Chapter 5

Quadrilateral Shell Elements.

Flat quadrilateral shell elements have use limitations even in linear analysis,
since a mesh that consists of strictly at elements may be impossibly to construct
over a doubly-curved shell surface. For large deection nonlinear analysis this
deciency becomes more pronounced. Even if the initial mesh satises the at
element restriction, the deformations can become so large that warping of the
elements can be signicant. Finding ways of handling warped element geometries
is thus of fundamental importance for quadrilateral shell elements.
The current research initially set out to develop a non-at quadrilateral
element that was able to satisfy self-equilibrium in a warped conguration. This
proved to be dicult. Finding a basic stiness Kb = A1 LCLT that maintained
self equilibrium was especially troublesome. To achieve this goal one must have
the lumped node forces from a constant stress state f = L to be in self equi-
librium. Two questions arise at this point: What is a constant stress state for a
warped shell element? Is a constant stress state for a warped surface an equilib-
rium state for arbitrary element shapes? These are key questions that remain
to be answered in order to develop satisfactory FF and ANDES elements for
such element geometries. It follows that those formulations have to be further
extended for developing warped shell elements. The construction of the basic
stiness matrix needs special attention because the Individual Element Test is
not clearly dened for curved or warped element geometries.
Restoring equilibrium in the warped element geometry can be done a
posteriori for elements developed with reference to the at projected footprint
by using a projector matrix. This is the approach that has been chosen for the
current quadrilateral element. It allows independent development of the mem-
brane and bending components since these two stiness contributions decouple,
which greatly simplies the development of the element. Both linear and nonlin-
ear projectors have been developed for the quadrilateral element. Both versions
give identical results for linear static problems, but the linear projector is recom-
mended for linear nite element codes because its formulation is much simpler
than that of the nonlinear projector.

1
5.1 Geometric denitions for a quadrilateral element.
This section describes geometric relationships for a quadrilateral element. The
development covers both warped geometry and the at best t element ob-
tained by setting the z coordinate for each node equal to zero. The at pro-
jection relationships are later used for the development of the membrane and
bending stiness of the shell element. The non-at relationships are needed for
the development of the nonlinear projector for quadrilaterals.
A vector r to a point on a nonat quadrilateral element can be parametrized
with respect to the natural coordinates and as

x N 0 0 x
r(, ) = y = 0 N 0 y , (5.1.1)

z 0 0 N z

where
x y1 z1
1


x2 y2 z2
x= , y= , z= , (5.1.2)

x3

y3

z3

x4 y4 z4
and xi , yi and zi denote the global coordinates of node i. Row vector N contains
the usual bi-linear isoparametric interpolation for a quadrilateral [00]. These
functions and their partial derivatives with respect to and are
1
N = [ (1 )(1 ) (1 + )(1 ) (1 + )(1 + ) (1 )(1 + ) ] ,
4
1
N, = [ (1 ) (1 ) (1 + ) (1 + ) ] ,
4
1
N, = [ (1 ) (1 + ) (1 + ) (1 ) ] .
4
(5.1.3)
Using these geometric relations the variation of the position vector r can be
written as
x d + x d x x
dx
 
y y y y d
dy = d + d =
z d + z d
d
dz z z

(5.1.4)

N, x N, x    
d d
= N, y N, y = [ g g ]
d d
N, z N, z
or
dr = Jd .

2
The Jacobian J introduced above is also used for computing partial derivatives
with respect to the natural coordinates and :
 ()   () () y
  x y   () 
x
x + y x
() = () x () y
= x y ()
x + y y
   ()  (5.1.5)
N, x N, y x
= ()
N, x N, y y

or
() ()
= JT . (5.1.6)
x
It should be noted that only the x and y components are included in the above
relationship for the partial derivatives. This is a consequence of assuming that
the element represents a best t of the warped quadrilateral in the (x, y)-plane.
If the z-coordinate is neglected the relationship between (x, y) and (, )
is isoparametric and the inverse relation can be formed as
     () 
() T T
x
Jx Jx
() = T T () . (5.1.7)
y
Jy Jy

5.1.1 Geometric quantities for a at quadrilateral element.

By dening lij to be length of side edge ij one obtains

lij = x2ji + yji
2 . (5.1.8)

The unit vector sij along side ij and the outward normal vector nij of side ij
can then be dened as

si x xji ni x si y
1
sij = si y = yji and nij = ni y = si x . (5.1.9)
lij
0 0 0 0

3
5.2 The quadrilateral membrane element.
Nygard [00] developed a 4-node membrane element with drilling degrees of
freedom based on the Free Formulation, which is called the FFQ element. The
element has given accurate results for plane membrane problems. Unfortunately,
the element is computationally expensive because the formation of each element
stiness requires the numerical inversion of a 12 12 matrix. The goal of the
present development is to construct a 4-node membrane element with the same
freedom conguration and similar accuracy as the FFQ, but that avoids those
expensive matrix inversion.
Recall that element stiness of the ANDES element is the sum of the
basic and higher order contributions:

1 T
K = Kb + Kh = LCL + BTh CBh dA . (5.2.1)
A A

These matrices are now developed for the membrane component.

5.2.1 Basic stiness.

The basic stiness for the membrane element is developed by lumping the con-
stant stress state over side edges to consistent nodal forces at the neighboring
nodes according to a boundary displacement eld. When the boundary dis-
placement eld is dened so that interelement compatibility is satised, pair-
wise cancelation of nodal forces for a constant stress state is assured, and thus
satisfaction of the Individual Element Test [00]. In turn, satisfaction of the In-
dividual Element Test ensures that the conventional multi-element Patch Test
is passed.
A very successful lumping scheme for membrane stresses was rst in-
troduced by Bergan and Felippa [00] in the paper describing the triangular
membrane FF element with drilling degrees of freedoms. This procedure has
since been used by Nyg ard [00] and Militello [00].
The presentation here rewrites the lumping matrix, in terms of nodal
submatrices. The expressions of the lumping matrix thus becomes valid for
elements of arbitrary number of corner nodes.
It is convenient to order the visible degrees of freedom as translations
along x, y and drilling rotation about z-axis for each node. This gives the
lumping of the constant stress state to nodal forces f as

xx
f = L where = yy , (5.2.2)

xy

4

L1 f1
fx i
L f2
L= 2 and f = where fi = fy . (5.2.3)
L3 f3
i
mz i
L4 f4
By using the Hermitian beam shape function for a side edge one obtains a bound-
ary displacement eld that is compatible between adjacent elements because the
displacements along an edge are only functions of the end nodes freedoms. This
again gives a lumping matrix L where each nodal contribution Lj is only a
function of its adjoining side edges ij and jk:

yki 0 xki
1
Lj = 0 xki yki . (5.2.4)
2 2
6 (yij ykj ) 6 (xij xkj ) 3 (xkj ykj xij yij )
2 2 2

The nodal indices (i, j, k, l) for a four node element undergo cyclic permutations
of (1, 2, 3, 4) in the equation above. Factor represents a scaling of the contri-
butions of the drilling freedom to the normal boundary displacements; see [00]
for details.

5.2.2 Higher order stiness.

To construct Kh a set of higher order degrees of freedoms that vanish for rigid
body and constant strain states is constructed.
Higher order degrees of freedom.
The 12 visible nodal degrees of freedom vx , vy and for each node are ordered
in an element displacement vector v as
vTx = [ vx1 vx2 vx3 vx4 ] ,
vx
v = vy , where vTy = [ vy1 vy2 vy3 vy4 ] , (5.2.5)

T = [ 1 2 3 4 ] .

The correct rank of the element stiness matrix for the quadrilateral membrane
element is 9, coming from 12 degrees of freedom minus 3 rigid body modes. The
basic stiness gives is rank 3 from the 3 constant strain modes. The higher order
stiness must therefore be a rank 6 matrix. This can be conveniently achieved
by introducing 6 higher order intrinsic degrees of freedom, which are collected
dened below.
in a vector v
Experience from the 3-node ANDES membrane element with drilling
degrees of freedom [00] and the 4-node ANDES tetrahedron solid element with

5

rotational degrees of freedom [00], suggests using the hierarchical rotations
as higher order degrees of freedom:

= 0 ,
T0 = [ 0 0 0 0 ] , (5.2.6)

where the subtracted 0 represents the overall or mean rotation of the element,
associated with rigid body and constant strain rotation motions. Furthermore,
splitting the hierarchical rotations into their mean and deviatoric parts as

+  ,
= T = [ ] ,
(5.2.7)

has the advantage of singling out the often troublesome drilling mode, or
torsional mode, where all the drilling node rotations take the same value with
all the other degrees of freedom being zero.
Unfortunately, the hierarchical rotations give only 4 higher order degrees
of freedom and at most a rank 4 update of the stiness matrix. Two more higher
order degrees of freedom must be found. The element has 8 translational degrees
of freedom represented by vx and vy . The 3 rigid body and 3 constant strain
modes can all be described by the translational degrees of freedom. There are
still 2 higher order modes which can be described by the translational degrees
of freedom. These two modes must be recognized so as to associate two higher
order degrees of freedom with them. The amplitudes of the six higher order
degrees of freedom are then represented as

T = [ 1
v 2 3 4 1 2 ] , (5.2.8)

where 1 and 2 are associated with the two higher order translational modes.
Although 7 degrees of freedom appear in this vector, the hierarchical rotation
constraint
4
i = 0 (5.2.9)
i=1

reduces (5.2.8) eectively to six independent components.

6
Higher order rotational degrees of freedom.
0 is evaluated as the rotation of the bi-linear quadrilateral computed at the
element center given by ( = 0, = 0), and is function of the translational
degrees of freedom only:
 
1 v u 1 vx
0 = ( ) = [ N,y N,x ] . (5.2.10)
2 x y 2 vy

By using the partial derivative expressions in equation (5.1.7) one obtains the
expressions for the rigid body and constant strain rotation as

T T 1
N,y = (Jy N, + Jy N, ) = [ x24 x31 x42 x13 ] ,
16|J|
(5.2.11)
T T 1
N,x = (Jx N, + Jx N, ) = [ y24 y31 y42 y13 ] ,
16|J|

where
1
|J| = ((x1 y2 x2 y1 ) + (x2 y3 x3 y2 ) + (x3 y4 x4 y3 ) + (x4 y1 x1 y4 )). (5.2.12)
8
The higher order rotational degrees of freedom h can be expressed in terms of
the visible degrees of freedoms as

h = Hv v, Th = [ 1 2 3 4 ] , (5.2.13)

0 0 0 0 0 0 0 0 3
4 14 14 14
0 0 0 0 0 0 0 0 14 3
14 14
4
Hv =
0 0 0 0 0 0 0 0 14 14 3
4 14

0 0 0 0 0 0 0 0 14 14 14 3
4
x42 x13 x24 x31 y42 y13 y24 y31 1 1 1 1
f f f f f f f f 4 4 4 4
(5.2.14)
and f = 16|J|.

7
Higher order translational degrees of freedom.
The translational degrees of freedom can be split into rigid body and constant
strain displacements and higher order displacements:
 
vx
v = vrc + vh where v= . (5.2.15)
vy

vrc can be written as a linear combination of the rc-modes as

vrc = Ra with R = [ rx ry rz cxx cyy cxy ] , (5.2.16)

where rx , ry and rz are the rigid translations in the x and y directions, and the
rigid rotation about the z axis, respectively. cxx , cxx and cxx are the constant
strain displacement modes. By combining equations (5.2.15) and (5.2.16), and
requiring that the higher order displacement vector be orthogonal to the rc-
modes, that is RT vh = 0, one obtains

RT v = RT Ra + RT vh a = (RT R)1 RT v . (5.2.17)

On the basis of this relation two projector matrices Prc and Ph that project
the displacement vector v on the rc and h subspaces, respectively, can be con-
structed:

vrc = Prc v , Prc = R(RT R)1 RT ,

where (5.2.18)
vh = Ph v , Ph = I R(RT R)1 RT .

The higher order translational modes can now be found either as the
null-space of Prc , or as eigenvectors of Ph with associated eigenvalues equal
to one, that is Ph vh = vh . The latter scheme is the simpler one because the
projector matrices enjoy the property PP = P. Every column in Ph is thus its
own eigenvector with eigenvalue one, and a higher order mode.
To write down these higher order modes in compact form it is convenient
to expresses the vector that goes from the element centroid to the nodes ri with
respect to the local coordinate system (g , g ). These base vectors are the -
and - gradient vectors computed at the element center according to equation
(5.1.4). The nodal vector ri can thus be expressed as
   
xi i
ri = = [ g g ] . (5.2.19)
yi i

8
The nodal coordinates i and i in the (g , g ) coordinate system are then
obtained from
   
i xi 1
= T , T = J1 = [ g g ] . (5.2.20)
i yi

The higher order translational mode is then given by

3 3 a
3 1
4
4 4 4 a
vh = = , a= i i . (5.2.21)
1
1 1 a
4 i=1
2 2 2 a

The higher order translational degrees of freedom are the components along this
higher order mode for the x and y nodal displacements vx and vy respectively.
Expressed in term of the visible degrees of freedom this becomes

t = Htv v ,
v (5.2.22)
   
vx 4 1 2 0 0 0 0 0 0 0 0
= 3 v. (5.2.23)
vy 0 0 0 0 3 4 1 2 0 0 0 0
If one denes the higher order translational degrees of freedom to be the higher
order translational components along the and directions the relationship
becomes
   
v 3 s x 4 s x 1 s x 2 s x 3 s y 4 s y 1 s y 2 s y 0
= v
v 3 s x 4 s x 1 s x 2 s x 3 s y 4 s y 1 s y 2 s y 0
(5.2.24)
where 0 = [ 0 0 0 0 ], and si and si denotes the unit vectors along the -
and -directions, respectively:
   
s x s x
s = , s = . (5.2.25)
s y s y

can then be obtained

The total higher order degrees of freedom vector v
from the visible degrees of freedom v as
 
Hv T = [ 1
v 2 3 4 v v ]
= Hv
v where H= and
Hvt T
v = [ vTx vTy ].T

(5.2.26)

3
4

1
2

Figure 5.1. Nodal strain gages for membrane.

Higher order strains.

The distribution of higher order strains is expressed in terms of natural strain
gage readings. The strain gage locations are placed at the 4 nodes (quadrilateral
corners). Readings along three directions are required. These directions are:
the and axis (quadrilateral medians) and the diagonal passing through the
neighboring nodes. See Figure 4.1.
The nodal natural strain readings are thus dened as

   
1 =  , 2 =  , 3 =  , 4 =  . (5.2.27)

24 13 24 13

The next step is to connect these readings to the higher order degrees of freedom.
This can be done by dening a generic template

,
i = Qi v (5.2.28)

where Qi are 3 7 matrices. These templates are worked out below.

10


3
4

l
l

d |1
d |1

1
2

Figure 5.2. Geometric dimensions for a quadrilateral element.

Higher order bending strain eld.

The main displacement and strain mode that the eld is trying to match is the
pure bending of the element along an arbitrary direction. The bending strain
eld is associated with the higher order degrees of freedom i , v and v . If one
considers pure bending of the element along the direction, it seems intuitive
that the strain should be proportional to the distance d from the -axis. The
strain should also be proportional to the curvature along the -axis. In terms of
the rotational degrees of freedoms this curvature will have the form /l where
l is the element length along the -axis. The strain thus gets coecients of
the form d /l associated with the rotational degrees of freedom. Following a
similar reasoning for the strains the strain distribution factors associated with
the and strains are established to be

d|i d|i
|i = , |i = , (5.2.29)
l l

where

1
d|i = (ri s ) (ri s ) , l = r r , r = (r2 + r3 r1 r4 ) ,
2

1
d|i = (ri s ) (ri s ) , l = r r ,
(r3 + r4 r1 r2 ) .
r =
2
The quantities d|i , d|i , l and l are illustrated in Figure 4.2.

11

4 3

2
1

Figure 5.3. Torsional mode for four node membrane element.

The strain distribution sensed by the diagonal strain gages are similarly
assumed to be proportional to the curvature along the diagonal, and proportional
to the distance from the diagonal as
d24 d13
24 = , 13 = , (5.2.30)
2l24 2l13
where

d24 = (r31 e24 ) (r13 e24 ) , l24 = r24 r24 , r24 = (r2 r4 ) ,

d13 = (r31 e24 ) (r13 e24 ) , l13 = r13 r13 , r13 = (r1 r3 ) .

Torsional strain eld.

The torsional strain eld is associated to the higher order degree of freedom.
As a guide for the construction of this strain eld one can use the torsional
displacement mode illustrated in Figure 4.3. This gure indicates that this dis-
placement mode should not induce shear strains, and that  should be positive
in 1st and 3rd quadrants and negative in 2nd and 4th. Similarly,  should be
positive in 2nd and 4th, and negative in 1st and 3rd quadrants. A simplied
strain distribution function for the strains  and  can thus be Nt = .
With a unit rotation at all the nodes, the maximum displacement in the
direction, u , will be proportional to the length l . Since the strain  is the
gradient of the displacement u in the direction this strain will be proportional
to 1/l . The torsional strain eld is thus assumed to be
l l
 = = t ,  = = t . (5.2.31)
l l

12
Generic nodal strain templates.
With the strain assumptions just described, the nodal strain gage readings can
be written down as

Q1 = (5.2.32)

1 |1 2 |1 3 |1 4 |1 t 1 |1
l 0
1 |1 4 |1 3 |1 2 |1 t 0

1 |1 ,
l
c24 c
5 24 6 24 7 24 8 24 0 2 l24 2 l24
24

Q2 = (5.2.33)

2 |2 1 |2 4 |2 3 |2 t 1 |2
l 0
4 |2 1 |2 2 |2 3 |2 t 0

1 |2 ,
l
c13 c13
8 13 5 13 6 13 7 13 0 2 l13 2 l13

Q3 = (5.2.34)

3 |3 4 |3 1 |3 2 |3 t 1 |3
l 0
1 |3
3 |3 2 |3 1 |3 4 |3 t 0 l

,
c13 c
7 13 8 13 5 13 6 13 0 2 l13 2 l13
13

Q4 = (5.2.35)

4 |4 3 |4 2 |4 1 |4 t 1 |4
l 0
2 |4 3 |4 4 |4 1 |4 t 0

1 |4 ,
l
c13 c
6 13 7 13 8 13 5 13 0 2 l13 2 l13
13

where c13 = sT13 s , c13 = sT13 s , c24 = sT24 s and c24 = sT24 s .
The cartesian strain displacement matrices at the nodes are obtained by
the transformations
2
s x s 2y s x s y
Bh1 = T13 Q1 ,
where T1 13 =
s 2 s 2 s x s y , (5.2.36)
x y
Bh3 = T13 Q3 , 2 2
s24 x s24 y s24 x s24 y

s 2x s 2y s x s y
Bh2 = T24 Q2 ,
where T1
24 = s 2x s 2y s x s y . (5.2.37)
Bh4 = T24 Q4 ,
s13 2x s13 2y s13 x s13 y

13
A higher order strain eld over the element can now be obtained by interpolating
the nodal Cartesian strains by use of the bi-linear shape functions dened in
(5.1.3).

Bh (, ) = (1 )(1 )Bh1 + (1 + )(1 )Bh2

(5.2.38)
+(1 + )(1 + )Bh3 + (1 )(1 + )Bh4 .

Interpolation of these nodal strains does not automatically give a deviatoric

higher order strain eld. Such a condition can be achieved by subtracting the
mean strain values:

h where B
Bd (, ) = Bh (, ) B h = B(, ) dA . (5.2.39)
A

Optimal coecients for the strain computation.

When computing the strain displacement expressions symbolically using Math-
ematica, the contributions of the dierent coecients i and i were evaluated
with respect to certain higher order strain modes. Based on pure bending of
rectangular element shapes the following dependencies between the coecients
were obtained:
2 = 1 , 3 = 2 , 4 = 1 ,
1
6 = 1 1 , 1 = + 1 , (5.2.40)
2
8 = 6 , 5 = 7 = 2 = 0 .
As seen this makes all the coecients a function of 1 . Optimizing 1 with
respect to irregular meshes for the cantilever described in the numerical section
suggests 1 = 0.1 , and the following set of optimal coecients:

1 = 0.1 2 = 0.1 3 = 0.1 4 = 0.1

5 = 0.0 6 = 0.5 7 = 0.0 8 = 0.5 (5.2.41)
1 = 0.6 2 = 0.0

14
 4 3

1 2

Figure 5.4. Spurious membrane mode for the four node

ANDES element.

Stiness computation for the membrane element.

According to the ANDES formulation the element stiness is computed as

1 T T
K = LCL + H Kd H where Kd = BTd CBd dA . (5.2.42)
A A

Numerical experiments, however, indicate that the element performs better when
the element stiness is computed as

1 T T h =
K = LCL + H Kh H where K BTh CBh dA , (5.2.43)
A A

that is when the non-deviatoric higher order strains are used. This is not strictly
justied according to the standard ANDES formulation since the higher order
strains displacement matrix Bh is not energy orthogonal with respect to the
constant strain modes for arbitrary element geometries. However, both of the
above element stiness matrices satisfy the Individual Element Test and thus
also the conventional Patch Test.
Rank of the stiness matrix.
Performing an eigenvalue analysis of the element stiness matrices given in equa-
tions (5.2.42) and (5.2.43) it was found that the element has one spurious zero
energy mode in addition to the correct three rigid body modes. This spurious
mode occurred using a 22 Gauss integration rule. It is expected that this spu-
rious mode would disappear with a 33 integration rule. For a square element

15
as shown in Figure 5.4 this spurious mode is dened by the nodal displacement
pattern

vT = [ vx1 vx2 vx3 vx4 vy1 vy2 vy3 vy4 z1 z2 z3 z4 ]

= [ 1 1 1 1 1 1 1 1 4 4
4 ] . 4
(5.2.44)
Analysis of a mesh of two elements shows that this the pattern (5.2.44) can
not occur in a mesh of more than one element. The spurious mode is then not
practically signicant for the performance of the element.

5.3 The quadrilateral bending element.

The current approach to deriving the quadrilateral plate bending element utilizes
reference lines. Hrenniko [00] rst used this concept for plate modeling where
the goal was to come up with a beam framework useful as a model for bending
of at plates.
Park and Stanley [ 00, 00] used the reference line concept in their devel-
opment of several plate and shell elements based on the ANS formulation. The
reference lines were used to nd beam-like curvatures; these curvatures were
then used to nd the plate curvatures through various Assumed Natural Strain
distributions. These plate and shell elements were of Mindlin-Reissner type, and
the reference lines were treated as Timoshenko beams.
The present element is a Kirchho type plate and the reference lines are
thus treated like Euler-Bernoulli (or Hermitian) beams.

5.3.1 Basic stiness.

The basic stiness for a at quadrilateral bending element has been developed
by extending the triangle element lumping matrices Ll and Lq of Militello to
four node elements. Ll and Lq denotes lumping with respect to a linear and
quadratic variation in the normal side rotation respectively.
By ordering the element degrees of freedom as rotation about x and y
axis and translation in z direction for each node one obtains the lumped forces
from bending as

mxx
f = Ll or f = Lq where = myy , (5.3.1)

mxy

Ll 1 Lq 1
L L
Ll = l 2 , Lq = q 2 (5.3.2)
Ll 3 Lq 3
Ll 4 Lq 4

16
and
f1

mx
f1
f= where f i = my . (5.3.3)
f1

fz
f1
The lumped node forces at a node j given by lumping matrix Lj , receive con-
tributions from the moments from adjoining sides ij and jk. The lumped force
vector at a node j is thus a function of the coordinates of the sides ij and jk
only. With linear interpolation of the normal and tangential rotations along a
side the lumping matrix becomes

0 0 0
1
Ll j = 0 xki yki , (5.3.4)
2
yki 0 xki

where superscript l denotes linear variation of normal rotation. If the normal

rotation is assumed to vary quadratically in accordance to Hermitian interpola-
tion whereas and the tangential rotation still varies linearly the lumping matrix
becomes

cjk sjk + cij sij cjk sjk cij sij (s2jk c2jk ) + (s2ij + s2ij )
Lq = 1 (s2jk xjk + s2ij xij ) 1 (c2jk xjk + c2ij xij )
j 2 2 c2jk yjk c2ij yij
1 2
2 (sjk yjk + s2ij yij ) 1 2
2 (cjk yjk + c2ij yij ) s2jk xjk s2ij xij
(5.3.5)
where superscript q is used to denote quadratic variation of normal rotations.
The nodal indices (i, j, k, l) in the equations above undergo cyclic permutations
of (1, 2, 3, 4) as for the membrane lumping.

5.3.2 Higher order stiness

The higher order stiness is computed as the deviatoric part of an ANS type
element using the Euler-Bernoulli beam as a reference line strain guide.

17
Nodal curvatures of a Euler-Bernoulli beam.
The transverse displacement of a Euler-Bernoulli beam, written as a function of
the nodal displacements and rotations is

w = Nw vbij , (5.3.6)

where
2 ( 2 + )(1 + )2 wi
T 1 l ( 1 + )(1 + )2
ni

N = and vb ij = .
8 2 ( 2 )( 1 + )2
wj
2

l (1 + )( 1 + ) nj
The beam curvatures are

6
2w 1 l (1 + 3)

= = 2 vb ij , (5.3.7)
x2
l 6

l (1 + 3)

The nodal curvatures are then

   
ij|i 1 6 4l 6 2l
= 2 v (5.3.8)
ij|j l 6 2l 6 4l bij

The nodal displacements of a reference-line from node i to j can be expressed

in terms of the visible degrees of freedom at those nodes as

vb ij = Tvij vij (5.3.9)

w

1
i


wi 0 0 0 0 0
xi


0 nij x nij y 0 0
0 y i
n i
=
0 . (5.3.10)

w j
0 0 1 0 0 wj
nij y y
n j 0 0 0 0 nij x

j
yj

The nodal curvatures expressed in terms of the visible dofs at node i and j then
become
   
ij|i 1 6 4l nij x 4l nij y 6 2l nij x 2l nij y
= 2 v.
ij|j l 6 2l nij x 2l nij y 6 4l nij x 4l nij y
(5.3.11)

18


3
4

1
2

Figure 5.5. Nodal curvature gages for bending.

Nodal natural coordinate curvatures for a quadrilateral.

When one collects all the nodal straingages in a vector g, the strain-gage dis-
placement relationship becomes

g = Qv = QF Fv , (5.3.12)

where denotes entry by entry matrix multiplication, and

gT = [ 41|1 12|1 13|1 12|2 23|2 24|2

(5.3.13)
23|3 34|3 13|3 34|4 41|4 24|4 ] ,

6 4 4 0 0 0 0 0 0 6 2 2
6 4 4 6 2 2 0 0 0 0 0 0

6 4 4 0 0 0 6 2 2 0 0 0



6 2 2 6 4 4 0 0 0 0 0 0

0 0 0 6 4 4 6 2 2 0 0 0

0 0 0 6 4 4 0 0 0 6 2 2

QF = ,

0 0 0 6 2 2 6 4 4 0 0 0

0 0 0 0 0 0 6 4 4 6 2 2

6 2 2 0 0 0 6 4 4 0 0 0



0 0 0 0 0 0 6 2 2 6 4 4

6 2 2 0 0 0 0 0 0 6 4 4
0 0 0 6 2 2 0 0 0 6 4 4
(5.3.14)

19

F41 F41 F41 F41
F12 F12 F12 F12

F13 F13 F13 F13
 
1 n12 x n12 y
F12 =
F12 F12 2
l12 l12 l12
F12 F12  
1 n23 x n23 y
F23 F23 F23 F23 F23 = 2
l23 l23 l23
 
F24 F24 F24 F24 F34 =
1 n34 x n34 y
2
l34 l34 l34
F= where  n41 y  . (5.3.15)
F41 =
1 n41 x
F23 F23 F23 F23 
2
l41 l41 l41

1 n13 x n13 y
F34 F34 F34 F34 F13 = 2

l13 l13 l13

F13 F13 F13 F13 1 n24 x n24 y
F24 = 2
l24 l24 l24


F34 F34 F34 F34

F41 F41 F41 F41
F24 F24 F24 F24

Cartesian curvatures for a quadrilateral.

The cartesian curvatures T = [ xx yy xy ] at the nodes can now be ob-
tained as
gC = QC v (5.3.16)

or
B1 T 1 Q1
|1
|2 B T Q
= 2 v = 2 2 v ,

|3
B3 T 3 Q3
|4 B4 T 4 Q4
where
s41 2x s41 2y s41 x s41 y
T 1
1 = s12 2x s12 2y s12 x s12 y ,
s13 2x s13 2y s13 x s13 y
2
s12 x s12 2y s12 x s12 y
T 1
2

= s23 2x s23 2y s23 x s23 y ,
s24 2x s24 2y s24 x s24 y
2
s23 x s23 2y s23 x s23 y
T 1
3

= s34 2x s34 2y s34 x s34 y ,
s13 2x s13 2y s13 x s13 y
2
s34 x s34 2y s34 x s34 y
T 1
4

= s41 2x s41 2y s41 x s41 y .
s24 2x s24 2y s24 x s24 y
The Cartesian curvatures over the element can then be obtained by interpolation
of the nodal values as
= B(, )v , (5.3.17)

20
where
B(, ) = (1 )(1 )B1 + (1 + )(1 )B2
. (5.3.18)
+(1 + )(1 + )B3 + (1 )(1 + )B4

Higher order stiness for the element.

The ANDES higher order stiness is computed as
 
1
Kd = Bd CBd dA where Bd = B
T
B dA . (5.3.19)
A A A

5.3.3 The ANS quadrilateral plate bending element.

Clearly one can form an ANS type element by

K= BT CB dA (5.3.20)
A

i.e. without extracting the mean part of the strain displacement matrix and not
including the basic stiness described in Section 5.3.1.

5.4 The linear non-at quadrilateral shell element.

The objective of this section is to develop a technique that allows the use of
the at quadrilateral membrane and bending element as parts of a non-at shell
element for linear problems. This is obtained by formulating a linear projec-
tor matrix, which for the linear case restores equilibrium at the undeformed
element geometry. This can also be obtained by using the nonlinear projector
with respect to the initial geometry. In fact the linear and nonlinear projec-
tor gives identical results for linear problems. However the linear projector is
recommended for linear nite element codes due to its greater simplicity.
The four node shell element is obtained by assembling the membrane
element and bending element to the appropriate degrees of freedom. This is
sucient as long as the shell element is strictly at since both the membrane
and bending elements are developed as at elements. Unfortunately, four node
shell elements on a real structure quite often end up being warped. To restore
or improve the behavior of the warped element one can use a projection technique
similar to that developed by Rankin and coworkers [ 00, 00].
The element stiness matrix does not have the correct rigid body modes
if the element geometry is warped since the element stiness has been developed
using the projected at positions of the element nodes. This causes two de-
ciencies of the element stiness:
1. The element picks up strains and thus forces from a rigid body displace-
ment vector i.e. f r = Kvr = 0.

21
 2. The element forces are not in self equilibrium and the force vector will
thus pick up energy for a rigid body motion. vTr f = vTr Kv = 0.
These two statements are equivalent for a symmetric element stiness matrix.
If an element stiness has columns that are in self equilibrium the element has
the correct rigid body modes and vice versa.
The foregoing deciencies lead to the investigation of the element inter-
nal energy
1 1
= vT Kv = (vTr + vTd )K(vr + vd )
2 2 (5.4.1)
1 T
= (vd Kvd + vTd Kvr + vTr Kvd + vTr Kvr ).
2
If the element fails the equilibrium and rigid-body conditions;

vTr Kvr = 0 , vTr Kvd = 0 and vTd Kvr = 0 . (5.4.2)

To extract the deformational energy, the total displacements are split into de-
formational and rigid body motions, the latter being spanned by the matrix
R:
v = vd + vr = vd + Ra. (5.4.3)

By requiring that the deformational displacement vector be orthogonal to the

rigid body modes one must have RT vd = 0. On pre-multiplying the equation
above with RT the rigid body amplitudes can be solved for:

RT v = RT Ra a = ( RT R )1 Rv , (5.4.4)

from which the deformational displacement vector can be extracted as

vd = v vr = ( I R ( RT R )1 R ) v = Pd v. (5.4.5)

If R is orthonormal the foregoing expression simplies to

Pd = I RRT . (5.4.6)

Applying this projection to gain invariance of the internal energy with

respect to rigid body motion (v) = (vd ) yields

1 T 1 1
(vd ) = vd Kvd = vT PTd KPd v = vT Kd v . (5.4.7)
2 2 2

22
5.4.1 Linear projector matrix for a general quad.
In order to express the rigid body modes one denes the vector ri from the
element centroid to node i as

xi 1
4
ri = ri r , where ri = yi and r = ri . (5.4.8)
4 i=1
zi

By ordering the element degrees of freedom as

vxi


v v
1
yi
v2 vzi
v= where vi = (5.4.9)

v3

xi




v4
yi

zi

the rigid body modes can be expressed as

1 0 0 0 zi
yi

R1   0 1 0
zi 0 xi

R2 I Spin(ri ) 0 0 1 yi
xi 0
R= , Ri = = .
R3 0 I 0 0 0 1 0 0

R4 0 0 0 0 1 0
0 0 0 0 0 1
(5.4.10)
The projector matrix becomes

Pd = I R(RT R)1 RT , (5.4.11)

where  
4I 0 
4
T
R R= with S = 4I Spin(ri )Spin(ri ) .
0 S
i=1

This simplies the computation of the projector matrix because only the lowest
33 submatrix of RT R is non-diagonal, and (RT R)1 can be eciently formed.

23
5.5 Nonlinear extensions for quadrilateral shell element.
The nonlinear extensions for an element consists of dening a procedure that
aligns the shadow element C0n as close as possible to the deformed element Cn .
This denes the element deformational displacement vector vd .
One also needs to form the rotational gradient of the shadow element
with respect to the visible degrees of freedom of the deformed element, as stated
in equation (0.0.0). In the local coordinate system this relation is

r

r =
vi = G v. (5.5.1)

vi

The local coordinate relationship is sought since this is is needed in forming the
geometric stiness of the element as expressed in equation (0.0.0) .
The rotation of the shadow element is most easily obtained from the
rotation of the shared or common local frame for the C0n and Cn congurations.
This orthogonal element coordinate frame with unit axis vectors e1 , e2 and e3
is rigidly attached to the shadow element C0n , since this element only moves
as a rigid body, and elastically attached to the deformed and elastic element
Cn . This local coordinate system for a quadrilateral element can be dened in
various ways. Most researchers select the element z-axis unit vector as the cross
product of the diagonals vectors d13 and d24

d13 d24
e3 = where Ap = (d13 d24 )T (d13 d24 ) (5.5.2)
Ap

This denes Ap as the area of the element projection on the local x y plane.
The positioning of the x and y axis unit vectors e1 and e2 diers among
researchers. Rankin and Brogan [00] chooses e2 to coincide with the projection
of the side edge 24 on the plane normal to e3 . This eectively lets only one of the
side edges determine the rigid rotation of the element about the local z axis. The
origin of the element coordinate system is chosen to coincide with node 1. When
this procedure is performed for both the C0 and Cn element congurations the
net result is that the shadow element C0n will be positioned relative to Cn so that
nodes 1 coincide and the projections of side edge 24 on the (x, y) plane coincide.
A consequence of this choice is that the element deformational displacement
vector vd , which is the dierence between the coordinate between the Cn and
C0n coordinates, is not invariant with respect to the element node numbering.
Bergan and Nyg ard [00] choose vector e1 and e2 to coincide with the
directions of side edge 12 and 14 for a rectangle that is positioned relative to
the quadrilateral element so that the sum of the angles between the side edges
of the quadrilateral and rectangle is zero. The origin of the coordinate system

24
is chosen at node 1. By applying this to both the C0 and Cn congurations the
shadow element C0n is positioned relative to the deformed element Cn so that
the element centroids coincide and so that the sum of the square of the angles
between the side edges of C0n and Cn is minimized. This represents a least
square t with respect to the side edge angular errors. This procedure gives a
element deformational displacement vector vd which produces an internal force
vector f e = Ke vd that is invariant with respect to the node numbering of the
element, provided that the element stiness matrix Ke satises the correct rigid
body translations.

5.5.1 Aligning side 12 of C0n and Cn .

The element frame is positioned at the element centroid. This change from
Rankins positioning at node 1 has been done in order to satisfy the orthogonality
condition for PT PR = 0 as expressed in equation (0.0.0) . Rankins formulation
did not contain PT so this requirement was ignored.
By expressing the nodal coordinates of the element in the local coordi-
nate system equation (5.5.2) gives

e3x y31 z42 y42 z31
1
3 = e3y =
e
x31 z42 + x 42 z31 , (5.5.3)
Ap
e3z 31 y42 x
x 42 y31

where

Ap = y31 z42 y42 z31 )2 + (
( x31 z42 + x x31 y42 x
42 z31 )2 + ( 42 y31 )2 .
(5.5.4)
These expressions simplify, but the full expressions has to be kept in order to
obtain the correct variation with respect to the nodal coordinates. The x and

y variation can now be obtained from the variation of e3y and e3x respectively

e3y
e3y
e3y
x = (
x
i + yi +
zi ) ,
xi yi zi
(5.5.5)

e3x
e3x e3x

y = ( x
i + yi +
zi ) .
xi yi zi

The variation of
x and y with respect to the in-plane coordinate components
of the nodes x
i and yi is zero since

e3x
e3x
e3y
e3y
= = = =0. (5.5.6)
xi yi xi yi

25
This gives the variation of the in-plane rotations
x and
y as function of the
out of plane displacements only:

e3x x4
lj
x =

zi =
zi ,
zi i=1
Ap
31 y42 x
with Ap = x 42 y31 , (5.5.7)

e3y  ylj
4

y =
zi =
zi ,
zi i=1
A p

where the nodal indices (i, j, k, l) takes cyclic permutations of (1, 2, 3, 4).
The e1 vector is chosen to lie along the projection of side 12 in the x-y
plane. This gives the y axis unit vector as
e3 r12
e2 = , (5.5.8)
l12
where l12 is the projected length of side 12 in the x-y plane. By expressing e2
in the local coordinate system the variation of
z can be obtained as

e2x
e2x
e2x
z = (
x
i + yi +
zi ) . (5.5.9)
xi yi zi
Carrying out the derivations gives

1 4
x
lj z21
z =
( y1 + y2 )
zi , (5.5.10)
l12 i=1
A p

where Ap is dened in equation (5.5.7).

The rotation gradient matrix in equation (5.5.1) can now be expressed
as
Gx

G = [ G1 G2 G3 ] = G
y (5.5.11)

Gz
where
0 0 x42 0 0 0
1 = 1
G 0 0 y42 0 0 0,
Ap A
0 l12p x42 z21 0 0 0

0 0 x13 0 0 0
2 = 1
G 0 0 y13 0 0 0,
Ap Ap
0 x 13 z21 0 0 0
l12
(5.5.12)
0 0 x24 0 0 0
3 = 1
G 0 0 y24 0 0 0,
Ap
0 0 x24 z21 0 0 0

0 0 x31 0 0 0
4 = 1
G 0 0 y31 0 0 0.
Ap
0 0 x42 z31 0 0 0

26
 The rotation gradient matrix as expressed here can be split as G = XA

where X contains all the coordinate dependencies and A is constant matrix.
is a function of more
This is possible since none of the vector components of
thus satises the consistency
than three distinct coordinate expressions. This G
requirement set forth in equation (0.0.0).

5.5.2 Least square t of side edge angular errors.

The orientation of e3 is the same as with Rankins procedure, which gives identi-
cal expressions for the rotations
x and
y . The positioning procedure of Bergan
and Nyg ards [00] positioning procedure can be viewed as a least square t of
the side edge angular errors between the C0n and Cn congurations. This gives
dierent expressions for the variation of the angle z with respect to the visible
degrees of freedom. The nodal submatrices Gi of G can then be dened as

0 0 xlj
1 0 0 ylj 0 , (5.5.13)
Gi =
Ap Ap nij x nki x Ap nij y nki y
4 ( lij lki ) 4 ( lij lki ) (xlj fx + ylj fy )

where
1 z21 x21 z32 x32 z43 x43 z14 x14
fx = ( + + + ),
4 l21 l32 l43 l14
(5.5.14)
1 z21 y21 z32 y32 z43 y43 z14 y14
fy = ( + + + ),
4 l21 l32 l43 l14
ij and lij is the outward normal and length of side edge ij respectively:
and n

1 yji 
nij = xji and lij = x2ij + yij
2 . (5.5.15)
lij
0

The nodal indices (i, j, k, l) undergo cyclic permutations of (1, 2, 3, 4).

The G derived above can not be expressed as G = XA where the 33
matrix X contains all the coordinate dependencies of G,
since the expressions for

z contains more than three distinct coordinate expressions. This will give a loss
in convergence rate if the deformed conguration Cn and shadow conguration
C0n are far apart since the present tangent stiness expressions have omitted
terms containing the unbalanced element forces. See Section 0.0.0.

27