INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING, VOL.

37, 793-812 (1994)

AN APPROACH TO REFINING THREE-DIMENSIONAL

TETRAHEDRAL MESHES BASED ON DELAUNAY
TRANSFORMATIONS

N. A. GOLIAS A N D T. D. TSIBOUKIS
Department of Electrical Engineering, Aristotle University of Thessaloniki. Thessaloniki, 54006 Greece

SUMMARY
A technique for refining three-dimensional tetrahedral meshes is proposed in this paper. The proposed
technique is capable of treating arbitrary unstructured tetrahedral meshes, convex or non-convex with
multiple regions resulting in high quality constrained Delaunay triangulations. The tetrahedra generated are
of high quality (nearly equilateral). Sliver tetrahedra, which present a real problem to many algorithms are
not produced with the new method. The key to the generation of high quality tetrahedra is the iterative
application of a set of topological transformations based on the Voronoi-Delaunay theory and a reposition
of nodes technique. The computational requirements of the proposed technique are in linear relationship
with the number of nodes and tetrahedra, making it ideal for direct employment in a fully automatic finite
element analysis system for 3-D adaptive mesh refinement. Application to some test problems is presented to
show the effectiveness and applicability of the new method.

INTRODUCTION
The main idea of the finite element method is the local nature of the approximating functions, the
shape functions’ as they are called. Exploiting this local nature, relatively simple shape functions
can be used to approximate complex fields with high accuracy. The secret lies in the generation of
an appropriate mesh of finite elements with varying density, so that the actual variation of the
unknown function is approximated. Element density should be high in regions where the field to
be approximated exhibits rapid variation. Further, the elements employed must have a high
aspect ratio (be of good quality). Quality is a very important factor since a bad-quality mesh can
result in inaccurate numerical modelling and ill-conditioned solution matrices, affecting the
accuracy of the approximation.’ Therefore, the generation of an appropriate mesh is of very
crucial importance to the successful application of the finite element method.
It is widely accepted that mesh generation is one of the most difficult processes of the finite
element method. This part of analysis may absorb as much as 80 per cent of the analyst’s time.
Mesh generation is a highly interactive job. In 3-D the situation is even worst since the
visualization of a 3-D mesh on a 2-D graphics terminal is not obvious, making the mesh
generation process a very difficult job. This is the reason why many analysts have shown their
preference to the use of the so-called brick element.’ The manipulation of a hexahedral mesh is
much easier than that of a tetrahedral mesh. However, simplex elements present certain proper-
ties that make their implementation attractive. Complete polynomial expansion functions are
very easily defined on these elements. Further, an arbitrary domain, no matter how complicated it
might be, can be decomposed ‘exactly’ as a set of simplex elements. Also, the density of simplex
elements can be varied, according to actual field variation.

CCC 0029-5981/94/050793-20 $9-00 Received 2 March 1992

0 1994 by John Wiley & Sons, Ltd. Revised 24 August 1993
794 N. A. GOLIAS AND T. D. TSIBOUKIS

The implementation of an adaptive refinement technique is important in two ways. First, the
high-error regions are recognized automatically, and then the element density is increased in
these regions, increasing the accuracy of the finite element method. Second it transforms the mesh
generation procedure from a highly interactive job to a batch one, which is very important since
mesh generation is the most time-consuming procedure in terms of man-hours. Adaptive mesh
refinement has proven to be an indispensable tool in 2-D3-6 and recently in 3-D7 finite element
analysis.
Up to now a lot of work has been contributed on automatic mesh generation. Indeed, various
efficient algorithms for triangulating an arbitrary distribution of nodes in two and three-
’
dimensions exist,8- offering almost full automation during the grid generation process. Al-
though the performance of these algorithms has been improved, they present certain drawbacks
that affect their applicability when employed in a 3-D mesh refinement procedure. Generation of
sliver elements, lack of an efficient technique for the automatic generation of interior nodes in the
problem domain and time complexity greater than linear make their application difficult and
troublesome.
A technique is presented in this paper for refining 3-D tetrahedral meshes. Insertion of new
nodes on selected regions of the mesh, preserving conformity, is direct. Delaunay conformity is
maintained by the application of three newly proposed local Delaunay topological transforma-
tions, which have no problem in dealing with degenerate cases as well as with complex domains
and multiple regions, resulting in constrained Delaunay triangulations. A technique for re-
positioning the nodes of the mesh is proposed for improving the quality of the tetrahedra.
Tetrahedra generated with the new technique are of high quality (nearly equilateral), improving
the approximation properties of the shape functions.’ The key to this end is the repeated
application of the topological transformations with the node relaxation technique until the mesh
is balanced. So an ‘optimal’ constrained Delaunay triangulation is generated, with high quality
tetrahedra. The presence of sliver elements is avoided completely. The extinction of bad-shaped
elements makes possible the approximation of curved surfaces directly in the refinement process.
Curved surfaces of arbitrary parametric representation, convex and non-convex, can be handled.
A very important advantage of the proposed technique is that it is implemented as an O(n)
process. With the introduction of certain connectivity structures no searching is needed, and this
makes possible the handling of large tetrahedral meshes as is shown in the application section.
Further, an acceleration of the procedure with the size of the tetrahedral meshes has been
observed. Application of the proposed technique in a completely automatic finite element
analysis system is described. The refining process in combination with an error estimation
technique can lead to optimal solutions with reduced computational cost. Application to some
test problems is presented to show the effectiveness of the proposed technique.

QUALITY OF TETRAHEDRONS

Tetrahedron’s quality
The quality factor of a tetrahedron is defined as three times the ratio of the radius r of the
inscribed sphere to the radius R of the circumscribed sphere, i.e.

Q=31
R
The quality factor is a number between 0 and 1. Tetrahedra with quality factors near 1 are
nearly equilateral while tetrahedra with low quality factors look like thin slices (flat tetrahedra). If
 DELAUNAY TRANSFORMATIONS 795

the ratio r / R is less than 0.01 ( r / R < 0.01) then the tetrahedron under consideration is character-
ized as a sliver element,',16 i.e. a very much degenerated tetrahedron.

Mesh quality
In order to have a picture of the mesh quality the mean quality Q, and the joint quality Qj are
introduced:

where N is the number of tetrahedra and Qi is the quality factor of the ith tetrahedron. As is
understood, the above means can give us a fairly good picture of the real mesh quality. Q, is
indicative of average element quality while Qj is indicative of the existence of sliver elements. The
joint quality factor Qj has a value that is very sensitive to elements with low quality factors. It is
much like the formula for resistances in parallel. If some sliver elements exist in the mesh, with
very low quality factors, the value of this mean is very near to zero.

Fast quality
A computationally cheaper formula (about 20 times cheaper) in comparison with (1) for
calculating a tetrahedron's quality is given by

Q=--
12 v
Jz L A x (4)

where V is the tetrahedron's volume and L,,, is the tetrahedron's longest edge.

VORONOI-DELAUNAY CONSTRUCTS
Delaunay triangulations and their dual structure, the Voronoi polyhedra, constitute one of the
more fundamental constructs of computational g e ~ m e t r y .-' I~9 Suppose we have a set of points
in R" called nodes, as is shown in Figure 1. The associated Voronoi polyhedron to a node P i is the
region of space closer to this node than to every other node. This fact is expressed mathematically
as
Vp8= { r E R" I d(r, P i ) < d(r, P j ) , j = 1, . . . , N, j # i} (5)
Each node has a Voronoi polyhedron associated with it. By connecting all nodes that have
neighbouring Voronoi polyhedra the complementary structure of the Voronoi polyhedra is
formed. This is the Delaunay triangulation of simplex elements (tetrahedra in three dimensions)
for the given set of nodes.
A very important property of a Delaunay triangulation is that, for a given tetrahedron, no
node, other than the four nodes of the tetrahedron, lies in the interior of the tetrahedron's
circumscribed sphere. This is referred to as the property of the empty sphere of a Delaunay
triangulation.'
The Delaunay triangulation is unique for a set of points except in the case when degeneracies
occur. A degenerate case is met when one or more nodes lie on the boundary of a tetrahedron's
796 N. A. GOLIAS AND T. D. TSlBOUKlS

Figure 1. Delaunay triangulation and its dual structure, the Voronoi diagram

circumsphere other than the tetrahedron’s nodes. While degenerate cases are very rare for
a random distribution of nodes they are very likely to occur in shapes which are usual in
engineering problems.

CONNECTIVITY STRUCTURES
A structure of the elements’ neighbours is allocated to hold each element’s neighbours, as is
presented in Table I. In this way the neighbours of a tetrahedron can be accessed directly and no
searching is needed.
In the implementation of NODE RELAXATION two other connectivity structures are
formed: NTON (node to node) and NTOE (node to element). In this way, the nodes and the
tetrahedra connected to a node are directly accessed and no searching is needed.
The algorithms for the formation of the NTON and NTOE structures are given below. The
tetrahedra are swept only once, making the process very fast. The procedure push stack c o places
the object o in the stack if o does not exist already in the stack.

Algorithm NTON formation

FOR i = 1 T O nrTetrahedra
Get the four nodes of tetrahedron i + 01,02,03,04
push stack[ol] c 02,03,04
push stack[o2] c 01,03,04
push stack[o3] t 01,02,04
push stack[o4] t 01,02,03
END FOR

Algorithm NTOE formation

FOR i = 1 TO nrTetrahedra
Get the four nodes of tetrahedron i + 01,02,03,04
push stack[ol] c i
 DELAUNAY TRANSFORMATIONS 797

push stack[o2] t i
push stack[o3] t i
push stack[o4] c i
END FOR

MESH REFINEMENT
Tetrahedron refinement
A new node K can be inserted in a tetrahedron’s interior, face or edge, dividing it into four,
three or two tetrahedra, respectively. Insertion of the new node K on an edge is a better choice
because the new tetrahedra produced are of better quality and nodes can be inserted everywhere
in the mesh, both in the interior and on external interfaces. Refinement of a tetrahedron is done by
refining its longest edge. Suppose the longest edge of tetrahedron ABCD is AB. Node K is
inserted on the middle of edge AB bisecting all the tetrahedra that have this edge in common to
preserve conformity of the mesh as is shown in Figure 2. In calculating the lengths of the
tetrahedron’s edges a favouring weighting should be given to boundary edges, so that the
insertion of new nodes on boundaries is favoured.
Refinement of edge AB (Figure 2) results in the refinement of all tetrahedra connected to this
edge. A recursive algorithm for accessing these tetrahedra without searching is given below. The
algorithm exploits the neighbours’ structure and the tetrahedra are stored on a stack structure.
The expression neighbour (ABCD, C) means the neighbour of tetrahedron ABCD opposite node C.

Algorithm (Find tetrahedra having edge AB in common with tetrahedron ABCD)

STACKTOP = 0
PROCEDURE Find-Tetrahedra-With-Edge ((ABCD), (AB))
BEGIN
T1 = neighbour ((ABCD), (C))
T2 = neighbour ((ABCD), (D))
TOP = STACKTOP
IF (T1 > 0) THEN PUSH STACK + T1
I F (T2 > 0) THEN PUSH STACK + T2
I F (STACKTOP > TOP) THEN
I F (T1 > 0) THEN Find-Tetrahedra-With-Edge (Tl, (AB))
IF (T2 > 0) THEN Find-Tetrahedra-With-Edge (T2, (AB))
END
Table I. Structures allocated to hold element nodes
and element neighbours

Element Element Element

number nodes neighbours
~~ ~ -

1 1 5 3 1 7 3 2 2 3 0
2 2 1 3 1 7 1 4 4 7 0
3 1 7 3 5 7 0 2 2 2
4 4 3 2 1 7 2 6 6 2 0
5 2 0 9 5 7 0 35 38 33
... ...
798 N. A. G O L I A S A N D T. D . TSIBOUKIS

A A

B 8

Figure 2. Insertion of a new node K on edge AB has, as a result, the bisection of all tetrahedra involved

Curved surfaces
Approximation of arbitrary curved surfaces is employed directly in the refinement procedure
with a simple technique. The method has proven to be versatile and easy to implement in two
dimension^.^ If the edge to be refined lies on a curved surface then the new node is inserted
directly on the curved surface. Suppose the curved boundary is represented by the following
parametric equations:
x = x(u, u), y = y(u, u), 2 = z(w, u) (6)
and a new node K is to be inserted in the middle of edge AB, where
A: x1 = X(U1, Ul), y , = Y ( U 1 , UI), z1 = Z(U1, u 1 ) (74
B: x 2 = x ( u 2 , ~ 2 )y 2~ = Y ( U ~ ~- 2 )z2 ~= z ( u 2 , ~ 2 ) (7b)
Then the new node K is inserted on the point with co-ordinates
x = X((U1 + u2)/2, (01 + v2)/2)
Y = Y((U1 + u2)/27 (01 + u2)/2) (8)
= z((u1 + u2)/27 (u1 + v2)/2)
which lies precisely on the curved boundary.
The approximation of curved boundaries is a very difficult task in the context of a refining
procedure. Serious troubles can come up when bad-shaped elements exist in the mesh. This is
shown in Figure 3, where the presence of a bad-shaped element results in the formation of
overlapping elements. In such a case the structure of the mesh must be modified with the
application of the proposed transformations and node repositioning (presented in the following
sections) so that the bad-shaped element is transformed and the movement of node K on the
curved boundary is possible.

MAINTAINING DELAUNAY CONFORMITY

The problem
Given a tetrahedral tessellation, transform it to conform to Delaunay’s property of the empty
sphere.
 DELAUNAY TRANSFORMATIONS 799

element

,,/ Oierlapping Region

Figure 3. The presence of a bad-shaped element leads to the formation of overlapping elements

( ABCD)
(BCDE)
B D
...._______

c
L
C
L

Figure 4. Diagonal swaping in 3-D destroys the tetrahedra boundary and is not allowed

Previous approaches
One way to proceed is by trying to make the Delaunay tessellation near the region of a point by
applying Watson’s algorithm” or variants of it.9*10*12 - 1 4 * 1 6 The process is theoretically strong

but presents some weaknesses: its time complexity is more than linear’6.20 and it needs special
treatment in degenerate cases.’. 1 2 * 1 3 *l 6 Further, it suffers from the formation of Sliuer
“9

elements. Sliver elements are produced in a very large percentage which can account from 5-10
per cent (Reference 9) to more than 40 per cent (Reference 16) of the total number of elements.
In two dimensions the exchange of diagonal technique proposed by Lawson” works very
ell,^*^*^^ producing Delaunay conforming triangular meshes and maximizing the minimum
angle of the mesh.22 This technique does not hold in general in three dimensions. Consider p.e.
the two tetrahedra of Figure 4 which form a non-convex hull. The exchange of diagonal has, as
a result, to alter the external boundary of the tetrahedra and the enclosed volume.

The new approach

A set of three local topological transformations is proposed that transform the given tetrahedra
so that they conform to the property of the empty sphere. The proposed transformations are fast,
800 N. A. GOLIAS AND T. D.TSlBOUKlS

reliable, having no problem with degenerate cases, making no distinction between convex and
non-convex domains and finally handling very easily different material regions resulting in an
‘optimal’ constrained Delaunay triangulation. These transformations are (a)exchange of interface
faces, (b) transformation of two tetrahedra to three and (c) transformation of three tetrahedra to
two, and are described below.

Exchange ojinterface jaces

Two neighbouring tetrahedra (tetrahedra that have a face in common) having two faces lying
on the same surface can be transformed by exchange of the surface diagonal as is shown in
Figure 5(a). Clearly, this is the extension of the 2-D exchange of diagonals technique4 to 3-D and
it is applied on tetrahedra having faces on a surface. Each tetrahedron is checked for the exchange
of interface face. The algorithm is described below.

PROCEDURE Exchange-.Interface. Face.-ofLTetrahedron((ABCD))

BEGIN
T1 = neighbour((ABCD), (A))
IF ((Tl = 0) OR (region (TI))! = region((ABCD))) THEN
/* check neighbour opposite node C*/
(BCDE) = neighbour((ABCD), (C))
IF (region((ABCD))= region((BCDE))) AND
node (E) is in the interior of CircumCircle((BCD)) THEN
[(ABCD), (BCDE)] + [(ABCE), (ACDE)]
ENDIF
IF (no transformation has taken place) THEN
check neighbour opposite node B and D
ENDIF
END IF
END

/’ A IABCEI

Figure 5. Topological Delaunay transformations: (a) exchange of interface faces; (b) transformation of two to three and
vice versa
 DELAUNAY TRANSFORMATIONS 801

Transformation of two tetrahedra to three and vice versa

Two tetrahedra that have a common face and do not satisfy the empty sphere property can be
transformed into three tetrahedra. On the contrary, three tetrahedra that have a common edge
can be transformed into two tetrahedra which satisfy the empty sphere property as is shown in
Figure 5(b). A little attention must be paid here since the exchange is valid only if edge AE is in the
interior space of the two tetrahedra. So a check has to be made as to where this edge belongs.
Transformation is done on an element basis.
PROCEDURE Transform-2-T0-3-ofLTetrahedron((ABCD))
BEGIN
/* check neighbour opposite node A */
(BCDE) = neighbour((ABCD), (A))
I F (region((ABCD))= region((BCDE)))AND
node (E) is in the interior of CircumSphere((ABCD)) THEN
[(ABCD), (BCDE)] -P [(ABCE), (ACDE), (ABDE)]
END I F
IF (no transformation has taken place) THEN
* check neighbours opposite node B, C, D
ENDIF
END

PROCEDURE Transfom-3--T0_2~ofLTetrahedron((ABCE))
BEGIN
/* check edge AE */
(ACDE) = neighbour((ABCD), (B))
(ADBE) = neighbour((ABCD), (C))
IF (neighbour((ACDE), (C))= neighbour((ADBE), (B))) AND
(region((ABCD)) = region((BCDE))) AND
node (E) is in the interior of CircumSphere((ABCD)) THEN
[(ABCE), (ACDE), (ADBE)] -+[(ABCD), (BCDE)]
END IF
IF (no transformation has taken place) THEN
check the other edges of tetrahedron (ABCE)
ENDIF
END
It has to be noted that sometimes these local topological transformations result in a decrease in
the local quality factor. The new tetrahedra formed sometimes have a joint quality factor worse
than the initial. It is a mistake to cancel the topological transformation and go back to the initial
state of better quality factor, since the initial situation does not conform to the Delaunay empty
sphere property and what we get is not a Delaunay tessellation. Although the second state has
a lower quality factor it satisfies the Delaunay property and the tetrahedra can be transformed
again with their neighbours and improve the overall quality factor, resulting finally in a Delaunay
tessellation.

Local transformation of tetrahedron

The above three transformations are engaged in a recursive technique to maintain Delaunay
conformity of the mesh around a tetrahedron’s region. The three transformations (exchange of
802 N. A. GOLIAS AND T. D. TSlBOUKlS

interface face, transform two to three and transform three to two) are applied to the tetrahedron
under consideration. If a transformation has taken place they are applied by a recursive call to the
tetrahedron’s neighbours, to the neighbours of the neighbours, etc., and this chain of transforma-
tions continues until equilibrium is achieved (no more transformations take place). Usually,
equilibrium is obtained in 2-5 steps. Since the neighbours of the tetrahedron are accessed directly
no searching is needed. The algorithm of the recursive procedure is outlined below.

PROCEDURE LocalLTransformation-of-Tetrahedron(T)
BEGIN
Exchange-Interface- Face-of-Tetrahedron(T)
Transform -2-TO-3 -of.-Tetrahedron (T)
Transform- 3-TO -2.- of-Tetrahedron (T)
IF (a Transformation has taken place) THEN
Local_Transformation_of-Tetrahedron(neighbours(T))
END IF
END

All tetrahedra of the mesh are processed with this local transformation recursive technique.
Sometimes it is better to first process the elements having a face on the boundary by the exchange
of interface face transformation and then process all elements with the recursive technique.
Although if the elements are swept again, a few transformations will take place, the successive
application of node relaxation will perturb node positions and Delaunay conformity will be lost.
So one sweep is enough since the major part of the elements is transformed to conform to
Delaunay’s property of the empty sphere.

NODE RELAXATION

Nodes attribute
Node relaxation or node smoothing is a technique of increasing overall mesh quality factor by
suitable reposition of nodes. There are four kinds of nodes according to their position in the
tetrahedral mesh: nodes lying in the interior of the mesh (internal nodes), nodes lying on the
boundary surface between different material regions, or on the outer boundary (surface nodes),
nodes lying on the junction of two interfaces (edge nodes) and finally nodes lying on the junction
of three or more interfaces (corner nodes) as is shown in Figure 6. Every time a new node is
inserted in the finite element mesh its attribute is set automatically. Node relaxation is applied to
internal and surface nodes. Edge or corner nodes are not moved.

The problem
Let us examine the case of a node numbered K and consider all the nodes connected to it. These
nodes form a polygonon if the node is a surface node (a curved polygonon if the node lies on
a curved surface) and a polyhedron if the node is an internal node. The problem is as follows:
Find the position node K must be moved to so that the joint quality factor Qjof the tetrahedra
connected to this node is maximized.
An analytical approach for the solution of the above problem would be impossible and
impractical at the same time. One could apply a Monte Carlo trial and error technique for finding
the best position, but the computational cost would be prohibitive.
 DELAUNAY TRANSFORMATIONS 803

3 2 3

NODES ATTRIBUTE
0 INTERNAL
1 SURFACE

2 ' EDGE
3 ' CORNER

Figure 6. Nodes attribute according to their position in the mesh

Another approach is to move the node to the centroid of the formed polyhedron with
co-ordinates
i n i n i n

where Xi, Yi, Z i are the Cartesian co-ordinates of node i and n is the number of nodes forming the
polyhedron under consideration.
While the application of the above technique (known also as the barycentre of Laplace
technique) to 2-D triangular meshes is very successful,4 the same is not true with 3-D tetrahedral
meshes. In fact, an implementation of this technique in 3-D has resulted in the deterioration of
mesh quality.23 In three dimensions it is very likely that some very bad-shaped elements exist
from the first steps of the mesh generation procedure and thus the movement of node K to the
above point will decrease the joint quality factor, producing even more bad-shaped elements.
Further, it is very likely that this polyhedron is non-convex, so that the reposition of node K out
of the polyhedron's internal region will result in the formation of overlapping elements.

The proposed technique

If the tetrahedra connected to node K are of about the same volume and nearly equilateral then
the reposition of node K to the centroid will maximize the joint quality factor. Since this is the
desired situation, we expect the movement of node K towards the centroid to be the right choice
even when the tetrahedra are of low quality. The proposed technique is to move the node towards
the centroid in varying steps and not directly. When the node is moved, check if the node still
remains in the polyhedron's interior. Further, check if the joint quality factor of the tetrahedra
forming the polyhedron is increased. If the above conditions hold, position the node there;
otherwise, move it half the initial step, perform the same checks, etc. Stop in three iterations. In
this way, the formation of overlapping elements is avoided and the joint quality factor improves
or, at worst, stays the same. The quick formula (4) is used for the calculation of the joint quality
factor to accelerate the procedure.
The flow chart of node relaxation is presented in Figure 7. First a node-to-node connectivity
matrix and a node-to-element connectivity matrix are allocated and formed. By the formation of
these structures no searching is needed since the nodes and the tetrahedra connected to node
K can be accessed directly and the movement of each node takes a definite time making the
804 N. A. GOLIAS AND T. D. TSlBOUKlS

MOVE EXIERNALNODES

WOE CONNECIlVrrY SIUUClURE

NTON CONNECIlVrrY SIUUCTl-lRE
DEALLOCATION

I
END

Figure 7. Flow chart of node relaxation

process O(n),although requiring some additional memory. But by the use of dynamic memory
allocation, this does not constitute a drawback, since the above structures are allocated and
deallocated after use, freeing computer memory, which is then used for the allocation of finite
element matrices, etc. All the nodes are swept and the maximum movement is stored. The nodes
are swept iteratively until the maximum movement is l/lOth of the 1st sweep’s maximum
movement or four iterations have been completed.
In Figure 7 it is seen that node relaxation is applied first on internal nodes and then on surface
nodes. Although node relaxation can be applied at the same time on internal and surface nodes, it
is preferable to relax the internal nodes first, increasing element quality, and then to relax surface
nodes. Even a very small movement of a surface node can affect remarkably the formation of the
tetrahedra that are connected to this node, especially if there are not many nodes distributed on
the surface and many bad-shaped elements exist in the mesh.

MESH STABILIZATION
The key to the formation of high-quality tetrahedra is the repeated application of the Delaunay
transformations and the node relaxation technique. First the topological transformations are
 DELAUNAY TRANSFORMATIONS 805

Figure 8. Flow diagram of the mesh stabilization (iterative application of Delaunay transformations and node relaxation)

applied, directing the mesh to conform to the Delaunay’s property. Then the application of node
ralaxation repositions the nodes and although quality is increased the mesh is disturbed and
Delaunay conformity is lost. A new cycle of topological transformations is needed to bring the
mesh to Delaunay conformity and a new application of node relaxation to increase mesh quality.
These cycles are repeated until convergence is achieved as is shown in Figure 8, where the flow
chart of this iterative stabilization technique is shown. Convergence is attained when a small
number of transformations take place or nodes are moved a small distance. These two situations
go together and usually 5-10 cycles are enough for the mesh to relax.
It should be noted that the actual number of iterations depends on the initial and final quality
of the mesh. For example, the effort to stabilize a mesh with a mean quality Qm from 0.3 to 0.8is
much greater than that required for a mesh with Qm = 0.6. During refinement better and better
meshes are produced, so that as the number of elements is increased the procedure accelerates.

IMPLEMENTATION OF THE REFINEMENT ALGORITHM IN AN

AUTOMATIC FINITE ELEMENT ANALYSIS SYSTEM
The proposed algorithm for refining and stabilizing tetrahedral meshes producing high-quality
tetrahedra can be employed directly in an automatic finite element analysis system as is shown in
Figure 9. The selective refinement generating varying element density and the high quality of the
generated tetrahedra make it ideal for 3-D automatic adaptive mesh generation in a finite element
con text. ’
Initially, a coarse mesh is produced with an automatic mesh generation scheme (Watson’s
algorithm). Boundary conditions and other finite element parameters are set on this initial mesh.
Then the mesh is processed by the finite element method. With the application of an error
estimator, regions that present large errors in the approximation procedure are recognized.
Accordingly, these regions are refined and the mesh is stabilized as is shown in the diagram. The
new mesh is processed again by the finite element method and this procedure is repeated until
certain termination criteria defined by the user are satisfied (accuracy, memory limits, etc.).
806 N. A. GOLlAS AND T. D. TSlBOUKlS

]- PRE - Po-
-
7
PROCESSING _] , ESnMATION
i

Figure 9. Implementation of the refinement algorithm in an automatic finite element analysis system

Table 11. Mean, joint and minimum quality for the

cube problem

Nodes Elements Qm Qj Qminimum

8 6 0.691 0.687 0.623

8 14 3572 0.635 0.589 0.104
814 3822 0.846 0.833 0.351
7485 39843 0,852 0.840 0,185
56470 318583 0.871 0.862 0.132

APPLICATIONS
The refinement algorithm is applied in the generation of an unstructured tetrahedral mesh of
a cube with unit length as is shown in Figure 10. An initial coarse mesh with six tetrahedra and
eight nodes is input to the proposed algorithm. The mean, joint and minimum quality of the
refined meshes are presented in Table I1 where it is seen that the quality factor increases with the
number of tetrahedra and no sliver elements are produced. It is observed that mean qualities of as
high as 0871 are attained, a number difficult to achieve even with triangular element^.^ For
further documentation, a histogram is presented in Figure 11 showing the distribution of element
quality factors for for the meshes of Figures 10(b) and (c), respectively. It is clear that the
application of the stabilization procedure has as, a result, to shift the bulk of tetrahedra to the
right, towards quality factor 0.8-1.0.The CPU time and memory requirements (HP Apollo
 DELAUNAY TRANSFORMATIONS 807

(d) (4
Figure 10. Automatic mesh refinement for a cube of unit length: (a) eight nodes, six tetrahedra; (b) 814 nodes, 3572
tetrahedra (before stabilization); (c) 814 nodes, 3822 tetrahedra (after stabilization); (d) 7485 nodes, 39 843 tetrahedra;
(e) 56 740 nodes, 31 8 583 tetrahedra

Workstation, 76 Mips 16 Mbytes RAM) for the refinement of the cube problem are presented in
Table 111. The application of Watson’s algorithm and variants of it on the same problem16 seems
to result in a large percentage of sliver elements (5-45 per cent). With the proposed algorithm,
sliver elements are avoided completely. Furthermore, the worst tetrahedron has a quality > 0.1
or 0.2, which is far from the value that characterizes a tetrahedron as sliver.
In a second application, an initial mesh with 27 nodes and 44 tetrahedra for a bracket element
is generated with Watson’s algorithm as is shown in Figure 12(a). Selective refinement follows and
the meshes shown in Figures 12(bHd) are generated. The mean, joint and minimum quality
808 N. A. GOLIAS AND T. D. TSlBOUKlS

Number of Tetrahedra
1800

1600

1400

1200

1000

800

600

400

200

0
0-0.1 01-0.2 0.2-0.3 0.3-0.4 0.4-0.6 0.6-0.8 0.8-0.7 0.7-0.8 0.8-0.0 0.0-1

Quality Factor
0Before Stablllzation After 8tablllzation

Figure 1 1 . Quality distribution for the meshes of Figures 10(b)and (c). It is obvious how the bulk of elements is moved
towards the right: (a) 814 nodes, 3572 tetrahedra (before stabilization); (b) 814 nodes, 3822 tetrahedra (after stabilization)

Table 111. CPU time and memory requirements for the cube
problem

Nodes Elements CPU time (min) Memory (Mbytes)

~- - ~ . -~

8 6 -

814 3822 1 :44 0.9

7485 39 843 19:38 3.8
56470 318583 210: 50 35.2

Table IV. Mean, joint and minimum quality for the

bracket element

Nodes
_ _ ~
Elements Qm
__
Q, Qminimum

24 47 0.198 0.075 00133

377 1431 0.679 0.600 0.0954
2776 12428 0.809 0.779 00251
25 000 135862 0.840 0824 0.0234

are presented in Table IV. The quality distribution for the mesh of Figure 12(d) is presented in
Figure 13.
Finally, mesh generation for a sphere in air is presented in Figure 14. The problem presents
some interest due to the direct implementation of the curved-surface approximation algorithm in
 DELAUNAY TRANSFORMATIONS 809

(4 (dl
Figure 12. Automatic mesh refinement for a bracket element: (a) 24 nodes, 47 tetrahedra; (b) 377 nodes, 1431 tetrahedra;
(c) 2776 nodes, 12428 tetrahedra; (d) 25 OOO nodes, 135 862 tetrahedra

Table V. Mean, joint and minimum quality for the

sphere problem

Nodes Elements Qm Qj Qminimum

11 9 0,570 0.318 0.123

287 1235 0.762 0735 0.342
204 1 10559 0.801 0.775 0.236
19721 110528 0.824 0.807 0,119

the refinement procedure. One-eighth of the problem is considered and an initial mesh with
11 nodes and 12 tetrahedra is input to the refinement procedure as is shown in Figure 14(a).Three
refined meshes are presented in Figures 14(bHd).The mean, joint and minimum quality for the
meshes of Figure 14 are presented in Table V.The formation of high-quality tetrahedra and the
810 N.A. GOLIAS AND T. D.TSIBOUKIS

Number of Tetrahedra (Thousands)

60
53.893
50

40

30

20

10

0.046 I
0.011 0.026 L 0.167
0 I
0-0.1 0.1-0.2 0.2-0.3 0.3-0.4 0.4-0.6 0.6-0.8 0.8-0.7 0.7-0.8 0.8-0.9 0.9-1

Quality Factor
Figure 13. Quality distribution for the mesh of Figure 12(d)

Figure 14. Automatic mesh refinement for the sphere problem (external view of the whole mesh and the sphere): (a) I t
nodes, nine tetrahedra; (b) 287 nodes, 1235 tetrahedra; (c) 2041 nodes, 10559 tetrahedra; (d) 19721 nodes, 110528
tetrahedra
 DELAUNAY TRANSFORMATIONS 81 1

--I
Number of Tetrahedra (Thousands)

50
43.024
40

30

20

10

..
n
0-0.1 0.1-0.2 0.2-0.9 0.9-0.4 0.4-0.6 0.6-0.0 0.6-0.7 0.7-0.8 0.8-0.9 0.0-1

Quality Factor
Figure 15. Quality distribution for the mesh of Figure 1qd)

complete elimination of bad-shaped elements makes the approximation of the curved surface
possible. For further documentation the quality distribution for the mesh of Figure 14(d) is
presented in Figure 15.

CONCLUSION
A technique for refining 3-D unstructured tetrahedral meshes has been developed and applied
successfully. The proposed technique presents certain advantages. The generation of sliver
elements is completely eliminated while the tetrahedra produced are of high quality (nearly
equilateral). The elimination of bad-shaped tetrahedra makes possible the approximation of
curved boundaries directly in the refinement process, which would otherwise be impossible.
Tetrahedral meshes of arbitrary structure are handled, making no distinction between convex or
non-convex domains and multiple regions. With the application of the proposed transformations
the boundaries between different material regions are respected and the major part of the
generated tetrahedra conforms to Delaunay’s criterion. Further, the computational requirements
of the proposed technique are in linear relationship with the size of the problem (number of nodes
and tetrahedra), making it very attractive for implementation in an automatic finite element
analysis system. Application to some test problems proves that it is a valuable scheme for the
automatic refinement of tetrahedral meshes.

REFERENCES
1. 0. C. Zienkiewicz and R. L. Taylor, The Finite Element Method, Vol. I, 4th edn, McGraw-Hill, London, 1989.
2. M. Krizek, ‘On the maximum angle condition for linear tetrahedral elements’, S I A M J . Numer. Anal., 29, 513-520
(1992).
3. Z. J. Cendes and D. N. Shenton, ‘Adaptive mesh refinement in the finite element computation of magnetic fields’,
I E E E Trans. Magnetics, 21, 181 1-1816 (1985).
812 N A. GOLIAS A N D -r u. TSIBOUKIS

4. N. A. Golias and T D. Tsiboukis. ‘Adaptive mesh refinement in 2-D finite element applications’. Int .I. Numur.
Modeling, Electronic Networks, Deuices and Fields, 4, 81 -95 (1991).
5. J. Penman and M. D. Grieve, ’Self-adaptive mesh generation technique for the finite element mrthod’, fEE Proc.
134 (Pt.A), 634-650 (19871.
6. A. M Pinchuk and P. P. Silvester, ‘Error estimation for automatic adaptive finite element mesh ge
Truns. Mugnetics. 21. 2551b2554 (1985).
7. N. A. Golias and ‘I.12 Tsiboukis, ‘Three-dimensional automatic adaptive mesh generation’. I EEE Trans. Mugnerics,
28, I7W- I703 ( 1992).
8. A. Bowyer, ‘Computing Dirichlet tesselations’, Comput. J . , 24, 162- 167 (1981).
9. J . C. Cavendish. 0 A ti ield and W. H. Frey, ‘An approach to automatic three-dimensional finite element generation’,
fnr. numer. method., en(/., 21, 329-341 (1985).
10. P. 1.. ( . m ) i p t . ail<: t Hermeline, ‘Delaunay’s mesh of a convex polyhedron in dimension d . Applica
polyhedra: Int. J. numer. methods m y . , 33, 975995 (1992).
1 1 . J. Peraire. M. Vahd.iti. K. Morgan and 0. C. Zienkiewicz, ‘Finite element Euler computations in three dimensions’.
Int. j . numrr. methods eng., 26, 21 35-2 I59 ( 1981).
12. W. J. Schroeder and M. S. Shepard, ‘Geometry-based fully automatic mesh generation and thc l k l a i i t . ~ , triangiila.
tion’, In/ 1. numer. methods eng.. 26, 2503-2515 (1988).
13. W. J. Schroeder and M. S.Shepard, ‘A combined octree/Delaunay method for fully automatic 3-D mesh generation’.
Int. .j numpr. mrthods eng., 2Y, 37-55 (1990).
14. ci. h 3tii r i t u i i and L. J . Cendes, ‘Three-dimensional finite element mesh generation using Delaunay tessellation’,
I E E E Trans. Magnetics, 21, 2535-2538 (1985).
15. M. S. Shepard. ‘Automatic and adaptive mesli grriirution’. !FF:l. TrunA Magnetics, 21, 2484-2489 (1985).
16. S. Kaganathan and N. B. Goldstein, ‘Comparison of four point adding algorithms for Delaunay-type. Three-
dimensional mesh generators’, IEEE Trans. Magnetics, 27, 3444 3451 (1991;.
17. F. Aurenhammer, ‘Voronoi diagrams-a ‘iui vry ol a fundamental geometric data structure‘, ACM Comput. S u r c q s ,
23, 345405 (1991).
18. B. Delaunay, ‘Sur la sphere vide’, Bull. Acad. Sci. USSR, 117, 793-800 (1934).
19. C. A. Rogers, Packing and Covering, Cambridge Tracts in Mathematics and Physics, No. 54, 1964.
20. D. F. Watson, ‘Computing the n-dimensional Delaunay tessellation with application to Voronoi polytopes’, Comput.
.I24,.,167-172 (1981).
21. C. L. Lawson, ‘Generation of a triangular grid with application to contour plotting’, Technical Memorandum 299, Jet
Propulstion Laboratory, California Institute of Technology, 1972.
22. R. Sibson, ‘Locally equiangular triangulation’, Cornput. J., 21, 243-245 (1978).
23. I. B. Albertini, These de Docteur, lnstitut National Polytechnique de Grenoble, 1988.