MACNEAL - The Treatment of Shell Normals in Finite Element Analysis

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

Finite Elements in Analysis and Design 30 (1998) 235—242

The treatment of shell normals in finite element analysis


Richard H. MacNeal*, Charles T. Wilson, Robert L. Harder, Claus C. Hoff
The MacNeal-Schwendler Corporation, 815 Colorado Blvd., Los Angeles, CA 90041, USA

Abstract

It is shown that the method used by MSC/NASTRAN in the past to treat the singularity of rotation about the normal
to a shell’s surface is defective. A new method is described in which a unique normal is constructed at each node and
automatically constrained. Good results are shown for a case which formerly gave excessive deformation due to the
non-equilibration of twisting moments. ( 1998 Elsevier Science B.V. All rights reserved.

Keywords: Shell normals; Shell elements; MSC/NASTRAN; QUAD4; P elements

1. Introduction

In shell analysis, strains are derivative functions of three components of translation and two
components of rotation. The third component of rotation, rotation about the normal to the shell
mid-surface, does not enter into the calculations. On the other hand, the degrees of freedom at
nodes, and the distributed hierarchical degrees of freedom on edges joining p-elements, include
three components of translation and three components of rotation. The question then arises as to
how to treat the nodal components of rotation so as to neither underconstrain nor overconstrain
the finite element model.
Consider a flat plate. In this case there will be no moment about the normal at any node, so that
rotation about the normal must be constrained to avoid singularity. If, on the other hand, we
consider the right-angled intersection of two plates, moments can be transferred about all three
axes along the intersection, and any constraint will be an overconstraint.
There are at least two solutions to the dilemma which has been presented. One solution is to
define a unique ‘normal’ at each node, make sure that the attached elements produce no moments
about it, and constrain it. In this case the grid-point ‘normal’ may not be normal to any of the
attached shell elements. For the example of two plates intersecting at a right angle, the ‘normal’ will

* Corresponding author.

0168-874X/98/$19.00 ( 1998 Elsevier Science B.V. All rights reserved


PII S 0 1 6 8 - 8 7 4 X ( 9 8 ) 0 0 0 3 5 - 3
236 R.H. MacNeal et al. / Finite Elements in Analysis and Design 30 (1998) 235—242

be at least 45° away from the normal to one of the plates. No matter. This solution requires that the
direction of the ‘normal’ be computed as a weighted average of the normal directions for the
elements connected to the node [1—3]. Vu-Quoc and Mora [1] describe a version of this method in
which the direction of the ‘normal’ is assumed to be a continuous field.
In the original development of MSC/NASTRAN’s shell elements in the 1970s, we considered the
computation of a special normal direction at each node to be an unnecessary task. Instead, we
allowed moments about all three axes at a node and invoked an existing subroutine, AUTOSPC,
which examines diagonal 3]3 partitions of the stiffness matrix for near-singularity. If the max-
imum ratio of the three principle stiffnesses for translation or rotation is high enough, AUTOSPC
constrains the component nearest to the direction of minimal principle stiffness. Otherwise it leaves
the node alone. In the case of the QUAD8 element [4,5], we took additional steps within each
element to ensure that the moment about the normal to the element’s mid-surface is zero at each
connected node. Note that this is not the same as ensuring that the moment about some unique
‘normal’ is zero.
Thus, there are at least two methods for treating shell normals at nodes. Does it make
a difference which one is used? For a long time we did not think so. We had discovered a related
problem with the QUAD4 element in running the MacNeal—Harder test problems [6] and fixed it.
The test problem which failed was the twisted ribbon, see Fig. 1, for which the displacement at the
free end was found to be orders of magnitude too large. The reason for the failure is that the
QUAD4 is a flat element parallel to the diagonals connecting nodes, see Fig. 2, so that, for
a warped element, bending moment transfer produces a component about the shell normal. The fix
was to rotate the moment about an edge into a moment about a line joining the two connected
nodes [7]. Note that this fix is local within each element and does not require any consideration of
the relative positions of adjacent elements.

Fig. 1. Twisted beam problem. Length"12.0, width"1.1, depth"0.32, twist"90°.

Fig. 2. QUAD4 element with offset nodes.


R.H. MacNeal et al. / Finite Elements in Analysis and Design 30 (1998) 235—242 237

In practical application, our treatment of shell normals appeared to work fine in linear problems.
In non-linear problems, distortion of a flat plate into a curved shell surface can cause a rapid
transition from singularity to substantial non-singularity of the 3]3 diagonal partitions of the
stiffness matrix. To cope with this difficulty we applied a user-supplied stiffness term, K6ROT, to
the normal rotation.
Then in 1991, Ingo Raasch [8], reported poor results with the linear example shown in Fig. 3. He
tried to get better results by refining the mesh, but instead the results got worse. The Raasch hook
was also studied by Knight [9], who reported convergence with decreasing element size only if
transverse shear flexibility was omitted. The difficulty turns out to be the transfer of twisting
moment between flat elements which are not in the same plane, see Fig. 4, or between curved
elements whose normal directions do not coincide along their common edge. In both cases, the
resultant of the twisting moments is in the normal direction. AUTOSPC is set so that no constraint
is imposed if the angle between the directions of the normals is greater than about 10~4 rad. This
limit was not reached, even by going to very small elements, with the result that normal rotation
could support no moment and twisting moment had to be transferred from one element to another
by a vertical couple. The vertical couple induced transverse shear strains whose effects became
larger as the elements were made smaller. This time there was no local fix and we had to admit that
we needed better treatment of the shell normal.

Fig. 3. The Raasch Hook problem.

Fig. 4. Transfer of twisting moment between elements not in the same plane.
238 R.H. MacNeal et al. / Finite Elements in Analysis and Design 30 (1998) 235—242

2. The new MSC/NASTRAN treatment of shell normals

In MSC/NASTRAN each gridpoint (node) has three components of translation and three
components of rotation which are oriented in a global coordinate system which is Cartesian but
arbitrarily oriented at each node by user input. Each shell element also has its own element
coordinate system which is fixed in orientation with respect to a flat median surface ‘near’
the element’s midsurface. To repair the problem just described, we needed to introduce a new
normal coordinate system at each node, whose ‘normal’ direction would be approximately normal
to the shell elements joined at the node, and then we had to make sure that no moments
were applied to rotation about the normal direction so that we could constrain it without affecting
the solution.
The transformations from global (g) to normal (n) to element (e) coordinate systems are formally
described as follows:

(1)

where Mu N is the vector of translations and Mh N is the vector of rotations in the global coordinate
g g
system, [¹ ]is a standard 3]3 coordinate rotation matrix, and
ng

(2)

where [A], [B] and [C] are non-standard 3]3 matrices. Note that Mu N and Mh N can be eliminated
n n
by combining Eqs. (1) and (2). Solution for Mu N and Mh N then gives a revised global-to-
e e
element coordinate transformation. The matrices [B] and [C] are constructed so that there
is no moment about h , the normal component of rotation. As a result the 3]3 diagonal partitions
n3
of the stiffness matrix for rotational components are singular. The subroutine AUTOSPC picks
this up and places a single point constraint on the component of Mh N nearest to the normal
g
direction.
The matrix [¹ ] is constructed so that the direction of the third component of Mu N is the
ng n
weighted average direction of the normals to the elements connected to the node. For curved
elements, the direction of the normal to each element is computed at the node. The contribution of
each element is weighted according to its included angle at the node.
To construct the matrices [A], [B] and [C] we note first that the third component of rotation in
the element coordinate system, h , exists as an independent variable only if the element uses
e3
drilling freedoms [10]. To compute actual rotations about the element’s normal, we define a new
variable, a , and compute it from translational motion parallel to the median (x,y) plane of the
3
element

A B
1 Lv Lu
a " ! . (3)
3 2 Lx Ly
R.H. MacNeal et al. / Finite Elements in Analysis and Design 30 (1998) 235—242 239

Likewise we can define

A B
1 Lw Lv
a " ! , (4)
1 2 Ly Lz

A B
1 Lu Lw
a " ! , (5)
2 2 Lz Lx

to form the vector Ma N, i"1,2,3.


i
Then, in place of the bottom-half of Eq. (2), we write

Mh N"[¹1]Mh N#[¹2]Ma N, (6)


n e i
where [¹1] and [¹2] are partitions of the standard 3]3 normal-to-element rotation matrix

C D
¹ ¹ ¹
11 12 13
[¹ ]" ¹ ¹ ¹ , (7)
ne 21 22 23
¹ ¹ ¹
31 32 33 ne
such that

C D
¹ ¹ 0
11 12
[¹1]" ¹ ¹ 0 (8)
21 22
0 0 ¹
33 ne
and

C D
0 0 ¹
13
[¹2]" 0 0 ¹ . (9)
23
¹ ¹ 0
31 32 ne
Substitution into Eq. (6) then gives

G H C DG H G H
h ¹ ¹ h ¹
n1 " 11 12 e1 # 13 a , (10)
h ¹ ¹ h ¹ 3
n2 21 22 e2 23
h "¹ h #¹ a #¹ a , (11)
n3 33 e3 31 1 32 2
which can be solved for the rotations in the element coordinate system,

G H C D AG H G H B
h h
12 ~1
¹ ¹ ¹
e1 " 11 n1 ! 13 a , (12)
h ¹ ¹ h ¹ 3
e2 21 22 n2 23
h " 133 (h !¹ a !¹ a ). (13)
e3 T n3 31 1 32 2
240 R.H. MacNeal et al. / Finite Elements in Analysis and Design 30 (1998) 235—242

If no stiffness is attached to the drilling freedom, h , then there will be no stiffness coupling h to
e3 n3
a and a . If, likewise, there are no drilling stiffnesses for the other elements attached to the node,
1 2
h will be a singular degree of freedom with no moment about it.
n3
Substitution into Eq. (12) for h and h from Eq. (1) and for a from Eq. (3) then gives the
n1 n2 3
components of rotation in element coordinates in terms of rotations in global coordinates and
translations in element coordinates. The inclusion of a in Eq. (12) is the key feature of the new
3
method. Rotation about the normal to the element is needed for the coordinate transformation but,
in contrast to h , a is a stiff degree of freedom supported by membrane action and will not
e3 3
produce excessive deformation.
Computation of a for a curved shell element requires care because the partial derivatives in
3
Eq. (3) are in directions parallel to the median plane which is not the same as the shell’s midsurface.
If we let v dx and u dy represent variations of v and u in the shell midsurface, then it can easily be
,x ,y
shown that

A B A B
1 Lv Lu 1 Lv Lz LuLz
a " ! " v !u ! # "1 (v !u #z h #z h ), (14)
3 2 Lx Ly 2 ,x ,y LzLx LzLy 2 ,x ,y ,x x ,y y

where z and z are the partial derivatives of the normal component of position in the mid-
,x ,y
surface relative to the median plane. The last form of Eq. (14) is exact only if transverse shear strain
is zero.
If we turn our attention to the coordinate transformation for translational degrees of freedom,
we note that [A] in Eq. (2) is the standard coordinate rotation matrix [¹ ] if the corners of the
ne
element are coincident with the nodes. This is not always the case for the MSC/NASTRAN
QUAD4 element which is flat. Thus, the nodes of QUAD4 may be offset in the z direction to its
plane (x,y).
Consequently,

A G HB
a
2
Mu N"[¹ ] Mu N#z !a , (15)
n ne e g 1
O
where z is the amount of offset. Also, by ignoring transverse shear strain relative to rigid-body
g
rotation, we can take
a "w ,
1 ,y (16)
a "!w ,
2 ,x
in place of Eqs. (4) and (5). w, and w, are computed from the values w at the corner nodes. It is
x y
important to note that the transformation indicated by Eq. (15) does not involve coupling to
rotational degrees of freedom. Thus, the new QUAD4 formulation avoids unnecessary and
undesirable coupling between membrane and bending deformations, just as the old one did [7]. It
can be used, even when warped, in a membrane-only mode.
Computation of the final global-to-element transformation requires consideration of the way in
which spatial derivatives, such as v , are computed. This is straightforward, but it is done
,x
differently in each element and will not be pursued here. We will only note that computation of the
R.H. MacNeal et al. / Finite Elements in Analysis and Design 30 (1998) 235—242 241

transformation for hierarchical degrees of freedom along the edges of p elements is done by
collocation at p!1 points along the edge.

3. Twist of a prismatic beam with a circular arc cross-section

A simple example which illustrates the failure due to unequilibrated twisting moment is the
torsion of a prismatic beam with an open cross-section in the shape of a circular arc. We chose
a test problem where the cross-section had an included angle of 64°, a radius of 10, and a thickness
of 1.0. A unit distributed twisting moment was applied at one end and the boundary conditions at
the other end were rigged to simulate infinite length, i.e., differential bending was prohibited.
Young’s modules was taken equal to 107 and Poisson’s ratio was set to 0.3. QUAD4 elements
were used.
Table 1 shows the measured value of rotation at the free end for the new method and the old
method versus the number of elements in the cross-section. It is seen that the solution quickly
converges in the new method and diverges when normal rotations are not constrained in the old
method.
The reason for divergence of the old method when normal rotation is not constrained (as it will
not be by AUTOSPC to avoid overconstraint) is that twisting moment must then be carried across
element boundaries in the cross-sectional direction by a vertical couple which induces transverse
shear strain. Analysis of this situation shows that the strain energy is proportional to (1#t2/w2)
where t is the thickness of the element and w is its width in the cross-sectional direction. w is
inversely proportional to the number of elements so that if t/w<1, the strain energy increases
approximately as the square of the number of elements. Comparison of this theoretical result with
Table 1 shows good agreement.
In the new method, the resultant of the twisting moment at the boundary between elements
is carried by moments about a in each element, i.e., by in-plane bending couples. As can
3
be seen from Table 1, this produces no excess flexibility. Application of the new method to

Table 1
Twist of a prismatic beam with a 64° circular arc cross-section. QUAD4 elements. t"1.0, R"10

Number of elements Included angle per element (°) End rotation! new method End rotation! old
method

4 16 0.9618 1.0909
8 8 0.9754 1.5214
16 4 0.9896 3.2486
32 2 0.9968 10.1581
64 1 0.9992 37.7967
128 0.5 0.9998 148.3510
256 0.25 1.0000 590.5681

!Normalized to the value for 1024 elements in the new method. The theoretical result, including 3D elastic effects not
represented in shell theory, is 0.860% lower.
242 R.H. MacNeal et al. / Finite Elements in Analysis and Design 30 (1998) 235—242

the Raasch hook produces satisfactory results [11], which converge to a valid solution with
decreasing element size.

References

[1] L. Vu-Quoc, J.A. Mora, A class of simple and efficient degenerated shell elements-analysis of global spurious-mode
filtering, Comput. Methods Appl. Mech. Eng. 74 (1989) 117—175.
[2] T.J.R. Hughes, W.K. Liu, Nonlinear finite element analysis of shells: Part I. Three-dimensional shells, Comput.
Methods Appl. Mech. Eng. 26 (1981) 331—362.
[3] H.C. Huang, E. Hinton, New nine-node degenerated shell element with enhanced membrane and shear interpola-
tion, Int. J. Numer. Methods Eng. 22 (1986) 73—92.
[4] R.H. MacNeal, Specifications for the QUAD8 quadrilateral curved shell element, MacNeal-Schwendler Corp.,
Memo RHM-46B, 1980.
[5] H.V. Lakshminarayana, K. Kailesh, A shear deformable curved shell element of quadrilateral shape, Comput.
Struct. 33 (1989) 987—1001.
[6] R.H. MacNeal, R.L. Harder, A proposed standard set of problems to test finite element accuracy, Finite Element
Anal. Design 1 (1985) 3—20.
[7] R.H. MacNeal, Finite Elements: Their Design and Performance, Marcel Dekker, New York, 1994, pp. 438—440.
[8] R.L. Harder, Charts illustrating the Raasch challenge, presented at the Structures Technical Forum at the 1991
MSC World Users’ Conference.
[9] N.F. Knight Jr., The Raasch challenge for shell elements, AIAA Paper No 96-1369, 375) SDM Conf, 1996.
[10] D.J. Allman, A compatible triangular element including vertex rotations for plane elasticity analysis, Comput.
Struct. 19 (1984) 1—8.
[11] C.C. Hoff, R.L. Harder, G. Campbell, R.H. MacNeal, C.T. Wilson, Analysis of shell structures using
MSC/NASTRAN’s shell elements with surface normals, Proc. 1995 MSC World Users’ Conf. Universal City, CA,
8—12 May, 1995, Paper d26.

You might also like