Engineering with Computers 8, 121-137 (1992) Engineering

C61 puters
9 Springer-Verlag New York Inc. 1992

An Automatic Coarse and Fine Surface Mesh Generation Scheme Based on

Medial Axis Transform: Part I Algorithms

H. Nebi Giirsoy
Intergraph Corporation, AnalysisApplications-Development,Huntsville, AL, USA

Nicholas M. Patrikalakis
Massachusetts Institute of Technology,Department of Ocean Engineering,Design Laboratory,Cambridge, MA, USA

Abstract. We present an algorithm for the generation of coarse extract global characteristics and topologically sim-
and fine finite element (FE) meshes on multiply connected sur- ple subregions, which are identified in Part II (this
faces, based on the medial axis transform (MAT). The MAT is volume), from a given complex domain. This initial
employed to automatically decompose a complex shape into
topologically simple subdomains, and to extract important shape
shape decomposition can be considered as a coarse
characteristics and their length scales. Using this technique, we FE mesh. Those simple subregions are then triangu-
can create a coarse subdivision of a complex surface and select lated to generate a fine FE mesh. Thus, a mesh
local element size to generate fine triangular meshes within those capturing important geometric characteristics of a
subregions in an automated manner. Therefore, this approach given domain can be created by our mesh genera-
can lead to integration of fully automatic FE mesh generation
functionality into FE preprocessing systems.
tion scheme in an automated manner. Numerous
case studies of complex and diverse meshing exam-
ples, which we have performed using our imple-
mentation, demonstrate the effectiveness of our al-
1 Introduction gorithm [2,3].
The rest of this article is structured as follows. The
The finite element method (FEM) is a widely used, second section presents some geometric aspects of
powerful technique in many scientific and engineer- FE meshing process and a brief review of existing
ing fields. An ongoing effort is to improve its capa- meshing techniques. The third section introduces
bilities and to make it more readily usable in diverse fundamental aspects of MAT and our MAT algo-
areas. The FEM addresses the solution of the rithm which is the underlying technique used in this
boundary value or initial value problems which are FE meshing scheme. The fourth section presents
discretized by means of FE meshes. The discretiza- our FE mesh generation algorithm. Finally, the fifth
tion of a domain into a set of finite elements is a section summarizes this work and points out related
geometrically based process. An automated prepro- and outstanding issues for future research. Imple-
cessor which creates the FE model by interrogating mentation aspects and various engineering applica-
geometry would eliminate manual user intervention tions of our meshing algorithm are presented in a
during the mesh generation process. Such a func- companion paper [3].
tionality is currently lacking, to some extent, in ex-
isting FE preprocessors. Thus, it would be a very
useful addition to available finite element analysis 2 Automation of Finite Element Modeling and
(FEA) systems. Discretization
In this paper, we present a novel mesh generation
scheme which comprises two stages: shape interro- In this section, we first introduce several shape
gation and area meshing. Our mesh generation characteristics which are important in automating
scheme first uses the medial axis transform (MAT) FE modeling and discretization. We also briefly re-
[1] as an automatic shape interrogation method to view existing FE meshing techniques and identify
some limitations of those techniques. This section
Offprint requests: Nicholas M. Patrikalakis, Massachusetts In- serves to motivate our meshing algorithm which au-
stitute of Technology, Department of Ocean Engineering, Design tomatically interrogates and triangulates multiply
Laboratory Cambridge, MA, 02139, USA connected surfaces.
122 H.N. Gtirsoy and N.M. Patrikalakis

2.1 Geometric Shape Characteristics Useful for 2.2 Overview of Finite Element Mesh Generation
Finite Element Modeling Methods
An automated preprocessor which identifies and ex- FE mesh generation is concerned with the subdivi-
tracts significant shape characteristics would be a sion of a geometric entity, such as a curve segment,
very useful addition to currently available FEA sys- a surface patch, or a volume into a set of geometri-
tems. To make a preprocessor most useful, the fol- cally simple shapes referred to as finite elements.
lowing capabilities are needed: This subdivision process must be controlled to en-
sure
9 Detection of constrictions and their length scales
in the problem domain allows the implementation 9 the accurate representation of all significant geo-
of physically motivated and more efficient dis- metric characteristics of the problem domain by
cretization of the problem domain. The choice of the mesh;
the initial FE mesh topology is an important factor 9 the proper matching of geometric features be-
from the viewpoint of efficiency. To achieve rap- tween finite elements; and,
idly convergent results, we usually have to refine 9 that the size and distribution of elements through-
a FE mesh in regions where the domain is narrow. out the domain being meshed satisfy the require-
Those regions, for example, are significant from ments specified by the analyst.
the structural analysis viewpoint, because stress
Over the last 25 years, various mesh generation
concentrations usually occur in such areas. In
schemes have been developed. Detailed reviews on
fluid dynamics, those regions most likely give rise
existing FE meshing schemes can be found in [21,
to flow separation.
[4], and [5]. None of those techniques has gained
9 Extraction of holes in the problem domain and
general applicability for FE meshing of complex ge-
proximity information permits more effective dis-
ometries. One reason is that most existing mesh
cretization of the domain and, would increase the
generators require a large amount of interactive
accuracy of numerical results. Depending on the
user input. Another reason is that although some
boundary and load conditions, a finer mesh should
mesh generators, in general, create meshes with
be used around holes in order to obtain accurate
"good" shape characteristics, they occasionally
results in a FEA.
generate meshes of poorly shaped elements or even
9 Decomposition of a complex shape into a set of
generate an unacceptable mesh in some regions.
topologically simple subdomains helps creation
Consequently, automatic FE mesh generation is an
of FE models in an efficient and automated man-
active research problem in computational geometry
ner. and CAD. In the authors' opinion there are substan-
tial opportunities for increased automation, and
Detection of the above characteristics from the
higher process reliability and efficiency.
geometric representation provides important infor-
We can group existing FE mesh generation
mation to the FE analysis process. This type of in-
schemes into two broad classes based on their inter-
formation could also lead to the development of
action with the geometric representation of a region
more automated FE mesh generators. If length
to be meshed:
scales of constrictions and other shape characteris-
tics of the problem domain are known, a mesh gen- 1. the geometric interrogation approach, in which
erator could be developed to adaptively select ini- mesh generators operate only by interrogating
tial mesh topology and local mesh density. To the the original geometric representation, and
best of our knowledge, these capabilities are not 2. the geometric interrogation and modification ap-
available in existing FE preprocessors. Existing proach, in which mesh generators operate by
preprocessors frequently require interactive user both interrogating and incrementally modifying
input for the specification of significant shape char- the geometric and topological representation of
acteristics, some of which are introduced above. the region during the meshing process.
With this motivation, we aim to demonstrate the
We can also classify existing FE mesh generators
feasibility of automation of a complete FE prepro-
depending on their underlying algorithmic ap-
eesser using shape interrogation and geometric al-
proaches:
gorithms based on MAT. In this section, we also
briefly review state-of-the-art technology in FE 9 mapping mesh generation [6-12];
meshing and discuss some limitations of existing 9 node insertion followed by area/volume triangula-
techniques. tion [13-201;
Automatic Surface Mesh Generation Scheme: Part I 123

9 topology decomposition [21-26]; the meshing process to lead to an acceptable mesh

9 spatial decomposition [27-29]; and of a complex multiply connected domain.
9 recursive subdivision of a region to element level These observations indicate that available meth-
[30,311 ods have, in general, limitations in providing effi-
cient, reliable and automated solutions to the FE
Mesh generation schemes based on the geometry
mesh generation problem for complex geometries.
interrogation approach include mapping mesh gen-
A possible solution to the general FE meshing prob-
eration and spatial decomposition. Meshing
lem may be the development of meshing schemes
schemes using the geometry interrogation and mod-
based on a hybrid approach. In such a hybrid
ification approach include the techniques of topol-
scheme, topologically simple large portions of a
ogy decomposition, recursive subdivision, and
complex domain could be extracted and a well
some forms of point insertion followed by area/vol-
shaped fine mesh would be generated quickly within
ume triangulation.
individual subregions. Thus, an efficient and robust
mesh generation system could be developed.
The FE mesh generation algorithm we develop
2.3 Summary of Limitations of Existing Meshing
here makes use of a two-step approach. The scheme
Techniques
first uses the MAT [1] as a shape interrogation
Although the mapping mesh generation techniques method to extract global geometric characteristics
are computationally efficient, they cannot handle and topologically simple subregions from a given
complex and multiply connected domains without a complex domain. This initial shape decomposition
substantial amount of user input. Meshing schemes can be considered as a coarse mesh useful for
based on the node insertion followed by area/vol- p-convergence FEA [3]. Next, those simple subre-
ume triangulation approach have been studied in gions are triangulated to generate a fine mesh for
detail; however, available techniques for generating h-convergence FEA. Thus, a FE mesh capturing
the initial set of nodes are not robust in general. important geometric characteristics of a given do-
Frequently, initial nodes picked can give rise to main can be efficiently created in an automated
badly shaped elements that require special treat- manner.
ment to generate an acceptable FE mesh. In these
meshing methods, the user might be required to in-
terpret the problem domain, decompose it into ap-
propriate subdomains and supply mesh gradation 3 Medial Axis Transform
information.
Mesh generation schemes using the topology de- In this section, we introduce main aspects of the
composition approach are not computationally effi- MAT and also our algorithm for MAT computation
cient for creation of fine FE meshes. These mesh for two-dimensional shapes. Such shapes may ei-
generators can create non-two-manifold topologies ther be physically planar shapes or represent the
during the meshing of multiply connected regions in parameter space of a curved, and generally
three dimensions. Thus, they require more complex trimmed, parametric surface patch. More detailed
data structures to allow the representation of non- discussions of these topics and related literature
two-manifold situations and more complex opera- surveys on MAT algorithms can be found in [2] and
tors that act on such data structures. Mesh genera- [32-34].
tors based on the spatial decomposition technique
can create badly shaped elements close to the
3.1 Overview of Medial Axis Transform
boundary of the region. The resulting mesh layout
depends on the position and orientation of the initial Blum [1,35] has proposed the technique of MAT to
enclosing box used to generate the uniform grid. describe biological shape. In this technique, a two-
These mesh generators also need appropriate mesh dimensional shape is described by using an intrinsic
gradation information which is usually provided by coordinate system. Every point p on the plane con-
the user. The mesh generation schemes based on taining the shape may be associated with a nearest
the recursive subdivision approach involve compu- point on the boundary contour B. The Euclidean
tationally complex operations during the meshing of distance from a point p to the boundary set B is the
domains with complex geometry. Existing prepro- distance from p to a nearest point P on B,
cessors based on this mesh generation approach, in
general, require interactive user input in order for d(p, B) = rnin{d(p, P): P ~ B) (1)
124 H.N. Gfirsoy and N.M. Patrikalakis

Such a nearest point exists because our shape is a

closed subset of the Euclidean space [36]. For a
particular set of points the minimum distance is not
achieved uniquely. Such points are equidistant from
two or more points on the boundary contour. This
set of points together with the limit points of this set
constitute the medial axis (MA) or skeleton or sym-
metric axis of the shape (see Fig. 1). We consider
here only points in the interior domain bounded by
B. For example, in Fig. 1 we have the relationship
for an interior point a, d(a, B) = d(a, b) = d(a, c).
This definition of the MAT is equivalent to Blum's
definition of the MAT [35,36]. Blum defines the MA
of a closed curve B in the Euclidean plane to be the
set consisting of the centers of all maximal disks
which fit into the domain bounded by B. The metric
interpretation of the MA provides a natural basis for Fig. 2. The Voronoi diagram and the medial axis of a planar
building a more complete description of the shape. nonconvex shape.
On the MA S of a boundary B, we define a function
r : S ~ R + as follows, where R + is the set of non-
negative real numbers. For every p E S sence, the VD of a set decomposes the region into a
finite number of subregions ("influence" regions of
r(p) = d(p, B) (2) elements in the set). Construction of the MA can be
referred to as the solution of the closest boundary
The function r(p) is called radius function (RF) or point problem. The MA of a planar convex polygon
disk function of MA. According to Blum, shape decomposes it into a set of subregions each of
may be described procedurally by means of its MA which is the nearest neighborhood of a boundary
and the associated RF. Namely, given the MA and edge. Thus, MA and VD are closely related. In the
associated RF of a shape, we can exactly recon- case of a convex polygon, they are identical to each
struct it. It can be shown that given the MA we can other. If a polygon is nonconvex, including reen-
uniquely recover the original shape by taking the trant corners on its boundary contour, then the VD
union of all disks with radius equal to the RF and of the polygon is a superset of the associated MA.
centered on the MA (see Fig. 1). In this case, the difference between the MA and VD
The concept of MA is related to the closest neigh- is that the MA does not include edges of the VD
borhood problems in computational geometry [37]. incident at the re-entrant corners (see Fig. 2). A re-
Given a set o f n elements ee (e.g., points) in a region entrant corner of a polygon is a vertex at which the
R, we can associate every element ei with a particu- internal (material) angle is greater than rr. If the
lar subregion re in which every point is closer to that angle is less than 7r the corner is called convex
element ee than to all other elements of the set. vertex.
These individual subregions re are referred to as Using the MAT as a shape interrogation tech-
Voronoi regions. The region R is the union of all nique, we can effectively decompose complex
Voronoi regions r;. shapes into subregions as we present in the sequel.
The boundaries of the Voronoi regions comprise Those subregions can be reduced to either four-
the Voronoi diagram (VD) of the given set. In es- sided or three-sided simple subdomains. This pro-

Fig. 1. The medial axis transform of a planar shape.

Automatic Surface Mesh Generation Scheme: Part I 125

cess constitutes the starting phase of our FE mesh- segments and arcs of conics (i.e., parabola, ellipse,
ing scheme, in which a complex shape is and hyperbola [35]). The MA branch S(ei, ej) of two
decomposed into coarse subdivisions. boundary elements e; and ej is the locus of the points
equidistant from ei and ei. Descriptions of conic
MA branches and their parametric representations,
3.2 A Computational Methodology for the
useful for tracing purposes, are presented in [2] and
Medial Axis Transform
[32]. The conic branches of MA can degenerate to
In this section, we briefly present our methodology straight line segments or circles.
to compute the MAT of connected planar shapes In our computational methodology, we can analyt-
bounded by closed curved boundaries. The bound- ically define the MA in terms of conic sections be-
ary of a region (or shape) is defined by an exterior tween two boundary elements. Determining end
loop and one or more interior loops (i.e., contours), points on the MA, we obtain the MA branch associ-
if the region of interest is multiply connected. Each ated with the two boundary elements. For this pur-
loop of the boundary is composed of an ordered set pose, we make use of the fundamental offset pro-
of boundary elements (curve segments and ver- cess directed towards the interior of a region. This
tices). The algorithm developed for the MAT com- process is analogous to propagation of a grass-fire
putation covers straight line, circular arc, and gen- wave front towards the interior of a shape. The off-
eral nonuniform rational B-spline (NURBS) bound- set of distance h of the boundary B of a planar re-
ary curves. Our method can also be easily extended gion R is the envelope of the union of all closed
to compute the MAT of the complement of a planar circular disks of radius h, the centers of which are
shape bounded by an arbitrary number of loops [2]. points of B. This definition accounts for two curves
A loop is a union of a finite number of boundary on both sides of the boundary, inside and outside.
elements which are ordered in such a manner that We are interested only in the offset of the boundary
when the loop is traversed in the positive direction in the interior of the shape.
the interior of the shape lies to the right. An element Using the sign convention adopted, we observe
of the boundary is either a reentrant vertex, which that on an inward offset of the boundary loop, con-
is associated with material angle greater than ~r, or a vex arcs shrink but concave arcs expand compared
straight line segment, or a circular arc segment with to the initial boundary shape. We also notice that a
arbitrary radius. Line and circular arc segments are reentrant vertex can be regarded as a degenerate
bounded by two end vertices. There are also two case of a concave circular arc with zero radius, be-
distinct types of circular arc segments. When we cause such a vertex gives rise to a finite arc segment
traverse a boundary loop in the positive direction, if on the offset contour.
a circular arc segment is traversed in the clockwise Given the boundary contour of a region, our objec-
direction with respect to its associated circle it is tive for the computation of the MAT is to determine
convex. On the other hand, if a circular arc is tra- inward offset distances and the associated branch
versed in the counterclockwise direction it is con- points at which the topology of the contour
cave. changes. These are so called effective offset dis-
In our approach, a free-form boundary curve (e.g., tances and effective branch points. Effective branch
a Bezier or B-spline curve) is approximated within a points are the end points of MA branches.
prescribed tolerance using these three boundary el- During the course of the offset process, there are
ement types [2]. In our algorithm, we use the curve three distinct types of branch points [39], (see Fig.
approximation technique presented in [38]. Approx- 3). An initial branch point of a contour is a vertex at
imation of a curved boundary in terms of a set of which precisely two nonadjacent elements of the
straight line segments can give rise to artificial offset contour are tangent to each other. An inter-
branches in MAT computation. Such artificial ef- mediate branch point of a contour is a vertex to
fects can be counteracted by using a threshold tech- which one or more elements of the nonvanishing
nique [39]. For this purpose, we use a threshold offset shrink. Afinal branch point of a contour is a
angle which specifies the maximum value of the an- vertex which represents a vanishing offset contour.
gle giving rise to an actual MA branch between two A final branch point is, in fact, a special case of an
adjacent segments [2]. intermediate branch point and indicates the end of
If the boundary contour of a planar shape is com- the offset process.
posed of reentrant vertices, straight line segments, For the computation of intermediate and final
and circular arcs, with arbitrary radii, then the MA branch points, the boundary contour is systemati-
of this shape, in general, consists of straight line cally analyzed by using triplets of boundary ele-
126 H.N. Giarsoyand N.M. Patrikalakis

circular arc. An analytical solution for such inter-

sections is again presented in [32] and [2].
From the solution of intersection problems dis-
mal branch point cussed above, we obtain a set of potential branch
points and offset distances associated with them.
The following criteria are used to identify admissi-
ble branch points. An admissible branch point p is a
branch point associated with triplets (ei, ei, e~) or
two nonadjacent boundary elements (eq, er) which
is within the interior of the region R and at an equal
distance h from those boundary elements. An ad-
missible branch point must also satisfy the condi-
intermediate branch point tion that there is not boundary element at a distance
less than the offset distance associated with the
branch point.

p ER (5)
h = d(p, el) = d(p, ej) = d(p, ek) or (6a)
h = d(p, eq) = d(p, er) (6b)
initial branch point
d(p, en) >- h(n = 1. . . . . N) (7)

where N is the number of boundary elements.

Once all admissible branch points are computed,
Fig. 3. Branchpoints (adapted from [39]). then effective branch points and associated effec-
tive offset distance H of the boundary contour are
determined from the set of admissible branch
ments. Triplets are a subset of the boundary of a points. For a given boundary contour or its offset
region consisting of three adjacent boundary ele- the set of effective branch points is determined by
ments. the minimum value of the offset distance associated
In this computation, the objective is to determine with admissible branch points.
the offset distance at which the middle element of
the triplet shrinks to a point. That point is the inter- H = rain{hi . . . . . hm} (8)
mediate branch point of the triplet. This point is
determined by computing the intersection of the where hi denotes offset distance associated with an
two MA branches generated by the triplets. This admissible branch point pi and m is the number of
intersection problem may be classified in one of the admissible branch points.
following types:
1. straight line to straight line intersection; 3.3 Medial Axis Transform as a Shape
2. straight line to conic intersection; and Decomposition Technique
3. conic to conic intersection.
We should note that from the offset process we ob-
An analytical solution of these problems is investi- tain branch points of the MA and also values of the
gated in [32] and [2]. For the computation of inter- RF, namely offset distances at those branch points.
mediate branch points, there is a special case which The branch points and convex vertices of the
does not involve triplets. A convex circular arc seg- boundary contours denote the end vertices of indi-
ment on the contour collapses to its center point vidual MA branches. Then we can define the RF
when the inward offset of the contour has a distance associated with every MA branch in a continuous
equal to the radius of the circle. Hence, for such a form as a function of offset distance or in a discrete
case, the center of the circular arc corresponds to sense as offset distance at distinct points on the MA
an intermediate branch point. branch.
The computation of initial branch points involves MA branches and the initial boundary contour de-
the computation of a tangent intersection point be- compose a shape into a set of subregions (i.e.,
tween two segments, at least one of them being a Voronoi regions). If the given shape is convex,
Automatic Surface Mesh Generation Scheme: Part I 127

are determined using the set of branch points deter-

mined in this computation. We note that a contour
may be associated with more than one effective
branch point, though their effective offset distance
must be the same. In the MAT computation, a con-
tour may create more than one contour during the
offset process, depending on the geometry of the
contour. The MAT computation results in a set of
Voronoi regions decomposing the shape. The Voro-
noi regions are represented as faces in a boundary
Fig. 4. Decomposition of a nonconvex shape into simple subdo- representation (B-Rep) model. The MA branches,
mains. Voronoi edges, and boundary elements are the
bounding edges of the Voronoi regions. Discussion
of data structures and representation techniques
those resulting subregions are also convex. For a used in our algorithm and other implementation as-
nonconvex shape, some subregions are not convex. pects of this scheme can be found in [2], [34], and
Introducing Voronoi edges and cuts at initial branch [3].
points as shown in Fig. 4, we can further subdivide
such nonconvex subregions to obtain convex or
pseudo-convex subregions. Voronoi edges of a 4 An Automatic Mesh Generation Algorithm
given VD are the edges that are incident at re-en- Based on the Medial Axis Transform
trant corners andflat vertices of the boundary con-
tour of a planar shape. A fiat vertex of a boundary In this section, we present an automatic mesh gen-
contour of a planar shape is a junction point of two eration scheme for multiply connected planar
adjacent boundary elements at which the interior shapes based on the MAT, see also [2] for a more
angle is equal to ~r. A cut at an initial branch point is detailed description. Our algorithm is similar to the
a straight line segment connecting two nonadjacent one proposed in [40] for polygonal domains. Our
boundary elements across the initial branch point mesh generation scheme can, however, directly
and orthogonal to these boundary elements. A handle shapes with curved boundaries. This mesh-
pseudo-convex region is an area whose closed ing process requires little user input or manual in-
boundary can be offset until the area becomes nil tervention, since it uses the MAT to automatically
without splitting the area into separate components. interrogate the geometric representation. There-
A pseudo-convex region involves no narrow bottle- fore, such a meshing scheme may prove to be a
neck type part, and the RF's associated with MA useful tool in design and analysis of complex engi-
branches have no local minima other than the end neering structures.
points of the MA branches that generated the We first introduce fundamental aspects of a gen-
region. eral two-step area meshing process. After that we
As shown in Fig. 4, subregions can be further di- present our methodology and meshing algorithm
vided into topologically simple subdomains using a which incorporates the MAT with the FE mesh gen-
set of straight line segments which are indicated by eration process.
the dotted lines in this figure. These lines are gener-
ated by projecting the branch points onto the
4.I A General Two-Step Finite Element
boundary elements associated with the branch
Triangulation Scheme
points. The resulting subdomains are either three-
sided or four-sided subdomains. We present basic aspects of a two-step mesh gener-
ation process to obtain a two-dimensional triangular
mesh. Theoretically, this process could be extended
3.4 An Algorithm for the Medial Axis Transform
in an analogous way to three-dimensional cases us-
Our computational methodology summarized in this ing tetrahedra to discretize volumes. This approach
section has led to an algorithm for MAT computa- is hybrid in the sense that it involves two distinct
tion, whose pseudo-code is given below (see Algo- processes during finite element mesh generation.
rithm 1). In our algorithm, boundary elements are The first process decomposes the two-dimensional
treated independently to compute branch points. domain to be meshed into a set of topologically sim-
Then the effective offset distance and branch points ple subdomains.
128 H.N. Gi~rsoy and N.M. Patrikalakis

Algorithm 1: Medial Axis Transform

input: Data of boundary contours (loops) of a region and threshold angle (TA).

output: List of medial axis (MA) branches, radius function, and list of Voronoi regions.

begin

Put initial contour(s) into contour queue;

while there is contour existing in contour queue {

Get a contour from contour queue as current contour;
/f interior of current contour is not nil {
for each segment, other than artificial ones {
Compute admissible intermediate branch point (BP);
if segment is a concave circular arc
Compute admissible initial BP(s);
Put BP(s) into admissible BP list;
}
Determine effective BP(s) using admissible BP list;
Determine offset distance using an effective BP;
for every convex vertex on current contour {
/f material angle of convex vertex -< TA {
Compute segment of MA branch (S);
Put S into MA branch list;
}
}
if effective BP is final
Stop computation for current contour;
else {
Compute new offset contour(s);
Put new contour(s) into contour queue;
}
}
else {
Compute MA branches of nil contour;
Put MA branches into MA branch list;
}

Construct list for Voronoi regions;

for every boundary element of initial contour(s) {

Create Voronoi region using MA branches;
Put Voronoi region into list;
)
end

A simple subdomain in a simply connected convex cess. Thus a triangular mesh of the domain is ob-
or pseudo-convex region with one boundary loop tained by taking the union of all triangulated
which is composed of a sequence of either three or subdomains.
four edges. The second process triangulates individ- The mesh generator as input requires the geome-
ual simple subdomains generated by the first pro- try of a region R to be discretized as its input. This
Automatic Surface Mesh Generation Scheme: Part I 129

input is usually in a B-Rep form which can be di- aries, etc can give rise to very irregular meshes. It is
rectly obtained from a B-Rep model or derived from common practice to apply smoothing to the mesh
a constructive solid geometry (CSG) representation generated by triangulation in order to improve irreg-
by means of boundary evaluation. ular shapes. Various schemes are available for the
Then initialization of discretization begins. A com- mesh smoothing process such as Laplacian and iso-
plex shape with reentrant corners and multiple in- parametric methods [9]. Our smoothing technique
ternal holes is decomposed into simpler subdomains to satisfy this objective is an iterative process in
so that an admissible mesh with triangular elements which the position of an interior node is incremen-
of good shape characteristics can be generated. As tally changed by averaging its coordinates and the
a heuristic rule we require all triangular elements to coordinates of all adjacent nodes.
approximately be equilateral triangles. The result of Depending on the problem at hand, a FE discreti-
the decomposition process is a set of n convex or zation may need local refinement in order to im-
pseudo-convex subdomains ri whose union is the prove numerical results. Before the first FEA, a
original shape R. coarse mesh should be locally refined in regions
The subdomains generated in the previous decom- close to boundary segments associated with bound-
positions process are organized into an appropriate ary conditions such as fixed nodes, concentrated
data structure for meshing. This representation con- loads, and also singularities arising from fixed and
tains adjacency information among all subdomains reentrant corners. Also in an adaptive FEA using
so that the triangular elements generated in adjacent a posteriori error indicators [41], finite elements
subdomains satisfy compatibility requirements. In which give rise to large error should be refined in
an admissible FE mesh composed of compatible el- order to obtain better numerical results in subse-
ements, two adjacent elements share all nodes on quent analysis steps. For local refinement based on
the interface edge. the h-convergence approach, triangular elements
After the region is decomposed into a set of subdo- can be bisected across their longest edge. This
mains, each subdomain can be triangulated individ- method allows generation of compatible meshes,
ually and at the same time does not degrade the shape
characteristics of triangular elements. The follow-
Ti = tz(ri) (9) ing pseudo-code summarizes the main steps in-
volved in this two-step finite element mesh genera-
where the function /z(r;) embeds a triangular FE tion process (see Algorithm 2).
mesh Ti in subregion ri. These subdomains can be
regarded as super finite elements from the FE dis-
cretization point of view. An approach for the trian- 4.2 The Finite Element Mesh Generation
Scheme Based on the Medial Axis Transform
gulation of a simple subdomain can be based on
discretization of the boundary of the subdomain fol- Using the two-step meshing methodology intro-
lowed by triangulation of the interior. The discreti- duced in the previous section, we have developed a
zation of the boundary requires specification of meshing scheme based on the MAT, which auto-
mesh size and density. The discretization of the matically discretizes a two-dimensional shape into a
subdomain boundaries also assures compatibility set of triangular elements. This meshing scheme ac-
between adjacent elements in different subdomains. cepts the MAT of a shape as its input and generates
The FE mesh T of the region R is the union of all FE complete FE mesh information. Major steps of this
meshes Ti embedded in all subdomains r~. scheme are presented in the following sections.

T = U~_l Ti (10) 4.2.1 Decomposition o f a region by means o f the

medial axis transform
It is worth noting that other mesh generation Due to the nature of the MAT, every boundary ele-
schemes (such as mapping, recursive decomposi- ment is associated with a unique Voronoi region of
tion, and point insertion followed by triangulation the shape. We can subdivide a Voronoi region into
techniques) can also be used to discretize these in- simple subdomains in such a way that a portion of
dividual simple subdomains. the boundary element is associated with only one
Automatic triangulation of a complex shape may MA branch on the boundary of the Voronoi region.
generate finite elements with unfavorable shape Thus, we obtain a coarse discretization of the
properties. Regions with highly nonuniform shape shape. Such a coarse discretization can be effec-
characteristics, closely spaced holes, wavy bound- tively used as a FE mesh in a p-version FEA. An-
130 H.N. Giarsoyand N.M. Patrikalakis

Algorithm 2: A General Two-Step Finite Element Meshing Scheme

input: Boundary representation model of a domain to be meshed.
output: Finite element mesh composed of triangles.
begin
Decompose domain into topologically simple subdomains;
Generate nodes along boundaries of subdomains;
for each subdomain
Create a triangular finite element mesh;
for each subdomain
/f refinement is required
Refine mesh of subdomain without violating compatibility;
if shape characteristics of mesh are not good enough
Smooth mesh;
end

other possible application of this approach is that sion of triangular finite elements discretizing the
we can easily identify subdomains associated with subdomain. We also make use of the value of radius
significant boundary conditions. Thus we determine function at initial branch points to determine the
areas for local mesh refinement in a direct and effi- local length scale of constriction.
cient manner. As a result, this meshing scheme di-
rectly yields discretizations with a spatial addressa- 4.2.2 Processing of the subdomains obtained fi'om
bility property. Such a feature is very useful for the medial axis transform
adaptive FEA methods. The MAT based process of the previous section
Given a Voronoi region, the region can be further decomposes a complex shape into a set of topologi-
subdivided so that each MA branch on the perime- cally simple three- or four-sided subdomains. Al-
ter of the Voronoi region can be mapped onto a though these subdomains are topologically simple,
unique finite portion of the boundary element asso- the lengths of their edges may, sometimes, turn out
ciated with the Voronoi region. The mapping is to vary significantly. For example, very narrow
done by means of a projection process. In a degen- subdomains involving angles which are very differ-
erate case, if the boundary element is a reentrant
corner, all MA branches of the Voronoi region asso-
ciated with this vertex are mapped onto this vertex.
The segmentation of the boundary is carried out
by computing the projections of the end points of
MA branches (i.e. branch points) onto the boundary
element associated with those branches, (see Fig.
5). This decomposition process results in a set of
subdomains with simple topology. These subdo-
mains are either three- or four-sided subdomains,
(see Fig. 5). Three-sided subdomains arise at termi-
nal branches of the MA and, possibly, at reentrant
(nonconvex) vertices (see Figs. 4 and 5). If a poten-
tial three-sided subdomain is very narrow, (i.e.,
with a small acute angle), this subdomain is not gen-
erated and the adjacent four-sided subdomain is
merged with this three-sided subdomain. Four-
sided subdomains arise for all other branches of the
MA. Such a decomposition process also allows us
to effectively use local lengths scales of a shape. In
this work, we use the largest value of the radius
function for a given subdomain as the local length
scale associated with that subdomain. This informa- Fig. 5. Shapedecompositionby means of the medialaxis trans-
tion, in turn, is used to determine the length dimen- form.
Automatic Surface Mesh Generation Scheme: Part I 131

ent from rr/3 may be occasionally generated. Pres-

ence of such badly shaped subdomains makes it dif- F- x~ . /1'

ficult to create a FE discretization composed of x~ / D

triangles with good shape characteristics. As a rule,
we require that all triangular elements closely re-
semble equilateral triangles. Although we do not
d=2a
use quadrilateral elements in our meshing scheme,
the optimum quadrilateral element shape is the
square. The objective of these heuristic rules is that
a well shaped finite element does not have a bias
towards a particular direction in a FE mesh layout.
Such a requirement can be expressed in terms the _A
aspect ratios of elements. A measure of aspect ratio
D=d
may be defined using the MA and radius function of
a triangular element. The aspect ratio (AR) defined
Fig. 7. Finite element mesh of the parametric shape.
as the ratio o f length o f longest MA branch to the
maximum value o f the radius function also effec-
tively accounts for skewness of elements.
We illustrate these abnormalities involved in tions. For the time being, let us consider only shape
shape decomposition based on the MAT using a characteristics of these triangular elements. We can
simple shape. Figure 6 shows a rectangle whose identify a situation which gives rise to badly shaped
longer dimension D is a variable and the shorter elements using the ratio, R, of the length of associ-
dimension is constant, d = 2a. In Fig. 7, corre- ated MA branch to the maximum value of radius
sponding to D = 2a, (i.e. a square), the solid line is function, in this particular case 0 -< R < 89 To rec-
the MA of the rectangle and dotted lines represent tify this abnormality, the short MA branch needs to
discretization of the region in terms of triangular be processed. For example, as seen in Fig. 9, we
elements. Here we observe that the triangular ele- can construct an approximate MA by eliminating
ments exhibit good shape characteristics. The as- the short MA branch, and obtain a modified decom-
pect ratio of the triangles is V~. If we start to in- position. In this example, the triangle is of aspect
crease D, the MAT decomposition will give rise to ratio AR = 1.6 which is better than the previous
two very narrow subdomains. If2a < D < 5a/2, the value.
discretization of those subdomains produces trian- Short boundary elements and re-entrant vertices
gles with poor shape characteristics (i.e., AR > with angles only slightly above 7r also give rise to
2.24, see Fig. 8). In Fig. 8, triangulation of the rec- very narrow triangular Voronoi regions, (see Figs.
tangular domain is carried out using our mesh gen- 10 and 11). Such cases arise frequently for high ac-
eration method presented in the subsequent sec- curacy approximations of a curved shape with lin-
ear segments. For the narrow subregion due to

sj I f

d-2a
d- 2a
j = 2.24

n d I

L L!
D - 5a

Fig. 6. A simple parametric shape. Fig. 8. Finite elements with high aspect ratios,
132 H.N. Giirsoy and N.M. Patrikalakis

.f I .....

Fig. 9. Modified medial axis and associated discretization.

Fig. 11. A narrow subregion associated with a re-entrant vertex.

short boundary elements, we can modify the MA nated using a process illustrated in Fig. 12. This
branches associated with the subregion so that the process is analogous to the "kill-edge-vertex" to-
triangle becomes almost equilateral (see Fig. 10). In pology operator which is used to manipulate two-
this process, the apex of the triangular subregion is manifold boundary models [42]. In this case, first
moved towards the boundary element until the as- new vertices A, B, C, and D are inserted on adja-
pect ratio of the triangle reaches an acceptable cent MA branches at distances equal to a fraction
value such as 3. In the case of a narrow subregion of the local length scale value and away from the
associated with a re-entrant vertex, it is possible to branch points. Then adjacent branches are modi-
completely eliminate such a subregion from the de- fied by creating a new segment on each of these
composition. As seen in Fig. 11, the narrow triangu- adjacent branches. A new straight line segment on
lar subregion is shrunk to a line segment connecting a MA branch is defined by connecting the new
the initial branch point to the re-entrant vertex. vertex (e.g. A, B, C, and D) to the middle point,
Based on these observations, we can carry out the M, of the short branch. Since this modification
following operations to rectify or eliminate subdo- process makes use of local length scale informa-
mains with unacceptable shape characteristics for tion, it can be made robust enough so that no in-
FE mesh generation purposes: tersections occur between new segments and
other boundary elements of the shape. Although
9 Given the Voronoi decomposition of a shape, we this process introduces an approximate represen-
first determine the ratio, R, of the lengths of the tation of the MA, such approximations are accept-
MA branches bounding the region with the associ- able for FE mesh generation purposes.
ated maximum radius function value, and identify If an abnormal subregion is associated with a
Voronoi regions with small value of R (e.g. R < k). boundary element, its bounding edges are modi-
9 If a MA branch with a small value of the ratio R is fied so that the subregion becomes more uniform
connected at its one end to a terminal branch, as shown in Fig. 10. Here an aspect ratio value
these two branches can be merged together. AR -< 3 can be used as a criterion for this modifi-
9 If a short branch has more than one adjacent cation process.
branch at each of its two ends, the branch is elimi- If an abnormal triangular subregion is associated

//i
J
VR2 VR2

Fig. 10. A narrow subregion and its modification.

Automatic Surface Mesh Generation Scheme: Part I 133

onto a unique point of each boundary element asso-

ciated with that MA branch.
In this mapping, a point on the MA branch is pro-
jected onto an associated boundary element. In a
degenerate case where a boundary element is a re-
entrant corner, then this mapping would result in
that corner vertex for every point on the MA branch
associated with the reentrant vertex. After this
mapping, we obtain a set of points on each bound-
Fig. 12. Elimination of a short branch. ary edge and perpendicular segments connecting
them to the corresponding points on the MA branch
of the subdomain. Thus a set of quadrilateral strips
with a reentrant vertex, the subregion is removed
are embedded into the subdomain. These quadrilat-
from the decomposition and the MA branches ad-
eral strips have two edges orthogonal to the bound-
jacent to the subregion are modified, as shown in
ary element of the subdomain (see Fig. 13). Trian-
Fig. 11.
gular elements are easily generated within such
This processing of a Voronoi decomposition at- simple quadrilateral strips. If the subdomain is tri-
tempts to eliminate poorly shaped subdomains of angular, at most two triangular strips (they can be
the decomposition by merging a portion or whole considered as individual elements) are also created
part of an irregular subdomain with other adjacent in addition to a set of quadrilateral strips (see Fig.
subdomains. After these processes, the resulting 13).
decomposition is expected to allow us to generate a
mesh with desirable shape characteristics. 4.2.4 Triangulation of subdomains and area
meshing
4.2.3 Discretization of boundaries and interiors of In the previous section we introduced a technique
subdomains to discretize the boundaries and the interior of a
Once we process and subdivide Voronoi regions, subdomain. The quadrilateral strips generated by
we obtain a set of subdomains. We can discretize the mapping have two edges which are orthogonal
bounding edges of subdomains, except their bound- to boundary elements of the subdomain. If there
ary elements, by generating nodes along those are, for example, n nodal points inserted along the
edges. During this automatic node insertion pro- MA branch, then (n - 1) strips will be generated
cess, a fraction of the value of radius function is after the mapping process. New nodal points with
used as distance between inserted nodes. Since uniform spacing can be inserted along the edges of
these inserted nodes are shared by adjacent subdo- these strips by interpolating the number of nodes
mains, the final mesh will satisfy compatibility re- along the subdomain (see Fig. 14). Node numbers
quirements between adjacent elements generated in on individual strip edges orthogonal to the bound-
different subdomains. ary may be determined by using the following quasi-
Next we further subdivide individual subdomains linear interpolation equation:
into a set of quadrilateral strips. In degenerate
cases, a quadrilateral strip becomes a single triangle
N~=k+|
[i n_
l - k ]l l O < i < n - I (11)
which appears within triangular subdomains. In this
process, a set of nodal points is specified on every
boundary element associated with a subdomain.
Nodal points on the MA branch edges of a subdo-
main are mapped onto the boundary element of the
subdomain. This process can be regarded as a dis-
crete dual of the MAT. Namely, in the MAT com-
putation, we determine the MA branch of two
boundary elements which is the set of points equi-
distant from those elements. On the other hand, in
subregion discretization, we determine points on strips
strips
the boundary elements corresponding to points of
the MA branch. Every point on a MA branch not Fig. 13. Strips created in quadrilateral and triangular subdo-
associated with a reentrant vertex can be mapped mains.
134 H.N. G~rsoy and N.M. Patrikalakis

n The triangulation process we use in discretization

of individual subdomains is an efficient and robust
technique. Our meshing algorithm, however, is not
I l

rUT,',',!l I I i

', :
limited only to this scheme. We could effectively
use other triangulation methods (e.g. Delaunay tri-
angulation [16]). In such an approach, we could de-
1 fine a individual subdomain using MA branches and
I I
I1 1 I
boundary element associated with the subdomain.
In that triangulation process, nodes on the bound-
I m lip

/ ary and in the interior of the subdomain would be

I-,,.i .t
NJ-k 1 "injected" into the mesh one at a time. In this case,
the meshing scheme based on our shape decomposi-
tion approach would also result in better computa-
tional efficiency, since we break down the overall
Fig. 14. Node generation procedure within a subdomain.
triangulation task into a set of smaller tasks. In the
meshing process, the triangulation of a set of subdo-
mains with a small fraction of the total number of
nodes could be carried out more efficiently than the
where N; is the number of nodes on the ith edge triangulation of the whole domain using all nodes.
connecting a point on the boundary element to a After all subdomains of a shape are triangulated,
corresponding point on the branch at integer posi- mesh smoothing and local refinement processes can
tion i in the point sequence; k and l are the numbers be applied to the resulting mesh in order to improve
of nodes on the extreme oblique edges of the subdo- its shape characteristics. The following pseudo-
main; n is the number of nodes on the MA branch; code summarizes the main steps involved in our
and [ ] denotes ceiling. In Fig. 14, the left and right finite element mesh generation process based on
"vertical" edges are oblique to the bottom edge be- MAT (see Algorithm 3). Data structures used and
cause of merging two triangular subdomains with techniques developed for these processes are re-
high aspect ratio with this quadrilateral subdomain. ported in [3].
After inserting all nodes along the edges of the Figures 15 through 18 illustrate the steps involved
strips of a subdomain, triangulation of a subdomain in our FE mesh generation scheme. Figure 15 illus-
can be carried out by triangulating individual strips. trates the MA of the multiply connected planar
If the subdomain is triangular it can be transformed shape. Subdivision of the region and nodes inserted
to a quadrilateral by removing a triangular element along the edges of the subdomains are illustrated in
(which is the degenerate strip of this subdomain) Fig. 16. Figure 17 shows the triangulation process
from the subdomain. Triangulation of the quadrilat- within a subdomain. Finally, the resulting mesh is
eral strips is a relatively straightforward process shown in Fig. 18.
which involves matching of appropriate nodes on
two opposite edges of the strip. Several techniques
have also been proposed for triangulation of such
strips [10]. The edges of the strips are perpendicular
to boundary elements of the subdomain. For a given
strip, the four corner vertices are interrogated to
determine the one associated with the largest inte-
rior angle. Once this vertex is identified, it is con-
nected to an appropriate vertex on the other edge of Fig. 15. A planar multiply connected shape and its medial axis.
the strip such that a triangular element is extracted
from the strip at the end of this process (see Fig.
17). Then, this process is applied to the remaining
part of the strip. Thus a set of triangles are gener-
ated by removing triangles from the strip one at a
time. In our current implementation, the difference
of node numbers on opposite edges of a strip is as
large as 3 but could easily be modified to other val-
ues. Fig. 16. Region subdivision and boundary discretization.
Automatic Surface Mesh Generation Scheme: Part I 135

Algorithm 3: A Two-Step FE Meshing Based on Medial Axis Transform

input: Data of boundary contour(s) of a region.
output: List of triangular finite elements.
begin
Construct medial axis, list of Voronoi regions using Algorithm 1;
Create list for simple subdomains;
for each three sided Voronoi region
Add region into subdomain list;
for each remaining Voronoi region {
Subdivide region into three four-sided subdomains;
Add subdomains into subdomain list;
}
for each edge between adjacent subdomains
Generate uniformly distributed nodes;
Create list for triangular elements;
for each subdomain in list {
/f subdomain is four sided {
Generate set of quadrilateral strips;
Triangulate strips in quadrilateral subdomain;
}
tf subdomain is three sided {
Generate set of strips;
Convert triangular subdomain into quadrilateral by extracting one
degenerate triangular strip as a triangular element;
Triangulate strips in remaining quadrilateral subdomain;
}
Add triangles into list;
}
/f shape of triangular elements requires improvement
Smooth mesh;
end

4.3 Complexity Estimates of the Medial Axis approach. Then all such points which pass local
Transform and Meshing Processes tests of being possible initial branch points, must be
checked against all other boundary elements O(n) to
A basic complexity analysis of our MAT algorithm
determine if they are admissible (i.e. not closer to
follows. We need to consider all O(n 2) pairs of
any other elements). At worst, therefore, the time
boundary elements to determine points of closest
complexity is O(n3). This estimate assumes that lo-
cal tests for possible initial branch points do not
reduce the order of magnitude of the number of
such points. Experiments indicate that this is an
overly pessimistic assumption.
The coarse triangular mesh generation process is

stops
\\// 4///!/
f
Sh - . v

Fig. 17. Triangulation of an individual subdomain.

\/// <//

Fig. 18. Finite element mesh of a planar multiply connected re-

gion.
136 H.N. Giirsoy and N.M. Patrikalakis

an O(m) process with respect to the total number of MAT in this case, the MA of an object in three
triangular elements, m, being generated. Triangula- dimensions, in general, involves mixed dimensional
tion of individual subdomains has a linear time com- entities (such a MA generally comprises connected
plexity with respect to the number of elements gen- distinct vertices, curved edge segments, and sur-
erated. The reason is that our meshing scheme is face patches as MA branches). Thus, representa-
geometrically based and no search operations are tion of such a complex structure within a volume
carried out during the meshing process as, for ex- would require sophisticated B-Rep techniques ca-
ample, in topologically based schemes. Thus this pable of handling non-two-manifold situations [44-
approach gives rise to a linear running time com- 46].
plexity. Even though there are major difficulties in MAT
Considering the overall performance of our imple- computation in higher dimensions, MAT is a very
mentation, the overhead involved in our MAT algo- rich topic involving interesting research problems.
rithm is subdominant with respect to the mesh crea- This technique promises elegant solutions to many
tion and refinement time complexity. In a typical potential applications in engineering design and
mesh generation problem n ~ m and therefore the analysis [2,34].
bulk of the computational effort is spent for actual
mesh generation and refinement processes, which
exhibit a linear time complexity. Timing results of Acknowledgments
several test cases obtained using our computer im-
Funding for this research was obtained in part from MIT Sea
plementation confirm these observations [2,3]. Grant College Program (grant number NA90AA-D-SG-424) and
the Office of Naval Research in the United States (grant numbers
N00014-87-K-0462 and N000-91-J-1014). Professor C. M. Hoff-
5 Summary and Conclusions mann provided valued assistance on the complexity analysis of
our MAT algorithm.
We have presented the MAT as a shape interroga-
tion method and an algorithm to compute the MA of
two-dimensional shapes with curved boundaries. References
The two-dimensional shape can either be planar ob- 1. Blum, H., (1967) A transformation for extracting new de-
jects or the parameter space of a curved trimmed scriptors of shape, models for the Perception of Speech and
surface patch. For the latter case, mapping of the Visual Form, Weinant Wathen-Dunn (Editor), MIT Press,
Cambridge, MA, 362-381
MA to three-dimensional space via the surface
2. Gursoy, H.N., (1989) Shape interrogation by medial axis
equation allows surface discretization. This scheme transform for automated analysis, PhD Dissertation, Massa-
can effectively extract several important shape chusetts Institute of Technology, Cambridge, MA
characteristics in an automated manner. We have 3. Gursoy, H.N., Patrikalakis, N.M., (1992) A coarse and fine
also presented a new FE mesh generation scheme element mesh generation scheme based on medial axis trans-
form: Part II implementation, Engineering with Computers,
based on MAT. Several important advantages of
Springer-Verlag, New York
this scheme include the capability to extract shape 4. Ho-Le K., (1988) Finite element mesh generation methods:
characteristics and length scales in a fully auto- A review and classification, Computer-Aided Design, 20, 1,
mated manner, a direct way of defining super finite 27-38
elements and a substructuring capability, and spa- 5. Shephard, M.S., (1988) Approaches to the automatic genera-
tion and control of finite element meshes, ASME Applied
tial addressability of resulting discretizations cre-
Mechanics Review, 41, 4, 169-184
ated by the scheme. 6. Zienkiewicz, O.C., Phillips, D.V., (1971) An automatic mesh
Extension of these techniques to higher dimen- generation scheme for plane and curved surfaces by isopara-
sions (MAT on curved surfaces and within three- metric coordinates, International Journal for Numerical
dimensional volumes) is feasible [2,43]. However, Methods in Engineering, 7, 461-477
7. Gordon, W.J., (1973) Construction of curvilinear coordinate
such extension is expected to be computationally systems and applications to mesh generation, International
intensive due to substantial increases in combinato- Journal for Numerical Methods in Engineering, 7, 461-477
rial and algebraic complexity of the process in 8. Cook, W.A., (1974) Body oriented (natural) coordinates for
higher dimensions. generating three dimensional meshes, International Journal
MAT computations on general curved surfaces us- for Numerical Methods in Engineering, 8, 27-43
9. Herrmann, L.R., (1976) Laplacian-isoparametric grid gener-
ing an appropriate distance metric would require ation scheme, Journal of Engineering Mechanics Division
geodesics. MAT of general three-dimensional Proceedings of the ASCE, 102, EM5, 749-756
closed volumes is an active research problem. In 10. Imafuku, I., Kodera, Y., Sayawaki, M., (1980) A general-
addition to the computational complexity of the ized automatic mesh generation scheme for finite element
Automatic Surface Mesh Generation Scheme: Part I 137

method, International Journal for Numerical Methods in En- 28. Kela, A., Perucchio, R., Voelcker, H., (1986) Towards auto-
gineering, 15,713-731 matic finite element analysis, ASME Computers in Mechani-
11. Cavendish, J.C., Hall, C.A., (1984) A new class of transi- cal Engineering, 5, 1, 57-71
tional blended finite elements for analysis of solid structures, 29. Baehmann, P.L., Wittchen, S.L., Shephard, M.S., Grice,
InternationalJournal for Numerical Methods in Engineering, K.R., Yerry M.A., (1987) Robust, geometrically based, au-
20, 241-253 tomatic two-dimensional mesh generation, International
12. Wellford, L.C., Gorman, M.R., (1988) A finite element tran- Journal for Numerical Methods in Engineering, 24, 8, 1043-
sitional mesh generation procedure using sweeping func- 1078
tions, International Journal for Numerical Methods in Engi- 30. Sluiter, M.L.C., Hansen, D.C., (1982) A general purpose
neering, 26, 2623-2643 automatic mesh generator for shell and solid finite elements,
13. Frederick, C.O., Wang, Y.C., Edge, F.W., (1970) Two-di- Computers in Engineering, L.E. Hulbert (Editor), ASME,
mensional automatic mesh generation for structural analysis, New York, 3, 29-34
InternationalJournal for Numerical Methods in Engineering, 31. Bykat, A., (1983) Design of a recursive, shape controlling
2, 1, 133-144 mesh generator, International Journal for Numerical Meth-
14. Van-Phal, N., (1982) Automatic mesh generation with tetra- ods in Engineering, 19, 9, 1375-1390
hedron elements, International Journal for Numerical Meth- 32. Patrikalakis, N.M., Gursoy, H.N., (1988) Skeletons in shape
ods in Engineering, 18,273-280 feature recognition for automated analysis, MIT Ocean En-
15. Lee, Y.T., De Pennington, A., Shaw, N.K., (1984) Auto- gineering Design Laboratory Memorandum 88-4
matic finite-element mesh generation from geometric 33. Patrikalakis, N.M., Gursoy, H.N., (1989) Shape feature rec-
model--a point-based approach, ACM Transactions on ognition by medial axis transform, MIT Ocean Engineering
Graphics, 4, 287-311 Design Laboratory Memorandum 89-1
16. Cavendish, J.C., Field, D.A., Frey, W.H., (1985) An ap- 34. Patrikalakis, N.M., Gursoy, H.N., (1990) Shape interroga-
proach to automatic three-dimensional finite element mesh tion by medial axis transform," Advances in Design Auto-
generation, International Journal for Numerical Methods in mation 1990, Volume One: Computer Aided and Computa-
Engineering, 21,329-347 tional Design, B. Ravani (Editor), ASME, NY, 77-88
17. Lo, S.H., (1985) A new mesh generation scheme for arbi- 35. Blum, H., (1973) Biological shape and visual science (Part
trary planar domains, International Journal for Numerical I)," Journal of Theoretical Biology, 38, 205-287
Methods in Engineering, 21, 8, 1403-1426 36. Wolter, F.-E., (1985) Cut loci in bordered and unbordered
18. Joe, B., (1986) Delaunay triangular meshes in convex poly- riemannian manifolds, PhD Dissertation, Technical Univer-
gons, SIAM Journal on Scientific and Statistical Computing, sity of Berlin, Department of Mathematics
7, 2, 514-539 37. Preparata, F.P., Shamos, M.I., (1985) Computational geom-
19. Field, D.A., (1986) Implementing Watson's algorithm in etry: An introduction, Springer-Verlag, New York
three dimensions, Proceedings of the Second Annual ACM 38. Patrikalakis, N.M., (1989) Approximate conversion of ra-
Symposium on Computational Geometry, 246-259 tional splines," Computer Aided Geometric Design, 2, 6,
20. Schroeder, W.J., Shephard, M.S., (1988)Geometry-based 155-165
fully automatic mesh generation and the delaunay triangula- 39. Montanari, U., (1969) Continuous skeletons from digitized
tion, International Journal for Numerical Methods in Engi- images, Journal of the Association for Computing Machin-
neering, 26, 2503-2515 ery, 16, 4, 534-549
21. Bykat, A., (1976) Automatic generation of triangular grid: 40. Srinivasan, V., Nackman, L.R., Tang, J-M, Meshkat, S.N.,
I--Subdivision of a general polygon into convex subregions. (1990) Automatic mesh generation using the symmetric axis
II--Triangulationof convex polygons, International Journal transformation of polygonal domains, Research Report,
for Numerical Methods in Engineering, 10, 1329-1324 IBM Research Division, RC 16132
22. Sadek, E., (1980) A scheme for the automatic generation of 41. Babuska, I., Zienkiewicz, O.C., Gago, J., Oliveira, E.R.A.,
triangular finite elements, International Journal for Numeri- (1986) Accuracy estimates and adaptive refinements in finite
cal Methods in Engineering, 15, 1813-1822 element computations, John Wiley, Chichester, England
23. Wordenweber, B., (1984) Finite element analysis for the na- 42. Mantyla, M., (1988) An Introduction to Solid Modeling,
ive user, Solid Modeling by Computers, From Theory to Computer Science Press, Rockville, Maryland
Applications, Plenum Press, NY., M.S. Pickett, J.W. Boyse 43. Hoffmann, C.M., (1990) How to construct the skeleton of
(Editors), 81-101 CSG objects, Technical Report, CSD-TR-1014, Purdue Uni-
24. Woo, T.C., Thomasma, T., (1984) An algorithm for generat- versity
ing solid elements in objects with holes, Computers & Struc- 44. Weiler, K.J., (1986) Topological structures for geometric
tures, 18, 2, 333-342 modeling, PhD Dissertation, Rensselaer Polytechnic Insti-
25. Joe, B., Simpson, R.B., (1986) Triangular meshes for regions tute, Troy, NY
of complicated shape, International Journal for Numerical 45. Gursoz, L., Choi, Y. and Prinz, F.B., (1990) Vertex-based
Methods in Engineering, 23, 5,751-778 representation of non-manifold boundaries, Geometric Mod-
26. Chae, S.W., (1988) On the automatic generation of near- eling for Product Engineering, M.J. Wozny, J.U. Turner and
optimal meshes for three-dimensional linear elastic finite ele- K. Preiss (Editors), North-Holland, NY, 107-130
ment analysis, PhD Dissertation, Massachusetts Institute of 46. Rossignac, J.R., O'Connor, M.A., (1990) SGC: A dimen-
Technology, Cambridge, MA sion-independent model for pointsets with internal struc-
27. Yerry, M.A., Shephard, M.S., (1984) Automatic three-di- tures and incomplete boundaries, Geometric Modeling for
mensional mesh generation by the modified octree tech- Product Engineering, M.J. Wozny, J.U. Turner and K. Pre-
nique, InternationalJournal for Numerical Methods in Engi- iss (Editors), North-Holland, NY, 145-180
neering, 20, 1965-1990