Reliable FE-Modeling with ANSYS

Thomas Nelson, Erke Wang

CADFEM GmbH, Munich, Germany

Abstract
ANSYS is one of the leading commercial finite element programs in the world and can be applied to a large
number of applications in engineering. Finite element solutions are available for several engineering
disciplines like statics, dynamics, heat flow, fluid flow, electromagnetics and also coupled field problems.
It is well known that a finite element solution is always an approximate solution of the considered problem
and one always has to decide whether it is a good or a bad solution. Hence, the question is how a finite
element solution can be validated. In this paper we mention several aspects of modelling to obtain a good
finite element solution for problems in structural mechanics. Although ANSYS gives quite a lot of
warnings to prevent the user from making modelling mistakes, there are still some aspects during the pre-
and postprocessing and also during the solution process we would like to discuss just to point out where
user knowledge might still be necessary to obtain good results.
In our investigations we use both ANSYS products: ANSYS and ANSYS Workbench. Several ways of
modelling a problem of structural mechanics are discussed, taking into account the different available
features of ANSYS and ANSYS Workbench.

Introduction
The finite element method (FEM) is the most popular simulation method to predict the physical behaviour
of systems and structures. Since analytical solutions are in general not available for most daily problems in
engineering sciences numerical methods have been evolved to find a solution for the governing equations
of the individual problem. Although the finite element method was originally developed to find a solution
for problems of structural mechanics it can nowadays be applied to a large number of engineering
disciplines in which the physical description results in a mathematical formulation with some typical
differential equations which can be solved numerically.
Much research work has been done in the field of numerical modelling during the last thirty years which
enables engineers today to perform simulations close to reality. Nonlinear phenomena in structural
mechanics such as nonlinear material behaviour, large deformations or contact problems have become
standard modelling tasks. Because of a rapid development in the hardware sector resulting in more and
more powerful processors together with decreasing costs of memory it is nowadays possible to perform
simulations even for models with millions of degrees of freedom.
In a mathematical sense the finite element solution always just gives one an approximate numerical
solution of the considered problem. Sometimes it is not always an easy task for an engineer to decide
whether the obtained solution is a good or a bad one. If experimental or analytical results are available it is
easily possible to verify any finite element result. However, to predict any structural behaviour in a reliable
way without experiments every user of a finite element package should have a certain background about
the finite element method in general. In addition, he should have fundamental knowledge about the applied
software to be able to judge the appropriateness of the chosen elements and algorithms.
This paper is intended to show a summary of ANSYS capabilities to obtain results of finite element
analyses as accurate as possible. Many features of ANSYS are shown and where it is possible we show
what is already implemented in ANSYS Workbench.
We will distinguish two different sources of errors within a finite element analysis. On the one hand some
mistakes might be introduced in the analysis because the user himself does not know enough about finite
elements in general. To minimize these errors we summarize important features of certain element types
and element formulations of ANSYS. We also discuss the quality of different element shapes with respect
to accuracy and finally provide some information of a correct coupling of different element types. On the
other hand errors might occur due to a poor quality of the used finite element program itself. A high-quality
program assists the user in reporting reasonable warnings and errors. We will discuss some typical error
messages from ANSYS, which allow the user to correct a finite element model immediately. Some reasons
possibly leading to poor finite element results are summarized in Tab.1 below to give a coarse overview:
Table 1: Possible sources of errors in a finite element analysis
Engineer Finite Element Software
- wrong element type - bad element formulation
Preprocessing - wrong element coupling - bad meshing algorithm
- wrong idealization - no warnings and errors
- wrong calculation discipline - wrong calculation algorithm
Solution - wrong boundary conditions - inaccurate equilibrium iteration
- wrong convergence criteria - no warnings and errors
- wrong result coordinate system - wrong result averaging
Postprocessing - wrong selection of components - wrong displaying of results
- wrong interpolation of results - no warnings and errors

We should not forget to point out, that ANSYS is a general purpose program, where many numerical
modelling techniques are implemented. However, it is sometimes not easy to learn especially for beginners
or even for designers, who are usually not finite element experts. With ANSYS Workbench some effort has
been done to offer a product, which has implemented by default the best algorithms of ANSYS and is
furthermore very easy to use. Hence we will always discuss in this paper as a first step some modelling
features of ANSYS and finally point out which of them are already available in ANSYS Workbench.

Preprocessing
First of all, we will summarize some important aspects every user should be familiar with when doing the
preprocessing of a finite element analysis with ANSYS. The following topics will be discussed very
briefly: consequences of different element shape functions, important features of different beam and shell
elements, results of different element shapes and element formulations and finally correct coupling of
different element types. In addition to that, we show what kind of help is available from ANSYS in terms
of warnings and errors during the preprocessing to minimize modelling mistakes.

User knowledge during preprocessing

Choosing an element with linear or a quadratic shape functions

To make it as easy as possible we will just look at elements with displacement degrees of freedom – no
rotational degrees of freedom are present. In ANSYS such elements are called SOLID… (acting in 3D),
PLANE…(acting in 2D) or LINK… (acting in 1D), respectively.
Let us consider the one-dimensional case. Within one element the displacements are supposed to vary in a
linear or quadratic manner:

u ( x) = a 0 + a1 ⋅ x (1)

u ( x) = a0 + a1 ⋅ x + a 2 ⋅ x 2 (2)

Hence, we talk about linear or quadratic elements. As a consequence of that assumption the strain and also
the stress distribution is either constant or linear within each element due to:
 du
ε ( x) = and σ ( x) = E ⋅ ε ( x) (3)
dx
This may be easily illustrated in the following one-dimensional truss example. The resulting displacement
and stress distribution is shown for a two element discretization with linear and quadratic elements:

F F

1 2 1 2
X
u ( x) = a0 + a1 ⋅ x u ( x) = a0 + a1 ⋅ x + a2 ⋅ x 2

ux Results: ux

analytic

du du
σ ( x) = E ⋅ σ ( x) = E ⋅
dx FEM dx

σx σx

Figure 1. One dimensional truss problem: linear and quadratic elements and their
consequences

With E as Young’s Modulus, F as applied force, and A(x) as the cross sectional area the analytic solution is
x
F 1 F
u ( x) =
E ∫ A( x) dx
0
and σ ( x) =
A( x)
(4)

coming from a direct integration of the governing differential equation of the above considered problem.
It is generally known that better results can be obtained using the same discretization with quadratic or
higher order elements compared with the results of linear or lower order elements. You should also keep in
mind that the degree of freedom solution always shows a smooth distribution from element to element
whereas the distribution of derived quantities (such as strains and stresses) is no longer smooth at the
element boundaries. This is a correct result in terms of the finite element method. Just the so-called “weak
formulation” of the concerning differential equation is solved by finite elements and the continuity
requirements for the governing variables of the problem are relaxed.
Recommendation:
ANSYS: In general, the user should prefer to take a quadratic element if possible.
ANSYS Workbench: By default, SOLID186/SOLID187 is used as a quadratic 3D Solid-Element

Beam elements – Bernoulli or Timoshenko

Let us now consider a different element type. To model the behaviour of a structure that is thin in two
dimensions relative to the third dimension, which is subjected to a bending load, a beam element is good
choice. Just the line of gravity of the beam-like-structure has to be discretized. In contrast to classical solid
elements beam elements have not only displacement degrees of freedom, but also rotations. A simple
supported beam structure under a uniform pressure load distribution is shown in Figure 2.
Y

X
Figure 2. Simple supported beam structure under uniform pressure load distribution

At this stage it is worth to distinguish two different beam theories: In ANSYS beam elements are available
according to the theory of Bernoulli or according to the theory of Timoshenko, respectively. The user
should know that shear stresses are not calculated for Bernoulli beam elements but only for Timoshenko
beam elements. Bending stresses (linear over the thickness) are available in both beam theories. Since the
effect of shear is neglected using Bernoulli beam elements the structure will show a stiffer behaviour as if a
Timoshenko beam model was used. This is illustrated in Figure 3.
BERNOULLI TIMOSHENKO
(default for BEAM3 / BEAM4) (standard for BEAM188 / BEAM189)

bending stress σx bending stress σx

shear stress τxy shear stress τxy

deformed configuration deformed configuration

Y

dmax_y= 0.007441 X dmax_y = 0.007626

Figure 3. Results of a beam structure using Bernoulli beam elements and Timoshenko
beam elements

It is worth to know about these both theories when using beam elements to choose the correct element type
if shear stresses are for example of interest. Considering the deformation it should be mentioned that
differences cannot be neglected any more if the ratio of the thickness to the length of the beam - d/l -
exceeds the limit of 1/10, i.e. if the beam cannot be considered any more as thin.
In [4] the analytical solution for the maximum displacement in y-direction according to Bernoulli is given:

5 ql 4
d max_ y = and with q=1, l=1 and EI=17500 d max_ y = 0.007440 . (5)
384 EI
We observe that the numerical solution with BEAM3/BEAM4 is almost exact. However, since shear
stresses are always present in reality, the user has a more accurate result, taking BEAM188/BEAM189.
Recommendation:
ANSYS: In general, the users should take BEAM188/189 if possible.
ANSYS Workbench: By default, BEAM188 is used.

Shell elements – Kirchhoff-Love or Reissner-Mindlin

Shell elements are quite similar to beam elements in the sense that they also have both degrees of freedom -
displacements and rotations. Shell elements are typically taken to model a structure subjected to a bending
load, that is thin in two dimensions relative to a third. When a thin structure is idealized using shell
elements one should know the following: by means of shell elements one actually predicts the structural
behaviour of the mid surface of the thin structure. A very easy shell example is shown in Figure 4.

Y X

Figure 4. Shell example: Simple plate subjected to a bending load with a fixed supported
edge

As for beam elements there also exist two different theories to formulate shell elements. The differences
between a typical KIRCHOFF-LOVE shell element like SHELL63 in ANSYS and a typical REISSNER-
MINDLIN shell element like SHELL181 is simple: Any shell element based on the KIRCHOFF-LOVE
shell theory does not calculate transverse shear stresses. Hence, the resulting deformations may be
underestimated, especially in thick shell structures. On the other hand, shell elements based on the
REISSNER-MINDLIN theory take into account the shear stress distribution over the thickness. As a
consequence of that these elements typically show a softer deformation behaviour due to the presence of
shear stresses. As for beam elements the effect of shear deformation can be neglected as long as the shell
structure can be considered as slender, i.e. if the ratio of the shell thickness to two typical lengths - d/l1 and
d/l2 - is less than 1/10. For both shell theories the bending stresses vary linearly with respect to the
thickness. The user should know which theory is implemented in which shell element.
As an example we analyse the simple shell structure shown in Figure 5. Due to the applied load case, we
can check both the correctness of the bending and membrane stiffness. An analytical solution for this
problem is derived based on the KIRCHOFF-LOVE theory in [3]. In Figure 6 we compare the analytical
result of the radial displacement with the numerical result using SHELL63. We use the default mesh
coming from ANSYS Workbench. The numerical solution is really good. Comparing the result of
SHELL63 and SHELL181 in Figure 7 it turns out that SHELL181 shows a softer behaviour. This result is
the best one, since the effect of shear stresses is also included in the analysis.
For the shell problem the following parameters are given: Q=10000, E=210000, µ=0.3, t=15, 0<s<223.61
and θ=63.43. From the geometry we can calculate rϕ and rϕ0. Note, that rϕ always depends on the position
of s. Note also, that the used finite element mesh is the default mesh the user gets in ANSYS Workbench:
 s Q
rϕ 0

rϕ
θ
r

Figure 5. Shell problem: Static system and finite element mesh

In [3] we find the analytical solution for the displacements in r-direction depending on the coordinate s:
s 2
2 Q sin 2 θ − λ0 rϕ 0
λ0 = 3(1 − µ )t
s
u r (s ) =
r
λ0 rϕ e ϕ 0 cos λ0 with 4 2
2
(6)
Et rϕ 0

ur(s=0) = 1.993 ur(s=0) = 2.046

Analytical solution Numerical solution

Figure 6. Shell problem: Analytical and numerical result of the displacement in r-direction
depending on s

KIRCHOFF–LOVE REISSNER–MINDLIN
(standard for SHELL63) (standard for SHELL181)

Figure 7. Shell problem: Displacement in r-direction (over-scaled) using SHELL63 and

SHELL181
Recommendation:
ANSYS: In general, the users should take SHELL181 if possible.
ANSYS Workbench: By default, SHELL181 is used.

Solid elements – element technology and element shape

Let us consider solid elements again. In this subsection we compare the results of linear and quadratic two-
dimensional solid elements when modelling a bending dominated problem. We will just focus on the solid
elements of the 18x class, i.e. PLANE182/PLANE183. We also have a look at different element shapes.
What is mentioned holds also for the three-dimensional case, using SOLID185/SOLID186/SOLID187.
For the problem in Figure 8 an analytical solution is available in [4]:

Y
B = 1 H = 1 M(F) = 1/6 W = 1/6 σx= M/W = 1
X
Figure 8. Simple supported structure subjected to pure bending: geometry model and
analytical solution

σx Let us first look at the results of quadrilaterals:

PLANE182 (linear) Using linear elements we can just obtain the correct
results using the Enhanced Strain Formulation
(instead of the Full Integration Technology). This
method prevents the elements from shear locking.
σx
PLANE182 (linear) This formulation should be activated by the user if it
is known from the beginning of the analysis that the
problem will be bending dominated.

σx Using quadratic elements results in a correct

PLANE183 (quadratic) bending stress distribution over the thickness.

Æ Æ Enhanced Strain Formulation

Figure 9. Results of bending stresses for linear and quadratic elements using a mesh of
quadrilaterals

PLANE182 (linear) σx We compare the results of quadrilaterals from Fig. 9

with the results of triangles in Fig. 10:
It is important to realize that linear elements are not
able to predict the correct results well when they are
PLANE183 (quadratic) σx used as triangles even if the Enhanced Strain
Formulation is used. Using quadratic elements the
results are correct even when triangles are used.

Figure 10. Results of bending stresses for linear and quadratic elements using a mesh of
triangles
Linear triangles never should be used whereas linear quadrilaterals can be used without any problems in a
structural analysis. If the problem is bending dominant the Enhanced Strain Formulation should be
activated. Quadratic triangles and quadrilaterals are always a good choice. The same thing is also valid for
the three-dimensional case.
Recommendation:
ANSYS: In general, the users should take PLANE183, SOLID186/SOLID187 if possible.
ANSYS Workbench: By default, SOLID186/187 is used.

Coupling of different element types

In this subsection we focus on coupling of different element types, especially those who have not the same
degrees of freedom. As an example we consider the following little example in Figure 11, where a beam or
shell element (displacements and rotations) is coupled with solid elements (displacements):

Incorrect coupling of elements with

different degrees of freedom:
Unintentionally a joint is created at the
Beam couple point.
= joint A rigid body motion is possible. The
Shell
resulting stiffness matrix will become
singular and the problem cannot be solved
like this.
Solids

Figure 11. Incorrect coupling of elements with different degrees of freedom

However, there are a few possibilities to correctly transmit the beam’s or shell’s rotation into the solid part
of the structure. Two finite element models are shown below in Figure 12, where additional beam or shell
elements are used to perform the coupling reasonably:

Beam Beam

Shell Shell

Solids Solids

Figure 12. Correct element coupling using additional finite elements

It is obviously that in both finite element models from Figure 12 the originally modelled joint from Figure
11 no longer exists and that the rotations are transmitted correctly by some additional elements. A quite
new technique to solve the above problem is given via the MPC (Mulit Point Constraint) method:
 Beam
MPC
Shell

Solids

Figure 13. Correct element coupling using the MPC method

Using the MPC technique, ANSYS generates internally some coupling equation’s to establish the correct
kinematics at the coupling point. In fact, the MPC technique is valid if a small and also a large deformation
analysis is performed. MPC is modelled via bonded contact together with a special contact solution
algorithm. The MPC technique is also already available in ANSYS Workbench.

Help from ANSYS during preprocessing

Element shape function and element shape

It was outlined above that linear elements should not be used as triangles or tetrahedra. However, if such
elements are still generated, ANSYS will give the following warnings from Figure 14. The user has
immediately the chance to correct the model to obtain better results. Since in ANSYS Workbench only
quadratic solid elements are available those warnings are not necessary in ANSYS Workbench:

Figure 14. ANSYS warnings to avoid linear elements as triangles or tetrahedra

Identification of rigid body motions in a structure

We have discussed above that a rigid body motion may be modelled in a structure due to an incorrect
coupling of different element types. Since the resulting stiffness matrix is singular ANSYS will not be able
to solve this problem. The error messages from Figure 15 appears in ANSYS and ANSYS Workbench:
 ANSYS

ANSYS Workbench

Figure 15. ANSYS and ANSYS Workbench messages due to incorrect element coupling
(rigid body motion)

However, it is also possible to introduce a rigid body motion into the system, if the structure is not
supported in a statically determined way. Such a situation is modelled in Figure 16 and it is shown how
ANSYS will identify such a modelling mistake, which has then to be corrected by the user:

Figure 16. ANSYS error messages to identify rigid body motions due to missing supports

Note, that ANSYS Workbench generates weak springs, if the program recognizes that a system is not
supported in a statically determined manner. Adding weak springs prevents the system from a rigid body
motion. If such a situation occurs, ANSYS Workbench gives the following information on screen:

Figure 17. ANSYS Workbench information to identify rigid body motions due to missing
supports
 Automatic shape checking control
A topic we did not discuss so far in this paper is the resulting mesh quality after discretizing the structure
with finite elements. It is well known that the accuracy of results within one element also depends on its
shape. Fortunately, there are a few algorithms implemented in ANSYS to check the quality of the resulting
element shapes to minimize the errors due to bad element shapes. For details we refer to [1].
If one of the shape checking criteria is not satisfied the user will get a warning in ANSYS like shown in
Figure 19. Furthermore, it is possible to indicate the bad elements in the mesh. So it is easy to identify
regions where remeshing is required.
To illustrate this we recall again the simple bending problem to compare the numerical results with the
analytical ones. In Figure 18 it is indicated that we expect a maximum bending stress of σx = 1:

Y
B = 1 H = 1 M(F) = 1/6 W = 1/6 σx= M/W = 1
X
Figure 18. Simple supported structure subjected to pure bending: geometry model and
analytical solution

In the following Figure 19 we show the results of a bad discretization using linear elements. The correct
results have already been shown in Figure 9 and Figure 10. We recognize, that especially within the
indicated bad elements the numerical results are poor. The shape warnings coming from ANSYS are also
shown for this case and the user has immediately the chance to correct the mesh:
σx
PLANE182 (linear)

Figure 19. ANSYS warnings to identify bad shaped elements and numerical results of bad
elements

At this stage we should note, that shape warnings will not be shown in ANSYS Workbench. However,
since ANSYS Workbench just uses quadratic solid elements this is not really a disadvantage, since those
elements are not as sensitive with respect to the resulting element shape. The above warnings would not
have been appeared if PLANE183 (quadratic) would have been used instead of PLANE182 (linear).
Solution
Solving a linear problem of structural mechanics meanwhile has become a standard task and can be
performed without any difficulties. The accuracy of the results just depends on the quality of the resulting
finite element model itself. By contrast a nonlinear problem has to be solved iteratively and a solution is
only obtained if convergence is achieved.
At this stage we would like to give a certain background about the solution of nonlinear problems. First of
all, we summarize briefly the basics of the iterative solution method which is implemented in ANSYS to
solve nonlinear problems. With this knowledge it should be possible to understand how certain user
settings on convergence criteria might influence the accuracy of a nonlinear solution.
In the last section we talk about the necessity of a geometric nonlinear calculation. We show what kind of
error is introduced in the analysis if the effect of large deflections is ignored numerically. We try to show
the limit of a geometric linear calculation considering the effect of large deflections.

General remarks about the solution of nonlinear problems

Every linear problem in structural mechanics results in the solution of the matrix equation K u = F, where
K is the stiffness matrix, u denotes the vector of displacements / rotations and F is the vector of forces /
moments. On the other hand in a nonlinear problem of structural mechanics the resulting matrix equation
can be formulated as K(u) u = F and it should be realized that the stiffness matrix depends on the
displacements / rotations. Due to this fact an iterative solution scheme is required.
In ANSYS the Newton-Raphson method is implemented to solve a nonlinear problem iteratively. The
iterative solution process can be described by the following equations together with Figure 20.

K T ∆u = F − F NR with F − FNR = R (7)

In every Newton-Raphson iteration the changing stiffness can be identified as the slope of the force
deflection curve from Figure 20. Hence we speak about a tangential stiffness KT. In every iteration a
displacement / rotation increment ∆u is calculated until the imbalance forces which are also called the
residual forces / moments R become acceptable small. Strictly speaking, a structure is only in equilibrium
if the residual forces /moments totally vanish. In the current version of ANSYS (8.0) the user has the
chance to postprocess the residual forces / moments to check the accuracy of the simulation.

F
Detailed consideration of the iteration:

R criterion
(small number)

∆u
Equilibrium is obtained if: |R| < criterion
u
u

Figure 20. The Newton-Raphson iteration scheme

To be able to compare the vector of residual imbalance forces or moments with a single scalar value a norm
of the vector is needed. Different vector norms are available in ANSYS and we refer to [2].

Help from ANSYS during the solution of nonlinear problems

ANSYS shows during a nonlinear solution a convergence monitor where the numerical behaviour during
the Newton-Raphson iteration can be studied. Hence, the user can observe whether the solution process
shows either a converging trend or an unstable behaviour. The iterative process which is graphically
displayed in the convergence monitor is also documented in the Output Window.
If the iterative process does not show a converging trend the user can stop the nonlinear iteration, think of
certain techniques to overcome the convergence difficulties and start the analysis again.
ANSYS Workbench has not yet implemented the convergence monitor. However, the Output Window of
ANSYS is available where the user is able to study the iterative solution procedure as well.

Iteration control of the residual forces and moments

By default for elements with displacement degrees of freedom the imbalance residual forces are iterated
below a certain criterion. However, if an element type supports rotational degrees of freedom, too, the
imbalance residual moments are also corrected in every iteration until they fall under a certain limit. Figure
21 shows a typical convergence diagram where the iteration process can be studied:
Convergence is obtained, if
the norm of the residual force
and moment vector (denoted
by F L2 / M L2) falls below
the criterion (called F CRIT /
M CRIT).

The considered structure

is then in equilibrium !!!

The value for the criterion is

calculated by ANSYS but can
be overwritten by the user if
desired.

Figure 21. The Newton-Raphson iteration of the imbalance residual forces and moments

Iteration control of the incremental displacements and rotations

It should be noted, that it is also possible to observe the evaluation of the incremental displacements and
rotations from iteration to iteration. Taking again a norm of the corresponding vectors allows the user to
stop with the iteration process if this norm is acceptable small. Hence, it is possible not to measure the
norm of the imbalance forces and moments but to stop with the iteration process as soon as the norm of the
incremental displacements and rotation falls under a certain limit. Such an iteration is show in Figure 22.
 The iteration stops, if the
norm of the incremental
displacements (U L2) and
rotations (ROT L2) falls
below the given criterion
(U CRIT / ROT CRIT).

The considered structure

is then not necessarily in
equilibrium !!!

The value for the criterion is

calculated by ANSYS but can
be overwritten by the user if
desired.

Figure 22. The Newton-Raphson iteration of the incremental displacements and rotations

Note carefully, that in case of activating only the incremental displacement and rotation control, you should
always check the results for equilibrium. In Figure 23 it is illustrated that especially for structures with a
stiffening behaviour a small increment in the displacements does not necessarily mean that also the
imbalance residual forces are already acceptable small. If there are still certain residual forces in the system
(imbalance forces) the structure is obviously not yet in equilibrium.

It can easily be observed that the incremental

displacement ∆u might already be acceptable small
but the imbalance residual force R might not.

∆u
u

Figure 23. Situation where just a displacement controlled iteration might lead to erroneous
results

The user should know that when activating the incremental displacement control an additional equilibrium
check should be performed after convergence has been achieved.
It is also possible to specify individually for which quantity the evaluation of the increments during the
Newton-Raphson iteration should be measured.
 About the necessity of a geometric nonlinear calculation
In this section we discuss the limit of a geometric linear calculation. We will show that with the classical
definition of strain like it is used within a geometric linear calculation it is not possible to calculate a rigid
body rotation in a correct manner.
However, by introducing a new nonlinear strain definition we are able to calculate correct results. It will be
shown that for large rigid body rotations the geometric linear theory looses its validity.
Let us focus on the following simple rod problem in Figure 24. The rod rotates as a rigid body with rotation
angle ϕ. Point A will move to point A’. The components of its movement are also calculated in Figure 23 in
an analytical manner:

y, v

A’
u ( x) = x (cos ϕ − 1)
v( x) = x sin ϕ

ϕ A
x, u

Figure 24. Single rod with rigid body rotation: system sketch and analytical results of rigid
body kinematics

In Eq. (3) the definition of engineering strain in x-direction was given for a geometric linear calculation.
Substituting the results from kinematics from Figure 24 we obtain the strain depending on ϕ:
∂u
ε ( x) = = cos ϕ − 1 (8)
∂x
Actually we expect a zero strain component in x-direction for the above problem. In the following Tab. 2
we show that this holds only approximately for very small angles ϕ. We think that the geometric linear
theory looses its validity if rotation angles are bigger the 10°:

Table 2: Resulting strain of a rigid body rotation using a geometric linear calculation

ϕ 1 5 10 45
ε -0,0002 -0,0038 -0,0152 -0,2929

One idea to solve the problem from above is to define the strain in a different way, i.e. using a different
strain measurement, like it is given for example in Eq. (9) taken from [5]. This definition of strain is called
Green-Lagrange strain. It is a nonlinear strain measurement with respect to the displacements:
2 2
∂u 1  ∂u  1  ∂v  1 1
ε ( x) = +   +   = cos ϕ − 1 + (cos ϕ − 1) + (sin ϕ )
2 2
(9)
∂x 2  ∂x  2  ∂x  2 2
Using this strain measurement results in zero strain for all angles ϕ so a reasonable correct result is
calculated. To use this definition of strain a geometric nonlinear analysis is required.

Table 3: Resulting strain of a rigid body rotation using a geometric nonlinear calculation

ϕ 5 10 45 90
ε 0 0 0 0

Recommendation:
ANSYS: If the user is not sure to include or exclude the geometric nonlinear effects in the
simulation, a geometric nonlinear calculation is for sure always the better
choice.
However, if a linear calculation has already been performed and the resulting
strains are small for example below 2% the error in the analysis will be small
and acceptable. If more than 5% strain is indicated a geometric nonlinear
analysis should be performed to obtain better results.

ANSYS Workbench: By default, ANSYS Workbench gives the following information if the effect of
geometric nonlinearities should be included in the analysis:

Figure 25. ANSYS Workbench information to activate geometric nonlinearities

Postprocessing
In this chapter we first summarize basic knowledge the user should have considering postprocessing
especially in ANSYS. Most features are not yet available for ANSYS Workbench.

User knowledge during postprocessing

Nodal and element solution

It follows directly from the theory of finite elements that discrete values of the degrees of freedom are
calculated and available only at the nodes of the model. Hence, they can only be displayed with the nodal
solution option of ANSYS. Derived quantities - such as element strains and stresses - are calculated at the
integration points which are located somewhere inside each element and values are extrapolated to the
nodes. Strains and stresses should first be displayed with the element solution option of ANSYS. The non-
smoothness of the strain and stress field can only be recognized if strains and stresses are shown with the
element solution option. However, it is possible to postprocess strains and stresses also with the nodal
solution option which simply means that the extrapolated element values of the integration points are
averaged at the nodes. This is illustrated in Figure 26.
Node Integration points
Calculation of the values of derived quantities
such as strains and stresses at the nodes:

∑ Value
Element
Value Node = Elem.
∑
Elem.

Figure 26. How element quantities like strains and stresses are calculated at the nodes

What we have outlined so far should be illustrated by a little example shown in Figure 27. A quarter of a
plate with a hole is modelled by linear elements and subjected to a traction force. Only one material is used.
The contour plots show the stresses in x-direction, displayed with the element and nodal solution option:

Y
System Element solution: σx Nodal solution: σx
(not averaged) (averaged)
X
Figure 27. Results from the element and nodal solution in ANSYS

It can be recognized that the stress field is not smooth in the element solution as it is characteristic for
finite element solutions. Displaying the same quantity with nodal solution results in a smooth contour
distribution due to the averaging process. The element solution is useful to identify high result gradients
within single elements. In those areas a finer mesh may be required.
A second very similar example is shown in Figure 28. This time the structure is made up of two different
materials – a soft material is combined with a stiff material. Clearly, it is not allowed to average the
element stresses at those nodes where elements with different materials touch each other. In ANSYS this is
recognized automatically if the postprocessing is done using the powergraphic option. If the full graphic
option is used all element values at the nodes are averaged even at the material boundary which can be
misleading. The powergraphic option is activated by default in ANSYS.
 soft

stiff

Y
System Element solution: σx Nodal solution: σx Nodal solution: σx
(not averaged) (averaged) (averaged)
X
FULL GRAPHIC POWERGRAPHIC

Figure 28. Results from element and nodal solution using full graphic and powergraphic in
ANSYS

Help from ANSYS during postprocessing

Estimating the discretization error:

The energy error approach
One possibility to estimate the discretization error in ANSYS is using the implemented energy error
approach. The basic ideas of this concept are taken from [6]. At this stage we will just very briefly describe
how the total energy error of a selected domain of the structure is calculated:

At every node n of an element i the following quantity can be computed:

∆σ ni = σ na − σ ni with σ na Æ averaged stress at the node (10)

σ ni Æ extrapolated stress to the node

This quantity is taken to calculate the total energy error of the structure according to:

e
E = 100 with U Æ total elastic strain energy in the selected domain (11)
U+e
e Æ total energy error in the selected domain

{ } { }
Elem
1
e=∑
T

i =1
∫
2 Ve
∆σ ni D -1 ∆σ ni dV

In other words the total energy error weights the different ∆σni using an energy formulation to obtain a
global measurement of the discretization error for the selected domain. It is calculated in percent.
 Estimating the discretization error:
Comparing the extrapolated and averaged stresses
Recall equation (10): ∆σni = ∆σna − σni. An easy way to estimate the discretization error is to take the
maximum value of the computed nodal quantity ∆σni in each element and define it as the characteristic
value for each element i. This local measurement of the discretization error is called SDSG:

SDSG = Max { ∆σni } (12)

Example of measuring the discretization error

The two outlined possibilities to measure the discretization error will be demonstrated in a little example.
The system with boundary conditions and the analytical solution is shown in Figure 29.

Y
F=20 σL=2 σR=20
X AR=1

AL=10

Figure 29. System with analytical solution to be compared with different finite element
discretizations

In Figure 30 the resulting stress distribution in x-direction is shown together with the local element quantity
SDSD and the global error energy to measure the discretization error:

σx σx

SDSG SDSG

E = 30.943 % E = 7.8457 %

Figure 30. Results of measuring the discretization error locally and globally in ANSYS
Classic
Clearly, the right discretization in Figure 30 is better than the left one. However, only postprocessing SGSG
gives one an idea in which regions the mesh should be refined. The total error energy just indicates how
good or bad a discretization is in general.
Unfortunately, both methods to measure the discretization error just work for linear static analyses.

Adaptive methods to minimize the discretization error

With an adaptive method you are able to minimize the error in a finite element analysis step by step
running a few analyses one behind the other. The following adaptive methods are implemented in ANSYS:
• p-method (ANSYS):
The idea of the p-method is to use higher order shape functions within the elements. However, the
discretization of the problem will be unchanged, i.e. for all analyses the same mesh is used. You may
for example start with an analysis using linear elements, followed by one with quadratic elements,
cubic elements and so on. Up to eight-order-polynomials may be used as shape functions to formulate
finite elements.
• h-method (ANSYS and ANSYS Workbench):
The idea of the h-method is to refine the mesh automatically in regions where certain results are of
interest. The order of the elements will remain unchanged, just an adaptive discretization is performed.
In Figure 31 we show the results of using the p-method and the h-method, respectively. The result quantity,
which is of interest, may be a characteristic stress. It can be recognized, that both methods iteratively
improve the accuracy of the obtained results:

• p-method: σ
25

24
2 3 4 5
• h-method: Order of the shape function

σ
25

24

23

22

21

20
Loop-1 Loop-2 Loop-3 Loop-4
Loop-1 …Æ… Loop-3 Mesh refinement loop

Figure 31. Adaptive methods in ANSYS and ANSYS Workbench

 Evaluating the residual forces / moments in a nonlinear analysis
Since for most nonlinear problems analytical solution are not available it is difficult to validate a nonlinear
numerical result coming from the finite element method.
But what the user always can do is to check the quality of the state of equilibrium in the system. Remember
the third chapter of this paper: Here we noted that during the nonlinear solution process the imbalance
residual forces / moments are iterated to become acceptable small. Numerically they will never be
perfectly zero. So a reasonable question would be how “big” the residual forces are in the end of a
converged nonlinear solution.
Clearly, the residual forces / moments should be closed to zero. Within ANSYS in its actual version (8.0)
the user has indeed the chance to postprocess the residual forces / moments. An example is shown in the
following Figure 32.

Figure 32. Residual forces in x-direction of a converged nonlinear solution step in ANSYS

What we see is a contour plot of the residual forces in x-direction of the last converged solution step of a
nonlinear analysis. It can be recognized that very locally there is a certain amount of disequilibrium in the
structure. However, if the user wants to improve the accuracy of the nonlinear result it is possible to iterate
the residual forces / moments to a smaller value.
Note, that the residual forces / moments can be displayed for every equilibrium iteration within the
Newton-Raphson iteration. Hence, is possible even to judge about the state of equilibrium or disequilibrium
in a solution step which has not yet converged.

Conclusion
In this paper we tried to discuss several modelling aspects to keep in mind when doing a finite element
analysis with ANSYS and ANSYS Workbench to improve the accuracy of the results. For the three typical
steps within a finite element analysis – the preprocessing, the solution and the postprocessing – we
provided some basic finite element knowledge the user should have to be able to build up a reasonable
finite element model for the considered problem.
Considering the preprocessing we discussed the finite element results of different element types, element
shapes and element orders. If available, numerical solutions have been compared with analytical results.
We also described briefly the correct element coupling of different element types. In addition to that, we
showed intelligent warning messages of ANSYS and ANSYS Workbench to minimize modelling mistakes.
Especially the solution of nonlinear problems requires some knowledge about the iterative solution scheme
of ANSYS and ANSYS Workbench. In this paper we discussed briefly the background of the Newton-
Raphson-Method. The limit between a geometrical linear and nonlinear calculation has also been shown.
Finally we mentioned some aspects of the powerful postprocessing of ANSYS and ANSYS Workbench.
Several features are available to check every finite element result for its numerical correctness like error
estimators or using adaptive finite element methods.
 References
[1] ANSYS Elements Reference, Release 6.1, Swanson Analysis Systems, Inc., 2001
[2] ANSYS Theory Reference, Release 6.1, Swanson Analysis Systems, Inc., 2001
[3] Pflüger, A., Elementare Schalenstatik, 5. Auflage, Springer-Verlag, 1981
[4] Schneider, K.-J., Bautabellen für Ingenieure, 11. Auflage, Werner-Verlag, 1994
[5] Wriggers, P., Nichtlineare Finite-Element-Methoden, Springer-Verlag, 2000
[6] Zienkiewicz, O.C., Zhu, J.Z., “A Simple Error Estimator and Adaptive Procedure for Practical
Engineering Analysis”, International Journal for Numerical Methods in Engineering,
Vol. 24, pp. 357-367, 1987