Sunteți pe pagina 1din 11

Thin-Walled Structures 41 (2003) 179–189

www.elsevier.com/locate/tws

The finite element method for thin-walled


members—basic principles
M.C.M. Bakker a, T. Peköz b,∗
a
Faculty of Architecture, Building and Planning, Eindhoven University of Technology,
The Netherlands
b
School of Civil & Environmental Engineering, Cornell University, Ithaca, NY 14853, USA

Abstract

The application of the finite element method to thin-walled structures often requires non-
linear analysis. Whereas in linear finite element analyses errors are easily made, this is even
more so in the non-linear analyses. This paper focuses on possible sources of error in linear
and non-linear finite element solutions, and gives suggestions how to check and prevent
these errors.
 2002 Published by Elsevier Science Ltd.

Keywords: Finite element method; Non-linear analysis; Errors

1. Introduction

The finite element method is a much used tool in current research. Although it is
relatively easy to get results, it requires care to guarantee good results. Table 1 gives
an overview of possible errors, which might occur during (linear and non-linear)
finite element analysis. These errors will be elaborated in this paper.
This paper is based on experiences in teaching the application of linear and non-
linear finite element method to structural engineering students. While the paper
focuses on practical aspects which may be simple and self-evident for experienced
finite-element users, it is hoped that the information given is useful for inexperienced
users. For more extensive guidelines to good finite element practice, see Refs. [1–3].
When users are unfamiliar with (non-linear) finite element analyses, it is best to


Corresponding author: Tel.: +1-607-255-6366; fax: +1-607-255-4828.
E-mail address: tp26@cornell.edu (T. Peköz).

0263-8231/03/$ - see front matter  2002 Published by Elsevier Science Ltd.


doi:10.1016/S0263-8231(02)00086-1
180 M.C.M. Bakker, T. Peköz / Thin-Walled Structures 41 (2003) 179–189

Table 1
Overview of possible errors

Reality

Idealization error

Mechanical model PREPROCESSING

Input error
Discretization error: equilibrium approximated
Geometry error: geometry approximated
Shortcomings in element formulation
Program bugs

Finite element model SOLVING

Solution error
Convergence error
Program bugs

Nodal displacements

Program bugs

Derived results according to FE model

Rendering error: postprocessor inter/extrapolates differently than FE model: integration


points→nodes→ contour plots
Program bugs

Results according to postprocessor POSTPROCESSING

Interpretation error: postprocessor shows something else than is expected, for instance
averaged instead of unaveraged stresses

Interpretation of results

get acquainted with the program to be used and the interpretation of results by run-
ning some simple problems for which exact (analytical) solutions are known. There-
fore, the types of errors discussed in this paper will be illustrated with very simple
examples, which provide insight in some basic issues encountered in the modelling
of thin-walled structures. In these examples, only beam and truss elements are used,
but similar results can be found when running the examples with shell elements.
The finite element method is an approximate method in which equilibrium and/or
compatibility conditions can be approximated. This paper will limit the discussions
to finite element programs in which only equilibrium is approximated.
M.C.M. Bakker, T. Peköz / Thin-Walled Structures 41 (2003) 179–189 181

2. Idealization errors

2.1. Choosing a mechanical model

Idealization errors are caused by a faulty idealization of the real structure, loads
and supports in a mechanical model. During modelling the following aspects have
to be decided:

앫 the choice of mechanical approach describing the behaviour, e.g. beam theory,
shell theory or 3D stress theory. This choice implies how the geometry of the
structure will be described, namely as a line, a plane or a solid.
앫 assumptions that may be made regarding the magnitude of deflections, strains and
rotations. This determines what type of geometrical non-linear analysis should
be run.
앫 how the material behaviour may be modelled.
앫 whether the influence of initial stresses should be accounted for, and how these
residual stresses are distributed through the model.
앫 how accurately the geometry of the structure must be modelled in order to get
accurate results. Magnitude and distributions of deviations from the ideal
geometry (initial imperfections) must be considered.
앫 how loads and supports should be modelled.

When idealising a structure it is important to know what one wants to learn from
the mechanical model. The objective is not to describe reality as accurately as poss-
ible, but to find the simplest model resulting in a sufficiently accurate description of
reality. Unfortunately, this will often require a lot of insight in the behaviour of the
structure. In case when it is necessary to build a very sophisticated model, it is best
to start the analysis with a simple model, and use a step by step approach to increase
the complexity of the model.
The choice of the kind of mechanical model needed (corresponding to the choice
of the type of element), the appropriate modelling of material properties, as well as
initial stresses and initial imperfections are discussed in an accompanying paper by
Sarawit, Kim, Bakker and Peköz. Only the types of nonlinear analysis will be dis-
cussed here.

2.2. Types of geometrical non-linear analysis

With respect to geometrical non-linear behaviour a distinction can be made


between first order analysis, second order analysis, large deflection analysis, large
rotation analysis, a large strain analysis and an eigenvalue buckling analysis (see
Table 2 and Fig. 1). The well known first order analysis is a geometrical linear
analysis, in which equilibrium is formulated with respect to the undeformed shape.
In the other, geometrical nonlinear types of analysis equilibrium is formulated with
respect to the deformed state. A second order analysis accounts for stress stiffening
effects (P-δ effects). In a large displacement analysis membrane stresses developing
182 M.C.M. Bakker, T. Peköz / Thin-Walled Structures 41 (2003) 179–189

Table 2
Types of geometrical nonlinear analysis

Type of analysis Displacements Rotations Strains

First order (linear) Very small Very small Small ( ⬍ 5%)


Second order (stress Small (negligible membrane Small ( ⬍ 100) Small ( ⬍ 5%)
stiffening) stresses due to out of plane
deflections, displacements smaller
than the thickness of the
beam/plate/shell)
Large displacement Large (significant membrane Small ( ⬍ 100)
stresses due to out of plane
deflections, displacements larger
than the thickness of the
beam/plate/shell)
Large rotation Large Large Small( ⬍ 5%)
Large strain Large Large Large

due to out of plane deflections can also be described, but rotations and strains still
have to be small. A large strain analysis is the most general geometrical non-linear
analysis, in which neither strains, nor rotations or displacements need to be small.
A large strain analysis is the only type of analysis in which changes in thickness or
area are accounted for.
An eigenvalue buckling analysis is a special kind of geometrical non-linear analy-
sis. Usually it is based on a second order formulation. Using the assumption of linear
load-deformation behaviour up to the attainment of the buckling load the factor with
which the loads applied to the structure should be multiplied in order to get a non-
unique solution (the eigenvalue) can be determined. For some structures the load
determined by this analysis represents the buckling load of the structure. However,
if prebuckling deformations are not small, the load determined does not have much
physical relevance (see for instance the eigenvalue calculation in the structure where
snap-through occurs). So one should keep in mind that an eigenvalue calculation is
a mathematical trick, which can always be performed, regardless of the physical sig-
nificance.
In most cases a linear analysis will not give useful results for the analysis of cold-
formed structures, but it is good practice to start every non-linear analysis with a
linear analysis. If one wants to model the non-linear load deflection behaviour of
thin-walled structures, including post-buckling behaviour, one needs at least a large
deflection analysis or a large rotation analysis, since a second order analysis cannot
describe membrane stresses which develop due to out-of plane deflections. It is also
possible to use a large strain analysis, but in most cases it is much simpler to use
a large displacement or large rotation analysis. Note that in defining loads in a large
rotation analysis (contrary to first and second order analysis) a distinction is made
between follower forces (element loads), where the direction in which the loads act
is dependent of deformations of the structure, and nodal loads, where the direction
of the loads does not change.
M.C.M. Bakker, T. Peköz / Thin-Walled Structures 41 (2003) 179–189 183

Fig. 1. Types of geometrical non-linear analysis.

It is the responsibility of the user to make sure that the right type of analysis is
used. A finite element program does not give warnings when displacements, rotations
or strains are so large that the assumptions on which the analysis is based are viol-
ated, but just gives unreliable results!
184 M.C.M. Bakker, T. Peköz / Thin-Walled Structures 41 (2003) 179–189

3. Input error

Input errors are errors made in the pre-processing phase, when building the finite
element model. It is best to check input errors by running a linear analysis, before
starting with the non-linear analysis.
Some simple checks on input error are to inspect a plot of the finite element mesh
and to run a linear analysis with the model loaded by unit loads. Furthermore the
material and geometry properties should be checked carefully, as well as the applied
loads and boundary conditions. For more checks see Nafems [1,2].

4. Discretization error

In the finite element method the nodal displacements and/or rotations are obtained
from nodal equilibrium equations, resulting in the best approximation of the equili-
brium conditions throughout the structure for the given finite element mesh. If the
mesh is too coarse, the equilibrium conditions are satisfied poorly. This is called a
discretization error.
A discretization error can be recognised from stress discontinuities between
elements. A discretization error may also be recognised at boundaries of the finite
element model, where it can be observed how much calculated stresses deviate from
known stresses. Furthermore, it is good practice to study the influence of mesh
refinement on the results.
When one builds a finite element model an important question is how dense the
mesh should be for sufficiently accurate results. The required mesh density depends
on the stress gradients (larger stress gradients need a finer mesh), and the question
as to whether the mesh should result in adequate stress results, or only in adequate
stiffness results. Poorly shaped quadrilateral elements may lose their accuracy, and
thus may increase the discretization error. Poorly shaped linear elements are less
shape sensitive, since both ideally shaped and poorly shaped linear element must be
able to describe constant stress states. In a physically non-linear analysis (using non-
linear material models) with shell elements, the discretization error is very much
dependent on the number of integration points used over the area of a shell element.
Due to discretization errors, the von Mises stresses in points different from the inte-
gration points may far exceed the yield stress. One should be aware, that both an
eigenvalue analysis and a physical and/or geometrical nonlinear analysis require a
finer mesh (sometimes much finer) than a linear analysis. Fig. 2 illustrates this for
a beam element.

5. Geometry error

A geometry error occurs when the geometry of a structure cannot be described


exactly by the finite element mesh used. An example of a geometry error is the
modelling of a beam with varying height by elements with constant stiffness proper-
M.C.M. Bakker, T. Peköz / Thin-Walled Structures 41 (2003) 179–189 185

Fig. 2. Discretization error and error due to shortcomings in element formulation.

ties over the length of the elements. Another example is the modelling of a curved
shell with flat shell elements.
A geometry error is of interest only when small changes in geometry have a large
influence on the results. For the modelling of thin walled structures, the correct
modelling of the corner radius might be an issue for some loading conditions (for
instance web crippling). In addition, one should be aware of the fact that small
changes in the shape of holes may have a large influence on the magnitude of
stress concentrations.
In general, a geometry error can only be checked by studying results for varying
mesh densities. A geometry error does not necessarily result in stress discontinuities.
186 M.C.M. Bakker, T. Peköz / Thin-Walled Structures 41 (2003) 179–189

The finite element analysis may result in a solution, which is in perfect equilibrium
(no discretization error), however a geometry error is present due to the use of a
wrong geometry!

6. Shortcomings in element formulations

When modelling physical non-linear behaviour of thin–walled structures with shell


elements, the number of integration points over the thickness of the shell element
is even more important than the number of integration points over the area of a shell
element. Whereas a limited number of integration points over the area of a shell
element may be compensated by using a fine mesh, this is not the case for a limited
number of integration points over the thickness of the shell element. Too few inte-
gration points over the thickness will result in an inaccurate prediction of the moment
of first yield, the fully plastic bending moment and the interaction between membrane
stresses and bending moment. This type of error is called a shortcoming in the
element formulation (see Fig. 2).
Other types of shortcomings in element formulations which might be encountered
in shell elements are shear locking and membrane locking. Also one should be aware
of the possibility of mesh locking due to large plastic strains (Zienkiewicz and Taylor
[4], Cook, Malkus and Plesha, [5]). If no results of appropriate benchmark tests (as
for instance published by NAFEMS) are available, one should assure oneself that
the elements to be used perform well for the intended application, by running appro-
priate tests.

7. Solution and convergence error

In literature concerning the solution of linear and non-linear nodal equilibrium


equations in finite element analysis (see for instance Bathe [6], Cook, Malkus and
Plesha [5] and Hinton [3]), much attention is given to the discussion of solution
errors, that is, errors occurring in the calculated nodal displacements due to ill con-
ditioning of the nodal equilibrium equations and the finite number of digits the com-
puter uses to represent floating-point numbers. In literature on non-linear analysis
the emphasis is laid on convergence errors, that is errors occurring in the calculated
nodal displacements in a nonlinear analysis due to the iterative solving of the nonlin-
ear equations, and not much is said about the influence of solution errors. It is prob-
able that solution errors do not directly affect the accuracy of the solution, but may
cause convergence problems.
Both in linear and nonlinear analysis if the equations are solved exactly, the
internal nodal forces calculated from the calculated nodal displacements are in exact
equilibrium with the applied external nodal forces. If the equations are not solved
exactly, the internal and external residual nodal forces will not be in exact equilib-
rium and residual nodal forces will occur.
Unfortunately, both in linear and nonlinear analysis small residual forces do not
M.C.M. Bakker, T. Peköz / Thin-Walled Structures 41 (2003) 179–189 187

guarantee accurate displacements (and stresses), and large residuals do not necessar-
ily imply inaccurate displacements (and stresses), see also Table 3. In linear analysis,
a well conditioned set of equations and small residuals will guarantee an accurate
solution. In nonlinear analysis, a combination of sufficiently tight residual force cri-
teria and displacement based convergence criteria will guarantee an accurate solution.
In other cases, engineering judgement will be needed to determine whether the sol-
ution is acceptable or not.

8. Rendering error

After the nodal displacements have been calculated, the displacements in the
elements, and hence the stresses and strains in the elements can be calculated. One
should be aware that the way in which a postprocessor presents these results does
not always correspond to the results according to the finite element theory. In a
postprocessor the stress distribution over an element will often be determined with
a linear interpolation between the stresses in the corner nodes, regardless of the
presence of intermediate nodes. Sometimes the stresses in the nodes are assigned
the values of the stresses in integration points. In many cases, rendering errors will
become apparent only in very coarse meshes, as may be used in simple problems
carried out in order to get accustomed to a program.

9. Interpretation error

When looking at stress results it is important to be aware whether one looks at


contour plots of element stresses (unaveraged stresses) or nodal stresses (averaged
stresses). Plots with averaged stresses may look much nicer but may obscure large
discretization errors.
Non-linear load-deflection calculations result in dots in a load-displacement dia-
gram. Interpolating a curve through these dots, as many post-processors do, can
result in erroneous results. Sharp bends in such curves should be mistrusted, until

Table 3
Influence of residuals on accuracy of solution

Nodal equilibrium Nodal displacements

Small residuals: Good condition/small Accurate Accurate


iterative displacements
Small residuals: Bad condition/large Accurate Possibly inaccurate
iterative displacements
Large residuals: Good condition/small Inaccurate Possibly accurate
iterative displacements
Large residuals: Bad condition/large Inaccurate Probably inaccurate
iterative displacements
188 M.C.M. Bakker, T. Peköz / Thin-Walled Structures 41 (2003) 179–189

explained by the model behaviour. It is especially important that a sufficiently large


number of dots is calculated to determine the ultimate load, otherwise the failure
load may be severely underestimated.

10. Program bugs

Program bugs may occur in any finite element program. It is good practice to
record the program version which has been used for an analysis, so that when a
program bug is reported in a newsletter, it is possible to check whether the results
of a certain analysis might have been affected by this bug.
Program bugs are more likely to be encountered in the more exotic options of a
program, for instance the application of initial stresses. Therefore when one is using
a new option for the first time, it is wise to check this option first on a simple
problem for which the exact solution is known.

11. Preferable order in checking errors

1. Check input errors (before starting a nonlinear analysis).


2. Check solution/convergence errors.
3. Check discretization errors.
4. Check idealization errors as far as possible (check assumptions on magnitudes of
deflections, rotations and strains). When one encounters unexplainable results,
keep in mind that there might be errors in the element formulation, program bugs,
rendering errors and interpretation errors. When comparing finite element results
with experimental or analytical results, the error might also be in the experimental
or analytical results! Note that stress discontinuities between elements may indi-
cate either a solution/convergence error, or a discretization error. Therefore it is
best to check solution/convergence errors before checking discretization errors.

12. Conclusions

The avoidance of errors and mistakes in finite element analysis requires a critical
attitude towards the obtained results. Much care should be given to checking possible
errors as discussed in this paper.

References

[1] Nafems. Guidelines to finite element practice. Glasgow, UK: Nafems, 1992.
[2] Nafems. A finite element primer. Glasgow, UK: Nafems, 1992.
[3] Hinton E. Introduction to nonlinear finite element analysis. Glasgow, UK: Nafems, 1992.
M.C.M. Bakker, T. Peköz / Thin-Walled Structures 41 (2003) 179–189 189

[4] Zienkiewicz OC, Taylor RL. The finite element method. Solid and fluid mechanics, dynamics and
non-linearity, 2. Maidenhead, UK: McGraw-Hill Book Company Europe, 1991.
[5] Cook RD, Malkus DS, Plesha ME. Concepts and applications of finite element analysis. New York:
John Wiley & Sons, 1989.
[6] Bathe KJ. Finite element procedures in engineering analysis. Englewood Cliffs, NJ: Prentice-Hall
Inc., 1982.

S-ar putea să vă placă și