Professional Documents
Culture Documents
jibum_thesis
jibum_thesis
A Dissertation in
by
Jibum Kim
Doctor of Philosophy
December 2012
The dissertation of Jibum Kim was reviewed and approved∗ by the following:
Suzanne M. Shontz
Adjunct Assistant Professor of Computer Science and Engineering
Dissertation Advisor
Chair of Committee
Padma Raghavan
Distinguished Professor of Computer Science and Engineering
Professor of Information Sciences and Technology
Director, Institute for CyberScience
Jesse Barlow
Professor of Computer Science and Engineering
Professor of Statistics
Qiang Du
Verne M. Willaman Professor of Mathematics
Professor of Materials Science and Engineering
Raj Acharya
Head and Professor of the Computer Science and Engineering
Abstract
High quality meshes are important for the accuracy, stability, and efficiency of numerical
techniques in computational simulations involving partial differential equations (PDEs), nonlo-
cal peridynamics, mesh deformation, or shape matching. Several applications involving these
mesh-based computational techniques include ocean dynamics (PDEs), surface cracks (nonlocal
peridynamics), hydrocephalus (mesh deformation), and object recognition (shape matching).
The first part of the dissertation explores the best combinations of mesh quality metrics,
preconditioners, and sparse linear solvers for solving various elliptic PDEs, multiobjective mesh
optimization methods, and the effect of mesh anisotropy, mesh refinement, and kernel functions
on the conditioning of nonlocal peridynamics models. Engineers use various mesh quality im-
provement methods for solving PDEs to improve the efficiency and accuracy as well as various
types of preconditioners and sparse linear solvers for various PDE problems. However, little
research has been performed with respect to choosing the most efficient combinations of these
three factors for solving various elliptic PDE problems. First, we investigate the effect of choos-
ing various combinations of mesh quality metrics, preconditioners, and sparse linear solvers on
the numerical solution of elliptic PDEs. Many PDE-based engineering and scientific applica-
tions have multiple requirements for the finite element mesh discretizing the geometric domain;
however, most traditional mesh optimization algorithms improve only one aspect of the mesh,
Second, we propose a multiobjective mesh optimization framework for simultaneous mesh qual-
ity improvement and mesh untangling for PDE-based applications for optimizing two or more
aspects of the mesh. Recently, a new paradigm called nonlocal peridynamics, which employs
integral equations, was proposed to model discontinuous domains. Third, we investigate the ef-
fect of mesh anisotropy, mesh refinement, and kernel functions on the conditioning of the global
stiffness matrix for a nonlocal peridynamic model.
The second part of the dissertation studies mesh deformation algorithms for robust anisot-
ropic mesh deformation and for shape matching. First, we propose a robust mesh deformation
algorithm using the anisotropy of the boundary deformation and multiobjective mesh optimiza-
tion. When mesh deformation occurs, it is challenging to preserve element shape and nonin-
verted mesh elements. In order to achieve this on the deformed domain, we use the direction
of the boundary deformation to estimate the interior vertex positions and employ multiobjective
mesh optimization for simultaneously preserving element shape and untangling the mesh.
iv
Second, we propose an improved shape matching algorithm for deformable objects mod-
eled by triangular meshes. We use dynamic programming to find the optimal mapping from the
source image to the target image.
i
Table of Contents
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv
Chapter 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Chapter 4. The Effect of Anisotropy, Mesh Refinement, and Kernel Functions on the
Conditioning of the Stiffness Matrix for Nonlocal Peridynamic Models . . 105
4.1 Introductions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.2 A Bond-based Nonlocal Peridynamic Model . . . . . . . . . . . . . . . . . 108
4.3 Quadrature Rules and Basis Functions . . . . . . . . . . . . . . . . . . . . 113
4.3.1 Quadrature rules for a triangle . . . . . . . . . . . . . . . . . . . . 113
4.3.2 Basis functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
4.4 Connections among the horizon, mesh refinement, anisotropy of the mesh
element and the condition number of the global stiffness matrix . . . . . . . 116
4.4.1 Condition number of the global stiffness matrix for a nonlocal peri-
dynamic model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
4.4.2 Condition number of the global stiffness matrix for general second-
order elliptic PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . 118
4.5 Numerical results on 2D rectangular meshes . . . . . . . . . . . . . . . . . 118
4.5.1 Piecewise constant basis function with an integrable kernel function 119
4.5.2 Piecewise linear basis function with an integrable kernel function . 125
4.5.3 Piecewise linear basis function with the nonintegrable kernel func-
tion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
4.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
4.7 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
4.8 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
4.8.1 FEMWARP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
4.8.2 FEMWARP and anisotropic boundary deformation . . . . . . . . . 143
4.9 Hybrid Mesh Deformation Algorithm . . . . . . . . . . . . . . . . . . . . 144
4.9.1 Step 1: Anisotropic FEMWARP using anisotropic PDE coefficients 145
4.9.2 Step 2: Multiobjective mesh optimization with shape and untangling 156
4.10 Numerical Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
iii
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
iv
List of Tables
2.1 Notation used in the definition of the 2D mesh quality metrics in Table 2.2. . . 16
2.2 The mesh quality metric definitions. . . . . . . . . . . . . . . . . . . . . . . . 16
2.3 Listing of numerical experiments and examples of PDE problems. The letters
(a) through (c) are representative examples of the three types of PDE problems
under consideration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.4 Properties of meshes on geometric domains. Initial angle distributions for the
meshes are given in Table 2.7. . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.5 The sixteen combinations of preconditioners and solvers. For example, 10
refers to using the ILU(0) preconditioner with the GMRES solver. . . . . . . . 25
2.6 Mesh smoothing time (sec) for various mesh quality metrics . . . . . . . . . . 26
2.7 Angle (θ ) distribution for various mesh quality metrics. The reported values
indicate a percentage of angles in the mesh. The initial mesh is the mesh with
50% of the interior vertices perturbed. . . . . . . . . . . . . . . . . . . . . . . 33
2.8 Linear solver time (secs) and number of iterations required to converge for Pois-
son’s equation (problem (A) in Table 2.3) as a function of mesh quality metric
for the 16 preconditioner-solver combinations (see Table 2.5) on the wrench
and hinge domains. A ’*’ denotes failure. For each quality metric, the num-
bers in the top and bottom rows represent the linear solver time and number of
iterations to convergence, respectively. . . . . . . . . . . . . . . . . . . . . . . 34
2.9 Linear solver time (secs) and number of iterations required to converge for Pois-
son’s equation (problem (A) in Table 2.3) as a function of vertex perturbation
for the 16 preconditioner-solver combinations (see Table 2.5) on the two ge-
ometric domains. A ’*’ denotes failure. For each percentage of vertices per-
turbed, the numbers in the top and bottom rows represent the linear solver time
and number of iterations to convergence, respectively. . . . . . . . . . . . . . . 36
2.10 Linear solver time (secs) and number of iterations required to converge for gen-
eral second-order elliptic PDEs (problem (B) in Table 2.3) as a function of mesh
quality metric for the 16 preconditioner-solver combinations (see Table 2.5) on
the two geometric domains. A ’*’ denotes failure. For each quality metric, the
numbers in the top and bottom rows represent the linear solver time and number
of iterations to convergence, respectively. . . . . . . . . . . . . . . . . . . . . 43
v
2.11 Linear solver time (secs) and number of iterations required to converge for the
linear elasticity problem (problem (C) in Table 2.3) as a function of mesh qual-
ity metric for the 16 preconditioner-solver combinations (see Table 2.5) on the
two geometric domains. A ’*’ denotes failure. For each quality metric, the
numbers in the top and bottom rows represent the linear solver time and num-
ber of iterations to convergence, respectively. . . . . . . . . . . . . . . . . . . 50
2.12 Linear solver time (secs) and number of iterations required to converge for the
linear elasticity problem (problem (C) in Table 2.3) as a function of vertex
perturbation for the 16 preconditioner-solver combinations (see Table 2.5) on
the two geometric domains. A ’*’ denotes failure. For each percentage of
vertices perturbed, the numbers in the top and bottom rows represent the linear
solver time and number of iterations to convergence, respectively. . . . . . . . 52
4.1 Cond(A) of the global stiffness matrix for piecewise constant basis functions
with an integrable basis kernel function (p=1). The number of elements is the
number of elements in Ω. Here, the ratio is the ratio of the condition number of
the current level of mesh refinement compared with that of the previous level
of mesh refinement for a fixed δ . Therefore, these ratios are not defined for the
200 elements and are denoted by ‘*’ for these cases. . . . . . . . . . . . . . . 122
vi
4.2 Cond(A) of a global stiffness matrix for piecewise linear basis functions with an
integrable basis kernel function (p=1). The total number of elements changes
with respect to the horizon, δ . The number of elements indicates a number of
elements in Ω. Here, the ratio is defined as the ratio of the condition number of
the current level of mesh refinement compared with that of the previous level
of mesh refinement for a fixed δ . Therefore, these ratios are not defined for the
200 elements and are denoted as ‘*’ for these cases. . . . . . . . . . . . . . . 126
4.3 Cond(A) of a stiffness matrix for piecewise linear basis functions with a non-
integrable basis kernel function (p=1). The number of elements indicates the
number of elements in Ω. Here, the ratio is defined as the ratio of the condi-
tion number of the current level of mesh refinement compared with that of the
previous level of mesh refinement for a fixed δ . Therefore, these ratios are not
defined for the 200 element case and are denoted as ‘*’ for such cases. . . . . 131
List of Figures
1.1 Output meshes (a) before applying mesh qualiy improvement and (b) after ap-
plying mesh quality improvement . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 One example of a tangled mesh which includes inverted elements. . . . . . . . 3
1.3 Fragmentation of the domain [55]. Discontinuous domains occur due to frag-
mentation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Bending bar deformation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5 Some of the binary images in the Brown dataset [108]. . . . . . . . . . . . . . 6
1.6 This flowchart represents the two parts of the dissertation and the dependency
among the chapters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1 Contour plots of the quality metric of a triangle as a function of a free vertex
when the two other vertices held fixed at (0, 0) and (0, 1). . . . . . . . . . . . . 17
2.2 Coarse initial meshes on the wrench and hinge geometric domains indicative of
the actual meshes to be smoothed. . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3 Effect of the mesh size on the number of iterations (a) and solver time (b) to
convergence as a function of GMRES restart value, m, for Poisson’s equation
(problem (A) in Table 2.3) on the wrench domain. . . . . . . . . . . . . . . . . 30
2.4 Percentage increase (PI) as a function of the solver time for different combi-
nations of preconditioners and solvers for Poisson’s equation (problem (A) in
Table 2.3). Preconditioner-solver combinations which fail to generate a pre-
conditioner or do not converge correspond to the missing bars in these figures.
Note that PI values for the hinge domain are significantly greater than those for
the wrench domain. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.5 The order of convergence for the solver time based on the number of iterations
for various combinations of preconditioners and solvers for Poisson’s equation
(problem (A) in Table 2.3). Meshes with 10K, 50K, 100K, 200K, and 500K
vertices on each domain are employed in computing s using the linear least-
squares method. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.6 Similar to Figure 2.5, this figure displays the order of convergence based on
the solver time for Poisson’s equation (problem (A) in Table 2.3). Meshes with
10K, 50K, 100K, 200K, and 500K vertices on each domain are employed in
computing s using the linear least-squares method. . . . . . . . . . . . . . . . 40
viii
3.1 Contour plots of the quality metric of a triangle as a function of a free vertex
when the two vertices are fixed at (0, 0) and (1, 0). In (e), β = 0.25. . . . . . . 66
3.2 Comparison among the exponential sum, objective product, and equal sum mul-
tiobjective mesh optimization methods . . . . . . . . . . . . . . . . . . . . . 76
3.3 Coarse initial meshes on the disk and barrier geometric domains indicative of
the actual meshes to be smoothed. . . . . . . . . . . . . . . . . . . . . . . . . 78
3.4 (a) Average element quality in terms of the shape metric (IMR) on the 3D disk
mesh; (b) average element quality in terms of the size metric (volume) on the
same mesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.5 (a) Worst element quality in terms of the shape metric (IMR) on the 3D disk
mesh; (b) worst element quality in terms of the size metric (volume) on the
same mesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.6 (a) Average element quality as measured by the shape metric (IMR) on the 3D
disk mesh; (b) average element quality as measured by the interpolation error
(SS) on the same mesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.7 (a) Worst element quality in terms of the shape metric (IMR) on the 3D disk
mesh; (b) worst element quality in terms of the interpolation error (SS) metric
on the same mesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
ix
3.8 2D barrier mesh: (a) the initial mesh with inverted elements (the elements cir-
cled in red); (b) the final mesh with inverted elements (the elements circled
in red) when employing the SI metric with the single objective method. The
number of inverted elements increases after mesh optimization; (c) the final
mesh when employing the untangling metric with the single objective method;
(d) the final mesh for the exponential sum multiobjective method; (e) the final
mesh for the objective product multiobjective method; (f) the final mesh for the
equal sum multiobjective method. The final mesh by employing the exponential
sum multiobjective method yields the best element quality while eliminating all
inverted elements. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
3.9 (a) Average element quality in terms of the shape metric (IMR) on the 3D disk
mesh; (b) worst element quality in terms of the size metric (volume) on the
same mesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.10 (a) Average element quality in terms of the shape metric (SI) on the 3D disk
mesh; (b) Average element quality in terms of the size (volume) metric on the
same mesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.11 Segmented images and meshes of initial and deformed domains. . . . . . . . . 98
3.12 Meshes of initial and deformed domains. . . . . . . . . . . . . . . . . . . . . . 100
4.1 Connection between horizon (δ ), force ( f ) between v and v0 , inside the neigh-
borhood, H, over the domain Ω. The point v does not interact with any points
beyond the distance δ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
4.2 One example of the sparsity pattern of a matrix A for the nonlocal peridynamic
model. Here nz is the number of nonzeros in A. . . . . . . . . . . . . . . . . . 111
4.3 Boundary conditions for (a) classical PDE models and (b) nonlocal peridy-
namic models. For classical PDE problems, the boundary condition is only
specified on the boundary points. For nonlocal peridynamic models, all mesh
elements belonging to BΩ belong to the boundary condition. . . . . . . . . . . 112
4.4 Three different cases for computing |T | when the intersection of T with δ is
computed. Here, we only compute the intersection areas inside the horizon. . . 114
4.5 The approximation of the area of intersection of the triangular element and the
horizon for case 2 in Fig. 4.4. . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
4.6 (a) Initial isotropic mesh with h=0.1 and δ =0.2; (b) refined mesh with h=0.05
and δ =0.2; (c) Two level-refined mesh with h=0.025 and δ =0.2 . . . . . . . . . 120
4.7 Anisotropic mesh (aspect ratio=4) with δ =0.2. . . . . . . . . . . . . . . . . . . 121
4.8 The definition of the aspect ratio of a triangle. The aspect ratio is defined as
W /w [75]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
4.9 Condition number of A for fixed h and varying anisotropy of the elements and δ . 123
4.10 Condition number of A as a function of δ −2 for fixed h and varying anisotropy
of the elements and δ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
4.11 Condition number of A for fixed anisotropy and varying h and δ . . . . . . . . . 124
4.12 Condition number of A as a function of δ −2 for fixed anisotropy and varying h
and δ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.13 Condition number of A for fixed h and varying anisotropy of the elements and δ . 127
4.14 Condition number of A for fixed h and varying anisotropy of the elements and δ . 128
x
4.15 Condition number of A for fixed anisotropy and varying h and δ . . . . . . . . . 129
4.16 Condition number of A for fixed anisotropy and varying h and δ . . . . . . . . . 130
4.17 Condition number of A when h=0.05 for fixed h and varying anisotropy and δ . 133
4.18 Condition number of A for fixed anisotropy (aspect ratio=1 and 4) and varying
h and δ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
4.19 Condition number of A for fixed h and anisotropy and various power of a kernel
function, p. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
4.20 (a) Initial mesh on the rectangular domain. (b) Deformed mesh using FEMWARP.
This mesh has 20 inverted elements. . . . . . . . . . . . . . . . . . . . . . . . 145
4.21 Generic element in structured triangular mesh with uniform elements. . . . . . 147
4.22 (a) Deformed mesh using FEMWARP. This mesh has 20 inverted elements. (b)
Deformed mesh using anisotropic FEMWARP with coefficients α=0 and β =1
in (4.11). This deformed mesh using anisotropic FEMWARP does not have any
inverted elements. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
4.23 Connection between the PDE coefficients (α and β ) and the direction of mo-
tion. If the deformation occurs (a) aligned with the x-axis or (b) aligned with
the y-axis, we only consider neighbors aligned with the same axis. Here, the
cross-out indicates that we do not consider those neighbors. If deformation
occurs that is not aligned with either the x or y axes (c), we use the angle of
direction of the deformation to choose the appropriate PDE coefficients. Here,
a and b are α and β in (4.19), respectively. . . . . . . . . . . . . . . . . . . . 155
4.24 Moving cylinder domain for anisotropic boundary deformation aligned with the
x-axis: (a) Initial mesh for a moving cylinder in a channel. (b) Deformed mesh
using FEMWARP for a moving cylinder. This mesh has 38 inverted elements.
(c) Deformed mesh using anisotropic FEMWARP for a moving cylinder. This
mesh has zero inverted elements. (d) Optimized mesh on the deformed domain
with no inverted elements using anisotropic FEMWARP followed by multiob-
jective mesh optimization with TMP shape and untangling. . . . . . . . . . . . 161
4.25 Mesh quality measured by the TMP shape metric on the deformed cylinder
domain. A smaller value indicates more similarity with zero meaning iden-
tical shape. The worst element quality of the deformed mesh which results
from our hybrid algorithm is 80.4% less than the quantity of the one using
UBN [95]. Note that FEMWARP includes 38 inverted elements, while the de-
formed meshes resulting from both our hybrid algorithm and UBN do not have
any inverted elements. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
4.26 Number of inverted elements with respect to the translation of inner bound-
ary after applying (a) FEMWARP (α=1 and β =1) and anisotropic FEMWARP
(α=1 and β =0). The inner boundary shifts right. (b) Time to untangle inverted
elements using multiobjective mesh optimization. This figure shows the time
to untangle inverted meshes using multiobjective mesh optimization. . . . . . . 162
xi
4.27 Moving bar domain for anisotropic boundary deformation aligned in y-axis:(a)
Initial mesh and (b) zoomed-in on the bar domain. (c) Deformed mesh with
FEMWARP and (d) zoomed in mesh with FEMWARP. Deformed mesh with
FEMWARP has 118 inverted elements. (e) Deformed mesh using anisotropic
FEMWARP and (f) zoomed-in mesh with anisotropic FEMWARP. The de-
formed mesh with anisotropic FEMWARP has 17 inverted elements. (g) Op-
timized mesh and (h) zoomed-in mesh with anisotropic FEMWARP followed
by multiobjective mesh optimization. This optimized mesh does not have any
inverted elements. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
4.28 Mesh quality measured by the TMP shape metric on the deformed bar domain.
The smaller value indicates more similarity with zero meaning identical shape.
The worst element quality of the deformed mesh which results from our hybrid
algorithm is 93.5% less than the quality of the mesh obtained from the use of
FEMWARP [95]. Note that FEMWARP includes 118 inverted elements while
the deformed mesh resulting from our hybrid algorithm does not have any in-
verted elements. For this domain, UBN fails to find the deformed mesh due to
numerical tolerance issues. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
4.29 Number of inverted elements with respect to the translation of inner bound-
ary after applying (a) FEMWARP (α=1 and β =1) and anisotropic FEMWARP
(α=0 and β =1). The inner bar moves down. (b) We apply multiobjective mesh
optimization to simultaneously smooth and untangle inverted elements. This
figure shows timing to untangle inverted elements using multiobjective mesh
optimization. Our multiobjective mesh optimization is able to improve ele-
ment qualities while eliminating inverted elements but anisotropic FEMWARP
takes up to 63.7% less time to eliminate inverted elements compared with
FEMWARP since anisotropic FEMWARP has fewer iterations to untangle in-
verted elements than FEMWARP.. . . . . . . . . . . . . . . . . . . . . . . . . 166
4.30 Moving gate domain for anisotropic boundary deformation aligned in y-axis:
(a) Initial mesh and (b) zoomed-in on the gate domain. (c) The deformed mesh
with FEMWARP and (d) zoomed-in mesh with FEMWARP. Deformed mesh
with FEMWARP has 103 inverted elements. (e) The deformed mesh using
anisotropic FEMWARP and (f) zoomed-in mesh with anisotropic FEMWARP.
The deformed mesh with anisotropic FEMWARP has 14 inverted elements.
(g) The optimized mesh and (h) zoomed-in optimized mesh on the deformed
domain with no inverted elements using anisotropic FEMWARP followed by
multiobjective mesh optimization with TMP shape and untangling. . . . . . . . 167
4.31 Mesh quality measured by the TMP shape metric on the deformed gate domain.
The smaller value indicates more similarity with zero meaning identical shape.
The worst element quality of the deformed mesh which results from our hy-
brid algorithm is 9.85% less than the one using FEMWARP [95]. Note that
FEMWARP includes 103 inverted elements while the deformed mesh result-
ing from our hybrid algorithm does not have any inverted elements. For this
domain, UBN fails to find the deformed mesh due to numerical tolerance issues. 168
xii
4.32 Number of inverted elements with respect to the translation of outer bound-
ary after applying (a) FEMWARP (α=1 and β =1) and anisotropic FEMWARP
(α=0 and β =1). The corner in the middle moves down. (b) We apply mul-
tiobjective mesh optimization to simultaneously smooth and untangle inverted
elements. This figure shows the time to untangle inverted elements using mul-
tiobjective mesh optimization. Our multiobjective mesh optimization is able to
improve element qualities while eliminating inverted elements but anisotropic
FEMWARP takes up to 90.6% less time to eliminate inverted elements com-
pared with FEMWARP. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
4.33 Moving cylinder domain for anisotropic diagonal boundary deformation: (a)
Initial mesh for a moving cylinder in a channel. (b) The deformed mesh using
FEMWARP for a diagonal deformation. It has 101 inverted elements. (c) The
deformed mesh using anisotropic FEMWARP for a diagonal deformation. It
has 15 inverted elements. (d) The optimized mesh on the deformed domain with
no inverted elements using anisotropic FEMWARP followed by multiobjective
mesh optimization with TMP shape and untangling. . . . . . . . . . . . . . . . 170
4.34 Mesh quality measured by the TMP shape metric on the deformed cylinder do-
main. The smaller value indicates more similarity with zero meaning identical
shape. The worst element quality of the deformed mesh which results from our
hybrid algorithm is 79.2% less than the one using FEMWARP [95]. Note that
FEMWARP includes 103 inverted elements while the deformed mesh resulting
from our hybrid algorithm does not have any inverted elements. . . . . . . . . 171
4.35 Bending bar domain for anisotropic boundary deformation:(a) Initial mesh and
(b) zoomed-in mesh on the bar domain. (c) Deformed mesh with FEMWARP
and (d) zoomed-in mesh with FEMWARP. Deformed mesh with FEMWARP
has 159 inverted elements. (e) Deformed mesh using anisotropic FEMWARP
and (f) zoomed-in mesh with anisotropic FEMWARP. Deformed mesh with
anisotropic FEMWARP has 48 inverted elements. (g) Optimized mesh and (h)
zoomed-in mesh with anisotropic FEMWARP followed by multiobjective mesh
optimization. This optimized mesh does not have any inverted elements. . . . . 172
4.36 Mesh quality measured by the TMP shape metric on the deformed bar domain.
The smaller value indicates more similarity with zero meaning identical shape.
The worst element quality of the deformed mesh which results from our hy-
brid algorithm is 92.5% less than the one using FEMWARP [95]. Note that
FEMWARP includes 159 inverted elements while the deformed mesh result-
ing from our hybrid algorithm does not have any inverted elements. For this
domain, UBN fails to find the deformed mesh due to numerical tolerance issues. 173
4.37 Moving sphere in a 3D cube domain: (a) Surface mesh on the initial domain
(b) Volume mesh on the initial domain (c) Optimized volume mesh using our
hybrid algorithm on the deformed domain. This optimized mesh does not have
any inverted elements. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
xiii
4.38 Mesh quality measured by the TMP shape metric on the deformed 3D sphere
domain. The smaller value indicates more similarity with zero meaning identi-
cal shape. The worst element quality of the deformed mesh which results from
our hybrid algorithm is 86.2% less than the one using FEMWARP [95]. Note
that FEMWARP includes 120 inverted elements while the deformed mesh re-
sulting from our hybrid algorithm does not have any inverted elements. For this
domain, UBN fails to find the deformed mesh due to numerical tolerance issues. 175
4.39 The number of inverted elements with respect to the translation of outer bound-
ary after applying (a) FEMWARP (α=1, β =1, and γ=1) and anisotropic FEMWARP
(α=0, β =0, γ=1). (b) We apply multiobjective mesh optimization to simulta-
neously smooth and untangle inverted elements. This figure shows the time to
untangle inverted elements using multiobjective mesh optimization. Our mul-
tiobjective mesh optimization is able to improve element qualities while elim-
inating inverted elements but anisotropic FEMWARP takes up to 71.1% less
time to eliminate inverted elements compared with FEMWARP. . . . . . . . . 175
5.1 Overview of the shape matching process. The function f maps triangles in the
triangular mesh on the source image to a triangular mesh on the target image.
(a) Equally-spaced boundary vertices are generated. (b) The triangular mesh is
created. (c) The detected image is illustrated on the target image. . . . . . . . 180
5.2 (a) A sample image shape from the Brown dataset [108]. (b) The dot in the
middle represents the center of mass for the image (this is denoted as C in the
target image), and the arrow represents the maximum distance from the center
of mass in the image to the boundary vertices (i.e., dmax,s or dmax,t ) . . . . . . . 182
5.3 Poor shape matching result of the algorithm in [100] (left) and improved match-
ing result using the center of mass in the target image (right). For these fig-
ures, the detected mesh is illustrated on the target image. For this experiment,
θ = dmax,t /3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
5.4 Sample images in the Brown dataset [108]. Three sample images are shown per
category. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
5.5 Good shape matching results for the Brown dataset on three source (query)
images and comparison with [100]. For each source image, the 10 best match-
ing results are shown with the smallest (left) to the largest (right) energy. The
top figures in each group represent the matching results obtained from our al-
gorithm, whereas the bottom figures in each group represent matching results
using the algorithm in [100]. For these experimental sets, only two matching
results of [100] (i.e., the bottom right images) fail to match. . . . . . . . . . . . 186
5.6 Poor shape matching results for the Brown dataset on three source (query) im-
ages and comparison with [100]. For each source image, the 10 best matching
results are shown with the smallest (left) to the largest (right) energy. The top
figures in each group represent the matching results obtained from our algo-
rithm, and the bottom figures in each group represent matching results using the
algorithm in [100]. For this experimental data set, both our algorithm and [100]
show poor matching results. However, our algorithm shows better matching
results than does the method in [100]. . . . . . . . . . . . . . . . . . . . . . . 186
xiv
5.7 Example matching results including detected meshes on the target image using
images from the Brown dataset [108]. For this experiment, Felzenszwalb’s
algorithm shows poor matching results because triangles are placed at poor
positions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
xv
Acknowledgments
for Applied Scientific Computing at the Lawrence Livermore National Laboratory, and through
instrumentation funded by the NSF through grant OCI-0821527.
1
Chapter 1
Introduction
mesh. It is well-known that mesh element shapes and mesh qualities affect both the efficiency
and accuracy for solving many scientific applications, especially partial differential equation
(PDE)-based applications [4, 5, 41]. Figure 1.1 shows a simple example of meshes (a) before
and (b) after mesh quality improvement. A mesh is said to be tangled if it has inverted elements
which have negative orientations. Figure 1.2 show an example of a tangled mesh which includes
inverted elements. Tangled meshes are often generated during mesh generation, mesh optimiza-
tion, and mesh deformation. Meshes with noninverted elements are important for PDE-based
applications since tangled meshes with inverted elements result in invalid PDE solutions [1].
Recently, a new finite element method to handle tangled meshes was proposed [2]. However,
this method can result in a system with a large condition number when the stiffness matrix is
close to singular [2]. Also, there is no guarantee that this method is able to handle highly tangled
meshes. In any case, it is desirable to eliminate inverted elements or to minimize the number of
improve the mesh qualities and to perform mesh deformations for various problems such as
PDEs, nonlocal peridynamics problems, mesh deformation problems, and shape matching prob-
lems.
for mesh quality improvements and mesh untangling. The accuracy and the efficiency of PDE
solution highly depends on the quality of the mesh, preconditioner, and sparse linear solver. In
Chapter 2, we study the effect the choice of mesh quality metric, preconditioner, and sparse
linear solver have on the numerical solution of elliptic PDEs. We smooth meshes on several
geometric domains using various quality metrics and solve the associated elliptic PDEs using
the finite element method. The resulting linear systems are solved using various combinations
of preconditioners and sparse linear solvers. We also investigate the effect of vertex perturbation
and the effect of increasing the problem size on the number of iterations required to converge
and on the solver time. We consider various elliptic PDE problems such as Poisson’s equation,
Many PDE-based engineering and scientific applications have multiple requirements for
the finite element mesh discretizing the geometric domain [41, 44, 45]. For example, such
requirements may include having non-inverted mesh elements, elements that are well-shaped,
elements with uniform element size, and/or elements which yield small PDE interpolation error.
Despite there being multiple mesh requirements for various PDE applications, most traditional
mesh optimization algorithms optimize only a single objective function and hence improve only
one aspect of the mesh. In Chapter 3, we propose a multiobjective mesh optimization framework
for use in simultaneous improvement of multiple aspects of the mesh with the goal of improving
3
Fig. 1.1 Output meshes (a) before applying mesh qualiy improvement and (b) after applying
mesh quality improvement
Fig. 1.2 One example of a tangled mesh which includes inverted elements.
4
the accuracy, efficiency, stability, and conditioning of the associated finite element solver. Our
framework combines two or more competing objective functions into a single objective function
to be solved using one of various multiobjective optimization methods. Methods within our
framework are able to optimize various aspects of the mesh such as the element shape, element
size, associated PDE interpolation error, and number of inverted elements, but the improvement
Many applications in solid mechanics such as damage, surface cracks, and fracture in-
clude discontinuities [56, 63, 64]. Figure 1.3 shows one example of a cracked domain which
includes discontinuities [55]. However, classical PDEs have some limitations in modeling and
computation of these discontinuous domains since we are not able to compute derivatives in
these domains. In order to improve these limitations of classical PDEs, a peridynamics nonlocal
model using integral equations was recently developed. In Chapter 4, we use the continuous
Galerkin finite element methods for discretizing a linear peridynamics system and study the ef-
fect the mesh anisotropy, mesh refinement, and kernel functions have on the conditioning of the
Fig. 1.3 Fragmentation of the domain [55]. Discontinuous domains occur due to fragmentation.
5
In the second part of the dissertation, we investigate meshing techniques for mesh de-
formation problems and shape matching problems. There are numerous applications where
discretized geometric domains vary with respect to time such as the solution of Arbitrary-
Lagrangian-Eulerian (ALE) flow simulations [76], deformation of the human face in computer
graphics [78], deformation of sequences of medical images [81, 82], and deformations of biomed-
ical applications [79, 80]. Figure 1.4 shows an example of a bending bar deformation. In Chapter
5, we propose a hybrid mesh deformation algorithm which uses the direction of the boundary
deformation. Our goal is to produce meshes on deformed domains which maintains element
shapes and possesses no inverted elements. The hybrid mesh deformation algorithm consists of
two steps, i.e., anisotropic finite element-based mesh warping (FEMWARP) followed by multi-
objective mesh optimization. The idea of multiobjective mesh optimization comes from Chapter
3, but we extend this idea to mesh deformation problems. The first step estimates the interior ver-
tex positions on the deformed mesh using the boundary deformation to choose appropriate PDE
coefficients in the anisotropic FEMWARP method. As a second step, we find a locally optimal
mesh with no inverted elements on the deformed domain by employing a multiobejctive mesh
optimization with one term controlling element shape and a second term designed to untangle
inverted elements.
object tracking and image-based searches [98, 99]. The goal of shape matching is to match
the source image to the target image, i.e., the deformed image. Figure 1.5 shows some binary
images which are frequently used for testing shape matching algorithms in the Brown dataset.
This image dataset includes various transformations and occlusions of shapes. In Chapter 6, we
propose a shape matching algorithm for deformable objects using both local and global shape
6
(a) Mesh on the initial domain (b) Mesh on the deformed domain
information. We focus on shape matching problems for binary images. For our approach, we
use triangular meshes to represent deformable objects and use dynamic programming to find
the optimal mapping from the source image to the target image which minimizes a new energy
function. Our energy function includes a new cost term that takes into account the center of mass
of an image. We use the well-known Brown dataset shown in Fig. 1.5 to test our algorithm.
Fig. 1.5 Some of the binary images in the Brown dataset [108].
Our goal is to improve the accuracy and efficiency for the solution of various applications
by employing our optimization-based meshing techniques. We will show that our optimized-
based meshing techniques significantly improve the accuracy and efficiency for the solution of
various problems: PDEs, nonlocal peridynamics, mesh deformation, and shape matching.
7
In the following paragraph, we summarize the software we use for our numerical experi-
ments. In Chapter 2, we use Mesquite [16], a mesh quality improvement toolkit, and PETSc [17],
a linear solver toolkit, to perform a numerical study investigating the performance of several
mesh quality metrics, preconditioners, and sparse linear solvers on the solution of various ellip-
tic PDEs. We use Mesquite and PETSc in their native state with the default parameters. Only
these two toolkits are employed so that differences in solver implementations, data structures,
and other such factors would not influence the results. In Chapter 3, we use Mesquite for our
also use PETSc to solve a sparse linear system, which is discretized from Poisson’s equation for
real-world applications. We use the default options of PETSc to compute the maximum eigen-
value, minimum eigenvalue, condition number of the stiffness matrix, and to solve the resulting
linear system. The Jacobi preconditioner (P) with the minimal residual (MINRES) solver is used
to solve the linear system. In Chapter 4, no existing software is used, but instead we use C++
warping algorithm (FEMWARP) [88] and Mesquite for our numerical experiments. We modify
the FEMWARP code, which was provided by Dr. Suzanne Shontz, to implement the first step
of our hybrid mesh deformation algorithm, i.e., anisotropic FEMWARP. We use Mesquite for
implementing the second step of our hybrid mesh deformation algorithm, i.e., multiobjective
mesh optimization. In Chapter 6, we do not use any existing software, but instead C++ is used
The outline of the dissertation is given on the next page. Figure 1.6 shows the flowchart
which describes the overall theme and subthemes of the dissertation as well as the connections
Dissertation Outline
Fig. 1.6 This flowchart represents the two parts of the dissertation and the dependency among
the chapters.
9
Chapter 2
A Numerical Investigation on the Interplay Amongst
2.1 Introduction
Discretization methods, such as the finite element (FE) method, are commonly used in
the numerical solution of partial differential equations (PDEs). The accuracy of the computed
PDE solution depends on the degree of the approximation scheme, the number of elements in
the mesh [4], and the quality of the mesh [5, 6]. More specifically, it is known that as the
element dihedral angles become too large, the discretization error in the finite element solution
increases [4]. In addition, the stability and convergence of the finite element method is affected
by poor quality elements [7]. In particular, it is known that as the angles become too small, the
Analytical studies have been performed at the intersection of meshing and linear solvers.
For example, mathematical connections between mesh geometry, interpolation errors, and stiff-
ness matrix conditioning for triangular and tetrahedral finite element meshes have been stud-
ied [41]. Quality metrics which determine the relevant fitness of elements for the purposes of
interpolation or for creating a global stiffness matrix with a low condition number have been
developed [41].
Further mathematical research has been performed at the intersection of finite element
meshes and linear solvers. A mesh and solver co-adaptation strategy for anisotropic problems
has been developed [8]. In addition, relationships between the spectral condition number of the
stiffness matrix and mesh geometry for second-order elliptic problems for general finite element
Several computational studies have been performed which examined the connections be-
tween finite element meshes and linear solvers in various contexts. For example, the effect of
unstructured meshes on the preconditioned conjugate gradient solver performance for the so-
lution of the Laplace and Poisson equations has been examined [10, 11]. In [12], the relative
performance of multigrid methods for unstructured meshes was studied on fluid flow and radi-
ation diffusion problems. Trade-offs associated with the cost of mesh improvement in terms of
solution efficiency has been examined for problems in computational fluid dynamics [13, 14].
Composite linear solvers which provide better average performance and reliability than single
linear solvers for large-scale, nonlinear PDEs have been designed [15].
We examine the connections between geometry, mesh smoothing, and linear solver con-
vergence for elliptic PDEs via an experimental approach. In particular, we seek answers to the
following questions pertaining to the solution of an elliptic PDE on a given geometric domain.
11
What is the most efficient combination of mesh quality metric, preconditioner, and solver for
solving each PDE problem? Which combinations are most and least sensitive to vertex pertur-
bation? What is the effect of increasing the problem size on the number of iterations required to
converge and on the solver time? Our goal is to determine the best combination of quality metric,
preconditioner, and linear solver which results in a small condition number of the preconditioned
matrix and fast solver convergence for a given PDE, geometric domain, and initial mesh.
To answer the above questions, we use Mesquite [16], a mesh quality improvement
toolkit, and PETSc [17], a linear solver toolkit, to perform a numerical study investigating the
performance of several mesh quality metrics, preconditioners, and sparse linear solvers on the so-
lution of various elliptic PDEs (i.e., Poisson’s equation, general second-order elliptic PDEs, and
linear elasticity) of interest. Mesh quality metrics used in this study are as follows: inverse mean
ratio (IMR) [18], radius ratio (RR) [41], a conditioning-based scale-invariant metric (SI) [41],
performance of the following preconditioners: Jacobi [19], symmetric successive over relax-
ation (SSOR) [19], incomplete LU (ILU) [19] and algebraic multigrid [20]; and linear solvers:
conjugate gradient (CG) [21], minimum residual solver (MINRES) [22], generalized minimal
residual solver (GMRES) [23], and bi-conjugate gradient stabilized solver (Bi-CGSTAB) [24].
The quality metric/preconditioner/linear solver combinations are compared on the basis of ef-
ficiency, robustness, and complexity in solving several elliptic PDEs on realistic unstructured
tetrahedral finite element meshes. We use Mesquite and PETSc in their native state with the
default parameters. Only these two toolkits are employed so that differences in solver imple-
mentations, data structures, and other such factors would not influence the results.
12
This chapter builds on our preliminary results, which were published in [25, 53]. In
particular, this chapter extends our previous work in two ways. First, we also consider the linear
elasticity problem, since this problem has a different sparsity pattern than our previously-studied
elliptic PDE problems. Second, we also investigate the effect of vertex perturbation and the
effect of increasing the problem size on the number of iterations required to converge and on the
solver time. The contents of this chapter have been published in [73].
We consider the solution of three elliptic PDE problems: Poisson’s equation, general
second-order elliptic PDE problems, and linear elasticity. On a given geometric domain, we
consider only homogeneous Dirichlet boundary conditions, since we observed in [25] that mod-
ifying the boundary conditions does not alter the efficiency ranking of the solver time for com-
binations of mesh quality metrics, preconditioners, and linear solvers. To solve the elliptic PDE
problems, triangular meshes are used to discretize the domain, Ω. The standard Galerkin finite
element (FE) method [26] is used to solve the given PDE problem resulting in the linear system
Aξ = b. (2.1)
The approximate solution, ξ , of the analytical solution, u, can be computed by solving (2.1). In
general, (2.1) is a sparse linear system, and iterative methods such as CG and GMRES are often
Poisson’s equation. Poisson’s equation is used to model many mechanical and electromagnetic
∂ 2u ∂ 2u
− − = f on Ω, (2.2)
∂ x2 ∂ y2
where f is a given function. For Poisson’s equation, the matrix A is given by A=K+N, where K
General second-order elliptic PDE problem. The general second-order elliptic PDE problem
on Ω is defined as
∂ 2u ∂ 2u
−α − β + au = f on Ω, (2.3)
∂ x2 ∂ y2
where α and β are PDE coefficients and a and f are given functions. If a = 0, (2.3) reduces to
Poisson’s equation. We consider the case when a is nonzero. The coefficients α and β form a
α 0
C=
.
0 β
For example, C is the identity matrix, I, for (2). The major difference between (2.3) and Pois-
son’s equation in (2.2) is the existence of a mass matrix, M, in the decomposition for matrix A.
For this problem, the matrix A is given by A=K+M+N, where K is the stiffness matrix, M is the
mass matrix, and N is a matrix containing boundary information. One application of the general
Linear elasticity. Linear elasticity is a common problem in structural mechanics and is used
to compute the displacement vector, u, given a body force, f , and (or) a traction load. We do
not consider a traction load. The linear elasticity problem with the Lamé parameters λ and µ is
defined as
− 5 ·τ = f
(2.4)
τ = λ α ∂ u + β ∂ u I + 2µε(u),
∂x ∂y
where α and β are PDE coefficients and τ and ε(u) are the stress and strain tensors, respectively.
1 ∂ ui ∂ u j
εi j (u) = + , i, j = 1, 2.
2 ∂ x j ∂ xi
For homogeneous materials, the Lamé parameters are given by λ =Eν/((1 + ν)(1 − 2ν)) and
µ=E/(2(1 + ν)), where E and ν are the elastic modulus and Poisson’s ratio, respectively [27].
Elliptic PDE coefficients and optimal element shape. The efficiency of elliptic PDE solvers
is highly connected with the condition number, κ, of the matrix A. A smaller κ(A) results in a
faster convergence time for solving (1). The optimal element shape that reduces κ(A) depends on
the coefficients of the elliptic PDEs [9, 28, 41]. It has been shown that for C = γI over Ω, where
γ is a constant, the optimal shape of a triangular element is an equilateral triangle [9, 28, 41].
If C is a constant, but is not given γI over Ω, the optimal shape is different from an equilateral
triangle, and anisotropic triangular elements can be used to obtain a smaller κ(A) [9, 41]. If
C is non-constant over Ω, the optimal shape of a triangular element varies over Ω [28]. We
consider elliptic PDEs with continuous, isotropic coefficients which are constant over Ω (i.e.,
15
C = γI). Therefore, for the PDEs considered in the thesis, the ideal triangular element shape is
an equilateral triangle.
Table 2.1 provides the notation used to define the following mesh quality metrics: in-
verse mean ratio (IMR) [18], radius ratio (RR) [41], a conditioning-based scale-invariant metric
(SI) [41], and an interpolation-based size-and-shape metric (SS) [41]. Table 2.2 defines IMR,
RR, SI, and SS. These four quality metrics were chosen based on their geometric features, which
result in varying contour plots as shown in Fig 3.1. The plots show the contour lines of the qual-
ity of a triangle as a function of a free vertex when the two other vertices held fixed at (0, 0) and
(0, 1).
IMR is one of the most well-known mesh quality metrics for mesh quality improvement
and is evaluated using the position vectors in the element. RR is a quality metric which is
computed using the radius of the triangular element’s circumscribing and inscribing circles. SI
is a quality metric based on stiffness matrix conditioning, whereas SS is a quality metric based
We denote the elements of a mesh and the number of mesh elements by E and |E|,
respectively. The overall quality of the mesh, Q, is a function of the individual element qualities,
qi , where qi is the quality of the ith element in the mesh. The mesh quality depends on the choice
of qi (Section 2.3) and the objective function used to combine the individual element qualities.
16
Notation in Definition
a, b, and c Position vectors for vertices in a
triangular element
C = [b− a; c1− a]
Jacobian of a triangular element
1 √2
W= Incidence matrix for an equilat-
0 23 eral triangle
Area Area of a triangular element
rcirc Radius of a triangular element’s
circumscribing circle
rin Radius of a triangular element’s
inscribing circle
s1 , s2 , and s3 Edge lengths of a triangular ele-
ment
Table 2.1 Notation used in the definition of the 2D mesh quality metrics in Table 2.2.
(c) SI (d) SS
Fig. 2.1 Contour plots of the quality metric of a triangle as a function of a free vertex when the
two other vertices held fixed at (0, 0) and (0, 1).
18
The value of the IMR quality metric lies between 1 and ∞, whereas the RR quality met-
ric’s value lies between 0.5 and ∞. For IMR and RR, a lower value indicates a higher quality
element. For the IMR and RR quality metrics, we define the overall mesh quality, Q, as the sum
|E|
Q = ∑ q2i . (2.5)
i=1
√
The SI quality metric’s value lies between 0 and 1/(4 3), whereas the SS quality met-
ric’s value lies between 0 and ∞. For SI and SS, a higher value indicates a higher quality element.
For the SI and SS quality metrics, we define the overall mesh quality, Q, as the sum of the squares
|E|
1
Q=∑ 2
. (2.6)
q
i=1 i
We use the reciprocal of the quality metric instead of the additive inverse to optimize the
mesh because the optimization process for the latter results in meshes with elements of both very
good and very poor quality, which is not desirable. Using the reciprocal (as was done in [29])
We use the Fletcher-Reeves nonlinear conjugate gradient method [30] in Mesquite [16]
to minimize Q (as defined by either (3.4) or (3.5)), since this method is the default NLCG method
in Mesquite. The local mesh smoothing technique is used for the mesh optimization procedure.
Mesquite employs a line search version of the nonlinear conjugate gradient method. The imple-
mentation ensures that the triangular elements do not tangle as a result of vertex movement.
19
Our preliminary experiments indicate that the total time, which is the sum of the smooth-
ing time and the solver time, is minimal when the reduction in the objective function is 95%
of the possible reduction (when the mesh is completely smoothed). The smoothing time signif-
icantly dominates the solver time when the mesh is completely smoothed. Thus, we perform
inaccurate mesh smoothing, i.e., we smoothe the mesh only to the extent required to yield the
optimal total time for numerically solving the PDE on the mesh (as in [31]).
Four iterative Krylov subspace methods are employed to solve the preconditioned linear
system. The conjugate gradient (CG) solver [21] is a well-known iterative method for solving
systems with symmetric positive definite matrices. It produces a sequence of orthogonal vectors
on successive iterations. Let P be the preconditioner (to be described in Section 2.6). The
convergence rate of the preconditioned CG method depends upon the condition number, κ, of
where λmax and λmin are the maximum and minimum eigenvalues of P−1 A, respectively. The
fastest convergence occurs when eigenvalues are clustered around the same non-null value, and
hence, κ is near 1 [3]. Some theoretical convergence results for the CG solver are given in [14,
32]. In particular, if the eigenvalues of A are uniformly distributed, the number of iterations
required to converge, n, to reduce the error in the energy norm by a factor of ε satisfies the
20
The minimal residual (MINRES) algorithm [22] solves linear systems with symmetric
indefinite matrices. It generates a sequence of orthogonal vectors and attempts to minimize the
residual in each iteration. Similar to CG, a small condition number for P−1 A and clustering of
For nonsymmetric matrices, the generalized minimal residual (GMRES) method [23] is
one of the most widely used iterative solvers. Similar to CG and MINRES, GMRES computes
orthogonal vectors on each iteration; however, the entire sequence needs to be stored. There-
fore, the version of GMRES which restarts GMRES every m steps, i.e., GMRES(m), is used in
practice. It is known that a large value of m is effective in decreasing the number of iterations
required to converge; however, the optimal value of m depends upon the problem [33].
technique, which generates two sets of biorthogonal vectors instead of producing long orthogo-
nal vectors. Bi-CGSTAB is known to have comparable or even faster convergence than other
biorthogonalization methods such as the conjugate gradient squared method. However, Bi-
methods [19]. Bi-CGSTAB and GMRES are the most widely-used iterative methods for solving
2.6 Preconditioners
make the system easier to solve, whereby reducing the convergence time of the iterative solver.
The reader is referred to [34] (and the references therein) for further information on iterative
solvers and preconditioners. In this thesis, four preconditioners are employed. The first is the
The second is the symmetric successive over relaxation (SSOR) preconditioner. SSOR
is similar to Jacobi but decomposes A into L (the strictly lower triangular part), D (the diagonal),
and U (the strictly upper triangular part), i.e., A=L+D+U. The SSOR preconditioner is given by
P = (D − ωL)D−1 (D − ωU),
The incomplete LU (ILU) preconditioner with level zero fill-in (ILU(0)) is the third pre-
conditioner. ILU is commonly used in the solution of elliptic PDE problems. The basic idea
of the ILU preconditioner is to determine lower (L̃) and upper triangular (Ũ) matrices such that
the matrix L̃Ũ-A satisfies certain constraints [34]. The ILU preconditioner works well for many
The fourth is the algebraic multigrid (AMG) preconditioner. Different from geometric
multigrid methods, the AMG preconditioner does not need any mesh information to generate
the preconditioner and hence is known as black-box technique. AMG only requires the matrix A
to generate the preconditioner. Therefore, the AMG preconditioner can be used to solve linear
systems which arise from unstructured meshes. The main idea of the AMG preconditioner is to
22
eliminate the smooth error using restriction and interpolation which is not removed by relaxation
on the fine grid [20]. We use the default options for HYPRE BoomerAMG in PETSc. The
default option in PETSc employs 25 levels of V-cycles. For further infomation on HYPRE
We consider the following questions which we investigate on the three elliptic PDE prob-
lems shown in Table 2.3. What is the most efficient combination of mesh quality metric, pre-
conditioner, and solver for solving each PDE problem? Which combinations are most and least
sensitive to vertex perturbation? What is the effect of increasing the problem size on the number
of iterations required to converge? Table 2.3 summarizes the experiments and corresponding
PDE problems to be solved. For all three PDE problems in Table 2.3, we consider only a sim-
ple homogeneous Dirichlet boundary condition with u=0 on the boundary because we already
observed that modifying boundary condition does not affect the efficiency ranking [25]. For the
elasticity problem in Table 2.3, E and ν are set to 1 and 0.3, respectively. The body force, f , is
set to [1 0]T . The machine employed for this study is equipped with an Intel Nehalem processor
We consider elliptic PDEs with isotropic coefficients, which is constant (i.e., C = αI)
over 2D geometric domains. The optimal element shapes for solving elliptic PDEs with anisotropic
coefficients are different from the optimal shapes for elliptic PDEs with isotropic coefficients [9,
28, 41]. Therefore, our experimental results cannot be generalized to the solution of elliptic
23
PDEs with anisotropic coefficients. The quality metrics used to determine the mesh element
quality are different for 2D and 3D elements. In addition, the sparsity patterns of the matrices
in the linear systems obtained from unstructured meshes are different for 2D and 3D meshes.
Table 2.3 Listing of numerical experiments and examples of PDE problems. The letters (a)
through (c) are representative examples of the three types of PDE problems under consideration.
Geometric Domains. The two 2D geometric domains, wrench and hinge, considered in our ex-
periments are shown in Figure 3.3. Triangle [36] is used to generate initial meshes on the two
domains. Half the interior vertices in each mesh are perturbed to create test meshes that are fur-
ther from optimal. Properties of the test meshes and the corresponding finite element matrices
Finite Element Solution. The FE method described in Section 2.2 is used to discretize the
domain, Ω, and to generate a linear system of the form Aξ =b. PETSc [17] is used to generate
the preconditioners, P, and to solve the linear system, P−1 Aξ =P−1 b. We employ the solvers and
preconditioners described in Sections 2.5 and 2.6, respectively, to solve the linear system. Table
24
Fig. 2.2 Coarse initial meshes on the wrench and hinge geometric domains indicative of the
actual meshes to be smoothed.
Table 2.4 Properties of meshes on geometric domains. Initial angle distributions for the meshes
are given in Table 2.7.
25
2.5 enumerates the 16 preconditioner-solver combinations used in our experiments. The default
Solver
Preconditioner
CG GMRES MINRES BI-CGSTAB
Jacobi 1 2 3 4
SSOR 5 6 7 8
ILU(0) 9 10 11 12
AMG 13 14 15 16
Table 2.5 The sixteen combinations of preconditioners and solvers. For example, 10 refers to
using the ILU(0) preconditioner with the GMRES solver.
The default stopping criteria in PETSc were employed. For example, the absolute toler-
ance, abstol, and the relative tolerance, rtol, are set to 1e-50 and 1e-05, respectively. The max-
imum number of iterations for solving the preconditioned linear system is set to 10,000. When
the preconditioned linear system is solved, ξ0 is set to the default value of 0. The preconditioned
linear system converges on the ith iteration if the following inequality is satisfied:
where ri is the residual at the ith iteration and r0 is the initial residual.
Accuracy of the Solution. The exact solutions of the PDE problems in Table 2.3 on the geo-
metric domains in Figure 3.3 are unknown. Therefore, we conduct a mesh-independence study
using meshes with 10K, 50K, 100K, 200K, and 500K vertices and verify that the angle distribu-
tion and resulting PDE solution is independent of the mesh size. Our experimental results show
that the finite element method converges to the same PDE solution if the mesh consists of more
26
than 10K vertices. We also observe that the finite element solution is not affected by choice of
quality metric, preconditioner, and linear solver for these mesh sizes. We will investigate the
accuracy of the finite element solution for boundary value PDEs as future work.
Timing. In our experiments, the total time is defined as the sum of the smoothing time and
the solver time. The smoothing time is the time to achieve an accurately smoothed mesh as
described in Section 2.4. The time required to smoothe the test meshes using various quality
metrics is shown in Table 2.6. The solver time is the time the solver takes to satisfy (2.7) and
For our experiments, we use Mesquite version 1.1.7 and PETSc version 3.0.0. Mesquite
and PETSc are widely used for solving linear systems [14, 15, 37] and mesh quality improve-
ment [25, 31, 38], respectively. Both PETSc and Mesquite are numerical libraries containing
data structures and routines. Algorithms in each software package are coded using similar data
structures and routines. We report both the timing (in seconds) and the number of iterations
required to converge. Note the latter is not affected by the software version or the hardware used
to solve the problem. However, the timing can be affected by these factors. We expect that the
relative difference amongst different combinations of quality metric, preconditioner, and linear
Table 2.6 Mesh smoothing time (sec) for various mesh quality metrics
27
Experiment 1: Best combination of mesh quality metric, preconditioner and solver. In this
experiment, we seek to determine the most efficient combination of mesh quality metric, pre-
conditioner, and solver for each type of PDE problem in Table 2.3. We will examine the most
efficient combinations for both the solver time and the total time.
reported that the CG solver is more robust to perturbation than is the GMRES solver for the
and solvers to such perturbations. We randomly perturb a certain number of vertices in each mesh
while ensuring that no perturbation results in a very poorly-shaped element (e.g., an inverted
element). We perturb the interior vertices such that they move less than half the distance at
which element inversion would occur. In our experiments, we perturb 5%, 25%, and 50% of the
elements and investigate the robustness of the preconditioner-solver combinations to the vertex
perturbations.
We define the relative increase, i.e., the percentage increase (PI), in the solver time due
Tmax − Tmin
PI = ,
Tmin
where Tmax and Tmin are the maximum and minimum convergence time among 0% (i.e., fully-
smoothed unperturbed mesh), 5%, 10%, 25%, and 50% perturbed elements, respectively. Note
28
that we observe monotonic convergence behavior with all four mesh quality metrics.
Experiment 3: Increasing the problem size. In this experiment, we examine the convergence
mesh for a given domain. We execute our numerical experiments for meshes with 10K, 50K,
100K, 200K, and 500K vertices on each domain. Let N be the number of vertices in the mesh.
We compute the order of convergence, denoted O(N s ), using the linear least-squares method for
both the number of iterations to converge and the solver time. Thus, the total work required to
solve the system is O(N s+1 ) for each combination of preconditioner and solver.
We now discuss the results for the three numerical experiments discussed in Section 2.7.1
2.7.3.1 Preliminary experiment for determination of restart value of the GMRES solver
The default restart value, m, in PETSc is 30, and we examine whether that value is
optimal to solve each of the PDE problems. We experimentally determine the optimal m value,
which yields the fastest solver time, for each application by varying it as follows: 10, 30, 50,
Figure 2.3(a) shows the number of iterations required to converge for different m val-
ues as the problem size increases for the wrench domain. For this experiment, the meshes are
smoothed with the IMR quality metric, and the SSOR preconditioner with the GMRES solver
is employed to solve Poisson’s equation (problem (A) in Table 2.3). The results obtained are
29
representative of the results obtained with other geometric domains, quality metrics, and pre-
conditioners. We observe that as we increase the values of m, the total number of iterations to
converge reduces. This is consistent with the theoretical analysis presented in [33].
Figure 2.3(b) shows the solver time for different m values as we increase the problem size.
We observe that the solver time is least when the restart value, m, is 30. As we reduce or increase
the restart value, the time taken to converge to a solution increases. Large restart values (e.g.,
m = 100 and 200) result in slower solver times because the cost of orthogonalization increases
with an increase in m. Small restart values (e.g., m = 10) also slow the solver since it needs a
For the other types of PDE problems in Table 2.3, we observe similar results. The GM-
RES solver with m = 30 yields the least solver time, although m = 200 results in the fewest
number of iterations to converge. Therefore, we choose the GMRES solver with m = 30 for all
and solver for Poisson’s equation. Table 2.8 shows the solver time and the number of iterations
required for convergence as a function of mesh quality metrics for different combinations of
preconditioners and solvers. The most efficient combinations (e.g., the IMR mesh quality metric
with the AMG preconditioner and the CG solver) are 97% faster than the least efficient combi-
nations (e.g., the IMR quality metric with the Jacobi preconditioner and the GMRES solver).
We observe that the AMG preconditioner with any choice of quality metric and solver is
faster to converge than other combinations of quality metric, preconditioner and solver. When
30
Fig. 2.3 Effect of the mesh size on the number of iterations (a) and solver time (b) to convergence
as a function of GMRES restart value, m, for Poisson’s equation (problem (A) in Table 2.3) on
the wrench domain.
31
we solve linear equations derived from a mesh, high frequency terms are not eliminated only
by relaxation on the fine grid. Thus, the AMG preconditioner used in the experiment employs
25 level of V-cycles and efficiently eliminates high frequency terms (which correspond to large
eigenvalues in the matrix A) using coarse grid correction techniques. The most time-consuming
step during the generation of the AMG preconditioner is an aggregation step which finds the
vertex-neighborhood information from the matrix A. However, the aggregation step is fast for
Poisson’s equation as it contains a single degree of freedom. Note that the number of iterations
required to converge for combinations with the AMG preconditioner represent the number of
We also observe that the Bi-CGSTAB solver (combinations 4, 8, 12, and 16 in Table 2.5)
is most sensitive to the choice of different mesh quality metric, whereas, the CG solver is least
affected by the choice of the quality metric. These results are related with the irregular conver-
gence behavior discussed in [25]. The number of iterations required to converge and the solver
time taken by the BiCG-STAB solver with a given preconditioner varies by more than 20% in
many cases. For other combinations, however, the variation is small and is restricted to less than
On the choice of mesh quality metrics, the SS metric is the least efficient mesh quality
metrics in most cases. The method using the SS metric often fails to converge because meshes
smoothed by the SS metric fail to yield a positive definite preconditioner. Note that both the
CG and the MINRES solvers require the preconditioner to be a positive definite matrix. In most
cases, the maximum eigenvalue of the matrix A is larger and the minimum eigenvalue of the
matrix A is smaller than the eigenvalues of the corresponding matrix A for the other quality
metrics. These eigenvalues are connected to the shape (angle) of the elements in the mesh.
32
Table 2.7 shows the angle (θ ) distributions for elements in the initial mesh (i.e., with
50% of the interior vertices perturbed) and different quality metrics on the wrench (500K) and
hinge (500k) domains. We observe that all four quality metrics try to generate nearly equilateral
elements, but the SS metric yields meshes containing elements with more small and large angles
than do the other quality metrics. This is because the SS metric penalizes small angles less than
do other quality metrics [41]. For elliptic PDEs with isotropic coefficients, it is well-known that
equilateral triangles are desirable for efficiency, whereas small angles have a bad effect on the
condition number of matrix A and on the efficiency [41]. This explains why the SS metric is
less efficient for mesh smoothing than other quality metrics. We also observe that the elements
with small angles or large angles (i.e., poorly-shaped) occur most often on the boundary of the
mesh. Interestingly, among the four quality metrics in Table 2.7, the RR metric has the fewest
In terms of the total time, the IMR metric outperforms the other quality metrics, be-
cause numerical computation of the IMR metric for mesh optimization is highly optimized in
Mesquite. Numerical computation of the other mesh quality metrics, i.e., RR, SI, and SS, are not
as optimized and hence are less efficient to compute. Similar to the solver time, the SS metric is
the least efficient quality metric in terms of the total time. The total time for the most efficient
combination (i.e., AMG with CG) for the IMR metric is 90% less than it is for SS. The least
efficient combination (i.e., Jacobi with GMRES) for the IMR metric is 45% faster than that of
Exp. 2: Effect of perturbation for Poisson’s equation. Table 2.9 shows the linear solver
time and the number of iterations required to converge as a function of the amount of vertex
33
Table 2.7 Angle (θ ) distribution for various mesh quality metrics. The reported values indicate a
percentage of angles in the mesh. The initial mesh is the mesh with 50% of the interior vertices
perturbed.
perturbation on the wrench and hinge meshes, respectively. Figures 2.4(a) and 2.4(b) show
the PI for different combinations of preconditioners and solvers when the number of perturbed
vertices is increased. For these experiments, the meshes are smoothed with the IMR quality
metric. The results obtained are representative of the results obtained with other quality metrics.
More preconditioner-solver combinations are able to solve the PDEs on the wrench do-
main than on the hinge domain. The PDE on the hinge domain is harder to solve because there
are many holes in the domain and the holes create thin areas near the boundary. The vertices
are highly constrained in these areas and perturbation produces poorly-shaped elements in many
cases.
Table 2.9 shows that the CG and MINRES solvers fail to converge in many cases, whereas
the GMRES and Bi-CGSTAB solvers do not. The CG and the MINRES solver often fail to
converge because they fail to generate a positive definite preconditioner for two reasons. First,
34
Table 2.8 Linear solver time (secs) and number of iterations required to converge for Pois-
son’s equation (problem (A) in Table 2.3) as a function of mesh quality metric for the 16
preconditioner-solver combinations (see Table 2.5) on the wrench and hinge domains. A ’*’
denotes failure. For each quality metric, the numbers in the top and bottom rows represent the
linear solver time and number of iterations to convergence, respectively.
35
generation of the preconditioners fails when there are poorly-shaped elements in the mesh and
when the minimum eigenvalue of the matrix A is too small relative to its maximum eigenvalue.
Poorly-shaped elements represent the elements which have large (e.g., angle > 100◦ ) or small
angles (e.g., angle < 20◦ ). This was also observed in the above experiment for the SS metric.
Secondly, preconditioners for the CG and MINRES solvers fail when the element lengths are
very small. In these cases, the minimum eigenvalue of the matrix A is very small and corresponds
to a large condition number. The minimum eigenvalue of the matrix A depends on the edge
lengths in the mesh. For these reasons, when the element lengths are very small or when the
meshes include lots of poorly-shaped elements, we see that GMRES and Bi-CGSTAB are more
perturbation: ILU(0) > SSOR ≈ AMG > Jacobi. The ranking is in order of most robust to least
robust. Note that the solver time for the AMG preconditioner is less than 10 seconds, whereas
it is greater than 40 seconds for the other preconditioners. Hence, the PI values are sensitive to
Exp. 3: Increasing the problem size for Poisson’s equation. We observe that the maximum
eigenvalues of P−1 A stay constant, but the minimum eigenvalue of P−1 A rapidly decreases as
N increases. This is consistent with the results observed in [41]. In [41], it was discussed that
the minimum eigenvalue in matrix A is a function of the edge lengths in the mesh. It was also
discussed in Section 2.5 that clustering of eigenvalues around the same non-null value results in
a faster convergence rate. Thus, as the minimum eigenvalues of P−1 A decrease, the eigenvalue
spectrum becomes larger, and the number of iterations required to converge increases.
36
Table 2.9 Linear solver time (secs) and number of iterations required to converge for Poisson’s
equation (problem (A) in Table 2.3) as a function of vertex perturbation for the 16 preconditioner-
solver combinations (see Table 2.5) on the two geometric domains. A ’*’ denotes failure. For
each percentage of vertices perturbed, the numbers in the top and bottom rows represent the
linear solver time and number of iterations to convergence, respectively.
37
Fig. 2.4 Percentage increase (PI) as a function of the solver time for different combinations of
preconditioners and solvers for Poisson’s equation (problem (A) in Table 2.3). Preconditioner-
solver combinations which fail to generate a preconditioner or do not converge correspond to the
missing bars in these figures. Note that PI values for the hinge domain are significantly greater
than those for the wrench domain.
38
Figure 2.5(a) shows the value of s in O(N s ) for the number of iterations required to
converge for different combinations of preconditioners and solvers on the wrench domain. In
this experiment, we use meshes smoothed by employing the IMR quality metric. The results are
similar for other quality metrics. The AMG preconditioner with any solver (i.e., combinations
13 − 16 in Table 2.5)yields values of s less than 0.1. The other combinations of preconditioners
and solvers yield s values around 0.5. The Jacobi-GMRES combination has the largest value with
s approximately 0.8. Figure 2.5(b) shows the same results on the hinge domain. Figures 2.6(a)
and 2.6(b) show the order of convergence of the solver time for the wrench and the hinge domain,
respectively.
The best combination of quality metric, preconditioner, and solver is not affected by
increasing the problem size. The AMG preconditioner with any solver beats the other combina-
tions. In terms of the quality metric, the RR metric is the most efficient, whereas the SS metric
and solver for general second-oder elliptic PDEs. Table 2.10 shows the solver time as a
function of quality metric for various combinations of preconditioners and solvers for problem
(B) in Table 2.3. We observe that the overall solver time for problem (B) is lower than it is for
the other problems. The results show the most efficient combination for the solution of problem
The most efficient combination is the AMG preconditioner with any choice of quality
metric and solver. Similar to Poisson’s equation, general second-order elliptic PDEs have one
39
(a) Wrench
(b) Hinge
Fig. 2.5 The order of convergence for the solver time based on the number of iterations for
various combinations of preconditioners and solvers for Poisson’s equation (problem (A) in Ta-
ble 2.3). Meshes with 10K, 50K, 100K, 200K, and 500K vertices on each domain are employed
in computing s using the linear least-squares method.
40
(a) Wrench
(b) Hinge
Fig. 2.6 Similar to Figure 2.5, this figure displays the order of convergence based on the solver
time for Poisson’s equation (problem (A) in Table 2.3). Meshes with 10K, 50K, 100K, 200K,
and 500K vertices on each domain are employed in computing s using the linear least-squares
method.
41
degree of freedom during the aggregation step for generation of the AMG preconditioner. There-
fore, it results in fast solver time. The solver time for the most efficient combination is 95% less
than the least efficient combination (e.g., the SI quality metric with the Jacobi preconditioner
and the GMRES solver). For the choice of mesh quality metric, RR is more efficient than other
quality metrics in most cases, and the SS quality metric is the least efficient quality metric. This
result is also related to the angle distribution of mesh elements as discussed before. Similar
to Poisson’s equation (Problem (A)), the Bi-CGSTAB solver is most sensitive to the choice of
different mesh quality metric, whereas, the CG solver is least affected by the choice of quality
metric.
We also observe that the use of any preconditioner significantly mitigates poorly-shaped
elements for the SS metric when solving problem (B). This implies that, for this problem, some
poorly-shaped mesh elements can be overcome, and the solver time does not increase when the
In terms of the total time, the overall trend is similar to that seen for Poisson’s equation
(problem (A)). In most cases, we observe the following rank ordering of quality metrics with
respect to the total time: IMR > RR > SI > SS. The ranking is in order of fastest to slowest.
The most efficient combination is the IMR quality metric and the AMG preconditioner with any
choice of the solver. Note that for problem (B), the smoothing time dominates the solver time
more when compared with problem (A) because the overall solver time for solving problem (B)
We also conducted experiments in which we modify the PDE coefficients in (2.3) such
that the coefficient matrix, C = γI, where γ is a constant. We verified that the efficiency rankings
are not affected by these modifications because they do not affect the optimal triangular shape,
42
and hence κ(A) is not affected. Also, the sparsity pattern of A is not changed by these mod-
ifications. We also modify the functions, f in (2.3) and observe consistent efficiency ranking
results because modifying f only affects b of the linear system in (1). This is consistent with the
a values. We investigate the existence of a value on the efficiency rankings. If a is too large (i.e.,
a > 100) or too small (i.e., a < -100) compared with f (e.g., f = 1), the problem is dominated
by the linear term and affects the efficiency rankings. Therefore, we assume -100 < a < 100
and study the presence of the mass matrix on the efficiency. Further discussion on the effect of
Exp. 2: Effect of perturbation for general second-oder elliptic PDEs. Figures 2.7(a) and 2.7(b)
show the changes in the solver time to compute the solution for the perturbed wrench (500K)
and hinge meshes (500K), respectively. For this experiment, the meshes are smoothed with the
IMR quality metric. The results obtained are representative of the results obtained with other
quality metrics. The Jacobi preconditioner is most sensitive to the vertex perturbation. Note that
the solver time for the AMG preconditioner is less than 10 seconds, whereas it is greater than 40
seconds for the other preconditioners. Hence the values are more sensitive to small changes.
Similar to Poisson’s equation, the SSOR preconditioner with the CG or MINRES solver
(combinations 5 and 7 in Table 2.5) shows the most robust performance with respect to pertur-
bation for both the wrench and hinge domains. A few poor quality elements can increase the
values of the maximum eigenvalues and the condition number of the linear system. We observe
that both CG and MINRES are able to circumvent the numerical difficulties associated with
43
Table 2.10 Linear solver time (secs) and number of iterations required to converge for general
second-order elliptic PDEs (problem (B) in Table 2.3) as a function of mesh quality metric for
the 16 preconditioner-solver combinations (see Table 2.5) on the two geometric domains. A ’*’
denotes failure. For each quality metric, the numbers in the top and bottom rows represent the
linear solver time and number of iterations to convergence, respectively.
44
large eigenvalues [41]. The GMRES solver (GMRES(30)) is most sensitive to the poor elements
because it restarts in every 30 iterations without executing the additional iterations needed to
converge faster. The least robust combinations are the Jacobi preconditioner with any choice of
the solver.
Exp. 3: Increasing the problem size for general second-oder elliptic PDEs. Figure 2.8(a)
and 2.8(b) show the order of convergence for the number of iterations required to converge as we
increase N. The overall trend is similar to that seen for problem (A). The Jacobi preconditioner
with the GMRES solver has the largest s value which is 0.63. The s value of the CG solver
with any preconditioner is approximately 0.5. We observe that the AMG preconditioner is not
sensitive to an increase in N. Four solvers when combined with the AMG preconditioner have
s values less than 0.1. Figures 2.9(a) and 2.9(a) show the same results for the solver times. The
results are similar to the results for problem (A). The best combinations (the RR quality metric
and the AMG preconditioner with any choice of the solver) is not affected by increasing the
problem size.
and solver for linear elasticity. Table 2.11 shows the solver time and number of iterations
for problem (C) in Table 2.3. We observe that the efficiency rankings are different from those
obtained from Poisson’s equation (problem (A)) and general elliptic PDEs (problem (B)).
45
Fig. 2.7 PI as a function of the solver time for various combinations of preconditioners and
solvers after vertex perturbation for general second-order elliptic PDEs (problem (B) in Ta-
ble 2.3).
46
(a) Wrench
(b) Hinge
Fig. 2.8 The order of convergence for the solver time based on the number of iterations for the
combinations of preconditioners and solvers for general second-order elliptic PDEs (problem
(B) in Table 2.3). Meshes with 10K, 50K, 100K, 200K, and 500K vertices on each domain are
employed in computing s using the linear least-squares method.
47
(a) Wrench
(b) Hinge
Fig. 2.9 Similar to Figure 2.8, this figure displays the order of convergence based on the solver
time for general second-order elliptic PDEs (problem (B) in Table 2.3). Meshes with 10K, 50K,
100K, 200K, and 500K vertices on each domain are employed in computing s using the linear
least-squares method.
48
The ILU(0)-preconditioned solvers are more efficient than the AMG-preconditioned solvers.
The reason for the difference is the difference in sparsity pattern of the A matrix for each appli-
cation. For Poisson’s equation, the matrix A has a sparsity pattern that corresponds to the mesh
connectivity. Thus, the AMG preconditioner works very well when an aggregation step is per-
formed. For the linear elasticity equations, A has twice the number of rows and columns as the
number of vertices in the mesh. Thus, the AMG coarsening algorithm generates aggregates of
physically-incompatible degrees of freedom [39]. This results in an increased solver time for the
Also, the BiCG-STAB solver shows an approximately 20% variation in the number of
variation of other solvers is in the range of 10%. We observe the power of preconditioners in
some cases. Although the meshes smoothed by the SS metric have more poorly-shaped elements
than those obtained by smoothing using other quality metrics, the solver time for the SS metric
Among the 16 combinations of preconditioners and solvers, the most efficient combina-
tions are the IMR or the SI quality metric with the ILU(0) preconditioner and the CG solver with
respect to the solver time. On the wrench domain, the most efficient combination is 92% faster
than the least efficient combination (i.e., the RR quality metric with the Jacobi preconditioner
In terms of the total time, the IMR quality metric with the ILU(0) preconditioner and
the MINRES solver outperforms other combinations. The least efficient combination is the SS
quality metric with the Jacobi preconditioner and the GMRES solver. The total time for the most
efficient combination (i.e., ILU(0) with MINRES) with the IMR metric is 71% less than it is
49
for this combination with SS. The least efficient combination (i.e., Jacobi with GMRES) for the
IMR metric is 16% faster than that of the least efficient quality metric, i.e., SI. In this case, the
We conducted experiments in which we modify the PDE coefficients in (2.4) such that
the coefficient matrix, C = γI, where γ is a constant. Similar to the second-order general elliptic
PDE problems, experimental results show that the sparsity pattern of A, the optimal triangular
shape, and hence κ(A) are not affected by these modifications as discussed in Section 2.
Exp. 2: Effect of perturbation for linear elasticity. Table 2.12 shows the effect of the vertex
perturbation on the solution of the linear elasticity problem. For these experiments, the meshes
are smoothed with the IMR quality metric. The results shown here are typical. We observe that
the effect of perturbation is different from problem (A) and (B). As was explained before, this is
because the sparsity pattern of the linear system in this application is different from that of the
previous applications. Figures 2.10(a) and 2.10(b) show the PI for different preconditioner-solver
combinations when the number of perturbed vertices is increased. The AMG preconditioner with
the GMRES solver (combination 14 in Table 2.5) is least sensitive to vertex perturbation. Similar
to problem (B), the SSOR preconditioner with the CG and MINRES solvers (combinations 5 and
We observe that the SSOR preconditioner with the GMRES solver or Bi-CGSTAB solver
and the ILU(0) preconditioner with the GMRES solver (combinations 6, 8, and 10 in Table 2.5)
are most sensitive to perturbation. These two solvers are not efficient to circumvent the numerical
difficulties associated with poor eigenvalues. The SSOR preconditioner with the Bi-CGSTAB
solver takes 52% more time to converge than does the most robust combination (combination 14
50
Table 2.11 Linear solver time (secs) and number of iterations required to converge for the linear
elasticity problem (problem (C) in Table 2.3) as a function of mesh quality metric for the 16
preconditioner-solver combinations (see Table 2.5) on the two geometric domains. A ’*’ denotes
failure. For each quality metric, the numbers in the top and bottom rows represent the linear
solver time and number of iterations to convergence, respectively.
51
Fig. 2.10 PI as a function of the solver time after vertex perturbation for the combinations of pre-
conditioners and solvers for the linear elasticity problem (problem (C) in Table 2.3). The missing
bar (combination 10) for the hinge domain corresponds to a preconditioner-solver combination
which does not converge.
52
Table 2.12 Linear solver time (secs) and number of iterations required to converge for the linear
elasticity problem (problem (C) in Table 2.3) as a function of vertex perturbation for the 16
preconditioner-solver combinations (see Table 2.5) on the two geometric domains. A ’*’ denotes
failure. For each percentage of vertices perturbed, the numbers in the top and bottom rows
represent the linear solver time and number of iterations to convergence, respectively.
53
(a) Wrench
(b) Hinge
Fig. 2.11 The order of convergence for the solver time based on the number of iterations for the
different combinations of preconditioners and solvers for the linear elasticity problem (problem
(C) in Table 2.3). Meshes with 10K, 50K, 100K, 200K, and 500K vertices on each domain are
employed in computing s using the linear least-squares method.
54
(a) Wrench
(b) Hinge
Fig. 2.12 Similar to Figure 2.11, this figure displays the order of convergence based on the solver
time for the linear elasticity problem (problem (C) in Table 2.3). Meshes with 10K, 50K, 100K,
200K, and 500K vertices on each domain are employed in computing s using the linear least-
squares method.
55
in Table 2.5). On both domains, the AMG preconditioner is less affected by vertex perturbation
than the other preconditioners because the coarse grid correction during the V-cycles removes
Exp. 3: Increasing the problem size for linear elasticity. Figures 2.11(a) and 2.11(b) show
the s values for the number of iterations required to converge for the wrench and hinge domains,
respectively. For this application, the experiments are performed on meshes smoothed using the
IMR quality metric. For the simple wrench domain, the s values for the AMG preconditioner are
less than for the other preconditioners. For the more complex hinge domain, however, the order
of convergence of the number of iterations for all the combinations are all close to 0.5. Note that
for the Jacobi preconditioner with the GMRES solver, s is approximately 0.8. Figures 2.12(a)
and 2.12(b) show the order of convergence of the solver time for the wrench and the hinge
domain, respectively.
In most cases, the best combinations are the RR quality metric with the ILU(0) precon-
ditioner and the MINRES (or CG) solver. The least efficient combinations are the SS quality
We studied the most efficient combinations of quality metric, preconditioner, and sparse
linear solver for the numerical solution of various elliptic PDEs on 2D geometric domains. This
is the first time these three important factors, which affect the efficiency of the solution for
various elliptic PDEs, have been studied. According to our experimental results, by choosing
56
the most efficient combination, solver time and total time can be reduced by 90% and 97%,
For all elliptic PDEs considered here, we observe that the radius ratio (RR) metric is the
most efficient metric for minimizing the solver time, as the mesh smoothed by the RR metric con-
tains the fewest poorly-shaped elements. Poorly-shaped elements increase both the maximum
eigenvalue and the condition number of the linear system. The interpolation-based size-and-
shape (SS) metric is the least efficient mesh quality metric in terms of its effect on the solver
time because meshes smoothed with the SS metric have more poorly-shaped elements than the
meshes smoothed by other quality metrics. We also observe that the choice of the preconditioner
is the most important factor that affects the solver time for all elliptic PDEs. The choice of a
good preconditioner is more important than the choice of a good quality metric or good linear
For solving Poisson’s equation and general second-order elliptic PDEs, the most efficient
combination for minimizing the solver time is the RR quality metric with the algebraic multigrid
(AMG) preconditioner and any choice of linear solver. For these problems, which have a single
degree of freedom, the AMG preconditioner efficiently eliminates large eigenvalues, and its
aggregation step is fast. This results in a solver time which is 97% faster than the least efficient
combination, i.e., the Jacobi preconditioner and the GMRES solver with any choice of quality
metric.
For linear elasticity problems, the RR metric with the ILU(0) preconditioner and the
MINRES (or CG) solver is the most efficient combination, which are up to 92% faster than the
least efficient combination, which is the Jacobi preconditioner and the GMRES solver with any
choice of the quality metric. Due to its higher degree of freedom and different sparsity pattern,
57
where displacements for x and y are computed in a coupled manner, the AMG preconditioner is
The most efficient combinations are also studied with respect to the total time, which
includes both the smoothing and the solver time. For all elliptic PDEs, choosing an efficient
quality metric is more important than the other factors in order to minimize the total time (as
opposed to just the solver time). The inverse mean ratio (IMR) metric computation is highly op-
timized in Mesquite, and thus combinations with the IMR metric appear in the best combination.
For Poisson’s equation and general second-order elliptic PDEs, the IMR metric with the AMG
preconditioner and the CG solver is 90% faster than the same combination with the SS metric.
For linear elasticity problems, the IMR metric with the ILU(0) preconditioner and MINRES (or
We also investigated the robustness of the combinations with respect to vertex perturba-
tion. These results are useful for choosing the most robust preconditioner-solver combination
when the initial mesh quality is poor and mesh smoothing is not performed. In most cases, the
SSOR preconditioner with the CG or MINRES solver is more robust than other combinations
because they are able to circumvent numerical difficulties associated with large eigenvalues as
discussed above. The AMG preconditioner is less affected by vertex perturbation than the other
preconditioners because the coarse grid correction during the V-cycles removes large eigenvalues
effectively.
Finally, we examined the effect of increasing the problem size on the number of iterations
required to converge and on the solver time. For all three PDEs, the AMG preconditioner with
any quality metric or solver exhibits up to 94% faster convergence than the other combinations
of preconditioners and solvers as the problem size is increased. The order of convergence for
58
Poisson’s equation and general second-order elliptic PDEs, combinations with the AMG precon-
Our experimental results can be generalized to the solution of other elliptic PDE prob-
lems with constant isotropic coefficients on various 2D geometric domains and homogeneous
boundary conditions. However, they cannot be generalized to elliptic PDE problems with anisotropic
coefficients or the 3D geometric domains for the following two reasons. First, elliptic PDEs with
anisotropic coefficients require different ideal element shapes. Second, the quality metrics for
In this chapter, we investigated the best combinations of mesh quality metric, precon-
ditioner, and sparse linear solver for solving various elliptic PDE problems. We focused on
optimizing only one aspect of the mesh, such as the element shape or interpolation error, when
we considered the effect the choice of various mesh quality metrics had on the solution of elliptic
PDEs. In the next chapter, we propose a multiobjective mesh optimization framework which is
Chapter 3
A Multiobjective Mesh Optimization Framework for Mesh
3.1 Introduction
have multiple requirements for the finite element mesh discretizing the geometric domain. For
example, such requirements may include having non-inverted mesh elements, elements that are
well-shaped, elements with uniform element size, and/or elements which yield small PDE in-
terpolation error. Several shape-based mesh optimization algorithms have been devised to yield
mesh elements which satisfy various geometric requirements or which improve the PDE solu-
tion accuracy or efficiency of the PDE solver [18, 40, 41]. Mathematical study of geometric
mesh quality metrics based on element shape has been performed using the Jacobian and re-
lated matrices [18, 40]. Some interpolation error-based quality metrics have been proposed [41].
Both theoretical interpolation error and extremal eigenvalue bounds have been used to develop
such mesh quality metrics. Although the size-based quality metric has been less studied by re-
searchers, elements should be smoothly graded for numerical methods to behave properly [42].
60
Many researchers have also proposed methods to eliminate inverted elements in a tangled mesh.
Knupp [43] proposed an untangling mesh quality metric by assigning a large penalty (cost) to
Despite there being multiple mesh requirements for various PDE applications, most tradi-
tional mesh optimization algorithms optimize only a single objective function and hence improve
only one aspect of the mesh. Optimizing only one aspect of the mesh often results in the deteri-
oration of other aspects of the mesh at the same time. For instance, the untangle single objective
method is able to eliminate inverted elements if the initial mesh is tangled. However, this method
often yields elements with poor shapes by focusing only on elimination of the inverted elements
while ignoring the element quality. These poorly-shaped elements ruin the accuracy of the PDE
solution, and the efficiency, stability, and conditioning of the associated finite element solver.
Moreover, the shape single objective method is able to improve element shape but cannot elim-
inate inverted elements in the tangled mesh. The resulting mesh yields an invalid finite element
solution which cannot be used in a computational simulation. Similar tradeoffs exist in regards
to optimizing other aspects of the mesh, such as the element shape and element size or the ele-
ment shape and the PDE interpolation error. Therefore, it should be clear that meshes obtained
by sequentially optimizing two or more aspects of the mesh do not simultaneously reflect all of
Researchers have also proposed a few techniques for optimizing two or more aspects of
the mesh. Escobar et al. [44] proposed a method for simultaneously untangling and improving
mesh quality by combining untangling and shape-based mesh quality metrics. However, their
algorithm can only be used for mesh untangling and for mesh quality improvement. In addition,
it is difficult to extend their method to improvement of other mesh aspects. Recently, Franks
61
and Knupp proposed new quality metrics within a target matrix paradigm for simultaneous mesh
untangling and quality improvement [45]. However, this method is also difficult to extend to
optimize other aspects of the mesh, as it develops two specific quality metrics for simultaneous
shape improvement and untangling. Combining two metrics to obtain a single objective function
based on the interpolation error and condition number of the stiffness matrix has also been pro-
posed [41]. An objective function formed in this manner needs to determine relative preferences
provement of multiple aspects of the mesh with the goal of improving the accuracy, efficiency,
stability, and conditioning of the associated finite element solver. Examples of such aspects of
the mesh which can be simultaneously improved by techniques within our framework include
the element shape, element size, associated PDE interpolation error, and the number of inverted
elements. Unlike the previous approaches, we are able to simultaneously improve several aspects
of the mesh. We are also able to consider new combinations of mesh aspects for optimization.
The biggest strength of our multiobjective framework lies in its flexibility and extensibility to
Techniques within our framework combine competing objective functions into a single
techniques satisfy a sufficient condition of weak Pareto optimality, which means that there is
no other solution that improves all of the objective functions simultaneously [91]. Multiobjec-
tive mesh optimization methods within our framework can be classified into a priori and no
not require articulation of preferences (weights), as weights for different objective functions do
not need to be specified, and often they are not known. The techniques which we develop in-
clude the exponential sum, objective product, and equal sum multiobjective mesh optimization
methods. We test our methods on challenging real-world applications such as problems in hy-
drocephalus treatment and shape optimization. The contents of this chapter have been accepted
We formulate and solve several mesh optimization problems with competing goals. These
competing goals can be categorized into four groups, i.e., improving mesh qualities based on the
element shape, element size, the associated PDE interpolation error, and number of inverted el-
ements. We assume the metrics will be used in the finite element solution of Poisson’s equation
for the isotropic physics case [41]. A similar approach can be taken for the solution of other
For each of the four goals, we describe mesh quality metrics which belong to each cat-
egory. For each mesh quality metric, the mesh quality of ith element on the mesh is denoted by
qi . The mesh qualities below are for the 3D case. Similar quantities can be defined in 2D.
63
Shape. The inverse mean ratio (IMR) quality metric considers both shape and size but focuses
more on the shape [18]. The IMR defines an incidence matrix for an ideal element shape, W ,
and transforms the incidence matrix of a non-optimal element to W. In the isotropic case, an
equilateral triangle or tetrahedron is considered to be the ideal element. The IMR metric also
uses an incidence matrix, C, and a barrier function, which prevents inverted elements. Let the
1 1
1 2 2
√ √
W =
0
3 3. (3.1)
2 6
√
6
0 0 3
2
CW −1 F
qi = 2 ,
3 |det(CW −1 )| 3
q
CW −1
−1
where F
= tr((CW )T (CW −1 )) is the Frobenius norm. For the IMR metric, a lower
qi value indicates a higher mesh quality. The value of the IMR metric lies between 1 and ∞ for
was proposed to minimize the condition number of the stiffness matrix which is associated with
Poisson’s equation [41]. The SI metric is defined as the reciprocal of the upper bound of the
do so, the ideal element shape is an equilateral tetrahedron for the solution of Poisson’s equation.
V
qi = 3 , (3.2)
2
Arms
where V is the volume of a tetrahedron, and Arms is the root-mean-square area of the tetrahedron’s
q
faces, which is given by Arms = 14 ∑4i=1 A2i . Here Ai is the ith unsigned area of a tetrahedron’s
face. Since a smaller maximum eigenvalue and condition number indicate a higher qi , a higher
qi value for the SI metric indicates a better mesh quality. The value of the SI metric lies between
√
0 and 6/9 for valid elements.
Size. The goal of optimizing the size-based metric is to adjust vertex positions to make the size
of the elements as uniform as possible. Let the volume of a tetrahedron be V. The size metric is
defined as
qi = V.
We employ the sum of squares of qi as the objective function. The value of the size metric lies
More generally, elements should be smoothly graded for numerical methods to behave
properly.
65
V ∑4k=1 Ak
qi = 2
,
∑1≤m≤n≤4 Am An lmn
where lmn is the edge length between the mth and nth vertices. The SS metric is not scale-
invariant; hence, the value of the SS metric lies between 0 and ∞ for valid elements.
Untangling. Knupp proposed an untangling mesh quality metric using a signed element area
or a volume [43]. The motivation behind his metric is to give high penalty (cost) to inverted
elements and no penalty to valid elements. Knupp’s untangling mesh quality metric is defined as
1
qi = (|V − β | − (V − β )), (3.3)
2
where β is a user defined parameter greater than 0. The value of the untangling metric lies be-
Contour Plots. Contour plots show different geometric features and can be used to identify
ideal shapes for each quality metric. Fig. 3.1 shows contour plots for the (a) IMR, (b) SI, (c) SS,
(d) size, and (e) untangling, respectively. In the plots, blue indicates higher mesh quality, and
red indicates lower mesh quality. Fig. 3.1 shows that the SS metric penalizes less a large or a
1 1 1
0 0 0
−1 −0.5 0 0.5 1 1.5 2 −1 −0.5 0 0.5 1 1.5 2 −1 −0.5 0 0.5 1 1.5 2
1.5 1.5
1 1
0.5 0.5
0 0
−1 −0.5 0 0.5 1 1.5 2 −1 −0.5 0 0.5 1 1.5 2
Fig. 3.1 Contour plots of the quality metric of a triangle as a function of a free vertex when the
two vertices are fixed at (0, 0) and (1, 0). In (e), β = 0.25.
In this section, we improve the average mesh quality by minimizing a single objective
function. We denote the number of mesh elements by |E| and the overall mesh quality by F.
For the IMR, size, and untangling metrics, we employ the sum of squares of individual element
qualities of qi :
|E|
F = ∑ q2i . (3.4)
i=1
|E|
1
F=∑ 2
. (3.5)
i=1 qi
We use the reciprocal of the quality metric instead of the additive inverse to optimize the mesh
because the optimization process for the latter results in meshes with elements of both very good
67
and very poor quality, which is not desirable [73]. Using the reciprocal results in smoothed
meshes with a more even distribution of element qualities. For the SI metric, we use a function
with the sum of the reciprocals of the element qualities, because inverted elements have negative
qi values, and the sum of squares function in (3.4) cannot differentiate between positive and
negative qi values.
optimize more than one aspect of a mesh such as the element size, shape, associated PDE in-
terpolation error, and number of inverted elements. Different from single objective optimization
Pareto optimality is a widely used concept to define an optimal solution for multiobjective opti-
A point, x∗ ∈ X, is Pareto optimal if and only if there does not exist another point, x ∈ X, such
that
F(x) ≤ F(x∗ ) and Fi (x) < Fi (x∗ ), for at least one function.
68
A point, x∗ ∈ X, is weakly Pareto optimal if and only if there does not exist another point, x ∈ X,
such that
F(x) ≤ F(x∗ ).
In this case, there is no other point that simultaneously decreases all of the objective functions.
There are various techniques for solving multiobjective optimization problems. Such
techniques can be categorized into three classes based on articulation of preferences. The three
articulation of preferences [91]. In methods with a priori articulation of preferences, the relative
importance (weight) of the objective functions are specified in advance before the optimization
problem is solved. These preferences can be viewed as specifying weights for each objective
function. One of the well-known methods, which belongs to this category, is a weighted-sum of
n
F̃ = ∑ wi Fi , (3.6)
i=1
erences, the preferences are chosen from a set of solutions, based on which solution is most
appealing to the user, without considering the relative importance of the objective functions.
The one method, which belongs to this category, is the genetic algorithm. In methods with no
articulation of preferences, preferences are not used in the solution of the optimization problem.
69
work combines n competing objective functions, F, into a single objective function, represented
by F̃ (e.g., (3.6)). Any multiobjective optimization methods, which can be combined into a sin-
gle objective function, can be employed to simultaneously optimize two or more aspects of the
mesh within our framework. We do not list all available methods, but, for example, the weighted
sum, weighted min-max, exponential weighted, lexicographic, and Rao’s methods can be em-
ployed. Multiobjective optimization methods which can be combined into a single objective
We develop three multiobjective mesh optimization methods within our framework, which
do not require articulation of preferences, namely the exponential sum, objective product, and
equal sum methods. One of the main advantages of multiobjective methods with no articulation
of preferences is that weights for different objective functions do not need to be specified, and
Exponential Sum Method. The first method we consider is the exponential sum method [90],
the maximum (worst) cost function. The general multiobjective form of the problem is shown
in (3.7). For the multiobjective mesh optimization problem, the goal of the min-max problem
is to minimize the maximum (worst) cost function. Min-max mesh optimization problems have
been studied in [46, 96]. However, researchers have only optimized one aspect of the mesh by
Li [90] proposed the exponential sum function to approximate the min-max problem by em-
ploying an exponential penalty function, because the min-max problem is not smooth. The
n
F̃ = c ln[ ∑ eFi /c ], (3.8)
i=1
where Fi is the ith objective function (described in Section 3.2.2), and c > 0 is the controlling
parameter. Typical values of c are between 104 and 106 . Li has shown that (3.8) is an approxima-
tion of maxi Fi . This implies that the exponential sum method is a smooth form of the min-max
problem. It was proven that the solution to (3.8) satisfies a sufficient condition of weak Pareto
optimality [47].
Objective Product Method. The second method we consider is the objective product method [47],
One advantage of the objective product method is that it does not require normalization of the
objective function. This is a nice property since many quality metrics are on widely different
scales. It is easy to see that reducing any objective function by some ratio yields the same prod-
uct value as does reducing any other objective function by that ratio. It was proven that (3.9)
Equal Sum Method. The third method we consider is the equal sum method [47]. If wi is set to
one in (3.6), the weighted-sum methods become an equal sum method. The equal sum method
71
It was proven that (3.10) satisfies a sufficient condition of Pareto optimality [48].
bining the mesh quality metrics in Section 3.2 with the multiobjective optimization problems
described in Section 3.3.1, one can define the specific optimization problem for the exponential
3 !
1 |E| 2 1 |E| A 2
∑i=1 12 (|Vi −β |−(Vi −β ))
[ ] +e ∑ j=1 rms
F̃ = c ln e c c V . (3.11)
Similar optimization problems can be defined for the objective product and equal sum
methods using other quality metrics. An examples of the optimization problem for the objective
product method using the untangling and SI quality metrics is shown in (3.12)
|E| 2 ! |E| 32
1 A
∑ rms .
F̃ = ∑ (|Vi − β | − (Vi − β )) (3.12)
i=1 2 j=1 V
An example of the optimization for the equal sum method using the same metrics is shown
in (3.13).
3
|E| 2 ! |E| 2
1 Arms
F̃ = ∑ (|Vi − β | − (Vi − β )) +∑ . (3.13)
i=1 2 j=1 V
Since (3.8), (3.9), and (3.10) are nonlinear optimization problems, existing nonlinear op-
timization methods can be employed to minimize them. We employ the nonlinear conjugate
gradient method (NLCG) to solve them, since NLCG requires only the gradient of the multiob-
Let xk be the solution (a vector of vertex positions) at the kth iteration. NLCG determines
a step length, αk by using a line search technique along a search direction, pk . The step length
xk+1 = xk + αk pk .
T
∇F̃(xk ) (∇F̃(xk ) − ∇F̃(xk−1 ))
βkPR = T . (3.15)
∇F̃(xk−1 ) ∇F̃(xk−1 )
At each iteration, the variation in NLCG is from the computation of βkPR . We employ the Polak-
Ribière NLCG method, since we observe in our preliminary experiments that the Polak-Ribière
NLCG method is more efficient in finding a local minimum for our problems.
73
3.3.3 Comparison among the Exponential Sum, Objective Product, and Equal Sum
Methods
We compare the exponential sum, objective product, and equal sum multiobjective mesh
optimization methods using a toy example. This example shows the effect of various multiob-
For this experiment, we initially generate a structured mesh on the rectangular domain
(Fig. 3.2(a)) and perturb one interior vertex in order to generate inverted elements (Fig. 3.2(b)).
We employ both the shape (SI) and untangling metric and observe the effect of using the expo-
nential sum and objective product multiobjective methods. We perform one iteration of mesh
optimization using the NLCG implemented in Mesquite [16] to find a Pareto optimal point for
Figs. 3.2(c), 3.2(d), and 3.2(e) show the resulting meshes after one iteration of the ex-
ponential sum, objective product, and equal sum multiobjective mesh optimization methods,
respectively. We observe that, on this problem, the exponential sum method eliminates in-
verted elements, whereas the equal sum and objective product methods are not able to do so.
Figs. 3.2(f), 3.2(g), and 3.2(h) show output meshes with inaccurate mesh optimization for the
exponential sum and objective product multiobjective methods, respectively. Here, inaccurate
mesh optimization means the mesh quality does not change up to two digits. All three methods
can eliminate inverted elements, but the exponential sum and the equal sum methods yield more
uniform meshes with good element quality. Here, the optimal mesh of the exponential sum and
Since the exponential sum method is used to solve a min-max problem, it improves the
worst objective function among the Fi . As a result, for this toy example, the exponential sum
method succeeds in eliminating inverted elements, because it focuses on eliminating the inverted
elements which correspond to the worst Fi . On the other hand, the objective product method
improves the overall element quality by placing equal weights on the Fi , but fails to eliminate
inverted elements, although it improves the overall element shapes. Similar to the objective
product method, the equal sum method also equally improves two competing quality metrics.
However, the equal sum method behaves a lot differently if there exist inverted elements in the
mesh. This is because the qi values for the inverted elements are negative for many quality
metrics (e.g., IMR and SI) but are positive for the untangling metric.
If the initial mesh has inverted elements or several poor quality elements, the exponential
sum multiobjective method will be effective for multiobjective mesh optimizations because it
focuses on improving the worst objective function. For some tangled meshes, which include in-
verted elements with very small areas or volumes (close to zero), the qi values for the untangling
metric in (3.3) could be very small compared with the qi for the SI (shape) metric in (3.2), which
is scale-invariant. For these cases, the exponential sum method is more effective for elimination
of inverted elements if the value of qi for the untangling metric is rescaled by multiplying a
constant such that the maximum qi value for the untangling metric is bigger than than the maxi-
mum qi value for the SI metric. This is because the exponential sum method improves the worst
The objective product multiobjective method will be more useful if the initial mesh qual-
ity has few poor quality elements and the goal is to improve the overall mesh qualities using
75
equal weights. If the initial mesh is close to optimal, we observe that the objective product mul-
tiobjective method sometimes converges faster to the optimal mesh than do the exponential sum
In this section, we describe our numerical experiments used to evaluate our mesh opti-
mization methods. Table 3.1 summarizes the experiments and goals and gives a description of
the goals for each problem. For Experiment 3.4.2, we use barrier functions, i.e., |A| > 0 for
2D meshes and |V | > 0 for 3D meshes, to prevent inverted elements when using the SS metric.
For Experiment 3.4.3, we start with tangled meshes and investigate whether our mesh opti-
mization methods are able to simultaneously improve the mesh quality and eliminate inverted
elements. For Experiment 3.4.4, we investigate whether our mesh optimization methods can
generate meshes which optimize three aspects of the mesh, such as the element shape and size,
and result in no inverted elements. For the size metric, we report root mean square (rms) values
of element qualities, since meshes with small rms values indicate meshes with uniformly-sized
elements.
Different from previous approaches, the strength of our multiobjective mesh optimization
framework lies in its ability to be extended to improve any aspects of the mesh and to untangle
the mesh. Our framework is not limited to the goals which are listed in Table 3.1; rather, it can be
easily applied to other categories, such as orientation and regularity [49], by simply employing
a different quality metric or objective function. Other element types can be also employed (e.g.,
(c) The output mesh of the ex- (d) The output mesh of the (e) The output mesh of the
ponential sum multiobjective objective product multiobjec- equal sum multiobjective
method after one iteration of tive method after one iteration method after one iteration of
mesh optimization of mesh optimization mesh optimization
(f) The optimal mesh of the (g) The optimal mesh of the (h) The optimal mesh of
exponential sum multiobjec- objective product multiobjec- the equal sum multiobjective
tive method after accurate tive method after accurate method after accurate mesh
mesh optimization mesh optimization optimization
Fig. 3.2 Comparison among the exponential sum, objective product, and equal sum multiobjec-
tive mesh optimization methods
77
For each experiment, we perform numerical experiments for both single and multiobjec-
tive mesh optimization algorithms (described in Sections 3.2 and 3.3, respectively). Note that our
multiobjective mesh optimization algorithms result in meshes which are weakly Pareto optimal,
whereas single objective algorithms result in meshes which are locally optimal. We also report
on results for the worst element quality, because the worst element quality plays a critical role
The machine employed for this study is equipped with an Intel Core2 Duo Processor
Geometric Domains. The four geometric domains considered in our experiments are shown in
Fig. 3.3. Freefem++ [50] software was used to generate the initial meshes. Half of the interior
vertices in each mesh were perturbed to create test meshes that were further from optimal. For
Experiments 3.4.1, 3.4.2, and 3.4.3, we employ meshes with approximately 1,000,000 elements.
For Experiment 3.4.4, we employ meshes with approximately 50,000 elements, because they are
(a) Disk (2D) (b) Barrier (2D) (c) Disk (3D) (d) Barrier (3D)
Fig. 3.3 Coarse initial meshes on the disk and barrier geometric domains indicative of the actual
meshes to be smoothed.
Mesh Smoothing, Untangling, and Nonlinear Solver. For Experiments 3.4.1 - 3.4.4, a local
implementation of the NLCG method (described in Section 3.3.2) was used to solve the mesh
optimization problems. We employed the Polak-Ribière NLCG method in Mesquite [16] to solve
each mesh optimization problem, because our preliminary experiments show that it is the fastest
amongst different versions of the nonlinear conjugate gradient method. We perform inaccurate
mesh smoothing as in [31, 73] such that the mesh qualities are the same up to two digits after the
decimal point. We do not perform accurate mesh smoothing because accurate mesh smoothing
(fully converged solution) significantly increases the smoothing time, but slightly increases the
mesh quality.
For this experiment, we consider two competing goals: improving both the shape (IMR)
and size metrics. Fig. 3.4 shows the average element qualities based on the (a) shape (IMR)
and (b) size metrics on the 3D disk mesh, respectively. We observe that the size with the single
objective mesh optimization method yields the worst result among the four different mesh opti-
mization algorithms when the shape metric is used to measure the mesh quality. Similarly, the
shape with the single objective mesh optimization method shows the worst results with respect
79
to the size metric. Our two multiobjective mesh optimization algorithms improve the average
Since the ideal element shape is an equilateral tetrahedron for optimization of the IMR
metric, we can expect that the elements in the output mesh employing the IMR metric will have
similar volumes. However, the size metric with the single objective method does not result in
equilateral elements, because two different elements can have the same volumes although they
have different shapes. Therefore, optimizing the mesh with respect to the size metric only can
Fig. 3.5 shows the worst element qualities based on the (a) IMR and (b) size metrics on
the 3D disk mesh, respectively. Fig. 3.5(a) shows that the size metric with the single objective
method has a negative effect on the element shape of the mesh. As the number of iterations in-
creases, the worst element quality of the size metric with the single objective method measured
by the IMR metric deteriorates significantly. Fig. 3.5(b) shows that the IMR metric with the
single objective method has a positive effect on the size metric for the first few iterations. Note
that the optimization of the IMR metric can be beneficial for improving the size metric. This
is because the ideal element shape when employing the IMR metric is an equilateral triangle or
tetrahedron. Experimental results show that our exponential sum multiobjective mesh optimiza-
tion algorithm improves the worst element quality by 96.6% when compared with the use of the
size metric with the single objective method when measured by the IMR metric.
Table 3.2 shows the smoothing times for various mesh optimization methods in terms of
the number of iterations of mesh smoothing. The size metric with the single objective method
shows the fastest smoothing time, because the numerical computation of the size metric is quite
80
simple compared with the computation for the other mesh optimization methods. For 20 iter-
ations of mesh smoothing, our objective product mesh optimization algorithm takes approxi-
mately two times longer compared with the size metric with the single objective mesh optimiza-
tion method. We also observe similar trade off between mesh quality improvement and efficiency
Number of iterations
Mesh optimization method
1 5 10 20
IMR single objective 24.6 75.9 149.0 290.3
Size single objective 21.0 59.3 105.6 196.5
Exponential sum multiobjective (IMR and size) 56.4 202.2 390.1 450.0
Objective product multiobjective (IMR and size) 28.9 100.3 197.8 392.2
Equal sum multiobjective (IMR and size) 33.1 117.5 233.2 439.5
Table 3.2 Smoothing time (secs) for various mesh optimization methods in terms of the number
of iterations of smoothing on the 3D disk mesh.
For this experiment, we consider two competing goals: improving both shape (IMR) and
interpolation error (SS) metrics. Fig. 3.6 shows the average element qualities based on the (a)
shape (IMR) and (b) interpolation error (SS) metrics on the 3D disk mesh. We observe that our
multiobjective mesh optimization algorithms significantly improve the average mesh qualities
based on both the IMR and SS metrics. Fig. 3.6 also shows that the SS metric with the single
objective method fails to improve the IMR metric. In our previous paper [73], and in the contour
plots in Section 3.2, we observed that the ideal shape when employing the SS metric is different
from the one when employing the IMR metric, because the SS metric penalizes small angles less
than do other mesh quality metrics. Therefore, the SS metric with the single objective method
can have a negative effect on the IMR metric. However, the IMR metric with the single objective
81
2.4
IMR single objective
2.3 Size single objective
Average element quality
1.9
1.8
1.7
0 5 10 15 20
Number of iterations of mesh smoothing
(a) Average element quality as measured by shape (IMR)
0.98
Average element quality (rms)
0.9
0.88
0.86
0.84
0 5 10 15 20
Number of iterations of mesh smoothing
(b) Average element quality as measured by size (volume)
Fig. 3.4 (a) Average element quality in terms of the shape metric (IMR) on the 3D disk mesh;
(b) average element quality in terms of the size metric (volume) on the same mesh
82
2
0 5 10 15 20
Number of iterations of mesh smoothing
(a) Worst element quality as measured by shape (IMR)
12
Worst element quality (log scale)
2
0 5 10 15 20
Number of iterations of mesh smoothing
(b) Worst element quality as measured by size (volume)
Fig. 3.5 (a) Worst element quality in terms of the shape metric (IMR) on the 3D disk mesh; (b)
worst element quality in terms of the size metric (volume) on the same mesh
83
method can be beneficial in terms of the SS metric. This is because the initial mesh has lots of
Fig. 3.7 shows the worst element qualities based on the (a) IMR and (b) SS metrics.
The overall trend is similar to the one with the average element qualities. Experimental results
show that our exponential sum multiobjective mesh optimization algorithm improves the worst
element quality by 95.4% compared with the SS metric with the single objective method ac-
cording to the IMR metric. Similar to the average element qualities, the IMR metric with the
single objective method can also improve the quality of the SS metric. However, after 10 itera-
tions of mesh smoothing, the worst element quality employing the IMR metric slightly increases
(deteriorates) because of the different ideal element shapes for the IMR and interpolation error
metrics.
Table 3.3 shows smoothing times for the various number of iterations of mesh smoothing
for the mesh optimization methods. The IMR metric shows the fastest smoothing time because
IMR is highly optimized in Mesquite. For the 20 iterations of mesh smoothing, our equal sum
multiobjective method takes four times longer than the IMR metric with the single objective
method, although it significantly improves both mesh quality metrics simultaneously. These
Number of iterations
Mesh optimization method
1 5 10 20
IMR single objective 27.1 77.3 143.5 309.6
SS single objective 66.1 306.2 583.1 1226.5
Exponential sum multiobjective (IMR and SS) 89.2 400.5 848.3 1423.2
Objective product multiobjective (IMR and SS) 88.9 382.4 734.6 1425.0
Equal sum multiobjective (IMR and SS) 70.7 301.1 599.2 1217.6
Table 3.3 Mesh smoothing time (secs) for various mesh optimization methods in terms of the
number of iterations of smoothing on the 3D disk domain.
84
0.34
10
0.32
10 IMR single objective
SS single objective
0.3 Exponential sum multiobjective
10
(IMR and SS)
Objective product multiobjective
0.28 (IMR and SS)
10
Equal sum multiobjective
(IMR and SS)
0.26
10
0 5 10 15 20
Number of iterations of mesh smoothing
(a) Average element quality as measured by shape (IMR)
2
IMR single objective
1.9 SS single objective
Average element quality
1.6
1.5
1.4
1.3
0 5 10 15 20
Number of iterations of mesh smoothing
(b) Average element quality as measured by interpolation
error (SS)
Fig. 3.6 (a) Average element quality as measured by the shape metric (IMR) on the 3D disk
mesh; (b) average element quality as measured by the interpolation error (SS) on the same mesh
85
1
10
0
10
0 5 10 15 20
Number of iterations of mesh smoothing
(a) Worst element quality as measured by shape (IMR)
4
10
Worst element quality (log scale)
1
10
0
10
0 5 10 15 20
Number of iterations of mesh smoothing
(b) Worst element quality as measured by interpolation er-
ror (SS)
Fig. 3.7 (a) Worst element quality in terms of the shape metric (IMR) on the 3D disk mesh; (b)
worst element quality in terms of the interpolation error (SS) metric on the same mesh
86
improve the element shape and eliminate inverted elements in tangled meshes. Approximately
1% of the elements in the initial mesh are inverted such that the average mesh quality and the
maximum mesh quality of the initial tangled mesh based on the untangling quality metric are
0.0027792 and 3.8103, respectively. This experiment is challenging because the SI metric with-
out a barrier often generates inverted elements during mesh optimization. For tangled meshes,
we employ the SI metric instead of the IMR metric, because the IMR metric includes a barrier
to prevent inverted elements and the IMR metric in Mesquite [16] is not able to compute valid
function gradients and function values for inverted elements. However, the SI metric can be used
to improve the shape of the mesh elements, because it does not include the barrier in the metric.
Also, the ideal element for both the IMR with W in (3.1) and SI metrics is the equilateral triangle
(tetrahedron).
Tab. 3.4 shows the numbers of inverted elements after employing various mesh optimiza-
tion methods. The untangling metric with the single objective method eliminates all inverted
elements after three iterations of mesh optimization. Our three multiobjective methods success-
fully eliminate all inverted elements within three iterations of mesh optimization. Note that both
our exponential sum and equal sum multiobjective algorithm even eliminates inverted elements
in fewer iterations of mesh optimization than does the untangling metric with the single objec-
tive method. This is because our multiobjective algorithm tries to simultaneously improve mesh
Fig. 3.8 shows the output meshes for four different mesh optimization methods after 10
iterations of mesh optimization on the 2D barrier mesh. The initial mesh includes 543 inverted
elements and exhibits poor mesh quality in terms of the SI (shape-based) metric. Since the ideal
element shape for the SI metric is assumed to be an equilateral element, the multiobjective mesh
optimization algorithms improve the element shape while eliminating inverted elements. Fig. 3.8
also shows that the exponential sum multiobjective mesh optimization algorithm yields the best
element shape among five mesh optimization methods. The elements in the final mesh resulting
from the exponential sum method are closer to equilateral compared with the elements of the
These experimental results are consistent with the toy example in Section 3.3.3 in that
the exponential sum method is more effective in eliminating inverted elements than the objective
product method, because it focuses on improving the worst objective function, which is the
In terms of the mesh optimization times, all three multiobjective methods were able to
simultaneously improve element qualities while eliminating inverted elements in less than one
minute.
Comparison with the Simultaneous Untangling and Smoothing (SUS) Algorithm of Esco-
bar et al. on a 3D Tangled Mesh. One of the existing algorithms for simultaneously improving
element shape and performing mesh untangling is due to Escobar et al. [44]. This algorithm is
called Simultaneous Untangling and Smoothing (SUS) and is able to optimize only two specific
aspects (element shape and untangling) of the 3D tetrahedral mesh. It cannot easily be extended
to optimize other aspects of the mesh. We compare our methods with the SUS algorithm using
88
Number of iterations
Mesh optimization method
0 1 2 3 10
SI single objective 543 1,019 1,030 1,030 1,030
Untangle single objective 543 10 1 0 0
Exponential sum multiobjective
543 2 0 0 0
(SI and Untangle)
Objective product multiobjective
543 31 2 0 0
(SI and Untangle)
Equal sum multiobjective
543 5 0 0 0
(SI and Untangle)
Table 3.4 Number of inverted elements as a function of the number of iterations of mesh opti-
mization. The initial mesh has 543 inverted elements.
Number of iterations
Mesh optimization method
0 1 2 3 4 5
SI single objective 111 1,392 1,390 1,389 1,389 1,389
Untangle single objective 111 5 0 0 0 0
Exponential sum multiobjective
111 2 0 0 0 0
(SI and Untangle)
Objective product multiobjective
111 2 0 0 0 0
(SI and Untangle)
Equal sum multiobjective
111 2 0 0 0 0
(SI and Untangle)
Simultaneous untangling and smoothing
111 0 0 0 0 0
(SUS) [44]
Table 3.5 Number of inverted elements as a function of the number of iterations of mesh opti-
mization. The initial mesh has 111 inverted elements.
89
(a) (b)
(c) (d)
(e) (f)
Fig. 3.8 2D barrier mesh: (a) the initial mesh with inverted elements (the elements circled in red);
(b) the final mesh with inverted elements (the elements circled in red) when employing the SI
metric with the single objective method. The number of inverted elements increases after mesh
optimization; (c) the final mesh when employing the untangling metric with the single objective
method; (d) the final mesh for the exponential sum multiobjective method; (e) the final mesh for
the objective product multiobjective method; (f) the final mesh for the equal sum multiobjective
method. The final mesh by employing the exponential sum multiobjective method yields the
best element quality while eliminating all inverted elements.
90
−1.2
10 Exponential sum multiobjective
Objective product multiobjective
Equal sum multiobjective
SUS [6]
−1.3
10
0 5 10 15 20
Number of iterations of mesh optimization
(a) Average element quality as measured by shape (SI)
2
10
Worst element quality
0
10
−1
10
0 5 10 15 20
Number of iterations of mesh optimization
(b) Worst element quality as measured by shape (SI)
Fig. 3.9 (a) Average element quality in terms of the shape metric (IMR) on the 3D disk mesh;
(b) worst element quality in terms of the size metric (volume) on the same mesh
91
their source codes posted in [51] on the tangled 3D disk domain, which includes 111 inverted
elements.
Table 3.5 shows the number of inverted elements in terms of the number of iterations of
mesh optimization on the 3D disk domain. We observe that the SUS algorithm is able to elimi-
nate all inverted elements after one iteration of mesh optimization, while our three multiobjective
mesh optimization methods each take one more iteration to remove inverted elements.
Fig. 3.9 shows both the average and worst element qualities based on the SI metric
(shape). Both the exponential sum and equal sum multiobjective methods significantly improve
the average and worst element qualities on the mesh and outperform the SUS algorithm. The
SUS algorithm improves the element quality after one iteration of mesh optimization; however,
the element quality does not improve and sometimes even deteriorates as the mesh optimization
proceeds. The shape metrics used in this work and by the SUS algorithm are different. However,
these results are also consistent with the results we obtained with the shape metric in [44] is
For other meshes, the SUS algorithm is able to eliminate inverted elements while improv-
ing element qualities. However, the element qualities are not as good as the element qualities
obtained by our methods. Also recall that our multiobjective methods are flexible and can opti-
mize any aspects of the mesh; however, the SUS algorithm is able to optimize only two specific
For this experiment, we consider three competing goals: improving the shape and size
metrics and eliminating inverted elements. The 3D disk mesh which include 111 inverted ele-
Tab. 3.6 shows the number of inverted elements in terms of the number of iterations of
mesh optimization. Our three multiobjective methods and the untangling metric with the single
objective algorithm are able to eliminate all the inverted elements. However, only the exponential
sum and the equal sum multiobjective mesh optimization methods improve the element shape
while eliminating all inverted elements. Similar to Experiment 3.4.3, our exponential sum mul-
tiobjective mesh optimization algorithm eliminates all inverted elements with fewer iterations.
This is because improving the SI metric also helps to eliminate inverted elements. In summary,
both our exponential sum and equal multiobjective mesh optimization algorithms improve two
competing mesh quality metrics while eliminating inverted elements in a tangled mesh.
Fig.10 shows the average element quality based on the (a) shape (SI) and (b) size metrics
for six mesh optimization algorithms on the 3D disk domain. We observe similar results for other
geometric domains. Among the six mesh optimization methods, only our exponential sum and
equal sum multiobjective mesh optimization algorithms are able to simultaneously improve two
different mesh quality metrics while eliminating inverted elements. The size single objective
method is able to simultaneously improve the element shape and size but it fails to eliminate
inverted elements.
93
0.07
Average element quality
0.065 SI single objective
Size single objective
0.06 Untangle single objective
Exponential sum multiobjective
Objective product multiobjective
0.055 Equal sum multiobjective
0.05
0.045
0.04
0 2 4 6 8 10
Number of iterations of mesh optimization
(a) Average element quality as measured by shape (SI)
−3
x 10
2.8
Average element quality (rms)
2.7
2.6
SI single objective
2.5 Size single objective
Untangle single objective
2.4
Exponential sum multiobjective
Objective product multiobjective
2.3
Equal sum multiobjective
2.2
2.1
1.9
0 2 4 6 8 10
Number of iterations of mesh optimization
(b) Average element quality as measured by size (volume)
Fig. 3.10 (a) Average element quality in terms of the shape metric (SI) on the 3D disk mesh; (b)
Average element quality in terms of the size (volume) metric on the same mesh
94
Number of iterations
Mesh optimization method
0 1 2 5 10
SI single objective 111 1,392 1,390 1,389 1,326
Size single objective 111 12 6 5 8
Untangle single objective 111 5 0 0 0
Exponential sum multiobjective
111 0 0 0 0
(SI, Size, and Untangle)
Objective product multiobjective
111 1 1 0 0
(SI, Size, and Untangle)
Equal sum multiobjective
111 1 0 0 0
(SI, Size, and Untangle)
Table 3.6 Number of inverted elements as a function of the number of iterations of mesh opti-
mization. The initial 3D disk mesh has 111 inverted elements.
Applications
In this section, we consider real-world applications and demonstrate how our multiobjec-
tive mesh optimization methods within our framework can be used to simultaneously optimize
multiple aspects of the mesh. First, we study the effect of employing the multiobjective mesh
optimization methods for problems in which warped meshes have created meshes with inverted
elements and elements with poor qualities. For the first two experiments, we investigate whether
the exponential sum, objective product, and equal sum multiobjective mesh optimization meth-
ods can simultaneously untangle the inverted mesh and improve the element qualities. We also
study the effect of employing the exponential sum, objective product, and equal sum multiobjec-
tive mesh optimization methods (shape and size) on the efficiency of the finite element solution
process.
In this process, we consider the solution of the following Poisson’s equation: −∆u = 1
with u = 0 on the boundary. We employ the piecewise linear Galerkin finite element method as
95
in [73] and use PETSc [17] to compute the maximum eigenvalue, minimum eigenvalue, condi-
tion number of the stiffness matrix A, and solve the resulting linear system. The Jacobi precon-
ditioner (P) with the minimal residual (MINRES) solver is used to solve the linear system.
We consider two application problems with deforming domains in Sections 3.5.1 and
3.5.2. For these problems, both mesh warping and remeshing can be used to generate meshes
on deformed domains. However, mesh warping is preferred to remeshing for many mesh de-
formation problems, because remeshing results in the accumulation of large interpolation errors
between successive time steps and less accurate PDE simulation results [52]. Also, similar
meshes, which have the same element connectivity, are desired between successive time steps
such that the solution varies smoothly during mesh deformations [54, 88].
Hydrocephalus (also called water in the brain) is a neurological condition and is one of
the most common neurological disorder treated by neurosurgeons [81]. A problem with the flow
of cerebrospinal fluid (CSF) results in hydrocephalus and causes the brain ventricles to become
enlarged as shown in Fig. 3.11(a). Medical treatment through CSF flow diversion by surgically
implanting a CSF shunt in the brain ventricles is one common treatment of decreasing the volume
of the fluid in the brain ventricles as shown in Fig. 3.11(b). The segmented images of the brain
were provided by the authors in [81]. More details on the segmentation of the medical images
We consider a 2D hydrocephalus mesh warping problem for tracking the evolution of the
ventricles after treatment. The initial and deformed domains for the mesh warping problem are
shown in Fig. 3.11(c) and Fig. 3.11(d), respectively. The initial brain mesh was provided by the
96
authors in [81]. The boundary vertices are known, and the deformed mesh was obtained by using
FEMWARP [88] to warp the initial mesh to the target brain as shown in Fig. 3.11(d). (Note that
the brain boundary (i.e., the external boundary) (not shown) was also provided by the authors
in [81].) For this example, FEMWARP generates inverted elements due to large deformations as
shown in Fig. 3.11(d). We do not compare with the SUS algorithm [44] on the 2D hydrocephalus
domain since the SUS algorithm and source codes posted in [51] are designed for 3D tangled
meshes.
Figs. 3.11(e) to 3.11(h) show close-up views of the meshes before smoothing, after ap-
plying 40 iterations of the exponential sum mutiobjective method and the untangle single objec-
tive method, respectively. The initial mesh has 8,166 elements and no inverted elements. The
warped mesh after employing FEMWARP has 13 inverted elements [81]. Even though smoothed
mesh in Figs. 3.11(f) has no inverted elements, the elements of the meshes in Fig. 3.11(g) and
Fig. 3.11(h) have better shapes and improve the worst quality element (i.e., the very skinny trian-
gle in the middle of the mesh). Note that this example is a challenging problem because there are
numerous interior vertices near the internal boundary on the deformed domain which are highly
Table 3.7 shows the number of inverted elements in terms of the number of iterations of
mesh optimization on the 2D hydrocephalus domain (Fig. 3.11(d)). The exponential sum multi-
objective method eliminates inverted elements after 9 iterations of mesh optimization, while the
untangling single objective method requires 33 iterations. As discussed earlier, the exponential
sum method minimizes the worst objective function, which is the untangling objective function
in this case. However, the objective product method improves the overall qualities with respect
97
to both shape and untangling. Although it slowly decreases the number of inverted elements, it
Number of iterations
Mesh optimization method
0 1 5 9 12 20 33
SI single objective 13 31 33 33 33 33 33
Untangle single objective 13 7 4 6 3 1 0
Exponential sum multiobjective (SI and Untangle) 13 9 1 0 0 0 0
Objective product multiobjective (SI and Untangle) 13 9 8 6 6 6 6
Equal sum multiobjective (SI and Untangle) 13 9 3 1 0 0 0
Table 3.7 Number of inverted elements as a function of the number of iterations of mesh opti-
mization. The initial mesh has 13 inverted elements.
The worst element quality on the mesh highly affects the efficiency for solving elliptic
PDEs (also Poisson’s equation), because both the condition number of the stiffness matrix and
the efficiency are closely related with the element shape on the mesh. Poorly-shaped elements
increase the maximum eigenvalue and condition number of the stiffness matrix [41]. The two-
Table 3.8 shows the maximum eigenvalue, minimum eigenvalue, and condition number
of the stiffness matrix when the above Poisson’s equation is solved on meshes using various
mesh optimization methods. Both the exponential sum and equal sum multiobjective methods
have smaller maximum eigenvalues because they have fewer poorly-shaped elements than does
the untangle single objective method. As a result, the exponential sum multiobjective method
has a 19% smaller condition number and is more efficient when solving Poisson’s equation
(a) Segmented image of the initial hydro- (b) Segmented image of the deformed hy-
cephalus domain (before treatment) [81] drocephalus domain (after treatment) [81]
(c) Initial mesh with no inverted el- (d) Deformed mesh on the hydrocephalus
ements on the hydrocephalus domain, domain (Fig. 3.11(b)) deformed with
Fig. 3.11(a) [81] FEMWARP. This mesh has 13 inverted
elements.
(e) Close-up view of (f) Close-up view of (g) Close-up view of (h) Close-up view
the mesh in 3.11(d) the mesh after apply- the mesh after apply- of the mesh after
with inverted elements ing 40 iterations of ing 40 iterations of applying 40 iterations
enclosed by the oval the untangling single the exponential sum of the equal sum
objective method on multiobjective method multiobjective method
the mesh shown in on 3.11(d). This mesh on 3.11(d). This
Fig. 3.11(d). This has no inverted ele- mesh has no inverted
mesh has no inverted ments. elements.
elements.
Fig. 3.11 Segmented images and meshes of initial and deformed domains.
99
Table 3.8 Maximum eigenvalue, minimum eigenvalue, condition number of P−1 A, and the num-
ber of iterations required to converge to solve Poisson’s equation on the 2D hydrocephalus do-
main. Here, P−1 A is the preconditioned stiffness matrix. A ’*’ denotes invalid PDE solution due
to inverted elements on the mesh.
For this experiment, we consider a problem from mesh warping for shape optimization on
3D geometric domains [94]. For shape optimization problems, geometric parameters need to be
determined automatically [94]. However, for general shape optimization problems on deforming
domains, it is not simple to create mapping functions from the initial domain onto the deformed
domain. Therefore, mesh warping algorithms are often used to find the new locations of vertices
on deformed domains. Similar to the mesh warping problem in Sec. 3.5.1, the boundary vertices
on the deformed domain are known and the meshes are warped from the initial domain to the de-
formed domain. The meshes were provided by the authors in [94]. In this case, FEMWARP [88]
was employed for mesh warping. The initial mesh, which is shown in Fig. 3.12(a), contains
109,535 elements; none of the elements are inverted. However, the deformed mesh, which is
shown in Fig. 3.12(b), contains 109,535 elements with 13 inverted elements. Note that the com-
putation time for using FEMWARP on these meshes is less than one second on an HP ProLiant
linux workstation with two X5570 Intel quad cores with 24 GB RAM [94].
100
The SI single objective mesh optimization method is not able to eliminate inverted el-
ements. However, the other three mesh optimization methods, which are the exponential sum,
the objective product, and untangle single objective methods, are able to untangle the inverted
elements in one iteration of mesh optimization. Table 3.9 shows the maximum eigenvalue, min-
imum eigenvalue, and condition number of the preconditioned stiffness matrix when Poisson’s
equation is solved using various mesh optimization methods. Similar to its performance on other
problems, both the exponential sum and equal sum multiobjective mesh optimization methods
outperform other mesh optimization methods. Similar to the comparison with the SUS algorithm
in Section 3.4.3, the element qualities of the output mesh from the SUS algorithm [44] are worse
than those obtained by our multiobjective methods. Therefore, the stiffness matrix generated by
the SUS algorithm has a larger condition number and takes more iterations to solve Poisson’s
(a) Initial mesh with no inverted elements (b) Deformed mesh on the 3D bore do-
on the 3D bore domain [94] main [94] deformed with FEMWARP [88].
This mesh has 13 inverted elements which
are located inside the domain.
Table 3.9 Maximum eigenvalue, minimum eigenvalue, condition number of P−1 A, and the num-
ber of iterations required to converge to solve Poisson’s equation on the 3D bore domain. Here,
P−1 A is the preconditioned stiffness matrix. A ’*’ denotes invalid PDE solution due to inverted
elements on the mesh.
3.5.3 Effect of Simultaneously Optimizing Both Shape and Size Metrics on the Efficiency
We study the effect of employing shape (IMR) and size multiobjective mesh optimization
methods on the condition number of the stiffness matrix and the number of iterations for solving
Poisson’s equation on the 2D disk domain (Fig. 3.3(a)) with 500K elements.
The dominant factor which affects the condition number of the stiffness matrix is the
element shape rather than element size. However, when two meshes have the same number of
elements, we can expect that a more uniform mesh (more smoothly graded mesh) has a smaller
condition number than a nonuniform mesh if the mesh element have the same shapes. This is
because the mesh with uniform-sized elements has a larger λmin than does a nonuniform mesh.
The smallest-sized element in the non-uniform mesh decreases the minimum eigenvalue of the
Table 3.10 shows the maximum eigenvalue, minimum eigenvalue, and condition number
of the preconditioned stiffness matrix when Poisson’s equation is solved using various mesh
optimization methods. The exponential sum multiobjective mesh optimization method has the
smallest condition number based on its small λmax (P−1 A) and its large λmin (P−1 A). The size
single objective mesh optimization method shows the slowest convergence rate because of its
Table 3.10 Maximum eigenvalue, minimum eigenvalue, and condition number of P−1 A, and
number of iterations to converge to solve Poisson’s equation. Here, P−1 A is the preconditioned
stiffness matrix.
improvement of various aspects of the mesh with the goal of improving the accuracy, efficiency,
stability, and conditioning of the associated finite element solver. The novelty of our framework
lies in its ability to combine multiple aspects of the mesh, besides just smoothing and untangling,
for mesh optimization. Specifically, we develop three methods within our framework, namely
the exponential sum, objective product, and equal sum mesh optimization methods, which do
not require articulation of preferences. Therefore, the relative preferences for different objective
103
functions do not need to be specified. Moreover, our solutions satisfy a sufficient condition of
Our experimental results show that our multiobjective mesh optimization methods im-
prove the worst element qualities by up to 96.6% compared with the corresponding single ob-
jective mesh optimization algorithms, respectively. Our experimental results also show that, in
most cases, the exponential sum multiobjective method is the most effective out of the three
methods we tested in simultaneously improving two or more aspects of the mesh. The expo-
nential sum multiobjective method is effective in improving the worst cost objective function
efficiently when the initial mesh quality is poor and the mesh contains inverted elements. The
equal sum multiobjective method is simple and shows comparable performance to the expo-
nential sum method. Experimental results show that both our exponential sum and equal sum
multiobjective methods outperform the existing simultaneous untangling and smoothing (SUS)
algorithm. Moreover, both our exponential sum and equal sum multiobjective methods improve
two competing mesh quality metrics while eliminating inverted elements in a tangled mesh.
our multiobjective mesh optimization methods are more effective in terms of simultaneously
untangling and improving the element shape than both the untangle single objective method
and SUS. In addition, our methods are more effective at decreasing the condition number of
the stiffness matrices and the number of iterations needed to solve elliptic PDEs on the meshes
than are the relevant single objective mesh optimization methods. In addition, for the 2D hydro-
cephalus treatment application, our exponential sum multiobjective mesh optimization method
significantly decreases the number of iterations required to untangle all inverted mesh elements
quality improvement and mesh untangling. In this dissertation, we have thus far focused on
mesh optimization techniques for PDE problems. However, traditional PDE methods are not
able to compute derivatives for discontinuous domains. However, new theory based on inte-
gral equations called a nonlocal peridynamics model has recently developed to address these
issues. In the next chapter, we investigate the effects that mesh anisotropy and size (i.e., mesh
refinement) have on the conditioning of the stiffness matrix for a nonlocal peridynamics model.
105
Chapter 4
The Effect of Anisotropy, Mesh Refinement, and Kernel
4.1 Introductions
Many applications in solid mechanics such as damage, surface cracks, and fracture have
discontinuities in their structures. Classical partial differential equations (PDEs) have some lim-
itations in their modeling and computation of discontinuous domains, since they are not able to
compute derivatives on such domains. In order to address this limitation of classical PDEs, a
nonlocal peridynamic model was recently developed by Silling [56]. The peridynamics nonlocal
theory uses integral equations rather than differential equations to model cracked surfaces and
deformations. It was reported that this peridynamics model was successfully applied to other
applications such as turbulence [57], nanofibers [61, 62], porous flow [60], and fracture and
damage modeling of membranes [61]. The reader is referred to [63, 64] for recent surveys on
nonlocal peridynamics models and applications. There have been various studies on nonlocal
peridynamics models. The connection between classical PDEs (such as an elasticity theory) and
106
nonlocal peridynamics was investigated [65]. This study showed that a nonlocal peridynamics
model is close to a classical local model when the length scale (horizon) goes to zero. The effect
of various kernel functions on the nonlocal advection problem was investigated for a 1D prob-
lem [67]. Researchers have also studied finite difference and finite element discretization of non-
local peridyanmics models [68–71]. Specifically, a posteriori error analysis and the connection
between the horizon and the condition number was studied in [68]. Condition number estimates
and upper bounds for the discretized linear system were studied [72]. The connections between
the horizon, mesh size, and condition number were also investigated for isotropic elements [72].
On the other hand, the interactions among the mesh geometry, mesh refinement, and the condi-
tion number of the global stiffness matrix for classical PDEs have been well studied [9, 41, 73].
The connection between the anisotropy of mesh elements and the condition number for elliptic
PDEs was investigated [9]. Various mesh quality metrics, interpolation error, and the condition
number for elliptic, parabolic, and hyperbolic PDEs were investigated [41]. The connections
between the mesh quality metric, preconditioner, and the linear solver for elliptic PDEs were
We use the Galerkin finite element method for discretizing a linear peridynamic system
and study the effect of the anisotropy of the mesh element (element shape), mesh refinement
(element size), and kernel functions on the condition number for a bond-based nonlocal peri-
dynamic model in 2D geometric domains. There exist various peridynamic models such as a
bond-based model [56, 69] and a state-based peridynamic model [58]. We consider a bond-based
peridynamic model which only considers central forces between particles [56, 59, 69]. We will
consider both integrable and nonintegrable kernel functions and numerically show the effect of
these kernel functions on the condition number for both piecewise linear and piecewise constant
107
basis functions. The study on the conditioning is important because it affects the accuracy of
the solution and the convergence rate of solving the discretized linear system. This study is the
first to explore the connections among anisotropy, mesh refinement, and the condition number
for 2D meshes with various kernel functions for a nonlocal peridynamic model. This work is
computationally challenging for two reasons. First, we need two different quadrature rules to
approximate the double integral terms in 2D geometric domains to avoid the singularity of the
denominator when we compute the condition number of the global stiffness matrix. Second, it
is desirable to compute approximately the area of intersection between the horizon (δ ) and the
We consider three different scenarios: (1) a piecewise constant basis function with an in-
tegrable kernel function, (2) a piecewise linear basis function with an integrable kernel function,
and (3) a piecewise linear basis function with a nonintegrable kernel function. For each scenario,
we investigate the effect of changing the anisotropy and mesh size (h) on the conditioning of the
nonlocal peridynamic model. The analytical results show that the condition number is bounded
by cδ −2 (c is a constant) [72] when an integrable (finite) kernel function is employed. For the
first two scenarios with an integrable kernel function, we will numerically show that the condi-
tion number is almost not affected by the choice of the basis function. We will also numerically
compute the constant c in the condition number bound (cδ −2 ) on uniform triangular and rectan-
gular meshes in 2D. For a piecewise linear basis function with a nonintegrable basis function,
we expect that the condition number is similar to that of general elliptic PDEs since this sce-
nario is similar to a local model. For general elliptic PDEs, it is well-known that the condition
number is proportional to h−2 when the mesh has the same anisotropy on uniform triangular and
rectangular meshes in 2D [41]. Mesh anisotropy also affects the condition number for general
108
elliptic PDEs. For instance, let θ be the smallest angle in the right triangle, then the condition
number is proportional to sin−1 (2θ ) [9]. Therefore, the condition number sharply increases as
θ approaches 0. For a piecewise linear basis function with a nonintegrable kernel function, we
expect that the condition number is affected by the anisotropy and the mesh size similar to the
general elliptic PDE case. We will investigate the effect of the anisotropy and the mesh size
on the condition number and study the difference between the nonlocal peridynamic model and
PDEs. The contents of this chapter will soon be submitted for review [110].
The peridynamic model replaces a PDE with an integral equation. The bond-based peri-
dynamic model assumes that the solid body consists of small particles and that these small par-
ticles interact with each other [59]. The interaction between particles is called a bond, and each
particle interacts with other particles within a distance, δ (horizon). We assume that δ is a sphere.
The equation of motion in the bond-based peridynamic model at point v = (x, y) is denoted by
Z
f u v0 − u (v) dv0 ,
Lu (v) = (4.1)
H
of particles which interact with a particle x, f is a pairwise peridynamic force, and dv0 is an
infinitesimal volume related with a particle v0 , respectively. From Newton’s Second Law,
where ρ is the mass density, and b(v) is an external force. Combining (4.1) and (4.2) results in
Z
f u v0 − u (v) dv0 + b (v) .
ρ ü (v) = (4.3)
H
Let ξ be the distance between v and v0 . Then, f is 0 if ξ is larger than δ due to the definition
of nonlocality. The horizon (δ ) and force ( f ) between v and v0 over the geometric domain Ω
are illustrated in Fig. 4.1. We only consider a microelastic material, which is commonly used in
Fig. 4.1 Connection between horizon (δ ), force ( f ) between v and v0 , inside the neighborhood,
H, over the domain Ω. The point v does not interact with any points beyond the distance δ .
the nonlocal peridynamic literature. For the microelastic material, the interaction between two
particles is considered to be an elastic spring [56, 69]. For the bond-based nonlocal peridynamic
model with a microelastic material, (4.3) can be reduced to the linearized peridynamic model
which is denoted as
(v0 − v) ⊗ (v0 − v)
Z
0
0
ρ ü (v) = c p u v − u (v) dv + b (v) , (4.4)
H |v0 − v|
110
where ⊗ is the Kronecker product. The reader is referred to [56, 69] for further information on
properties of microelastic models. Since we assume a steady-state model, ü(v)=0. Then, (4.4)
(v0 − v) ⊗ (v0 − v)
Z
(u v0 − u (v)) dv0 + b (v) = 0.
c 0 p (4.5)
H |v − v|
For a one-dimensional domain (e.g., v = x) with a Dirichlet boundary condition, (4.5) is ex-
pressed as
u (v) − u (v0 )
Z v+δ
1
dv0 = b (v) , v ∈ Ω,
δ 4−p v−δ |v − v0 | p
where u (v) = g (v) on the boundary, ∂ Ω, The global stiffness matrix, A, is computed by using a
standard Galerkin finite element method and basis functions, φ (v). The basis functions will be
φ j (v) − φ j (v0 )
Z
1
Ai j = , φi (v) ,
δ 4−p H |v − v0 | p
where M is the number of mesh elements for a piecewise constant basis function and can be
1 φ j (v) − φ j (v0 ) 0
Z Z
Ai j = φi (v) dv dv. (4.6)
δ 4−p I H |v − v0 | p
111
Here I represents all the neighborhoods of a point v where φi (v) is nonzero. We define 1/ |v − v0 | p
to be the kernel function. The matrix A is a symmetric, positive-definite matrix and is sparse. Fig-
ure 4.2 shows an example of the sparsity pattern of the global stiffness matrix on a 2D geometric
domain. Compared with the PDE sparsity pattern, more nonzeros appear due to nonlocality. The
Another important difference between classical PDEs and nonlocal peridynamic models
is the definition of the boundary condition. Figure 4.3 shows the difference between the classical
PDE boundary condition and the nonlocal peridynamic boundary condition. For classical PDEs,
the boundary condition, BΩ, is defined for boundary points. However, the boundary conditions
are defined for a boundary area, BΩ, for nonlocal peridynamic models. Therefore, the additional
mesh generation on the boundary area, BΩ, is required with respect to the horizon, δ .
Fig. 4.2 One example of the sparsity pattern of a matrix A for the nonlocal peridynamic model.
Here nz is the number of nonzeros in A.
112
Fig. 4.3 Boundary conditions for (a) classical PDE models and (b) nonlocal peridynamic models.
For classical PDE problems, the boundary condition is only specified on the boundary points.
For nonlocal peridynamic models, all mesh elements belonging to BΩ belong to the boundary
condition.
113
Our goal is to compute the condition number of (4.6) for 2D geometric domains. It is
computationally inexpensive to compute (4.6) using quadrature rules similar to [69]. In partic-
ular, we need to choose two different quadrature rules for inner and outer integrals because the
denominator term, i.e., |v − v0 |, could be singular if we choose the same quadrature rule twice.
Z m
f (x, y) dx dy ≈ |T | ∑ f (xi , yi ) wi ,
T i=1
where T is a triangle in Ω, |T | is the area of T , and m is the number of points used in the
associated quadrature rule [74]. We employ with one point and three point quadrature rules to
approximate the outer and inner integrals, respectively. For the unit triangle whose vertices are
(0,0), (0,1), and (1,0), the one point quadrature rule (centroid), which has degree of precision 1,
is computed as
(x1 , y1 ) = (1/3, 1/3) ,
w1 = 1.
For the unit triangle whose vertices are (0,0), (0,1), and (1,0), the three point quadrature rule
(x1 , y1 ) = (2/3, 1/6) , (x2 , y2 ) = (1/6, 2/3) , (x3 , y3 ) = (1/6, 1/6) ,
w1 = 1/3, w2 = 1/3, w3 = 1/3.
114
Since the horizon is a circle and the mesh element is a triangle, we use another ap-
proximation to compute the intersection area of T with the horizon (circle). We consider three
different cases when the intersection of T with δ is computed. The first case arises when the
entire triangle is inside the horizon. For this case, the |T | is exactly the area of a triangle and is
shown in case 1 in Fig. 4.4. The numerical difficulty arises when only part of a triangle is located
inside the horizon as shown in cases 2 and 3 in Fig. 4.4. The exact intersection area is expensive
to compute for cases 2 and 3. Therefore, we use approximations to compute the intersection
areas of T with δ for both cases 2 and 3. For case 2, we approximately compute the intersection
area inside the horizon by adding the areas of the two triangles inside the horizon as shown in
Fig. 4.5. For case 3, we compute the intersection area of a triangle inside the horizon rather than
Fig. 4.4 Three different cases for computing |T | when the intersection of T with δ is computed.
Here, we only compute the intersection areas inside the horizon.
115
Fig. 4.5 The approximation of the area of intersection of the triangular element and the horizon
for case 2 in Fig. 4.4.
The continuous piecewise constant (PC) basis function cannot be used for classical PDEs
because the derivative of zero is not defined. However, the nonlocal peridynamic model is able
to use the continuous PC basis function since it uses an integral equation instead of a derivative.
The PC basis function is 1 if x is on the jth element, otherwise, it is 0. The continuous PC basis
function is defined as
1, if (x, y) ∈ Tj
φ j (x, y) =
0,
otherwise.
The continuous piecewise linear (PL) basis function with respect to the jth vertex is defined as
− [(yl − yk ) (x − xk ) + (xk − xl ) (y − yk )] / [(x j − xk ) (y j − yl ) − (x j − xl ) (y j − yk )] ,
φ j (x, y) = if (x, y) ∈ neighborhood of jth vertex
0, otherwise,
where k and l are two other vertices of an element to which (x, y) belongs.
116
4.4 Connections among the horizon, mesh refinement, anisotropy of the mesh
4.4.1 Condition number of the global stiffness matrix for a nonlocal peridynamic model
The connection between the horizon, mesh size, and condition number of the global
stiffness matrix is studied analytically for an integrable kernel function [72]. The condition
where c is a constant and h is the mesh size. Since many nonlocal problems assume that δ is
greater than h, this equation indicates that the condition number of the global stiffness matrix is
bounded by the horizon, δ . In other words, the condition number is proportional to δ −2 when
δ is greater than h. This equation also indicates that decreasing the horizon increases the upper
We consider three different choices of basis functions and kernel functions in (4.6). The
power of a kernel function, p, is chosen based on the following equation. In d-dimensional space
We consider the case where d=2. If the power of a kernel function p ≥ 2, (4.6) is nonintegrable
Piecewise constant basis function with an integrable kernel function. For the integrable
kernel function, we set p to be 1 in (4.7). For this kernel function, we expect that the condition
number is bounded by
where c is a constant and the condition number converges to a stable level for the mesh refinement
and anisotropy. Since we assume that the δ is bigger than the mesh size, h, the condition number
is bounded by
cond(A) ≤ c δ −2 .
Piecewise linear basis function with an integrable kernel function. For the piecewise linear
basis function with an integrable kernel function, we set p to be 1 in (4.7). Similar to the
piecewise constant basis function with an integrable kernel function case, we expect that the
cond(A) ≤ c δ −2 ,
Piecewise linear basis function with a nonintegrable kernel function. For this nonintegrable
kernel function, we assume that k is greater than two. For this type of peridynamic model, we
expect that the condition number behavior is similar to the classical PDE behavior, since the
nonlocal peridynamimcs model is close to a local model such as a PDE model for this case. This
118
means that the anisotropy and the mesh refinement increase the condition number similar to the
PDE model.
4.4.2 Condition number of the global stiffness matrix for general second-order elliptic
PDEs
The effect of the mesh size and the mesh anisotropy on the conditioning of the global
stiffness matrix for general second-order elliptic PDEs has been studied [9, 41]. Suppose we
solve the general second-order elliptic PDEs on a uniform structured triangular mesh with right
triangles on 2D rectangular domains. Let θ be the smallest angle in the right triangle, then the
condition number is proportional to sin−1 (2θ ) [9]. This equation indicates that the condition
that the mesh elements have the same anisotropy, the condition number is proportional to h−2 .
We investigate the effect of anisotropy of the mesh element, mesh refinement, and vari-
ous kernel functions on the condition number of the global stiffness matrix of A in (4.6) on 2D
structured rectangular meshes, i.e., Ω = (0, 1) × (0, 1). We assume that δ is larger than the mesh
size, h. We use the 2-norm condition number to compute the condition number of the global
stiffness matrix A.
Mesh generation and boundary condition. We consider structured triangular meshes on the
rectangular geometric domain as shown in Fig. 4.6. The element sizes (edge lengths) in the
initial mesh, after one level of mesh refinement, and after two levels of mesh refinement, are (a)
119
0.1, (b) 0.05, and (c) 0.025, respectively. Additional triangular elements were generated in the
area surrounding Ω inside the area where the boundary conditions are defined. One example of
the anisotropic mesh, whose aspect ratio is 4, is shown in Fig 4.7. We fix the total number of
elements for both isotropic and anisotropic meshes for given h and δ . The aspect ratio of a right
where w and W are shortest and the second shortest edge lengths, respectively. Figure 4.8 shows
Computation time. We use an OpenMP library to speed up the computation time. Note that we
parallelize our code such that the parallelization of our code using an OpenMP library does not
affect the condition number. The machine employed for this study is equipped with a 48 core
AMD opteron 2.3 GHz processor. When δ =0.2 and piecewise constant basis functions are used,
the total time to compute cond(A) is 1 second (h=0.1), 11 seconds (h=0.05), and 341 seconds
(h=0.025), respectively. The running time for piecewise linear basis functions is much greater
than the the running time for piecewise constant basis functions since we must consider all the
neighborhood elements of each vertex. When δ =0.2 and piecewise linear basis functions are
used, the total time to compute cond(A) is 1 second (h=0.1), 11 seconds (h=0.05), and 100,825
Fig. 4.6 (a) Initial isotropic mesh with h=0.1 and δ =0.2; (b) refined mesh with h=0.05 and δ =0.2;
(c) Two level-refined mesh with h=0.025 and δ =0.2
121
Fig. 4.8 The definition of the aspect ratio of a triangle. The aspect ratio is defined as W /w [75].
122
For these experiments, we employ piecewise constant basis functions and fix the power
of the kernel function, p in (4.6), to be 1 for an integrable kernel function. Table 4.1 shows the
condition number of the global stiffness matrix, A, for various mesh sizes and horizons (δ ) for
both isotropic and anisotropic meshes. Here, the ratio is defined as the ratio of the condition
number of the refined mesh (after one level of mesh refinement) divided by the condition num-
ber of the coarser mesh for the same δ . We also observe that the overall ratio decreases as δ
increases for a fixed level of mesh refinement. This is because the numerical computation and
the condition number for larger δ values are less sensitive than is the computation with smaller
δ values. When δ is small (i.e., δ =0.2), a large portion of triangular elements within the horizon
intersects the horizon (e.g., cases 2 and 3 in Fig. 4.4), and the increased ratio of intersections
between the horizon and the triangular elements decreases the numerical stability.
Table 4.1 Cond(A) of the global stiffness matrix for piecewise constant basis functions with an
integrable basis kernel function (p=1). The number of elements is the number of elements in
Ω. Here, the ratio is the ratio of the condition number of the current level of mesh refinement
compared with that of the previous level of mesh refinement for a fixed δ . Therefore, these ratios
are not defined for the 200 elements and are denoted by ‘*’ for these cases.
123
Fix the mesh size (h) and vary the anisotropy of the elements. We compute the condition
number while fixing the mesh size (h) and vary the anisotropy (aspect ratio) of the elements.
Figures 4.9 and 4.10 show the connection between δ , anisotropy, and the condition number. We
observe that the condition number of the global stiffness matrix A is proportional to δ −2 when
δ is bigger than h, and the increasing anisotropy of the elements does not affect the condition
Fix the anisotropy of the elements and vary the mesh size (h). We fix the anisotropy of the
elements and compute the condition number for different mesh sizes ( h). Figures 4.11 and 4.12
show the connection between the horizon (δ ), mesh refinement, and the condition number of the
global stiffness matrix. Similar to the previous numerical results, the condition number of the
global stiffness matrix A is proportional to δ −2 when δ is larger than h. We observe that the
condition number of the global stiffness matrix A converges to a stable level as the number of
mesh refinement levels increases (as h decreases by half). We observe that the condition number
For these experiments, we use piecewise linear basis functions and fix the power of the
kernel function, p, in (4.6) to be 1 for an integrable kernel function. Based on the analytical
results in Sec. 4.4, we expect that the overall trend is similar to that seen for piecewise constant
basis functions with an integrable kernel function. Table 4.2 shows the condition number of the
global stiffness matrix for various mesh sizes and horizons (δ ) for both isotropic and anisotropic
124
Fig. 4.9 Condition number of A for fixed h and varying anisotropy of the elements and δ .
Fig. 4.10 Condition number of A as a function of δ −2 for fixed h and varying anisotropy of the
elements and δ .
125
(a) Isotropic mesh (aspect ratio=1) (b) Anisotropic mesh (aspect ratio=4)
Fig. 4.11 Condition number of A for fixed anisotropy and varying h and δ .
(a) Isotropic mesh (aspect ratio=1) (b) Anisotropic mesh (aspect ratio=4)
Fig. 4.12 Condition number of A as a function of δ −2 for fixed anisotropy and varying h and δ .
126
meshes. Here, the ratio is defined as the ratio of the condition number of the refined mesh (after
one level of mesh refinement) divided by the condition number of the coarser mesh for the same
δ . We observe that the overall ratio decreases as δ increases for a fixed level of mesh refinement
in most cases. Similar to what we observed for piecewise constant basis functions, a large por-
tion of triangles within the horizon intersects the horizon (e.g., cases 2 and 3 in Fig. 4.4) and the
Table 4.2 Cond(A) of a global stiffness matrix for piecewise linear basis functions with an in-
tegrable basis kernel function (p=1). The total number of elements changes with respect to the
horizon, δ . The number of elements indicates a number of elements in Ω. Here, the ratio is
defined as the ratio of the condition number of the current level of mesh refinement compared
with that of the previous level of mesh refinement for a fixed δ . Therefore, these ratios are not
defined for the 200 elements and are denoted as ‘*’ for these cases.
Fix the mesh size (h) and vary the anisotropy of the elements. We compute the condition
number of the global stiffness matrix while fixing the mesh size (h) and varying the anisotropy
of the elements. Figures 4.13 and 4.14 show the connection between the horizon (δ ), anisotropy,
and the condition number of the global stiffness matrix. Similar to the piecewise constant basis
function with the integrable function, the condition number of the global stiffness matrix A is
127
proportional to δ −2 when δ is bigger than h. These results are consistent with the analytical
Fix the anisotropy of the elements and vary the mesh size (h). We fix the anisotropy and com-
pute the condition number for different values of h. Figures 4.15 and 4.16 show the connection
between the horizon (δ ), mesh refinement, and the condition number. We also observe that the
also observe that the condition number of A converges to a stable level as the number of mesh
4.5.3 Piecewise linear basis function with the nonintegrable kernel function
For these experiments, we fix the power of the kernel function, i.e., p in (4.6) to be 2.5
for a nonintegrable kernel function. Table 4.3 shows the condition number of the global stiff-
ness matrix for various mesh sizes and horizons (δ ) for both isotropic and anisotropic meshes.
Here, the ratio is defined as the ratio of the condition number of the refined mesh (after one level
of mesh refinement) divided by the condition number of the coarser mesh for the same δ . We
observe that the condition number increases as δ decreases. Previously, we observed that the
overall ratio is close to 1 when δ is 0.4. However, for these cases, the ratio is not close to 1, even
Fix the mesh size (h) and vary the anisotropy of the elements. We compute the condition
number of the global stiffness matrix while fixing the mesh size, h, and varying the anisotropy
128
(a) h=0.05
(b) h=0.025
Fig. 4.13 Condition number of A for fixed h and varying anisotropy of the elements and δ .
129
(a) h=0.05
(b) h=0.025
Fig. 4.14 Condition number of A for fixed h and varying anisotropy of the elements and δ .
130
Fig. 4.15 Condition number of A for fixed anisotropy and varying h and δ .
131
Fig. 4.16 Condition number of A for fixed anisotropy and varying h and δ .
132
Table 4.3 Cond(A) of a stiffness matrix for piecewise linear basis functions with a nonintegrable
basis kernel function (p=1). The number of elements indicates the number of elements in Ω.
Here, the ratio is defined as the ratio of the condition number of the current level of mesh refine-
ment compared with that of the previous level of mesh refinement for a fixed δ . Therefore, these
ratios are not defined for the 200 element case and are denoted as ‘*’ for such cases.
of the elements. Figure 4.17 shows the connection between the horizon (δ ), anisotropy, and the
condition number of the global stiffness matrix. Figure 4.17 shows that the condition number
of the global stiffness matrix A increases when the element shape is not isotropic. However,
further increase in anisotropy of the elements does not increase the condition number, which
means the effect of anisotropy is limited. Recall that the condition number sharply increases and
Fix the anisotropy of the elements and vary the mesh size (h). We fix the anisotropy of the
elements and compute the condition number for various mesh sizes, h. Figure 4.15 shows the
connection between the horizon (δ ), mesh refinement, and the condition number. Figure 4.11
shows that the condition number of the global stiffness matrix A is proportional to h−1 . Note
that for general elliptic PDEs, the condition number of the global stiffness matrix is proportional
to h−2 if the mesh has the same aniostropy. Therefore, the mesh refinement for a nonlocal peri-
dynamic model increases the condition number less compared with the one for general elliptic
133
(a) h=0.05
(b) h=0.025
Fig. 4.17 Condition number of A when h=0.05 for fixed h and varying anisotropy and δ .
134
Fig. 4.18 Condition number of A for fixed anisotropy (aspect ratio=1 and 4) and varying h and
δ.
135
Fig. 4.19 Condition number of A for fixed h and anisotropy and various power of a kernel func-
tion, p.
136
PDEs.
Fix the anisotropy of the elements and mesh size (h), and vary p. We fix the anisotropy of
the elements and h but vary p and see the effect of various values of p on the condition number.
We observe that as p increases, the condition number rapidly increases. Simple regression re-
sults show that the condition number exponentially increases as p ≥ 2, which means the kernel
4.6 Conclusions
We have investigated the effects of anisotropy, mesh refinement, and various kernel func-
tions on the condition number for a 2D nonlocal bond-based peridynamic model. This is the first
study to examine the connections between the anisotropy and the conditioning on 2D geometric
domains. We employ the Galerkin finite element method to discretize the nonlocal peridynamic
model and investigate the effects that various choices of basis functions have on the condition
number of the global stiffness matrix. We observe consistent numerical results with the analytical
results when we employ an integrable kernel function. As far as the integrable kernel function is
concerned, the condition number is bounded by cδ −2 (where c is a constant) and is not affected
by the choice of the basis function, when δ is bigger than the mesh size (h). Our numerical re-
sults show that the constant c is close to 1 on the 2D uniform triangular and rectangular meshes
on geometric domains for an integrable kernel function. For these scenarios, mesh anisotropy
and the mesh refinement affect the condition number very little. For piecewise linear basis func-
tions with a nonintegrable kernel function, we observe that the condition number is proportional
137
to h−1 when we fix the mesh anisotropy. We observe that anisotropic mesh elements increase the
condition number compared to the isotropic mesh elements when we fix h. However, the effect
of mesh anisotropy is limited since further increase in the mesh anisotropy does not increase
the condition number. We also observe that the condition number exponentially increases as the
power of a kernel function increases. Our results can be used to generate meshes with a small
condition number for a bond-based nonlocal peridynamic model. For future research, we will
consider other peridynamic models on unstructured geometric domains. The sparsity pattern of
a nonlocal peridynamic model is different from PDEs since the sparsity pattern is affected by the
horizon radius. We also plan to develop a preconditioner to decrease the condition number for a
In this dissertation, we have thus far focused on mesh quality improvement techniques
for PDEs and nonlocal peridynamic models. In the next part of dissertation, we will focus
on optimization-based meshing techniques for mesh deformation problems and shape matching
problems. In the next chapter, we will propose a hybrid deformation algorithm using the di-
rection of boundary motion and multiobjective mesh optimization. We find the optimal interior
vertex positions on the deformed domain when the boundary deformations are known.
138
4.7 Introduction
There are numerous applications where discretized geometric domains vary with respect
to time such as the solution of Arbitrary-Lagrangian-Eulerian (ALE) flow simulations [76], de-
formation of the human face in computer graphics [78], deformation of sequences of medical
images [81, 82], and deformations in biomedical applications [79, 80]. When these domain de-
formations occur, the meshes approximating the domain should be also updated appropriately
with respect to time such that the mesh remains valid. This mesh deformation (update) method
is called mesh warping [88, 95] (mesh morphing [77] or moving meshes [92]) with respect to
the applications. There exist many mesh deformation methods; however, we focus on mesh de-
formation algorithms which recompute interior mesh vertex positions after the mesh boundary
has been deformed. Our algorithm requirements are twofold: to produce meshes elements of
good element quality including no inverted elements, and second to maintain ‘similarity’ dur-
ing mesh deformation. An untangled deformed mesh is important since tangled meshes with
inverted elements result in physically invalid solutions when standard finite element shape func-
tions are used. Similarity means identical mesh connectivities and topologies in the deformed
mesh as were present in the initial mesh along with the preservation of general mesh character-
istics (e.g., shape, size, anisotropy) of the original mesh in the deformed mesh. This is important
because many applications need a smooth variation of meshes between each time step or itera-
tion [43], and the initial mesh may have been constructed with anisotropies designed with the
particular simulation in mind. For deforming domains, both mesh deformation and remeshing
can be used to generate meshes on deformed domains. However, mesh deformation is preferred
139
to remeshing, because remeshing results in the accumulation of large interpolation error be-
tween successive time steps and less accurate PDE simulation results [52]. Also, remeshing is
Researchers have proposed various mesh deformation algorithms based on Laplace equa-
tion methods, e.g., finite element-based mesh warping (FEMWARP) [84, 88], weighted Lapla-
cian smoothing [97], biharmonic partial differential equations [76], elasticity based approaches [92,
93, 95], by utilizing an inverse distance weighting function [83], and by combining vertex move-
ment with other techniques which alter mesh topology [85–87]. FEMWARP creates elements
with good quality for small or moderate boundary deformations and is exact for affine transfor-
mations, is easy to implement and scales well with mesh size. However, FEMWARP ignores
information about the nature of the deformation and often generates many inverted elements
for anisotropic deformations. The authors in [88] proposed three modifications of FEMWARP,
based untangler to address this problem. Small-step FEMWARP can be used to generate un-
tangled meshes on the deformed domain by introducing a pseudo time-stepping of the bound-
ary deformation with steps small enough to prevent the creation of inverted elements in each
and is able to prevent inverted elements, it does not preserve the original mesh element shapes
on the deformed domain. Also, it may not be practical to use this boundary pseudo time-stepping
for some applications since it requires a continuous function from the old to new boundary con-
ditions, which is not always available [88]. The mesh refinement strategy is not always practical
since the mesh generation source code is not always available. In addition, more refinement
alter the mesh topology. The combined approach with FEMWARP and mesh untangling is also
140
successful in eliminating inverted elements on the deformed mesh. However, it is also unable to
The biharmonic operator method to solve mesh warping problems [76] controls both the
normal mesh spacing and the boundary discretization. However, the method is computationally
expensive and also does not maintain ’similarity’ on the deformed domain. Finally, it does not in-
clude a mechanism for eliminating inverted elements if the deformed mesh includes any inverted
elements. The elasticity equation-based methods described in [92, 93] provide a mechanism for
decreasing the creation of inverted elements by choosing stiffness coefficients in a problem de-
pendent fashion. Another related elasticity based approach called Untangling Before Newton
(UBN) is proposed for handling large boundary deformations by using the iterative stiffness
method [95]. Recently, the log-barrier approach for the worst element mesh quality improve-
ment and untangling is also proposed and shows better untangling performance compared with
UBN [96]. UBN is also not able to preserve similar element shapes on the deformed domain.
There is another category of optimization-based mesh warping algorithms which use the target
matrix paradigm (TMP) [43]. These methods are capable of maintaining similarity between
initial and deformed meshes for various aspects of the mesh (e.g., shape, size, shape and size).
However, these methods often have poor convergence characteristics if the mesh undergoes large
boundary deformations.
All of the methods mentioned thus far suffer from at least one of the following prob-
lems: a tendency to produce tangled elements for large boundary deformations, or an inability
to preserve features of the initial mesh in the deformed mesh. In this chapter, we examine tech-
niques for robustly warping meshes subject to anisotropic boundary deformations which address
both of these problems. In Section 4.8, we review FEMWARP which our method generalizes.
141
Section 4.9 introduces our two-step hybrid algorithm. The first step is to estimate the internal
vertex positions of the deformed mesh using knowledge of the nature of the boundary defor-
mation. We choose the relative weighting of neighbor vertices in our anisotropic FEMWARP
method based on the direction of the boundary deformation. We prove that for some simple
mesh configurations and certain types of anisotropic deformations, the method we propose will
maintain desired properties of the initial mesh independent of the magnitude of the boundary
deformation. In more complicated cases the mesh produced in the first step often possesses
some inverted elements, so, as a second step, we propose a multiobjective mesh optimization of
element shape and element untangling. This step is designed to produce meshes with no inverted
elements while maintaining element similarity with the initial mesh. In Section 4.10, we demon-
strate that for several complicated boundary deformations the first step of our method produces
meshes with fewer inverted elements than other existing methods. We also demonstrate that our
multiobjective optimization is capable of producing meshes with no inverted elements and pre-
serving characteristics of the initial mesh. The contents of this chapter will soon be submitted
4.8 Background
In this section, we briefly describe the FEMWARP algorithm and motivate our anisotropic
FEMWARP which uses anisotropic PDE coefficients. The reader is referred to [84, 88] for fur-
4.8.1 FEMWARP
2D geometric domains. The equations are similar for 3D. FEMWARP solves Laplace’s equation
−∆u = 0 on Ω,
where u = u0 on the boundary, ∂ Ω. FEMWARP updates interior vertex positions given a bound-
ary deformation discretized using piecewise linear finite elements [88]. Let B and I be the sets
of boundary and interior vertices, respectively. Let NI and NB be the numbers of interior and
boundary vertices, respectively. FEMWARP is composed of three steps and represents each in-
terior vertex as an affine combination of its neighbors. The first step of FEMWARP forms the
(NI + NB ) × (NI + NB ) global stiffness matrix A by assembling the element stiffness matrices on
the undeformed domain. By ordering the interior unknowns first and the boundary unknowns
second, the matrix A can be partitioned into boundary and interior sub-matrices
AI AB
A=
,
(4.8)
(AB )T C
where AI is the (NI × NI ) submatrix of A which denotes the connections of interior vertices
to interior vertices; AB is the (NI × NB ) submatrix denoting the connections between interior
and boundary vertices, and C is the (NB × NB ) submatrix representing the connections among
boundary vertices. Let (xI , yI ) and (xB , yB ) be the coordinates of the interior and boundary
143
vertices on the initial undeformed domain, respectively. Then, since linear functions are in the
AI xI + AB xB = 0
(4.9)
AI yI + AB yB = 0.
Let (x̂I , ŷI ) and (x̂B , ŷB ) be the coordinates of interior and boundary vertices on the deformed
domain, respectively. The second step is to apply a known boundary deformation to (4.9), i.e.,
The final step is to compute the coordinates of interior vertices on the deformed domain (x̂I , ŷI )
AI x̂I + AB x̂B = 0
(4.10)
AI ŷI + AB ŷB = 0,
FEMWARP has several advantages over other mesh warping algorithms in that it is
simple to implement and is exact for affine boundary deformations [88]. It was reported that
FEMWARP is successful in yielding noninverted elements during the mesh updating process for
some applications such as a beating canine heart from an atrial pacing experiment [97] and also
for shape optimization problems [94]. However, FEMWARP often generates a large number
144
of inverted elements if deformations are not affine and anisotropic deformation occurs in one
direction. This is because FEMWARP is not designed to consider the direction of boundary de-
formation and equally considers all vertex neighbors even if the deformation occurs in only one
direction.
We present a motivating example which was used in [76]. The initial mesh and the
deformed mesh using FEMWARP are shown in Figure 4.20. The initial mesh is uniform and
structured on a rectangular domain. Here, the the deformation only occurs in the y direction.
Figure 4.20 shows the deformed mesh produced by FEMWARP includes 20 inverted elements
and has several valid elements with poor quality. This mesh also fails to preserve the element
spacing of the initial mesh on the deformed domain. We point out that these inverted elements
with poor element shapes occur because FEMWARP equally considers all neighboring vertices
when it computes the stiffness matrix and does not consider the direction of the boundary defor-
mation. For this kind of one-directional deformation, it is obvious that we should only connect
neighbors aligned in the y direction. In the next subsection, we will show that similarity (here,
uniform spacing on the deformed domain) can be achieved using our anisotropic FEMWARP
Our hybrid mesh deformation algorithm is composed of two steps. First, we apply
anisotropic FEMWARP for determining an initial guess for the interior vertex positions on the
deformed domain. The goal of this first step is to preserve the element spacing and to reduce the
number of inverted elements on the deformed domain. For this step, we consider the direction
145
Fig. 4.20 (a) Initial mesh on the rectangular domain. (b) Deformed mesh using FEMWARP. This
mesh has 20 inverted elements.
of deformation when choosing the PDE coefficients. Similar to other mesh deformation algo-
rithms, anisotropic FEMWARP may produce inverted elements. So as a second step, we apply
the multiobjective mesh optimization using a Target Matrix Paradigm (TMP) shape term and
an untangling term to find the optimal interior vertex positions on the deformed domain. Note
that meshes with noninverted elements are important since meshes with inverted elements yield
which is better suited to cases where the initial mesh is anisotropic and/or the deformation is
aligned with one coordinate axis. However, this idea is easily extended to other kind of defor-
mations (such as a diagonal deformation). For anisotropic FEMWARP in 2D, the PDE used to
146
generate the linear system defining the connection between the interior and boundary vertices is
∂ 2u ∂ 2u
−α − β = 0 on Ω, (4.11)
∂ x2 ∂ y2
where u = u0 on the boundary of Ω, i.e., ∂ Ω. Here, we assume that α > 0 and β > 0. From
here, we consider 2D meshes to simplify the notation; however, the extension to 3D meshes is
quite natural, and we will explain how to extend our anisotropic FEMWARP to 3D.
different PDE coefficients are employed. It follows the same three steps as FEMWARP which
are explained in Sec. 4.8.1 after the stiffness matrix is formulated. Similar to FEMWARP,
anisotropic FEMWARP also represents each interior vertex as an affine combination of neigh-
bors, but anisotropic FEMWARP adaptively changes its weights for each interior vertex with
respect to the PDE coefficients, α and β . We will show that this adaptive strategy significantly
helps to reduce the number of inverted elements compared with FEMWARP, which always fixes
α=1, β =1. We will also show that for extreme cases, if either α or β is zero, we only consider
How to preserve similar element spacing on the deformed domain? We begin with a dis-
cussion of how to maintain similar element spacing between the initial and deformed domains
by appropriately choosing the PDE coefficients in (4.11). We show that for some simple mesh
configurations and certain types of anisotropic deformations, setting one of the PDE coefficients
to one and other zero in (4.11) is optimal to maintain element spacing. In all our results in this
147
section, we assume that the initial domain is rectangular and axis aligned and the mesh is a tri-
angulated structured grid with the grid lines also axis aligned. We begin with the simplest case,
i.e., that of uniform vertex spacing equal to h in x and k in y. Referring to the generic triangle
Hi×h, Hj+1L×kL
k
Tij
A h B
Hi×h, j×kL HHi+1L×h, j×kL
Fig. 4.21 Generic element in structured triangular mesh with uniform elements.
in this assumed initial mesh shown in Figure 4.21, it is a straightforward exercise to derive the
αk β h
2h + 2k − αk − β2kh
2h
. (4.12)
− αk2h
αk
2h 0
− β2kh βh
0 2k
Proposition 1. If α = 0 and β = 1 in (4.11), then the positions of the interior vertices on the
deformed domain are determined only by the boundary vertices which had the same x coordinate
Proof. Since α = 0 we can see that the element stiffness matrix in (4.12) has only non-zeros in
the positions connecting vertices A and C, those vertices with the same x coordinate.
Since each element stiffness matrix is structured so that only vertices with the same x
coordinate in the original mesh are connected, this characteristic also holds in the global stiffness
matrices AI and AB .
With the same assumptions for the domain and mesh characteristics and with α = 0 and
Proposition 2. The mesh spacing in the y direction is uniform in the deformed mesh.
Proof. As explained earlier, anisotropic FEMWARP basically follows the same three steps as
does FEMWARP in Sec. 4.8.1. By (4.10), AI ŷI + AB ŷB = 0. Proposition 1 allows us to rearrange
the linear system AI ŷI = −AB ŷB into a block tridiagonal system with each independent block
consisting of unknown interior vertex coordinates with the same x coordinate and each block
ordered by increasing mesh y coordinate. After appropriate scaling, each block then has the
form
2 −1 0 · · · 0 ŷ1 ŷ0
−1 2 −1 · · · 0 ŷ 0
2
. .. .. .. = ..
..
. . . . . (4.13)
0 · · · −1 2 −1 ŷN−3 0
0 ··· 0 −1 2 ŷN−2 ŷN−1
149
ŷN−1 − ŷ0
ŷ j = ŷ0 + j .
N −1
If we relax our grid assumptions slightly and allow variable grid spacing in y as a form
of anisotropy, then our proposed method still satisfies our desired properties. We assume that the
Proposition 3. The FEMWARP method with α = 0 and β = 1 will maintain the relative grid
Proof. First introduce a parameter γ = (ŷN−1 − ŷ0 )/(yN−1 −y0 ) to carry the relationship between
the deformed mesh y-axis extents and the original mesh extents. The element stiffness matrix and
global stiffness matrix construction from Proposition 2 carry though with minor modifications
for the varying grid spacing in y and the block tridiagonal system to solve for each block of
1 1
+
k0 k1 − k11 0 ··· 0 ŷ1
1
k0 ŷ0
−1 1
+ 1
− k12 ··· 0 ŷ 0
k1 k1 k2 2
.. .. .. .
. .. =
.
..
.
. . .
1 1 1 1
0 ··· − kN−4 kN−4 + kN−3 − kN−3 ŷN−3 0
1 1 1 1
0 ··· 0 − kN−3 kN−3 + kN−2 ŷ N−2 ŷ
kN−2 N−1
(4.14)
150
Therefore the deformed mesh has the same relative grid spacing in y as the undeformed mesh.
and β = 0. The observation we make is that the choice of setting one coefficient to 0 and the
other to 1 in (4.11) in anisotropic FEMWARP will preserve the original mesh spacing in that
direction. Furthermore, for these axis aligned meshes, the decoupling that occurs in the global
be easily treated while still preserving the original relative mesh spacing along each grid line in
that dimension.
Anisotropic FEMWARP is exact for affine boundary transformations. One of the main
advantages of FEMWARP is that this method is exact for affine boundary transformations [88].
We will show that anisotropic FEMWARP is also exact for affine boundary transformations.
Proposition 4. Let AB and AI be the submatrices generated using anisotropic FEMWARP (see
Sec. 4.8.1). We define [xB , yB ] and [xI , yI ] to be the known boundary and interior coordinates on
the initial domain, respectively. We also define [x̂B , ŷB ] and [x̂I , ŷI ] to be the known boundary
and unknown interior coordinates on the deformed domain, respectively. Suppose the affine
transformation occurs on the each boundary coordinate, k ∈ B, then there exist a nonsingular
x̂k xk
= M + w.
ŷk yk
Then, each interior vertex, l ∈ I, [x̂l , ŷl ], can be represented as an affine transformation such that
x̂l xl
= M + w.
ŷl yl
Proof. Since anisotropic FEMWARP follows the same three steps as does FEMWARP explained
in Sec. 4.8.1, the interior vertex positions of anisotropic FEMWARP can be computed as
AI x̂I + AB x̂B = 0
AI ŷI + AB ŷB = 0.
x̂B = LxB + w
(4.16)
ŷB = LyB + w,
where eB is a vector of size NB with all 1’s. In order to show that interior vertices are exact, we
where eI is a vector of size NI with all 1’s. By combining (4.17) and (4.18), we need to show
that
Similar to FEMWARP, the weights of each interior vertex sums to one, i.e.,
AI eI + AB eB = 0.
AI xI = −AB xB
AI yI = −AB yB .
that setting α=0 and β =1 in anisotropic FEMWARP is a good strategy for preserving element
spacing for an anisotropic boundary deformation aligned with the y-axis. Similarly, anisotropic
FEMWARP with α=1 and β =0 is desired when a boundary deformation occurs which is aligned
153
with the x-axis. We now apply this idea to our motivating example shown in Figure 4.20 and
compare with FEMWARP, which fixes α=1 and β =1. Figure 4.22 shows a comparison of the
meshes produced by FEMWARP and anisotropic FEMWARP, respectively. For this example,
we chose α = 0 and β = 1 for anisotropic FEMWARP, since the deformation occurs only in
the y direction. We observe that the deformed mesh using anisotropic FEMWARP does not
have inverted elements, whereas the deformed mesh generated by FEMWARP has 20 inverted
elements. Also, we observe that anisotropic FEMWARP is able to preserve similar element
Fig. 4.22 (a) Deformed mesh using FEMWARP. This mesh has 20 inverted elements. (b) De-
formed mesh using anisotropic FEMWARP with coefficients α=0 and β =1 in (4.11). This de-
formed mesh using anisotropic FEMWARP does not have any inverted elements.
We now consider more general cases where the boundary deformation occurs not aligned
with the x-axis or y-axis (e.g., a diagonal deformation). We adaptively change α and β values
with respect to the direction of boundary deformation. We have demonstrated that α controls
the strength of the x-axis coupling between adjacent vertices, and β controls the strength of
154
the y-axis coupling. Therefore, increasing α corresponds to putting more weights on neighbors
aligned with the x-axis than on neighbors aligned with the y-axis when we formulate the global
stiffness matrix A. This strategy is appropriate when more deformation occurs in the x-axis than
in the y-axis. Similarly, β should be bigger than α when more deformation occurs along the
y-axis than the x-axis. We compute the cumulative boundary vertex displacements in the x and
y directions and observe in which direction more deformation occurs by computing the relative
ratio. If the cumulative boundary vertex displacement in the x direction is larger than the one in
the y direction, we say that more deformation occurs along the x-axis and use the relative ratio of
these deformations to compute α and β . The relative ratio can also be understood as the angle
α = ∑Nk=1
B
(xˆk − xk ) , k ∈ B
(4.19)
β = ∑NB (yˆk − yk ) , k ∈ B,
k=1
where N is a total number of boundary vertices on the mesh. Also, recall that [xk , yk ] are the
known boundary coordinates on the initial domain, and [xˆk , yˆk ] are the known boundary coor-
dinates on the deformed domain. Figure 4.23 summarizes our ideas and shows the connection
between PDE coefficients (α and β ) and the direction of boundary motion. Here, a and b in
Extension to 3D meshes. The extension to 3D meshes is natural since we just add one more
coefficient term, i.e.,γ, for the z-axis. The PDE used to generate the linear system for anisotropic
FEMWARP in 3D is
155
Fig. 4.23 Connection between the PDE coefficients (α and β ) and the direction of motion. If the
deformation occurs (a) aligned with the x-axis or (b) aligned with the y-axis, we only consider
neighbors aligned with the same axis. Here, the cross-out indicates that we do not consider those
neighbors. If deformation occurs that is not aligned with either the x or y axes (c), we use the
angle of direction of the deformation to choose the appropriate PDE coefficients. Here, a and b
are α and β in (4.19), respectively.
∂ 2u ∂ 2u ∂ 2u
−α − β − γ = 0 on Ω, (4.20)
∂ x2 ∂ y2 ∂ z2
where u = u0 on the boundary, ∂ Ω. Here, we assume that α > 0, β > 0, and γ > 0. Similar to
the 2D cases, we set only one of the PDE coefficients as 1 and the other two coefficients as 0 if
deformation occurs that is aligned with the x-,y-, or z-axis in order to preserve similar element
spacing. For example, if the deformation only occurs in z-axis, we set α = 0, β = 0, and γ = 1.
If the deformation that occurs is not aligned with any of the axes, we follow the same ideas
in (4.19), i.e.,
α = ∑Nk=1
B
(xˆk − xk ) , k ∈ B
β = ∑Nk=1
B
(yˆk − yk ) , k ∈ B (4.21)
γ = ∑NB (zˆk − zk ) , k ∈ B.
k=1
156
deformed domain and produces fewer inverted elements on the deformed domain than does
FEMWARP, large deformations still often cause element inversion on the deformed domain.
We propose a multiobjective mesh optimization of element shape and untangling to correct this
problem. The idea of using multiobjective mesh optimization for mesh quality improvement and
mesh untangling was proposed in our previous paper [73]. We employ this multiobjective mesh
optimization framework and apply this framework to mesh deformation problems. We focus on
preserving similar element shape by employing a TMP shape metric [54]. Let Ai be the Jacobian
matrix of the mapping from the reference element to the ith mesh element. The fundamental ob-
ject used to construct TMP quality metrics is a dimensionless Jacobian matrix denoted Ti defined
by
where (Adef )i and (Ainit )i are the Jacobians of the mappings from the reference element to the ac-
tual elements in the deformed and initial domains, respectively. The scale- and rotation-invariant
|Ti − (adj(Ti )t ))|2F , T ∈ R2 × R2
qi = (4.22)
|Ti |F
√ Ti − (adj(Ti )t )2 , T∈ R3 × R3 .
3 F
The TMP shape metric is zero when the quality of the deformed element is the same as in the
initial mesh, which means the element shape is identical. Note that other similarity (e.g., size
157
and shape and size) could be preserved by simply changing (4.22) into other TMP metrics such
as TMP size or TMP shape and size metric, which are defined in [16].
In order to eliminate inverted elements, we employ the untangling beta quality metric
which is defined
q j = V j − β − (V j − β ),
where V j is the area of the jth element, and β is a user-defined parameter greater than 0. The
untangling beta metric is zero when the mesh element is not inverted. The overall mesh quality
|E|
F1 = ∑ q2i ,
i=1
where |E| is the number of mesh elements. Similarly, the overall mesh quality computed by the
untangling metric is
|E|
F2 = ∑ q2j .
j=1
Our goal is to simultaneously optimize these two objective functions to simultaneously untangle
and improve the element shape on the deformed domain. We find the maximum cost function
between F1 and F2 and also minimize the worst cost function between F1 and F2 . Then, the
However, (4.23) is not smooth and has neither a Jacobian nor a Hessian. Therefore, we use the
exponential sum multiobjective method to approximate the solution to (4.23). The exponential
sum multiobjective function approximates the min-max problem by employing the exponential
" #
2
F = c ln ∑ eF /c
i
, (4.24)
i=1
where c is a controlling parameter and is typically chosen between 104 and 106 . It was proven
that the solution to (4.24) satisfies a sufficient condition of Pareto optimality [91]. One of the
main strengths of employing this exponential sum function is it does not require any articulation
of preferences between the two objective functions. Similar to [89], we employ the nonlinear
conjugate gradient method to find a local optimal point, since this is the default NLCG method
in Mesquite. Here, the local optimal point means the optimal vertex locations on the deformed
domain.
We present more realistic unstructured mesh examples in this section. For our second
step, i.e., multiobjective mesh optimization, we use Mesquite (version 2.99) [16]. We first con-
sider three examples where anisotropic deformation occurs aligned with the x or y axis. We also
apply our hybrid mesh deformation algorithms when deformation occurs that is not aligned with
either the x or y axis (e.g., diagonal deformation). Finally, we apply our hybrid algorithm to the
We make comparisons to several existing mesh warping methods and compare three
key features: (1) the number of inverted elements produced, (2) the time to untangle the de-
formed mesh using our multiobjective mesh optimization with both FEMWARP and anisotropic
FEMWARP, and (3) the similarity of the deformed mesh to the original mesh as measured by the
TMP shape metric for three different algorithms: FEMWARP, our hybrid algorithm (anisotropic
FEMWARP + multiobjective mesh optimization), and UBN [95]. In (1), we will show that the
anisotropic FEMWARP significantly helps to reduce the number of elements as a first step and
our second step, i.e., multiobjective mesh optimization, is able to eliminate inverted elements
from tangled meshes both with FEMWARP and anisotropic FEMWARP. In (2), we will show
that anisotropic FEMWARP reduces the number of inverted elements which significantly helps
the convergence time of our second step, i.e., multiobjective mesh optimization. Finally in (3),
we show that our hybrid algorithm is able to preserve the element shape by showing the simi-
larity measure by the TMP shape metric. Note that when we compare with UBN, we only use
the iterative stiffening aspect of UBN which comes after FEMWARP in the UBN code. The
machine employed for this study is equipped with an AMD Quad-core Opteron processor (2.3
4.10.1 Moving cylinder domain for anisotropic boundary deformation aligned in x-axis
We first consider a moving cylinder example in a channel as [76] for testing anisotropic
deformation. We set the coefficients of anisotropic FEMWARP in (4.11) as α=1 and β =0 for
this problem since deformation only occurs which is aligned with the x-axis. Figure 4.24 shows
the initial mesh, the deformed mesh generated by FEMWARP, the deformed mesh generated by
anisotropic FEMWARP (after the first step of hybrid algorithm), and the optimized mesh (after
160
the second step of the hybrid algorithm) on the deformed domain. Figure 4.24 shows that the
deformed mesh for FEMWARP has 38 inverted elements around the inner boundary, while our
hybrid algorithm does not have any inverted elements. Finally, Fig. 4.24(d) demonstrates that
our hybrid algorithm is able to recover similar element shapes on the deformed domain.
Figure 4.25 compares element qualities measured by the TMP shape metric, which mea-
sures similarity between the deformed and the initial elements. A smaller value indicates more
similarity with zero meaning the shapes are identical. The quality measure of the worst ele-
ment using our hybrid algorithm is 80.4% less than the one using UBN and 45.5% less than the
one produced by FEMWARP. Figure 4.26(a) shows the number of inverted elements for various
amounts of translation and (b) time to untangle inverted elements using multiobjective mesh op-
timization. This figure shows the number of inverted elements is reduced up to 91.2% when we
employ anisotropic FEMWARP compared with FEMWARP. Figure 4.26(b) shows the time to
untangle meshes (from (a)) using multiobjective mesh optimization. We observe that anisotropic
FEMWARP reduces the running time up to 90.6% over that of FEMWARP, when we employ
our multiobjective mesh optimization to untangle meshes with inverted elements. This is be-
cause anisotropic FEMWARP requires fewer iterations to untangle inverted meshes than does
FEMWARP.
4.10.2 Moving bar domain for anisotropic boundary deformation aligned in y-axis
We consider a moving bar example derived from a mechanical engineering problem [92]
on unstructured meshes. We choose the coefficients for anisotropic FEMWARP to be α=0 and
β =1 in (4.11) for this problem since deformation occurs which is aligned with the y-axis. Fig-
ure 4.27 shows the initial mesh, the deformed mesh generated by FEMWARP, the deformed mesh
161
Fig. 4.24 Moving cylinder domain for anisotropic boundary deformation aligned with the x-axis:
(a) Initial mesh for a moving cylinder in a channel. (b) Deformed mesh using FEMWARP for
a moving cylinder. This mesh has 38 inverted elements. (c) Deformed mesh using anisotropic
FEMWARP for a moving cylinder. This mesh has zero inverted elements. (d) Optimized mesh
on the deformed domain with no inverted elements using anisotropic FEMWARP followed by
multiobjective mesh optimization with TMP shape and untangling.
162
Fig. 4.25 Mesh quality measured by the TMP shape metric on the deformed cylinder domain. A
smaller value indicates more similarity with zero meaning identical shape. The worst element
quality of the deformed mesh which results from our hybrid algorithm is 80.4% less than the
quantity of the one using UBN [95]. Note that FEMWARP includes 38 inverted elements, while
the deformed meshes resulting from both our hybrid algorithm and UBN do not have any inverted
elements.
(a) Number of inverted elements after initial guess (b) Time (sec) to untangle inverted elements using
multiobjective mesh optimization
Fig. 4.26 Number of inverted elements with respect to the translation of inner boundary after
applying (a) FEMWARP (α=1 and β =1) and anisotropic FEMWARP (α=1 and β =0). The
inner boundary shifts right. (b) Time to untangle inverted elements using multiobjective mesh
optimization. This figure shows the time to untangle inverted meshes using multiobjective mesh
optimization.
163
generated by anisotropic FEMWARP (after the first step of hybrid algorithm), and the optimized
mesh (after the second step of hybrid algorithm) on the deformed domain. These figures clearly
show our hybrid algorithm simultaneously preserves element shape and eliminates inverted ele-
ments on the deformed domain. Also, our anisotropic FEMWARP (first step) has significantly
fewer inverted elements than does FEMWARP and helps to decrease the optimization time in
the second step of our hybrid algorithm. Figure 4.28 shows element qualities on the deformed
domain for our hybrid algorithm and for FEMWARP. For this domain, UBN is not able to find
the deformed domain due to the singular stiffness matrix which arises when numerical tolerances
are reached. More specifically, this is the result of numerical tolerance issues related to elements
with nearly zero area around the inner bar. The worst element quality measure produced by
our hybrid algorithm is 93.4% smaller than that produced by FEMWARP. FEMWARP produced
118 inverted elements, while our anisotropic FEMWARP produced only 17. Figure 4.29 shows
the number of inverted elements with respect to amount of translation and the time to untangle
inverted elements using multiobjective mesh optimization, respectively. Similar to the cylinder
domain, our anisotropic FEMWARP has significantly (up to 94.2%) fewer inverted elements af-
ter the initial guess step (step 1) and resulted in taking less time (up to 63.7%) to remove the
4.10.3 Moving gate domain for y axis aligned anisotropic boundary deformation
We consider a moving gate domain similar to one described in [92] where deformation
α=0 and β =1 in (4.11) for this anisotropic boundary deformation, since deformation only occurs
that is aligned with the x-axis. Figure 4.30 shows the initial mesh, the deformed mesh generated
164
Fig. 4.27 Moving bar domain for anisotropic boundary deformation aligned in y-axis:(a) Ini-
tial mesh and (b) zoomed-in on the bar domain. (c) Deformed mesh with FEMWARP and (d)
zoomed in mesh with FEMWARP. Deformed mesh with FEMWARP has 118 inverted elements.
(e) Deformed mesh using anisotropic FEMWARP and (f) zoomed-in mesh with anisotropic
FEMWARP. The deformed mesh with anisotropic FEMWARP has 17 inverted elements. (g)
Optimized mesh and (h) zoomed-in mesh with anisotropic FEMWARP followed by multiobjec-
tive mesh optimization. This optimized mesh does not have any inverted elements.
165
Fig. 4.28 Mesh quality measured by the TMP shape metric on the deformed bar domain. The
smaller value indicates more similarity with zero meaning identical shape. The worst element
quality of the deformed mesh which results from our hybrid algorithm is 93.5% less than the
quality of the mesh obtained from the use of FEMWARP [95]. Note that FEMWARP includes
118 inverted elements while the deformed mesh resulting from our hybrid algorithm does not
have any inverted elements. For this domain, UBN fails to find the deformed mesh due to
numerical tolerance issues.
by FEMWARP, the deformed mesh generated by anisotropic FEMWARP (after the first step of
hybrid algorithm), and the optimized mesh (after the second step of the hybrid algorithm) on
the deformed domain. This example clearly shows that anisotropic FEMWARP outperforms
FEMWARP. A very large number of inverted elements is generated around the center corner
(as shown in Fig. 4.30(b)), since FEMWARP connects vertices isotropically when the stiffness
preferentially connecting neighbors aligned with the y axis. Similar to previous examples, our
shown in Fig. 4.31. Similar to the moving bar example, UBN fails to find the deformed mesh due
to numerical tolerance issues. This is the result of numerical issues related to refined elements
(a) Number of inverted elements with respect to the (b) Time (sec) to untangle inverted elements using
amount of translation multiobjective mesh optimization
Fig. 4.29 Number of inverted elements with respect to the translation of inner boundary after
applying (a) FEMWARP (α=1 and β =1) and anisotropic FEMWARP (α=0 and β =1). The inner
bar moves down. (b) We apply multiobjective mesh optimization to simultaneously smooth
and untangle inverted elements. This figure shows timing to untangle inverted elements using
multiobjective mesh optimization. Our multiobjective mesh optimization is able to improve
element qualities while eliminating inverted elements but anisotropic FEMWARP takes up to
63.7% less time to eliminate inverted elements compared with FEMWARP since anisotropic
FEMWARP has fewer iterations to untangle inverted elements than FEMWARP..
We consider a moving cylinder example similar to that of Section 4.10.1; however, now
the inner cylinder moves diagonally along the vector [0.6, 0.1]T . Based on our strategies for set-
ting the coefficients of anisotropic FEMWARP described in Fig. 4.23, we choose the coefficients
of α=6 and β =1. Figure. 4.33 shows the initial, the deformed mesh generated by FEMWARP,
the deformed mesh generated by anisotropic FEMWARP (after the first step of the hybrid algo-
rithm), and the optimized mesh (after the second step of the hybrid algorithm) on the deformed
domain. Compared with FEMWARP, the number of inverted elements is reduced by 86%. Sim-
ilar to the previous examples, our hybrid algorithm results in meshes with better mesh quality
with no inverted elements and similar element shapes as shown in Fig 4.34. The worst element
quality of the deformed mesh which results from our hybrid algorithm is 79.2% less than the one
167
(c) Mesh on deformed domain with (d) Zoomed-in mesh on the de-
FEMWARP formed domain with FEMWARP
(e) Mesh on deformed domain with (f) Zoomed-in mesh on the de-
anisotropic FEMWARP (α = 0 and β = 1) formed domain with anisotropic
FEMWARP
(g) Optimized mesh on the deformed do- (h) Zoomed-in mesh on the de-
main with anisotropic FEMWARP followed formed domain with anisotropic
by multiobjective mesh optimization FEMWARP followed by multi-
objective mesh optimization
Fig. 4.30 Moving gate domain for anisotropic boundary deformation aligned in y-axis: (a) Initial
mesh and (b) zoomed-in on the gate domain. (c) The deformed mesh with FEMWARP and (d)
zoomed-in mesh with FEMWARP. Deformed mesh with FEMWARP has 103 inverted elements.
(e) The deformed mesh using anisotropic FEMWARP and (f) zoomed-in mesh with anisotropic
FEMWARP. The deformed mesh with anisotropic FEMWARP has 14 inverted elements. (g) The
optimized mesh and (h) zoomed-in optimized mesh on the deformed domain with no inverted
elements using anisotropic FEMWARP followed by multiobjective mesh optimization with TMP
shape and untangling.
168
Fig. 4.31 Mesh quality measured by the TMP shape metric on the deformed gate domain. The
smaller value indicates more similarity with zero meaning identical shape. The worst element
quality of the deformed mesh which results from our hybrid algorithm is 9.85% less than the
one using FEMWARP [95]. Note that FEMWARP includes 103 inverted elements while the
deformed mesh resulting from our hybrid algorithm does not have any inverted elements. For
this domain, UBN fails to find the deformed mesh due to numerical tolerance issues.
Fig. 4.32 Number of inverted elements with respect to the translation of outer boundary after ap-
plying (a) FEMWARP (α=1 and β =1) and anisotropic FEMWARP (α=0 and β =1). The corner
in the middle moves down. (b) We apply multiobjective mesh optimization to simultaneously
smooth and untangle inverted elements. This figure shows the time to untangle inverted ele-
ments using multiobjective mesh optimization. Our multiobjective mesh optimization is able to
improve element qualities while eliminating inverted elements but anisotropic FEMWARP takes
up to 90.6% less time to eliminate inverted elements compared with FEMWARP.
169
using UBN and 41.8% less than the one using FEMWARP. Note that FEMWARP also includes
We consider a bending bar deformation which is introduced in [92]. The initial mesh is
same as Fig. 4.10.1, but the bar undergoes a bending deformation. For this nonlinear deforma-
tion, we basically follow the same idea shown in Fig. 4.23. We compute the cumulative vertex
displacements in the x and y directions and use this information to choose values of the coef-
ficients for α and β . For this problem, our methodology results in coefficients of α = 1 and
β = 4. Figure 4.35 shows the initial mesh, the deformed mesh generated by FEMWARP, the
deformed mesh generated by anisotropic FEMWARP (after first step of hybrid algorithm), and
(d) the optimized mesh (after second step of hybrid algorithm) on the deformed domain. Com-
pared with FEMWARP, the number of inverted elements is reduced to 63.2%. Figure 4.36 shows
that our hybrid algorithm significantly improves mesh quality (by greater than 96%) compared
with FEMWARP. These results are consistent with other displacement magnitudes of bending.
Similar to the moving bar and gate domains, UBN is not able to find the deformed domain due
4.10.6 3D moving sphere domain for z axis aligned anisotropic boundary deformation
ner sphere is moving along the z axis. We choose the coefficients of anisotropic FEMWARP
to be α=0, β =0, and γ=1 in (4.20) for this problem since deformation only occurs aligned with
170
(d) Optimized mesh on the deformed domain with anisotropic FEMWARP followed by
multiobjective mesh optimization
Fig. 4.33 Moving cylinder domain for anisotropic diagonal boundary deformation: (a) Initial
mesh for a moving cylinder in a channel. (b) The deformed mesh using FEMWARP for a di-
agonal deformation. It has 101 inverted elements. (c) The deformed mesh using anisotropic
FEMWARP for a diagonal deformation. It has 15 inverted elements. (d) The optimized mesh
on the deformed domain with no inverted elements using anisotropic FEMWARP followed by
multiobjective mesh optimization with TMP shape and untangling.
171
Fig. 4.34 Mesh quality measured by the TMP shape metric on the deformed cylinder domain.
The smaller value indicates more similarity with zero meaning identical shape. The worst ele-
ment quality of the deformed mesh which results from our hybrid algorithm is 79.2% less than
the one using FEMWARP [95]. Note that FEMWARP includes 103 inverted elements while the
deformed mesh resulting from our hybrid algorithm does not have any inverted elements.
the z-axis. Figure 4.37 shows the initial surface mesh, the initial volume mesh, and the de-
formed volume mesh (after the second step of the hybrid algorithm) on the deformed domain.
Table 4.38 shows mesh quality statistics on the deformed domain using our hybrid algorithm
and FEMWARP, respectively. Here, a smaller value indicates more similarity with zero mean-
ing identical shape. We observe significant quality improvements of our hybrid algorithm over
FEMWARP. The worst element quality of the deformed mesh which results from our hybrid al-
gorithm is 86.2% less than the one using FEMWARP. Figure 4.39 shows the number of inverted
elements for various amounts of translation and the time to untangle inverted elements using
multiobjective mesh optimization. The overall trend is similar to those seen for 2D examples.
Anisotropic FEMWARP significantly helps to reduce the number of inverted elements as a first
step. The reduced number of inverted elements decreases the time needed to eliminate inverted
elements as a second step. Note that our second step, multiobjective mesh optimization, is able
172
Fig. 4.35 Bending bar domain for anisotropic boundary deformation:(a) Initial mesh and (b)
zoomed-in mesh on the bar domain. (c) Deformed mesh with FEMWARP and (d) zoomed-
in mesh with FEMWARP. Deformed mesh with FEMWARP has 159 inverted elements.
(e) Deformed mesh using anisotropic FEMWARP and (f) zoomed-in mesh with anisotropic
FEMWARP. Deformed mesh with anisotropic FEMWARP has 48 inverted elements. (g) Op-
timized mesh and (h) zoomed-in mesh with anisotropic FEMWARP followed by multiobjective
mesh optimization. This optimized mesh does not have any inverted elements.
173
Fig. 4.36 Mesh quality measured by the TMP shape metric on the deformed bar domain. The
smaller value indicates more similarity with zero meaning identical shape. The worst element
quality of the deformed mesh which results from our hybrid algorithm is 92.5% less than the
one using FEMWARP [95]. Note that FEMWARP includes 159 inverted elements while the
deformed mesh resulting from our hybrid algorithm does not have any inverted elements. For
this domain, UBN fails to find the deformed mesh due to numerical tolerance issues.
to simultaneously untangle inverted elements and improve qualities on the deformed domain for
4.11 Conclusions
We have presented a new hybrid mesh deformation algorithm for computing the interior
vertex positions of unstructured meshes subject to anisotropic boundary deformation. The aim
of this method is to overcome the shortcomings of existing mesh warping methods, namely
the propensity for creating inverted elements when the boundary deformation is large and the
inability to maintain qualities of the initial mesh in the deformed mesh. Our first step is an
anisotropic FEMWARP method which computes an initial guess for the interior vertex positions
and has proven to exactly preserve initial mesh characteristics for a restricted class of meshes
174
Fig. 4.37 Moving sphere in a 3D cube domain: (a) Surface mesh on the initial domain (b)
Volume mesh on the initial domain (c) Optimized volume mesh using our hybrid algorithm on
the deformed domain. This optimized mesh does not have any inverted elements.
175
Fig. 4.38 Mesh quality measured by the TMP shape metric on the deformed 3D sphere domain.
The smaller value indicates more similarity with zero meaning identical shape. The worst ele-
ment quality of the deformed mesh which results from our hybrid algorithm is 86.2% less than
the one using FEMWARP [95]. Note that FEMWARP includes 120 inverted elements while the
deformed mesh resulting from our hybrid algorithm does not have any inverted elements. For
this domain, UBN fails to find the deformed mesh due to numerical tolerance issues.
Fig. 4.39 The number of inverted elements with respect to the translation of outer boundary
after applying (a) FEMWARP (α=1, β =1, and γ=1) and anisotropic FEMWARP (α=0, β =0,
γ=1). (b) We apply multiobjective mesh optimization to simultaneously smooth and untangle
inverted elements. This figure shows the time to untangle inverted elements using multiobjective
mesh optimization. Our multiobjective mesh optimization is able to improve element qualities
while eliminating inverted elements but anisotropic FEMWARP takes up to 71.1% less time to
eliminate inverted elements compared with FEMWARP.
176
and deformation types and numerical examples demonstrate the suitability of this method for
more general meshes and deformations. Of particular importance is that this first step produces
meshes with fewer inverted elements then the original FEMWARP method. The second step
of our method performs a multiobjective optimization of element shape and untangling to both
untangle the mesh produced in the first step and to recover the initial mesh characteristics in the
deformed mesh. Our numerical examples show the robustness of this optimization method to
produce valid meshes which maintain many of the initial mesh characteristics. Our numerical
experiments have also demonstrated the importance of reducing the number of inverted elements
passed to the multiobjective optimization when attempting to minimize the time to solution.
There are several interesting ideas worth exploring which will be the subject of subse-
quent papers. These include attempting to capture some aspects of the initial mesh characteristics
in the linear system solved to generate the initial guess for the interior vertex positions. The TMP
framework provides a natural way to express these characteristics and can be used to set up such
a linear system. Additional work also remains in providing either better heuristics or a more
formalized mechanism for computing coefficients in the anisotropic FEMWARP method based
on the nature of the boundary deformation. We suspect that our efforts in this direction are but a
In this chapter, we have developed a hybrid mesh deformation algorithm which finds
the optimal interior vertex positions on the deformed domain when boundary deformations are
known. In the next chapter, we solve a more challenging mesh deformation problem, which is a
shape matching problem. We use optimization-based meshing techniques to solve this problem.
We focus on finding the optimal boundary points on the target image using triangular shapes and
image gradients.
177
Chapter 5
An Improved Shape Matching Algorithm for Deformable
5.1 Introduction
object tracking and image-based searches [98, 99]. The goal of shape matching is to match
the source image to the target image, i.e., the deformed image. Felzenszwalb proposed a shape
matching algorithm for deformable objects using triangular meshes and dynamic programming
in [100]. Felzenszwalb’s algorithm was novel in that it does not require any initialization to de-
tect the target image unlike previous algorithms (e.g., [101]). A modification of Felzenszwalb’s
algorithm using flexible shape priors was proposed in [102]. In [102], large deformations on
flexible regions were allowed through the use of the shape priors. However, it is hard to know
which parts should be made flexible in advance, and thus significant user knowledge is required.
Moreover, [100, 102] do not use global image features to detect deformable objects.
Recently, several papers have used global image features for shape matching. For exam-
ple, hierarchy-based shape matching algorithms for deformable objects were proposed in [103,
178
104]. In [103], the authors identified shape parts, composed of multiple salient points, and used
many-to-many matching for shape matching. The authors in [104] used a shape tree to capture
both local and global shape information and employed dynamic programming to perform shape
matching.
We use local and global shape information to perform shape matching in a complemen-
tary method compared with that of [104]. In particular, we use triangular meshes and dynamic
programming for shape matching, as in [100, 102], and extend the algorithm in [100] through
the use of an added global image feature. The rest of this paper is organized as follows. In
Section 5.2, we describe the three-step shape matching process. In Section 5.3, we present ex-
perimental results, and compare our experimental results with those of algorithm in [100] (i.e.,
the algorithm to which ours is the most similar). Finally, we give some conclusions and plans
for future work in Section 5.4. The contents of this chapter have been published in [109].
Our shape matching process is composed of three steps. The first step is to determine
boundary vertices which approximate the source image boundary. The second step is to gen-
erate a triangular mesh using the constrained Delaunay triangulation method to represent the
deformable object. The third step is to find the optimal mapping from the source image to the
target image which minimizes our energy function. We now describe each of the three steps in
more detail.
179
5.2.1 Determination of the boundary vertices approximating the source image boundary
This step determines the number of boundary vertices to use in the source image. In
order to find a polygonal approximation to a boundary curve in the source image, S, we create
nearly equally-spaced vertices on the boundary of S. We select vertices such that the distance
between them is as close to an ideal distance parameter as possible, and the boundary curve is
5.2.2 Generation of the triangular mesh on the source image using the constrained De-
This step combines the boundary vertices into triangles to represent the non-rigid objects.
The constrained Delaunay triangulation method [105] implemented in Triangle [36] is used to
generate a triangular mesh, M, that respects the boundary of S, without adding any interior
vertices, to represent the deformable parts of the image. This step is shown in Figure 5.1(b).
Note that a triangular mesh without interior vertices can be represented using a dual graph.
The vertices in the dual graph of M can be ordered and eliminated using a perfect elimination
scheme [106].
In order to determine the optimal placement of the boundary vertices on a target image,
T , we formulate and solve an optimization problem with the goal of determining the mapping,
f , from the triangles in S to the triangles in T , which has the lowest energy. Unlike the energy
(cost) functions in [100, 102], our energy function is composed of three terms: an edge cost, a
triangular deformation cost, and a triangular edge length distortion cost. The edge cost of the ith
180
Fig. 5.1 Overview of the shape matching process. The function f maps triangles in the triangular
mesh on the source image to a triangular mesh on the target image. (a) Equally-spaced boundary
vertices are generated. (b) The triangular mesh is created. (c) The detected image is illustrated
on the target image.
triangle, Eedge,i , corresponds to a simple edge detector by assigning high costs to a low image
gradient magnitude in T . Thus, Eedge,i increases if edges detected on T are not placed on the
1
Eedge,i = ,
λ + |OI|
where λ is a constant.
The triangular deformation cost, Ede f ,i , represents how far the original triangle is trans-
formed from a similarity transform [107] when the mapping from the source image to the target
image occurs. The affine transformation of each triangle from S to T takes a unit circle to an
ellipse with major and minor axes of lengths, α and β , respectively. Here, α and β are the sin-
gular values of the matrix associated with the affine transformation. The squared log-anisotropy
of α and β is used to measure the triangular deformation as in [100]. The triangular deformation
α 2
Ede f ,i = log .
β
Unlike [100, 102], we introduce a new cost, the triangular edge length distortion cost,
Edis,i , which penalizes the sum of the edge length distortions of each triangle. Let l j,s and l j,t
be the jth edge lengths of a triangle in S and T , respectively. Then, the sum of the three edge
3 3
lsum,s = ∑ l j,s , lsum,t = ∑ l j,t .
j=1 j=1
Then, γ is defined as
l sum,s /lsum,t if lsum,s > lsum,t
γ=
lsum,t /lsum.s if lsum,t > lsum,s .
To make the triangular edge length distortion cost invariant to uniform scaling from the
source image to the target image, we use the center of mass of an image. Let dmax,s and dmax,t be
the maximum distances from the center of masses to the boundary vertices in S and T , respec-
tively. Figure 5.2 illustrates dmax,s and dmax,t for a fish shape in the Brown dataset. Then, δ is
defined as follows:
d max,s /dmax,t if dmax,s > dmax,t
δ=
dmax,t /dmax,t if dmax,s < dmax,t .
The triangular edge length distortion cost, Edis,i , is large when δ and γ are different. This
occurs when the transformation of the triangles from the source image to the target image does
182
(a) (b)
Fig. 5.2 (a) A sample image shape from the Brown dataset [108]. (b) The dot in the middle
represents the center of mass for the image (this is denoted as C in the target image), and the
arrow represents the maximum distance from the center of mass in the image to the boundary
vertices (i.e., dmax,s or dmax,t )
log(γ/δ ) if γ > δ
Edis,i =
log(δ /γ) if γ < δ .
The energy function, E( f ), is defined as a weighted sum of the three cost terms over the
N N
E( f ) = ∑ Ei = ∑ (Eedge,i + a ∗ Ede f ,i + b ∗ Edis,i ), (5.1)
i=1 i=1
where, a and b are scalars. Note that small values of a and b are desirable for highly deformable
objects. Let n be the number of vertices in M. Our goal is to determine the f that minimizes E( f ).
To solve the optimization problem, the shape matching algorithm uses a dynamic programming
method given in Algorithm 1. For computational efficiency, a discrete set of grid locations, G,
in T is assumed. In line 3 of Algorithm 1, vertices i and j are the parents of the kth vertex. The
The optimal matching image (detected mesh) with the lowest possible energy is found
using the dynamic programming method shown in Algorithm 1. Note that, for some cases,
matching based on the smallest energy does not guarantee a good match. For those cases, we
choose the matching image with the next smallest energy. In order to choose the criteria for the
sub-optimal cases, we employ the center of mass of an image, which is a global feature of an
image. Let the center of mass in T be C and the center of the matched vertices in T be Ĉ. We
define D as follows:
D = C − Ĉ 2
.
poor as shown in Figure 5.3. The threshold value, θ , can be defined using dmax,t , the maximum
distance from C to the boundary vertices in T . In Figure 5.3, θ is set to dmax,t /3. In this case, we
select instead the next smallest energy, and find a new optimal mapping (i.e., by executing the
D ≤ θ. (5.2)
Fig. 5.3 Poor shape matching result of the algorithm in [100] (left) and improved matching
result using the center of mass in the target image (right). For these figures, the detected mesh is
illustrated on the target image. For this experiment, θ = dmax,t /3.
5.3 Experiments
This section describes the experimental evaluation of our shape matching algorithm and
provides a comparison of our results with those in [100], i.e., the shape matching algorithm
which is the most similar to ours. We use the well-known Brown [108] dataset to test our
algorithm. The Brown dataset has 99 images divided into 9 categories (each of which has 11
images). The Brown dataset is challenging because it includes both occluded images and images
with missing parts. Sample images from the Brown dataset are illustrated in Figure 5.4. We use
the standard evaluation method to test our algorithm. For each image, we count how many of
the 10 best matches (i.e., as defined by the smallest cost) belong to the same category. For our
algorithm, the detected mesh with the smallest cost must also satisfy (2). We use 25 pixels for
the distance between successive boundary vertices. The grid size in the target images is set to
185
30x30. We use a=4b in (1). The threshold value θ is set to dmax,t /3. Our experimental results
show that choosing a value of θ between dmax,t /3 and dmax,t /2 gives good matching results.
Fig. 5.4 Sample images in the Brown dataset [108]. Three sample images are shown per category.
Figure 5.5 shows some of the good matching results obtained with our algorithm. For
these source images, both our algorithm and Felzenszwalb’s algorithm show good matching re-
sults. However, our algorithm shows slightly better matching results. Felzenszwalb’s algorithm
shows incorrect matching results for some cases (e.g., the bottom right images). Interestingly,
we see that our algorithm and Felzenszwalb’s algorithm show different rank orderings among
the 10 best matching results. Figure 5.6 shows some poor matching results for our matching
algorithm on the Brown data set. For these images, Felzenszwalb’s algorithm fails in most cases
The recognition rates comparing our algorithm with Felzenszwalb’s algorithm are shown
in Table 5.1. The recognition rate is defined as the ratio of the total number of correct hits to
the total number of correct hits possible [104]. For the Brown data set, the total number of
correct hits possible is 99*10 since there are 99 images, and we find the 10 best matches. Our
algorithm yields a 7.8% higher recognition rate when compared with Felzenszwalb’s algorithm
on the Brown data set. Several matching results, which include detected meshes on the target
186
Fig. 5.5 Good shape matching results for the Brown dataset on three source (query) images
and comparison with [100]. For each source image, the 10 best matching results are shown
with the smallest (left) to the largest (right) energy. The top figures in each group represent
the matching results obtained from our algorithm, whereas the bottom figures in each group
represent matching results using the algorithm in [100]. For these experimental sets, only two
matching results of [100] (i.e., the bottom right images) fail to match.
Fig. 5.6 Poor shape matching results for the Brown dataset on three source (query) images and
comparison with [100]. For each source image, the 10 best matching results are shown with the
smallest (left) to the largest (right) energy. The top figures in each group represent the matching
results obtained from our algorithm, and the bottom figures in each group represent matching
results using the algorithm in [100]. For this experimental data set, both our algorithm and [100]
show poor matching results. However, our algorithm shows better matching results than does
the method in [100].
187
image, are shown in Figure 5.7. We observe that the algorithm in [100] produces poor matching
results when detected meshes (triangles) are placed in poor positions within a target image. This
often results in a large D value. This occurs because [100] matches the target image only by
considering the local properties of each triangle such as shape similarity and the edge bound-
ary. However, our algorithm produces better matching results by using global image features as
We have proposed a shape matching algorithm for deformable objects using both local
and global shape information. In particular, we employ the center of mass of an image as a global
image feature. Similar to the algorithm in [100], our algorithm does not need initialization to
detect deformable objects. Experimental results show that Felzenszwalb’s algorithm [100] some-
times produces poor matching results because it only considers the local properties of each tri-
angle such as shape similarity and the edge boundary. However, our method produces improved
matching results using a new energy function term and improved dynamic programming based
upon global image features. Experimental results show a 7.8% higher recognition rate compared
Fig. 5.7 Example matching results including detected meshes on the target image using images
from the Brown dataset [108]. For this experiment, Felzenszwalb’s algorithm shows poor match-
ing results because triangles are placed at poor positions.
189
Chapter 6
Conclusions and Future Work
This dissertation explores optimization-based meshing techniques for improving the mesh
quality for partial differential equations (PDEs) and nonlocal peridynamics problems, shape
matching and mesh deformation problems. In the first part of the dissertation, we studied mesh
quality improvement and mesh untangling methods for the solution of partial differential equa-
tions and nonlocal peridynamics models. We observed the importance of improving the mesh
quality and the effect of these improved mesh qualities on the solution of PDEs and nonlocal
peridynamic models. In the second part of this dissertation, we explored mesh deformation
problems. We investigated how to update finite element meshes when geometric domains vary
with respect to time and how to utilize triangular meshes for shape matching problems.
preconditioners, and sparse linear solvers, for solving various elliptic PDE problems. We ob-
served the importance of choosing appropriate preconditioners and mesh quality metrics for effi-
ciently solving these elliptic PDE problems. In terms of mesh quality metrics, the interpolation-
based size and- shape (SS) metric is the least efficient mesh quality metric in terms of its effect
on the solver time. This is because meshes smoothed with the SS metric have more elements
190
with very small or large angles than do meshes smoothed by other quality metrics. We observed
that in most cases, the radius ratio (RR) metric is the best quality metric which results in fastest
time to solve the sparse linear system. We plan to investigate the most efficient combinations for
solving other time-dependent PDE problems such as parabolic and hyperbolic PDEs and also
plan to study the best combinations which result in the smallest error for solving PDEs.
ity improvement and mesh untangling in order to satisfy multiple requirements of the mesh.
Among various multiobjective mesh optimization methods in our framework, the exponential
sum multiobjective method was the most effective method we tested in simultaneously improv-
ing two or more aspects of the mesh. The exponential sum multiobjective method is effective in
improving the worst cost objective function efficiently when the initial mesh quality was poor
and the mesh contained inverted elements. The equal sum multiobjective method is simple and
shows comparable performance to the exponential sum method. We expect that our multiobjec-
tive mesh optimization framework is able to be extended to other mesh element types such as
hex and mixed elements. We also showed that this multiobjective mesh optimization framework
In Chapter 4, we observed the connections that mesh anisotropy, mesh refinement, and
kernel functions have on the conditioning of the stiffness matrix for the bond-based nonlocal
peridynamics problems. Different from general elliptic PDE problems, we observed that both
anisotropy and mesh refinement do not affect the condition number when we employ integrable
kernel functions. We also observed that the condition number behavior is similar to general
elliptic PDEs when the piecewise linear basis function and the nonintegrable kernel function
191
were used. This study can be easily extended to other nonlocal peridynamics problems and also
In Chapter 5, we proposed a hybrid mesh deformation algorithm which uses the anisotropy
of the boundary deformation and multiobjective mesh optimization for deforming domains. We
observed that our hybrid mesh deformation algorithm outperforms existing mesh deformation
algorithms in terms of mesh quality and number of inverted elements and is able to preserve
similar element shapes on the deformed domain while eliminating inverted elements on the de-
formed domain.
jects. We observed that the deformable objects can be discretized into triangular meshes and
these triangular meshes with new cost functions can be successfully used to improve the recog-
nition rates for shape optimization problems. We plan to consider nonconstant PDE coefficients
For large scientific problems, parallel computing libraries such as OpenMP or MPI are
often used to speed up the solver. Currently, MPI is implemented in the mesh optimization
software package, Mesquite. We plan to implement Mesquite in OpenMP and compare the two
implementations on mesh quality, speed-up, and power consumption. Also, we plan to simulta-
neously use OpenMP and MPI in hybrid algorithms and to determine the optimal combinations
of OpenMP and MPI for parallel mesh optimization problems. Domain decomposition is nec-
essary for using MPI, and the optimal domain decomposition for mesh optimization is another
interesting topic. These domain decompositions affect the communication costs between mul-
tiple cores and the mesh quality, speed up, and power consumption of the HPC machines used.
192
We also plan to propose a new domain decomposition method that is able to achieve high-quality
meshes while reducing the solver time and consuming less power.
193
References
[1] S. Bhowmick and S.M. Shontz, Towards high-quality, Untangled meshes via a force-directed
graph embedding approach, in Proc. of the 2010 International Conference on Computational
Science, Procedia Computer Science, pp. 357-366, May 2010, .
[2] J. Danczyk and K. Suresh, Finite element analysis over tangled meshes, in Proc. of the
ASME 2012 International Design Engineering Technical Conferences & Computers and
Information in Engineering Conference, Chicago, IL, pp. 1-7, 2012.
[3] G. H. Golub and C.F. Van Loan, Matrix Computations, 3rd Edition, 1996.
[4] I. Babuska and M. Suri, The p and h-p versions of the finite element method, basic principles,
and properties, SIAM Rev., 35: 579-632, 1994.
[5] M. Berzins, Solution-based mesh quality for triangular and tetrahedral meshes, in Proc. of
the 6th International Meshing Roundtable, Sandia National Laboratories, pp. 427-436, 1997.
[6] M. Berzins, Mesh quality - Geometry, error estimates, or both?, in Proc. of the 7th Interna-
tional Meshing Roundtable, Sandia National Laboratories, pp. 229-237, 1998.
[7] E. Fried, Condition of finite element matrices generated from nonuniform meshes, AIAA
Journal, 10: 219-221, 1972.
[8] Q. Du, Z. Huang, and D. Wang, Mesh and solver co-adaptation in finite element methods for
anisotropic problems, Numer. Methods Partial Differential Equations, 21: 859-874, 2005.
[9] Q. Du, D. Wang, and L. Zhu, On mesh geometry and stiffness matrix conditioning for general
finite element spaces, SIAM J. Numer. Anal., 47(2): 1421-1444, 2009.
[10] A. Ramage and A. Wathen, On preconditioning for finite element equations on irregular
grids, SIAM J. Matrix Anal. Appl., 15: 909-921, 1994.
[11] A. Chatterjee, S.M. Shontz, and P. Raghavan, Relating mesh quality metrics to sparse lin-
ear solver performance, in Proc. of the SIAM Conference on Computational Science and
Engineering, Costa Mesa, CA, 2007.
[12] D. Mavripilis, An assessment of linear versus nonlinear multigrid methods for unstructured
mesh solvers, J. Comput. Phys., 175: 302-325, 2002.
194
[13] M. Batdorf, L. Freitag, and C. Ollivier-Gooch, Computational study of the effect of un-
structured mesh quality on solution efficiency, in Proc. of the 13th CFD Conference, AIAA,
Reston, VA, 1997.
[16] M. Brewer, L. Freitag Diachin, P. Knupp, T. Leurent, and D. Melander, The Mesquite mesh
quality improvement toolkit, in Proc. of the 12th International Meshing Roundtable, Sandia
National Laboratories, pp. 239-250, 2003.
[17] S. Balay, K. Buschelman, W.D. Gropp, D. Kaushik, M.G. Knepley, L.C. McInnes, B. F.
Smith, and Hong Zhang, PETSc Webpage, http://www.mcs.anl.gov/petsc, 2009.
[18] T. Munson, Mesh shape-quality optimization using the inverse mean-ratio metric, Math.
Program., 110: 561-590, 2007.
[19] R. Barrett, M. Berry, T.F. Chan, J. Demmel, J. Donato, J. Dongarra, V. Eijkhout, R. Pozo,
C. Romine, and H.Van der Vorst. Templates for the solution of linear systems: Building
blocks for iterative methods, 2nd Edition., SIAM, 1994.
[20] A. Brandt, S. McCormick, and J. Ruge, Algebraic Multigrid for Sparse Matrix Equations,
Cambridge Univ. Press, Cambridge, 1985.
[21] M.R. Hestenes and E. Stiefel. Methods of conjugate gradients for solving linear systems.
J. Res. Natl. Bur. Stand., 49:409-436, 1952.
[22] C.C. Paige and M.A. Saunders. Solution of sparse indefinite systems of linear equations.
SIAM J. Numer. Anal., 12:617-629, 1975.
[23] Y. Saad and M. Schultz. GMRES: A generalized minimal residual algorithm for solving
nonsymmetric linear systems, SIAM J. Sci. Comput., 7(3):856-869, 1986.
[24] H.A. van der Vorst. Bi-CGSTAB: A fast and smoothly converging variant of Bi-CG for the
solution of nonsymmetric linear systems. SIAM J. Sci. Comput., 13:631-644, 1992.
[25] J. Kim, S.P. Sastry, and S.M. Shontz, Efficient solution of elliptic partial differential equa-
tions via effective combination of mesh quality metrics, preconditioners, and sparse linear
solvers, in Proc. of the 19th International Meshing Roundtable, Sandia National Laborato-
ries, pp. 103-120, 2010.
[26] E.B. Becker, G.F. Carey, and J.T. Oden. Finite Elements: An Introduction. Prentice-Hall,
Englewood Cliffs, New Jersey, 1981.
[27] Y. Kwon and H. Bang, The Finite Element Method using Matlab, CRC Press, 2nd Edition.,
2000.
195
[28] S. Oh and J. Yim, Optimal finite element mesh for elliptic equation of divergence form,
Appl. Math. Comput., 162:969-989, 2005.
[29] L. Diachin, P. Knupp, T. Munson, and S. Shontz, A comparison of two optimization meth-
ods for mesh quality improvement, Eng. Comput., 22(2): 61-74, May 2006.
[30] J. Nocedal and S. Wright, Numerical Optimization, Springer-Verlag, 2nd Edition, 2006.
[31] M. Brewer, Obtaining smooth mesh transitions using vertex optimization, Internat. J. Nu-
mer. Methods Engrg., 75: 555-576, 2008.
[32] C. Johnson. Numerical Solution of Partial Differential Equations by the Finite Element
Method, Cambridge University Press, 1987.
[33] A.H. Baker, E.R. Jessup, and Tz.V. Kolev, A simple strategy for varying the restart param-
eter in GMRES(m), J. Comput. Appl. Math. 230: 751-761, 2009.
[34] Y. Saad, Iterative methods for sparse linear systems, Society for Industrial and Applied
Mathematics, 2nd Edition., 2003.
[36] J.R. Shewchuk, Triangle: Engineering a 2D quality mesh generator and Delaunay trian-
gulator, Lecture Notes in Computer Science, vol. 1148, pp. 203-222, 1996.
[37] B. Norris, L. McInnes, S. Bhowmick, and L. Li, Adaptive numerical components for PDE-
Based simulations, in Proc. of the 6th International Congress on Industrial Applied Mathe-
matics, 2007.
[38] S.M. Shontz and P.M. Knupp, The effect of vertex reordering on 2D local mesh optimization
efficiency, in Proc. of the 17th International Meshing Roundtable, Sandia National Laborato-
ries, pp. 107-124, 2008.
[39] P. Vaněk, J. Mandel, and M. Brezina, Algebraic multigrid by smoothed aggregation for
second and fourth order elliptic problems, Computing, 56(3):179-196, 1996.
[40] P.M. Knupp, Algebraic mesh quality metrics, SIAM J. Sci. Comp. 23:193-218, 2001.
[41] J. Shewchuk, What is a good linear element? Interpolation, conditioning, and quality mea-
sures, in Proc. of the 11th International Meshing Roundtable, Sandia National Laboratories,
p. 115-126, 2002.
[42] P.M. Knupp, Algebraic mesh quality metrics for unstructured initial meshes, Finite Elem.
Anal. Des., 39: 217-241, 2003.
[43] P.M. Knupp, Hexahedral and tetrahedral mesh untangling, Eng. Comput., 17(3): 261-268,
2001.
[45] J. Franks and P. Knupp, A new strategy for untangling 2D meshes via node-movement,
CSRI Summer Proceedings, Sandia National Laboratories, p. 152-165, 2010.
[46] L.A. Freitag and P. Plassmann, Local optimization-based simplicial mesh untangling and
improvement, Int. J. Numer. Methods Engng. 49:109-125, 2000.
[47] F.Y. Cheng and D. Li, Multiobjective optimization of structures with and without control,
J. Guid. Control. Dyn., 19: 392-397, 1996.
[49] B. Pellenard, P. Alliez, and J. Morvan, Isotropic 2D quadrangle meshing with size and
orientation control, in Proc. of the 20th International Meshing Roundtable, Sandia National
Laboratories, p. 81-98, 2011.
[53] S.P. Sastry, Dynamic Meshing Techniques for Mesh Quality Improvement, Untangling, and
Warping, Ph.D. Thesis, The Pennsylvania State University, 2012.
[55] M.L. Parks, P. Seleson, S.J. Plimpton, R.B.Lehoucq, and S.A. Silling, Peridynamics with
LAMMPS: A User Guide, Technical Report SAND2010-5549, Sandia National Laborato-
ries, August 2010.
[56] S.A. Silling, Reformulation of elasticity theory for discontinuities and long-range forces,
J. Mech. Phys. Solids 48: 175-209, 2000.
[57] O. G. Bakunin, Turbulence and diffusion: Scaling versus equations, Springer Series in
Synergetics, Springer, 2008.
[58] S. Silling, M. Epton, O. Weckner, J. Xu, E. Askari, Peridynamic states and constitutive
modeling, J. Elast. 88: 151-184, 2007.
[59] Y. Kebing, X.J. Xin, K.B. Lease, A new method of adaptive integration with error control
for bond-based Peridynamics, in Proc. of the World Congress on Engineering and Computer
Science, San Francisco, 2010.
197
[60] G. Dagan, The significance of heterogeneity of evolving scales to transport in porous for-
mations, Water Resour. Res. 30: 3327-3336, 1994.
[61] F. Bobaru, S. A. Silling, and H. Jiang, Peridynamic fracture and damage modeling of mem-
branes and nanofiber networks. in Proc. of XI International Conference Fract., Turin, pp.
5748:1-6, 2005.
[62] F. Bobaru, Influence of van der Waals forces on increasing the strength and toughness in
dynamic fracture of nanofiber networks: A peridynamic approach, Model. Simul. Mater. Sc.
15: 397-417, 2007.
[63] E. Emmrich, R.B. Lehoucq, and D. Puhst, Peridynamics: A nonlocal continum theory, to
appear in Meshfree Methods for Partial Differential Equations VI, 2012.
[64] S. A. Silling and R. B. Lehoucq, Peridynamic theory of solid mechanics, Adv. Appl. Mech.
44: 73-168, 2010.
[65] S.A. Silling and R.B. Lehoucq, Convergence of peridynamics to classical elasticity theory,
J. Elasticity 93: 13-37, 2008.
[66] M. Gunzburger and R.B. Lehoucq, A nonlocal vector calculus with application to nonlocal
boundary value problems, Multiscale Model. Simul. 8: 1581-1598, 2010.
[67] Q. Du, J.R. Kamm, R. B. Lehoucq, and M.L. Parks. An approach to nonlocal, nonlinear
advection, Technical report 2011-3164J, Sandia National Laboratories, 2011.
[68] B. Aksoylu and M.Parks, Variational theory and domain decomposition for nonlocal prob-
lems, Applied Mathematics and Computation 217: 6498-6515, 2011.
[69] X. Chen and M. Gunzburger, Continuous and discontinuous finite element methods for a
peridynamics model of mechanics, Comput. Methods Appl. Mech. Engrg. 200: 1237-1250,
2010.
[70] R.W. Macek and S.A. Silling, Peridynamics via finite element analysis, Finite Elem. Anal.
Des. 43: 1169-1178, 2007.
[71] S.A. Silling and E. Askari, A meshfree method based on the peridynamic model of solid
mechanics, Comput. Struct. 83, Advances in Meshfree Methods, 1526-1535, 2005.
[72] K. Zhou and Q. Du, Mathematical and numerical analysis of linear peridynamic models
with nonlocal boundary conditions, SIAM J. Numer. Anal. 48: 1759-1780, 2010.
[73] J. Kim, S.P. Sastry, and S.M. Shontz, A numerical investigation on the interplay among
geometry, meshes, and linear algebra in the finite element solution of elliptic PDEs, Eng.
Comput., 28(4): 431-450, 2012.
[75] M. Picasso. Adaptive finite elements with large aspect ratio based on an anisotropic error
estimator involving first order derivatives, Comput. Method Appl. M., 196(1):14-23, 2006.
198
[76] B.T. Helenbrook, Mesh deformation using the biharmonic operator, Internat. J. Numer.
Methods Engrg., 56: 1007-1021, 2003.
[77] Sigal, I.A., M.R. Hardisty, C.M. Whyne, Mesh-morphing algorithms for specimen-specific
finite element modeling, J. Biomech., 41(7):1381-1389, 2008.
[78] A.W. F. Lee, D. Dobkin, W. Sweldens, P. Schroder, Multiresolution mesh morphing, in
Proc. of the 26th SIGGRAPH Conference, pp. 343-350, 1999.
[79] M.T. Bah, P.B. Nair, and M. Browne, Mesh morhping for finite element analysis of implant
positioning in cementless total hip replacements, Med. Eng. Phys., 31:1235-1243, 2009.
[80] M.A. Baldwin, J.E. Langenderfer, P.J. Rullkoetter, and P.J. Laz, Development of subject-
specific and statistical shape models of the knee using an efficient segmentation and mesh-
morphing approach, Comput. Meth. Prog. Bio, 97:232-240, 2010.
[81] J. Park, S.M. Shontz, and C.S. Drapaca, A combined level set/mesh warping algorithm for
tracking brain and cerebrospinal fluid evolution in hydrocephalic patients, Invited submis-
sion to Image-Based Geometric Modeling and Mesh Generation, Springer, Lecture Notes in
Computational Vision and Biomechanics, Volume 3, 107-141, 2013.
[82] J. Park, S.M. Shontz, and C.S. Drapaca, Automatic boundary evolution tracking via a com-
bined level set method and mesh warping technique: Application to hydrocephalus, in Proc.
of the MICCAI Workshop on Mesh Processing in Medical Image Analysis, pp. 122-133,
2012.
[83] Luke, E., E. Collins, and E. Blades, A fast mesh deformation method using explicit inter-
polation, J. Comput. Phys., vol. 231:586-601, 2012.
[84] T.J. Baker, Mesh movement and metamorphosis, in Proc. of the 10th International Meshing
Roundtable, 387-396, 2001.
[85] J. Antaki, G. Blelloch, O. Ghattas, I. Malcevic, G. Miller, and N. Walkington. A parallel
dynamic mesh Lagrangian method for simulation of flows with dynamic interfaces. in Proc.
of the 2000 Supercomputing Conference, pp. 26, 2000.
[86] D. Cardoze, A. Cunha, G. Miller, T. Phillips, and N. Walkington. A Bezier-based approach
to unstructured moving meshes, in Proc. of the 20th ACM Symposium on Computational
Geometry, 2004.
[87] D. Cardoze, G. Miller, M. Olah, and T. Phillips. A Bezier-based moving mesh framework for
simulation with elastic membranes, in Proc. of the 13th International Meshing Roundtable,
pp. 71-80. Sandia National Laboratories, 2004.
[88] S.M. Shontz and S.A. Vavasis, Analysis of and workarounds for element reversal for a finite
element-based algorithm for warping triangular and tetrahedral meshes, BIT, Numerical
Mathematics, 50: 863-884, 2010.
[89] J. Kim, T. Panitanarak, and S.M. Shontz, A multiobjective mesh optimization framework
for mesh quality improvement and mesh untangling, Submitted to the Internat. J. Numer.
Methods Engrg, Accepted, September 2012.
199
[90] X.S. Li, An entropy-based aggregate method for minimax optimization, Eng. Optim., 18:
277-285, 1992.
[91] R.T. Marler and J.S. Arora, Survey of multi-objective optimization methods for engineering,
Struct. Multidisc. Optim., 26: 369-395, 2004.
[92] K. Stein, T. Tezduyar, and R. Benney, Mesh moving techniques for fluid-structure interac-
tions with large displacements, Trans. ASME, 70: 58-63, 2003
[93] K. Stein, T. Tezduyar, and R. Benney, Automatic mesh update with the solid-extension mesh
moving technique, Comput. Method Appl. M., 193:2019-2031, 2004.
[94] M. Staten, S. Owen, S.M. Shontz, A.G. Salinger, and T.S. Coffey, A comparison of mesh
morphing methods for 3D shape optimization, in Proc. of the 21st International Meshing
Roundtable, 293-312, 2011.
[95] S.M. Shontz, S.A. Vavasis, A robust solution procedure for hyperelastic solids with large
boundary deformation, Eng. with Comp., 28:135-147, 2012.
[96] S.P. Sastry, S.M. Shontz, and S.A. Vavasis, A log-barrier method for mesh quality im-
provement and untangling, Invited submission; Submitted to Engineering with Computers;
Accepted, September 2012.
[97] S.M. Shontz and S.A. Vavasis, A mesh warping algorithm based on weighted Laplacian
smoothing, in Proc. of the 12th International Meshing Roundtable, pp. 147-158, 2003.
[98] A. Yilmaz, O. Javed, and M. Shah, Object tracking: A survey, ACM Comput. Surv., 38(4),
pp. 1–45, 2006.
[99] T. Yeh, K. Tollmar, and T. Darrell, Searching the web with mobile images for location
recognition, in Proc. of the 17th IEEE Conference on Computer Vision and Pattern Recog-
nition, pp. 76-81, 2004.
[100] P.F. Felzenszwalb, Representation and detection of deformable shapes, IEEE PAMI.,
27(2), pp. 208-220, 2005.
[101] M. Kass, A. Witkin, and D. Terzopoulos, Snakes: active contour models, Int. J. Comput.
Vision, 1(4), pp. 321-331, 1988.
[102] T. Chang and T. Liu, Detecting deformable objects with flexible shape priors, in Proc. of
the 17th IEEE International Conference on Pattern Recognition, pp. 155-158, 2004.
[103] N. Payet and S. Todorovic, Matching hierarchies of deformable shapes, GBR, pp. 1-10,
2009.
[104] P.F. Felzenszwalb and J.D. Schwartz, Hierarchical matching of deformable shapes, in
Proc. of the 20th IEEE Conference on Computer Vision and Pattern Recognition, pp. 1-8,
2007.
[105] R. Seidel, Constrained Delaunay triangulations and Voronoi diagrams with obstacles,
Technical Report 260, Inst. for Information Processing, Graz, Austria, pp. 178-191, 1988.
200
[106] M.C. Golumbic, Algorithmic Graph Theory and Perfect Graphs, Academic Press, New
York, 1980.
[107] B. Widrow, The rubber mask technique, Pattern Recognition, 5(3), pp. 174-211, 1973.
[108] T.B. Sebastian, P.N. Klein, and B.B. Kimia, Recognition of shapes by editing their shock
graphs, IEEE PAMI., 26(5), pp. 550-557, 2004.
[109] J. Kim and S.M. Shontz, An improved shape matching algorithm for deformable objects
using a global image feature, in Proc. of the 6th International Symposium on Visual Com-
puting, Part III, LNCS 6455, pp. 119-128, 2010.
[110] J. Kim, L. Tian, S.M. Shontz, and Q. Du, The effect of anisotropy, mesh refinement, and
kernel functions on the conditioning of the stiffness matrix for a nonlocal peridynamics
model, to be submitted to Applied Mathematics and Computation, 2012.
[111] J. Kim, B.J. Miller, and S.M. Shontz, A hybrid mesh deformation algorithm using
anisotropy and multiobjective mesh optimization, to be submitted to the International Journal
for Numerical Methods in Engineering, 2012.
Vita
Jibum Kim received his B.S. degree in Electrical Engineering and M.S. degree in Electrical
Engineering from Yonsei University, Korea, in 2003 and 2005, respectively. He also received
his M.S. degree in Electrical Engineering from The Pennsylvania State University in August
2008. He continued his Ph.D study in the Computer Science and Engineering Department at The
Pennsylvania State University from September 2008. During his Ph.D study, he did internships
at Samsung Electronics in 2009 and at Lawrence Livermore National Laboratory in 2012. His
research interests lie in mesh optimization, mesh untangling, mesh generation, mesh warping