You are on page 1of 11

Comparative Study of Unstructured Meshes

Made of Triangles and Quadrilaterals


Anish Malanthara Walter Gerstle
anishm@arc.unm.edu  gerstle@unm.edu y

Abstract.
Extensive studies have been conducted on the generation of unstructured meshes of triangles for
the purpose of nite element analysis. Delaunay triangulation has been the basis of several such
methods, and has lead to algorithms that can produce unstructured meshes of triangles for any planar
domain. However, there have not been many robust methods for producing unstructured meshes of
quadrilaterals. This is unfortunate, because for plain stress and plain strain problems solved using
nite element analysis, 4-noded quadrilateral elements perform much better than 3-noded triangular
elements, especially when the discretisation is not dense.
In this paper, we present an extremely simple and guaranteed method to generate meshes of quadri-
laterals from meshes of triangles generated using a robust Delaunay triangulation algorithm. The
elements thus obtained have less than optimal aspect ratios. In order to determine if the less than
optimal aspect ratios of the elements a ect the quality of results when these meshes are used for
nite element modeling, we applied these meshes to the solution of a problem. We are of the opinion
that a mesh of quadrilaterals generated in the inexpensive manner presented in this paper produces
results that compare well even with those from mapped meshes. Thus, the presented method provides
an ecient, simple and e ective way to generate unstructured meshes of quadrilaterals of reasonable
quality.
Keywords. Delaunay unstructured quadrilateral triangular mesh nite element mapped

1 Background
In the past, a number of algorithms for generation of unstructured triangular meshes have been reported
[Chew 89, Cavendish 83, Ruppert 92] (Structured or mapped meshes are developed through mappings of a
mesh de ned in a logical domain into a geometric domain, whereas unstructured meshes do not depend upon
a topologically similar logical domain. However, it has been the analysts' experience that in nite element
analyses, for a given number of degrees of freedom (DOF), 4-noded quadrilateral elements provide better
results than 3-noded triangular elements [Cook et al 74]. The 3-noded constant strain triangle behaves
poorly in bending because the stresses and strains are constant within the element. However, the results
improve when the mesh becomes ner [Cook et al 74].
There have been attempts to come up with algorithms yielding good-quality meshes of quadrilaterals. The
Paving algorithm is one of them [Blacker 90]. There are a number of algorithms that produce mapped
(structured) meshes of quadrilaterals. However, a disadvantage of mapped meshes is that they cannot
be adapted to t any region. There have been quite a few algorithms to improve the quality of meshes
 Graduate student, Dept. of Civil Engineering, University of New Mexico
y Associate Professor, Dept. of Civil Engineering, University of New Mexico
Triangle Three Quadrilaterals

Figure 1: Conversion of a triangle into quadrilaterals

of triangles too, and sometimes these approaches talk about using meshes of quadrilaterals in place of
triangles[Cannan 96, Mavriplis 97].

1.1 Investigations Using Unstructured Meshes


Although unstructured quadrilateral mesh generation is still in its developing stage, much research has gone
into improving the quality of unstructured triangular meshes, especially the ones produced using Delaunay
triangulation [Cavendish 83, Bern 90, Chew 89, Ruppert 92]. Arbitrary planar regions can be meshed using
an unstructured mesh generated using Delaunay triangulation [Cavendish 83]. The density and quality of
elements in the region may be predictably controlled.
A mesh made of quadrilateral elements can be derived from a mesh made of triangular elements. The
conversion is as shown in Figure 1. The conversion is simple, and involves placing nodes at the centroid of
the triangle and the midpoints of the edges, and then joining them appropriately to form three quadrilaterals.
It is noted that the conversion time is small (proportional to the number of elements in the original triangular
mesh) compared to the creation time of the triangular mesh itself. One also obtains increased number of
degrees of freedom (DOF) and a ner mesh for a quadrilateral mesh derived in this manner.
Thus, with a proven triangular mesh generation technique, and a proven method of deriving quadrilateral
elements from triangular elements, one obtains a proven method for obtaining an unstructured quadrilateral
mesh of any planar region. However, the quadrilateral elements thus developed all have less than optimal
aspect ratios. The question we try to answer in this study is whether these less-than optimal aspect ratios
a ect the results obtained when the meshes mentioned above are used in nite element analyses. The
objective of this study is to apply the meshes created in this manner to the solution of a problem with a
clear analytical solution and to study the convergence of results.
The problem of a cantilever beam loaded with a uniformly distributed shear traction at the free end is chosen
for the discretisation and analysis. Vertical displacement at the top of the free end is the response that is
considered for this study. The geometry and dimensions of the problem are as shown in Figure 2.
The material assumed is linear elastic, with the properties given below:

Young's Modulus (E) - 29000 lb in?2


Poisson's ratio () - 0:3
cantilever beam, with a thickness of 1" in the direction perpendicular to this plane 10"

P = 1 lb
1"

l = 70"

Figure 2: The cantilever beam problem

2 Mesh Generation and Finite Elements Used for the Study


The algorithm we used for generating unstructured triangular meshes through Delaunay triangulation takes
as input a value called , used to control the the mesh density. The implementation of this algorithm was
done using an object-oriented toolkit for computational mechanics research (CoMeT) being developed at the
Dept. of Civil Engineering, University of New Mexico [Atencio et al 97]. The smallest edge of the planar
geometry is divided into subdivisions of length l  , l being the length of that edge. All the other edges are
also divided into subdivisions the same length as subdivisions on the smallest edge. The meshing proceeds
from there, introducing nodes in the interior of the geometric domain as required [Cavendish 83].
First and second order Gauss integration schemes were used for the elements, although in the case of 4-noded
quadrilateral elements it was found that rst order integration results in under-integrating the elements and
may result in hourglass modes [Cook et al 74]. For 3-noded triangular elements, the rst-order scheme is
sucient to give the exact integration results.
A few structured meshes were also created using quadrilateral elements and triangular elements. The results
obtained using these meshes were also compared with the results obtained using the Delaunay meshes
of triangles and quadrilaterals. For purposes of comparison and error estimation, 'accurate' answers were
obtained using an appropriate analytical solution and using a very ne mesh of mapped rectangular elements
and using 8 noded quadrilateral nite elements.

3 Analytical Solution
The analytical solution for the vertical displacement at the top of the free end of a cantilever beam loaded
at the free end is made up of two parts:

1. Simple bending displacement - this is calculated using Timoshenko beam theory.


2. Displacement due to shear.
This solution too is approximate because the beam considered is not exactly a shallow beam, although
the inclusion of shear displacement corrects this approximation to some degree. In any case, the value of
displacement obtained analytically is within 1% of a re ned nite element analysis result, indicating that
the values obtained by both methods are probably accurate to within 1%.
3.1 Vertical Displacement Due to Bending
The cross-section of the beam in the problem is as shown in Figure 2. Using the values given in the problem
section,
Moment of Inertia of the cross ? section; I = bh
3

12 (1)
= 1 1210 = 10
3 3

12
The vertical displacement at the free end due to bending is given by:
b = 3EI Pl3 (2)
where E is the Young's Modulus of elasticity of the material of the beam. Thus,
b = 1  703
3  29000  10123
b = 4:731  10?2 in
3.2 Vertical Displacement Due to Shear
According to Equation. 6.52 in [Wempner 95], the de ection of a beam due to shear is given by the formula
ds = V (x)Q0 (3)
dx GIt
s - vertical displacement due to shear
V (x) - shear at point x
where GQ0 - moment of the cross-sectional area about its centroid
- shear modulus
I - moment of inertia of the cross-section
t - thickness of the cross-section
Using the values calculated from the problem data, subject to the boundary condition that s = 0 at the
xed end (x = 0), it is determined that
ds = 1:3448  10?5 (4)
dx
s = 1:3448  10?5x (5)
Then, x = 70 gives s = 0:0941  10?2 in:

3.3 Total Vertical De ection : Analytical Result.


The total vertical de ection is the sum of the components mentioned above.
 =  b + s (6)
= (4:731 + 0:0941)  10?2 in
= 4:8251  10?2 in
As it will be shown later, this is within 1% of the value of vertical de ection obtained through a very re ned
nite element analysis.
# element gauss num. of num. of num. of v (in) mesh
order nodes elements DOF from FEA type
1 3-noded triangle 1 22 20 44 0.01518 Delaunay
2 3-noded triangle 1 28 26 56 0.01603 Delaunay
3 3-noded triangle 1 60 76 120 0.03378 Delaunay
4 3-noded triangle 1 206 330 412 0.04396 Delaunay
5 3-noded triangle 1 329 548 658 0.04536 Delaunay
6 3-noded triangle 1 444 762 888 0.04610 Delaunay
7 3-noded triangle 1 1088 1972 2176 0.04719 Delaunay
8 4-noded quadrilateral 1 83 60 120 0.04920 Delaunay
9 4-noded quadrilateral 1 107 78 214 0.04967 Delaunay
10 4-noded quadrilateral 1 271 228 542 0.04807 Delaunay
11 4-noded quadrilateral 1 1071 990 2142 0.04796 Delaunay
12 4-noded quadrilateral 2 83 60 120 0.03433 Delaunay
13 4-noded quadrilateral 2 107 78 214 0.04027 Delaunay
14 4-noded quadrilateral 2 271 228 542 0.04545 Delaunay
15 4-noded quadrilateral 2 1071 990 2142 0.04736 Delaunay
16 3-noded triangle 1 74 120 148 0.03720 Mapped
17 3-noded triangle 1 267 480 534 0.04466 Mapped
18 3-noded triangle 1 1013 1920 2026 0.04706 Mapped
19 4-noded quadrilateral 2 44 30 88 0.03976 Mapped
20 4-noded quadrilateral 2 147 120 294 0.04550 Mapped
21 4-noded quadrilateral 2 533 480 1066 0.04731 Mapped
22 4-noded quadrilateral 2 1111 2110 2222 0.04779 Mapped

Table 1: Discretisations used in the study

4 Finite Element Analysis of the Problem


As mentioned before, the problem was solved using the nite element method. The triangular meshes gen-
erated using Delaunay triangulation, quadrilaterals derived from Delaunay triangulation, and a few mapped
meshes of triangles and quadrilaterals were used for this purpose.

4.1 Unstructured (Delaunay) Triangular Meshes


The meshes made using Delaunay triangulation are given in Figure 3. They were prepared using various
values. The values 0.8, 0.6, 0.4, 0.2, 0.15 and 0.13 were used for triangular meshes. A smaller results in a
ner mesh. The value , as it can be seen, is related to the size of the elements, but arriving at a relation
between the two would not be trivial.

4.2 Unstructured (Delaunay) Quadrilateral Meshes


Quadrilateral meshes created from triangular meshes using Delaunay triangulation are given in Figure 4.
The additional e ort involved in converting a triangular mesh to a quadrilateral mesh is minimal, and is
proportional to the number of elements in the original triangular mesh, as far as time for computation is
concerned.
Y

3-noded triangular mesh created by Delaunay triangulation: 20 elements


22 nodes, beta = 0.8. discretisation # 1
Y

3-noded triangular mesh created by Delaunay triangulation: 26 elements


28 nodes, beta = 0.6. discretisation # 2
Y

3-noded triangular mesh created by Delaunay triangulation: 76 elements


60 nodes, beta = 0.4. discretisation # 3
Y

3-noded triangular mesh created by Delaunay triangulation: 330 elements


206 nodes, beta = 0.2. discretisation # 4
Y

3-noded triangular mesh created by Delaunay triangulation: 548 elements


329 nodes, beta = 0.15. discretisation # 5
Y

3-noded triangular mesh created by Delaunay triangulation: 762 elements


444 nodes, beta = 0.13. discretisation # 6
Y

3-noded triangular mesh created by Delaunay triangulation: 1972 elements


1088 nodes, beta = 0.08. discretisation # 7

Figure 3: Delaunay triangulation meshes: 3-noded triangles


Y

4-noded quadrilateral mesh created using Delaunay triangulation: 60 elements


83 nodes, beta used for underlying triangular mesh = 0.8. discretisation # 8,12

4-noded quadrilateral mesh created using Delaunay triangulation: 78 elements


107 nodes, beta used for underlying triangular mesh = 0.6. discretisation # 9,13

4-noded quadrilateral mesh created using Delaunay triangulation: 228 elements


271 nodes, beta used for underlying triangular mesh = 0.4. discretisation # 10,14

4-noded quadrilateral mesh created using Delaunay triangulation: 990 elements


1071 nodes, beta used for underlying triangular mesh = 0.2. discretisation # 11,15

Figure 4: Delaunay triangulation meshes: 4-noded quadrilaterals

4.3 Structured Triangular and Quadrilateral Meshes


For comparing results, several mapped triangular and quadrilateral meshes were made. These meshes were
also used in the study in order to see if the mesh being mapped had any e ect in the convergence of
results obtained. For error calculations, the results obtained using 8-noded quadrilaterals on one of the
fairly re ned mapped meshes were used as `exact' values, after verifying against analytical calculations. The
mapped meshes are given in Figure 5 and Figure 6.

5 Results From Finite Element Analysis, and Discussion


The vertical displacement at the top of the loaded side of the beam was obtained from all discretisations.
The results were then plotted against the number of degrees of freedom for each discretisation. Table 1
shows the details of each discretisation, and the 'number' assigned to each discretisation.
Discretisation #19 (Figure 6) was converted into a mesh with 8-noded quadrilateral elements, and analyzed.
Y

mapped 3-noded triangular meshes: 74 nodes,


120 elements. discretisation # 16

mapped 3-noded triangular meshes: 267 nodes,


480 elements. discretisation # 17

mapped 3-noded triangular meshes: 1013 nodes,


1920 elements. discretisation # 18

Figure 5: Structured meshes: 3-Noded triangles


Y

mapped 4-noded quadrilateral meshes: 44 nodes,


30 elements. discretisation # 19

mapped 4-noded quadrilateral meshes: 144 nodes,


120 elements. discretisation # 20

mapped 4-noded quadrilateral meshes: 533 nodes,


480 elements. discretisation # 201

Figure 6: Structured meshes: 4-noded quadrilaterals


This gave a vertical displacement of  = 0.04793 in. This value of  is within 0.67% of the value obtained
analytically, and has been used as the `exact' answer for error calculations. This value may be compared
with the theoretically estimated result of  = 0.048251 in.
Discretisations #7 through 10, although included in the study, are not to be considered reliable. This is
because when a 4-noded quadrilateral element is used with a rst degree Gauss integration, it has been
shown to exhibit hourglass instability. Plots of the deformed mesh con rm this in the case of the beam
considered. The plots of vertical displacement and percentage error v=s # DOF are given in Figure 7 and
Figure 8 respectively.
It may be seen from Figure 7 and Figure 8 that in general the Delaunay meshes do not perform as well
as mapped meshes with comparable number of degrees of freedom. However, it is obvious that a Delaunay
mesh of quadrilaterals outperforms a Delaunay mesh of triangles when the number of degrees of freedom
(DOF) involved are low (up to  400). For higher number of DOF, both meshes give similar results. It may
be seen that a Delaunay mesh of quadrilaterals which takes about the same e ort to generate as a Delaunay
mesh of triangles, performs better than the latter in general.

6 Conclusions
The unstructured meshes performed satisfactorily when the meshes were dense. The Delaunay meshes of
quadrilaterals gave results comparable to the ones obtained using structured meshes at almost all mesh
densities. However, it was noted that in order to reach the same accuracy as a structured mesh, the
Delaunay meshes required a higher number of DOF, although the di erence in error at any stage was only
a few percent. In the case of the 3-noded triangles, it was noted that the Delaunay meshes seem to perform
better than the mapped meshes.
In the comparisons given above, the generation time also has been included in the `cost' of a mesh. Comparing
between triangular and quadrilateral Delaunay meshes, it was found that the quadrilateral meshes performed
better than triangular meshes at smaller numbers of DOF. At higher numbers of DOF, the di erence in results
became small. However, it has to be remembered that the e ort involved in generating a Delaunay mesh of
quadrilaterals of a certain neness is much less than the e ort required in generating a Delaunay mesh or
triangles of similar neness.
It is concluded that meshes of quadrilaterals generated from a Delaunay mesh of triangles do compare
well even with mapped meshes. The creation of meshes of quadrilaterals from meshes of triangles may be
employed when one wishes to make use of quadrilateral elements but does not wish to create a mesh of
quadrilaterals from start or the region to be meshed is complicated and generation of a structured mesh is
dicult. This is an advantage that the Delaunay meshes of quadrilaterals have over structured meshes of
quadrilaterals.
It is hoped that from this study one can get insights into how the meshes of triangular and quadrilateral
elements behave, and at which stage should one opt for a mesh of quadrilaterals created either using a
specialized algorithm or use a structured mesh if possible. The authors feel that the mesh of quadrilaterals
created from a mesh of triangles in the inexpensive manner mentioned above gives good results, and the
method may be adopted whenever one wishes to make use of the ecient triangular mesh generators and
take advantage of the quadrilateral elements' better performance in nite element analyses.
In the study conducted, only one test problem has been considered. Alhough the problem we chose is a
simple one but brings out the problems with 3-noded triangular elements, other problems could conceivably
show di erent results.
vertical displacement at the top of free end
against the number of degrees of freedom

8 9
0.050 10
11
22
6 21 7
14 5 18 15
vertical displacement (in) 20
4
17
19 13
0.040
16

3 12

0.030
3-Noded Triangular Elements, Gauss Order 1 (Delaunay)
4-Noded Quadrilateral Elements, Gauss Order 1 (Delaunay)
4-Noded Quadrilateral Elements, Gauss Order 2 (Delaunay)
3-Noded Triangular Elements, Gauss Order 1 (Mapped)
0.020 4-Noded Quadrilateral Elements, Gauss Order 2 (Mapped)
exact
2
1

0.010
0.0 1000.0 2000.0 3000.0
number of DOF

Figure 7: Vertical displacement at loading point v/s # DOF

percentage error in vertical displacement at the top of free end


against the number of degrees of freedom
20.0

9
8
10
11
0.0 21
22
5 6 15 7
20 14
18
4
17

19 13
percentage error

-20.0
16

3 12

-40.0
3-Noded Triangular Elements, Gauss Order 1 (Delaunay)
4-Noded Quadrilateral Elements, Gauss Order 1 (Delaunay)
4-Noded Quadrilateral Elements, Gauss Order 2 (Delaunay)
3-Noded Triangular Elements, Gauss Order 1 (Mapped)
-60.0 4-Noded Quadrilateral Elements, Gauss Order 2 (Mapped)
2 exact
1

-80.0
0.0 1000.0 2000.0 3000.0
number of DOF

Figure 8: Percentage error in vertical displacement at loading point v/s # DOF


7 Acknowledgements
The nancial support provided by Sandia National Laboratories contracts AU-9606 and AT-9325; the Center
for Autonomous Control in Engineering (ACE); the Maui High-Performance Computing Education and Re-
search Center; the Waste Management Education and Research Consortium (WERC); and National Science
Foundation Award No. EEC-9322063 is gratefully acknowledged.

References
[Atencio et al 97] Atencio, Lance., Gerstle, Walter H., Panthaki, Malcolm J., Sahu, Raikanta. Object-
Oriented Design and Development of a Computational Mechanics Toolkit, pp 73-79,
Proceedings, URC-TC 1997.
[Bathe 96] Bathe, Klaus-Jurgen., (1996), Finite Element Procedures. Englewood Cli s, NJ:
Prentice-Hall.
[Bern 90] Bern, Marshall., Eppstein, David., Gilbert, David., Provably Good Mesh Generation,
1990 Symposium on Foundations of Computer Science.
[Blacker 90] Blacker, Ted D., Stephenson, Michael B. (1990) Paving: A New Approach to Auto-
mated Quadrilateral Mesh Generation, Sandia Report # SAND90-0249. UC-705.
[Cannan 96] Cannan, S.A., Muthukrishnan, S.N., Phillips, R.K., Topological Re nement Proce-
dures for Triangular Finite Element Meshes, Engineering With Computers, v.12(#3-
4), pp. 243-255 1996 .
[Cavendish 83] Cavendish, James C., Frey, WilliamH. (1983) Fast Planar Mesh Generation Using the
Delaunay Triangulation, Research Publication GWR-4555, General Motors Research
Laboratories, Warren, MI.
[Chew 89] Chew, L. Paul, Guaranteed-Quality Triangular Meshes, (1989) Publication No: TR-
89-983, Department of Computer Science, Cornell University, Ithaca, NY.
[Cook et al 74] Cook, Robert D., Malkus, David S., Plesha, Michael E., (1974) Concepts and Appli-
cations of Finite Element Analysis, Edn. 3, New York: John Wiley & Sons.
[Mavriplis 97] Mavriplis, D.J., Unstructured Grid Techniques, Annual Review of Fluid Mechanics,
v. 29, pp. 473-514, 1997.
[Ruppert 92] Ruppert, Jim. (1992) A New and Simple Algorithm for Quality 2-Dimensional Mesh
Generation, Report No: UCB/CSD 92/694, University of California, Berkeley, CA.
[Wempner 95] Wempner, Gerald. (1995) Mechanics of Solids, Section 6.6, De ection Due to Shear,
Boston, MA: PWS Publishing Company.

You might also like