant
acceleration, 331333
numerical results with different material
properties, 336338
numerical results with free surface,
335336
numerical results with nonlinear
acceleration, 333335
hydrodynamic ram, 338342
description, 338
drag-phase fluid pressure, 342
entry wall x displacement, 339
hemispherical shock wave, 341
numerical models, 338339
Index
numerical results, 339342

low-velocity impact with FSI, 298306
comparison to experimental data,
304306
continuous Galerkin technique, 299
discontinuous Galerkin technique, 299
single-layer plate model, 299300
three-layer plate model, 302304
two-layer plate model, 300301
vibration with FSI, 306318
fast Fourier transform, 307
finite element eigenvalue analysis, 307
numerical modal analysis, 307
velocimeters, 307
verification study, 307308
Composite structures, multiscale analysis of,
207254
binding material, 207
common fibers, 207
coordinate transformation, 222224
global and local coordinate axes, 223
stiffness transformation, 222, 223
stress and strain transformations, 222
transformation matrices, 223
elastoplastic analysis of composite materials,
243246
Hollomon equation, 243
Kronecker delta, 243
Ludwik equation, 243
Newtons iteration method, 244
return mapping algorithm, 243
state matrix, 245
subcell stress increment, 243
elastoplastic analysis of composite materials,
examples of, 246254
fibrous composite plate with preexisting
crack, 251254
fibrous metal matrix composite, 250251
laminated composite plate, 254
microflaws, 251
particulate composite made of SiC and
Cu alloy, 248
particulate metal matrix composite made
of SiC and aluminum, 246248
residual stress analysis, 253
strengthening mechanisms, 247
thermal residual stresses for whisker
composite, 249250
von Mises yield criterion, 251
exchange of critical information, 207
Index
fabric model, 214222

bidirectional passage, 222
functions, 214
plain weave composite, 215219
shear components, 218
stress/strain components, 215
subcells, strain compatibility conditions
among, 216
twill composite, 219222
unit cell, 215, 220
weaving information, 214
fabric model, examples of, 239242
boron/epoxy plain weave composites, 241
coefficients of thermal expansions, 240
glass/epoxy plain weave composites, 241
thermal stresses, comparison of, 242
fiber model, 209214
constitutive equations, 212
elastoplastic deformation, 211
incremental stress and strain tensors, 210
material stiffness matrix, 211
nonelastic problems, 210
shear stresses, 210211
subcells, 209
unit cell level, thermal strain at, 214
fiber model, examples of, 225239
coefficients of thermal expansions, 229
cubic inclusion, variational asymptotic
method for, 226
delamination zone, 240
fibrous composite made of boron fibers
and aluminum matrix, 226227
fibrous composite made of graphite fiber
and epoxy matrix, 227
fibrous composite plate with a hole,
234235
fibrous composite with statistical
consideration, 237239
fibrous composite for strength, 239
filament wound cylinder, 233
Gaussian statistical distribution, 237
glass fiber/epoxy matrix fibrous
composite under tensile load, 231232
hierarchical composite, 229230
particulate composite made of Al2O3
particle/aluminum matrix, 225226
particulate composite made of SiC
particle/aluminum matrix, 226
particulate composite plate with a hole,
236
Poisson ratio, 237
397
porous material made of Al2O3, 228

rule of mixtures, 231
short-fiber composite made of SiC whiskers
and aluminum matrix, 227228
spherical inclusion, theory for, 226
thermal stress of glass fiber/epoxy matrix
fibrous composite, 229
ultrafine grain aluminum, 229
lamination model, 224225
bending stiffness properties, 224
constitutive equation, 224
numerical integration points, 224
postloop, 225
preloop, 225
matrix materials, 207
multiscale modeling of composite
structures, 207209
fabric model, 208
fiber model, 208
fiber splitting, 208
interlayer delamination, 208
lamination model, 208
matrix cracking, 208
stiffness loop, 208
strength loop, 208
transverse matrix cracking, 208
reinforcing material, 207
Constrained particles, 111
Continuous Galerkin (CG) technique, 299
Copper, behavior under conditions of cyclic
loading, 183
Cortical bone, 284285
Coulomb interaction among charged particles,
69
Coulombs frictional law, 349
Coupling techniques, 187205
example problems of coupled finite element
and atomic models, 188194
atomic arrangement, kinds of, 191
atomic array embedded in finite element
mesh with a crack, 191192
atomic behavior at crack tip, 193194
hexagonal array of atoms with
dislocation, 189190
iteration procedure, 189
nodal displacements, 192
shear loading, 190
example problems for smeared atomic
model, 196200
axial mode vibration shape, 199
bending mode vibration shape, 199
continuum model, mass matrix of, 197
398
interatomic potential, 198

natural frequencies, 197
shear mode vibration shape, 199
vibration of atoms in one-dimensional
arrangement, 196198
vibration of atoms in two-dimensional
arrangement, 198200
finite element and atomic models, 187188
continuum domain, 187
exchange of information, 187
matrix equation, 187
nodal displacement, 188
overlapped interface domain, 187
finite element and cellular automata models,
203205
coupling acoustic domains, 203204
coupling the acoustic and structural
domains, 204205
nodal variables, 204
time step size, 204
finite element and lattice Boltzmann models,
200203
coupling fluid and structural domains,
201203
element shape functions, 202
fluid stress tensor, 202
Kronecker delta, 202
LBM grid points, 200
two-way coupling problem, 201
velocity compatibility condition, 202
homogenization of atomic model into
continuum model, 194196
atomic model, stiffness matrix of, 195
discrete atoms, 195
finite element stiffness matrix, 195
matrix equation, 195
shape functions, 195
wave propagation problem, 200
C programming language, 96
CPUs, see Central processing units
CTEs, see Coefficients of thermal expansions
D
DFT, see Discrete Fourier transform
DG technique, see Discontinuous Galerkin
technique
Index
Diagonal mass matrices, 23

Dirac delta functions, 63
Discontinuous Galerkin (DG) technique, 299
Discrete Fourier transform (DFT), 168
D2Q9 lattice, 56, 60
E
Effective carbon bond length, 151
EFLBM, see Element-free-based lattice
Boltzmann method
EFM, see Extrafibrillar mineralization
Elasticity (solid element), theory of, 2830
force equilibrium, 29
isotropic material, 29
kinematic equations, 30
normal stresses, 28
shear stresses, 28
Electromechanical system, multiscale analysis
of, 343358
analysis model, 352353
armature, 352
surface conditions, 355
contact theory, 349352
asperities, 350, 351
expected total load, 351
height distribution, 352
Hertz contact solution, 350
interface parts, 350
electromagnetic theory, 345348
Coulomb gauge condition, 347
equations, 348
field variables in Lagrangian description,
346
Fouriers law, 348
integral laws, 345346
magnetic field, changes in, 348
Stokes theorem, 346
armature materials, 355
barrel, 355
electric resistance, 357
kinetic energy, 357
Lorentz forces, 355
rail gun, 353
Lorentz force, 343
mathematical models for multiphysics
analysis, 344345
electric and thermal conductivities, 245
Joules law, 344
Lorentz force, 344
Index
rail gun launchers

previous study of, 344
principle of operation of, 343
thermal stress analysis, 348349
contact theory, 348
Coulombs frictional law, 349
Euler integration scheme, 349
Hookes law, 349
Lorentz force, 349
Element-free-based lattice Boltzmann method
(EFLBM), 6465
coefficient vector, 64
interpolation function vector, 64, 65
spline function, 65
weighting function, 64
Embedded atom
method, 183
potential, 154
Euler integration scheme, 349
Explicit solver, 23
Extrafibrillar mineralization (EFM), 278, 280
F
Fabric model, 208, 214222, 239242
bidirectional passage, 222
examples of, 239242
functions, 214
plain weave composite, 215219
shear components, 218
stress/strain components, 215
subcells, strain compatibility conditions
among, 216
twill composite, 219222
unit cell, 215, 220
weaving information, 214
Fast Fourier transform (FFT), 307
Fatigue of metals, 183186
FELBM, see Finite elementbased lattice
Boltzmann method
FEM, see Finite element method
FFT, see Fast Fourier transform
Fiber model, 208, 209214
elastoplastic deformation, 211
examples of, 225239
incremental stress and strain tensors, 210
material stiffness matrix, 211
nonelastic problems, 210
shear stresses, 210211
subcells, 209
unit cell level, thermal strain at, 214
Fiber splitting, 208
399
Fibril bundles, 277

Fibrillogenesis, 272, 274
Fibrous metal matrix composites (FMMCs),
250251
Finite elementbased lattice Boltzmann method
(FELBM), 6163
Crank-Nicholson time integration technique,
63
finite element mesh, 62
Galerkin method, 63
nodal variables, 62
single-relaxation time operator, 62
unknown variable, 62
weighted residual equation, 62
Finite element method (FEM), 2, 553
acoustic wave equation, 4750
ambient fluid density, 47
bulk modulus, 48
condensation, 47
Newtons 2nd law, 47
triangular element, area of, 49
atomic models and, 187188
axial bar and beam, 1324
axial member, 1317
backward difference technique, 23
beam, 1721
deflection and slope, 18
diagonal mass matrices, 23
explicit solver, 23
free-body diagram, 13
governing equation, 18
Hookes law, 14
load intensity function, 21
moment equilibrium, 17
Newmark solution method, 24
nodal variable, 15, 19
solution accuracy, 24
solution techniques, 2124
unknown reaction force, 21
cellular automata models and, 203205
comparison of CA and, 114
cortical bone model, 288
400
fabric model, 239

fluid-structure interaction modeling, 298
frame, 2628
diagonal mass matrix, 28
element stiffness matrix, 26
global coordinate system, element matrix
in, 27
transformation matrix, 26
axial member, 14
linear shape functions, 10
nodal coordinates, 12
nodal variable, 9
shape functions, 9
strong formulation, 11
test function, 11
interaction of structure with acoustic
domain, 5053
final matrix expression, 50
fluid-solid interface, 51
fluid speed of sound, 51
fluid-structure interaction interface, 53
harmonic solutions for the rod,
assumption of, 51
incompressible fluid, 51
nodal variables, 53
one-dimensional case, 5051
submatrix, 53
two-dimensional case, 5253
vibration theory, 51
isoparametric formulation, 3641
chain rule, 39
Gauss-Legendre quadrature rule, 40
Jacobian matrix, 39
natural coordinate domain, 36
nodal variable, 38
physical coordinate domain, 36
metallic materials, multiscale analysis of,
257
method of weighted residual, 59
cubic function, 7
Galerkin method, 7
polynomial function, 6
residual, 6
test function, 7
trial function, 6, 8
weighted residual, 7
Index
plate and shell structures, 4147

advantage, 42
bending energy, 46
constitutive equation, 41, 43
elastic modulus, 44
in-plane displacements, 42
natural coordinate system, 43
Poissons ratio, 44
shell element, 42
transverse displacement, 42
transverse normal strain, 45
transverse shear strains, 43
rail gun launcher, 343, 344
solid element, 2836
discretized problem domain, 31
equations of motion, 30
finite element formulation, 3036
kinematic equations, 30, 33
nodal variable, 31
normal stresses, 28
shape functions for triangular element, 31
shear stresses, 28
test functions, 30
theory of elasticity, 2830
traction force vector, 35
triangular element, 32
structural domain, 204
transverse vibration of string, 114
truss, 2426
coordinate systems, 24, 25
displacements in coordinates, 24
element, 25
global coordinate system, 24
local coordinate system, 24
strain energy, 25
structure, 25
Fluid-structure interaction (FSI), 201
fatigue loading with, 318330
beam system, 320
problem description, 318320
spring-to-mass ratio, 320
hybrid lattice Boltzmann formulation, 68
interface, 53
401
Index
low-velocity impact with, 298306

comparison to experimental data, 304306
continuous Galerkin technique, 299
discontinuous Galerkin technique, 299
single-layer plate model, 299300
three-layer plate model, 302304
two-layer plate model, 300301
modeling, composite structures, 297298
vibration with, 306318
fast Fourier transform, 307
finite element eigenvalue analysis, 307
numerical modal analysis, 307
velocimeters, 307
verification study, 307308
FMMCs, see Fibrous metal matrix composites
Fouriers law, 348
Frame, 2628
global coordinate system, element matrix
in, 27
transformation matrix, 26
Free-energy model, 69
FSI, see Fluid-structure interaction
G
linear shape functions, 10
nodal coordinates, 12
nodal variable, 9
shape functions, 9
strong formulation, 11
test function, 11
Game of life, 103
Gear predictor-corrector method, 147
Generalized lattice Boltzmann equation, 60
Glass/epoxy plain weave composites, 241
GPUs, see Graphics processing units
Graphics processing units (GPUs), 96100
array of structures, 98
basic implementation on graphics processing
units, 9799
central processing units, 96
coalesced memory access, 99
collision operator, 96
computational requirements, 9697
data layout, 9899
graphics processing units, 96

LBM routine, 97
Message Passing Interface, 96
NVIDIA CUDA GPU computing platform,
96
performance benchmark, 99100
structure of arrays, 98
Visualization Toolkit file formats, 97
H
HA, see Hydroxyapatite
Hamiltonian dynamics, 147
Hamiltonian formalism, 107
Haversian systems, 284
HE material, see Hyperelastic material
Hertz contact solution, 350
HLBM, see Hybrid LBM
Hollomon equation, 243
Hookes law, 14, 349
HRAM, see Hydrodynamic ram
Hybrid LBM (HLBM), 6568
Hydrodynamic loading, 330338
Hydrodynamic ram (HRAM), 338342
description, 338
drag-phase fluid pressure, 342
entry wall x displacement, 339
hemispherical shock wave, 341
numerical models, 338339
numerical results, 339342
Hydroxyapatite (HA), 267269, 294
chemical formula of, 268
crystal
dimensions of, 293
lattice, 268
description of, 267
nanoindentation testing on, 269
Hyperelastic (HE) material, 360
I
Interatomic potential energy function, 149154
Abell-Tersoff-Brenner potential, 150154
bond energy, 151
effective carbon bond length, 151
embedded atom potential, 154
equilibrated carbon nanotube structure, 151
kinetic molecular theory, 152
Lennard-Jones potential, 149
Morse potential, 150
short-range interaction potential, 153
402
Interlayer delamination, 208

Interparticle potential model, 6970
Isoparametric formulation, 3641
chain rule, 39
Jacobian matrix, 39
natural coordinate domain, 36
nodal variable, 38
physical coordinate domain, 36
J
Jacobian matrix, 39
Joules law, 344
K
Kinetic molecular theory, 152, 159
Korteweg-de Vries equation, 55
Kronecker delta, 202, 243
L
Lamellar bone, 281282
Laminated plate theory, 224
Lamination model, 208, 224225
bending stiffness properties, 224
constitutive equation, 224
numerical integration points, 224
postloop, 225
preloop, 225
Lattice Boltzmann equation, generalized, 60
Lattice Boltzmann method (LBM), 2, 55100, 200
bounce-back boundary conditions,
on-grid version of, 73
cumbersome data, 78
D2Q9 lattice, 74
dry nodes, 73
expressions requiring intensive
bookkeeping, 78
fixed rigid boundary, 7273
nodal variable, 71
periodic boundary, 72
pressure boundary condition, 7374
required bookkeeping, 78
velocity boundary condition, 7479
Brinkman equation, 55
Burgers equation, 55
cellular automata, 103
Index
coupling between FEM and, 200

element-free-based lattice Boltzmann
method, 6465
coefficient vector, 64
interpolation function vector, 64, 65
spline function, 65
backward-facing step, 8788
benchmark stipulation, 89
channel flow over cylinder, 9296
halfway bounce-back boundary
condition, 86
lattice units, fluid viscosity in, 87
lid-driven cavity, 8892
on-grid bounce-back technique, 89
Poiseuille flow, 8587
Reynolds number, 94
Strouhal number, 94
trailing vortex, 92, 94
Von Karman street, 94
vorticity plot for cylinder, 95
finite elementbased lattice Boltzmann
method, 6163
Crank-Nicholson time integration
technique, 63
Galerkin method, 63
nodal variables, 62
single-relaxation time operator, 62
unknown variable, 62
weighted residual equation, 62
graphics processing units, implementation
on, 96100
array of structures, 98
basic implementation on graphics
processing units, 9799
central processing units, 96
coalesced memory access, 99
computational requirements, 9697
data layout, 9899
LBM routine, 97
Message Passing Interface, 96
NVIDIA CUDA GPU computing
platform, 96
performance benchmark, 99100
structure of arrays, 98
403
Index
hybrid lattice Boltzmann formulation, 6568

classical LBM, 65
dispersion errors, 66
dissipation error, 66
first-order time integration schemes, 66
hybrid LBM, 65
subdomain, 66
time and space integration methods, 66
Korteweg-de Vries equation, 55
macroscale conservation laws, 55
multicomponent flow, 6871
artificial recoloring process, 68
color fluid model, 6869
computational fluid dynamics methods, 69
Coulomb interaction among charged
particles, 69
free-energy model, 69
immiscible multicomponent lattice
Boltzmann procedures, 7071
interparticle potential model, 6970
LBM theory, models using, 68
mean field theory model, 69
nonequilibrium thermodynamics,
Cahn-Hilliards approach for, 69
Rayleigh-Taylor instability, simulation
of, 69
recoloring step, 68
surface tension effect, 69
Vlasov equation, 69
multiple relaxation lattice Boltzmann
formulation, 6061
D2Q9 lattice, 60
fixed Prandtl number, 60
generalized lattice Boltzmann equation, 60
Navier-Stokes equations, 55
scaling, 8185
benchmark values, 82
inlet boundary condition, 82
lattice units, 81
LBM time step flowchart, 84
pressure boundary condition scaling,
8385
Reynolds number, 81
velocity boundary condition scaling, 83
viscosity scaling, 83
Schrodinger equation, 55
standard lattice Boltzmann method, 5560
cellular automaton, 55
collision process, 60
computational procedure, 60
discrete velocity vector, 56

D2Q9 lattice structure, 56, 57
lattice definition, 56
lattice site update, rule for, 56
redistribution process, 60
streaming process, 60
turbulent flow, 7980
eddy viscosity, 79
Prandtl mixing length approach, 80
relaxation constant, 79
von Karman constant, 80
Visualization Toolkit file formats, 97
wave equation, 8081
refraction index, 81
wave attenuation factor, 81
wave propagation speed, 81
LBM, see Lattice Boltzmann method
LE material, see Linear elastic material
Linear elastic (LE) material, 360
Linear shape functions, 10, 11
LJ potential method, 171, 175
Lorentz force, 343, 349, 355
Ludwik equation, 243
M
Maxwell-Boltzman velocity distribution, 163
MD, see Molecular dynamics
Mean field theory model, 69
Message Passing Interface (MPI), 96
Metallic materials, multiscale analysis of,
255265
deformation state, 261
discrete atomic region, 264
elastic modulus, 261
grain boundary stiffness, 262
interface spring constants, 260, 262
no-defect case, 263
Poisson ratio, 261
smeared continuum region, 265
macroscale analysis, 257258
displacements, 258
information obtained, 257
mesoscale analysis, 258260
grain boundaries, 259
unit cell, 258
Voronoi diagram, 258
microscale analysis, 260
404
nanoscale analysis, 260261

inner subdomain, 260
polycrystalline metals, 256257
coupling, 257
intergrain boundary characteristics, 256
single-grain domain, 256
previous study, 255256
asymptotic homogenization theory, 256
atomistic modeling, 256
Cauchy-Born hypothesis, 256
enhanced mesh superposition method,
256
grain deformation, 256
linking of models, 256
Metals, fatigue of, 183186
Method of weighted residual (MWR), 59
choices, 63
cubic function, 7
Galerkin method, 7
polynomial function, 6
residual, 6
test function, 7
trial function, 6, 8
weighted residual, 7
Molecular dynamics (MD), 145186
carbon nanotubes (elastic modulus),
156168
bamboo shape nanotube, 156
basic structures of carbon nanotubes,
157158
chiral nanotubes, 158
comparative results of equilibrium and
nonequilibrium simulations, 165168
elastic moduli, difference in, 165
elastic modulus of CNTs under
equilibrium, 162165
equilibrium and vibration motion of
CNTs, 159162
freestanding thermal vibration method,
159
Maxwell-Boltzman velocity distribution,
163
multiwall carbon nanotubes, 157
simulation time step, 158
single-wall carbon nanotube, 157
thermal vibration method, 161
time evolution method, 163
Youngs modulus ratio concept, 165
Index
carbon nanotubes (vibrational mode shapes),

168173
comparison of radial breathing modes of
armchair SWCNTs, 168
discrete Fourier transform approaches, 168
hollow beam, bending of, 171
BSCNTs, 172173
MWCNTs, 171172
SWCNTs, 168170
radial breathing modes, 172
Tersoff-Brenner potential, 171
classical formulation, 145146
atoms, interactive forces among, 145
Hamiltonian, 145
fatigue of metals, 183186
box boundary, atoms moving outside of,
183
copper, behavior under conditions of
cyclic loading, 183
embedded atom method, 183
examples, 184186
impurity models, 185
modeling, 183
output data, 183
simulation units, 183
interatomic potential energy function, 149154
Abell-Tersoff-Brenner potential, 150154
bond energy, 151
effective carbon bond length, 151
embedded atom potential, 154
equilibrated carbon nanotube structure, 151
kinetic molecular theory, 152
short-range interaction potential, 153
molecular mechanics formulation, 154156
displacement vectors, 154
external loads, 154
matrix expression, 156
nanoscale fluid flow, 179183
fluid flow simulation, 179
geometry used, 179
particle-fluid interaction, 180183
velocity profiles, 180, 182
polymers, 173178
allocation of polymers, 174175
autocorrelation, 176
405
Index
composite material, simulation

representing, 178
cross-linking of polymers, 173174
examples, 176178
LJ potential method, 175
potential energy for polymers, 175176
quasi-atom concept, 175
randomly dispersed particles, 173
time integration technique, 146149
correction, 148149
evaluation, 148
Gear predictor-corrector method, 147
Hamiltonian dynamics, 147
prediction step, 148
Mooney-Rivlin type model, 360
Moore neighborhood, 101
MPI, see Message Passing Interface
Multiphysics and multiscale modeling,
introduction to, 13
book organization, 23
computational methods, 2
cellular automata, 2
overview, 12
aneurysms, investigation of, 1
blood circulation system, 1
living systems in nature, 1
man-made systems, 1
Multiwall carbon nanotubes (MWCNTs), 157,
171172
MWCNTs, see Multiwall carbon nanotubes
MWR, see Method of weighted residual
N
Nanoropes (SWCNT), 157
Nanoscale fluid flow, 179183
fluid flow simulation, 179
geometry used, 179
particle-fluid interaction, 180183
velocity profiles, 180, 182
Natural coordinate domain, 36
Navier-Stokes equations, 55, 297
Newmark method (FEM), 24
Nodal variable, 9
axial member, 15
beam, 19
FELBM, 62
fictitious particle densities as, 298

isoparametric formulation, 38
LBM boundary condition, 71
structural domain (FEM), 204
Nonequilibrium thermodynamics,
Cahn-Hilliards approach for, 69
NVIDIA CUDA GPU computing platform, 96
O
On-grid bounce-back technique, 89
Osteons, 284
P
Particulate metal matrix composite (PMMC),
246
Physical coordinate domain, 36
Plate and shell structures, 4147
advantage, 42
bending energy, 46
constitutive equation, 41, 43
elastic modulus, 44
in-plane displacements, 42
Poissons ratio, 44
shell element, 42
transverse displacement, 42
transverse normal strain, 45
transverse shear strains, 43
PMMC, see Particulate metal matrix composite
Polycrystalline metals, 256257
Polymers, 173178
allocation of, 174175
autocorrelation, 176
composite material, simulation representing,
178
cross-linking of, 173174
examples, 176178
LJ potential method, 175
potential energy for, 175176
quasi-atom concept, 175
randomly dispersed particles, 173
Q
Quasi-atom concept, 175
406
R
Rail gun launcher, 343
Rayleigh-Taylor instability, simulation of, 69
Return mapping algorithm (RMA), 243
Reuss model, 294
RMA, see Return mapping algorithm
RMS pressure, see Root mean square pressure
ROM, see Rule of mixtures
Root mean square (RMS) pressure, 127
Rule of mixtures (ROM), 231
S
Schrodinger equation, 55
Shape functions
atomic model, 195
beam element, 18
boundary integral, 35
continuum model, 195
coupling fluid and structural domains, 202
derivatives, 15, 39
displacements, 33
FELBM, 62
finite element nodal displacements, 188
Galerkin method, 9, 63
isoparametric formulation, 36, 37
linear, 10, 11
Simpsons rule, 123
Single-wall carbon nanotube (SWCNT), 157,
168170
SoA, see Structure of arrays
Solid element, 2836
discretized problem domain, 31
equations of motion, 30
kinematic equation, 33
nodal variable, 31
shape functions for triangular element,
31
test functions, 30
traction force vector, 35
theory of elasticity (solid element), 2830
Index
normal stresses, 28
shear stresses, 28
Stokes theorem, 346
String
infinite, 115
momentum of, 107
physical properties of, 113
transverse vibration of, 111114
velocity, 107
vibration, modeling of, 103
Strouhal number, 94
Structure of arrays (SoA), 98
SWCNT, see Single-wall carbon nanotube
T
TEM, see Transmission electron microscope
Test function, 7
Theory of elasticity (solid element), 2830
normal stresses, 28
shear stresses, 28
Time-scaling factor, 110
Transmission electron microscope (TEM), 159
Transverse matrix cracking, 208
Trial function, 6
Tropocollagen, 267, 269, 271
Truss, 2426
coordinate systems, 24, 25
displacements in coordinates, 24
element, 25
global coordinate system, 24
local coordinate system, 24
strain energy, 25
structure, 25
U
UFG aluminum, see Ultrafine grain aluminum
Ultrafine grain (UFG) aluminum, 229
Underwater acoustics, 126133
air-to-waterline boundary, 129
confined water channel, wave propagation
in, 133
propagation losses, 127, 128
reflected wave, 132
root mean square pressure, 127
transmission loss development, 127129
various boundary conditions, 129
407
Index
wave propagation across flat-bottom ocean

floor, 129131
wave propagation over curved hill, 131132
wave propagation over sloping bottom, 132133
V
Velocimeters, 307
Visualization Toolkit (VTK) file formats, 97
Vlasov equation, 69
Voigt model, 294
Von Karman street, 94
von Mises yield criterion, 251
von Neumann neighbors, 101
Voronoi diagram, 258
VTK file formats, see Visualization Toolkit
fileformats
W
Wave propagation
across flat-bottom ocean floor, 129131
CA rule for, 118
one-dimensional, 117
over curved hill, 131132
over sloping bottom, 132133
problem, 200
speed, 81
underwater acoustics, 126
Weighted residual, 7
Y
Youngs modulus (CNTs), 152, 159

Bz5xn Multiphysics Anfd Multiscale Modeling Techniques and Applications

Diunggah oleh

Informasi Dokumen

Judul Asli

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

Bz5xn Multiphysics Anfd Multiscale Modeling Techniques and Applications

Diunggah oleh

Hak Cipta:

Format Tersedia

MULTIPHYSICS

TECHNIQUES AND APPLICATIONS

Boca Raton London New York

CRC Press is an imprint of the

2016 by Taylor & Francis Group, LLC

Multiscale and Multiphysics Modeling

1.3Organization of This Book

2016 by Taylor & Francis Group, LLC

2016 by Taylor & Francis Group, LLC

2.2Method of Weighted Residual

u(0) = 0 and u(1) = 1

This is a simple two-point boundary value problem. The exact solution is

Multiscale and Multiphysics Modeling

(x) = a0 + a1x + a2x2 (2.4)

where ai (i = 0, 1, 2) is the coefficients to be determined, and denotes an approximate

(x) = (1 a2)x + a2x2 (2.7)

R(x) = 2a2 (1 a2)x a2x2 (2.8)

2016 by Taylor & Francis Group, LLC

Finite Element Method

w(x)R(x) dx = w(x){2a (1 a )x a x } dx = 0 (2.9)

Inserting Equation 2.10 into Equation 2.9 gives

( x + x ){2a (1 a )x a x }dx = 0 (2.11)

(x) = b0 + b1x + b2x2 + b3x3 (2.13)

Applying the boundary conditions to this expression yields

2016 by Taylor & Francis Group, LLC

Multiscale and Multiphysics Modeling

Elimination of two variables from the equations gives

Then, the trial function is expressed as

R(x) = 2b2 + (1 + b2 + 7b3)x b2x2 b3x3 (2.19)

The resulting two weighted residual expressions are

2016 by Taylor & Francis Group, LLC

Finite Element Method

2.3Galerkin Finite Element Formulation

(xi) = a + bxi = ui (2.25)

(xj) = a + bxj = uj (2.26)

2016 by Taylor & Francis Group, LLC

Multiscale and Multiphysics Modeling

Solving these two equations gives

Substitution of Equations 2.27 and 2.28 into Equation 2.24 yields

The shape functions are expressed as

Hi(xi) = 1, Hi(xj) = 0, Hj(xi) = 0, Hj(xj) = 1,

2016 by Taylor & Francis Group, LLC

Finite Element Method

Now, the weighted residual formulation becomes

where e denotes an element domain. For the present example,

Substitution of Equations 2.332.39 into Equation 2.37 gives

2016 by Taylor & Francis Group, LLC

Multiscale and Multiphysics Modeling

for the first element and

for the first element and

2016 by Taylor & Francis Group, LLC

Finite Element Method

2.4Axial Bar and Beam

2016 by Taylor & Francis Group, LLC

Multiscale and Multiphysics Modeling

If the elastic modulus is constant along the bar, it is simplified as

where the speed of sound in the material is

u( x , t) = H i ( x)ui (t) + H j ( x)u j (t) = H i ( x)

2016 by Taylor & Francis Group, LLC

Finite Element Method

Evaluation of the matrices in Equation 2.58 results in the following expression:

2016 by Taylor & Francis Group, LLC

Multiscale and Multiphysics Modeling

1.3Organization of This Book

2.2Method of Weighted Residual

2.3Galerkin Finite Element Formulation

2.4Axial Bar and Beam

2.5Truss and Frame