FULLY AUTOMATIC ADAPTIVE MESH REFINEMENT INTEGRATED

INTO THE SOLUTION PROCESS

Joseph R. Tristano, Zhijan Chen, D. Alfred Hancq, Wa Kwok

ANSYS Incorporated, Canonsburg, PA U.S.A.

{joe.tristano,james.chen,al.hancq,wa.kwok } @ansys.com

ABSTRACT

Finite element analysts and designers need to feel confident in the results of their analyses before sending a product to prototype
or production. Mesh discretization can greatly influence the desired results. In this paper we present framework for adaptive
mesh refinement to obtain FEA results with a desired accuracy. The process involves adaptively refining the mesh based on
solution error norms until the result desired converges to certain accuracy. The adaptive refinement/meshing process must be
fully automatic and very robust. We present an exhaustive method to create a fully automatic and integrated process that takes
advantage of many of the mesh refinement and mesh optimization algorithms found in literature. The results of the process
provide the user with the desired accuracy in the smallest number of iterations possible.
Keywords: h-adaptivity, adaptive mesh refinement, adaptive finite element

1. INTRODUCTION SOLVER COMPONENT solve

Finite element analysts and designers need to feel confident DRIVER COMPONENT YES
converged stop
in the results of their analyses before sending a product to or num iterations met
prototyping or production in order to save their companys
time and money. This is is especially true for users who are get nodelist to be
not experts in FEA. The discretization of the model marked and nodelist to
exclude from
intended for analysis can greatly influence the outcome of refinement
their desired results. In order to feel confident with the
desired results we have developed a fully automatic REFINE COMPONENT
nodes marked > % of total
adaptive mesh refinement process that is integrated into the or PGR key is set

solution process. The mesh is refined based upon error NO YES

norms and other information output by the solver. In order
to make the process reliable and robust, multiple methods delete
unmark nodes that may cause
for refining the mesh have been implemented. The process problems
tetrahedron

also needs to be efficient, i.e., the process needs to mark

enough nodes for refinement such that the iterations are not template Preprocess
refinement surface mesh
wasted in the solver. The importance of this work lies in
the method to integrate the tools and procedures found in
refine surface
literature into a fully embedded automatic and robust SET PGR mesh using
KEY
adaptive solution process. templates

The remainder of this paper will first discuss the overall mesh interior with
tetrahedron and refine
framework of the process. We will then discuss the interior nodes using
templates
previous work in the area on which the building blocks for
the process were built. We will then go into detail of the conventional mesh
sub processes within the process such as methodology for optimization methods

selecting nodes and elements for refinement and the

YES
refinement process itself. We will finally show some quality ok
examples and discuss areas for future work. This paper will
NO
focus on refinement of tetrahedral meshes. However, the
method can easily be extended to other element types. quadratic tetrahedron
improvement

NO YES
quality ok

Figure 1 Overview of Adaptive Solution Process

1.1 Process Overview complex real world geometry and quadratic tetrahedron, this
process is never trivial.
The adaptive solution process is an iterative process
illustrated Figure 1. 1.2.1 Adaptive Refinement
It is clearly seen that this adaptive solution process does not Diaz-Morcillo [7] presented an adaptive mesh refinement
lend itself to failure. Various techniques are used to protect process for electromagnetic problems. Since the elements are
the refine process from producing poorly shaped elements. It usually linear and the geometry is usually simple for this
is imperative that the mesh generated from the refinement class of problems, the framework for the algorithm is allowed
process is of adequate quality. to be quite simple: solve, estimate error, refine. Jones [8]
The process is as follows. The initial solve takes place. The also describes a method very similar to Diazs. His work also
solver driver then determines which nodes must be refined concentrates more on the refinement algorithms than
and which nodes must be excluded from refinement (to creating a fail-safe process that can run on real world
prevent refinement of singularities). Depending on the geometry.
amount of nodes marked, two paths may be taken. The first
path is the template based tetrahedron refinement based on 1.2.2 Template Based Refinement
Statens work. The second path uses a Pseudo Global
Refinement (PGR) method that: deletes the tetrahedron, We have chosen to use Statens [6] (Figure 2) template based
refines the surface mesh and performs some optimization of refinement over existing methods of refinement. This
the surface mesh, re-meshes with tetrahedron and refines any method was chosen over Delaunay[5] refinement because it
marked interior nodes using templates. Extensive tetrahedron has shown the ability to maintain overall mesh quality after
optimization is performed using conventional methods of successive refinements.
B.Joe [1] for each method of refinement. If the conventional
optimization methods fail, more specialized quadratic
tetrahedron optimization methods are called (Kwok et al.
[2]). This process iterates until the solution has converged or
the number of adaptive iterations has been met.
If either method fails, various styles of the PGR algorithm
work to refine the model to create a high quality tetrahedron
mesh.
The processes communicate via data passed through COM Figure 2 Statens Template Based Refinement
Interfaces. These interfaces allow us to maintain each
procedure of the process independently. The interfaces are
implemented in several components (DLLs). The solver
component solves the FE model. The driver component 2. ADAPTIVE REFINEMENT SUBPROCSESES
determines what elements should be refined based on
analysis type specific criterion. This component
communicates between the solver and mesh refiner and 2.1 Methodology for Choosing Candidates for
determines when the adaptive loop should stop. This Adaptive Refinement
component also charts the progress of adaptivity for the user,
which enables the user to track the convergence of the result Adaptive H refinement is supported for structural, thermal,
per refinement iteration along with displaying the number of and modal analysis. The result quantity , the expected
nodes and elements for each iteration. The mesh refiner asks accuracy E (expressed as a percentage), and the region R on
the driver for a mesh object and the nodes and elements to be the geometry that is being subjected (scoped) to adaptive
refined and excluded from refinement. This component is analysis may be selected. The user-specified accuracy is
self-encapsulated and determines whether templates or PGR achieved when convergence is satisfied as follows:
refinement should be executed. If the quality of the output
mesh is not adequate, more refinement and mesh
optimization are performed to increase the quality to (1)
100 i + 1 i < E , i = 1,2,3, K n (in R)
acceptable levels. This componentization also enables quick 1 ( + )
prototyping and implementation of new refinement i +1 i
2
algorithms, solvers, and error estimators.
Elements will be flagged for refinement on some basis
measure. The criterion for which elements and nodes are
1.2 Previous Work selected for adaptive refinement depends on analysis type and
on what results quantities are requested. In addition,
The literature contains many works regarding error estimates
depending on analysis type and the requested result type of ,
[3],[4] and refinement methods[5],[6], however, the methods
nodes may be flagged for refinement to augment the element
used to fully automate and integrate these tools into a robust
list. This is done to help ensure that true convergence is
failsafe algorithm have been trivialized. When dealing with
achieved on the result .
 2.1.1 Element Selection Methodology, element error throughout the element set. Although highly
Structural and Thermal Analysis dependent on the loading and geometry, typical percentages
of elements flagged for refinement range from 5 to 20
A basis measure must be employed to identify poorly shaped percent.
elements that are causing analysis inaccuracies. The
procedure to select the elements for refinement, aside from
the error calculation itself, is the same for both structural and 2.1.2 Element Selection Methodology, Modal
thermal analysis and will be discussed first. For structural Analysis
analysis, a displacement based error measure developed by
Zienkiewicz-Zhu[3] is used as this basis. Essentially an In selecting elements for modal refinement, a different basis
elemental stress error is calculated from the difference mechanism as well as different cutoff technique is employed.
between the average nodal stress and the element nodal The basis function is simply largest element volume. The
stress. This stress error is then integrated over the element element set is sorted on that basis and then a cutoff element
volume to calculate an energy error that is then used as the technique is employed so that a set percentage of the largest
structural error basis. This calculation is performed over the elements are flagged for refinement. Typically, 25 percent of
element set and can be seen mathematically by [9]: the elements will be flagged for refinement but this value is
controllable by the user.
{ ni } = { na } { ni }
where:
2.1.3 Node Augmentation Selection
{ ni } = stress error vectorat node n of elementi
methodology

N en
{ ni }
{ a
} = averaged stress vectorat node n = i =1 Although the element selection methods described above
n
N en perform well in obtaining an accurate global solution, node
N en = numberof elementsconnectedto node n augmentation based on the result may be required to ensure
that a local converged result is achieved. This is applicable
{ ni } = stress vectorof node n of elementi
to results where local refined meshes are required to achieve
an accurate solution. Consider the case where is based on
Then for eachelement: equivalent stress and a stress concentration exists. Consider
1 also that in that region the element error is low enough
ei = { }T [ D]1{ }d (vol) compared to another region such that the other region in the
2 vol
where:
model is flagged for refinement. Thus, as a result, the change
in will be likely small enough to satisfy the requested
ei = energy error for elementi accuracy E, although true convergence has not been reached.
vol = elementvolume To aid in this, a list of nodes on the region R will be flagged
[ D] = stress strain matrixof the element (2) for refinement that have values near and thus ensures that a
{ } = stress error vector false convergence will not occur. This is node augmentation
is only applicable to certain type of results where this false
A similar calculation for temperature-based problems using convergence can occur and includes:
elemental heat flux was given by Huang and Lewis[4] and is
used a basis for thermal energy error. Structural stresses and strains

After all the element errors are collected, a cutoff range Structural post tools based on stress and/or strain
technique is employed where the elements with the largest Structural contact results
error are flagged for refinement. A critical value is calculated
and elements with an error above this value will be flagged Thermal heat fluxes
for refinement according to the equation:
Nodes are selected for refinement based on a cutoff range
e crit = e max (e max e min ) technique. However, since convergence may be applied to
where : (3) either the minimum or maximum of , and in addition, may
e crit = minimum element error required be positive or negative, logic is required to handle each
t o be flagged for refinement permutation. Each node in the set will test the following
logic statement and will be flagged for refinement if the
e max = maximum element error in element set
statement is true:
e min = minimum element error in element set
= cutoff factor, 10 default of .75 Converging on max Converging on min
< 0 (n > ) < 0 (n < )
As can be seen from the equation above, a cutoff factor of
0 (n > ) 0 (n < )
zero would cause only elements with the highest error to be
flagged while a factor of one would flag all elements for where : (4)
refinement. At first glance, the default of .75 would seem = result quantity to be converged upon
to allow a disproportionately large number of elements to be n = result quantity at node n
flagged but experience has shown this value to be valid. This
= cutoff factor, 10 , default of .8
is due to a typically large gradient in the distribution of
Care must be taken when compiling this node list. Artificial
singularities may exist in the model due to the applied loads
that could cause a divergent refinement series. A list of
boundary conditions known to cause these singularities is
compiled and any nodes in the region R that exist on these
boundaries are excluded as candidates for node
augmentation. Note that geometric singularities may exist in
model that the program cannot detect and thus care must be
taken by the user to ensure that a proper region R is selected
for the result quantity where convergence is of interest.

2.2 Refinement Process

2.2.1 Refinement Driver Figure 4 Bending Interior Edges

After the nodes and elements are marked for refinement, the
refinement driver loads the refine component. Data, such as There may be cases where there is not enough room on the
the previous mesh, marked and excluded nodes and elements, interior of the model to bend an interior element edge to
and the BRep are passed into the refine component via a create an acceptable quality element (Figure 4). In this case,
COM Interface (Figure3). we determine if the tetrahedron is a poor element that lies on
the boundary of the topological body. If this poor element
lies on the boundary, a few operations may be performed to
either modify the tetrahedron or simply delete it. Some of
these operators include:
Splitting the longest edge to create better-shaped
elements
Swapping element edges to create a better
configuration
Moving interior corner nodes to improve resulting
element shapes
Deleting the element if it is a kite or cap on the
Figure 3 COM Interface for Refinement Component
boundary
Modifying a mid-side node of the elements on the
2.2.2 Node Marking and Element boundary.
Improvement If all of these methods fail to produce a valid configuration,
The mesh is converted into a local data structure and the the nodes marked on the poor element will be unmarked for
appropriate nodes are marked. During the node marking leg, refinement. Although the driver component wanted the
the algorithm may determine that some quadratic tetrahedron nodes to be refined, the unmarking of nodes has been proven
may invert during the refinement process. At this time a few to be an acceptable practice via empirical studies.
things may happen: Mid-side nodes on element edges
interior to the topological body are perturbed, poor boundary
2.2.2 Template Based Refinement
elements may be deleted, nodes may be unmarked.
The template base refinement follows Statens work with
When bending interior edges, we examine the diagonal of the
slight modifications. There are cases where after some
tetrahedron that is chosen to be refined that will invert the
refinement that the diagonal to refine the tetrahedron upon
children tetrahedron. The mid-side node is perturbed until the
will create invalid elements. In this case, we again try to
children elements will have an acceptable shape.
bend interior mid-side nodes to create an acceptable quality
element.

2.2.3 Pseudo Global Refinement (PGR)

In the cases where template based refinement fails, we try a
new method called Pseudo Global Refinement (PGR). The
PGR method is an innovative technique that utilizes template
based surface mesh refinement and local remeshing
techniques to perform mesh refinement during an adaptive
solution process.
The PGR method is comprised of the following steps.
1. Remember marked interior node locations
(memory nodes) and delete tetrahedron
2. Split boundary edges that have very high
transitions as to improve the quality of the resultant
mesh
3. Improve quality of surface mesh via triangle quality Figure 6 Acceptable Linear Element Becomes
swapping Invalid with the Addition of Mid-Side Nodes

4. Refine marked nodes on surface using template

refinement and clean surface mesh
2.2.3.3 PGR Template Surface Mesh
5. Fill body with tetrahedron Refinement
6. Refine Interior Nodes closest to memory nodes
7. Optimize Mesh

2.2.3.1 PGR Surface Mesh Aspect Ratio

Adjustment
To obtain a successful refinement using the PGR method, the
surface mesh is massaged to prevent poor element quality Figure 7 Template Triangle Refinement
later in the process (Figure 5). This preventative method
reduces high transitions in the mesh by looping through PGR Template triangle refinement uses Statens method [6]
marked nodes and determining if any edge, Ei @ node M, with of refinement to refine the marked nodes (Figure 7). This
length, Li @ node M, exceeds the ratio when compared to the refinement is then followed by topological optimization of
smallest edge length at the node, Lmin @ node M. If this edge the triangles [10] and mixed Laplacian/Optimization
ratio exceeds , the edge is split to reduce the high transition. smoothing[11] of the refined surface mesh.
In practice, we have found that =10.0 is a suitable number.

2.3 Tetrahedron Optimization Methods

After the topological body bound by the refined surface mesh
is filled with tetrahedron, mesh optimization is performed on
the mesh. B. Joes [1] tetrahedron optimization techniques
have been implemented for the initial optimization leg of the
code. For cases where the mesh cannot be improved by Joes
techniques, more advanced and specialized tetrahedron
Figure 5 Splitting High Transition Edges optimization operators are used. These methods consist of the
same methods used to improve tetrahedron quality during the
node marking leg of the process.
2.2.3.2 PGR Surface Mesh Optimization
Improving poor boundary element tetrahedron by reposition
We perform topological optimization of the mesh along with their mid-side nodes is one of the most common specialized
quality optimization of the mesh. The quality optimization of tetrahedron operators called in the process (Figure 8). The
the mesh is required for quadratic elements. An element may operator is quite simple. The smallest or inverted angle is
have acceptable quality when it is linear but adding mid-side found on the boundary face. The unconstrained edge is
nodes to the element may invalidate it (Figure 6). Therefore, determined, i.e., the edge that does not lie on a topological
the quality swap is extremely important. A quality swap is edge is found. The direction of the edge, E , is crossed with
one in which a swap is performed if the quality of the the surface normal, N surface , to determine the direction, Vmove ,
elements will improve in the new configuration. This step
provides a high quality surface mesh to send to the template in which the mid-side node of the moving edge should move.
based surface mesh refinement. This process moves the node a small based on the distance
between the mid-side nodes of the adjacent edges until the
quality of the tetrahedron adjacent to the edge are acceptable.
 8a bad 8b corrected 8c correcting
configuration configuration configuration

Figure 9 Improving Bad Boundary Element

3. EXAMPLES AND FUTURE WORK

Figure 10 Pressure Vessel Initial Mesh

3.1 Example

3.1.1 Pressure Vessel

The following example illustrates our process on a moderate
complexity CAD geometry of a pressure vessel. The
intersections of the fillets in the model contain many sliver
surfaces, which have been defeatured via Mesh Based
Defeaturing [12]. The loading condition is shown in Figure
8. The initial mesh (Figure 10) was intentionally set to the
coarsest setting possible to show the robustness of the
refinement process. It is clearly seen that the initial mesh Figure 11 Convergence Control on Solution
contains many high aspect ratio tetrahedron. The model was
set to adapt on the maximum normal stress of the whole
model with a convergence criterion of 7% allowable change
in the maximum normal stress (Figure 11). After four
iterations, the model converged to .5% change in maximum
normal stress (Figure 12, Table 1). The large increase in the
number of nodes is due to the massaging of the surface mesh
during PGR. PGR allows the complex fillets to be refined
while maintaining the overall quality of the mesh.

Figure 12 Pressure Vessel Convergence

Figure 8 Pressure Vessel Loading Condition

 Table 1 Pressure Vessel Convergence

Solution Number Normal Stress (Pa) Change (%) Nodes Elements

1 1.3980E+07 2080 1056
2 1.9378E+07 32.062 13099 4643
3 2.0886E+07 7.8 35255 22418
4 2.0994E+07 0.51585 123451 83988

Figure 15 Plate with Hole Initial Mesh

Figure 13 Pressure Vessel Converged Normal

Stress

3.1.2 Plate with Hole

The next example illustrates an academic problem of a plate
with a hole. The interesting item with this model is the
scoping of results to the top surface (region R as discussed Figure 16 Plate with Hole Convergence
above). As stated above, the scoping control limits the
marking of nodes to nodes lying on the scoped surface
(Figure 17). The model converges to 1.88 % change in Table 2 Plate with Hole Convergence
maximum equivalent stress in four iterations (Figure 16,
Table2). Solution Number Equivalent Stress (psi) Change (%) Nodes Elements
1 1088.4 438 174
2 1430.5 27.161 2388 1183
3 1502.9 4.9363 12494 7543
4 1531.6 1.8895 35341 22925

Figure 14 Plate with Hole Loading Condition

Figure 17 Plate with Hole Scoped Equivalent Stress

Result

3.2 Future Work

This paper presented a framework for successfully integrating
fully automatic adaptive refinement into the solution process.
This process is quite robust. However, as with any process,
there is always room for improvement. Future work in this International Meshing Roundtable, Sandia National
area may consist of: Laboratories, pp.147-155, (2000 )
Determining better heuristics for when to choose [8] Mark T. Jones and Paul E. Plassmann, Adaptive
PGR over Template refinement Refinement of Unstructured Finite-Element Meshes,
Finite Elements in Analysis and Design, Num 25 pp.
Using an optimization method to place the mid-side 41-60, (1997)
nodes when bending interior edges or moving the [9] ANSYS Inc, Theory Manual
mid-side node of a poor surface tetrahedron [10] Scott A. Canann, S. N. Muthukrishnan and R. K.
Developing new specialized tetrahedron cleanup Phillips, Topological Refinement Procedures for
operators Triangular Finite Element Meshes, Engineering with
Computers, Springer-Verlag, Vol 12, pp.243-255,
Improve refinement for thin models when the (December 1996)
aspect ratio of the opposite side of the model is [11] Zhijian Chen, J. Tristano, W Kwok, Combined
much larger than the side being refined (Figure 18) laplacian and optimization-based smoothing for
quadratic mixed surface meshes, Submitted to 12th
International Meshing Roundtable
[12] Anton V. Mobley, M. P. Carroll, and S. A. Canann,
An object oriented approach to geometry defeaturing
for finite element meshing, In Proc. 7th Intl Meshing
Roundtable '98, Sandia National Laboratories, (1998)

Figure 18 Cross-section of thin model with high

aspect ratio tetrahedron

REFERENCES

[1] Barry Joe, Construction of Three-Dimensional

Improved-Quality Triangulations Using Local
Transformations Siam J. Sci. Comput., Vol 16,
pp.1292-1307 (1995)
[2] Wa Kwok, Z. Chen, J. Tristano, Quadratic
Tetrahedral Mesh Improvements, Submitted to 12th
International Meshing Roundtable
[3] O. C Zienkiewicz,., and J. Z Zhu, "A Simple Error
Estimator and Adaptive Procedure for Practical
Engineering Analysis", International Journal for
Numerical Methods in Engineering, Vol. 24, pp. 337-
357 (1987).
[4] H.C. Huang and R.W. Lewis, Adaptive Analysis for
Heat Flow Problems Using Error Estimation
Techniques, Paper presented at the 6th International
Conference on Numerical Methods in Thermal
Problems. Also University of Wales, University College
of Swansea Internal Report CR/635/89 (April 1989)
[5] Maria-Cecilia Rivara, Local Modification of Meshes
for Adaptive and/or MultiGrid Finite-Element
Methods, Journal of Computational and Applied
Mathematics, Elsevier, Num 36, pp.79-89, (1991)
[6] Matthew Staten, Selective Refinement of Two and
Three-Dimensional Finite Element Meshes, Master's
Thesis, Department of Civil and Environmental
Engineering, Brigham Young University, (1996 )
[7] Alejandro Diaz-Morcillo, L. Nuo, J. V. Balbastre and
D. Snchez-Hernndez, Adaptive Mesh Refinement in
Electromagnetic Problems, Proceedings, 9th