Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
&Structures, Press1974.Printed
Vol.4, PP.1207-1222.Pergamon in Great Britain
Abstract-A method for determining realistic error estimates for conforming finite element solutions is
presented. The method requires solution of the problem by at least two, and preferably three mesh schemes
that yield monotonic solution covergence. This in turn will automatically yield one solution bound, upper
or lower. The paper describes a simple and practical scheme for obtaining the other bound by utilizing
the solutions from the multiple mesh schemes. These bounds bracket the exact solution within relatively
narrow limits and provide the basis of the error estimate. The solution quantities considered are the system
energy quantities; and for eigenvalue problems these correspond to the eigenvalues themselves. As in
convergence proofs, it is expected that the displacement and stress quantities will follow the behavior of the
energy quantities. The proposed bounding method is applicable to eigenvalue and static problems devoid
of stress singularities, and considers only the discretization error of conforming finite element models. The
validity of the proposed bounding method has not been proved mathematically; however, extensive mrmer-
ical applications of the method indicate its workability in every case tested. Results of some applications
are included in this article.
NOMENCLATURE
-4 B half width and half depth of rectangular bar
b beam width
D flexural rigidity of plate, Eh3/12( 1 - v2)
E modulus of elasticity
EO exact solution
F.E. finite element
G shearing rigidity
12 plate thickness; mesh size parameter
6 j integer subscripts
I strain energy
k buckling coefficient
k, torsional rigidity parameter
K constant
L length of plate
L.B., U.B. lower bound, upper bound
Mt twisting moment
integer number denoting mesh subdivision
; plate in-plane force per unit length
NO maximum (central) value of parabolic load
Y number which designates the rate of convergence
1. INTRODUCTION
UTILIZATION of the finite element method began shortly after the advent of the digital
computer, and today it is probably the most effective and widely used approximate solution
method. If one reviews the major requirements of an approximate solution method he finds
them to be as follows [I] :
(1) Provide convergence to the exact solution (preferably at a rapid rate),
(2) Provide a means of improving the approximate solution, and
(3) Provide an error estimate of the approximate solution.
Requirements (1) and (2) above can be accomplished in the finite element method by
the use of conforming displacement elements. For these elements, mathematical proofs
(based on variational principles) of convergence, rates of convergence, and conditions for
monotonic convergence may be found in recent literature. It is requirement (3) which
proves most difficult, yet is probably the most important. That is, since we can only work
with some finite mesh size, we are more concerned about the resulting solution error than
we are with eventual convergence as the element size goes to zero.
One method of estimating the error is to bracket the exact solution between two
approximate solutions, i.e. determine upper and lower bound solutions, and this is the
approach used in this paper. Although displacement and equilibrium models can bound the
exact solution from opposite sides, it is highly desirable from practical considerations that
both solution bounds be obtained through the use of one finite element model.
A simple and practical method of determining both solution bounds, by utilizing con-
forming elements, is presented in this paper. By using conforming elements, one solution
bound is obtained immediately from the finite element numerical solution. The other bound
is generated by using solution results from two mesh sizes. The validity of this other bound,
which may be referred to as the “generated bound”, cannot be proved mathematically and
hence the method should be viewed as a practical rather than theoretical means of deter-
mining a realistic error estimate. The method is applicable to eigenvalue and static problems
Error Estimates for Conforming Finite Element Solutions 1209
where stress singularities are not present and considers only the discretization error of the
finite element approximations. The solution quantities bounded are the system energy
quantities which correspond to the eigenvalues for eigenvalue problems.
The proposed method of determining the generated bound will be described in detail in
the following section and this will be followed by illustrative numerical applications.
r 1
Exoct=X,
I I I I
I/2 l/4 I/a I /I6 I/O 3
n-1 scale
approximate curve from above. The generated bound will be discussed with particular
reference to eigenvalue solutions 1, although they would be equally applicable to the
potential energy IT,and, in a modified form, to the strain energy I. For practical purposes,
it may be assumed that procedures for bounding the energy quantities would be applicable
to bounding displacements and stresses also.
Exact = E,,
I I
1234567
I I I I I I
ss9
I I
IO
F-J
/m
n scale
Fro. 2. Error bounding curves, Kn -P, plotted on normal n scale.
n’ scale
FIG. 3. Error bounding curves, Kn -p, plotted on n -1 scale.
In Fig. 2 the scale is tbe normal arithmetic n scale which increases from left to right; whereas
in Fig. 3 we use an n-l scale which increases from right to left. These are the two scales
commonly used in presenting results from finite element solutions in graphical form.
Note the difference in appearance of the convergence curves in the two figures. For the
arithmetic n scale all the curves are of one concavity; however, for the n-l scale the con-
cavity changes depending on the value ofp. For example, for a p of 1 the curve reduces to a
straight line which converges to the exact value at an n of co, or n-l of 0; for p less than
unity the bounding curve is concave downward; for p greater than unity it is concave
upward.
If we use an n sequence such that ni+ 1 equals 2n,, it can be shown mathematically for
p greater than unity (independent of the K value) that when the difference between two
successive points on the error bounding curve, i.e. (Kn, -P- Knlypl), is subtracted from
Kni+‘l the results will be an error term of sign opposite from that of Kn;;i. Mathematically
we have
(1)
Error Estimates for Conforming Finite Element Solutions 1211
which reduces to
A+&+ K
np c 2p
cz -1
1
, (2)
where 2: denotes a lower bound to 5. Hence, for p greater than unity, equation (2) gives
a lower bound solution for any value of K or n. (Note that when p is unity, the lower bound
is the exact solution.) It must be emphasized that the lower bound generated above is the
lower bound for the upper bound curve which we know in a qualitative sense only and hence
is of no value in arriving at a numerical bound. However, we can determine values of
approximating finite element solutions in a quantitative sense; and we now look to these in
order to generate the other numerical solution bound.
where the inequalities result from equation (4). To show that the bounds (AF)mjare indeed
lower bounds, we form the difference from equation (5), i.e.
or
<~f->,j-<~~),j=C(;I+*>j-(~),*)4jl-33(~~*)Zj-(Izr*)4jl. (7)
The right hand expression in equation (7) can be shown to be greater than or equal to zero
on the basis of the upward concavity of the (12:)” curve. By referring to the similar
triangles (broken lines) in Fig. 4, we note that
(J,*jj-(;I,*),j=36, (8)
1212 GEORGEE. RAMEY and NATARAJANKRISHNAMURTHY
J i I f I
where A is the length of the line segment indicated in Fig. 4. The upward concavity of the
($), curve implies that
A~($),j-(G’)4jr (10)
and hence, from equation (9),
(*Ih.j-faf)2j20* (10
Equations (5) and (11) imply that the generated lower bound (A:), curve is mono-
tonically increasing, and it can never be higher than the upper bound (;I:), curve. Since
the finite element upper bound solutions (g), converge monotonic~y from above, the
(A:)-,,values obtained in this manner are indeed the corresponding lower bound solutions.
This means then, as an example, that
or
Hence, two confo~ing finite element approximate solutions are needed to generate
the lower bound. When more than two finite element solutions are available, it is possible
to construct a sequence of lower bound solutions. It must be emphasized however, that the
generated lower bounds discussed above are based on the assumption that the (A:), curve
is concave upward. The justifications and requirements for the validity of this assumption
wit1 now be considered.
Error Estimatesfor Conforming Finite Elemmt Solutions 1213
(Yj-YJ (Yj-Yi>
Y=Yj+ (14)
(h,lh,Y’-’ = yj+ (nj/ni)P- 1
(154
or,
Wb)
or,
(154
in which superscriptsj and 2j refer to the finest and next to the finest meshes.
Error Estimates for Conforming Finite Jknent Solutions 1215
The proposed method has been successiully tested on a broad class of problems, and
in every case, the validity of the method has been demons~a~d. In these test problems,
only rarely was it found necessary to solve for more than three meshes to assure upward
concavity of the convergence curve. Results of some of the applications of the proposed
bounding method are presented in the next section.
t- U,VCfor n=2-3-4-5.. .
,” Q
$I
?=i 12,000-
x
L’
,o
---mv
E” ,9,00_ Exact (10446.2) using 5-10 data
e /
/
/
:: /
/ ous~ng 3-6 data
6
/
z
WOO- Generated LB /
z Using 2-4-edat&/
f!
IA.
k
I I
l/2 l/3 l/4 l/5 l/6 I/B l/IO I W
n-’ scale
FIG. 5. Frequency parameter for a square plate with two opposite edges simply supported and the
other two edges clamped, solved with RI6 elements.
3.97 - Generated LB /
396
n-’ scale
Ro. 6. Buckling cawcient for a simply supported square plate under &axial comprcusion, solved
with RI6 and CFQ elements.
Error Estimates for Conforming Finite Element Solutions 1217
389.5
,FE solution
Exoct(38963636) 0
Usr ng 3-Sdot
389.6
I I I
‘2 l/3 l/4 116 fk
n-t stole
FIG. 7. Frequency parameter for a simply supported square plate, solved with Tl8 cleinults.
(4x4 Shown)
I I i
l/4 l/6 l/8 l/l2 ‘0)
n-’ scale
Strain energy parameter for a clamped square plate under uniform load solved with T18
and R16 elements.
94 -‘. Generated U 6.
L\
Generated U.B. ‘~,.,(-‘(R6 elements)
92 (RP16 elements)
\ .x
\
,A&___
90
A,.
0’0
\
6.6 L- F.E. solutlan
(RPl6 elements)
64
I-
62
6.0 t
/‘ FE. solution
(R8 elements)
t7-’ scale
Fro. 9. Strain energy parameter for an cantilever beam under uniform end load solved with R8
and RP16 elements.
Error Estimates for Conforming Finite Element Solutions 1219
Exact (0 2290)
--------
l/8 l/l6 I Lo
FIG. IO. Torsional rigidity parameter for a rectangular bar, solved with ‘IT3 and ZEE elements
(A/B=2-0).
One was Herrmann’s [IS] linear triangular displacement element (TT3), and the other was
Zienkiewicz’s [ 191linear triangular equilibrium element (ZEE). On the basis of variational
principles of theory of elasticity, these elements should yield upper and lower bound
solutions respectively to the torsional rigidity of the bar [20], and both bounds are shown
in the figure by the solid lines bounding the exact solution obtained by Fourier series
analysis. The broken lines represent the generated lower and upper bounds determined
respectively from the displacement and equilibrium solutions. Although the generated
bounds are somewhat inferior to those directly obtained from the finite element solutions
themselves, the validity and adequate accuracy of the bounding technique proposed are
evident.
1220 GEORGEE. RAMEY and NATARAJANKRISHNAMURTHY
4. CONCLUSIONS
The bounding method proposed in this study requires two conforming finite element
solutions in an n-sequence, with n,, i being equal to 2ni, and provides a simple and pract,cal
means of determining a realistic error estimate. Additional solutions are needed only to
confirm the desired concavity of the convergence curve, if unknown. With this error estimate
a decision concerning the necessity for further mesh refinement may be mdde based on the
accuracy desired. The validity of the proposed bounding method has not been proved
mathematically; however, all numerical applications of the method yielded excellent results
and confirm its workability. These applications further indicate that for a tight bracket of
the exact solution it is desirable to use an n-sequence such that the finest mesh gives an
anticipated good approximation to the exact solution; and use the results of the finest and
next to the finest meshes to generate the other solution bound.
In utilizing the method, the decision on the finite element mesh sequence to be employed
in the analysis should be bared primarily on (1) the value of p (in Knmp) as determined
from a theoretical study, and (2) the criterion that the finest two meshes provide a good
approximate solution such that a tight bracket on the exact solution can be obtained. The
proposed method lends itself well to automation, including automated mesh refinement
techniques.
Acknowlcdgcmcnts-Much of the development and results reported in this paper formed the basis of the
first author’s doctoral work, conducted under the guidance of Professor P. P. Lynn of the University of
Colorado, Boulder, Colorado, U.S.A.
REFERENCES
[l] S. W. KEY, A convergence investigation of the direct stiffness method, Ph.D. Thesis, University of
Washington, Seattle, Washington (1966).
[2] M. W. JOHNSON,JR. and R. W. MCLAY, Convergence of the finite element method in the theory of
elasticity. J. Appl. Mech., ASME, 274-278 (1968).
[3] R. W. MCLAY, Completeness and convergence properties of tinite element displacement functions-
a general treatment. AIAA Paper No. 67-143, AIAA 5th Aerospace Sciences Meeting, New York
(1967).
[4] P. P. LYNN and A. P. BORESI,Kinds of convergence and improved convergence of conforming finite
element solutions in plate bending. Nuclear Engineering and Design 11, 159-176 (1970).
[fl P. P. LYNNand B. S. DHILLON,Convergence of eigenvahre solutions in conforming plate bending tinite
elements. Inr. J. Num. Met/t. Engng., 4, No. 2, 217-234 (1972).
[q P. Tout and T. I-LH. PIAN,The convergence of finite element method in solving linear elastic problems,
ht. J. solidr Struct. 3, 865-879 (1967).
[7j J. E. WALZ, R. E. FUL~N and N. Y. CYRUS.Accuracy and convergence of finite element approxi-
mations. Proc. 2nd Cot& on Matrix Methods in Structural Mechanics. Wright-Patterson Air Force
Base, Ohio, 9954025 (1968).
[8] G. R. COWPER, E. Kosrco, G. M. LINDBERGand M. D. OLSON, A high-precision triangular plate-
bending element. National Research Council of Canada Aeronautical Report LR-514, Ottawa (1968).
[9] G. E. RAMEY, A Study of the Convergence Characteristics and Accuracy of Conforming Finite
Blemcnt Solutions. Ph.D. Thesis, Department of Civil and Environmental Engineering, University of
Colorado, Boulder, Colorado (1972).
[lo] ~~.)MELosH, Basis for derivation of matrices for the direct stiffness method. AIAA J. 1, 1631-1637
[ll] H. R.&STAD, Convergence and numerical accuracy with special reference to plate bending. Finite
Efemetu Methoak in Stress Analysis, Edited by I. HOLANDand K. Bw, Tapir, Norway (1969).
[12] W. G. CARSONand R. E. NEV~TCN,Plate buckling analysis using a fully compatible finite element.
AIAA J. 7, 527-529 (1969).
[13] R. W. CLOUGHand C. A. FEL~PPA.A refined quadrilateral element for analysis of plate bending.
Proe. 2nd Con/ on Matrix MetW in Structural Mechanics, Wright-Patterson Air Force Base, Ohio,
399-440 (19as).
Error Estimates for Conforming Finite Element Solutions 1221
[14] B. S. DHILLON, Triangular finite elements for the bending analysis of thick elastic plates. Ph.D.
Thesis, Department of Civil Engin- University of Colorado, Boulder, Colorado (1970).
[IS] G. M. LINDBERG, M. D. OLSONand H. A. m, Closed form finite element solutions for plate
vibrations. National Research Council of Canada Aeronautical Report LR-518, Ottawa (1969).
[16] F. K. Boom, R. L. Fox and L. A. SCHMIT, JR, The generation of inter-element compatible stiffness
and mass matrices by the use of interpolation formulas. Conferenceon Matrix Merho& in Srrucrural
Mechanics. Wright-Patterson Air Force Base, Ohio, 397-443 (1965).
[lT] G. R. COWPEFC, G. M. LINDBJZRG and M. D. OWN, A shallow shell finite element of triangular shape.
Znt. J. Solids Srruct. 6, 1133-1156 (1970).
[18] L. R. HWRMANN,Elastic torsional analysis of irregular shapes. J. &gng Mech. Div. ASCE 91,
No. EM6, 11-19 (1965).
[191 0. c. ZIENKEWICZ,The Finite Element Medwd in Structural and Continuum Mechanics. McGraw-Hill,
London (1967).
[20] S. H. CRANDALL, Engineering Analysis. McGraw-Hill, New York (1956).
(Received 17 December 1973)
APPENDIX
Following is a listing of the code names and trial displacement functions for the con-
forming finite element models considered in this article.
RI 6: 16 degree-of-freedom rectangular thin plate element
A plate bending element reported on in [8, 12, 151, but generally credited to [16].
+ a205.
LDT: Lynn and Dhillon’s thick plate triangular element
A thick plate bending element presented in [5j and [14].
(w*)I=aI +a2x+a3y+a4x2+a,xy+r,y2
(8*)i=a,+ol~X+CtgJt
1222 GEORGEE. RAMEYand NATARAJANKRISHNAMURTHY