You are on page 1of 13

Pergamon

0045-7949(94)30172-X

Computers & S~ucrures Vol. 53. No. 4. pp. 787-799. 1994 Copyright 0 1994 Elwicr Science Ltd Printed in Great Britain. All rights resmwd 0045-7949/w 57.00 + 0.00

ON IMPROVING THE MEMBRANE CAPABILITY OF A FOUR-NODED QUADRILATERAL ELEMENT


M. V. V. Mm-thy Structures Division, National Aeronautical Laboratory, Bangalore, India
(Received 23 July 1993)

reduced integration version of a four-noded, isoparametric, quadrilateral membrane element is re-examined in the light of its poor performance in the cantilever beam test for a nonrectangular mesh. The deficiency is sought to be overcome through a reformulation based on the method of assumed strain distribution and minimization of error in satisfying a so-called edge compatibility requirement in a least squares sense. The new element reduces to the reduced integration element for the special case of a rectangle, but also passes the cantilever beam test for a parallelogram mesh which the reduced integration element does not. The element design procedure ensures passing of the patch test for an arbitrary quadrilateral.

Abstract-The

1. INTRODUCTION

In any structural analysis package, three-noded triangular and four-noded quadrilateral elements (generally named TRIA3 and QUAD4) are the workhorse elements. They are generally used in a complementary fashion; QUAD4 to model large portions of the structure over the interior and TRIA3 to cover the boundary to enable easy modelling of arbitrary geometry, corners, etc. Several formulations of the four-noded quadrilateral element were carried out in the early stages of finite element development but with the introduction of isoparametric elements, QUAD4 development received special attention as a potential work-horse element. When Taig presented the first isoparametric version of QUAD4 [ 11, there was initial excitement but it soon proved to be a disappointment. The element was discovered to suffer from serious drawbacks of locking under a variety of situations. This paper is confined to one such situation which corresponds to in-plane bending. For example, it fails if it is used to model cantilever beam bending under tip loading with a single row of elements through the thickness and with elements of large aspect ratio as in Fig. 1. Here, the element gives (almost) zero deflection at the tip for all the three meshes shown. This phenomenon, known as parasitic shear locking, is widely discussed in the literature. Locking, in this case, is due to the development of parasitic shear, leading to a spurious constraint on certain terms in the in-plane shear. A detailed account of the reasoning and an error model is presented by Prathap [2]. One of the earliest efforts in overcoming parasitic shear locking was due to Doherty et al. [3] who carried out a modification of the isoparametric element through reduced (one point Gaussian) inteCM 53,4-B

gration of the shear term in the strain energy. Reduced integration works well for rectangular elements (Fig. la). The element also passes the patch test for arbitrary quadrilateral but locks for both arbitrary quadrilateral and parallelogram meshes (Figs 1b and lc). This type of locking for nonrectangular shapes is generally known as distortional locking, a terminology which is also used in this paper. The modified isoparametric element due to Doherty et al., in spite of its known drawbacks, continues to be commonly used. For the sake of brevity, we shall herein refer to this element simply as the reduced integration element. The problem of parasitic shear in QUAM4 (used henceforth to designate the membrane part of QUAD4) has also been the topic of many other investigations in the past. These efforts were mostly based on use of one of the following approaches: incompatible bubble modes [4,5], natural or skew coordinates [6,7], hybrid finite elements [S-10], Hu-Washizu principle and stabilization matrix [ 111, (v) energy orthogonal functions [ 121, (vi) drilling degree of freedom [13-151. Wilsons bubble function approach [4] did not lock for distorted configurations, but failed the patch test except in the case of a parallelogram. A partial remedy for this was found by Taylor er al. [S]. This was based on the assumption that elements of the Jacobian are constant over the element and have the same values as at the origin of the isoparametric coordinates. This is not true for an arbitrary quadrilateral and is, therefore, an approximation. The approach based on the use of natural or skew (i) (ii) (iii) (iv)

787

788

M. V. V. Murthy which exists in the current general purpose programs is possible for nonrectangular configurations. The improvement pertains to extending the applicability to a parallelogram shape for the element without locking taking place in the straight beam test [17] as in Fig. 1. In this paper, we present a new formulation which offers advantages over both the commonly used reduced integration and bubble mode methods. The advantage over reduced integration is that the applicability is extended to the parallelogram shape. Unlike the bubble mode approach [4], the present method does not require static condensation. The element passes the patch test for an arbitrary quadrilateral. For the rectangular configuration, the element developed here becomes identical with the reduced integration element [3]. The mathematical formulation is based on a modified form of the method of assumed strain distribution. A novel feature of the method proposed is that passing of patch test is ensured in the element formulation. The patch test due to Irons and Razzaque [ 181 is now widely recognized as an essential feature for acceptance in general purpose programs. Another criterion used in the element formulation is designing against distortional locking. This is of particular significance in view of the fact that many commonly used elements suffer from this problem as a consequence of performing some variational trick (like reduced integration) during the element development. A typical example is the eight-noded brick. The method proposed here is expected to have considerable potential in reexamining the formulations of such elements.
2. ELEMENT DEVELOPMENT

(a)

Rectangular

mesh 0.5

-0.5
(b) Trapezoidal mesh

(c)

Predominantly

parallelogram

mesh

Thickness All

= O-I,

E,

= I.0 equal

x IO7 volume

, tr = O-3

elements unit

have tip

Loading:

load

Fig. 1. MacNeal and Harders cantilever beam test examples. coordinates in defining element shape functions [6,7] led to the elimination of locking for parallelogram shaped elements, but these formulations were based on an assumed stress or matrix force method. Efforts using other approaches as listed in (iii)-(v) also provided a partial remedy for the problem. The problem of an arbitrary quadrilateral, however, remained unsolved. Most general purpose programs use the reduced integration element. A few also use Wilsons bubble function approach. Recent research efforts appear to be centered around the introduction of an additional degree of freedom in the drilling direction [13-151. For purposes of comparison, we restrict ourselves to elements with two degrees of freedom (DOFs) per node. The problem addressed in this paper is to find a four-noded quadrilateral element which

2.1.

Approach and outline of the method

(9 has two DOFs per node, i.e., the in-plane


displacements u and v, (ii) does not lock in the case of an arbitrary quadrilateral under in-plane bending, (iii) passes the patch test, and (iv) does not involve static condensation in its development. In this context, it must be mentioned that MacNeal [16] published a proof that an element satisfying all the above requirements fully does not exist. Studies reported in this paper also confirmed MacNeals findings but revealed that an improvement over that

We start by examining why the reduced integration element [3] locks for nonrectangular shapes in the cantilever beam test. The change in length of a side of the quadrilateral can be obtained either by determining the strain component along the side from the strain field and integrating along the length or directly from the displacements of nodes on that side. If both yield the same value, we shall say that the element satisfies the condition of edge compatibility. If this condition is not satisfied, the element experiences distortional locking. It is easy to see that the original isoparametric element [ 1] without modifications satisfied the condition of edge compatibility, but the situation changed after reduced integration [3]. Reduced integration was equivalent to using a smoothened value of the in-plane shear. The consequence was a change in the effective value of the strain components along the sides of the quadrilateral and hence in a violation of edge compatibility. This does not happen in the particular case of rectangular element, provided the Cartesian coordinate axes in which the shear is represented are oriented in the

Improving the membrane capability of a quadrilateral element direction of the sides. The reason is that in such a case, the shear strain does not contribute to strain along any side. We, therefore, see that distortional locking is a secondary type of locking which follows as a consequence of reduced integration which was necessary to eliminate locking due to parasitic shear. In the method being proposed here, the satisfaction of edge compatibility is built in as part of the element deveiopment. Reduced integration is avoided altogether by using the method of assumed strain distribution and by choosing shear stress in the proper functional form, i.e., as a constant. The element development here is based on the method of assumed strain distribution which is now widely recognized as an accepted procedure for finite element formulationsjl9j. In the form in which the method is commonly used, the strains are matched at certain discrete points between the assumed strain field and the original strain field derived from the displacement field. The method is used here with a modification. The modification introduced is that we simulate what we consider as the essential features of the original strain field rather than the strains at discrete points. The essential features are, in turn, chosen from the desired performance requirements of the modified element. In the present problem, the isoparametric element is taken as the parent element and the essential features considered are: (i) functional forms of the strains in the rectangular version of the parent element. (ii) passing the patch test, and (iii) satisfying the condition of edge compatibility. A further comment on the first of the above essential features is required. In choosing the functional form of the strains from the rectangular version of the parent element, we modify it by choosing the shear as a constant, which is equivalent to reduced integration of the shear term. Here, we are making use of our experience that reduced integration overcomes locking due to parasitic shear. Effectively, we use the same strain distribution form as in the rectangular version of the reduced integration element for our arbitrary quadrilateral element, irrespective of whether it is rectangular or not. The reason is that the rectangular version of the reduced integration element is known to work well and passes the beam test of Fig. 1. When the procedure for the element development to be outlined below is carried out, it will turn out that, for the rectangular case, the element becomes identical with the rectangular version of the reduced integration element. In addition, the element works better than the reduced integration element for the nonrectangular case. The assumed strain distribution is of the following form in Cartesian coordinates (s, y):

789

(1)
Ed= $o) + +(d)

12)

Here, subscript (0) refers to strain components at the origin, which will be referred to as the basic strain components. Subscript (d) refers to differential strain defined as the difference between the strain at the point under consideration and the corresponding strain at the origin. This definition proves convenient from a mathematical point of view, as will be seen later. The quantities 6,(,), &(d),6y(0), ty(d)and ylVcO) are linear expressions in terms of the nodal displacements, u,, z+,i = l-4 with constant coe~cients and U, D are displacements in the .Y and y directions respectively. We further stipulate that L,(,) , -$@), $,.(oj, cp(dj and yxVcO) are linearly independent of each other and can be looked upon as independent unknowns. It is to be noted that there was a restriction on the number of unknowns which could be introduced in the assumed strain distribution, eqns (i)-(3). We have eight DOFs but. effectively, only five are independent on account of the necessary provision for rigid body displacements and rotation. Accordingly, we have five unknowns in L.~~(,), +,), &r(oj,$,,(?n;d) and y._(0). We cannot have anything more than this number of unknowns as it would overdetermine the problem and result in locking. This was also confirmed by numerical trials. Now, let us examine the number of conditions to be satisfied to meet the requirements we have set for ourselves in respect of the element performance. To avoid distortional locking, edge consistency should be satisfied on all the four sides which leads to four conditions. We will later see that the requirement of satisfying patch test leads to three conditions. We, therefore, have a total of seven conditions to be satisfied by only five unknowns. The number of unknowns cannot be increased for the reason already mentioned. This explains why the problem has defied a solution all along. An advantage of the method proposed in this paper is that the five unknowns are sufficient for the special case of a parallelogram. It will be shown that for this case, if the edge consistency is satisfied on any one side of the parallelogram, it is automatically satisfied on the opposite side. The problem, of course. persists for the general quadrilateral. Although we would have achieved our objective only partially, the procedure leads to an element superior to the reduced integration element which locks even for the parallelogram configuration. Further, the method developed has potential for resolving the distortional locking problem in other elements where the count matches in respect of unknowns and the number of conditions to be satisfied. The mandatory constant strain requirement to be satisfied by the element leads to the following condition:

790

M. V. V. Murthy lems to test the FE accuracy proposed by MacNeal and Harder [17] and is applicable, in particular, to plane stress elements. The same test was also used earlier by Robinson and Blackham [20] to test commercial finite elements and is used, in this paper, as a criterion for passing the element developed. The patch test implies that the correct uniform stress should be exactly represented by an arbitrary patch of elements. There are several ways of carrying out the patch test [21], but one simple and effective method is the so-called single element test [21] or the individual element test (the IE-test, for short) [22]. Although the test shown in Fig. 3 will be used for finally checking the element, the IE-test method will be used for ensuring the passing of the patch test at the formulation stage. Bergan and Hanssen [22] showed that the standard patch test will be automatically satisfied if all elements of a patch during a state of constant strain are subjected to the condition that nodal force contributions for adjoining element sides should cancel by pairs. The single element test concept was evolved from this. Thus, passing of the IE-test means that for an arbitrary state of constant strain, the same value should be arrived at for a nodal force component if calculated from either of the following two methods: (i) from the stiffness matrix directly, or (ii) by considering the virtual work done by the traction forces on the element boundary due to a variation in the corresponding degree of freedom, the traction forces being same as the resultants of internal stresses in the constant strain field considered. This condition can be expressed mathematically F& ds as

0 x ond y oxes

1 Centre

of to I-3

gravity bisectors and of angles

: Parallel
between

2-4

Fig. 2. Coordinate system.


6 r(d) =

O, %(d) =o

(4)

if ui, vi, i = l-4 satisfy linear distribution of u and v with respect to x and y. In view of the definitions of the patch test and the edge compatibility condition and the mathematical procedure followed in this paper for element development, no separate effort is required to ensure that eqn (4) is satisfied; it will be satisfied automatically. A formal mathematical proof of this is given in Appendix A. 2.2. Mathematical formulation
2.2.1. System of coordinates. A Cartesian coordinate system is chosen with the x and y axes oriented along the bisectors of the angles between diagonals of the quadrilateral as in Fig. 2, which ensures invariance. The origin is chosen at the area centre of gravity of the element. It will be seen later that this results in considerable simplification in the mathematical handling of the problem.

(5)

2.2.2. Numbering of nodes. The nodes are numbered with increasing order in anticlockwise direction (Fig. 2). 2.2.3. Matrix notation. Square brackets [] denote a matrix. A column matrix is distinguished by braces { }. The size of the matrix is indicated at the bottom of the symbol where required. Superscript T denotes transpose of a matrix. 2.2.4. Determination of unknowns. The unknowns in the assumed strain distribution (l)-(3) are now determined by enforcing the patch test and edge compatibility conditions. The patch test, as it was originally conceived [ 181, requires that if the exterior nodes of a patch of elements (5-8 in Fig. 3) are subjected to displacements following a linear distribution, the calculated displacements of the interior nodes (l-4) should follow the same distribution. In fact, Fig. 3 is an extract from a standard set of prob-

for a constant strain condition, where U is the strain energy of the element, c( represents either u or LT. i is the nodal number, Ci is the element boundary formed by the two sides intersecting at the ith node, F, is the resultant of internal stresses in the direction corresponding to u per unit length along C, and 6a is the variation in tl over C, due to a unit variation in LY,. The left hand side of eqn (5) represents the nodal force corresponding to cc,at the ith node as calculated from the stiffness matrix. At this stage, let us recall that we are trying to construct the assumed strain distribution from the original strain field of the isoparametric element by using the three criteria identified earlier, out of which passing the patch test is one. The concept of the patch test itself is based on a state of constant strain which is possible only in the case of a plate of constant thickness. So, we consider the element to be of a notional, uniform thickness t in enforcing the patch test equation (S), whereas the actual thickness

Improving the membrane capability of a quadrilateral element

791

a=

O-12,

b =

O-24,

t=

O*OOl,

E,=

I-Ox

IO6

ti

= 0.25

Y 0.02 18 0.03 0.08 0.08

I 2 3 4

0.04 0.

0.16 0.08

Boundary

conditions

Displacements 5 to 8 lC3( to

at follow

nodes

IO3

(x+

y/2

1 )

y +

x/2)

Fig. 3. Membrane plate patch test.

on the right hand side also represent products of different strain components with the same subscript (0) or (d) i.e., ~~~~~~~~~ and Q,c.~~~), respectively. The integral sign indicates area integration over the element. Note that it is area integration rather than volume integration because, for purposes of the patch test, we are now considering strain energy due to a notional uniform thickness t. As the patch test relates to a constant strain condition and the differential strain terms have a special property as defined by eqns (4), let us examine u W C(O)$)+ [x9Ylqo,qd, the contribution from such terms to the patch test s 5 equations. The strain energy U appears in the patch test equations (5) only in the form of its derivatives + [x2,v9 Y2k(&d,. (6) with respect to the degrees of freedom C(~. Hence, the s differential strain terms cannot be straightaway set The above form is only symbolic and indicates the to zero in U in satisfying the patch test equations. The three different types of terms occurring in U. Here, cc,,) third term on the right hand side of eqn (6) involving denotes any one basic strain component, 6(d)denotes purely differential strain terms does not contribute to any one differential strain component and the square the patch test equations as it involves the product of brackets indicate that any one quantity inside should two such terms. However, the second term involving he taken. The integrands in the first and third terms coupling between basic and differential strain terms

t could be variable. The magnitude oft is immaterial, because it cancels out on both sides of eqn (5). If the thickness is variable, it would be taken into account in computing the stiffness matrix after the assumed strain field is determined because the thickness comes under the integral sign while calculating the strain energy. The strain energy U is quadratic in strain terms and in view of eqns (l)-(3), can be written in the form

192

M. V. V. Murthy unless the following area integrals

does contribute, vanish:

li -S As we have chosen the area centre of gravity of the element as the origin, eqns (7) are satisfied. Thus, only basic strain terms appear in the patch test equations. The unknowns to be determined in our problem are the basic and differential strain components in the assumed distribution. By an appropriate choice of the origin, we have been able to render the patch test equations independent of the differential strain components. The basic strain components are thus determined exclusively from the patch test equations. The differential strain components are determined later from the edge compatibility condition using the basic part already determined. This leads to considerable simplification in the mathematical handling of the problem. We shall now proceed with the first step, i.e., determining the basic strain part of the assumed distribution. The vector of nodal degrees of freedom {d} is written as {d} =[u, lA? uj zlq 21 Z$ 2) I#. (8)
li

Yi-1 J

I I

Yi +I 1

B(Xi

yi)

Equation (5) can be written for each of the nodal degrees of freedom in the order in which they appear in {d) as (9) where {FsM } is the vector of nodal forces as calculated from the stiffness matrix, i.e., from the left hand side of eqn (5) and {Fv,} is the vector of nodal forces as calculated from the virtual work done by traction forces on the element boundary, i.e., from the right hand side of eqn (5). The stress-strain-displacement relations are defined by

Fig. 4. Virtual work done by traction forces on element boundary.

where A = area of the element = k $ (XJ, - -u,rl), ,-I j=i+l and j=l for i=4, (13) of the element, for i#4

{c}= [El{c), {c}= [Bl{d),

(10)

t= a notional uniform thickness as already defined. From eqns (5) and (12)

where {o} is the stress vector [a, ov r,,lT, (c} is the strain vector [tl t, yl,,lT, [E] is the 3 x 3 matrix of material stiffnesses representing the genera1 case of anisotropy and [B] is a 3 x 8 matrix with basic and differential components defined in eqns (l)-(3),

= AWlT,,blc,,.

(14)

PI = [~I,,,+ Wd,

(11)

and [B],,, has only constant elements. We have already seen that it is necessary to consider only the basic strain components in the patch test equation (5). The basic component of the strain energy U is

Let us now determine the right hand side of eqn (9). Figure 4 shows the two sides intersecting at node i. The line joining nodes i and (i + 1) (BC in the figure) is defined as side i of the element. The traction forces F, and F,, per unit length at a point D on side i are given by F, = t(~,~(,)sin Oi- T,,.(,)cos Oi) cv = t ( - a~v~o~ cos Oi+
T,~~~~, sin

(15) (16)

UC,,, = fAr{d)TIBl:o,[El[Bl(,){d}, (12)

ei).

Improving the membrane capability of a quadrilateral element Variation in a DOF of the node i produces variation in the corresponding displacement only on the element sides (i - 1) and i (AB and BC in Fig. 4). These variations over the two sides should be linear as there are no mid-side nodes. Note that eqn (5) is written for a unit variation in ai. Considering a unit variation in ui as shown in Fig. 4, variation in u at the point D is (li - s)/l,. The virtual work done due to variation in u over the side i is G,i=i(x,-+i) =f(yj-y,) j=i+l
= 1

793 for i <4 for i >4 for i 24 for i =4 for i # 1 for i = 1. (22)

k=i-1 =4

7.y(o)(xi

xtt 1119

(17)

As the patch test condition (9) should be satisfied for an arbitrary strain field, it follows from eqns (14) and (21) that

whereas the virtual work done over the side (i - 1) is 1,- I sW,,[for au, = l] = s0 FX+r
, 1

[Bl,,, = ; PI.

(23)

=ft'[cx(o)(Yi-Yi-I)

7xy(&-

I - 411.

(18)

In eqns ( 17) and ( 18), s is measured counterclockwise along the element boundary from the trailing end of the side considered (see Fig. 4). The total virtual work done by the traction forces on the element boundary due to unit variation in ui, which is the element i of {F,,} for i < 4, is obtained by adding eqns (17) and (18):
Fvw,t)=6Wfor h=

1l=:t[a,~,,(y,+,-y,-,)
-xi+()1, for i

+7,(,)(xj-1

<4.

(19)

The corresponding quantity for unit variation in oi, which is the element (i + 4) of {F,,} for i < 4, is
Fvwc,t4~=SW[for6u,=1]=ft[a,,,,(xi_,-x,t,)

+ r,(,j(yi + I - yi - 1)I, for i G 4.

(20)

With the derivation of eqn (23) we have been successful in determining the basic strain components in terms of nodal displacements. The passing of the patch test is thus ensured. It is to be noted that the BcO,matrix involves only geometric parameters of the element and not the material parameters. It is, in fact, as it should be as &,, only defines the kinematic relationship between strains and displacements. The definition of basic strains through eqn (23) is therefore valid even in case of anisotropy. The differential strain counterpart, namely [B],,, in equation (1 l), is now determined by making use of the edge compatibility requirement. As per this requirement, the change in length of any side of the element calculated from the assumed strain field should be the same as that determined directly from the displacements of nodes on the side considered. As the assumed strain field is linear in x and y, the change in length of a side is the product of its length and the strain at its mid-point. Thus, for the side i of the element, the edge compatibility requirement is written as (see Fig. 4)
Ii+, = [(ui+, cos 8, + vi+, sin 0,)

The subscripts (i + 1) and (i - 1) in eqns (19) and (20) refer to the leading and trailing nodes of the node i where the variation is considered. Hence, (i + 1) for i = 4 and (i - 1) for i = 1 should be interpreted as 1 and 4, respectively. {F,,} can thus be written as {&,J = ~K7Tblc,,, where [G] is a 3 x 8 matrix defined by
I G,,=T(Y,-Y~) for 1 <

- (ui cos Bi+ vi sin f3,)], (24) where the subscript i(m) denotes the mid-point of the side i and Q,,,, is calculated from the assumed strain field. cicrnj has basic and differential components: $10 = Li(m)(o) + %n)(d)~
(25)

(21)

= 0
Gzi=O

for i >4 for i (4 for i >4

Using eqns (l)-(3) and noting that the shear strain Y.~~ has no differential component, one gets
'%(m)(d) = ih + Yi, ,k(d) cos2 &

=+(xk-xi)

f(xi +

xi+

I)6y(d) sin* &.

194

M. V. V. Murthy many independent equations as there are unknowns. We, therefore, take advantage of this situation and handle the problem of overdetermined unknowns in eqn (27) by satisfying all the equations in a least squares sense. Transposing the right hand side of eqn (26) to the left, the equation can be written as e, = 0. Satisfying eqn (27) in a least squares sense is then equivalent to minimizing the sum of the squares of errors e, with respect to the unknowns: F=O I(d) where
e=

c~,,,,)(~) is known, because the basic strain components are already determined in terms of the nodal displacements. Using strain transformation relations, Q,,,)+,)is expressed in terms of the Cartesian components of the basic strain and is transposed to the right hand side in eqn (24) to arrive at an edge compatibility equation for side i in the following form:

fh + h+ ,) cos* ~,~.v,d~ + f(x,+ 4, ,) sin2 4~~)


=

Ku, +1 cos Bi+ u, + , sin 0,)


- (u, cos Bi+ v, sin ei)]/ri cos* 8, + tvcO, sin* 8, - (%O) + Y.~~(~)

and

&=O, ,(dt

(28)

C ef.
i=l

sin 6 ~0s 0.

(26)
The minimum possible value of e is zero because e can only be positive in view of eqn (29). The error minimization process will capture whatever minimum value of e is possible for the given configuration of the quadrilateral. For the particular case of a parallelogram, the minimum possible value of e is zero because eqn (27) can be exactly satisfied. The error minimization equations (28) will, therefore, lead to this value. The condition e = 0 can be satisfied only if e, = 0 for all i, that is when the edge compatibility equations for all the sides are satisfied. Thus, for a parallelogram, a completely satisfactory solution is automatically achieved. In the case of an arbitrary quadrilateral for which all the edge compatibility equations represented by the matrix equation (27) cannot be exactly satisfied, the error in doing so is minimized through eqns (28). Equations (28) are equivalent to pre-multiplying both sides of the matrix equation (27) by [PI, i.e., [PIT [P] c<(d) 2 X4 4 X2 [ t?(d)

Equation (26) can be written for all the four sides in the matrix form (27) where P,, =

f(yi + y,) COS* ei,

Pi2 = f(xi + x,) sin e,,

the elements of [R] are zero except those covered by the following for i = l-4:
R,, =

--..-2

cos e

Rv=---l

cog e

4 Ri,i+,l) = -sin ei 1, and

4
Ric,+4, = 2

sin e
I,

1
ix;

(30)

which yields two equations for the two unknowns. Let %(d) and $v(d) thus determined be written as %(d)

Hi, = CO$ ei,

Hi2 = sin e,,

Hi3 = sin 8, cos ei

[ %(d)

= (4. 1

(31)

and j is the serial number of the node ahead of node i in the anticlockwise direction as in eqn (13). The matrix equation (27) represents four equations for the two unknowns +) and tYcd,,i.e., there are more equations to be satisfied than the number of unknowns involved. However, as shown in Appendix B, an encouraging aspect of the situation is that for the special case of a parallelogram (not necessarily a rectangle), eqn (26) becomes identical for any pair of opposite sides. In other words, if the equation is satisfied for any one side, it is automatically satisfied for the opposite side also and there would be only as

Using eqns (l)-(3), (lo), (11) and (31), the [B] matrix defining the strain-nodal displacements relationship is obtained as B,j = &,,I, + Yf,,
B2, = 42, + xr2,

(32) (33) (34)

4,

Bw3,.

The stiffness matrix for the element with reference to the nodal displacements vector {d}, as defined by eqn (8), is then given by

Improving the membrane capability of a quadrilateral element

195

[KB = x

vc 8x3

PIT 3x33~8 [El PI dvv

where V, denotes the volume of the element.


3. RESULTS AND DISCUSSION

The assumed strain field was constructed in two parts, namely, basic and differential. The basic strain field was constructed exclusively from the patch test consideration, whereas the differential strain field was determined from the condition of edge compatibility. For the arbitrary quadrilateral, the edge compatibility condition was satisfied in a least squares sense. The condition could be satisfied exactly for the special case of a parallelogram, for which the edge compatibility equations for each pair of oppposite sides become identical. It must be noted in this context that in the reduced integration element, the edge compatibility condition is satisfied only for the more specific case of a rectangle. It is shown later in Sec. 3.1 that the element developed in this paper is identical to the reduced integration element for the particular case of a rectangle. The basic and differential strains are only terminologies introduced here, but they should exist in any finite element formulation which satisfies the patch test criterion, including conventional, direct formulations based on an assumed displacement field. Here, the basic strain field is to be understood as the constant part of the strain field satisfying the patch test criterion. It will also be the strain at some point in the element, which happens to be the area centre of gravity in this particular formulation. The differential part is to be understood as the difference between the total strain and the basic strain, which also satisfies the condition represented by eqn (4). From the way in which the basic strain field was determined, it is clear that it is independent of the functional form of the differential strain field (which is linear in x and y in the present case). The basic strain field should therefore be the same for any element which passes the patch test, irrespective of the way the element was formulated. For example, it can be easily verified that the basic strain field obtained here and defined by eqns (22) and (23) is identical with the strain field at the origin of the isoparametric coordinate system in the original isoparametric element [I] or its reduced integration version [3]. This is true for an arbitrary quadrilateral. However, the point at which the basic strains occur is different in the two elements because the functional forms of the differential strains are different. 3.1. Particular case of a rectangle The reduction of the strain field constructed to the special case of a rectangle is examined in Appendix C. The basic components of the direct strains are given by eqns (C4) and (CS), whereas the corresponding differential components are given by eqns (C6) and

(C7). Substituting from these equations into eqns (1) and (2), it is seen that the direct strains are the same as those in the rectangular version of the reduced integration element [3]. The shear strain in the assumed field has only a basic component as in eqn (3). We have already seen that all the basic strain components are the same as in the reduced integration element even for a general quadrilateral. Thus, the complete strain field coincides with that of the reduced integration element for the rectangular case. The present element is therefore identical with the reduced integration element for the special case of a rectangle.

3.2. Numerical results The element was first subjected to the patch test shown in Fig. 3. Values of the displacements at the exterior nodes 5-8 following the linear law u = 10e3(x + fv) and v = 10e3(y + fx)

were prescribed as the boundary conditions. Finite element analysis of the patch yielded values of displacements at the interior nodes l-4 which also followed the same linear law exactly (correct to eight significant digits). The element thus passes the patch test. This was only to be expected as the formulation procedure ensured the passing of patch test. The element was shown to be mathematically identical with the reduced integration element for the rectangular configuration. The performance of the reduced integration element is known to be good for the rectangular shape. We therefore restrict the numerical study here to the cantilever beam test problem in which the reduced integration element fares badly in respect of nonrectangular shapes. For this purpose, we choose the test examples proposed by MacNeal and Harder as shown in Fig. 1. These examples provide quite a severe test because of the large aspect ratio of the elements used. Also, in the case of nonrectangular shapes, the departure from a rectangle is significant. The results of the cantilever beam test are shown in Table 1 in terms of a normalized tip deflection 5, where 5 = 6 /S lheoret,ca,, 6 is the actual tip deflection and 6theoretical is the theoretical value of the tip deflection as calculated from the beam formula. It is seen that for the case of a rectangular mesh, the present element yields exactly the same result as the reduced integration element as expected. For a trapezoidal mesh, the present element gives slightly better results but the predicted tip deflections in both cases are very small compared to the theoretical value. A substantial difference can be seen in the parallelogram mesh results. For the mesh in Fig. l(c), the reduced integration element locks, giving only 8% of the correct deflection whereas the present element gives 55.9% of the correct deflection. However, it must be. noted that the mesh is not a pure parallelogram mesh although it is captioned parallelogram shape elements in [17].

796 Table 1.Comparison of performance in cantilever with reduced integration element

M. V. V Murthy
beam test

Nondimensional tip deflection. 8 = 6 l%wcucal Reduced integration element 0.904 0.071 Present element 0.904 0.081

that for a rectangular mesh. This is as expected, because the edge compatibility is satisfied in all the elements and over all the sides as in a rectangular mesh.
4. CONCLUDING REMARKS

Type of mesh I. Rectangular mesh Fig. I(a) 2. Trapezoidal mesh Fig. l(b)

3. Predominantly parallelogram mesh I


Fig. I(c) 4. Predominantly parallelogram mesh 2 Fig. 5(a) 5. Pure parallelogram mesh Fig. 5(b)

0.080
0.068

0.559
0.752

0.034

0.900

There are two trapezoidal elements at each end with four parallelogram elements in between. The mesh can only be considered to be a predominantly parallelogram mesh as labelled in Fig. 1. Considering the relative number of parallelogram and trapezoidal elements, it is still quite far from being a pure parallelogram mesh. Let us therefore consider another mesh with a larger proportion of parallelogram elements as in Fig. S(a) with six parallelogram elements in between the two trapezoidal end elements. It is seen that for this mesh, the accuracy of prediction of the tip deflection increases considerably, the predicted value being 75.2% of the theoretical value. On the other hand, prediction from the reduced integration element worsens. Here, the relative proportion of parallelogram and trapezoidal elements is immaterial; the element performance is equally bad in respect of both. For a pure parallelogram mesh as shown in Fig. S(b), the result obtained from the present element is as accurate as

(a)

Predominantly

parallelogram

mesh

A problem typically encountered in lower order finite element design is to ensure that the element performs equally well for regular as well as irregular (like nonrectangular) meshes and also passes the patch test. The four-noded membrane element is only an example. A systematic procedure to achieve this objective is presented here. This is made possible by the introduction of the concept of basic and differential strains with certain special properties which considerably simplifies the mathematical handling of the problem. Although the method is applied specifically to the four noded quadrilateral problem, it can, in principle, be generalized to other lower order finite element formulations. An attractive feature of the element developed is that for the rectangular configuration, the element reduces identically to the commonly used reduced integration element which is known to be satisfactory for that particular configuration. In addition, the locking is completely eliminated under inplane bending for a parallelogram mesh. The problem, however, still persists for an arbitrary quadrilateral, but the error minimization done in respect of the edge compatibility requirement ensures that whatever best can be achieved within the limitations of the number of arbitrary constants available is actually achieved. The problem with the four noded membrane element is that as shown, we have two constants too few. This is the reason why the problem has defied a solution all along. The method proposed here has potential in cases where the count matches in respect of the number of DOFs and the number of conditions to be satisfied. In fact, there need not be an exact match. One could also have more DOFs than the minimum required to prevent distortional locking. The extra unknowns could then be used to collocate strains at certain desired points over the element domain or satisfy such other criteria as may be prescribed.

0.5 REFERENCES 33 I. C. Taig, Structural analysis by the matrix displacement method. English Electric Aviation report No. 5017 (1961). G. Prathap, The poor bending response of the fournode plane stress quadrilateral. In!. J. Numer. Meth. Engng 21, 825-835 (1985). W. P. Doherty, E. L. Wilson and R. L. Taylor, Stress analysis of axisymmetric solids utilizing higher order quadrilateral finite elements. SESM Report No. 69-3, Structural Engineering Laboratory, University of California, Berkeley, California (1969).

(b) Thickness All

Pure = 0.1,

parollelogrom

mesh lJ= 0.3

E, load

I.0

x IO, volume

elements

have equal

Loading

unit tip

Fig. 5. Further examples on cantilever beam test.

Improving the membrane capability of a quadrilateral element 4. E. L. Wilson, R. L. Taylor, W. P. Doherty and J. Ghaboussi, Incompatible displacement models. In
Numerical and Computer Methods in Structural Mechanics (Edited by S. J. Fenves et al.), p. 43. Academic
APPENDIX A: A SPECIAL PROPERTY OF DIFFERENTIAL STRAIN COMPONENTS

797

Using eqns (10) and (I I), eqn (26) can be rewritten as


$(Y, + Y,, ,) cos*&ctid,+ fh, + x,+ ,) sin2Q,w = e, = p, - 4,. (AlI

Press, New York (1973). 5. R. L. Taylor, P. J. Beresford and E. L. Wilson, A nonconforming element for stress analysis. Int. J. Numer.
Merh. Engng 10, 1211-1219 (1976).

6. J. H. H. Pian and K. Sumihara, Rational approach for assumed stress finite elements. Int. J. Numer. Meth.
Engng 20, 168551695 (1984).

where
p, = [(u;+ , cos0,+ ui+ , sin ei)

7. J. Robinson and G. W. Haggenmacher, Some new developments in matrix force analysis. In Recent
Advances in Matrix Method of Structural Analysis and Design, pp. 183-228. University of Alabama Press,

- (u, cos 0, + v, sin 0,)1/l, (A2) qi = [cos* 0, sin* Bi sin ei cos e,] [B],,, {d} (A3)

University of Alabama (197 I). 8. R. D. Cook and Jaafar K. Al-Abdulla, Some plane quadrilateral hybrid finite elements. AIAA Jnl 7(1 I), 2184-2185 (1969). 9. R. D. Cook, Improved two-dimensional finite element. J. struct. Div. ASCE 100 (ST9), Proc. Paper 10808, 1851-1863 (September 1974). 10. R. D. Cook, Avoidance of parasitic shear in plane element. J. Struct. Div., AXE lOl(ST6), 1239-1253 (June 1975). 11. T. Belytschko and W. E. Bachrach, Simple quadrilaterals with high coarse-mesh accuracy, Winter Annual Meeting of the American Society of Mechanical Engineers, Miami Beach Florida, November 17-22, 39-56. AMD-Vol. 73, (1985). 12. P. G. Bergan, Finite elements based on energy orthogonal functions. Inr. J. Numer. Meth. Engng 15,
1541-1555 (1980).

and [B],,, is defined by eqns (22) and (23). In order to prove eqn (4), it is to be shown that e, vanishes if the nodal displacements follow a linear distribution with respect to x and y because in that case, Q,,) and c,(d)have to satisfy a set of homogeneous equations. Let u = a + bx + cy,
v = d +fx + gy. (A4)

Applying eqn (A4) to each of the nodes 1, 2 and 3 (Fig. 2) and solving for a, b and c, we get
b

=U,(Y,-Y,)+u,(Y,-YY,)+u,(Y,-Y,)

2423
where Au3 = area of triangle l-2-3 and

(A5)

13. R. D. Cook, On the Allman triangle and related quadrilateral element. Comput. Struct. i2(6), 1065-67 (1986). 14. A. G. Razaauur and M. Nofal. Performance of a new quadrilateral finite element with rotational degrees of freedom. Quality Assurance in FEM Technology,
Proc. 5th World Congress on Finite Element Methoak,

Salzburg, Austria, pp. 121-130 (5-9 October 1987). 15. R. H. MacNeal and R. L. Harder, A refined four-noded membrane element with rotational degrees of freedom.
Quality Assurance in FEM Technology, Proc. 5th World Congress on Finite Element Methorls, Salzburg, Austria,

Repeating the process for the triangle l-34 b=u(Y3-Y,)+u3(Y,-Y,)+u,(Y,-Y3) 2A,,

yields

pp. 191-201 (5-9 October 1987). 16. R. H. MacNeal, A theorem regarding the locking of tapered four-noded membrane elements. Int. J. Numer. Meth. Engng 24, 1793-1799 (1987). 17. R. H. MacNeal and R. L. Harder, A proposed standard set of problems to test finite element accuracy. Finite Elements in Analysis and Design 1, pp. 3-20. NorthHolland, Amsterdam (1985). 18 B. M. Irons and A. Razzaque, Experience with the patch test for convergence of finite elements. In
Mathematical Foundations of the Finite Element Method

. (46)

From the property of equal ratios, if N,/D, = N2/D2, then each ratio is equal to (N, + N2)/(D, + D2). Applying this to eqns (A5) and (A6) leads to ~=!]u,(Y,-Y.,)+u,(Y,-Y,) +
u,(Y, -Y*) + Q(Y, -Y,W

(A7)

19. 20.

21.

22.

(Edited by A. K. Aziz), pp. 557-587. Academic Press, New York (1972). R. H. MacNeal, Derivation of element stiffness matrices by assumed strain distributions. Nucl. Engng Design 70, 3-12 (1982). J. Robinson and S. Blackham, An evaluation of lower order membranes as contained in the MSC/ NASTRAN, ASAS, and PAFEC FEM systems. Robinson and Associates, Dorset, U.K. (1979). R. L. Taylor, J. C. Simo, 0. C. Zienkiewicz and A. C. H. Ghan, The patch test-a condition for assessing FEM convergence. Int. J. Num. Meth. Enpnp II 22.39-62 _ (1986). P. C. Bergan and L. Hanssen, A new approach for deriving element stiffness matrices. Mathematics - good of Finite Elements and Applications, (Edited by J. R. Whiteman), pp. 483-497. Academic Press, New York (1976).

since A = area of the element = A,,, + A,34. Similarly, one gets c=;[u,(xq-x*)+u2(x,-x3) +
f=f[u,(Y2-Y4)+U*(Y3-Y,) + v,(Y, -YZ) + o,(Y, -~3)llA
g=f[u,(x,-x,)+o,(x,-x3) u3(x2 xd + u4@3 XI

)1/A

648)

(A9)

v3@2 - 4

+ u4&3 -xl

WA.

(AI01

Using eqns (A4) and the relations xi+1 -x,=l,cos0, and yi+,-y,=l,sin0,,

798 eqn (A2) becomes p, = h co? 8, + g sir? 8, + (c +flsin 8, cos 8,.

M. V. V. Murthy
Yalo,=(-X*~,+XIUz+X2U3--XIU4

(Al 1) where

+Y,v,-Y,v,-Y,v,+Y,~,)/A,

VW

With the help of eqns (A7))(AIO), (22), (23) and (8), eqn (All) can be written as p, = [co? 0, sin2 6, sin 8, cos %,][B],,,{d), (A12)

A = area of the parallelogram = 2(x, y2 - .x2 y, ). Equation (26) can be written as which means. from eqns (A3) and (A12), that e, vanishes identically if u,. t, follow a linear distribution in x and y. It is seen from the matrix equation (27) that er(,,)and e,(d) are overdetermined. i.e.. we have four equations for the two unknowns. The situation was overcome by the least squares satisfaction of eqn (27) through eqn (30). However, the right hand sides of the two equations represented by the matrix equation (30) also vanish if ui and a, follow a linear variation in x and y. because they are only a linear combination of e,. e2. e3 and e4. Hence, eqn (4) is proved. f(r, + Y,, , )cos2 Qddl + f@,+ x,+ , ) sin2kid,

(B7)

= e,= P,- (I,,


where ~,=Ku,+,cos%,+v,+,sin%,)

- (u, cos 8, + v, sin %,)I/[, (B9) sit? 8, + yl,+,,sin 8, cos %,). (BlO) 8, + LvC0, 9, = ($0, COG

APPENDIX B: REPEATED EQUATIONS IN CASE OF A PARALLELOGRAM

We now apply eqn (B8) to side 1. Equation (B9) becomes p, = 7 (u* - u,) + y (v* L, ). @I 1)

In the special case of a parallelogram (Fig. 6), the centre of gravity, which is the origin for our cartesian coordinate system, is at the point of intersection of the diagonals. Hence,
X) = -

x,,

y,=

-y,,

x4=

-x*.

y,= -y,.

We put i = 1 in eqn (BlO) and carry out a few algebraic manipulations. The right hand.side is first multiplied and divided by I and is further simplified using the relations lcos% =x2-x,, Isin% =y,-y,

@I)
and eqns (B4-(B7). cos e
q, = 21 c-u, + 4 + LQ -u,)

As per our convention, side i is the line joining nodes i and (i + I) (Fig. 4). Let us consider two opposite sides of the parallelogram, say sides 1 and 3 (Fig. 6). As the opposite sides should be equal in length,
I, = I, = say. 1.

We get

(B2)
+~(-o,+L~~+L~,-v,).

Also, noting that side i is reckoned to be pointing from node i towards node (i + 1) and 8, appearing in the edge compatibility eqn (26) is measured from the x-axis to the side i in the anticlockwise direction. 8, =% and %,=% +2n, (B3)

(B12)

The edge compatibility equation for side 1 can be written from eqns (B3), (B8), (Bll) and (B12) as f(y, + y2) co? %tYtd, + f(x, + x,) sin %e,,,, =c~(-U,+U:-ul+u~) sin % + 21 ( - u, +

where %is the angle of inclination to the x-axis as shown in Fig. (22), (23), (13) and (Bl), the can be expressed in terms of the 1 and 2:
~,f,) =

of l-2 or 4-3 with respect 6. Using eqns (IO), (1 I), basic! strain components coordmates of only nodes

L2 -

v3 +

L4 ).

(813)

012UI -

y1u2-~2u1+~,~d/~

(B4)
-

t,,,, = (-x2v,

x,v2

+ x2vl

x,v,)/A

(BS)

Equations (B9) and (BlO) are now applied to side 3. Making use of eqns (B2) and (B3) and carrying out similar algebraic manipulations as for side 1, there follows:
P3=~(u3-4)+~(u-~4)

cos e

sin % (814)

and q3=41. (BlS)

Substituting from eqns (Bl4), (BlS), (Bl2) and (B3) into eqn (B8), the edge compatibility equation for side 3 is obtained as i(y, + y4) cos2 %Q, + f& +xJ =c$(u,-U*+UI-Un) sin %c,~~~

Fig. 6. Special case of a parallelogram.

+ 31

sin e

(0, -

02 +

03 -

%I.

(B16)

Improving the membrane capability of a quadrilateral element From eqns (BI), it is Seen that eqns (B13) and (Bl6) are identical. Hence, the edge compatibility equations of the opposite sides of a parallelogram are identical.
APPENDIX c: SPECIAL CASE OF A RECTANGLE Let us consider a rectangle of side lengths I and b (Fig. 6). We recall that the origin is at the centre of the rectangle and the x and y axes are oriented along the bisectors of the angles between the diagonals. With reference to Fig. 6, there follows

799

Substituting from eqns (C2) into eqns (B4) and (B5) yields the basic componenents of direct strains as

Using eqns (Cl)-(C3), the edge compatibility condition for sides I and 3 defined by equation (Bl3) or (B16) becomes
I

e=o,

/I =x/2 -b/2),

(Cl)

(x,..Y,)=(-l/2,
hyd =

%d)=

(I(, -

112 + U, -

Uq).

(112, -b/2)

(x,. ~3)= UP.9 b/2)> (xd.yd = (--l/L b/2)


A = lb

The edge compatibility condition for side 2, which is the same as that for side 4, can be derived as in Appendix B from eqn (26). Using eqns (Cl)-(C3), this condition becomes

w
L>(d)

(C3)

(0, -

02 + 0) -

0,).

You might also like