INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING

Int. J. Numer. Meth. Engng 2000; 48:1165}1185

Achieving "nite element mesh quality via optimization

of the Jacobian matrix norm and associated quantities.
Part II*A framework for volume mesh optimization and
the condition number of the Jacobian matrixS

Patrick M. Knupp*R
Parallel Computing Sciences Department, Sandia National ¸aboratories, M/S 0847, P.O. Box 5800, Albuquerque,
NM 87185-0847, ;.S.A.

SUMMARY
Three-dimensional unstructured tetrahedral and hexahedral "nite element mesh optimization is studied
from a theoretical perspective and by computer experiments to determine what objective functions are most
e!ective in attaining valid, high-quality meshes. The approach uses matrices and matrix norms to extend the
work in Part I to build suitable 3D objective functions. Because certain matrix norm identities which hold
for 2;2 matrices do not hold for 3;3 matrices, signi"cant di!erences arise between surface and volume
mesh optimization objective functions. It is shown, for example, that the equality in two dimensions of the
smoothness and condition number of the Jacobian matrix objective functions does not extend to three
dimensions and further, that the equality of the Oddy and condition number of the metric tensor objective
functions in two dimensions also fails to extend to three dimensions. Matrix norm identities are used to
systematically construct dimensionally homogeneous groups of objective functions. The concept of an ideal
minimizing matrix is introduced for both hexahedral and tetrahedral elements. Non-dimensional objective
functions having barriers are emphasized as the most logical choice for mesh optimization. The performance
of a number of objective functions in improving mesh quality was assessed on a suite of realistic test
problems, focusing particularly on all-hexahedral &whisker-weaved' meshes. Performance is investigated on
both structured and unstructured meshes and on both hexahedral and tetrahedral meshes. Although several
objective functions are competitive, the condition number objective function is particularly attractive.
The objective functions are closely related to mesh quality measures. To illustrate, it is shown that the
condition number metric can be viewed as a new tetrahedral element quality measure. Published in 2000
by John Wiley & Sons, Ltd.

KEY WORDS: unstructured grid generation; "nite element mesh; mesh optimization; smoothing; condition
number

*Correspondence to: Patrick M. Knupp, Parallel Computing Sciences Department, Sandia National Laboratories, M/S
0847, P.O. Box 5800, Albuquerque, NM 87185-0847, U.S.A.
R E-mail: pknupp@sandia.gov
S This article is a U.S. Government work and is in the public domain in the U.S.A.

Contract/grant sponsor: United States Department of Energy; contract/grant number: DE-ACO4-94AL85000

Received 16 April 1999

Published in 2000 by John Wiley & Sons, Ltd. Revised 25 October 1999
1166 P. M. KNUPP

1. INTRODUCTION

Optimization of three-dimensional unstructured meshes is studied from both a theoretical

perspective and via computer experiments to determine what objective functions are most
e!ective in achieving high-quality "nite element meshes. Unstructured mesh optimization is not
a new subject, although much of the e!ort has been restricted to two dimensions. In three
dimensions [1}5] and others, have tackled the problem with varying degrees of depth. Some have
concentrated upon using &smart' Laplacian smoothers in which a set of heuristics is developed to
adjust the basic algorithm to move towards better meshes, often using tetrahedral quality
measures as a guide (a good summary of tetrahedral measures is given in Reference [6]). A more
systematic approach is attempted in the present work.
Currently, production meshing relies heavily on Laplace smoothing for unstructured meshes.
Although frequently adequate for tetrahedral meshes, Laplace smoothing regularly fails to
produce even a validA mesh when applied to the all-hexahedral meshes arising from the &whisker
weaver' [7]. Moreover, even when the mesh is valid, the quality, as measured by aspect ratio,
skew, and scaled-Jacobian quality metrics, is often very poor. This lack of quality serves as
a major motivation for the present work. In structured meshing, considerable reliance is placed
upon solving the partial di!erential equations that result from minimizing the harmonic map
objective function. Although this is sometimes e!ective, it is much more so in two dimensions than
in three, probably due to the lack of a theoretical guarantee of valid meshes in three dimensions.
Thus, even in structured meshing there is room for improvement in 3D smoothing techniques.
Optimization of unstructured hexahedral meshes is largely unexplored territory since such
meshes have not been available until recently. The problem addressed is thus both novel and
di$cult because it requires automatic smoothing of three-dimensional structured or unstructured
hexahedral, tetrahedral, and mixed element meshes.
This paper extends work reported in Reference [8] which introduced the idea of using matrix
norms to generate mesh quality objective functions and in Part I of this series of papers on mesh
optimization [9]. In Part I the matrix norm idea was applied to unstructured surface triangular
and quadrilateral "nite element meshes. Surface objective functions were implemented and
compared on various test problems. The Smoothness objective function was discovered to be
equal to the condition number of the Jacobian matrix while the Oddy objective function was
equal to the condition number of the metric tensor. The Oddy objective function gave the best
over-all results.
Part I on surface mesh optimization is extended in this paper to volume mesh optimization.
A separate paper on volume optimization is needed because many relationships which hold in 2D
do not hold in 3D, and vice versa. This issue has not been critically examined before. The subject
also deserves attention because in practice it is much harder to achieve good mesh quality in 3D
than in 2D. This paper thus has two somewhat con#icting goals, namely, to systematically
explore the space of possible 3D mesh objective functions derived from matrix norms, and to
argue on theoretical grounds that there is one particular objective function, namely the condition
number (no longer equivalent to smoothness), that is best suited for optimizing meshes. Although
most readers will be interested primarily in the latter goal, the former goal is important because it
shows how objective functions are related to one another.

By valid it is meant that the elements are properly oriented with locally positive Jacobian determinant.
A

Published in 2000 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 48:1165}1185
 FINITE ELEMENT MESH QUALITY 1167

Section 2 de"nes the 3D Jacobian and metric tensor matrices in terms of element edges.
Section 3 shows that many well-known volume objective functions can be expressed in terms of
matrix norms. Section 4 uses 3D matrix norm identities to form dimensionally homogeneous
groups of objective functions and de"nes several new objective functions. Section 5 describes the
ideal element types in terms of matrices and introduces the idea of di!erentiation of a scalar
function with respect to a matrix in order to create objective functions which have the ideal
elements as stationary points. Because non-dimensional objective functions appear to be the most
e!ective in improving mesh quality, Section 6 focuses on identifying as many non-dimensional
objective functions as possible and discusses relationships between them and gives reasons why,
among all the non-dimensional objective functions, the Jacobian condition number is one of the
most attractive. Section 7 considers optimization of tetrahedral meshes and studies the Jacobian
condition number as a tetrahedral quality measure. Section 8 presents the results of numerical
optimization experiments performed with the CUBIT code. Section 9 is the summary and
conclusion.

2. BUILDING BLOCKS FOR NODALLY BASED VOLUME OBJECTIVE FUNCTIONS

For reasons discussed in Part I, attention is focused not on the mesh elements but rather on the
nodes of the mesh and the edges emanating from a given node. The mesh optimization is a series
of local optimization problems, one for each free node of the mesh.
Let M be the number of elements attached to a given interior node of the mesh whose spatial
position is to be &optimized' and let M be the set of integers m"0, 1, 2, . . . , M!1.B Assume for
the rest of this paper that m3M. Let the given node be associated with the vector x3R. It is
assumed that each of the M elements is attached to the interior node by 3 neighbouring vertices.#
Let the three neighbour nodes associated with the mth element be xKI 3R with k"1, 2, 3. To
achieve control over mesh quality, the local objective function f (x) needs to be based, not directly
on x, but rather on the important geometric entities associated with the node. The critical
quantities for the mth element are:
1. The 3 edge vectors, eKI:

eKI"xKI!x

2. The discrete Jacobian matrix:**


xK!x xK!x xK!x
JK"[eK, eK, eK]" yK!y yK!y yK!y
zK!z zK!z zK!z

B
According to Carey [10], the total number of edges or vertices attached to the interior node is  (M#4). M can be as

large as 40 in the all-hexahedral whisker-weaved meshes.
#
This assumption eliminates unusual element types such as pyramids and knives for the present.
** The name &Jacobian' refers to the underlying mapping from a master element to a physical element that exists for the
most commonly used element types. If the Jacobian of the map is evaluated at the corners of the element, one obtains
the Jacobian matrix de"ned here.

Published in 2000 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 48:1165}1185
1168 P. M. KNUPP

and the metric matrix:

GK"J2 K JK
3. Finally, de"ne for each of the m-elements:
gK"det GK
and the Jacobian determinant
j "det J "e ) (e ;e )
K K K K K
Then jK"gK and jK is six times the local volume of the mth tetrahedron enclosed by the
corresponding triad of edges. For shorthand, let " jK ""(g . It is important to distinguish between
K
j and (g because the former can be negative or positive, with negative j signifying inverted (also
K K K
called tangled, folded, or invalid) elements. A mesh with all-positive j is generally the minimum
K
quality criterion for a mesh. If j )0 for some element then the mesh is considered invalid.
K
Recent work in continuum variational methods of mesh generation has shown that a powerful
approach to the construction of objective functions involves the use of matrices and matrix norms
[8]. The Jacobian matrix is the fundamental quantity that describes all the "rst-order mesh
qualities (length, areas, and angles) of interest, therefore, it is appropriate to focus the building of
objective functions on the Jacobian matrix or the associated metric tensor. Many of the
well-known objective functions in structured mesh optimization can be cast in the form of norms
of matrices. One advantage of expressing objective functions in terms of matrix norms is that it is
relatively straightforward to generalize a surface mesh objective function to a volume mesh
objective function. New non-geometric interpretations of previously known objective functions
become available using matrix norms. Another advantage is that matrices permit easy introduc-
tion of weighted forms of objective functions for anisotropic mesh quality measures. For these
reasons, objective functions constructed from matrices are emphasized.
The matrices of interest must be converted to scalars to create objective functions. This can be
done using the trace, determinant, or matrix norms. There are several matrix norms but, for mesh
generation, the Frobenius norm has proven the most useful and most easily implemented. Let
A be a 3;3 matrix, then the Frobenius norm of A is "A" "(tr (A2 A)). The norm-squared is the
$
sum of the square of the elements in the matrix, which is why the Frobenius norm lends itself so
readily to the construction of objective functions for meshing. For the rest of this paper subscript
F is omitted since all the norms will be Frobenius norms. The following matrix properties prove
useful in analysing the objective functions used in the present approach. Let A3;3 be a real matrix
with determinant a non-zero and let adj A"aA\ be the adjoint matrix of A. Then
"A\"""adj A"/"a"
Note that, in contrast to the case for A2;2 , "A"O"adj A" for A3;3. The condition number of A is the
dimensionless quantity
i(A)""A" "A\""i(A\)
The following matrices are useful:


A if i"j
Diag A" GH
0 else

Published in 2000 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 48:1165}1185
 FINITE ELEMENT MESH QUALITY 1169

which is the matrix formed from the diagonal elements of A, and O!diag A"A!Diag A, which
is the matrix consisting of the o!-diagonal elements of A.
Let f be a scalar quality metric for the mth triad attached to the centre node, derived from the
K
building blocks in this section. For example,

f (x, y, z)"f (A (x, y, z))

K K

where the A are the Jacobian matrices constructed from the adjacent edges. Consider the vector
K
f"( f , f , . . . , f )3R+. The p-norm of f is
  +\


N
" f " " " f "N
N K
K

with p*1. Usually p"1, 2 will be used, but the l norm:


"f " "max +" f ",
 K
K
is also a useful norm because it enables optimization of worst-case mesh quality measures instead
of average measures.
The general form of the local mesh objective functions in this work is "f " . For p"2, the
N
minimizations are unconstrained, three-variable, and have smooth objective functions on some
domain. The necessary condition for a minimum to exist are (i) the gradient of the objective
function is zero at the minimum, and (ii) the Hessian is positive semi-de"nite at the minimum
[11]. Su$cient conditions require that the Hessian be positive de"nite.
With this form of the objective function chosen, it remains only to determine the choice of the
function f which takes A3;3 to a scalar. Objective functions will thus be posed in the form f"f (A),
where A is tacitly recognized as the Jacobian matrix.

3. KNOWN VOLUME OBJECTIVE FUNCTIONS EXPRESSED AS MATRIX NORMS

A number of previously proposed 3D objective functions can be expressed in terms of matrices

and matrix norms. Some of these are summarized in this section.

¸ength (length):
f (A)""A"
*
In the 2-norm, Length is the sum of the squares of the three edge lengths. Length is important
because it is the objective function for Laplacian smoothing. This objective function is widely
used due to its simplicity but it is recognized that it does not consistently produce valid meshes.
Length is also the sum of the eigenvalues (trace) of the metric tensor.
In the 4-norm, this objective function is closely related to, but not identical with, the &Length-
weighted Laplacian' smoother [12].

Published in 2000 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 48:1165}1185
1170 P. M. KNUPP

Norm of the metric tensor (length):

f (A)""A2A"
,+2
NMT has no simple geometric interpretation (it is the sum of the three edge lengths to the fourth
power plus the sum of the squares of the pairwise projected edge lengths). However, as shown in
Reference [8], because A2A is symmetric, the square of the norm of the metric tensor is the sum of
the squares of its eigenvalues. Minimization of this functional should make the eigenvalues of the
metric tensor equal (in a least-squares sense). This, in turn, means that the metric tensor will be
made proportional to a rotation matrix, giving near-equal aspect ratios and orthogonality. To
our knowledge, this objective function has never been tried in 3D, but it is an obvious extension of
the work in Reference [13].
O+-Diagonal (length): The O!-Diagonal objective function:

f (A)"" O! diag A2A"

-" 
attempts to equidistribute the face angles about a node. Although geometrically appealing,
O!-Diagonal su!ers the great drawback that it is non-convex in general, i.e. the minimal mesh
does not always exist. To make practical use of O!-Diagonal requires combining it with other
objective functions to get convexity. For this paper O!-Diagonal is combined with 2-Length since
they are both have dimensions of (Length). This objective function has been proposed by some
for structured 3D mesh generation but has not been used extensively.
<olume (length): The Volume objective function makes locally equal element volumes:

f (A)"a
4
In the continuum the volume objective function is non-elliptic, so non-smoothed grids result. This
property carries over to the discrete optimization function for volume. Volume has been
proposed in References [1, 14], and others.
Adjoint of the Jacobian matrix (length):

f (A)""adj A""a"A\"

 

The geometric interpretation of this objective function can be found by noting that

" adj J """e ;e "#"e ;e "#"e ;e "

K K K K K K K
i.e. "adj J " is twice the sum of the squares of the areas of the three triangular faces lying between
K
the three edges. The adjoint is also the second-invariant of the characteristic polynomial of G .
K
This objective function was used in a group of objective functions in Reference [15].
Smoothness (length): The Smoothness objective function derives by analogy to variational
structured mesh generation [16]:

f (A)"a "A\"
1
Published in 2000 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 48:1165}1185
 FINITE ELEMENT MESH QUALITY 1171

When writing the Smoothness objective function for surface meshes in Part I in terms of matrix
norms, it was found that f (A)"f (A)"i(A) because the norm of the adjoint was equal to the
1 '1
norm of the matrix. This does not hold in three dimensions, and therefore, there are three
objective functions ( f , f and i) to consider, whereas in 2D there was only one.
1 '1
Inverse smoothness (length)\: This objective function is an ad hoc generalization that appears
in Reference [17] and others:

f ""A"/a
'1
Note that f (A)"f (A\).
'1 1
¹he Oddy metric (length): In Part I on surface meshes, the condition number of the metric
tensor was shown to be equivalent to an objective function based on the Oddy metric [18]. This
equivalence does not hold in three dimensions, thus there is another objective function to
consider

f (A)"a\ +"A2 A"! "A",

- 
This objective function is in"nite for invalid meshes and is zero for the identity or rotation
matrices.
These previously known objective functions are included in the numerical tests given later in
this paper, but "rst several new objective functions suggested by the matrix norm approach are
introduced.

4. VOLUME OBJECTIVE FUNCTIONS FROM DIMENSIONALLY

HOMOGENEOUS GROUPS

The objective functions of the previous section do not exhaust the myraid possibilities that arise
by approaching mesh optimization via matrix norms. This section seeks to answer two questions
(1) are there some other potentially useful objective functions that have so far been over-looked,
and (2) is there a way to group objective functions in some rational way? Brackbill suggested
combining continuum objective functions into a group to control smoothness, area, or ortho-
gonality [16] but found that it was di$cult to select the constants in these combinations due to
dimensional inhomogeneity of the objective functions. As shown in Part I, this limitation can be
partly overcome by combining objective functions having the same dimension.RR This idea is
extended further here by giving a systematic way to construct dimensionally homogeneous
groups via certain identities which hold for arbitrary 3;3 matrices (these identities do not hold
for 2;2 matrices). There are two identities that appear rather fundamental, and from which many
others can be constructed. The "rst has dimensions (length):

"A"#2 tr (adj A),(tr A)# "A!A2"



RR By dimension reference is made to the idea that if the matrix A has a particular set of units (say length), then the units of
other related matrices and norms are determined.

Published in 2000 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 48:1165}1185
1172 P. M. KNUPP

and can be derived from a trivial matrix identity. The second has dimensions (length):

(tr A)#3a,tr(A)#3 (tr A) tr (adj A)

and can be derived from the fact that every matrix satis"es its own characteristic polynomial.
These identities are not particularly useful in themselves because the objective functions they
suggest (e.g. tr A) have geometrical interpretations that do not correspond to a useful element
quality measure. Replacing A with A2 A in these identities, one can obtain two other identities
having dimensions of (length) and (length) which contain norms upon which useful objective
functions can be based:

"A2A"#2a"A\","A"
3"A" "A2A"!"A"#6a,2 "AA2A"

The identities given in this section suggest dimensionally homogeneous groups of objective
functions. Let c , c , c , and c be arbitrary real constants.
   
Group six (length):

f (A)"3c "A""A2A"!c "A"#6c a

   
Note that if c "c "c " then one obtains the other objective function, "AA2A", in the
   
(length) identity, and hence left out of this group.
Group four (length):

f (A)"c "A2A"!c "A"#c a

   
Group two (length):

f (A)"c "A"#2 c tr (adj A)!c (tr A)#c a

    
A method for choosing the constants in these groups is given in the next section.

5. THE IDEAL ELEMENT*A GEOMETRIC VIEW

Every matrix A with column vectors a , i"1, 2, 3, has the following factorization:
G
A"DQ2

where

A"[a , a , a ]
  
D"diag (d , d , d )
  
Q"[a /d , a /d , a /d ]
     
Published in 2000 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 48:1165}1185
 FINITE ELEMENT MESH QUALITY 1173

and d ""a ". The normalized column vectors of A are contained in Q and the lengths in D. Thus
G G
"Q""3 because the column vectors comprising Q are of unit length. The quantity det (Q) is
referred to in the CUBIT code [12] as the scaled Jacobian quality measure. It varies from minus
one to plus one. Positive scaled Jacobian is considered the minimal quality permitted for a mesh,
while negative values of the scaled Jacobian signify that invalid elements exist.
The matrix A determines the geometric quality of the corresponding element. In the absense of
anisotropy, each edge of the element should have equal length. This requirement corresponds to
making the diagonal matrix D have equal positive entries, i.e. D "* for some *'0. The proper
GG
choice for Q depends on the element type. For a hexagonal element, the edges of the element
should be orthonormal, hence we require Q2"R for some rotation matrix R where R2R"I and
det (R)"#1 to maintain proper orientation.
Thus, for an isotropic hexahedral element the ideal matrix has the form

A"*R

where R is an orthogonal matrix with det(R)"1 and *'0. If A has this form then
(1) the column vectors of A are orthogonal,
(2) the lengths of the column vectors are equal, and
(3) the volume of the region spanned by the 3 vectors is positive, and
(4) the corner of the element de"ned by A matches the corner of a cube.
Although the form of the ideal A has been determined, the scalar * and the matrix R have not
been speci"ed. These control the element size and orientation, respectively. With isotropic
meshes, it is not desirable to specify the size and orientation of each element of the mesh, rather,
the optimization procedure should do this. To eliminate these two quantities, note that the
following relationship holds for the ideal element:

A2A"*I

The determinant of the left-hand side of this expression is a and therefore

*"a

The ideal matrix is then

A"aR

Because the Frobenius matrix norm is invariant to rotations, i.e.

"MR"""RM"""M"

for any matrix M, the rotation matrix in the ideal will not need to be speci"ed.
The ideal matrix for tetrahedral elements will be discussed in Section 7.
Not all of the objective functions described so far are minimized by the ideal A. Although it is
neither necessary nor su$cient to require that an objective function be minimized by the ideal,
doing so is a rational approach to designing objective functions. Consider the function f (A)

Published in 2000 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 48:1165}1185
1174 P. M. KNUPP

taking 3;3 matrices to a scalar. Di!erentiation of f with respect to A can be de"ned to be the
following 3;3 matrix:SS

[* f/*A] "*f/*A
GH GH
Derivatives for relevant functions f (A) are given in Appendix A. The matrix function has
a stationary point when * f/*A"0, with 0 denoting the 3;3 zero matrix.
Applying the di!erentiation formula to the Length objective function we "nd that A"0 is the
minimizer. Of course, A"0 is not attained in practice because the admissible set of A1s are
required to satisfy the boundary data which holds the surface mesh nodes "xed. It is reasonable to
ask if the Length objective function can be modi"ed in some way to give the ideal as the
minimizer. Length is part of the Group 2 objective function of the previous section, so let us ask if
one can choose the constants c to give the desired minimizer. This is indeed possible, and can also
G
be done for the other Group objective functions:

f (A)""A"!3a
*
f (A)""A2A"!3a
*
f (A)""AA2A"!3a
*
The smoothness and inverse smoothness objective functions must also be modi"ed slightly to
have the ideal as the minimizer:

f (A)"a "A\"
1
f (A)"a\ "A"
'1
Empirical results con"rm that, in general, mesh quality is improved with these modi"ed objective
functions, when compared to their original forms.

6. NON-DIMENSIONAL OBJECTIVE FUNCTIONS

The non-dimensional objective functions are interesting because they are scale-independent and
are the lowest-order objective functions having barriers.AA Furthermore, they can be symmetrized
so that f (A)"f (A\). This means that both the local map from the logical to the physical region
and the local inverse map from the physical to the logical region are well conditioned. Begin by
de"ning the following three non-dimensional objective functions:

f (A)"a\ "A"

f (A)"a\ "A2A"

f (A)"a\ "AA2A"


SSOf course, su$cient smoothness must be assumed.

AA
By a barrier it is meant a set of matrices in the domain for which the objective function f (A) becomes in"nite.

Published in 2000 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 48:1165}1185
 FINITE ELEMENT MESH QUALITY 1175

and their three counterparts:

fK (A)"f (A\2)"a "A\"

 
fK (A)"f (A\2)"a " (A2A)\"
 
fK (A)"f (A\2)"a " (AA2A)\"
 
The functions f and fK are recognized as the dimensionless versions of inverse smoothness and
 
smoothness, respectively.
These objective functions serve as building blocks for some of the other known non-dimen-
sional objective functions, suggest new non-dimensional functions, and help create some interest-
ing non-dimensional identities involving mesh objective functions. For example, one can build
symmetric objective functions by combining f with f) . Three condition numbers are immediate:
G G
i(A)"f fK
 
i(A2A)"f fK
 
i(AA2A)"f fK
 
The building blocks are related to one another through the following identities:

f "f !2 fK
  
fK "fK !2 f
  
f "f !3i(A)#3
 
fK "fK !3i(A)#3
 
The condition number objective functions are related through the following identities:

i(A)"3#f f !f "3#fK fK !fK

     
i(A2A)"i(A)#4i(A)!2 f !2fK 
 
i(AA2A) can be similarly expressed in terms of lower-order objective functions.
Let f be the Oddy objective function, then
-
f (A)"(2/3) f !2fK
-  
fK (A)"f (A\)"( ) fK !2 f
-    
( ) f fK "( ) i(A)#i(A)!( ) ( f #fK  )
 - -    
The last objective function will be referred to as the symmetric Oddy objective function. Let
w "(n!3)/3 with n an integer and de"ne the following symmetric objective functions:
L
S "(w f !2 fK ) (w fK !2 f )
L L   L  
Published in 2000 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 48:1165}1185
1176 P. M. KNUPP

Then for n"0, 1, 2, 3 one obtains the i(A2A), Symmetric Oddy, S *a new symmetric objective

function, and i(A), respectively, showing that these objective functions are all part of the same
family. Another interesting family can be obtained from the three objective functions f , f , f ,
* * *
and their respective counterparts to give three more symmetric dimensionless objective functions:

f fK "( f !3) ( fK !3)"i(A)!3( f #fK )#9

* *    
f fK "( f !3) ( fK !3)
* *  
f fK "( f !3) ( fK !3)
* *  
Proposition. The ideal matrix is a stationary point of i(A), i(A2A), and the other non-
dimensional objective functions given in this section.
Proof. All of the dimensionless objective functions can be expressed as combinations of f and

fK and the ideal is a stationary point of the latter two functions. )

Observation: Let A3;3 with a'0. Let f (A)'0 be a non-dimensional objective function such
that lim f (A) is unbounded but f (A) is bounded for A in some set of matrices having positive
?
determinant. Then the volume a associated with A is bounded below by a positive constant.
The non-dimensional objective functions described in this section appear quite attractive,
having a number of favourable theoretical properties. The objective function i(A) in particular
enjoys all of these nice properties, including symmetry. A geometric interpretation of this
objective function can be found by nothing that

i(A)""A" "adj A"/"a"

i.e. the condition number is the square root of the product of the sum of the squares of the
edge-lengths and the sum of the squares of the adjacent face areas, divided by six times the volume
of the tetrahedron de"ned by the edges. This is similar to, but not identical with, the objective
function given in Reference [19, p. 11], for a tetrahedral quality measure.
Minimizing the condition number of A would seem to be a good idea because this maximizes
the distance to the set of singular matrices [20, p. 26]. In view of this, i(A) is the "rst objective
function we know of that directly states that invalid meshes are to be avoided.BB
Because condition number is the simplest symmetric non-dimensional objective function, close
attention was paid to its performance in the empirical tests given in Section 8.

7. OPTIMIZATION OF TETRAHEDRAL MESHES

Although our main focus is on hexahedral mesh optimization, the theory presented applies
particularly well to tetrahedral elements and suggests the ideal of introducing weight matrices to
drive elements toward the proper ideal. The ideal isotropic tetrahedral element is the unit
equilateral tetrahedron. For isotropic tetrahedral elements the ideal matrix A again has D"*I.

BB
We speculate that all of the non-dimensional objective functions having barriers measure distance from the set of
singular matrices.

Published in 2000 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 48:1165}1185
 FINITE ELEMENT MESH QUALITY 1177

However, the columns of Q should not be orthonormal. Instead let us require Q2"R= where
R is a rotation matrix and


1 1/2 1/2
=" 0 (3/2 (3/6 (1)
0 0 (2/(3

This matrix is derived from examining the edge vectors of the ideal tetrahedron. Since
det (=)"(2/2, the ideal must have the form

A"((2a)R=

A2A"((2a)=2=

The objective functions given in the previous sections should be weighted for the tetrahedral
case in order that the tetrahedral ideal is a stationary point. If f (A) is minimized by A"A then
 
f (A=\) is minimized by A"A =. Thus, for example, the modi"ed length objective function
 
for tetrahedra reads

f (A)""A=\"!3 ((2a)
*
The matrix A=\ forms the linear transformation between the ideal tetrahedral element and the
element de"ned by the matrix A.
In this section, however, attention is mainly focused on the Jacobian condition number
objective function as it applies to tetrahedral elements. The weighted condition number objective
function for tetrahedral is

i(A=\)""A=\" "=A\"

Proposition. Let A , i"0, 1, 2, 3 be the matrix associated with node i of a tetrahedral element.
G
Then f (A =\)"C, where C is a constant independent of i and f is any of the objective functions
G
discussed in this paper.
Proof. A sketch of a proof goes like this: one can "rst show that there exists matrices M such
G
that A "A M . One can also show that there exist rotation matrices such that M =\
G  G G
"=\R . Then
G
"A =\"""A M =\"""A =\R """A =\"
G  G  G 
A similar result holds for " (A =\)\ " and for det (A =\). )
G G
Thus, for each objective function a single number can be associated with any given tetrahedron.
Adopting the de"nition of a tetrahedral-shape measure given in Reference [6], any of the
dimensionless objective functions are valid shape measures provided (i) 1/ f is used, and (ii) there is
a unique, global maximum. The latter requirement is not proven here but is likely to hold for

Published in 2000 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 48:1165}1185
1178 P. M. KNUPP

Table I. Sensitivity of condition number quality measures to tests A, B, C, and D.

Apex dist (A) i Angle (B) i LM (C) i Angle (D) i

0.25 2.03 0 1.00 1.00 1.00 90 1.00

0.50 1.22 15 1.02 0.75 1.04 75 1.02
1.00 1.00 30 1.11 0.50 1.22 60 1.11
2.00 1.22 45 1.29 0.30 1.74 45 1.29
3.00 1.61 60 1.73 0.20 2.47 30 1.73
4.00 2.03 75 3.21 0.10 4.77 15 3.21

most of the dimensionless objective functions. If the weight = were not included in the argument
of the objective function, the function would no longer be invariant under a change of node.
i(A) is somewhat similar to the Q tetrahedral measure reported in Reference [19, p. 11]. But
)
i(A=\) is neither identical to Q nor can it be expressed in terms of any combination of the
)
quality metrics given in Reference [21]. Thus i(A=\) is a new tetrahedral element quality
measure. To emphasize this point, tests A}D from Reference [21] were performed, to show how
the condition number quality measure varies with distortions of a tetrahedron## (see Table I).
The key feature to note in the table is that the condition number quality measure behaves
similarly to other tetrahedral quality measures.*** We suspect, but have not proved, that
i(A=\) can be shown to be an equivalent tetrahedral quality measure to Q in the sense de"ned
)
in Reference [22]. At the very least, then, the situation for using condition number as a tetrahedral
quality measure is paraphrased from Reference [20]: &since it is impossible to "ll an arbitrary
volume with equilateral tetrahedral, equivalent quality measures will perform similarly, but with
somewhat di!erent results'. The &somewhat di!erent' results will be seen in section 10 where
numerical experiments are performed to determine which objective function gives the best overall
mesh quality.
As a "nal observation, note that most of the tetrahedral quality measures given in References
[21, 6] cannot be expressed in terms of matrix norms. Three exceptions are the mean ratio g,
which is roughly (a\ "A")\, c, approximately "A"/a, and the i measure, which is roughly
(a"A\")\.

8. EMPIRICAL TESTS

The theory presented has identi"ed a number of promising objective functions, but is unable to
determine, for example, whether i(A) or i(A2A) will produce superior mesh quality. To explore
this question, computer experiments on realistic problems are needed. Many of the objective
functions considered in this paper were implemented within the CUBIT code. All of the objective
functions can be evaluated in terms of just three matrix functions, the determinant, the norm, and
the norm of the adjoint. Evaluating the gradient of an objective function entails an indirect
approach because if one attempts to write out the complete expression an unwieldy number of

##
A normalization factor of 1/3 was included in the de"nition so that the ideal element gave i"1.
***Note that i(A=\) is symmetric about the apex distance and has values close to c.

Published in 2000 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 48:1165}1185
 FINITE ELEMENT MESH QUALITY 1179

terms results. For example, "A" contains 9 terms which must be di!erentiated, "adj A" has 27,
while a has 6. When di!erentiated, these objects contain 9, 54, and 18 terms, respectively. i(A)
contains upwards of 324 non-di!erentiated terms in the numerator alone, if everything is
expanded fully. The di$culty may be avoided by making use of the idea of di!erentiating a scalar
function of a matrix introducted in the previous section. Let

F(x, y, z)" f (J (x, y, z))

K
K
be a mesh objective function. Then the chain rule shows that

*F/*x" tr +(*f / *J )2 (*J /*x),

K K
K
and, because *J /*x is readily computed
K

F"! (*f / *J ) u
K
K
where u2"[1, 1, 1]. A su$cient condition for a stationary point is thus *f/*J "0 for all m. This
K
is one reason why objective functions should be zero at the ideal element. Of course, the gradient
method is only applicable for 1(p(R. For the non-di!erentiable cases, a non-gradient method
is employed.
Several sets of optimization tests were performed focusing on di!erent geometries, element
types, and mesh connectivities. These tests included (1) all-hexahedral unstructured, whisker-
weaved meshes [7] (three geometries, including these shown in Figures 1 and 2), (2) all-hexahedral
semi-structured swept meshes (three geometries, including that in Figure 3), (3) all-tetrahedral
meshes (three geometries, including those in Figures 1 and 2) and (4) a tetrahedral mesh using an
1-in"nity norm (the same geometry as in Figure 1). The investigation began with about twenty
objective functions which were tested on the &knee' geometry. The results are shown in Table II.

Figure 1. Initial WW hexahedral mesh on sphere Figure 2. Initial WW hexahedral mesh on knee
geometry. geometry.

Published in 2000 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 48:1165}1185
1180 P. M. KNUPP

Figure 3. Initial swept hexahedral mesh on curved geometry.

Table II. Average quality metrics for hexahedral mesh (knee geometry).

Method Aspect Skew Oddy Scaled J Valid?

Whisker weave 1.496 0.4594 7119 0.6225 No

Laplacian 1.493 0.4597 7116 0.6229 No
Equipotential 1.976 0.5386 12 900 0.5394 No
Length, "A" 1.520 0.4106 4270 0.6680 No
"A"!3a 1.486 0.4082 4267 0.6757 No
2-Length, "A" 1.482 0.4357 5690 0.6530 No
NMT "A2A" 1.513 0.4468 8532 0.6334 No
Volume, a 1.608 0.5019 4271 0.5983 No
" O!Diag A2A" 1.480 0.4648 4273 0.6278 No
"adj A" 1.453 0.4172 1430 0.6729 No
"AA2A"!3a 1.834 0.4860 3708 0.5479 No
"A2A"!3a 1.688 0.4722 55 460 0.5427 No
Smoothness, a "A\" 1.556 0.4127 25.7 0.6669 Yes
a "A\" 1.472 0.3872 9.396 0.6861 Yes
"A\" 1.446 0.3994 6.755 0.6911 Yes
"A"/a 1.419 0.3905 5.836 0.6962 Yes
a\ "A" 1.441 0.3888 2847 0.6911 Yes
a\ "A2A" 1.551 0.4153 13.41 0.6592 Yes
Oddy (A) 1.448 0.4109 1426 0.6761 Yes
Oddy (A\) 1.669 0.4371 23.2 0.6328 Yes
i(A) 1.453 0.3929 8.760 0.6889 Yes
i (A2A) 1.465 0.4096 8.115 0.6778 Yes

The list was then reduced to include only those objective functions having barriers since this is
a critical requirement of any robust optimizer (see Table III). Among those having barriers there
were several which performed particularly well, so the list was further reduced. To compare
results between objective functions several quality measures were used. For hexahedral elements
these were: aspect ratio [23], skew [23], Oddy's metric [18], condition number, and minimum

Published in 2000 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 48:1165}1185
 FINITE ELEMENT MESH QUALITY 1181

Table III. Average quality metrics for == hexahedral mesh (sphere geometry)*barriers only.

Method Aspect Skew Oddy i (A) Scaled J Valid? Rel CPU

Whisker weave 1.533 0.3452 1867 622 0.6877 No *

Laplacian 1.528 0.3438 1867 622 0.6888 No *
Equipotential 2.113 0.5007 5600 1865 0.5482 No *
"A"/a 1.434 0.2955 3.559 1.424 0.7441 Yes 0.86
i (A) 1.467 0.2782 4.133 1.447 0.7529 Yes 1.00
a "A\" 1.480 0.2744 4.733 1.469 0.7549 Yes 1.51
i(A2A) 1.470 0.2829 4.178 1.451 0.7522 Yes 3.18
a\ "A2A" 1.514 0.2822 5.374 1.501 0.7513 Yes 7.47
Smoothness, a " A\" 1.523 0.2735 5.003 1.492 0.7533 Yes 2.26
Oddy (A\) 1.548 0.2877 6.084 1.532 0.7476 Yes 12.8
Oddy (A) 1.466 0.3072 3.579 1.457 0.7319 Yes 3.02
a\ "A" 1.467 0.2854 3.770 1.441 0.7471 Yes 0.60
"A\" 1.438 0.3068 3.801 1.429 0.7418 Yes 2.27

scaled Jacobian. For tetrahedral elements, these were: gamma [21], condition number, and
minimum scaled Jacobian. Except for the minimum scaled the Jacobian, smaller numbers in the
tables indicate superior mesh quality. Results were also compared against two volume smoothers
already in CUBIT namely, Laplacian smoothing and Equipotential [24].
Although only a very limited set of tests could be performed, the following observations are
made:
(1) As expected, objective functions having barriers maintain valid meshes whereas those
without barriers did not consistently do so.
(2) All of the objective functions having barriers gave meshes notably superior to those
smoothed with CUBIT's original Laplacian and equipotential algorithms. This was espe-
cially true on the whisker-weaved meshes.
(3) The objective functions that were modi"ed to make the ideal a stationary point (like f
1
and f ) generally performed better than the original objective functions, but not always,
'1
(4) The weighted objective functions, f (A=\), consistenty improved tetrahedral mesh quality
over the unweighted objective functions, f (A), through not dramatically,
(5) It is currently rare to achieve excellent quality by optimizing the whisker-weaved meshes,
although the quality is much improved with the non-dimensional objective functions
compared to the unoptimized mesh. For example, the minimum scaled Jacobian was
typically #0.003 in the unoptimized mesh and #0.300 in the optimized mesh. Hexahedral
element swapping techniques to change mesh connectivity may be considered in the future
to further improve quality.
(6) In terms of relative e$ciency, it was found that objective functions which use the metric
tensor were noticeably slower than those which use the Jacobian matrix (see Table III). The
fastest objective function based on the metric tensor was Oddy (A).
(7) On Swept meshes, Oddy performed very well, whereas on whisker-weaved meshes it lagged
quite a few others (see Table IV).
(8) Somewhat surprisingly, quality achieved with the smoothness objective function was
noticeably less than a half dozen other objective functions also having barriers. Smoothness
was also the slowest of the objective functions based on the Jacobian matrix.

Published in 2000 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 48:1165}1185
1182 P. M. KNUPP

Table IV. Average quality metrics for swept hexahedral mesh (curved geometry).

Method Aspect Skew Oddy i (A) Scaled J Valid?

Sweep 7.68 0.130 524 4.25 0.930 Yes

Laplacian 3.89 0.600 3e#05 9e#04 0.228 No
Equipotential 6.21 0.314 1023 288 0.823 No
"A"/a 6.00 0.246 168 3.22 0.880 Yes
i (A) 6.38 0.167 230 3.42 0.919 Yes
a "A\" 6.75 0.148 791 3.88 0.896 Yes
i(A2A) 6.31 0.224 240 3.50 0.891 Yes
a\ "A2A" 7.51 0.167 468 4.14 0.918 Yes
Smoothness, a "A\" 7.68 0.140 508 4.21 0.931 Yes
Oddy (A\) 7.56 0.168 491 4.19 0.916 Yes
Oddy (A) 5.71 0.363 148 3.26 0.804 Yes
a\ "A" 5.93 0.264 162 3.21 0.870 Yes
"A\" 6.36 0.177 231 3.43 0.913 Yes

Table V. Worst-case quality metrics for tetrahedral mesh

(sphere geometry)*l norm.

Method b c i (A) Scaled J

Tetrahedral mesh 4.232 4.764 3.579 0.1377

Laplacian 5.750 376 297 0.0024
"A"/a 3.965 4.345 3.790 0.1613
i (A) 4.263 5.155 2.989 0.1004
a "A\" 4.757 5.855 3.033 0.0883
i (A2A) 4.262 5.146 2.989 0.1013
Smoothness, a"A\" 4.405 4.777 3.044 0.1136
Oddy (A\) 6.124 6.406 3.156 0.0779
Oddy (A) 3.864 4.153 3.607 0.1740
a\ "A" 3.866 3.986 3.300 0.1574

(9) Condition number of the Jacobian matrix, condition number of the metric tensor, inverse
smoothness, and Oddy (all in their weighted forms) appeared to give the best quality of all
the objective functions tried on tetrahedral meshes (see Table V). An l norm was used in

this test. These objective functions also performed well on hexahedral meshes.

9. SUMMARY AND CONCLUSIONS

This paper applied the matrix norm idea to design objective functions for "nite element volume
mesh optimization. Traditional volume objective functions such as length, volume, and smooth-
ness can be expressed in terms of matrix norms. In two dimensions, smoothness, inverse
smoothness, and condition number of the jacobian matrix are identical objective functions, while

Published in 2000 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 48:1165}1185
 FINITE ELEMENT MESH QUALITY 1183

in three dimensions they become three distinct objective functions because the norm of the
adjoint matrix is no longer equal to the norm of the matrix. Similarly, in two dimensions oddy
and the condition number of the metric tensor give the same objective function while in three
dimensions there are distinct.
Two fundamental matrix norm identities were presented having dimensions of length and
length. These show the relationship between a number of objective functions and how dimen-
sionally homogeneous groups of objective functions can be formed. The arbitrary constants in
these groups can be determined by introducing the matrix corresponding to the ideal mesh
element. The constants were chosen so that the ideal matrix is a stationary point of any given
objective function. For example, the power of a in the smoothness and inverse smoothness
objective functions was adjusted to make the ideal a stationary point.
Non-dimensional objective functions with barriers seem to be the most logical choice for
mesh objective functions because they are scale invariant and avoid inverted elements. Several
non-dimensional matrix norm identities were derived in terms of smoothness and inverse
smoothness to show how other non-dimensional objective functions were related to one
another. For example, the condition number of the metric tensor can be expressed in terms of
the condition number of the Jacobian matrix, smoothness, and inverse smoothness. Given the
wealth of potential non-dimensional objective functions the condition number of the Jacobian
matrix was preferred because it is the simplest of the symmetric non-dimensional objective
functions and because it is a measure of the distance of a given matrix to the set of singular
matrices.
The ideal matrix corresponding to tetrahedral elements was related to a matrix = derived from
considering the equilateral tetrahedral element. Objective functions for tetrahedral elements
expressed in terms of the matrix A=\ have the ideal matrix as a stationary point. For each such
objective function it was shown that a single number, independent of the node at which it is
computed, could be assigned to a given tetrahedral element. To illustrate, the condition number
of the Jacobian for tetrahedral elements was subjected to the distortions given in Reference [21],
the numerical results being very similar to the other tetrahedral quality measures given in that
reference.
Finally, many of the objective functions were implemented in the CUBIT code and empirically
tested using unstructured all-hexahedral whisker-weaved meshes, semi-structured swept meshes,
and an all-tetrahedral mesh. Ultimately, the list of objective functions could not be reduced to just
one objective function as several were competitive. In terms of dimension, the non-dimensional
objective functions performed best. Of these, Oddy, condition number of the metric tensor, and
others based on the metric tensor gave good mesh quality but, due to their relative slowness, are
less attractive. Probably the two best overall objective functions were condition number of the
Jacobian matrix and inverse smoothness (with a).
Although this study has clari"ed a number of issues, there remain several others that must be
addressed before the matrix norm approach can realize its full potential. E$ciency issues related
to the optimization algorithm need further consideration. In addition, barrier-based objective
functions require that one optimize beginning with a valid mesh. If this is not done, the resulting
mesh will likely contain inverted elements. Part III will consider ways to create valid meshes from
invalid meshes so that barrier-based objective functions can further improve mesh quality.
Because the objective functions in this paper are based only on interior nodes, the current
approach does not guarantee good mesh quality on the boundary of the domain. This also will be
addressed in Part III.

Published in 2000 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 48:1165}1185
1184 P. M. KNUPP

APPENDIX: DERIVATIVES OF OBJECTIVE FUNCTIONS

*
"A""2A
*A
*
trace(A)"I
*A
*
trace (AL)"n(AL\)2
*A
*aL
"naLA\2
*A
*
"A""4 "A" A
*A
*
"A2A""4AA2A
*A
*
"O!Diag A2A""4A O!Diag (A2A)
*A
*
"adj A""2A +"A" I!A2A,
*A
*
"A\""2a\A +"A" I!A2A!a "A\" (A2A)\,
*A
*
trace (adj A)"(trace A) I!A2
*A
*
"AA2A""6AA2AA2A
*A
*
"A=2""2A=2=
*A
*
"=A""2=2=A
*A

ACKNOWLEDGEMENTS

This work was made possible through the MICS program o$ce directed by Fred Howes and managed at
Sandia Laboratories by David Womble. We would also like to thank department manager Rob Leland for
his support and CUBIT team members Scott Mitchell and Nate Folwell for providing many whisker-
weaved mesh examples.
This work was funded by the Department of Energy's Mathematics, Information and Computational
Sciences Program (SC-31) and was performed at Sandia National Laboratories. Sandia is a multiprogram
laboratory operated by Sandia Corporation, A Lockheed Martin Company, for the United States Depart-
ment of Energy under contract DE-ACO4-94AL85000.

Published in 2000 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 48:1165}1185
 FINITE ELEMENT MESH QUALITY 1185

REFERENCES
1. Chen CL, Szema KY, Chakravarthy SR. Optimization of unstructured grid. 33rd Aerospace Sciences Meeting, AIAA
95-0217, Reno NV, January 1995.
2. Freitag L, Jones M, Plassman P. An e$cient parallel algorithm for mesh smoothing. In Proceedings of the Fourth
International Meshing Round¹able, Sandia National Laboratories, 1995; 47}58.
3. Parthasarathy VN, Kodiyalam S. A constrained optimization approach to "nite element mesh smoothing.
Finite Elements in Analysis and Design 1991; 9:309}330.
4. Canann S, Tristano J, Staten M. An approach to combined Laplacian and optimization-based smoothing for
triangular, quadrilateral, and quad-dominant meshes. Proceedings of the 7th International Meshing Round ¹able,
Detroit, 1998; 479}494.
5. Zavattieri P. Optimization strategies in unstructured mesh generation. International Journal for Numerical Methods in
Engineering 1996; 39:2055}2071.
6. Dompierre J, Labbe P, Guibault F, Camerero R. Proposal of benchmarks for 3D unstructured tetrahedral mesh
optimization. Proceedings of the 7th International Meshing Round¹able 198, Dearborn MI, October 26}28, 1998;
459}478.
7. Folwell NT, Mitchell SA. Reliable whisker weaving via curve contraction. Proceedings of the 7th International
Meshing Round¹able, Dearborn, MI, 1998; 365}378.
8. Knupp P, Robidoux N. A framework for variational grid generation: conditioning the Jacobian matrix with matrix
norms. SIAM J. Sci. Comput. 2000, to appear.
9. Knupp P. Achieving "nite element mesh quality via optimization of the Jacobian matrix norm and associated
quantities, Part I*A framework for surface mesh optimization. International Journal for Numerical Methods in
Engineering 2000; 48:401}420.
10. Carey G. Computational Grids. Taylor & Francis: London, 1997.
11. Gill P, Murray W, Wright M. Practical Optimization. Academic Press: New York, 1981.
12. CUBIT Mesh Generation Environment. SAND94-1100, Sandia National Laboratories, Albuquerque, NM 87185.
13. Liao G. Variational approach to grid generation. Numerical PDE1s 1992; 8:143}147.
14. Castillo J. A discrete variational method. SIAM Journal on Scienti,c and Statistical Computing 1991; 12(2):454}468.
15. Jacquotte OP, Cabello J. A variational method for the optimization and adaptation of grids in computational #uid
dynamics. In Numerical Grid Generation in Computational Fluid Mechanics 188, Sengupta S et al. (eds). PineRidge
Press: Swansea UK, 1988; 405}413.
16. Brackbill J, Saltzman J. Adaptive zoning for singular problems in two dimensions. Journal of Computational Physics
1982; 46:342}368.
17. Tinoco JG, Barrera P. Smooth and convex grid generation over general plane regions. Mathematics and Computers in
Simulation. Elsevier: Amsterdam, 1998.
18. Oddy A, Goldak J, McDill M, Bibby M. A distortion metric for isoparametric "nite elements. ¹ransactions of CSME,
No. 38-CSME-32, Accession No. 2161, 1988.
19. George PL, Borouchaki H. Delaunay ¹riangulation and Meshing. Hermes: Paris, 1998.
20. Golub G, Van Loan C. Matrix Computations. Johns Hopkins University Press: Baltimore, MD, 1983.
21. Parthasarathy VN, Graichen CM, Hathaway AF. A comparison of tetrahedron quality measures. Finite Elements in
Analysis and Design 1993; 15:255}261.
22. Liu A, Joe B. Relationship between tetrahedron quality measures. Bit 1994; 34:268}287.
23. Robinson J. CRE method of element testing and the Jacobian shape parameters. Engineering Computations 1987;
4:113}118.
24. Tipton RE. Grid optimization by equipotential relaxation. Unpublished, Lawrence Livermore National Laboratory,
1990.

Published in 2000 by John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 48:1165}1185