3331158

NORTHWESTERN UNIVERSITY
The Extended Finite Element Method for Special Problems with Moving
Interfaces
A DISSERTATION
SUBMITTED TO THE GRADUATE SCHOOL
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
for the degree
DOCTOR OF PHILOSOPHY
Field of Applied Mathematics
By
Bryan G. Smith
EVANSTON, ILLINOIS
December 2008
3331158

3331158

2009
Copyright 2008 by
Smith, Bryan G.

All rights reserved
2
c Copyright by Bryan G. Smith 2008
All Rights Reserved
3
ABSTRACT
The Extended Finite Element Method for Special Problems with Moving Interfaces
Bryan G. Smith
Problems involving irregularly shaped domains or embedded interfaces occur in a wide
range of mathematical models. These include models in computational biology, uid
mechanics, and solidication. Numerical solutions for this type of problem can be dicult
to obtain, and therefore a variety of methods have been devised to solve them. One method
that has been shown to perform well in comparison with other methods is the eXtended
Finite Element Method (X-FEM).
In this thesis, a variety of modications are made to the X-FEM in order to apply
it to problems with special interfacial features. First, a custom enrichment function is
shown to increase the accuracy of the X-FEM for a problem containing a boundary layer
at an internal interface. This technique is then applied to the problem of biolm growth
in order to explore the relationship between the shape of a biolm colony and the prole
of its growth. Next, the X-FEM is applied to lm thinning in foam dynamics, and we
present a method for solving the Stokes equations while accounting for surface tension
4
along a free surface. Finally, we extend the model for lm thinning to include a solidica-
tion front. The process of foam solidication is challenging to model computationally due
to the presence of a triple junction where the solidication front intersects the gas-liquid
interface. We propose two procedures for determining the subgrid location of the triple
junction and demonstrate how to properly prescribe boundary conditions along each in-
terface. This method is then applied to a simplied model for foam solidication in order
to generate some preliminary results.
5
Acknowledgements
I am not the only one to be pleased that this dissertation is complete. Many people
helped me get to this moment, and I would like to take this opportunity to thank them
in the only section of my dissertation that they will probably read.
First, to my advisor, Dave Chopp, and my committee, Steve Davis, and Sasha Golovin,
who guided me in the research and writing of this dissertation. Your guidance and support
were essential to the formulation of the ideas in this dissertation. I hope that you are
pleased with our end result.
Second, to my colleagues Ben Vaughan and Brian Merkey, who shared my oce and
became my friends. Our discussions of our work sparked some of my best ideas and kept
me thinking even when I was frustrated with this project.
Third, to the US Department of Defense and the ladies of the ARCS Foundation,
whose nancial support made all of this possible.
Fourth, to my parents, Bill and Cyrilla Smith, and my brother Andy, who have been
waiting for this day just as long as I have. You pushed me to do my best and to work
hard. You always believed in me and encouraged me, telling me that I could do anything
even nish school. I hope that I have made you proud.
Finally, to my wife, Jane, for your endless patience and support as I slowly nished
writing this dissertation. I couldnt have done it without you. I hope youre as happy as
I am that its nished.
6
Table of Contents
ABSTRACT 3
Acknowledgements 5
List of Tables 8
List of Figures 10
Chapter 1. Introduction 13
Chapter 2. The Extended Finite Element Method 17
2.1. Enrichments 18
2.2. Interface Approximation 24
2.3. Construction of Element Matrices and Vectors 27
2.4. Boundary Conditions 29
2.5. Assembly and Solution of the Global System 36
2.6. Level Set Coupling and Evaluation of Interface Derivatives 37
2.7. Summary 38
Chapter 3. Convergence Studies of the X-FEM 39
3.1. Elliptic Examples 39
3.2. Convergence with a Specialized Enrichment Function 47
7
Chapter 4. Biolm Growth 56
4.1. Mathematical Model 56
4.2. Tip Splitting 58
4.3. Shadowing 62
Chapter 5. Stokes Flow and Film Thinning 66
5.1. Stokes Flow Around a Solid 66
5.2. Stokes Flow With a Free Surface 70
5.3. Film Thinning in Foam Dynamics 73
Chapter 6. Triple Points and Foam Solidication 84
6.1. The Triple Point 84
6.2. Interface Conditions 87
6.3. Freezing in a Simple Geometry 88
6.4. Foam Solidication 91
Chapter 7. Conclusions 100
Bibliography 103
8
List of Tables
3.1 Domain results for Ex. 1. 41
3.2 Interface results for Ex. 1. 41
3.3 Interface derivative results for Ex. 1. 42
3.4 System sizes for Ex. 1. 43
3.11 Domain results for Ex. 4 with Neumann jump interface conditions. 50
3.12 Interface results for Ex. 4 with Neumann Jump interface conditions. 50
3.13 Domain results for Ex. 4 with Dirichlet interface conditions. 51
3.14 Interfacial gradient results for Ex. 4 with Dirichlet interface conditions. 51
3.15 Domain results for Ex. 5 with Neumann jump interface conditions 53
3.16 Interface results for Ex. 5 with Neumann jump interface conditions. 54
3.17 Interface gradient results for Ex. 5 with Dirichlet interface conditions. 54
9
3.18 Domain results for Ex. 5 with Dirichlet interface conditions. 54
10
List of Figures
2.1 The 1-D Linear Isoparametric basis function. 19
2.2 The Heaviside Enrichment. 20
2.3 The Absolute Value Enrichment. 21
2.4 The Exponential Enrichment. 22
2.5 Enrichment Strategies. 24
2.6 An X-FEM mesh with a void. 25
2.7 Valid Interface Congurations. 26
2.8 Element Partitioning. 30
2.9 Lagrange Multiplier Mesh. 33
3.1 Solution of Example 1. 40
3.6 Cross sections generated using various mesh sizes. 55
4.1 The domain and governing equations for the biolm problem. 57
11
4.2 Representative velocity proles. 59
4.3 The eect of aspect ratio on mushrooming. 60
4.4 The eect of radius of curvature on mushrooming. 61
4.5 Eect of domain period and lm thickness on mushrooming. 62
4.6 Normalized substrate penetration depths. 62
4.7 Natural growth rates for selected colonies. 63
4.8 The eect of period length on shadowing. 64
4.9 The 50% shadowing region. 64
4.10 The 97% shadowing region. 65
5.1 Flow past a cylinder: the computational domain. 69
5.2 Flow past a cylinder: Horizontal Velocities. 69
5.3 Flow past a cylinder: Vertical Velocities. 69
5.4 Flow past a cylinder: Pressure Field. 70
5.5 Flow past a cylinder: Streamlines. 70
5.6 The lm thinning domain. 74
5.7 Sample Initial Condition. 76
5.8 Interface shapes of an evolving lm. 77
5.9 Interface curvature of an evolving lm. 78
5.10 Film Height vs. Time. 78
5.11 Sample Initial Conditions: Varying initial curvatures. 79
12
5.12 The eect of initial curvature conguration on lm evolution. 79
5.13 Sample Initial Conditions: Varying initial lm heights. 80
5.14 The eect of initial lm height on lm evolution. 80
5.15 Sample Initial and Final Conditions: Varying volume fractions. 81
5.16 The eect of volume fraction on lm evolution. 82
5.17 Sample Evolving Interface Shapes: Varying boundary conditions. 83
5.18 The eect of boundary conditions on lm evolution. 83
6.1 The solidication domain. 85
6.2 Determining the Triple Point Location: Strategy 1. 86
6.3 Determining the Triple Point Location: Strategy 2. 86
6.4 Triple Point Partitioning. 88
6.5 Results of simulating freezing in one dimension. 90
6.6 Film congurations at t = 0 and t = t
final
. 97
6.7 Time evolution of the lm height in the solidication problem. 98
6.8 Final lm congurations when applying extreme undercooling. 98
6.9 Time evolution under extreme undercooling. 99
6.10 Comparison of interfacial curvatures. 99
13
CHAPTER 1
Introduction
Problems involving irregularly shaped domains or embedded interfaces occur in a
wide range of mathematical models. These include models in computational biology,
uid mechanics, and solidication. Numerical solutions for this type of problem can be
dicult to obtain, and therefore a variety of methods have been devised to solve them.
A common technique for solving problems with embedded interfaces or irregular do-
mains is the construction of a conforming mesh. This allows the problem to be dened and
solved in a straighforward manner using the Finite Element Method. Numerous available
numerical libraries make this method particularly attractive, and as such it has been used
in a multitude of applications [7, 8, 9, 10, 11, 12, 27, 28]. Unfortunately, the need for
a conforming mesh can be a major drawback in some cases. Mesh construction is often
time consuming for dicult domain shapes, and poorly constructed meshes can produce
ill conditioned approximations. This is an even more serious drawback for problems with
evolving interfaces, as the domain must be remeshed as the geometry changes.
In response to these drawbacks, a number of methods have been developed to solve
problems with irregular domains on a regular Cartesian mesh. An early method using
this approach is the Immersed Boundary Method (IB) [39]. Here, a formulation is used
in a thin band around the interface to solve a variable coecient Poisson equation. While
it is simple to implement, the Immersed Boundary Method always gives a continuous
14
approximation even if the solution is discontinuous. Further details can be found in
[41, 42, 43].
Another method for solving elliptic equations on a domain with an internal interface is
the Immersed Interface Method (IIM) [33]. The IIM achieves global second-order conver-
gence by modifying the standard nite dierence stencil near the interface. The stencil is
modied by inserting the specied interfacial jump conditions into the Taylor expansion
at a point on the interface. This allows the method to capture discontinuities in the so-
lution and its derivatives. However, while this method exhibits second-order convergence
globally, the convergence at the interface is only rst order. This can create diculties
when interfacial derivatives must be computed, such as when a velocity potential is used
to compute the velocity at an interface. More information can be found in [33, 34, 35].
A third method that has shown promise for solving these types of problems is the Ghost
Fluid Method (GFM). The GFM was introduced in [23] to handle contact discontinuities
at an interface when solving the inviscid Euler equations. It has since been generalized and
used to model shocks and deagrations [24], solidication [25], uid-structure interaction
[22], and Poissons equation [26, 36]. This method is easy to implement and extends to
three dimensions, but like the IIM, it is typically only rst order at the interface.
One method that has been shown to perform well in comparison with other methods is
the eXtended Finite Element Method (X-FEM) [51]. The X-FEM is a Partition of Unity
Method [37] in which enrichment functions are added to the conventional nite element
function space in order to capture discontinuous behavior at an internal interface. It was
introduced in [4] to solve crack propagation problems in linear elastic fracture mechanics
[17, 19, 38, 49, 50]. It was extended to solve crack propagation problems involving
15
frictional contact [20] and cohesive cracks [52], and was rst utilized in combination with
the Level Set Method in [48]. It has since been used for a variety of applications, including
solidication [21], multiphase ow [14], and biolm growth [46]. What distinguishes the
X-FEM from conventional nite element methods is the ability to account for an internal
interface or irregularly shaped external domain without the need for a conforming mesh;
both the location of the interface and the local solution behavior are embedded in the
enrichment functions. This allows the method to solve moving interface problems without
the need for remeshing as well as to capture large solution gradients near an interface
without the need for extremely ne grids [46].
In this work, we demonstrate the versatility of the X-FEM by applying it to three
problems with unique interfacial characteristics. We rst consider the problem of bacterial
biolm growth. Biolms are microbial communities that grow on solid surfaces and are
one of the most ubiquitous forms of life on the planet [13]. In this case, we use the X-FEM
to solve the linearized biolm equations to investigate the relationship between the shape
of a biolm colony and the prole of its growth. The form of these equations leads to the
development of large gradients in the velocity potential near the biolm surface, and a
new enrichment function is introduced to accurately capture these gradients and compute
the growth rate.
The second application under consideration is lm thinning in foam dynamics. A foam
is composed of gas bubbles within a liquid, and these bubbles are separated by thin lms
connected by junctions. For foams with small liquid fractions, the lifetime of the foam is
determined by the thinning and rupture of the lms, and an understanding of this process
is necessary to develop a large-scale model for foam evolution [6]. We use the X-FEM to
16
model the drainage of a lm within a periodic array of gas bubbles for several initial and
boundary conditions. We then present a method for solving the Stokes equations while
accounting for surface tension along a free surface.
For the nal application, we extend the model for lm thinning to include a solidica-
tion front. The process of lm solidication is potentially important in the manufacture
of solid metallic foams. This process is challenging to model computationally due to the
presence of a triple junction where the solidication front intersects the gas-liquid inter-
face. A method for capturing the motion of a triple point was presented in [47], and a
modication of this method is used here in conjunction with the X-FEM. We propose two
procedures for determining the location of a triple junction at the subgrid level and also
demonstrate how to properly apply boundary conditions at each of the three interfaces.
Finally, we apply this method to a simplied model for foam solidication and present
some preliminary results.
17
CHAPTER 2
The Extended Finite Element Method
The Extended Finite Element Method (X-FEM) is a variation of the standard nite
element method that encodes information about an internal interface into the nite ele-
mement function space in order to eliminate the need for a conforming mesh. This infor-
mation is included in the function space by using extra basis functions, called enrichment
functions, at carefully selected, enriched, nodes. The ability to model solution behav-
ior near an interface without a conforming mesh is particularly useful in moving interface
problems as it eliminates the potentially expensive step of mesh regeneration only the
placement of the enriched nodes changes as the interface evolves.
Consider the equation:
(2.1) (u) +u = f,
on a two-dimensional domain, , containing an interface, . The coecients , , and f
may be discontinuous across , leading to discontinuities in the solution, u, as well as its
normal derivative. The X-FEM approximation for the solution u is written as
(2.2) u
h
(x, y) =
n
i
N
i
(x, y) u
i
+
n
j
N
E
j
(x, y)
k
() a
j
,
where N is the set of all nodes in the domain, N
E
is the set of enriched nodes, n
i
and n
j
are
the i
th
and j
th
nodes of their respective sets, is a standard nite element test function,
k
18
is the enrichment function, and is the signed distance function representing the interface.
The variables u
i
and a
j
are the coecients for the unenriched and enriched degrees of
freedom respectively. Note that the rst summation, representing the contribution from
the unenriched degrees of freedom, does not depend on the location of the interface, and
therefore remains unchanged as the interface moves.
2.1. Enrichments
In order to account for an internal interface, enrichment functions are added to the
standard nite element function space, appearing as the second sum in (2.2). An enrich-
ment can be constructed as a function only of the signed distance from the interface:
= min
X
||x X|| ,
where the sign depends on whether x is inside or outside the region enclosed by the
interface . The fact that the enrichments depend only on the signed distance to the
interface allows the X-FEM to be easily coupled to the Level Set Method [40].
2.1.1. Enrichment Functions
The enrichment functions added to the standard nite element approximation in the X-
FEM can serve two purposes. First, they encode the location of the interface into the
function space itself, which allows for the application of both Dirichlet and Neumann
interface conditions without the need for a conforming mesh. Second, it is possible to
include information about the known asymptotic behavior of the solution near the inter-
face, allowing the method to accurately capture boundary layer behavior without the need
19
for a very ne mesh. Therefore, the ideal enrichment function for a given application is
dependent on both the type of solution behavior expected near the interface and the types
of boundary conditions to be applied. A few examples of enrichment functions are given
below. For comparison, the value and derivative of a 1-D isoparametric basis function is
given in gure 2.1.
-1 -0.5 0 0.5 1
0
0.5
1
x
(
x
)
-1 -0.5 0 0.5 1
-1
-0.5
0
0.5
1
x
'
(
x
)
(a) (b)
d
dx
Figure 2.1. The value, (a), and derivative, (b), of a 1-D linear isoparametric
basis function. The node is located at x = 0 with neighboring nodes at
x = 1. Note that the basis function is continuous on its support but has
a jump in the derivative at the node.
The simplest and most exible enrichment function is the Heaviside enrichment [17,
49]:
Heaviside
() = H () .
This generic function can be used to solve problems with a jump in the solution value, the
normal derivative, or both. Because the enrichment itself does not specify a jump in the
solution or its derivative, two separate conditions must be applied in order to maintain
20
well-posedness. A second, related discontinuous enrichment is the step enrichment [38]:
Step
() = H () H () .
This enrichment has many of the same properties as the Heaviside enrichment but it
cannot be used as easily when the the domain contains void regions (Section 2.1.3).
-1 -0.5 0 0.5 1
0
0.5
1
x
(
x
)

(
x
)
-1 -0.5 0 0.5 1
-1
-0.5
0
0.5
1
x
(

(
x
)

(
x
)

)
'
(a)
Heaviside
(b)
d
dx

Heaviside ( )
Figure 2.2. The value, (a), and derivative, (b), of the Heaviside enrichment
function multiplied by the 1-D linear isoparametric basis function. The
interface is located at x = 0.5 with > 0 for x < 0.5. Here both the
function and its derivative contain jumps at the interface.
For solutions which are continuous across the interface but feature a jump in the
normal derivative, a continuous enrichment function can be employed. Because these
functions are continuous across the interface, only one interfactial constraint is necessary.
21
Two commonly used continuous enrichments are the ramp enrichment [19]:
Ramp
() =
_
_
0 > 0,
0,
and the absolute value enrichment [5]:
AbsV al
() =
_
_
> 0,
0.
It is important to note that in each case, the enrichment decays to zero at the interface.
This is an important criterion for stability when a Dirichlet interface condition is applied.
-1 -0.5 0 0.5 1
0
0.25
0.5
x
(
x
)

(
x
)
-1 -0.5 0 0.5 1
-1.5
-1
-0.5
0
0.5
1
1.5
x
(

(
x
)

(
x
)

)
'
(a)
AbsVal
(b)
d
dx

AbsVal ( )
Figure 2.3. The value, (a), and derivative, (b), of the absolute value enrich-
ment function multiplied by the 1-D linear isoparametric basis function.
The interface is located at x = 0.5 with > 0 for x < 0.5. The function
is continuous on the entire domain, with jumps in the derivative at the in-
terface and the node. Note that this enrichment increases the order of the
approximation near the interface.
22
Finally, we present one example of an enrichment function that is tailored to the
solution behavior near the interface. The exponential enrichment function, dened as:
exp
() =
_
_
1 e
> 0,
0,
can be used for problems which display exponential behavior within an interfacial bound-
ary layer [46]. The enrichment contains a parameter
1
which is the thickness of the
boundary layer. This thickness is problem dependent and is generally determined by as-
ymptotic analysis. Here we present the exponential enrichment as a continuous function,
but it can be adapted to be discontinuous if required.
-1 -0.5 0 0.5 1
0
0.25
0.5
0.75
1
x
(
x
)

(
x
)
-1 -0.5 0 0.5 1
-1
-0.5
0
0.5
1
x
(

(
x
)

(
x
)

)
'
(a)
Exponential
(b)
d
dx

Exponential ( )
Figure 2.4. The value, (a), and derivative, (b), of the exponential enrich-
ment function multiplied by the 1-D linear isoparametric basis function.
The interface is located at x = 0.5 with > 0 for x < 0.5. Once again,
the function is continuous with derivative jumps at the interface and the
node. Note that while the function is O(1), the derivative near the interface
becomes very large. This allows the enrichment to capture boundary layers
without the need for a ne mesh.
23
2.1.2. Nodal Enrichment
Once a suitable enrichment function has been selected, a method for selecting nodes for
enrichment is needed. The most common approach is topological enrichment, where a
node is enriched if the interface cuts across its support. This method is most eective
when using enrichments that do not increase the order of the nite element approxima-
tion, such as the Heaviside or step enrichments. Another approach, which has been used
to model both crack propogation [3] and biolm growth [46], is geometric enrichment.
With this method, all nodes within a specic signed distance of the interface are enriched.
This is generally done when using enrichment functions that do increase the order of the
nite element approximation, such as the absolute value enrichment or the exponential
enrichment, and serves two purposes. First, if there is a boundary layer, this ensures
that the enriched region captures the entire layer. Second, it allows higher-order enriched
approximations to blend with the standard approximation farther away from the inter-
face. This preserves the accuracy of the solution near the interface without the need for
complicated blending elements.
2.1.3. Voids
In many problems, a solution is only required over a portion of the domain. For example,
when solving for the uid ow around a solid, only the solution within the liquid is
needed - there is no ow within the solid. To handle these types of voids within the
domain, a Heaviside enrichment function is applied topologically to the interface that
denes the void. This enrichment evaluates to zero inside the void and one in the desired
computational domain. This is similar to the technique described in [17, 49]. However,
24
Topologically Enriched Node
Geometrically Enriched Node
Figure 2.5. Topologically enriched nodes have their support cut by the in-
terface. Geometrically enriched nodes lie within a specied signed distance
of the interface.
in contrast to that method, all of the unenriched degrees of freedom (DOFs) in both
the enriched region and the voided region are removed. Therefore, each node contains at
most one degree of freedom - nodes near the interface have a single enriched DOF, nodes
entirely in the computational domain have a single unenriched DOF, and nodes entirely
within the void have no DOFs. The resulting mesh for the ow around a cylinder is
shown in Figure 2.6. The eect of this technique is to generate an approximation that
looks like a standard nite element approximation but is only nonzero in a specic region
of the domain.
2.2. Interface Approximation
Due to its ability to handle evolving interfaces accurately and eciently, the interior
interface, , is tracked using the Level Set Method [40]. With this method, the location
of the interface is stored implicitly as the zero contour of a signed distance function.
25
Unenriched DOFs
Enriched DOFs
Removed DOFs
Figure 2.6. An X-FEM mesh with a void. The region within the circle is
a void. All unenriched DOFs within the circle are removed, as well as all
unenriched DOFs at nodes containing enriched DOFs.
However, in order to partition elements for integration and to apply interfacial boundary
conditions, the X-FEM requires an explicit representation for the interface. As with every
other aspect of nite elements, the construction of this representation must be done on
an element by element basis.
The rst step towards this end is determining whether or not an interface segment
exists within a given element. Consider a rectangular element with a signed distance
funtion, , dened on the four corners. If has the same sign at each corner, the
interface cannot pass through the element.
If the element does contain an interface segment, the sides containing the entry and
exit points can be determined by the sign of at the corners. A side contains the
entry point of the interface if is positive at the clockwise endpoint and negative at
the counterclockwise endpoint. Similarly, if is negative at the clockwise endpoint and
positive at the counterclockwise endpoint, the side contains the exit point. If more than
26
one side contains an entry or exit point, the conguration is considered invalid. Figure
2.7 shows the various valid interface congurations. After the interface conguration is
determined, the entry and exit points are found by locating the zeros of the signed distance
function along the appropriate interface boundaries using a bisection method.
+
+ +
-
+
- -
-
-
- +
-
+
- +
+
+
+ -
-
+
- -
+
North-East South-East North-South
North-West South-West East-West
Figure 2.7. Valid interface congurations. The signs at the corners corre-
spond to the sign of the level set function at the nodes. In each case, the
same conguration (with the opposite direction) is obtained by changing
the signs at all of the corners.
Once the entry point, x
0
, and exit point, x
1
, have been found, the interface is
approximated as a line segment between these two points with the parameterization:
I
: x(s) = ms +b for 1 x 1, with:
x(s) =
_
_
x(s)
y(s)
_
_
, m =
_
_
1
2
(x
1
x
0
)
1
2
(y
1
y
0
)
_
_
, and b =
_
_
1
2
(x
1
+x
0
)
1
2
(y
1
+y
0
)
_
_
.
27
The interfacial unit normal vector needed to compute normal derivatives can be found
from this parameterization and is dened as:
n =
1
_
m
2
x
+m
2
y
_
_
m
y
m
x
_
_
.
2.3. Construction of Element Matrices and Vectors
To create the global system matrix used to nd a solution, individual element matrices
are constructed using the weak formulation of the governing equation. These matrices
are all block matrices of the form:
A =
_
_
A
uu
A
ua
A
au
A
aa
_
_
,
where A
uu
is the standard nite element matrix, and A
ua
, A
au
, and A
aa
arise from the
inclusion of the enriched degrees of freedom. These terms are only present when an
element contains enriched degrees of freedom, therefore they do not signicantly increase
the size of the global system. The global system vector is similarly constructed from
individual element matrices that have the form:
v =
_
_
v
u
v
a
_
_
,
where v
u
is the standard nite element vector and v
a
comes from the enrichments.
28
2.3.1. Integral Forms of Common Terms
There are several terms which commonly appear in nite element calculations, including
Laplacian terms, mass terms, force terms, and gradient terms. For the Laplacian term,
(u), the element matrices in Cartesian coordinates are:
K
uu
i,j
=
_
E
[
i

j
]
E
,
K
ua
i,j
= K
au
j,i
=
_
E
[
i
(
j
j
)]
E
,
K
aa
i,j
=
_
E
[(
i
i
) (
j
j
)]
E
.
For the mass term, u, the matrices in Cartesian coordinates are:
M
uu
i,j
=
_
j

E
,
M
ua
i,j
= M
au
j,i
=
_
j

E
,
M
aa
i,j
=
_
j

E
.
For the gradient term, v
x
k
u/x
k
, the matrices in Cartesian coordinates are:
_
G
uu
x
k
_
i,j
=
_
E
v
x
k
j
x
k
E
,
_
G
ua
x
k
_
i,j
=
_
E
v
x
k
j
x
k
E
,
_
G
au
x
k
_
i,j
=
_
E
v
x
k
x
k
(
j
j
)
E
,
_
G
aa
x
k
_
i,j
=
_
E
v
x
k
x
k
(
j
j
)
E
.
29
Finally, for the force operator, f, the vectors in Cartesian coordinates are:
f
u
i
=
_
i
f (x, y)
E
,
f
a
i
=
_
i
f (x, y)
E
.
2.3.2. Element Integration
Due to the fact that many of the element integrals contain discontinuous terms, care must
be taken to ensure the accuracy of the numerical quadrature. Gauss-Legendre quadrature
can generate large errors when integrating even simple discontinuous functions, and these
errors do not necessarily decrease as the number of integration points is increased. In order
to eliminate this error, elements containing interface segments must be partitioned into
subelements that contain only continuous functions. Unlike the method employed in [20],
here the elements are subdivided into both rectangles and triangles. After the element has
been partitioned, the rectangles are partitioned using standard 2-D Gaussian quadrature
while the triangles are converted to rectangles for integration with the method described
in [44]. Note that this partitioning is not the same as meshing around the interface as no
new elements or degrees of freedom are created.
2.4. Boundary Conditions
In order for a problem to be well posed, boundary conditions must be applied both
at the internal interface and at the domain boundaries. There are several methods for
applying these conditions depending on the type of condition to be imposed as well as on
30
Interface
Partition Boundary
Figure 2.8. An element containing an interface segment is partitioned into

triangles and rectangles using the linear approximation of the interface.
the type of boundary. These methods include weakly enforced line source terms, Lagrange
multipliers, and direct application.
2.4.1. Line Sources
Due to the integration by parts that occurs when constructing the weak form of the
Laplacian, it is frequently convenient to apply Neumann or Robin conditions weakly
using a line source term. Consider rst the equation:
(2.3)
2
u = 0,
with a jump in the normal derivative across an interface
I
:
__
u
n
__
= v (x, y) .
31
This condition can be enforced by adding a line source term to the right hand side of
equation (2.3) of the form:
(2.4)
_
I
v (x, y) (x X (S)) (y Y (S))
I
,
where X (S) and Y (S) are the parameterized coordinates of the interface. The weak form
of this source term is added as the vector terms:
u
i
=
_
i
v (x, y)
I
,
a
i
= (0)
_
i
v (x, y)
I
.
This idea can be extended to general Robin boundary conditions of the form:
u
n
= w(x, y) u.
This type of condition can be enforced by replacing the term:
_
i
u
n

I
,
from the weak form with terms from the weak form of the boundary condition. These
include the matrix terms:
(Z
uu
I
)
i,j
=
_
j

I
,
(Z
ua
I
)
i,j
= (Z
au
I
)
j,i
=
_
j

I
,
(Z
aa
I
)
i,j
=
_
i

j
j

I
,
32
and the vector terms:
u
i
=
_
i
w(x, y)
I
,
a
i
=
_
i
w(x, y)
I
.
Here, the term
(0) indicates that the enrichment function is evaluated according to

the side of the interface where the condition is being applied. Note that this technique
can be used to apply pure Neumann conditions using = 0.
2.4.2. Lagrange Multipliers
Another common technique for applying boundary conditions within the nite element
method is the use of Lagrange multipliers. Unlike the line source method, Lagrange
multipliers add degrees of freedom to the system, but multipliers are the only method for
enforcing all types of interface conditions. Consider again the PDE:
(2.5)
2
u = 0,
this time with a Dirichlet boundary condition on :
u = g (x, y) .
In order to enforce this condition, equation (2.5) is rewritten including the boundary
condition as a restraint:
2
u +B() = 0,
B(u) = g (x, y) .
33
Here, B is the constraint operator (a Dirichlet condition in this case), g (x, y) is the value
of the constraint, and is the Lagrange multiplier.
To add Lagrange multipliers to the system, the rst step is the construction of a one
dimensional mesh along the interface. Due to the rst order interface approximation
provided by the level set, piecewise linear segments are placed in each element containing
the interface. An example is shown in Figure 2.9.
Figure 2.9. A lagrange multiplier mesh overlayed on a circular interface
within an 11 11 nite element mesh.
34
The Lagrange multiplier, , is then approximated using one dimensional nite ele-
ments:
=
m
i
M
i
,
where the
i
are the one dimensional basis functions and
i
is the value of the multiplier
on each segment [29].
The matrices generated by this approximation vary according to the type of boundary
condition being applied. For a Dirichlet condition, the Lagrange multiplier matrix is:
C
D
=
_
_
C
u
D
C
a
D
_
_
,
with:
(C
u
D
)
i,j
=
_
j

I
,
(C
a
D
)
i,j
=
_
i

j

I
,
where
i
corresponds to the side of the interface where the condition is being applied. A
similar construction is used to apply a Dirichlet jump condition, with:
(C
u
DJ
)
i,j
= 0,
(C
a
DJ
)
i,j
=
_
i
[[
i
]]
j

I
.
To apply conditions on the normal derivative, Lagrange multipliers can be used in con-
junction with line sources. This can signicantly increase the accuracy of the solution
when using an enrichment with a nonzero derivative. To apply a boundary condition on
35
the normal derivative, the Lagrange multiplier matrix is:
(C
u
N
)
i,j
=
_
i
n

j

I
,
(C
a
N
)
i,j
=
_
n
_
i
_
j

I
.
Similarly, the matrix used to apply a jump in the normal derivative is:
(C
u
NJ
)
i,j
= 0,
(C
a
NJ
)
i,j
=
_
I
__

n
(
i
i
)
__
j

I
.
In all cases, the vector term that applies the value of the constraint is the same:
g
i
=
_
I
g (x, y)
I
.
Robin boundary conditions can be applied using a combination of the Dirichlet and Neu-
mann Lagrange operators. The condition u + n u = g (x, y) is enforced using
C = C
D
+C
N
.
2.4.3. Domain Boundaries
Domain boundary conditions are enforced in a very similar way to interfacial boundary
conditions. Neumann and Robin conditions are applied using line sources in the same
manner. Dirichlet conditions, however, are somewhat dierent. First, when applying a
homogeneous Dirichlet condition at a boundary, all degrees of freedom at that boundary
are set to zero. For a nonhomogeneous boundary condition, Lagrange multipliers are used.
If the boundary is not intersected by an internal interface, the Lagrange multipliers are
36
constructed in the manner described above. However, if an internal interface is present,
care must be taken to avoid large errors at the point of intersection. In order to apply the
domain boundary condition properly, two Lagrange elements are placed in the element
that contains the intersection - one on each side of the internal interface. This allows the
proper boundary condition to be applied on each side of the interface and eliminates the
error associated with improper integration across a discontinuity.
2.5. Assembly and Solution of the Global System
The global system is assembled in the same manner as the element matrices, with a
block structure as a result of the enrichments and the Lagrange multipliers. The linear
system has the form:
(2.6)
_
_
A
uu
A
ua
C
u
A
au
A
aa
C
a
(C
u
)
T
(C
a
)
T
0
_
_
_
_
u
a
_
=
_
_
b
u
b
a
g
_
_
,
where the A
uu
, A
au
, A
ua
, A
aa
, b
u
, and b
a
are the X-FEM matrices resulting from the
element by element approximation of the PDE, and C
u
, C
a
, and g are the Lagrange
multipliers used to enforce the boundary constraints.
Once the element matrices have been assembled into the global system given in (2.6),
a method for solving the linear system must be selected. Unfortunately, for certain mesh
congurations, this system can be poorly conditioned, making the use of an iterative solver
dicult. Eorts to alleviate this diculty through preconditioning have been made [3],
but here the system is solved using a parallel sparse LU decomposition [45].
37
2.6. Level Set Coupling and Evaluation of Interface Derivatives
In later chapters, a variety of problems are presented, each involving the evolution of
an interface over time. In each case, a set of elliptic equations is solved using the X-FEM
at each timestep to determine the velocity eld that drives the motion of the interface.
This result is then fed into the level set method and the interface is marched forward in
time. The updated level set is fed back to the X-FEM to be used at the next time step. In
some cases, the velocity eld is a direct result of the X-FEM calculation, but sometimes
the interfacial velocity must be derived from the elliptic solution. In particular, many
problems involve the solution for a velocity potential, so the velocity at the interface is
the jump in the normal derivative of the resulting solution at the interface. Therefore, it
is necessary to develop an accurate method for approximating these derivatives.
In order to accurately determine the derivative jump across the interface, a domain
integral method, similar to the one described in [20], is employed. We rst consider a
section of interface,
i
, within the support,
i
, of a test function,
i
. Equation (2.1) is
multiplied by
i
and integrated over
i
. Integrating by parts yields:
(2.7)
_
i
(u
i
+u
i
f
i
) d
i
=
_
i
[[ n u]]
i
d
i
.
Using a rst order approximation, the derivative jump is assumed to be constant along
i
. Equation (2.7) can be then be rewritten as:
(2.8) [[ n u]]
i
=
1
_
i
d
i
_
i
(u
i
+u
i
f
i
) d
i
.
38
To then determine the jump at a single point on the interface, x
d
, the jump across the
interface segment is determined using each test function,
i
, which has support containing
x
d
. In order to avoid oscillations due to roundo error, interface segments are discarded
if the integral of
i
over the segment is less than 10
13
. These values are then weighted
appropriately and summed:
[[ n u]]
x
d
=
j
[[ n u]]
j
(x
d
) .
2.7. Summary
In this Chapter we have given the procedure for implementing the X-FEM in two
dimensions. There are three essential decisions to be made when implementing the X-
FEM for a specic problem. First, the type of enrichment function must be chosen. The
choice of enrichment function depends on the type of interface conditions as well as the
expected solution behavior near the interface. Once the enrichment function has been
decided, the method of enrichment can be determined. For most problems, topological
enrichment is appropriate. However, for problems that include boundary layers or require
continuous enrichment functions, geometric enrichment is preferable. The nal decision
is choosing a technique for enforcing interface conditions. This depends primarily on the
type of condition. Neumann or Robin conditions are typically enforced weakly through the
addition of a forcing vector, and Dirichlet conditions are enforced by Lagrange multipliers.
Once these choices have been made, constructing the stiness matrix and solving the linear
system is relatively straightforward.
39
CHAPTER 3
Convergence Studies of the X-FEM
In order to test the convergence of the X-FEM, the method is applied to a collection
of test problems for which exact solutions are available. The rst three examples are the
elliptic problems proposed in [33], and the others are chosen to highlight the eectiveness
of exponential enrichment function discussed in 2.1.
For each problem the interface
I
is the circle
_
x
1
2
_
2
+
_
y
1
2
_
2
=
1
16
within the
square domain 0 x, y 1.
3.1. Elliptic Examples
3.1.1. Example 1: Singular Sources
The rst example includes a singular source placed along the interface
I
. The dierential
equation to be solved is:
(3.1)
2
u =
_
I
(r R
I
)
I
,
where is the Dirac Delta function, and R
I
is 1/4. The solution is continuous, with
[[u]] = 0, and has a jump in the normal derivative at the interface due to the line source:
__
u
n
__
= 4.
40
The exact solution to ((3.1)) is:
u(x, y) =
_
_
1, r
1
4
1 + log (4r) , r >
1
4
.
Figure 3.1. Solution of Example 1.
The problem is solved using 4-node bilinear nite elements with both step and ramp
enrichments. In both cases, the jump in the normal derivative is applied with a line source
term, and Lagrange multipliers are used to enforce the continuity condition when using
the step enrichment. Table 3.1 gives the convergence results for both types of enrichment.
The error given is the maximum error at the nodes:
||T
n
||
= max
i
n
i
N
{
u(x
i
, y
i
) u
h
i
},
where N is the set of all nodes, (x
i
, y
i
) is the location of the i
th
node, and u
h
i
is the
computed value at the node. The ratios of successive errors are included as well to show
convergence rates.
41
n
Step Absolute Value
||T
n
||
ratio ||T
n
||
ratio
19 3.8397 10
3
7.8138 10
3
39 9.3782 10
4
4.094 3.9577 10
3
1.974
79 2.3034 10
4
4.072 1.9029 10
3
2.080
159 6.4061 10
5
3.596 9.3797 10
4
2.029
319 1.5619 10
5
4.102 4.7646 10
4
1.969
Table 3.1. Domain results for Example 1.
These results show that the X-FEM has rst order convergence using ramp enrichments
and second order convergence with step enrichments. It is important to note that the
choice of enrichment function does inuence the overall accuracy and convergence of the
X-FEM.
Of greater importance for moving interface problems is the accuracy of the solution at
the interface. This can be problematic for some methods even though they exhibit good
accuracy away from the interface [51]. Table 3.2 shows the error interpolated along the
interface. This error is computed at 1,000 evenly spaced points along the parameterized
interface. The results are similar to those shown in Table 3.1, but the dierence in
accuracy between the step and ramp enrichments is even more pronounced.
n
Step Absolute Value
||T
n
||
ratio ||T
n
||
ratio
19 5.1857 10
3
2.1871 10
2
39 1.2444 10
3
4.167 1.1708 10
2
1.868
79 3.0043 10
4
4.142 6.0996 10
3
1.948
159 8.8146 10
5
3.408 3.1101 10
3
1.961
319 1.9315 10
5
4.564 1.6142 10
3
1.927
Table 3.2. Interface results for Example 1.
42
Table 3.3 shows the convergence of the normal derivative along the interface. As stated
in Section 2.6, this quantity is important when the speed of the interface is derived from
a velocity potential. Due to the rst order global convergence of the X-FEM with ramp
enrichments, the error in the normal derivative when using the ramps does not converge.
When using step enrichments, however, the method exhibits rst order convergence.
n
Step Absolute Value
||T
n
||
ratio ||T
n
||
ratio
19 4.1828 10
1
1.8292 10
0
39 1.6067 10
1
2.603 1.6479 10
0
1.110
79 9.3826 10
2
1.712 1.3096 10
0
1.258
159 4.5301 10
2
2.071 1.4733 10
0
0.889
319 2.2290 10
2
2.032 1.3818 10
0
1.066
Table 3.3. Interface derivative results for Example 1.
One aspect of performance that has not yet been discussed is the increase in the size of
the global linear system as a result of adding enrichments and Lagrange multipliers. Table
3.4 gives the linear system size as well as its sparsity for dierent versions of the X-FEM.
This shows that in terms of the total number of degrees of freedom, the addition of the
enrichments to the standard nite element approximation has only a small computational
cost ( <2% for a 319319 grid). Due to the fact that the step enrichments are signicantly
more accurate than the ramp enrichments and add very little to the system size, the
remaining elliptic examples will only consider the convergence of the X-FEM with step
enrichments.
43
n
FEM (No Enrichments) X-FEM (Step Enrichments) X-FEM (Ramp Enrichments)
System Size % Sparse System Size % Sparse System Size % Sparse
19 400 1.87000% 520 2.07396% 480 2.15625%
39 1,600 0.51375% 1,840 0.54277% 1,760 0.55191%
79 6,400 0.13445% 6,880 0.13840% 6,720 0.13940%
159 25,000 0.03438% 26,560 0.03490% 26,240 0.03501%
319 102,400 0.00869% 104,320 0.00876% 103,680 0.00877%
Table 3.4. System sizes for Example 1.
3.1.2. Example 2: Discontinuous Coecients
The second elliptic example adds a discontinuous coecient to a PDE with a singular
source term:
(3.2) (u) = f +C
_
I
(x X (s))
I
,
where:
f (x, y) = 16
_
r
2
+ 1
_
,
(x, y) =
_
_
4r
2
+ 1, r
1
4
,
b, r >
1
4
,
,
with the jump conditions:
[[u]] = 0,
__
u
n
__
= 4C.
The exact solution to (3.2) is:
u(x, y) =
_
_
4r
2
, r
1
4
,
1
4
_
1
9
8b
_
+
4r
2
b
(2r
2
+ 1) +
C
b
log (4r) , r >
1
4
,
44
with C = 0.1 and b = 10.
As in the rst example, the jump in the normal derivative is enforced using the line
source, and continuity is enforced with Lagrange multipliers. Table 3.5 shows the conver-
gence on the nodes. As in the rst example, the X-FEM with step enrichments achieves
second order convergence. Tables 3.6 and 3.7 show that the convergence of the solution
and its derivative at the interface remain the same as well.
n
Step
||T
n
||
ratio
19 1.7613 10
3
39 4.1771 10
4
4.217
79 1.0289 10
4
4.060
159 3.0164 10
5
3.411
319 6.7960 10
6
4.439
45
n
Step
||T
n
||
ratio
19 1.6517 10
3
39 3.3824 10
4
4.883
79 8.2238 10
5
4.113
159 3.1568 10
5
2.605
319 7.4612 10
6
4.231
n
Step
||T
n
||
ratio
19 2.7307 10
1
39 1.2776 10
1
2.137
79 6.1203 10
2
2.088
159 4.8216 10
2
1.269
319 2.4790 10
2
1.945
3.1.3. Example 3: Discontinuous Solution
The nal elliptic example has no source term or discontinuous coecient but contains a
jump in the solution at the interface
I
:
(3.3)
2
u = 0.
The interfacial jump conditions are
[[u]] = e
x
cos y,
__
u
n
__
= 4e
x
( x cos y y sin y) ,
where x = x
1
2
, and y = y
1
2
.
46
The exact solution to (3.3) is:
u(x, y) =
_
_
e
x
cos y, r
1
4
,
0, r >
1
4
.
In order to apply the condition on the normal derivative at the interface, a line source
term is added to the governing equation so that equation (3.3) becomes
(3.4)
2
u =
_
I
4e
x
( x cos y y sin y) (x X (s))
I
.
The Dirichlet jump is then enforced with Lagrange multipliers.
The convergence results for Example 3 are given in Tables 3.8, 3.9, and 3.10. The
conclusions are the same as in the previous examples with second order convergence for
the value of the solution on the nodes and the interface and rst order convergence for
the normal derivative.
47
n
Step
||T
n
||
ratio
19 1.7648 10
4
39 6.0109 10
5
2.936
79 1.7769 10
5
3.383
159 4.8626 10
6
3.654
319 1.2362 10
6
3.934
n
Step
||T
n
||
ratio
19 4.7842 10
4
39 1.0659 10
4
4.488
79 2.8361 10
5
3.758
159 7.3603 10
6
3.853
319 2.0634 10
6
3.567
n
Step
||T
n
||
ratio
19 5.6520 10
2
39 2.4190 10
2
2.337
79 9.4512 10
3
2.560
159 7.1671 10
3
1.319
319 2.6865 10
3
2.668
3.2. Convergence with a Specialized Enrichment Function
In this section, dierent types of enrichment functions are applied to the solution of
two example problems. In each problem, exponential enrichment functions are compared
to both absolute value enrichments and step enrichments when nding a solution that
contains a boundary layer.
48
We solve each example problem using both types of interface conditions discussed
in Section 2.2. These two types of boundary conditions demonstrate how the custom
enrichment function can improve the accuracy of both the solution and its derivative at
the interface.
3.2.1. Example 4: Enriching with the Exact Solution
The rst example shows the improvement in the nite element approximation when the
exact solution is included in the enriched function space. The dierential equation is
(3.5)
2
u +

_
1

1
r
_
u = 0,
where
(x, y) =
_
_
0 r
1
4
,
1 r >
1
4
.
The exact solution to (3.5) is
u(x, y) =
_
_
1 r
1
4
,
exp
_
1
_
1
4
r
_
r >
1
4
,
where is
1
200
.
The expression
1
4
r is equivalent to the level set variable , so the exact solution on
the exterior of the circle is contained within the exponential enrichment function.
49
(a)
0 0.1 0.2 0.3 0.4 0.5
0
0.2
0.4
0.6
0.8
1
r
(b)
Numerical
Exact
Figure 3.4. Solution of Example 4. (a) A surface plot of the solution on
the entire domain. (b) A cross section of the solution, taken at an angle of
one radian. The numerical approximation overlaps the exact solution and
is computed using the exponential enrichment function on a 799799 grid.
Equation (3.5) is solved with two sets of boundary conditions. The rst prescribes
continuity and sets a derivative jump condition on :
[[u]] = 0,
__
u
n
__
=
1
.
Due to the fact that the derivative jump along the interface is already known, the im-
portant simulation results to consider are the solution values along the interface and
throughout the domain.
Table 3.11 shows the convergence results for the X-FEM using three types of enrich-
ment functions: the step enrichment, absolute value enrichment, and exponential enrich-
ment. As Table 3.11 shows, the exponential enrichment is more accurate than both the
absolute value and step enrichments, and it converges more uniformly as well.
50
n
Step Absolute Value Exponential
||T
n
||
ratio ||T
n
||
ratio ||T
n
||
ratio
49 1.3367 10
1
2.8126 10
1
3.6922 10
2
99 7.8584 10
2
1.701 4.9647 10
2
5.665 8.1788 10
3
4.514
199 2.6213 10
2
2.998 9.8545 10
3
5.038 1.8534 10
3
4.413
399 8.5907 10
3
3.051 3.5610 10
3
2.760 4.6404 10
4
3.994
799 2.9094 10
3
2.953 9.7458 10
4
3.663 1.1565 10
4
4.013
Table 3.11. Domain results for Example 4 with Neumann jump interface conditions.
Table 3.12 gives the errors on the interface. The results are similar to those seen in Ta-
ble 3.11, but the disparity between the exponential enrichment and the other enrichments
is more signicant.
n
||T
n
||
ratio ||T
n
||
ratio ||T
n
||
ratio
49 3.4331 10
1
4.5436 10
1
3.6860 10
2
99 1.4444 10
1
2.377 1.3363 10
1
3.400 9.1137 10
3
4.045
199 3.9447 10
2
3.662 4.5511 10
2
2.936 1.9652 10
3
4.638
399 1.1170 10
2
3.531 1.4101 10
2
3.228 5.3682 10
4
3.661
799 3.1200 10
3
3.580 3.6471 10
3
3.866 1.3277 10
4
4.043
Table 3.12. Interface results for Example 4 with Neumann Jump interface conditions.
The second boundary condition we consider is a Dirichlet condition on both sides of
the interface:
u = 1.
The exact solution remains the same as above. Here, the value at the interface is known,
so the important computational results are the solution values within the domain and the
jump in the derivative across the interface.
Table 3.13 contains the convergence results for the X-FEM using this boundary condi-
tion. Once again, the asymptotic enrichments perform much better than the other two, by
51
more than an order of magnitude. Table 3.14 shows the gradient results on the interface.
Due to the magnitude of the gradient jump ( 10
2
), this table presents relative errors
rather than absolute errors:
||T
n
||
=
1
1/
max
i
n
i
N
{
u(x
i
, y
i
) u
h
i
}.
n
||T
n
||
ratio ||T
n
||
ratio ||T
n
||
ratio
49 2.5623 10
2
1.9174 10
1
1.1620 10
2
99 2.2492 10
2
1.135 6.9352 10
2
2.765 4.4235 10
3
2.627
199 1.4385 10
2
1.564 2.1398 10
2
3.241 8.7237 10
4
5.071
399 6.0928 10
3
2.361 5.9632 10
3
3.588 3.1170 10
4
2.799
799 1.9458 10
3
3.131 1.5402 10
3
3.872 6.0066 10
5
5.189
Table 3.13. Domain results for Example 4 with Dirichlet interface conditions.
n
||T
n
||
ratio ||T
n
||
ratio ||T
n
||
ratio
49 4.2738 10
1
2.2663 10
1
3.1433 10
2
99 1.9041 10
1
2.245 9.1789 10
2
2.469 1.5427 10
2
2.038
199 6.8870 10
2
2.765 3.4742 10
2
2.642 3.8923 10
3
3.964
399 3.1636 10
2
2.177 1.7591 10
2
1.975 1.4946 10
3
2.604
799 1.3817 10
2
2.290 8.5566 10
3
2.056 5.5810 10
4
2.678
Table 3.14. Interfacial gradient results for Example 4 with Dirichlet inter-
face conditions.
3.2.2. Example 5: Enriching with an Asymptotic Solution
In the second example, the custom enrichment function does not contain the exact solu-
tion, but instead includes the asymptotic approximation of the solution near the interface.
52
The equation in this example is the Helmholtz Equation on the interior of a circle with
radius
1
4
:
(3.6)
2
u +
2
u = 0,
where is given by
(x, y) =
_
_
200 r
1
4
,
0 r >
1
4
.
The exact solution to (3.6) is
u(x, y) =
_
_
I
0
(r)
I
0(
4
)
r
1
4
,
1 r >
1
4
,
where I
n
(x) is the modied Bessel function of the rst kind.
(a)
0 0.1 0.2 0.3 0.4 0.5
0
0.2
0.4
0.6
0.8
1
r
(b)
Numerical
Exact
Figure 3.5. Solution of Example 5. (a) A surface plot of the solution on
the entire domain. (b) A cross section of the solution, taken at an angle of
one radian. The numerical approximation overlaps the exact solution and
is computed using the exponential enrichment function on a 799799 grid.
53
An asymptotic analysis of (3.6) is contained in the appendix and shows that the
behavior of the solution in the boundary layer near the interface is of the form u e
r
.
Therefore, the exponential enrichment function is appropriate here as well, using as the
parameter. It is important to note that while an exact solution to (3.6) is available, we
only use information from the asymptotic analysis to construct our approximation.
As in Example 4, we once again consider two types of interface conditions. First, the
solution is again prescribed to be continuous, with a jump in the normal derivative at the
interface:
[[u]] = 0,
__
u
n
__
=
I
1
(r)
I
0
_
4
_ .
Table 3.15 shows the error on the domain, and Table 3.16 shows the error on the interface.
The results are similar to those in the Example 4.
n
||T
n
||
ratio ||T
n
||
ratio ||T
n
||
ratio
49 1.6496 10
1
2.8971 10
1
1.8642 10
2
99 7.7465 10
2
2.130 5.8281 10
2
4.971 5.4660 10
3
3.429
199 2.8396 10
2
2.728 1.1361 10
2
5.130 1.5259 10
3
3.582
399 8.8436 10
3
3.211 2.4660 10
3
4.607 3.8795 10
4
3.933
799 2.8451 10
3
3.108 8.1249 10
4
3.035 9.7831 10
5
3.966
Table 3.15. Domain results for Example 5 with Neumann jump interface conditions.
Like the rst example, the second set of boundary conditions applies a Dirichlet con-
dition at the interface, and the solution gradient at the interface is the desired result:
u = 1.
54
n
||T
n
||
ratio ||T
n
||
ratio ||T
n
||
ratio
49 3.2720 10
1
3.6044 10
1
3.4735 10
2
99 1.2621 10
1
2.593 1.1859 10
1
3.040 7.1808 10
3
4.837
199 3.9873 10
2
3.165 4.0367 10
2
2.938 1.9359 10
3
3.709
399 1.0574 10
2
3.771 1.2436 10
2
3.246 4.9710 10
4
3.887
799 2.9643 10
3
3.567 3.3563 10
3
3.705 1.2127 10
4
4.106
Table 3.16. Interface results for Example 5 with Neumann jump interface conditions.
Table 3.17 shows that the asymptotic enrichment again outperforms the other generic
enrichment functions with an error of less that 0.1% for the largest system. For the sake
of completeness, Table 3.18 charts the maximum error on the domain.
n
||T
n
||
ratio ||T
n
||
ratio ||T
n
||
ratio
49 3.5875 10
1
2.3265 10
1
3.6010 10
2
99 1.5830 10
1
2.266 9.0700 10
2
2.565 1.6945 10
2
2.125
199 5.7291 10
2
2.763 4.0960 10
2
2.214 4.0448 10
3
4.189
399 2.4665 10
2
2.323 1.8754 10
2
2.184 1.6265 10
3
2.487
799 1.2950 10
2
1.905 9.9692 10
3
2.067 5.9443 10
4
2.736
Table 3.17. Interface gradient results for Example 5 with Dirichlet interface conditions.
n
||T
n
||
ratio ||T
n
||
ratio ||T
n
||
ratio
49 3.0534 10
2
1.3923 10
1
6.6740 10
3
99 2.2035 10
2
1.386 6.1419 10
2
2.267 3.1346 10
3
2.129
199 1.4526 10
2
1.517 1.6752 10
2
3.666 7.4375 10
4
4.215
399 6.6449 10
3
2.186 4.8119 10
3
3.481 2.8600 10
4
2.601
799 2.0239 10
3
3.283 1.2758 10
3
3.772 5.6394 10
5
5.071
Table 3.18. Domain results for Example 5 with Dirichlet interface conditions.
55
Finally, the exponential enrichment function not only increases the accuracy at the
nodes and the interface, it improves the approximation within the elements as well. Be-
cause the asymptotic solution is contained in the interpolant, fewer elements are needed
to capture the boundary layer. Figure 3.6 shows the exact solution plotted next to the
numerical approximation for varying mesh sizes. Each plot is generated by evaluating
the exact solution and the numerical approximation at 10,000 points along a line from
r = 0.23 to r = 0.26 at an angle of one radian. This clearly demonstrates the advantage
of using a customized enrichment function, as the exponential enrichment accurately ap-
proximates the solution on a 25 25 grid, where the entire boundary layer is contained
within one element.
0.23 0.24 0.25 0.26
0
0.5
1
S
t
e
p

E
n
r
i
c
h
m
e
n
t
25 x 25
0.23 0.24 0.25 0.26
0
0.5
1
E
x
p
o
n
e
n
t
i
a
l

E
n
r
i
c
h
m
e
n
t
x = 0.04
0.23 0.24 0.25 0.26
0
0.5
1
50 x 50
0.23 0.24 0.25 0.26
0
0.5
1
x = 0.02
0.23 0.24 0.25 0.26
0
0.5
1
100 x 100
0.23 0.24 0.25 0.26
0
0.5
1
x = 0.01
0.23 0.24 0.25 0.26
0
0.5
1
200 x 200
0.23 0.24 0.25 0.26
0
0.5
1
x = 0.005
Figure 3.6. Cross sections generated using various mesh sizes. Each cross
section is taken at an angle of one radian, to avoid alignment with the mesh.
Note that in the coarsest plot, the entire boundary layer is contained in one
element. The solid line is the computed solution, and the dashed line is the
exact solution.
56
CHAPTER 4
Biolm Growth
We now consider the problem of bacterial biolms. The addition of substrate nutrients
to the top of the lm drives the growth of the biolm. Here we explore how the geometry
of the biolm colonies aects the prole of their growth. We rst consider the phenomenon
of tip-splitting [15], where the colony does not grow directly toward the substrate, but
develops ngers that grow at an angle into the lm. We then analyze how the shadowing
eect of a large colony decays with distance.
4.1. Mathematical Model
The system is considered on a domain , periodic in x and consisting of two regions,
the interior of the bacterial colonies
b
and the exterior uid lm
f
. The substrate
concentration s and velocity potential are governed by
(4.1) D
2
s = s
and
(4.2)
2
= s,
57
where
D(x, y) =
_
_
120 (x, y)
b
,
150 (x, y)
f
,
(x, y) =
_
_
3.6 10
6
(x, y)
b
,
0 (x, y)
f
,
and
(x, y) =
_
_
10
6
(x, y)
b
,
0 (x, y)
f
.
D
b
s = s
= s
D
f
s = 0
= 0
[[D n s]] = 0
= 0
s
y
=
y
= 0
s =
= 0
f
b
Figure 4.1. The domain and governing equations for the biolm problem.
Figure 4.1 shows the boundary conditions applied to each equation. This follows the
description given in [18] without erosion, with the parameters chosen to approximate the
behavior of the nonlinear system described in [16]. Both the substrate concentration and
the velocity potential are continuous throughout the domain. In the substrate equation,
a derivative jump condition is applied at the interface between the colony and the liquid,
and in the velocity equation, a Dirichlet condition is prescribed:
[[D n s]] = 0, = 0.
58
In both cases, a Dirichlet condition is applied at the top boundary:
s = 10
5
, = 0,
and a Neumann condition is applied at the bottom:
s
y
=

y
= 0.
We solve for the growth rate prole for a given colony via a three step process. First,
the substrate concentration is found by solving (4.1) with the given boundary conditions.
This solution is then fed back into (4.2) to nd the velocity potential within the colony,
and nally the rate of growth at the biolm surface is given by the normal derivative
of the velocity potential at the interface, n . We calculate this quantity at a large
number of points along the parameterized interface to obtain an overall prole of the
colonys growth. Both the substrate and velocity equations are solved using the X-FEM
with the exponential enrichment function (Section 2.1.1). The relevant boundary layer
exists on the interior of the colony, and we use =
_
/D
b
as the parameter in the
exponential enrichment denition.
4.2. Tip Splitting
As biolm colonies grow, they do not always grow in a uniform fashion. Sometimes
the tip of the colony splits into separate ngers instead [15]. This instability appears
to be distinct from the ngering instability described in [18], where it is observed that
peaks have a greater growth rate than valleys. Due to random uctuations, one of the
59
ngers generally overtakes the other resulting in an irregularly shaped colony. The con-
ditions driving this instability are unknown, but a possible mechanism is the splitting of
a mushroom-shaped tip due to the instability described in [18]. This is similar to the
well known tip-splitting instability in the Stefan problem, where a attened dendritic tip
splits due to the Mullins-Sekerka instability [32]. Here we look at the ways in which the
shape of a biolm colony can lead to mushrooming of the tip.
0 0.5 1 1.5 2 2.5 3
0.5
1
1.5
2
2.5
3
3.5
4
4.5
x 10
-3
Angle
V
e
l
o
c
i
t
y
(a)
Aspect Ratio = 2.25
Aspect Ratio = 16
-0.1 0 0.1
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
x
y
(b)
Figure 4.2. Representative velocity proles. Shown here are two examples
of velocity proles, from colonies with dierent aspect ratios but the same
mass. (a) Normal velocity vs. surface parameterization. In the second
case, the maximum velocity is not attained at the tip, so mushrooming
may occur. (b) Velocity vectors along the biolm surface. The taller colony
has a large aspect ratio, and the shape of the velocity eld near the peak
indicates that mushrooming may occur.
In order to study the onset of mushrooming, we examine the velocity proles of in-
dividual colonies on a periodic domain. The period of the domain is large enough to
prevent interaction between neighboring colonies, and all of the colonies are hemielliptical
in shape. Due to this simple geometry, the interface is easily parameterized from 0 to
, and Figure 4.2 shows some examples of parameterized velocity proles. In the rst
60
case, the biolm has an aspect ratio (height/width) of 2.25, and the maximum velocity
is obtained at the peak of the biolm colony, precluding a change in the shape of the
tip. In the second case, the colony is much taller and thinner, with an aspect ratio of 16,
and the maximum velocity is obtained on either side of the peak, which may allow for
mushrooming to take place. Thus, to determine whether or not mushrooming can occur,
we consider the ratio of the maximum velocity to the peak velocity. If this ratio is one,
the maximum occurs at the peak, and the colony will grow normally. The strength of the
mushrooming instability increases as this ratio rises above unity.
To demonstrate, we rst explore how the aspect ratio of the hemielliptical colony
aects the velocity prole. As we change the aspect ratio, we keep the area of the colony
constant so that the total amount of biomass is unchanged. The colony shape is dened
as
a
2
(x x
0
)
2
+
y
2
a
2
= R
2
,
where a
2
is the aspect ratio of the ellipse, and R
2
/2 is the area. We vary the aspect ratio
between 1.0 to 25.0 (height to width) for colonies with three dierent masses.
5 10 15 20 25
1
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
Aspect Ratio
M
a
x

V
e
l
o
c
i
t
y

/

P
e
a
k

V
e
l
o
c
i
t
y
R=0.1
R=0.2
R=0.3
R=0.4
Figure 4.3. The eect of aspect ratio on mushrooming.
61
Figure 4.3 shows the relationship between the aspect ratio and the strength of the
mushrooming instability. While the curves are qualitatively similar, the mass of the
colony is clearly important. This mass eect can be removed by considering the radius of
curvature at the colony peak, R/a
3
, rather than the aspect ratio. Figure 4.4 contains the
same velocity data as Figure 4.3 but plotted against the radius of curvature. In this case,
all three curves lie on top of each other, indicating that tip splitting is a function of the
radius of curvature at the tip of the colony. So that we may ensure that other factors are
not also important, the height of the lm and the period of the domain were varied. As
Figure 4.5 shows, these variations have no eect.
0 0.02 0.04 0.06 0.08 0.1
1
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
Peak Radius of Curvature
M
a
x

V
e
l
o
c
i
t
y

/

P
e
a
k

V
e
l
o
c
i
t
y
R=0.1
R=0.2
R=0.3
R=0.4
Figure 4.4. The eect of radius of curvature on mushrooming.
This dependence on radius of curvature is probably due to elevated substrate concen-
trations within the biolm colony. As the tip radius decreases, less biomass is present
near the peak, and unconsumed substrate can diuse farther down into the colony. To
quantify the substrate penetration, we calculate the depth within the colony at which
the concentration of substrate is 10% of the concentration at the colony tip. In order to
compare biolms containing dierent amounts of biomass, we normalize all of the values
62
0 0.02 0.04 0.06 0.08 0.1
1
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
Peak Radius of Curvature
M
a
x

V
e
l
o
c
i
t
y

/

P
e
a
k

V
e
l
o
c
i
t
y
Standard
Shorter Period
Thinner Film
Figure 4.5. Domain period and lm thickness have no eect on mushrooming.
by the substrate penetration depth into a circular colony with radius R. As shown in
Figure 4.6, the substrate penetration depth is correlated with the radius of curvature.
0 5 10 15 20 25
1
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
3
(a)
Aspect Ratio
N
o
r
m
a
l
i
z
e
d

S
u
b
s
t
r
a
t
e

P
e
n
e
t
r
a
t
i
o
n

D
e
p
t
h
R=0.1
R=0.2
R=0.3
R=0.4
0 0.1 0.2 0.3 0.4
1
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
3
(b)
Tip Radius of Curvature
N
o
r
m
a
l
i
z
e
d

S
u
b
s
t
r
a
t
e

P
e
n
e
t
r
a
t
i
o
n

D
e
p
t
h
R=0.1
R=0.2
R=0.3
R=0.4
Figure 4.6. Normalized substrate penetration depths. (a) The normalized
penetration depths for colonies of various sizes against the aspect ratio of
the colony. (b) The normalized penetration depths against the radius of
curvature at the colony tip.
4.3. Shadowing
Another important feature of biolm growth is the interaction between neighboring
colonies. Competition for substrate creates a shadowing eect near large colonies as
nearby colonies are unable to attain their optimal growth rate. This shadowing eect
63
depends on the relative sizes of the colonies as well as the separation distance. In order to
determine the relative importance of these two parameters, we consider a system consisting
of alternating large and small colonies along a periodic domain. In each case the large
colony remains a standard size, with a = 2.0 and R = 0.2. The smaller colony maintains
the same aspect ratio, but R is varied from 0.02 to 0.18. The dierent size cases are then
simulated using a wide range of period lengths.
0 1 2 3 4
1
1.5
2
2.5
3
3.5
4
4.5
5
5.5
6
x 10
-3
Period Length
M
a
x
i
m
u
m

I
n
t
e
r
f
a
c
e

V
e
l
o
c
i
t
y
(a)
0 0.05 0.1 0.15 0.2
3
3.5
4
4.5
5
5.5
6
x 10
-3
R
V
e
l
o
c
i
t
y
(b)
R=0.2
R=0.1
R=0.02
Maximum Value
Value at Peak
Figure 4.7. Selected natural growth rates. (a) The maximum growth rate
for each colony size is reached on a domain with a relatively small period.
(b) While the velocity at the peak decreases monotonically with colony size,
the maximum velocity begins to increase again for the smallest colonies.
This eect is caused by mushrooming.
Proper determination of the eects of the interaction requires the maximum growth
rate of an uninuenced single colony, the natural growth rate for that colony. Due
to the fact that the majority of a biolm colonys growth occurs at or near the tip, this
maximum growth rate is a reasonable measure of the overall growth rate of the colony. To
nd this rate, simulations are performed using a single colony on a periodic domain and the
same lm height as in the combined system. The period of the domain is increased until
maximum growth rate asymptotes to a constant value. Figure 4.7 shows maximum growth
64
rate plotted against the period size for several representative colony sizes. Interestingly,
the maximum growth rate of the smallest colonies is greater than that of the medium sized
colonies. This eect is caused by mushrooming. When the colonies are relatively large,
growth occurs primarily at the tip, and therefore the maximum growth rate scales with
the size of the colony. However, in small colonies, growth occurs throughout the colony
due to the greater substrate availability. This allows for large growth rates in areas far
from the peak and creates the eect in Figure 4.7.
0 0.5 1 1.5 2 2.5 3
0
10
20
30
40
50
60
70
80
90
100
A = 0.002
A = 0.016
Period Length
P
e
r
c
e
n
t

o
f

N
a
t
u
r
a
l

G
r
o
w
t
h

R
a
t
e
Figure 4.8. The eect of period length on shadowing. For all colony sizes,
the shadowing eect of the large colony decreases for longer periods.
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
C
o
l
o
n
y

H
e
i
g
h
t
Position Relative to Central Colony
Figure 4.9. The 50% shadow. Any colony contained within this region
around the large central colony will achieve less than 50% of its maximal
growth rate.
65
Once the natural growth rate for each colony size has been found, it is possible to
quantify the shadowing that takes place in the combined system. For each colony size
and period length, the maximum growth rate can be expressed as a percentage of the
natural rate for that colony. Figure 4.8 shows this ratio for each colony size, plotted
against the length of the period. Using this data, it is then possible to create pictures
of the shadow for a given inuence percentage. For example, the portrait of a 50%
inuence shadow is shown in Figure 4.9. In the center of the gure is the large colony,
surrounded by colonies placed so that their maximum growth rate is 50% of their natural
rate. The tops of the colonies are connected to create a shadow region. Any colony which
is contained within this shadow region will grow at a maximum of 50% of its natural
rate. Of particular interest is the 97% shadow, shown in Figure 4.10, because rather than
decaying asymptotically, the region of inuence appears to have a sharp cuto outside of
which the presence of the large colony is virtually undetectable.
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
C
o
l
o
n
y

H
e
i
g
h
t
Position Relative to Central Colony
Figure 4.10. The 97% shadow. Any colony which falls outside of this region
will achieve at least 97% of its maximal growth rate. While this shadow is
signicantly wider than the 50% shadow, it exhibits a steep drop o at the
edges, so the eect of the central colony is largely conned to a nite area.
66
CHAPTER 5
Stokes Flow and Film Thinning
In order to understand the dynamics of foam solidication, a method for simulating
the dynamics of an evolving liquid lm must rst be developed. In the foams considered
here, the Reynolds numbers are suciently small that the Stokes equations can be used
to approximate it. Here we develop a method for solving these equations on a domain
with xed boundaries and then consider how to modify it to incorporate the dynamics
of a free boundary. This method is then used to study the eect of initial and boundary
conditions on foam evolution.
5.1. Stokes Flow Around a Solid
Consider the Stokes equation for low Reynolds number ow:
2
u = p, (5.1)
u = 0. (5.2)
Due to the presence of the incompressibility constraint, (5.2), a mixed nite element
method must be utilized to approximate both the uid velocity, u, and the pressure, p.
67
The variational forms of equations (5.1) and (5.2) for the velocity and pressure are:
u
_
p = 0, (5.3)
_
( u) = 0, (5.4)
for all velocity test functions and pressure test functions . In order to ensure the
existence and uniqueness of the solution to (5.3) and (5.4), the function spaces for the
velocity and pressure must satisfy the inf-sup, or LBB, condition [1, 2, 31].
The X-FEM approximation described in chapter 2 is used to discretize the velocities,
(u, v) and the pressure, p, yielding:
u
h
(x, y) =
n
i
N
i
(x, y) u
i
+
n
j
N
E
j
(x, y)
k
() a
j
,
v
h
(x, y) =
n
i
N
i
(x, y) v
i
+
n
j
N
E
j
(x, y)
k
() b
j
,
p
h
(x, y) =
n
i
N
i
(x, y) p
i
+
n
j
N
E
j
(x, y)
k
() c
j
.
These equations are dened on a rectangular domain containing a solid bounded by
an interface
I
. No slip and no pentration conditions are applied on the solid boundaries,
, as well as the interface. The homogeneous natural boundary condition:
(5.5)

n
(u n) +P = 0,
68
is applied at the outow boundary. In order to properly apply the outow condition, the
term:
_
p
from (5.3) must be integrated by parts. This yields the gradient terms:
(G
x
)
i,j
=
_
i
x
,
_
G
y
_
i,j
=
_
i
y
.
The full linear system is therefore:
(5.6)
_
_
K 0 G
x
0 K G
y
G
T
x
G
T
y
0
_
_
_
_
u
v
p
_
_
=
_
_
0
0
0
_
_
,
where K is the Laplacian and G
T
x
, G
T
y
, G
x
, and G
y
are gradient terms.
In order to ensure the stability of the method, the function spaces for the velocity and
pressure must be chosen to satisfy the LBB condition. Integrating the pressure gradient
term by parts allows for the use of bilinear velocity and piecewise constant pressure
elements (Q
1
Q
0
). While (Q
1
Q
0
) does not strictly satisfy the LBB condition, it is still
usable, [28], but can generate pressure oscillations. Fortunately, the presence of the
interface and enrichments destroys the symmetry that generates the oscillations, so these
are not a concern when using the X-FEM.
69
5.1.1. Flow past a cylinder
Consider ow past a cylinder in a channel as an example of Stokes ow around a solid.
A cylinder with a radius of 0.25 is placed in the center of a channel with height 1.0. A
constant inow condition is applied at the left end of the channel, with no slip and no
penetration conditions at the channel walls, and a homogeneous outow condition, (5.5)
at the exit.
0 1 3 4
0
1
u = v = 0
u = v = 0
u = v = 0
u = U y 1-y ( ) [ ] 0
v = 0
r=1/4
du
dx
+ P = 0
Figure 5.1. Flow past a cylinder: the computational domain.
Figure 5.2. Flow past a cylinder: horizontal velocities.
Figure 5.3. Flow past a cylinder: vertical velocities.
70
Figure 5.4. Flow past a cylinder: pressure eld.
Figure 5.5. Flow past a cylinder: streamlines.
5.2. Stokes Flow With a Free Surface
Now consider the introduction of a free surface to the Stokes ow problem. Instead
of the no slip and no penetration conditions that are applied at a solid boundary, a free
surface requires normal and tangential stress conditions at the interface:
(5.7)
n n = ,
t n = 0,
where n and

t are the unit normal and tangent vectors to the interface, is the surface
tension at the interface, and is the interfacial curvature. The uid stress tensor, , is
dened as:
=
_
_
p + 2
u
x

_
u
y
+
v
x
_
_
u
y
+
v
x
_
p + 2
v
y
_
_
.
71
Equation (5.7) can be expanded in terms of u, v, and p, and expressed in Cartesian
coordinates as:
(5.8)
p n
x
+
_
u
y
+
v
x
_
n
y
+(u n) = n
x
,
p n
y
+
_
u
y
+
v
x
_
n
x
+(v n) = n
y
.
Unfortunately, these types of boundary conditions are nearly impossible to impose
using Lagrange multipliers and are not natural boundary conditions when using the form
of the Stokes equation given in section 5.1. In order to apply these conditions properly,
the stress divergence form of the Stokes equation is used:
=
_
I
, (5.9)
u = 0. (5.10)
This can be rewritten as:
p
x
+
2
u +

x
_
u
x
+
v
y
_
= 0,
p
y
+
2
v +

y
_
u
x
+
v
y
_
= 0.
In general, the last term in each equation can be eliminated using the continuity condition,
but in this case they are utilized to apply the free surface boundary conditions. These
terms, as well as the pressure terms, are integrated by parts to obtain the variational
72
forms of equation (5.7):
_
x

_
u
_
u
x
x

_
v
x
y
=
_
I
p n
x
I
u
n
I

_
u
x
n
x
I

_
v
x
n
y

I
+
_
I
n
x
I
,
_
y

_
v
_
u
y
x

_
v
y
y
=
_
I
p n
y
I
v
n
I

_
v
y
n
y

I

_
u
y
n
x
I
+
_
I
n
y

I
,
_
u
x
+
_
v
y
= 0.
This set of equations describes the uid motion within the bulk and correctly applies
the free surface boundary conditions without the need for additional degrees of freedom.
The implementation of this method requires the development of some additional X-FEM
matrices for the second order operators other than the Laplacian. These include:
(S
xx
)
i,j
=
_
i
x
j
x
,
(S
xy
)
i,j
=
_
i
x
j
y
,
(S
yx
)
i,j
=
_
i
y
j
x
,
(S
yy
)
i,j
=
_
i
y
j
y
.
73
In addition, the force vectors take the form:
f
x
=
_
i
n
x
I
,
f
y
=
_
i
n
y

I
.
The overall X-FEM system matrix for Stokes ow with a free surface is therefore:
_
_
K S
xx
S
yx
G
x
S
xy
K S
yy
G
y
G
T
x
G
T
y
0
_
_
_
_
u
v
p
_
_
=
_
_
f
x
f
y
0
_
_
.
This system is solved at each timestep. Once the linear system has been solved, u and
v are evaluated along the interface and the level set method uses this velocity data to
evolve the interface.
5.3. Film Thinning in Foam Dynamics
A foam is composed of gas bubbles within a liquid [6]. If the volume fraction of the
liquid is suciently small, the bubbles are polygonal and separated by thin lms connected
by junctions. The thinning and rupture of these lms determines the lifetime of the foam.
As the foam evolves, the bubbles coalesce and ultimately combine to form a single gas
bubble within the liquid. This process is one of the fundamental building blocks of the
theory of large scale foam behavior.
The foams considered here have volume fractions around ten percent. This is small
enough that the bubbles cannot relax to a circle, but large enough for the numerics to
be tractable. As shown in gure 5.6, these foams are assumed, at t = 0, to exist on
74
a periodic square array and consist of thin lm regions, lamellae, separated by rounded
corners, plateau borders, of nearly constant curvature. As the foam evolves, it can be
described by the radius of curvature of the plateau border regions and the thickness of
the lamellae.
2h
Computational
Domain
Liquid
Gas Bubble
Gas-Liquid Interface
a
Plateau Border
Lamella
L
0
Figure 5.6. The lm thinnig domain. A schematic representation of a cross
section of a square array of bubbles showing the lamellar (L) and plateau
border (PB) regions. The dashed line gives the boundary of the period
computational domain.
5.3.1. Nondimensionalization
In order to study general trends within the system, the governing equations must be
nondimensionalized. This is done with the following nondimensional variables:
( x, y) =
_
x
L
0
,
y
L
0
_
, ( u, v) =
_
u
U
0
,
v
U
0
_
,
t =
_
U
0
L
0
_
t, p =
_
L
0
_
p,
75
where L
0
is half the period length of the bubble array and U
0
is a characteristic velocity.
The velocity U
0
is chosen to specify the capillary number C = U
0
/, which gives the
relative importance of viscous eects versus that of surface tension eects. Dropping the
tildes, the full nondimensional system to be solved is therefore:
p
x
+C
_
2
u +

x
_
u
x
+
v
y
__
=
_
I
(x X (S)) n
x
I
, (5.11)
p
y
+C
_
2
v +

y
_
u
x
+
v
y
__
=
_
I
(x X (S)) n
y

I
, (5.12)
u
x
+
v
y
= 0. (5.13)
5.3.2. Results
In order to understand the evolution of these types of foams, the system is simulated using
a variety of initial conditions as well as several types of boundary conditions. Two primary
types of initial conditions are implemented: the rst uses a fourth-order polynomial in the
corner so that curvature is continuous across the entire lm, and the second uses a quarter
circle to attach two at lms. These two types of initial conditions are run with several
dierent initial lm heights in order to study the eect of varying volume fractions.
The rst initial condition to be considered is designed to have continuous curvature
throughout the interface. The thin lm regions at the edges have height 0.05, and the
corner region is dened as:
x(s) =
1
64
_
5 + 16s + 15s
2
5s
4
+s
6
_
_
1
2
h
0
_
+ 1 h
0
,
y (s) =
1
64
_
5 16s + 15s
2
5s
4
+s
6
_
_
1
2
h
0
_
+ 1 h
0
.
76
for 1 s 1. The volume fraction of uid is 11.19%. Figure 5.7 shows the shape of this
initial condition as well as the parameterized curvature along the interface. Beginning
0 0.5 1
0
0.5
1
y
x
Figure 5.7. Sample Initial Condition. The shape of the initial condition to
be used as a reference to explore the eects of curvature conguration, lm
height, and volume fraction on the long time behavior of a gas bubble.
with this shape, the lm is evolved from t = 0 to t = 200. The shape of the interface
at various times is shown in Figure 5.8, and the curvature proles are given in Figure
5.9. As these gures show, the lm relaxes rather quickly to a conguration with two at
surfaces connected by a corner region of nearly constant curvature. Once the lm has
taken on this general shape, uid slowly drains from at lm into the corner, leading to a
decrease in the minimum height of the lm. This eect is shown in Figure 5.10, as there
is an initial period of relatively rapid change, followed by a slow decrease in

h = h/h
0
.
Note that for long times,

h t
2
, as predicted by the asymptotic analysis in [6]. Using
the results from this initial conguration as a reference, the eects of varying the initial
curvature eld, lm height, and volume fraction are explored.
77
0 0.5 1
0
0.5
1
t = 0
0 0.5 1
0
0.5
1
t = 1
0 0.5 1
0
0.5
1
t = 10
0 0.5 1
0
0.5
1
t = 50
0 0.5 1
0
0.5
1
t = 100
0 0.5 1
0
0.5
1
t = 200
Figure 5.8. Interface shapes of an evolving lm. The progression of the
interface shape as the lm evolves throught time, showing both the rapid
initial transitory period and the slow drainage at long time.
Consider the case where the initial lm conguration consists of two at lms con-
nected by a quarter circle. Here, the initial lm height is again 0.05 and the radius of
curvature of the corner region is selected so that the volume fraction is 11.19%. Figure
5.11 compares this conguration to the reference initial condition. As this gure shows,
these initial conditions are very similar, and this is conrmed when comparing the evo-
lution of the lm height over time (Figure 5.12). This shows that the initial curvature
conguration is relatively unimportant to the long time behavior of the lm. In order to
determine the relative importance of the initial volume fraction, three initial conditions
with the same volume fraction, 11.19%, but varying initial lm heights, 0.05, 0.055, and
0.025, are considered. The rst condition listed is the reference described above with
78
0 0.5 1 1.5
-1
0
1
2
3
4
5
6
t = 0
s
0 0.5 1 1.5
-1
0
1
2
3
4
5
6
t = 1
s
0 0.5 1 1.5
-1
0
1
2
3
4
5
6
t = 10
s
0 0.5 1 1.5
0
0.5
1
1.5
2
t = 50
s
0 0.5 1 1.5
0
0.5
1
1.5
2
t = 100
s
0 0.5 1 1.5
0
0.5
1
1.5
2
t = 200
s
Figure 5.9. Interface curvature of an evolving lm. The progression of the

interface curvature as the lm evolves through time.
10
-1
10
0
10
1
10
2
10
-3
10
-2
10
-1
10
0
log t
l
o
g

h ^
h ~ t
-2 ^
Figure 5.10. Film Height vs. Time.
79
0 0.5 1
0
0.5
1
Continuous Curvature
Discontinuous Curvature
Figure 5.11. Sample Initial Condition: Varying initial curvatures.
10
-1
10
0
10
1
10
2
10
-4
10
-3
10
-2
10
-1
log t
l
o
g

h
Continuous Curvature
Discontinuous Curvature
Figure 5.12. The eect of initial curvature conguration on lm evolution.
continuous curvature, but the other two begin with discontinuous curvature along the
interface. Figure 5.13 shows these three interfaces. Note that in contrast to the initial
conditions shown in gure 5.11, these interfaces are qualitatively much dierent. However,
as these lms evolve, they all converge to the same ultimate shape. Figure 5.14 shows the
80
0 0.5 1
0
0.5
1
h
0
= 0.025
h
0
= 0.05
h
0
= 0.055
Figure 5.13. Sample Initial Conditions: Varying initial lm heights.
behavior of the lm height, h, and the radius of curvature at the center of the interface,
a, as the interface evolves. In both plots, despite the dierent initial conditions, all three
10
-1
10
0
10
1
10
2
10
-4
10
-3
10
-2
10
-1
log t
l
o
g

h
10
0
10
1
10
2
10
-0.6
10
-0.5
10
-0.4
10
-0.3
10
-0.2
10
-0.1
log t
l
o
g

a
h
0
= 0.025
h
0
= 0.05
h
0
= 0.055
Figure 5.14. The eect of initial lm height on lm evolution. A loglog
plot of the lm height and radius of curvature vs. time. For long times, the
three lms converge to the same shape despite varying initial heights.
lms reach the same shape at long times. This result suggests that the most important
parameter for determining the ultimate shape of the foam is the volume fraction. To test
this, the behavior of three lms with dierent volume fractions is simulated. Figure 5.15
81
shows the initial and nal congurations for lms having volume fractions of 7.5%, 11.2%,
and 15.1%. The long time behavior of the lm height and radius of curvature is given
in gure 5.16. As predicted, the resulting lm congurations are very dierent. This is
0 0.5 1
0
0.5
1
t = 0
0 0.5 1
0
0.5
1
t = 200
Figure 5.15. Sample Initial and Final Conditions: Varying volume fractions.
due to the fact that for long times, very little of the uid volume is contained in the thin
lm region. Therefore, the radius of curvature at the center is eectively determined by
the volume fraction. It is worth noting that the transitory period before the lm height
converges to the power law phase, h t
2
, is longer for lms with larger volume fractions.
In addition to studying the eect of varying initial conditions, dierent types of bound-
aries are considered as well. First, a no slip condition is applied at the right-hand edge
to simulate the presence of a vertical solid wall between cells. This greatly changes the
shape of the evolving lm, as uid drains from the edge near the wall much more slowly.
The second type of varied boundary condition is the application of a no slip condition at
the bottom boundary. This simulates the presence of a static solidication front midway
through the lm. As with the wall condition, this signicantly changes the shape of the
82
10
-1
10
0
10
1
10
2
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
log t
l
o
g

h
10
0
10
1
10
2
10
-1
10
0
log t
l
o
g

a
Figure 5.16. The eect of volume fraction on lm evolution.
evolved lm. The xed point along the bottom boundary prevents the lm from thinning
as quickly as in the standard system, and approximates the manner in which a slowly
freezing foam might evolve. Figure 5.17 shows the shapes of these two types of lms at
various times. As predicted, the lm at the right of the domain, near the altered bound-
ary condition, drains much more slowly than in the reference foam. However, the lm
at the top of the domain, where conventional boundary conditions are applied, drains at
the same rate as in the reference foam. This is conrmed by the plot of h vs. t, given in
Figure 5.18.
83
0 0.5 1
0
0.5
1
t = 0
0 0.5 1
0
0.5
1
t = 10
0 0.5 1
0
0.5
1
t = 100
0 0.5 1
0
0.5
1
t = 0
0 0.5 1
0
0.5
1
t = 10
0 0.5 1
0
0.5
1
t = 100
Figure 5.17. Sample Evolving Interface Shapes: Varying boundary condi-
tions. The top row compares the shape of the interface generated using a
wall condition at the right to the reference lm. The bottom row compares
the shape of the interface generated with a solid boundary at the bottom
to the reference lm. In both cases, the evolution of the lm layer at the
top of the domain remains the same while the lm layer at the right takes
on a much dierent shape.
10
-1
10
0
10
1
10
2
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
log t
l
o
g

h
10
0
10
1
10
2
10
-1
10
0
log t
l
o
g

a
h
top
h
right
Figure 5.18. The eect of boundary conditions on lm evolution.
84
CHAPTER 6
Triple Points and Foam Solidication
The nal step in the process of modeling foam solidication is the development of a
method for handling triple points and applying boundary conditions along segments of
an interface.
6.1. The Triple Point
As shown in Figure 6.1, the domain is divided into three domains, a liquid domain, a
solid domain, and a gas domain, using two interfaces,
g
and
f
. The interface
g
, the
zero contour of the level set function
g
, splits the domain into gas and non-gas regions,
and
f
, the zero contour of the level set function
f
, represents the freezing front. In order
to pass the locations of the three phases within to the X-FEM, each phase i is dened
by a level set function
i
, where
i
0 inside the phase and
i
< 0 outside the phase.
These functions are constructed from the level set functions,
g
and
f
. In this case, the
gas phase is the region where
g
> 0, the liquid phase is the region where
f
< 0 and
g
< 0, and the solid phase is the region where
f
> 0 and
g
< 0. At every point, x, in
the domain there is a vector containing the values of the level sets dening each phase:
(x) = (
1
(x) ,
2
(x) ,
3
(x)). To avoid errors around the triple junction, constraints are
85
g
Liquid
Solid
Gas
Gas
s
>0
L
>0
g
>0
Figure 6.1. The domain used in the solidication problem.
placed on the values of (x) such that:
i
(x) 0,
j
(x) =
i
(x) ,
k
(x)
j
(x) ,
for k = i, j where x
I
. These constraints restrict (x) to a two dimensional manifold,
, embedded within R
3
. While the three level set functions determine the element, N
trip
,
which contains the triple point, the location of the triple point within N
trip
is determined
by the map used to project onto R
2
. There are many well dened maps for this process,
but only two are considered here. As shown in Figure 6.2, the rst map chooses the triple
point so that the freezing front intersects both the liquid-gas and solid-gas interfaces at
right angles. The second map, illustrated by Figure 6.3, simply chooses the intersection of
86
g
x
Tr
Figure 6.2. One strategy for determining the subgrid location of the triple
point is choosing the point so that the line segments intersect at a right
angle.
the zero contours of
f
and
g
as the triple point without considering the angle between
the interfaces.
g
x
Tr
Figure 6.3. A second strategy for determining the subgrid location of the
triple point is choosing the intersection of the level sets
g
and
f
.
87
6.1.1. Integration Around a Triple Point
Accurate integration around the triple point requires that the level set functions
s
,
l
,
and
g
be evaluated properly at the Gauss-Legendre points within the element N
Tr
. In
the absence of a triple point, the level set function is approximated at the subgrid level
with a linear interpolant generated by the values of at the element corners, eectively
modeling the interface as a line segment within the element. When an element contains
a triple point, each interface is represented as two line segments rather than one. The
level set value at an interior point x
i
is then calculated in two steps. First, the absolute
value is determined by the minimum of the distances to each interface segment. To then
determine the correct sign, line segments are created connecting x
i
to each corner of the
element. At least one of these articial segments does not intersect an interface segment,
and the sign of the level set function at that corner is copied to (x
i
). This procedure
ensures that each interior point is considered to be inside exactly one phase. In addition,
the line segment approximations for the interior interfaces allows for a relatively simple
integration scheme, similar to the scheme described in Section 2.3.2 for single interfaces.
An example is shown in Figure 6.4.
6.2. Interface Conditions
In addition to the complications created by the triple point, the presence of three
phases means the method must be able to distinguish between the dierent types of
interfaces in order to correctly apply interface conditions. Due to the three level set
representation for the three phases, a scheme similar to that described in Section 2.2
can be used in triple point calculations as well. Conventional single interface problems
88
g
x
Tr
g
Figure 6.4. An element containing an a triple point is partitioned into tri-
angles and using a linear approximation of the interfaces.
create an interface segment whenever the zero level set passes through an element. In this
case, an interface segment is created when two zero level sets pass through an element.
All three interfaces only exist in one element at the triple point, and the appropriate
line segment constructed by the method in Section 6.1.1 is used to apply the condition.
This allows the method to apply dierent conditions on the liquid-solid, liquid-gas, and
solid-gas interfaces.
6.3. Freezing in a Simple Geometry
An example of a problem involving the motion of a triple point is the freezing of a
at vertical lm. In this simple problem, the solid-gas, liquid-gas, and initial liquid-solid
interfaces are initially modelled as straight lines, and the temperature is assumed to reach
89
steady state at each time step. The governing equation for the temperature, T, is:
(6.1)
2
T = 0,
in both the solid and the liquid, with insulating temperature conditions at the liquid-gas
and solid-gas interfaces:
T
x
= 0,
and the condition:
T =
sl
,
at the solidication front where
sl
is the curvature of the solidication front. Dirichlet
conditions are applied at the top and bottom boundaries,
T = s, y = 1,
T = 1, y = 0,
where s is related to the degree of undercooling, and the speed of the front is given by:
(6.2) v
sl
=
__
T
n
__
.
Near the triple point, values of the three level set functions are determined so that the
interfaces are forced to intersect at right angles. To move the solidication front in the
gas as well as the liquid, v is extrapolated linearly into the gas. This linear extrapola-
tion preserves the continuity of the level set curvature eld and is necessary to prevent
the interface from becoming unstable. The resulting system should behave like a one
90
dimensional problem, with the interface advancing according to the equation:
(6.3) v (y) =
1
y

s
1 y
.
To test the eectiveness of the triple point algorithm, equation (6.1) is solved on a domain
with the liquid-gas interface at x = 0.5, the initial solication front at y = 0.1, and
s = 0.2. The liquid and solid phases are each enriched with a Heaviside enrichment
function, and the gas phase is treated as a void, as described in Section 2.1.3. In addition,
the unenriched degrees of freedom around the liquid-solid interface are removed, so that
the temperature eld in each phase is essentially solved separately. Figure 6.5 shows the
computed two-dimensional solution as well as the exact solution. In the second plot, the
exact position was obtained by numerically integrating the one-dimensional exact velocity
given by (6.3). The computed solution agrees with the exact solution, indicating that the
triple point algorithm performs correctly.
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
-2
0
2
4
6
8
10
y
sl
v
s
l
0 0.5 1 1.5 2 2.5 3
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
t
y
s
l
Computed
Exact
Computed
Exact
Figure 6.5. Panel (a) shows the velocity as a function of the height of the
solid-liquid interface, and panel (b) shows the height of the solid-liquid
interface as a function of time. Note that the computed and exact solutions
are shown overlapping in the plot.
91
6.4. Foam Solidication
The nal problem considered here is the eect of a solidication front on the rate of
lm thinning in a liquid foam. For the sake of simplicity, we again simulate the motion
of one corner of a bubble within a larger array.
6.4.1. Governing Equations
Simulating the solication system involves solving for both the temperature eld and the
uid velocities. To simplify the calculation, we assume that the system reaches steady
state at each timestep. In the liquid, we solve both the heat and Stokes equations:
= 0, (6.4)
u = 0, (6.5)
2
T = 0. (6.6)
In the solid, only the temperature eld is relevant, so only the heat equation is considered:
(6.7)
s
2
T = 0.
Interface conditions must be prescribed on all three interfaces. At the solid-gas interface
(
sg
), an insulating heat condition is applied:
T
n
= 0.
92
At the liquid-gas interface (
lg
), normal and tangential stress conditions are applied to
the free surface, as well as another insulating heat condition:
n n = ,
t n = 0,
T
n
= 0,
where is the interfacial curvature and is the surface tension. At the solidication front
(
ls
), no slip and no penetration conditions are applied to the uid, and the temperature
condition is derived from the phase diagram:
u = v = 0,
T = T
m
_
1 +
ls
sf
L
v
_
,
where
ls
is the curvature of the front and
sf
is the surface energy. At the domain
boundaries, mirror conditions are applied on all sides to the uid and at the left and right
boundaries in the temperature eld. Dirichlet temperature conditions are applied at the
top and bottom boundaries:
T = T
0
y = L,
T = T
y = 0.
93
Finally, expressions for the motion of the liquid-gas and liquid-solid interfaces are given
by:
v
lg
= u n,
v
ls
=
1
L
v
__
k
T
T
n
__
.
The level set functions dening the phase geometries are derived from two master level
set functions
g
and
f
. As discussed in Section 6.1,
g
denes the gas/non-gas regions
of the domain, and
f
denes the solidication front. In this problem, the location of
the triple point is dened as the intersection of the zero contours of these two level sets.
Due to the fact that part of the solidication front lies in the gas phase, an extrapolation
procedure is necessary for the interface speed in this region. Because the phase boundaries
each form a corner at the triple junction, the level set curvature calculation is unstable in
the neighborhood of the triple point. Unfortunately, this numerical instability propagates
across the solidication front over time, causing the approximation to diverge. To counter
this instability, we model the front as a straight line, with the front speed across the entire
interface
f
given by the average speed across the solid-liquid interface.
In order to more easily compare the eects of various parameters, the equations are
nondimensionalized before simulations are performed. The nondimensional variables are:
u =
u
U
0
, x =
x
L
,

t =
U
0
L
t,
and
p =
_
L
0
_
p, =
T T
m
T
m
T
.
94
After substituting the nondimensional variables into the governing equations and dropping
the tildes, the problem is dened as:
p
x
+Ca
_
2
u +

x
_
u
x
+
v
y
__
=
_
lg
(x X (S)) n
x
lg
, (6.8)
p
y
+Ca
_
2
v +

y
_
u
x
+
v
y
__
=
_
lg
(x X (S)) n
y

lg
, (6.9)
u
x
+
v
y
= 0 (6.10)
2
= 0, (6.11)
in the liquid, and:
(6.12)
2
= 0.
in the solid. In this formulation, the normal and tangential stress conditions at the liquid-
gas interface are applied weakly. Therefore, the only interface condition to be applied at
lg
is the insulation condition:
n
= 0.
The same condition is applied at the solid-gas interface
sg
. Because the solidication
front is assumed to be at, the temperature condition at the front is simplied:
T = 0.
The no slip and no penetration conditions remain the same. The temperature in the liquid
is taken to be the melting temperature T
m
so that the Dirichlet boundary conditions on
95
the heat equation are:
= 0, y = 1,
= 1, y = 0.
Finally, the liquid-gas interface evolves at each time step according to:
v
lg
= u n
x
+v n
y
,
and the solidication front moves with speed:
v
ls
=
T
T
m
1
Ca
k
T
T
m
L
v
L
__
n
__
.
Here, T/T
m
is the applied undercooling and can be varied to control the speed of the
front. The remaining parameters are determined by the uid system and the material.
The following examples consider the case where Ca = 0.001, and the foam is aluminum.
The material parameters, taken from [30], are:
k
T
= 210W/mK,
T
m
= 933.6K,
L
v
= 9.5 10
8
J/m
3
,
= 8.4 10
1
kg/s
2
,
= 4.0 10
3
kg/ms,
L = 1.0 10
3
m.
96
Using these values, the expression for the front velocity becomes:
(6.13) v
s
=
T
T
m
A
0
__
n
__
.
where A
0
0.9827.
6.4.2. Solidication Simulation
For the rst simulation of a freezing foam cell, we use the physically realistic value of 10%
for the undercooling, and the initial solidication front is placed at y = 0.1. Because the
temperature condition at the top boundary of the domain is set to T
m
, the nondimensional
temperature in the liquid is uniformly zero. Therefore, only the velocity eld is computed
in the liquid, and only the temperature eld is computed in the solid. In each case, a
Heaviside enrichment function is used to apply the interface conditions, and the remaining
phases are treated as voids. The conditions at the liquid-gas and solid-gas interfaces
are applied weakly, and the conditions along the solidication front are applied using
Lagrange multipliers. At each time step, the liquid velocity and temperature elds are
calculated, and the local velocity eld at the liquid-gas and liquid-solid interfaces must be
computed. The velocity of the liquid-gas interface can be evaluated directly, but because
the solidication front velocity depends on the jump in the derivative in the temperature
eld (6.13), it must be computed using an integral method (2.6). Once the speed has
been determined across the liquid-solid interface, an average is taken in order to move the
entire front.
The solidication front begins at y = 0.1 and moves up over the course of the simu-
lation. Figure 6.6 shows the conguration of the lm and the front at time zero and at
97
t = 0.19. At this nal time, the lm at the top of the domain has evolved as far as the
method can resolve. Figure 6.7 shows the time evolution of the lm thicknesses at both
0 1
0
1
t = 0
0 1
0
1
t = 0.19
Figure 6.6. Film congurations at t = 0 and t = t
final
for the solidication
problem with 10% undercooling.
the top and the side of the domain, and compares it to the evolution of the liquid lm in
Chapter 5. Note that the results are given in non-dimensional time. As shown here, the
solidication front prevents the lm at the right edge from fully evolving. However, due
to the fact that the front moves very slowly, it has almost no eect on the lm at the top.
To determine whether the front can aect the lm at the top of the domain using this
physical model, simulations were performed using extremely high, physically unrealistic
values for the undercooling including 75%, 100%, and 200%. In these simulations, the
front starts at y = 0, with the bottom boundary conditions applied at y = 0.1. The
nal lm conguration for each case is shown in Figure 6.8. While the nal position of
the solidication front is well into the corner region of the lm, it still does not reach this
region quickly enough to aect the thin lm. Figure 6.9 shows the time evolution of the
thin lms for all three values of the undercooling. As in the other case, the evolution
of the lm on the right is quickly halted by the solidication, but the lm at the top is
98
10
-4
10
-3
10
-2
10
-1
10
0
10
-4
10
-3
10
-2
10
-1
Thin Film
log hy
log hx
Figure 6.7. Time evolution of the lm height in the solidication problem
with 10% undercooling.
0 0.5 1
-0.1
0
0.5
1
T/T
m
= 200%
0 0.5 1
-0.1
0
0.5
1
T/T
m
= 100%
0 0.5 1
-0.1
0
0.5
1
T/T
m
= 75%
Figure 6.8. Final lm congurations when applying extreme undercooling.
unaected. This lack of inuence is due to the fact that the uid ow in this problem is
driven almost entirely by the local curvature at the intersection of the lamellar regions
and the plateau borders. Due to the fact that no constraint is applied to the intersection
angle at the triple junction, the solidication front has little eect on the evolution of the
interfactial curvature within the liquid phase. It can be seen in Figure 6.10 that there is
very little dierence between the interfacial curvature in the solidication problem and
99
10
-2
10
-1
10
0
10
1
10
2
10
-5
10
-4
10
-3
10
-2
10
-1
log t
l
o
g

h
h
x
h
y
Figure 6.9. Time evolution of the lm height in the solidication problem
with extreme undercooling. The degree of undercooling determines the
nal lm height h
x
along the side of the domain, but has no eect on the
evolution of h
y
along the top of the domain.
that in the pure lm thinning problem, and it is this that leads to the nearly identical
behavior seen above.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
-0.2
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
x
System with no Solidification

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
-0.2
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
x
System with Solidification

t = 1.98
t = 19.98
t = 1.98
t = 19.98
Figure 6.10. Comparison of interfacial curvatures in lm thinning systems
with and without a solidication front.
100
CHAPTER 7
Conclusions
This thesis summarized the formulation of the Extended Finite Element Method (X-
FEM), and discussed ways in which this method can be extended to problems with irreg-
ular interface features, such as boundary layers or triple points. The X-FEM is a Finite
Element Method that is modied to include enriched basis functions so that internal
interfaces can be represented by the nite element function space rather than the mesh
geometry. This allows the method to solve problems with moving interfaces on a static
Cartesian mesh, eliminating the need for costly mesh generation steps.
In Chapter 3, the X-FEM was shown to display second order convegence in the solution
and rst order convergence in the solution gradient for conventional problems using the
Heaviside and step enrichment functions. In addition, it was shown that the exponential
enrichment could be used to attain the same convergence order for problems containing
boundary layers that were asymptotically exponential. This custom enrichment function
was then applied to the problem of biolm growth in Chapter 4.
Biolms are microbial communities that grow attached to solid surfaces, and are one
of the most ubiquitous forms of life on the planet. While they are important in a variety
of engineering applications, the factors governing the patterns of their growth are poorly
understood. The X-FEM was used to study the relationship between the geometry of a
biolm colony and its subsequent growth prole. In particular, the commonly observed
tip-splitting and shadowing phenomena were explored. It was shown that the onset of
101
tip-splitting in a colony is primarily governed by the radius of curvature at the tip. An
increase in the penetration depth of the substrate allows for mushroom-shaped growth
at the colony tip when the local curvature radius is small. The existence of a shadowing
eect around a single large colony was conrmed, but it was shown that outside of a nite
region of inuence, the presence of a large colony is virtually undetectable. This could
play a role in the observed spacing of mature biolm colonies in experiments.
Chapter 5 demonstrated how the X-FEM can be used to solve Stokes ow problems,
and showed how to properly implement interface conditions along a free surface. This
technique was then applied to study the evolution of thin lms within a periodic foam
structure. The thinning and rupture of these lms determines the lifespan of the foam, so
an understanding of this process is necessary to understand larger scale lm coarsening.
This system was examined using asymptotic methods in [6], and the lm height was shown
to evolve according to a power law for long times. This long time behavior was conrmed
here numerically. In addition, it was shown that the length of the transition period prior
to the onset of power law thinning is relatively independent of initial conditions and
determined primarily by volume fraction.
Finally, in Chapter 6, a method was proposed for solving three-phase problems con-
taining a triple point. This method properly applies boundary conditions along interface
segments, and allows for two dierent types of conditions to be applied at the triple point.
The rst is an angle condition, and the second is a passive intersection condition. Both
are applied through the choice of subgrid location for the triple point. This method was
then used to solve a simple model for foam solidication. These simulations show the
movement of a solidication front through an evolving foam and the eect this has on
102
the lm thinning. While these simulations show how the front prevents the vertical foam
along the side of the domain from fully evolving, the simplied model is unable to su-
ciently capture the complex physics in the physical problem, and few conclusions can be
drawn about the eect of the front on the horizontal lm at the top of the domain.
While it initially appeared that the choice of a at line to represent the solidication
front would have a signicant eect on the accuracy of the model, the fact that the
evolution of the thin lm is largely driven by the curvature at the edge of the plateau
border indicates that other eects are responsible for the lack of interesting results near
the top of the domain. This suggests two main avenues for future work. The rst is the
development of a better method for extending the front velocity into the gas phase so
that the curvature remains smooth as the front evolves. This would allow the condition
that the solidication front remain at to be relaxed, and a realistic angle condition
could be applied at the triple point. The second is the addition of physical realism to
the governing equations. The heat equation should be made time dependent rather than
assumed to reach steady state at each time step. This would allow for the temperature
eld in the liquid phase to be nonzero with an insulating boundary condition at the top of
the domain. This in turn would add Marangoni eects due to the temperature dependence
of the surface tension along the liquid-gas interface. Because the uid ow in this system
is surface tension driven, the addition of this eect would probably have a larger eect
on the evolution of the lm far from the front than the propagation of the front itself.
While the results obtained with this simplied model are inconclusive in terms of the
physics, they do conrm that the combination of the X-FEM and the level set method
can eectively model multiphase systems with triple junctions.
103
Bibliography
[1] I. Babuska. Error bounds for the nite element method. Numer. Math., pages 322333, 1971.
[2] I. Babuska. The nite element method with lagrangian multipliers. Numer. Math., pages 179192,
1973.
[3] E. Bechet, H. Minnebo, N. Moes, and B. Burgardt. Improved implementation and robustness study
of the x-fem for stress analysis around cracks. International Journal for Numerical Methods in En-
gineering, 45:10331056, 2005.
[4] T. Belytschko and T. Black. Elastic crack growth in nite element with minimal remeshing. Inter-
national Journal of Numerical Methods in Engineering, 45:601620, 1999.
[5] T. Belytschko, N. Moes, S. Usui, and C. Parimi. Arbitrary discontinuities in nite elements. Inter-
national Journal of Numerical Methods in Engineering, 50:9931013, 2001.
[6] L. N. Brush and S. H. Davis. A new law of thinning in foam dynamics. Journal of Fluid Mechanics,
534:227236, 2005.
[7] G. F. Carey and J. T. Oden. Finite Elements: An Introduction, Volume I. Prentice-Hall, Inc., 1981.
[8] G. F. Carey and J. T. Oden. Finite Elements: A Second Course, Volume II. Prentice-Hall, Inc.,
1983.
[9] G. F. Carey and J. T. Oden. Finite Elements: Mathematical Aspects, Volume IV. Prentice-Hall,
Inc., 1983.
[10] G. F. Carey and J. T. Oden. Finite Elements: Computational Aspects, Volume III. Prentice-Hall,
Inc., 1984.
[11] G. F. Carey and J. T. Oden. Finite Elements: Special Problems in Solid Mechanics, Volume V.
Prentice-Hall, Inc., 1984.
104
[12] G. F. Carey and J. T. Oden. Finite Elements: Fluid Mechanics, Volume VI. Prentice-Hall, Inc.,
1986.
[13] W. G. Characklis and K. C. Marshall. Biolms. John Wiley & Sons, Inc., 1990.
[14] J. Chessa and T. Belytschko. An extended nite element method for two-phase uids. Transactions
of the ASME, 70:1017, 2003.
[15] D. L. Chopp. Simulation of biolms using the level set method. In Parametric and Geometric De-
formable Models: An Application in Biomaterials and Medical Imagery. Springer, 2006.
[16] D. L. Chopp, M. J. Kirisits, B. Moran, and M. R. Parsek. The dependence of quorum sensing on
the depth of a growing biolm. Bulletin of Mathematical Biology, 65:10531079, 2003.
[17] C. Daux, N. Moes, J. Dolbow, N. Sukumar, and T. Belytschko. Arbitrary branched and intersecting
cracks with the extended nite element method. International Journal of Numerical Methods in
Engineering, 48:17411760, 2000.
[18] J. Dockery and I. Klapper. Finger formation in biolm layers. SIAM Journal of Applied Mathematics,
62(3):853869, 2001.
[19] J. E. Dolbow, N. Moes, and T. Belytschko. Discontinuous enrichment in nite elements with a
partition of unity method. Finite Elements in Analysis and Design, 36:235260, 2000.
[20] J. E. Dolbow, N. Moes, and T. Belytschko. An extended nite element method for modeling crack
growth with frictional contact. Computational Methods in Applied Mechanics and Engineering,
190:68256846, 2001.
[21] J. Chessa et. al. The extended nite element method (xfem) for solidication problems. International
Journal of Numerical Methods in Engineering, 53:19591977, 2002.
[22] R. P. Fedkiw. Coupling an eulerian uid calculation to a lagrangian solid calculation with the ghost
uid method. Journal of Computational Physics, 175(1):200224, 2002.
[23] R. P. Fedkiw, T. Aslam, B. Merriman, and S. Osher. A non-oscillatory eulerian approach to interfaces
in multimaterial ows (the ghost uid method). Journal of Computational Physics, 152(2):457492,
1999.
105
[24] R. P. Fedkiw, T. Aslam, and S. Xu. The ghost uid method for deagration and detonation discon-
tinuities. Journal of Computational Physics, 154(2):393427, 1999.
[25] F. Gibou, R. Fedkiw, R. Caisch, and S. Osher. A level set approach for the numerical simulation
of dendritic growth. Journal of Scientic Computing, 19(1-3):183199, 2003.
[26] F. Gibou, R. Fedkiw, L. Cheng, and M. Kang. A second order accurate symmetric discretization of
the poisson equation on irregular domains. Journal of Computational Physics, 176:205227, 2002.
[27] P. M. Gresho and R. L. Sani. Incompressible Flow and the Finite Element Method: Volume 1,
Advection-Diusion. John Wiley & Sons Ltd., 1998.
[28] P. M. Gresho and R. L. Sani. Incompressible Flow and the Finite Element Method: Volume 2,
Isothermal Laminar Flow. John Wiley & Sons Ltd., 1998.
[29] H. Ji, D. Chopp, and J. Dolbow. A hybrid extended nite element/level set method for modeling
phase transformation. International Journal of Numerical Methods in Engineering, 54:12091233,
2002.
[30] W. Kurz and D.J. Fisher. Fundamentals of Solidication. Trans Tech Publications, 1986.
[31] O. A. Ladyshenskaya. The Mathematical Theory of Viscous Incompressible Flow. Gordan and Breach
Science Publishers, Inc., 1969.
[32] J. S. Langer. Instabilities and pattern formation in crystal growth. Review of Modern Physics,
52(1):128, 1980.
[33] R. J. LeVeque and Z. Li. The immersed interface method for elliptic equations with discontinuous
coecients and singular sources. SIAM Journal Numerical Analysis, 31:10191044, 1994.
[34] Z. Li. A fast iterative algorithm for elliptic interface problems. SIAM Journal of Numerical Analysis,
35:230254, 1998.
[35] Z. Li. An overview of the immersed interface method and its applications. Taiwanese Journal of
Mathematics, 7:149, 2003.
[36] X. . Liu, R. P. Fedkiw, and M. Kang. A boundary condition capturing method for poissons equation
on irregular domains. Journal of Computational Physics, 160(1):151178, 2000.
106
[37] J. M. Melenk and I. Babuska. The partition of unity nite element method: Basic theory and
applications. Computational Methods in Applied Mechanics and Engineering, 139:289314, 1996.
[38] N. Moes, J. Dolbow, and T. Belytschko. A nite element method for crack growth without remeshing.
International Journal of Numerical Methods in Engineering, 46:131150, 1999.
[39] Y. Mori and C. S. Peskin. Implicit second-order immersed boundary methods with boundary mass.
Computer Methods in Applied Mechanics and Engineering, 197:20492067, 2008.
[40] S. Osher and J. A. Sethian. Fronts propagating with curvature-dependent speed: Algorithms based
on hamilton-jacobi formulations. Journal of Computational Physics, 79:1249, 1988.
[41] C. Peskin and B. Printz. Improved volume conservation in the computation of ows with immersed
boundaries. Journal of Computational Physics, 105:3346, 1993.
[42] C. S. Peskin. Numerical analysis of blood ow in the heart. Journal of Computational Physics,
25:220252, 1977.
[43] C. S. Peskin. Lectures on mathematical aspects of physiology. Lectures in Applied Mathematics,
19:69107, 1981.
[44] H. T. Rathod, K. V. Nagaraja, B. Venkatesudu, and N. Ramesh. Gauss-legendre quadrature over a
triangle. Journal of the Indian Institute of Science, 84:183188, 2004.
[45] O. Schenk and K. Gartner. Solving unsymmetric sparse systems of linear equations with pardiso.
Journal of Future Generation Computer Systems, 20:465487, 2004.
[46] B. G. Smith, B. L. Vaughan, and D. L. Chopp. The extended nite element method for boundary layer
problems in biolm growth. Communications in Applied Mathematics and Computational Science,
2(1):3556, 2007.
[47] K. A. Smith, F. J. Solis, and D. L. Chopp. A projection method for motion of triple junctions by
level sets. Interfaces and Free Boundaries, 4:263276, 2002.
[48] M. Stolarska, D. L. Chopp, N. Moes, and T. Belytschko. Modelling crack growth by level sets in
the extended nite element method. International Journal of Numerical Methods in Engineering,
51:943960, 2001.
107
[49] N. Sukumar, D. Chopp, N. Moes, and T. Belytschko. Modeling holes and inclusions by level sets
in the extended nite element method. International Journal of Numerical Methods in Engineering,
48:15491570, 2000.
[50] N. Sukumar, D. L. Chopp, and B. Moran. The extended nite element method and fast marching
method for three-dimensional fatigue crack propagation. Engineering Fracture Mechanics, 70:2948,
2003.
[51] B. L. Vaughan, B. G. Smith, and D. L. Chopp. A comparison of the extended nite element method
for elliptic equations with discontinuous coecients and singular sources. Communications in Applied
Mathematics and Computational Science, 1(1):207228, 2006.
[52] G. Zi and T. Belytschko. New crack-tip elements for xfem and applications to cohesive cracks.
Internation Journal for Numerical Methods in Engineering, 57:22212240, 2003.

3331158

Caricato da

Informazioni sul documento

Descrizione originale:

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

3331158

Caricato da

Copyright:

Formati disponibili

NORTHWESTERN UNIVERSITY

Figure 2.8. An element containing an interface segment is partitioned into

(0) indicates that the enrichment function is evaluated according to

Figure 5.9. Interface curvature of an evolving lm. The progression of the

System with no Solidification

System with Solidification

Potrebbero piacerti anche