ABSTRACT
This paper presents a mesh quality metric for hexahedral and wedge elements. A series of studies was carried out to relate the
proposed metric with the solution errors. During the studies, the metric was also compared with the aspect ratio, the warping
factor and the control number. The results show that the metric can be used to identify invalid as well as potential poor quality
elements, which may cause larger solution errors.
Keywords: mesh quality, mesh metric, hexahedral, wedge, mixed elements
1. INTRODUCTION
Recently, there have been great amounts of progress in
hexahedral mesh generation research. Several approaches [14] for various types of geometric domains were successfully
developed. However, these methods still cannot be directly
applied to general complicated models or assembly parts,
which have more than one body or volume. To overcome this
difficulty, both tetrahedral and hexahedral mesh generation
techniques are used together. First, the bodies or regions,
whose topologies are hexahedral meshable or which may be
critical to the analysis, are filled with hexahedral elements [57]. The rest of the domains are meshed with tetrahedral
elements. In order to make sure conformability between
these two different element shapes, authors [8] proposed to
use pyramids as transition elements to connect different types
of elements. The final mesh contains not only tetrahedral and
hexahedral elements but also pyramids and wedges. Figure 1
shows an assembly part, which has two solid cylinders on top
each other contained inside a larger hollow cylinder. The
inside cylinders are meshed with hexahedra, while the outside
is filled with tetrahedra due to the topology difficulty. The
pyramids are used as transition elements between tetrahedra
and hexahedra. Figure 2 shows a portion of interfaces
between these two different types of elements
Before this type of mixed element meshes can be used for
analysis, the first question which we may ask ourselves is:
Do the meshes contain any invalid or bad (poor quality)
elements, which may cause invalid or incorrect finite element
solutions? In order to answer this question, we need some
kind of metrics to measure the quality of different types of
elements. The metrics should be co-related each other. In
other words, they should have the same or similar physical
Q=
J min
R (1 Wmax ) J
max
J min 0
J min > 0
(1)
Aspect ratio for this element is equal to the abovecalculated ratio multiplied with a factor, which is used to
normalize the ratio between [0, 1]. The value of one
means a regular element. This factor depends on the
element shape.
2.
Move an edge along and off the edge (Along Edge / Off
Edge).
3.
Move a face diagonal on and off the face (On Face / Off
Face)
4.
The beam was meshed with 20-node hexahedra, whose midnodes located at the middle of each element edge. The mesh
of the model is shown in Figure 10. The mesh was created in
such a way that the corresponding solution had less than
0.05% errors. The centerpiece of the model is the part (Figure
11), which was under study. The elements in this region were
uniform mesh.
3. METRIC EVALUTATION
Generally, a better quality mesh means a better finite element
solution, i.e. lower solution errors. A shape quality metric
should be able to inform users if there are any elements in a
given mesh, which may likely introduce higher errors in the
solution. For this reason, a simple problem is used to
evaluate the proposed mesh quality metric Q in terms of the
solution errors. The basic idea is to modify the shapes of
some selected elements in various ways and then observe the
Modify the shape of a selected (key) element with predefined manners outlined in the previous section, i.e.
(a) Move one corner node in X, Y, and Z directions.
(b) Move one edge in X, Y, and Z directions.
(c) Move two diagonal nodes in the three directions.
(d) Rotate one element face.
2.
The same set of the data (stress errors) was also plotted
against the aspect ratio, the warping factor and the control
number (CN) [14], as shown in Figures 17~19. The aspect
ratio ranges from one to infinity [17]. In Figure 17, the range
was normalized between 0 and 1 for a better comparison.
The normalized range for the warping factor was also used to
plot the data. The control number ranges from 3 to infinity
[14], which is difficult to make a fair comparison with the
other metrics. For this reason, it was also normalized between
0 and 1, called the modified control number. It can be seen
that the aspect ratio and warping factor do not show as a clear
relationship as the proposed metric shown in Figure 16.
Figure 19 shows that the modified CN values relatively
uniformly scattered along most of the range. There is no a
clear correlation between the solution errors and the modified
control numbers. For instance, when the modified CN is 0.8,
some the errors are still higher and the errors at this value are
no significantly different from the values at 0.4. In other
words, when the modified CN approaches to 0.8 from 0.0, the
solution errors are not necessarily getting lower.
One important fact from this test is that both the aspect ratio
and the warping factor failed to recognize the invalid
elements. In this test, there were 96 cases, which had
negative Jacobian determinants at the corner nodes. These
negative determinants resulted in zero values of the proposed
metric and the modified control number. Further studies
show that three of these cases had negative Jacobian
determinants at the Gauss points, which generated invalid
finite element solutions. The solutions from these three cases
must be thrown away. Table 1 summaries the values of the
various shape metrics for negative Jacobians at Gauss points.
From Table 1 and Figures 17 and 18, it can be seen that for
invalid elements, the values of the aspect ratio as well as
warping factor did not show particularly different. In other
words, these two shape measures cannot identify invalid
elements.
One interesting fact is that when sampling at the corner
nodes, there were more negative Jacobians than at the Gauss
points, which resulted in the zero values of the metric. It
seems that using the corner nodes to sample is too
conservative or overkill. As pointed out before, the proposed
metric is designed as a guideline to identify poor elements,
which likely cause higher solution errors. It is believed that
this kind of the conservative approach is acceptable.
Modified
Control
Number
Aspect
Ratio, R
Warping
Factor, Q
Stress Errors
(%)*
0.42604
0.106929
51.04
0.60890
0.078586
18.46
0.59004
0.074114
15.48
The same set of the data (stress errors) was also plotted
against the aspect ratio, the warping factor and the modified
control number, as shown in Figures 22 ~ 24. The
normalized ranges of these shape measures were used for a
better comparison. It can be seen that the aspect ratio and the
warping factor do not show as a clear correlation as the
proposed metric shown in Figure 21. Figure 24 illustrates
how the modified control number varies with the solution
errors. Unlike the hexahedral case, the control number shows
a similar trend as the proposed metric. From Figures 21 and
24, the latter one has a slightly clearer relationship with the
errors.
Like the hexahedral case, the aspect ratio and warping factor
failed to recognize the invalid elements, which produced
unusable solutions. Also, using the corner nodes, as the
Acknowledgments
The authors would like to thank all the members of the Ansys
meshing team and B. Joe for many helpful suggestions.
[6]
[7]
[8]
[9]
REFERENCES
[1]
[2]
[3]
[4]
[5]