Chapter 4

Controlling
Discretization Errors

Modeling errors, discretization errors, and solution errors, summarized in Figure 2.10,
all affect FEA results. However, only discretization errors are specific to the FEA and
only discretization errors can be controlled using FEA tools. For this reason, we will
discuss them first and the discussion of other errors will follow.

The objective of discretization error control is to find out how is the data of interest such
as displacement or stress, dependent on the choice of discretization (mesh). Note that the
objective is not to minimize the discretization error but to obtain a solution where the
data of interest do not significantly depend on the choice of discretization. The solution
cannot be considered as reliable unless we have a reliable estimate of the discretization
error. The analysis of a discretization error is done in a convergence process.

An example of different choices of discretization is shown in Figure 4.1 where the same
model is meshed with four meshes, each one with a different element size.

35

6650_Book.indb 35 10/31/16 10:31 AM

 Chapter 4

Figure 4.1 The same model meshed with shell elements of different sizes; this illustrates a
global mesh refinement.

4.1 Presenting Stress Results

Before we discuss convergence process and its variations, we need to look at different
ways of presenting stress results.

As we already know, nodal displacements are computed first. Strains and then stresses
are calculated based on displacement results. Stresses are first calculated inside the
element at certain locations called Gauss points. Next, stress results are extrapolated

36

6650_Book.indb 36 10/31/16 10:32 AM

 Controlling Discretization Errors

to the elements’ nodes. If one node belongs to more than one element (which is always
almost always the case), then the stress results from all the elements sharing a given
node are averaged and one stress value, called a node value, is reported for each node.
This stress value is called a nodal stress or averaged stress.

An alternate procedure to present stress results is by obtaining stresses at Gauss points,

then averaging them in between themselves. This means that one stress value is calcu-
lated for the entire element. This stress value is called an element stress or non-averaged
stress because stresses are not averaged between neighboring elements. Gauss points are
locations used in numerical generations of the element stiffness matrix; the location of
Gauss points in an element is schematically shown in Figure 4.2.

Figure 4.2 Schematic locations of the Gauss point in the second-order 2D element. The
red dots denote element nodes, and the black dots denote Gauss points.

Nodal stresses are used more often because they offer smoothed out, continuous stress
results. However, examination of element stresses provides an important feedback on
the quality of the results. If element stresses in two adjacent elements differ too much,
it indicates that the element size at this location is too large to properly model the
stress gradient.

Nodal and element stresses in mesh 1 from Figure 4.1 are shown in Figure 4.3. The
model is restrained along the left vertical edge and subjected to a uniform tension
applied to the right vertical edge. Large elements are used for clarity of this illustration.

37

6650_Book.indb 37 10/31/16 10:32 AM

 Chapter 4

Figure 4.3 Nodal stress plot (top) and element stress (bottom) produced by a coarse
mesh of second-order elements. The model represents a hollow plate in tension. Large
elements are used for clarity of illustration; these elements are too large to produce any
meaningful results.

4.2 Types of Convergence Process

Discretization errors can be found in the process of making systematic changes to the
choice of discretization and studying the impact of those changes on the data of interest.
The process is called the convergence process; there are many variations in the conver-
gence process.

4.2.1 h Convergence by Global Mesh Refinement

On way to make systematic changes to the choice of discretization is to modify the
element size through mesh refinement. As h denotes the characteristic element size
as shown in Figure 3.8, the convergence process through mesh refinement is called h
convergence process. In this process, the size of elements is gradually reduced. We will
illustrate it with the already familiar example of the hollow plate. The plate is supported
along the left vertical edge and loaded with the tensile load uniformly distributed along
the right vertical edge. Figure 4.1 shows the model meshed with four different meshes.
With every mesh refinement, the element size is halved, meaning that each triangular
element is replaced with four smaller triangular elements.

38

6650_Book.indb 38 10/31/16 10:32 AM

 Controlling Discretization Errors

Step 1 in the h convergence process uses the element size of 20 mm. Step 4, which is the
last step, uses the element size of 2.5 mm. The displacement results are summarized in
Figure 4.4 and von Mises stress results are summarized in Figure 4.5. Shell elements
were used to produce these results, but as the model is flat, loaded in plane, and thin, the
2D plane stress element could have been used as well.

Figure 4.4 Displacement results produced by four different meshes representing the
same problem; this illustrates the h convergence process by global mesh refinement.

39

6650_Book.indb 39 10/31/16 10:32 AM

 Chapter 4

Figure 4.5 Von Mises stress results produced by four different meshes representing the
same problem. this illustrates the h convergence process by global mesh refinement.

Having performed four iterations in the h convergence process, we can summarize the
results in Table 4.1.

40

6650_Book.indb 40 10/31/16 10:32 AM

 Controlling Discretization Errors

Table 4.1 Summary of the displacement and stress results produced in four steps of the h convergence
process
Max von
Element Max horizontal Displacement Mises Stress
Mesh size # of # of displacement convergence stress convergence
# mm 1/h nodes elements #DOF mm error MPa error
1 20 0.05 194 78 1098 0.11723 unknown 308.8 unknown
2 10 0.1 822 374 4806 0.11792 0.59% 297.1 3.94%
3 5 0.2 3142 1498 18606 0.11821 0.25% 337.9 12.07%
4 2.5 0.4 12284 5596 73218 0.11824 0.03% 364.3 7.25%

Notice that the total number of degrees of freedom is not exactly equal to the number of
nodes times the number of degrees of freedom per node (here 6 DOF/ per node) because
some degrees of freedom are eliminated by the support and, therefore, not counted. The
data of interest (here the maximum displacement and the maximum von Mises stress)
may be plotted against the number of degrees of freedom or some other related measure.
The graph in Figure 4.6 shows the maximum displacement in the direction of load as a
function of the inverse of the element size. The graph in Figure 4.7 shows the maximum
von Mises stress as a function of the inverse of the element size.

Figure 4.6 Global mesh refinement and convergence of maximum displacement. The
maximum displacement is graphed as a function of 1/h, where h is the element size.

41

6650_Book.indb 41 10/31/16 10:32 AM

 Chapter 4

Figure 4.7 Global mesh refinement and convergence of maximum von Mises stress. The
maximum stress is graphed as a function of 1/h, where h is the element size.

In a discretized model, displacements have to follow displacement pattern designed

in the elements used for meshing. This adds some (very low) stiffness, which we call
artificial stiffness. Every mesh refinement reduces that artificial stiffness and this
explains why the maximum displacement increases, as shown in Figure 4.6. Strain
is found as a derivative of displacement and stress is calculated once strain has been
found; therefore, stress also increases with mesh refinement. Note the “dip” in the nodal
stresses graph shown in Figure 4.7 caused by stress averaging, which is a part of nodal
stress calculations.

Having completed the four steps of the h convergence process by global mesh refine-
ment, we notice that displacements converge faster than stresses. To explain this, we
need to differentiate between the global and local FEM results. The maximum displace-
ment is a global result; stiffness of the entire model contributes to the maximum
displacement results. The global results converge fast with mesh refinement. The
maximum stress (here von Mises stress) is a local result and is modeled only by a few
elements; for a faster convergence of the maximum stress, we would need a more aggres-
sive refinement in the area of stress concentration.

4.2.2 h Convergence Process by Local Mesh Refinement

An alternative to refining mesh everywhere is a local mesh refinement that produces
meshes shown in Figure 4.8. Local mesh refinement adds fewer degrees of freedom to
the model compared with the global mesh refinement and consecutive models solve
faster. The h convergence process by local mesh refinement is easier to execute compared
to using global mesh refinement as long as we know where to apply the mesh refine-
ment; the h convergence process by local mesh refinement requires a prerequisite

42

6650_Book.indb 42 10/31/16 10:32 AM

 Controlling Discretization Errors

knowledge of stress pattern in the analyzed model. There is a risk of not finding stress
concentrations if local mesh refinement is applied in an incorrect location.

Figure 4.8 The same model meshed with shell elements of different sizes along the
controlled entity (edge of the hole). This illustrates a local mesh refinement. Mesh bias
expressed in mm characterizes the element size on the entity where mesh bias has
been applied.

The von Mises stress results produced by four meshes used in the h convergence
process by local mesh refinement are shown in Figure 4.9; the convergence graph of the
maximum von Mises stress (nodal) is shown in Figure 4.10.

43

6650_Book.indb 43 10/31/16 10:32 AM

 Chapter 4

Figure 4.9 Von Mises stress results produced by four different meshes representing the
same problem. This illustrates the h convergence process by local mesh refinement.

44

6650_Book.indb 44 10/31/16 10:32 AM

 Controlling Discretization Errors

Figure 4.10 Convergence of maximum nodal von Mises stress during the h convergence
process by local mesh refinement. The maximum stress is a function of 1/h, where h is the
element size. “Dip” is stress in the second iteration is caused by stress averaging.

4.2.3 Adaptive h Convergence Process

The h convergence process by global mesh refinement is time consuming; the number of
degrees of freedom increases with every mesh refinement and the model grows is size
very fast. The h convergence process by local mesh refinement is faster because local
mesh refinement has less effect on the number of degrees of freedom in the analyzed
model, but it requires the user’s decision on where to refine the mesh. This disadvan-
tage of local h convergence process is eliminated if the adaptive h convergence process
is used.

In the adaptive h convergence process, the mesh is refined automatically during an itera-
tive solution. The initial mesh is defined by user and a solution is obtained; let us call it
iteration #1. Based on errors found in this first solution, the mesh is automatically refined
in the locations characterized by high errors and a new solution is obtained; this is itera-
tion #2. Mesh refinements continue in the subsequent iterations until the user specified
accuracy has been satisfied or the number of the allowed iterations has been reached.
The user does not have a direct control over how the final mesh will look like.

The name “adaptive” derives from the fact that mesh refinement is adapted the stress
errors found. The measure of error is related to the difference between nodal stress and
element stresses.

Consecutive mesh refinements during the h adaptive convergence process are illustrated
in Figure 4.11. A hollow plate is subjected to a uniform tension causing stress concentra-
tions on the cylindrical surface of the hole and the mesh is automatically refined there.
There are important differences between this problem and that shown in Figures 4.5 and
4.9. First, the model is subjected to a tension on both end faces and stresses are perfectly

45

6650_Book.indb 45 10/31/16 10:32 AM

 Chapter 4

symmetric. Second, because of a larger thickness, the problem requires 3D solid elements
rather than 3D shell elements or 2D plate elements. Lower stresses are reported because
the thickness has increased while the load has remained the same.

Figure 4.11 Von Mises stresses in five iterations of the h adaptive convergence process.

A summary of the von Mises stress results from Figure 4.11 is shown in graph in Figure
4.12, where von Mises stress is as a function of the iteration number.

46

6650_Book.indb 46 10/31/16 10:32 AM

 Controlling Discretization Errors

Figure 4.12 Convergence of the maximum von Mises stress in the h adaptive process; the
maximum von Mises stress is shown as a function of the iteration number.

4.2.4 p Convergence Process

We have discussed three variations of the h convergence process: global, local, and h
adaptive. In all cases, the element size is reduced, but the element order stays the same.
Now, we will keep the same mesh, but we will upgrade the element order. Because the
polynomial order of element displacement interpolation functions changes, this type
of convergence process is called p convergence (p stands for polynomial). The p conver-
gence process is possible only using elements capable of upgrading their order.

In a direct analogy to global mesh refinement (uniform refinement) and local mesh
refinement (nonuniform refinement) used in the h convergence process, the p conver-
gence process can also be uniform or nonuniform. In a uniform p convergence process,
the order of all elements is upgraded until the desired accuracy is obtained. The advan-
tage of uniform element upgrade is that mesh compatibility is automatically assured
because the same displacement interpolation functions are used on all shared faces
and shared edges of all elements. To assure mesh compatibility in a nonuniform (called
adaptive) p-convergence process, different displacement interpolation functions must
be used on different edges and faces of the same element; this is done to assure that
displacements on the adjacent edges and faces of the neighboring elements are described
by the same displacement interpolation functions. The face and edge order is “adapted”
to the actual stress pattern and “quiet” portions of the model can be left at lower p orders
for faster solution. To take a full advantage of p elements, a dedicated p element auto-
mesher is required. As a nonadaptive p convergence would not be numerically efficient,
all commercial p element programs use p adaptive convergence.

47

6650_Book.indb 47 10/31/16 10:32 AM

 Chapter 4

The p element mesh shown in top illustration in Figure 4.13 is a result of discretiza-
tion of a surface; it consists of only six shell elements: two triangular elements and four
quadrilateral elements. It looks very different from the h element mesh. The model is
subjected to a uniform tensile load applied to both vertical edges. The model is solved
with program that can use element order anywhere between p = 1 and p = 9. In this case,
the number between 1 and 9 refers to the order of stress interpolation function, not the
displacement interpolation function. The final element order is found after a number of
iterations during which the edge order is increased selectively based on errors found
in the previous iteration; this happens in direct analogy to the h adaptive conver-
gence process.

Figure 4.13 The p element mesh (top), the final edge element order (middle), and the von
MIses stress results found in a p adaptive solution process.

Figure 4.13 (middle) shows the final p order of element edges. Figure 4.13 (bottom)
shows the von Mises stress plot; this result was produced in seven iterations. Given the
accuracy requirements used in this solution, the solver did not have to reach the highest
available p order and solution stopped at p = 7. The convergence of the maximum von
Mises stress is shown in Figure 4.14, where the maximum von Mises stress is plotted for
all seven iterations.

48

6650_Book.indb 48 10/31/16 10:32 AM

 Controlling Discretization Errors

Figure 4.14 Convergence of the maximum von Mises stress in the p adaptive process; the
maximum von Mises stress is shown as a function of the iteration number.

4.2.5 The Choice of Convergence Process

The majority of commercial FEM programs use h elements; they are native h element
programs. For those programs, the h adaptive convergence, if available for the given
type of analysis, offers advantages of an automatic detection of stress concentrations and
mesh refinements adapted to the identified stress patterns. The users benefit from auto-
matically generated convergence graphs of the data of interest. Some native h elements
programs do support p elements even though those p elements are created using an h
element automesher. This makes for a very useful learning tool that allows the users’
familiarization with both h and p element technologies and leads to a full understanding
of the convergence process. However, the best performance of native h element programs
is usually achieved in the h element mode.

4.3 Discretization Error

All the graphs presented in this chapter show the data of interest (here displacements
and stresses) converging to a finite value and indicate that discretization errors dimin-
ishes with the increase in the number of degrees of freedom in the model.

A simplified convergence curve shown in Figure 4.15 summarizes the results of the
convergence process. For clarity of this illustration, only three steps are shown. This
graph could have been produced by any type of convergence processes: h convergence
global, h convergence local, h convergence adaptive, or p convergence adaptive. On the
ordinate, there is the iteration number and on the abscissa, the selected data of interest,
which is the convergence criterion. We will now present different ways of calculating the
discretization errors in the FEM solution.

49

6650_Book.indb 49 10/31/16 10:32 AM

 Chapter 4

Figure 4.15 Definitions of convergence error and solution error.

4.3.1 Convergence Error

We can define convergence error as the difference in results between two consecutive
mesh refinements. It is convenient to normalize the error to make it dimensionless. If we
do that, the convergence error for iteration n is defined as

result(n) − result(n − 1)
Convergence error = . (4.1)
result(n)

Figure 4.15 shows the convergence error of the last performed iteration, which is the
third iteration.

Using the definition (4.1), the convergence error can be calculated for all steps of the
convergence process except step 1. Convergence error for step 1 is unknown because no
prior results exist. We can rephrase this important observation saying that a single run
produces results with unknown discretization error.

4.3.2 Solution Error

With adding degrees of freedom to the model during the convergence process, we
approach the exact solution of continuous mathematical. That unknown solution is the
limit to which the data of interest converge. This asymptotic value may be used to define
the discretization error another way; this will be called solution error. A solution error is
the difference between the results produced by the FE model and the results that would
be produced by a hypothetical FE model with an infinite number of infinitesimally small
elements. This hypothetical FE model with an infinite number of elements would not be
any different from continuous mathematical model:

result(n) − asymptotic result

Solution error = (4.2)
result(n)

where result(n) is the result of the last performed iteration.

50

6650_Book.indb 50 10/31/16 10:32 AM

 Controlling Discretization Errors

As asymptotic solution is not known, we can only estimate the solution error. The
convergence error and solution error are depicted in Figure 4.15.

4.4 Problems With Convergence

The finite-element solution converges to the exact solution of the mathematical model
on which the FE model is based. Both h and p convergence processes demonstrated so
far showed that the data of interest converge to finite values. But what happens if the
solution fails to converge to a finite value?

4.4.1 Stress Singularity

Say we wish to find the maximum von Mises stress in a thin L-shape bracket. The
bracket is modeled as a 2D plane stress problem and is meshed initially with large h
elements as in Figure 4.16.

Figure 4.16 A thin L Bracket (left) may be represented by a flat surface and meshed with
2D plate elements (right). The mesh shown has an element size of 8 mm. Notice that 2D
model may be used for stress analysis but not for buckling analysis.

51

6650_Book.indb 51 10/31/16 10:32 AM

 Chapter 4

We conduct the h convergence process by global mesh refinement starting with an

element size of 8 mm as shown in Figure 4.16. The element size will be split in half in
each iteration until it reaches 0.25 mm in the sixth and final iterations. A summary of
the maximum von Mises stress results is shown in Figure 4.17. In all six iterations, the
maximum stress is located in the sharp reentrant edge. The maximum stress 73.7 MPa
(element size 0.25mm) is over 2.5 times higher than the yield strength of material (1060
aluminum alloy), which is 27.6 MPa. Figure 4.18 shows the results of the last performed
iteration with an element size of 0.25 mm.

Figure 4.17 The maximum von Mises stress (located in the sharp reentrant edge) as a
function of the number of degrees of freedom in the model with the element sizes of 8, 4,
2, 1, 0.5, 0.25 mm.

To find the accuracy of this result, we need to examine if the stress convergence curve
converges to a finite value. However, a quick examination of the curve in Figure 4.17
reveals no sign of converging to a finite value; instead, each iteration brings ever higher
stress results; the maximum stress diverges to infinity. The reason is that the mathemat-
ical model does not offer a solution for the maximum stress. The sharp reentrant corner
constitutes a singularity where stress is infinite. As the objective of analysis is to find
the maximum stress, we conclude that the FE model has been based on a wrong math-
ematical model. The mathematical model with stress singularity that coincides with the
location of maximum stress cannot be used as a basis for the FE model if the analysis
objective is to find the maximum stress.

52

6650_Book.indb 52 10/31/16 10:32 AM

 Controlling Discretization Errors

Figure 4.18 The maximum von Mises stress plot produced by mesh with an element size of
0.25 mm.

Why did the FE model produce a high but finite stress instead of infinite stress predicted
by the mathematical model? This is because the modeling error (using wrong math-
ematical model) was masked by discretization error, which causes that finite stress is
produced where mathematical model predicts infinite stress.

The stress singularity in the model in Figure 4.16 is caused by 270° sharp reentrant
corner, but any reentrant corner (in 2D model) or a reentrant edge (in 3D model)
produces stress singularity.

The strength of a singularity increases with the angle of the sharp reentrant edge. For
example, a 270° edge causes a stronger singularity than a 225° edge and 340° causes
stronger singularity than a 270° edge. Singularities manifest themselves as “hot spots”;
the stronger the singularity, the easier it is to notice that “hot spot.” To visualize singular
stresses caused by the 225° edge, a more aggressive mesh refinement is required
compared with the stress singularity caused by the 270° edge (Figure 4.19).

53

6650_Book.indb 53 10/31/16 10:32 AM

 Chapter 4

Figure 4.19 All sharp reentrant edges (or sharp reentrant corners in 2D) produce stress
singularity. The closer is the angle to 360°, the stronger is the singularity. Comparing
these three models, the singularity in the model with a 225° edge angle is the weakest
and the singularity in the model with a 340° edge angle is the strongest. The strength of
singularity indicates how fast stress diverges to infinity.

By showing the divergence of the data of interest (here the maximum von Mises stress),
the convergence process revealed the modeling error. We can remedy this situation
using a different mathematical model, one that does not have a stress singularity. The
most obvious way is to model a fillet, which is always present in a real part even if
the edge is very sharp. The result may show very high stress, but that result will be
bounded; it will converge to a finite value.

Another way of eliminating stress singularity is to use material model capable of

modeling plasticity. To illustrate this, we will use an elastic–perfectly plastic model char-
acterized by a strain–stress curve in Figure 4.20. The elastic–perfectly plastic material
model limits the value of the maximum von Mises stress to the plasticity limit; this
eliminates stress singularity even though strain at the sharp reentrant corner continues
diverging. The von Mises stress results using an elastic–perfectly plastic material and an
element size of 0.5 mm is shown in Figure 4.21.

54

6650_Book.indb 54 10/31/16 10:32 AM

 Controlling Discretization Errors

Figure 4.20 Elastic–perfectly plastic material, von Mises type; this illustration shows the
strain–stress curve of 1060 aluminum alloy. The material behaves linearly until the von
Mises stress reaches 27.6 MPa and then the modulus of elasticity becomes zero. Von Mises
stress is used as a measure to control the switch of the modulus of elasticity to zero.

Figure 4.21 The von Mises stress solution using an elastic–perfectly plastic material and an
element size of 0.5 mm. This illustration shows the results of step 14, which is the last step.

55

6650_Book.indb 55 10/31/16 10:32 AM

 Chapter 4

The solution shown in Figure 4.21 has been reached in 14 steps during which the load
was being changed from 0 to 100%. The maximum von Mises stress for each step is
shown in Figure 4.22.

Figure 4.22 Changes of the maximum von Mises stress during 14 steps of the solution
using elastic–perfectly plastic material and an element size of 0.5 mm.

Stress results in the vicinity of singularity are entirely dependent on the choice of
discretization and, therefore, are meaningless. By manipulating the element size,
element order or both, we can produce any stress result we want. Using the geometry
with sharp reentrant corners in 2D models or sharp reentrant edges in 3D models, while
the objective is finding stress in that location, is a severe yet common modeling error.
An erroneous model is shown in Figure 4.23. This model can be used for displacement
analysis because sharp reentrant edge does not pose displacement singularities. It can
also be used for stress analysis in a location distant from sharp reentrant edges.

It is important to mention that fillets cannot be ignored even if stresses along the edge
are not of interest, if removal of fillets changes model stiffness significantly.

If stresses along the edge are of interest, then fillets, no matter how small, must be
modeled (Figure 4.24).

56

6650_Book.indb 56 10/31/16 10:32 AM

 Controlling Discretization Errors

Figure 4.23 This model cannot be used for analysis of the maximum stress because
those stresses (most likely) will coincide with the stress singularity caused by the sharp
reentrant edge.

Figure 4.24 This model has a fillet (red) added in place of the former sharp reentrant
edge. It can be used for analysis of the maximum stress. Stresses in the fillet may be very
high, but during a convergence process, they will converge to a finite limit. Notice a mesh
control defined on the fillet to produce correctly sized and correctly shaped elements.

4.4.2 Displacement Singularity

Now we wish to find displacements of a thin plate in bending supported by two spot
welds (Figure 4.25). Model geometry and boundary conditions (restraints and loads)
lend themselves to a 2D plane stress representation. Because the weld size is small

57

6650_Book.indb 57 10/31/16 10:32 AM

 Chapter 4

in comparison with the overall beam dimensions, we decide to model them as point
supports; this is mistake, as we will soon find out. The stress results shown in Figure
4.26 indicate high stresses around supporting points. Divergence of the maximum stress
reveals stress singularity as shown in Figure 4.27.

Figure 4.25 A thin plate supported in two points (green dots) and loaded along the top
edge. This model lends itself to a 2D plane stress representation.

Figure 4.26 Three stress results in the global mesh refinement process produce diverging
stress results caused by stress singularities at point supports. Undeformed stress plots
are shown.
58

6650_Book.indb 58 10/31/16 10:32 AM

 Controlling Discretization Errors

Figure 4.27 Summary of the maximum stress results from three iterations shown in
Figure 4.26. The maximum von Mises stress is plotted as a function of 1/h, where h is the
characteristic element size. The stress at the point support diverges.

Can we still rely on this model to produce meaningful displacement results? To answer
this question, we need to examine the convergence of displacements. The displace-
ment results for three different meshes are shown in Figure 4.28 and are summa-
rized in Figure 4.29. Examination of graph in Figure 4.29 proves that displacement
results diverge.

59

6650_Book.indb 59 10/31/16 10:32 AM

 Chapter 4

Figure 4.28 Three displacement results in the global mesh refinement process produce
diverging displacement results caused by displacement singularities at point supports.
Deformed displacement plots are shown.

Figure 4.29 Summary of the maximum displacement results from three iterations shown
in Figure 4.28. The maximum resultant displacement is plotted as a function of 1/h, where
h is the characteristic element size. Displacement results diverge.
60

6650_Book.indb 60 10/31/16 10:32 AM

 Controlling Discretization Errors

The maximum displacement tends to infinity with mesh refinement and finite displace-
ment results reported by the FE model are because of discretization errors that conceal
both displacement and stress singularities. Displacement and stress singularities are
summarized in Table 4.2.

Table 4.2 Types of singularities encountered in finite-element models

Type 1 Type 2
Stress Infinite Infinite
Strain energy Finite Infinite
Displacement Finite Infinite
Examples Point load in 2D Point support in 2D
Line load in 3D Line support in 3D
Sharp reentrant corner in 2D
Sharp reentrant edge in 3D

Point support is a mathematical abstract and just like a sharp reentrant corner, it can
never exist in a real structure. Point supports can be correctly used in FEA only if they
are not expected to generate nonzero reactions like in the case of restricting rigid body
motions (RBMs). Figure 4.30 illustrates eliminating RBMs for a 2D model, Figure 4.31
for a 3D solid element model and Figure 4.32 for a 3D shell element model. All these
three cases prepare the plate model to be loaded with tensile load applied to both short
edges without the use of fixed restraints applied to a face (3D model) or edge (2D model).
This is often done to eliminate stress singularities created by fixed restraints also called
rigid supports.

Figure 4.30 Point supports applied to two corners of this 2D model prevent RBMs but do
not generate any reaction forces. Restraints shown in the lower left corner (green arrows)
eliminate two translational degrees of freedom leaving the model free to rotate about
this corner. Restraint eliminating one degree of freedom, shown in the lower right corner,
eliminates that rotation. Load on one side may be replaced by restraint acting in horizontal
direction only. 2D elements in this model can be plane stress or plane strain elements; they
have 2 DOF/ per node; both are translations.

2D model in Figure 4.30 is loaded with tensile load applied to both short sides.

61

6650_Book.indb 61 10/31/16 10:32 AM

 Chapter 4

The following two 3D models, one using solid elements and the other one using shell
elements, have restraints applied to one side and tensile load to the opposite side.

Figure 4.31 Point restraints used to eliminate RBMs in model meshed with solid elements.
Solid elements have 3 DOF/ per node; all three are translations. Notice that applying
fixed (rigid) restraints to the end-face would cause stress singularities in the four corners.
Tensile load is applied to the opposite short face.

62

6650_Book.indb 62 10/31/16 10:32 AM

 Controlling Discretization Errors

Figure 4.32 Point restraints used to eliminate RBMs in model meshed with shell elements.
Shell elements have 6 DOF/ per node: three translations and three rotations. Tensile load is
applied to the opposite short edge.

Singularities cannot be “fixed” using FEM methods. Singularities are modeling errors
introduced by the formulation of mathematical model, and not by the finite-element
approximation. Singularities, like other modeling errors, are introduced before the FEM
enters the stage. Singularities can be revealed in convergence process, but no corrective
action is possible unless mathematical model is changed.

The presence of singularity does not necessarily invalidate the FE model or make results
incorrect, but we must be aware of the limitations imposed by the existence of singu-
larity. For example, a model with sharp reentrant edges is incorrect if the analysis objec-
tive is to find the maximum stress or stress close to that edge. However, if the data of
interest are displacements or modes of vibration, then the model may be used and accep-
tance of sharp reentrant edges in model geometry allows us to simplify the model. This
is valid as long as the elimination of sharp reentrant edges from model geometry does
not change stiffness significantly.

63

6650_Book.indb 63 10/31/16 10:32 AM

 Chapter 4

4.5 Hands-On Exercises

4.5.1 Hollow Plate (Figure 4.33)

Figure 4.33 Hollow plate exercise. Restraints defined on the flat end face (red arrows)
generate reaction forces and restraints defined on two corners (blue and green arrows)
eliminate RBMs.

Model name
• 4.01.HOLLOW_PLATE.x_t
• 4.01.HOLLOW_PLATE.sldprt

4.5.1.1 Objective: Demonstrate the convergence of displacements and stresses using

h and p convergence processes and first- and second-order elements.

4.5.1.2 Description: A plate with a circular hole is loaded with 100000-N tensile load
uniformly distributed to one ends and supported as shown in Figure 4.33; restraints are
identical to those shown in Figure 4.31. The geometry is suitable for meshing with solid
elements. This exercise illustrates the h and p convergence processes and demonstrates
how data of interest (maximum displacement in the direction of load and maximum von
Mises stress) change with the element size during several steps of mesh refinement (h
convergence) or several steps of element order upgrade (p convergence).

64

6650_Book.indb 64 10/31/16 10:32 AM

 Controlling Discretization Errors

The following are the required steps:

1. Apply the material properties (steel).

2. Apply sufficient restraints to balance the load and eliminate RBMs but not to create
stress singularities; follow the steps shown in Figure 4.31.
3. Apply a 100000-N uniformly distributed tensile load to the opposite face.
4. Mesh using first-order solid elements.
5. Obtain displacement and stress solutions.
6. Repeat steps 4 and 5 each time meshing with smaller elements, this is the h
convergence process by global mesh refinement.
7. Plot the displacement and stress results as a function of the number of degrees of
freedom in the model. Each curve will consist of three points corresponding to three
steps in the mesh refinement process. If your software does not provide information
of the number of degrees of freedom in the model, use the umber of nodes or
number of elements or the inverse of the characteristic element size.
8. Repeat the exercise using second-order solid elements, and note that the results
converge faster when using second-order elements.
9. Repeat the exercise using the h adaptive solution with second-order elements if your
program supports this option; only one mesh will be required and this initial mesh
may be a coarse one. The number of automatic steps will depend on the accuracy
requirements specified in the analysis definition.
10. Repeat the exercise using p-type elements if your software supports this option;
only one mesh will be required and this initial mesh may be a coarse one. The
number of automatic steps will depend on the accuracy requirements specified in
analysis definition.
11. Observe the convergence of displacement and stress in all solutions.
12. Delete all restraints and define a fixed restraint applied on the face opposite to
where the load is applied. Notice that stress singularities are now present in the
corners of the supported face. These singularities are difficult to spot when global
mesh refinement is used because stress concentration becomes visible only when
a very fine mesh is used. In this exercise, they may show in the h adaptive and p
adaptive convergence processes.

65

6650_Book.indb 65 10/31/16 10:32 AM

 Chapter 4

4.5.2 L Bracket (Figure 4.34)

Figure 4.34 L bracket exercise. Restraints are defined on the top edge; the bending load
is evenly distributed over the right vertical edge.

Model name
• 4.02.L_BRACKET.x_t
• 4.02.L_BRACKET.sldprt

4.5.2.1 Objective: Demonstrate the convergence of displacements and the

divergence of stresses in the h and p convergence processes in the model with
stress singularities.

4.5.2.2 Comment: The L shape bracket is represented by a surface intended for

meshing with 2D plane stress elements or 3D shell elements. This exercise illustrates the
lack of convergence of the maximum stress because of stress singularity in the sharp
reentrant corner. The stress singularity makes the model useless if the objective is to
find the maximum stress because the location of the maximum stress is coincident with
stress singularity.

The following are the required steps:

1. Apply the material properties (1060 aluminum alloy).

66

6650_Book.indb 66 10/31/16 10:32 AM

 Controlling Discretization Errors

2. Apply the shell properties (thickness 10mm).

3. Define a fixed support to the top edge.
4. Define a uniformly distributed bending load of 200 N, as shown in Figure 4.34.
5. Use second-order 2D plate elements or 3D shell elements.
6. Mesh with coarse, medium, and fine meshes.
7. Plot the displacement and stress results as a function of the number of degrees of
freedom in the model. Each curve will consist of three points corresponding to three
steps in the mesh refinement process.
8. Repeat the exercise using the h adaptive solution.
9. Repeat the exercise using the p adaptive solution.

4.5.3 2D Beam (Figure 4.35)

Figure 4.35 2D beam exercise; two point restraints (green dots); bending load (red
arrows) evenly distributed over the top horizontal edge.

Model name
• 4.03.2D_BEAM.x_t
• 4.03.2D_BEAM.sldprt

4.5.3.1 Objective: Demonstrate the divergence of displacements and stresses in

model with displacements and stress singularities.

4.5.3.2 Comment: A thin beam is supported by two spot welds. This model is
intended for meshing with 2D plane stress elements or 3D shell elements. Because of the
small size of spot welds, compared with the overall model size, somebody decided to
model spot welds as point supports, which was a bad mistake as we soon will discover.

The following are the required steps:

1. Apply steel material properties.

2. Define thickness (10 mm).
3. Apply fixed supports to two points.

67

6650_Book.indb 67 10/31/16 10:32 AM

 Chapter 4

4. Apply the load of 1000 N as shown in Figure 4.35.

5. Use the second 2D plate elements.
6. Mesh with coarse, medium, and fine meshes decide on the actual element size.
7. Plot the displacement and stress results as a function of the number of degrees of
freedom in the model. Each curve will consist of three points corresponding to three
steps in the mesh refinement process.
8. Repeat the exercise using the h adaptive solution; only one mesh will be required (it
may be the coarse one).
9. Repeat the exercise using the p adaptive solution; only one mesh will be required (it
may be the coarse one).
10. Observe the divergence of displacements and stresses as demonstrated by the global
h convergence, h adaptive convergence, and p adaptive convergence processes.

68

6650_Book.indb 68 10/31/16 10:32 AM