Professional Documents
Culture Documents
10,243-251 (1976)
SUMMARY
The existence of optimal points for calculating accurate stresses within finite element models is discussed. A
method for locating such points is proposed and applied to several popular finite elements.
INTRODUCTION
The existence of unique points within certain finite element models, at which the stresses have
higher accuracy than at any other points, first came to the authors attention when using simple
structural elements in the representation of airframe structures. A similar phenomena was noted
by Zienkiewicz et a1.* in more advanced elements and the wider implications of optimal stress
points have been expounded by Irons3 and St~-ang.~
The least squares nature of the finite element
approximation as used by Moan5 to determine optimal locations and procedures for smoothing
stress distributions by Hintoq6 also using a least squares method, incorporates similar points.
The current note attempts to rationalize the reasons for the presence of optimal stress points and
outlines a process by which the locations of such points may be determined.
POLYNOMIAL DISPLACEMENT FUNCTIONS
Polynomial functions are commonly used to represent the displaced shapes of finite element
models in the direct displacement method. These functions contain the complete polynomial
terms up to a given order n but may involve selected terms of higher order. The popular isoparametric elements of Irons are typical of this type. When such elements are used to represent
any displacement field of orde: n then the representation is exact in all its attributes. However,
when the displacement field to be represented contains general functions of order n+ 1 some
approximation is evident as the required field will generally contain all the polynomial terms
of order n + l , only some of which are present in the element. This approximation is most
apparent in the erroneous stress distributions, calculated from the derivatives of the element
functions, although the coefficients of the functions, i.e. the nodal displacements, may still be
very accurate. As the representation of higher order terms is attempted, using a model with low
order functions, the accuracy of the nodal displacements continues to be far better than that
of the element stress distributions. This type of behaviour is inherent in the displacement method
wherein the nodal displacements are the prime variables used in the energy minimization.
-___
243
244
JOHN BARLOW
The behaviour of some popular elements is now investigated by subjecting each to a complete
polynomial field of one order higher than that which is complete in the shape function polynomial. The objective is to locate unique positions at which the stresses have the same degree
of accuracy as the nodal displacements. The technique is first demonstrated with a simple beam
element, where the phenomenon of optimal stress points was first observed, then extended to
more complicated plane, solid and shell elements.
BEAM ELEMENT
The commonly used displacement functional for a simple beam element (Figure 1) is a complete
cubic polynomial of the form,
wa
<7
5'7
t31{a)
where a are coefficients and 5 is the non-dimensional co-ordinate measured from the centre.
Nodes
IOptimol stress
-1
-1
47
points
43
The suffix a is used to denote the displacement field which can be exactly represented, in all
respects, by the function. Suppose that this element was to be used to represent a quartic displacement field of the form
c3,
[l, 5,
t4i {b)
where b are coefficients and suffix b is used to denote the displacement field which is required
to be represented.
The nodal displacements are the values of w and dwldx at the two ends so that the nodal displacement, 6,, in the element field may be derived in terms of a, by inserting the relevant nodal
co-ordinates in the first equation
wb
5'3
where [A]is the matrix normally used in conjunction with the former equation to derive the
displacement functions in terms of the nodal displacements. Similarly in the required field
{6b)
[B1{b)
If the element is capable of representing the nodal displacements in the field to a good degree
of accuracy then the values of the two nodal displacements may be equated approximately.
[A1{a) == P
I{b)
245
so that
where
The bending strains in the two fields will be equal when the curvatures are equal
d2wa d2wb
dt2
dt2
or when
[O,0,29651{a} = [O,0,2965,
m21{b)
LO,~ 2 ~ 6415{b)
,
LO,0,2,6t, m21{b)
5=*P
i.e. at the two point Gauss quadrature points.
Hence if the element size is chosen such that the actual displacements are effectively quartic
the cubic model will give stresses, at the 2 point Gauss quadrature points, which have accuracy
of the same order as the nodal displacements.
Note that a similar procedure applied to a quadratic displacement (3 node) end load carrying
bar element, when subjected to a cubic displacement field yields the same optimal stress points
as for the beam.
EIGHT NODE ISOPARAMETRIC PLANE ELEMENTS
The same technique is now applied to the eight node isoparametric plane element' (Figure 2).
The displacement function for this element contains polynomial terms up to a complete quadratic (n = 2) with two additional cubic terms thus,
[ u a , 0aI
[1,<, V , t2,
where the a's are again the coefficients and 5, v are the non-dimensional co-ordinates.
Suppose that this element were used to represent a full cubic (n+ 1 = 3) displacement field
of the form
[Ub, ub]
Again assuming that the 16 nodal displacements (the u and u translations at the corners and midsides) of the element are accurate in the representation of the full cubic field, inserting the nodal
co-ordinates and equating the nodal displacements in the two fields, as in the previous example,
246
JOHN BARLOW
-1 -1
43
Nodes
Optirnol stress
mints
(b)
Figure 2. Isoparametric plane quadratic element : (a) x, y space; (b) <,9 space
then
1
1
1
1
1
1
1
1
3,
247
Now the Jacobian of 5 and q with respect to x and y is a function of the element geometry only
and is independent of the strain field. The strains, and hence the stresses, in the two fields are
equal when the derivatives of the displacements, with respect to the non-dimensional coordinates, are equal. These derivatives are
Equating the derivatives and substituting for the as in terms of the bs gives
t,
~3
These functions are complete to n = 2. The polynomial for a complete cubic (n+ 1 = 3) field are
LUb, Ob? wb1
= [1,(, ?,
Following the same procedure as for the plane element a relationship between the as and bs
is obtained by equating the three displacements at the corners and midside nodes. The nine
248
JOHN BARLOW
Nodes
# Optimal stress
points
t'
Figure 3. lsoparametric solid quadratic element: (a) x, y , z space; (b) 15q, ( space
Nodes
)t.
Optimol stress
points
249
(C)
Figure 4. Linear elements: (a) 2 node end-load element; (b) 4 node quadrilateral element; (c) 8 node solid element
t-t3,
1-v3
are present in the b field only, and are zero at the nodes. The optimal stress points are those at
which the strains, due to these unrepresented functions, are zero, i.e. at
1-3c2 = 1-3q2 = 0
This alternative approach is convenient when the unrepresented modes can be readily identified.
THREE NODE FLEXURAL TRIANGLE
The non-conforming cubic triangle of Bazeley et a19 (Figure 5) is notorious for giving violently
oscillating stresses. This element uses all ten cubic polynomial terms but has only nine nodal
displacements. The tenth coefficient is that of the bubble function
w
L , .L , . L ,
The coefficient of this function was eliminated, in terms of the nodal displacements, from the
requirement of constant curvature representation. In consequence the element is not capable of
250
JOHN BARLOW
Nodes
Optimal strain
points
Figure 5. Non-conforming cubic triangle: (a) Nodal displacements; (b) Non-dimensional co-ordinates (L, , L,,L 3 ) ;
(c) Optimal strain locations
representing a general cubic displacement field as the coefficient of the bubble function is only
correct for constant curvature modes.
Following the approach of the previous section, it is evident that the bubble function is that
which has zero values at the nodes and which is not correctly represented by the model. The
optimal stress points are those at which the curvatures of this function are zero. Unfortunately
this condition cannot be satisfied completely at any single point. However examination of the
individual conditions reveals that the curvatures tangential to each edge are zero along each
edge and the twist normal to each edge is zero at the mid-sides. Thus at the mid-side locations
the direct and shear strains tangential to the edges will be accurate (Figure 5(c)) but the strains
normal to the edges will not. This is not surprising considering the non-conformity of normal
slopes in this element. Judicious smoothing of the optimal strains, before calculation of the
stresses may be satisfactory.
The possible occurrence of optimal locations for individual strains, as found in this element,
was predicted by Strang and
25 1
COMMENTS
Finite element programmes can only be used on a large economic scale, if the users are protected
from the apparent idiosyncrasies of the method. Output of violently oscillating stresses in a
smooth displacement field can only confuse the user. The alternatives are to smooth the stresses
or to output only stresses which have some guarantee of accuracy. As smoothing methods using
optimal stresses are appropriate, it is sufficient to supply only those stresses in the primary output from the analysis, for subsequent use in the smoothing programme. A preference for this
type of reliable stress output has been shown in the use of the Rolls-Royce G-SAS finite element
system which features optimal stress locations.
ACKNOWLEDGEMENTS
The author wishes to thank Rolls-Royce (1971) Limited and Inter-G (Software Marketing)
U.K. Limited for permission to publish this work.
REFERENCES
1. J. Barlow, A stiffness matrix for a curved membrane shell, Conf.Recent Advances in Stress Analysis, Roy. Aeron. SOC.
(1968).
2. 0. C. Zienkiewicz, R. L. Taylor and J. M. Too, Reduced integration techniques in general analysis of plates and
shells, Int. J . num. Meth. Engng, 3, 275-290 (1971).
3. B. M. Irons and A. Razzaque, Shape functions for elements other than displacement models, Conf. Variational
M e f h . in Engng, Southampton Univ. (1972).
4. G. Strang and G. Fix, An Analysis of the Finite Element Method, Prentice-Hall, Englewood Cliffs, N.J., 1973,
pp. 168-169.
5 . T. Moan, On the local distribution of errors by finite element approximations, Seminar on Finife Elemenf Analysis,
Tokyo (1973).
6. E. Hinton and J. S. Campbell, Local and global smoothing of discontinuous finite element functions using a least
squares method, Int. J. num. Meth. Engng, 8, 4 6 1 4 8 0 (1974).
7. B. M. Irons, Engineering applications of numerical integration in stiffness methods, AIAA J. 4, 2035-2037 (1966).
8. S. Ahmad, B. M. Irons and 0. C. Zienkiewicz, Analysis of thick and thin shell structures by curved finite elements,
Int. J. num. Meth. Engng, 2,419-451 (1970).
9. G. P. Bazeley, Y. K. Cheung, B. M. Irons and 0. C. Zienkiewicz, Triangular elements in plate bending, conforming
and non-conforming solutions, First Con$ Matrix Meth. Sfrucf. Mech., Wright-Patterson A.F.B., AFFDL-TR66-80 (1965).