Professional Documents
Culture Documents
2004 Int Ansys Conf 24
2004 Int Ansys Conf 24
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 Preprocessing - wrong element type - wrong element coupling - wrong idealization - wrong calculation discipline - wrong boundary conditions - wrong convergence criteria - wrong result coordinate system - wrong selection of components - wrong interpolation of results Finite Element Software - bad element formulation - bad meshing algorithm - no warnings and errors - wrong calculation algorithm - inaccurate equilibrium iteration - no warnings and errors - wrong result averaging - wrong displaying of results - no warnings and errors
Solution
Postprocessing
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) (2)
u ( x) = a0 + a1 x + a 2 x 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:
( x) =
du dx
and
( x) = E ( x)
(3)
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
1
u ( x) = a0 + a1 x
F
1 X
u ( x) = a0 + a1 x + a2 x 2
ux
Results: analytic
ux
( x) = E
du dx
FEM
( x) = E
du dx
Figure 1. One dimensional truss problem: linear and quadratic elements and their consequences With E as Youngs Modulus, F as applied force, and A(x) as the cross sectional area the analytic solution is
F u ( x) = E
A( x) dx
0
and
( x) =
F 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: ANSYS Workbench: In general, the user should prefer to take a quadratic element if possible. By default, SOLID186/SOLID187 is used as a quadratic 3D Solid-Element
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 (default for BEAM3 / BEAM4) bending stress x TIMOSHENKO (standard for BEAM188 / BEAM189) bending stress x
shear stress xy
shear stress xy
deformed configuration
Y X
deformed configuration
dmax_y= 0.007441
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:
d max_ y =
5 ql 4 384 EI
d max_ y = 0.007440 .
(5)
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: ANSYS Workbench: In general, the users should take BEAM188/189 if possible. By default, BEAM188 is used.
Z 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 REISSNERMINDLIN 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, =0.3, t=15, 0<s<223.61 and =63.43. From the geometry we can calculate r and r0. Note, that r 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:
r 0
r
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:
0 2 Q sin 2 s r 0 r e 0 cos 0 u r (s ) = Et r 0
with
0 = 3(1
4
)t
r 0
2
(6)
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
Figure 7. Shell problem: Displacement in r-direction (over-scaled) using SHELL63 and SHELL181
Recommendation: ANSYS: ANSYS Workbench: In general, the users should take SHELL181 if possible. By default, SHELL181 is used.
Y X
B = 1 H = 1 M(F) = 1/6 W = 1/6 x= M/W = 1
Figure 8. Simple supported structure subjected to pure bending: geometry model and analytical solution PLANE182 (linear) x Let us first look at the results of quadrilaterals: 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. 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. Using quadratic elements results in a correct bending stress distribution over the thickness. Enhanced Strain Formulation Figure 9. Results of bending stresses for linear and quadratic elements using a mesh of quadrilaterals x
PLANE182 (linear)
PLANE183 (quadratic)
PLANE182 (linear)
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 used as triangles even if the Enhanced Strain Formulation is used. Using quadratic elements the results are correct even when triangles are used.
PLANE183 (quadratic)
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: ANSYS Workbench: In general, the users should take PLANE183, SOLID186/SOLID187 if possible. By default, SOLID186/187 is used.
Beam Shell
Solids
Figure 11. Incorrect coupling of elements with different degrees of freedom However, there are a few possibilities to correctly transmit the beams or shells 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 Shell
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:
MPC
Figure 13. Correct element coupling using the MPC method Using the MPC technique, ANSYS generates internally some coupling equations 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:
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
Y X
B = 1 H = 1 M(F) = 1/6 W = 1/6 x= M/W = 1
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: PLANE182 (linear) x
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.
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: F R criterion (small number) u Equilibrium is obtained if: u u |R| < criterion
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].
Figure 21. The Newton-Raphson iteration of the imbalance residual forces and moments
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.
R F 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.
y, v A
u ( x) = x (cos 1) v( x) = x sin
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 :
( x) =
u = cos 1 x
(8)
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 . 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 1 -0,0002 5 -0,0038 10 -0,0152 45 -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:
(9)
Using this strain measurement results in zero strain for all angles 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 5 0 10 0 45 0 90 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:
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.
Node
Integration points Calculation of the values of derived quantities such as strains and stresses at the nodes:
Value
Value Node
= Elem.
Element
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
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
System
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:
i a i n =n n
with
a n i n