Beruflich Dokumente
Kultur Dokumente
53--65, 1983
Printed in Great Britain
0045-7949!83{010053--13503.00/0
Pergamon Press Ltd
INTRODUCTION
//I
OriginoL mesh
h- Refinemen+~ (subdivision)
_2/f
54
~). C. ZiENK,.'EWiCZ et
aL
the same kind? The answer lies in the fact that generally
the sequence of meshes derived in subdivision is NOT
hierarchical while the Fourier series expansion is. By
hierarchical we mean the following: if u is the unknown
function (e.g. the structural displacement) and its approximation is
u = ~ = ~ ] Nia,
(l)
Ns
(o1 Standard
N~
[~i Nier~rc~J
Fig. 3. Increase of polynomialsin element (p-refinement).Standard (a) and hierarchical(b) shape functions.
(domain)
(2)
Mn + S = 0 on F (boundary).
In the above L and M are linear differential operators.
When an approximate solution is sought in the form
n
= ~ N~a~= Nat"'
(3)
i=t
(4)
Ns
I~.@" =
( a ) Standard
( b ) Hierarchic
f"'
~5)
55
DEGENfRA'~rE
ELEMEN1
la~
h~ERA~J,
eC.~E~.EME~
Fig. 4. Three dimensional analysis of a pressure vessel indicating advantages in idealization provided by
hierarchical elements.
KI,.,,~/~
i, =
g,,,, J I.a"~'J [ f " ' J
(7)
to start the next solution. Peano[19] discusses the possibility of storing the triangularized form of (K,,>) -~ in
attempting solution of eqn (7) and such a procedure has
much to commend it. More importantly however a~,,*
could be used as a start of an iterative solution giving the
first approximation for a '"> as
a I~'* = K . , , ~ ( f
-Kl~.,.a
).
(9i
~6
:: - Ic
I :lbBIC ELEMENT
i
f
25 0
25 0
e = u-Li
AFTER NORMALISATION OF K
C0nd K ' ~ = .Lmc,,, = 390 27
~m,n
Cond Km' = 36 28
PATCH OF ~I~MF,NTS
~ CUBlC ELEMENTS;
e=tV"~'-6
,,
:
tC>
Ill
J
25 0
21 666
,:=,
25 0
Cond K~%I6t~3
21
....
Ill)
I>~
i 0)
666
(12)
~2
Cond K
~ m'=12378
e ~ N"'a ")*
~- N " ) K
(-j/f"~
,.~
- K .,,..) a"*) .
(13)
_L
Li i i
t= 15.0
2.4 Summary
In the preceding we have noted that the hierarchical
forms have the merits of
(1) Utilizing previous solutions and computation when
attempting a refinement:
(2) Permitting a simple iteration in solving for successive refinements;
(3) Always resulting in improved equation conditioning; and
(4) Giving an immediate error measure.
While items (1) and (2) are only relevant if successive
refinements are to be attempted (3) and (4) are present
even if a single solution only is to be used.
i, i_
h
This
Table 1.1. Pure traction example on cantilever beam with varying thickness t. Stress evaluation at point (7.7.0.50
~,,~, = 1.0000, tip displacement 8oxac,-- 15.0
I"
1,
i!
0.0%
0.0%
0.01000
0.0%
0.0%
I 0.00100
1.6%
0.1%
I
I
0.27,
0.0%
O. 00010
15.7%
7.0%
17.4%
O. 5%
0. 00001
FALLS
FALLS
I.
CASE A - I0 e i g h ~ - n o d e d i s o p a r a ~ e t r i e e t e m e n t s
CASE 8 - i0 e i g h t - n o d e d h i e r a r c h i c e l o m e n : s
CASE I - relative error in compu=ed Cxx s=ress
CASE
IT - relative error
in computed d i s p l a c e m e n t
57
Table 1.2. Applied shear at the free end with varying thickness: stress evaluation at point (8.25, 0.3873t)
,r+x~ct= 31.371t, tip displacement 6~x~c~= 0.135 (theory of beams)
I
A
O. 0%
0.02%
0.02%
0.07%
FALLS
t
I.OO
0.O%
O. I0
3.9%
3.3%
4.58%
0.01
FALLS
FAILS
FALLS
CASE A - iO e i g h t - n o d e d i s o p a r a m e t r i c elements
CASE B - IO e i g h t - n o d e d h i e r a r c h i c elements
CASE I - relative error in acomputed stesses
CASE Ii - relative error
h'[lilllliltlilllillll~lllllll]llllllllllllllllllllllllllllll~
,/,///,///,/,-,,:~ ~ ,
,////:~ ~
in computed displacement
@j
or~+A,
"//:
N;
f
/
3. ERRORINDICATORSANDFA'TIMATORS
~
....
f
f
Finite element
suoerposeO
3.1 Definitions
If we are to adopt the procedures of refinement two
types of question will arise. First, we shall need an
INDICATION WHERE additional refinement is most
effective and second a MEASURE OF ERROR determining whether the refinement is necessary.
In the previous section we have defined the "error" by
eqn (10) as
l ~ o b o l function
e=u-fl.
Fig. 8. Local-global finite element method, a hierarchical form.
-E
(14)
05)
(16)
58
O . C . ZIENKIEWICZ et ai.
(17)
I]ei]~"= - (
df~
(18)
Jfl
ile-,tl~: ~ '7~+,
(20)
e,,+) --- g n . l a , , + t
and
a . . l = (./,,+,- K.+I.,,#"))IK. +,...,
(21)
(22)
,2,,
and that the quantity on the r.h.s, is always positive. To
estimate the error we shall therefore evaluate
111
rl.+, = -
N..lrd
Ijl
K,,., ....
"
(26)
,,.-i,,,+,
(24a)
or
rl~,*, = 0 c . + , -
(23)
e . . i r dfZ.
If,
K..,.,,a(")):/K.+,..+~.
(24b)
While the form (24b) may appear more direct and has
been/dent/fled as an energy error indicator by Peano[19]
we find that the alternative form given by (24a) and
introduced by the present authors [20, 271 has a special
sign/ficance since we can avoid the computation of some
stiffness terms of the additional solution.
We stress once again that no additional solution is
required to evaluate the indicators and this can be done
by expression (24) separately for each possible new
degree of freedom INDICATING where corrections are
most desirable.
59
Inter - element
zone I
L2"
II
Fig. 9. Interior elements and interface zones.
difficulty arises in determining such terms as
I!ehc = e~
r 2 dO +
elements
f Nr dft.
(28)
O2(dan:
fo
~ n~-dn = dn o
(33)
(27)
fi N;rdf~=frfo'NirdndF
(29)
Ell - v
1---L=f N2.,df~/K,+,.,l.
fo
* r dn m J
(30)
(31)
E(fN:o+ldr)(fS'-dr))/Ko.,.o.,.
(32)
(34)
Strategy
4.1
A finite element solution will generally be started on a
specified mesh. This may be deemed by the analyst to be
sufficiently accurate or may represent a very crude subdivision machine generated with sufficient accuracy to
model the boundaries only. In both cases the first step of
additional computation must be the estimation of error
and if this is found sufficiently small the computation
will be terminated. Error estimators are of primary importance here.
If the error estimate is too high, further refinement is
necessary and here error indicators will tell us where
best to refine. Various strategies are here obviously
possible--in all attention turning to the degree of
freedom where indicators are largest. Some of these are
outlined below:
(1) We could for instance fix
an indicator
value 77 = ~ above which all the degrees of freedom with
higher indication would be activated;
(2) A variant of (1) would be to define the threshold
value ~ after consideration of the total error; or
(3) At each stage we could activate all degrees of
freedom which show an indicator higher than a prespecified proportion y of the maximum indicator value
and refine when "0/>, (max).
a priori
O . C . ZIENKIEV,[CZet aL
60
This third alternative has been adopted in our examples with ,/ being in the range of 0.25-0.5. Obviously,
v = 1 corresponds simply to a complete refinement of aJl
elements and is a procedure adopted in one of the
recently launched commercial programs.
In all procedures of adaptive refinement (at least with
7 < 1) we shall, after many steps, reach a situation of
nearly constant error indicators r/. If the process was
started from a good subdivi
ais stage will obviously
be reached faster than if ;
start ~s made, but the
final result is the same. At ..... ~,oint we shall make the
assertion that the mesh is optimal when a (nearly) uniform distribution of error indicator is reached. It appears
reasonable to argue that at this stage the d.o.f.'s are
placed so that for their given number the error in energy
is least. While the formal proof of this equivalence with
the early efforts of attaining mesh optimality has only
been given for a simple line element (Ref. [26, 29]) the
assertion is reasonable--and we note that whatever the
starting point the near optimal solution will always be
reached adaptively.
There are other features of this mesh optimality. It can
be shown (at least for h convergence with linear elements) that only at that stage the error estimators
become fully accurate [26, 29].
More importantly it is only at this final stage that the
full theoretical rate of convergence appears to be
reached and the effects of singularities which always
reduce the convergence rate is eliminated.
4.2 Tests of adaptive and complete refinement. Con-
vergence rates
In Fig. 10 we show an example which due to the
1.0 '~
~ ; ;
;
=
25.0
16 T
"8.3
"85
167
L
St~
250
conditions. = 25.8%.
$
Oy ~yOy
f~$
CL*M~L"
.. ~.,.cL
CL_,
**
Ct
Step 6 ,
a d a p t i v e h - version
II~llZE = 30.841,
= 9.8%. 4 5 2 OOF
72 DOF
'
~$
~J
..,l 3~
Step 5, a d o p t i v e p - v e r s i o n 11~1t~- 3 0 9 6 3 7 ,
= 7.2%. 1 4 4 DOF
Fig. 10. Initial and adaptive meshes (h and p-refinement) for L-shaped domain. Relative error (e%).
61
error eit(l )
error eet(2 )
error eet(3 )
E1
I III
Complete
p-refinement
adaptive
p-refnement
7.2%
1.57
0.77
1.03
0.3
6.1%
1.49
0.74
0.88
1.4
25.3%
1.70
0,7
0,97
1.0
19Z
1.4
198
0.39
7.0%
0.62
11%
0.3
5.4%
324
0.34
6.1%
0.50
9%
0.24
4.5%
72
1.44
25.8%
2.45
41%
1.0
19%
0.97
82
1.12
20.0%
1.75
30%
0.73
13%
1.01
19.0%
l,Sb
0.65
0.89
i01
0.81
14.4%
1.13
20%
0.50
9%
0.64
12.0%
1.40
0.62
0.80
118
0.58
10.5%
0.88
15Z
0.39
7%
0.48
9.0%
1.5
0.67
0.83
144
0.40
7,2%
0.65
11%
0.30
6%
0.36
6.5%
1.6
0,75
0.90
error
estimate(2
) - error
estimate
from
( 33 )
error
estimate(3
) - error
estimate
from
( 33 ) w i t h
error/exact
corrective
//
L-Lineor elements
Q-OuoOrotic elements
C-Cul)ic elements
~U- Quortic elements
0'9~
4o./.
L,;,:,::/
o-
I. LirL,~r h - version
/
(complete)
/ . ~
..-,.
i///~
'/p/
iv
1.8
factors
(20)
error.
NDF
0.4
41%
fro~= ( 32 )
0.7
~.45
estimate
1.70
25.8%
) - error
:J
25.31
1.44
estimate(1
'
72
error
e = predicted
2 Ouolrotie h-ver$io
5. HIERARCHICALAPPLICATIONIN TRANSIENTAND
NON-LINEARANALYSIS
6. EIGENVECTORSASHIERARCHICALVARIABLES
An extreme form of hierarchical formulation is
obviously the choice of eigenvectors of the system as the
basic variables. As is well known the choice of such
shape functions will result in a diagonal form of equations and the optimal condition for the formulation
would arise. (This indeed was the case with the Fourier
series techniques we mentioned at the outset of this
paper.)
In many solutions of dynamic problems approximate
values of such egienvectors are used (viz. Guyan-Irons
reduction techniques, etc.[14]). Similar techniques have
been used in non linear analysis with so called "reduced
basis" techniques[33, 34]. While such techniques often
call for substantial initial computation, much could be
accomplished we believe with the use of global functions
of "reasonable basis" type combined with the use of
more automatic hierarchical forms.
?. CONCLUDIng~MAP,KS
The objective of this paper was two fold. One, to
advocate most strongly the very general adoption of
hierarchical form and second, to address the question of
adaptive refinement and error estimation. We hope that
//
-2
4. Adoptive p
5. Acloptive h
LS"/.
I/"~D-~rsion
//
.....
(overooe st.ope I
1
-3
l
-2
LOO, I / N
-L
f).
r:"
C. /~IE~KiEWIC2 :'t
~IL
___.j
/ /
\,,\
.}I~/
Fig. 12aJ.
/
'
.<-.~.~ " ~
.....
/"
----~-~-~
:--
PredlCtlOfl of error
e = 20.4 (32)
-- -- ~--v-'='" T
(-)
I~L-- ___..__
!+)
Fig. 12(b).
--
Quadratic d o t
"
(x direction)
63
I
)
~.L_--~
~'~.~'~\
I --
.....
=
~~'~
'Exoct'principal stress
Prediction of error
~ = 8.4 (32)
(*!
(-)
Fig. 12(d).
- - - C ~ i c d.o.f (x direction)
...... Q u a r t i c d o.f. ( x direction)
Ox Centre quodrotic d.I~f Ix direction)
~"
-~-Ox \ Ox ~ ,
~
t-).'~
o -- Oxy
Quoclrotics on element boufclor,es not
represented
Fig. 12(e).
!) C. ZIENKIEW[CZ ef a,1,
, U~r'---._
(+)
(-)
Fig. 12(0.
Fig. 12. (a) Mesh for imtial solul~on O18 d~Fees of freedom). (b) Adaptive refinement, initialst~---solutionon
Acknowledgements--The second author acknowledges the support of Instituto Superior Tecnico, Lisbon, Portugal, and Gulbenkian Foundation, Lisbon, Portugal, Fellowship No. 18t79/B.
RF_cEIIENCgS
t. N. R. Buell and B. A. Bush, Mesh ganeration--a survey.
ASME Publ.--Paper No. 72--WAtDE-2. 1972.
2. W. C. Thacker, A brief review of techniques for generating
irregular computational grids. Sea-Air Interaction Laboratory, 15 Rickenbauer Causeway, Miami, Florida 33149, 1980.
3. O. C. Zienkiewicz and D. V. Phillips, An automatic mesh
generation scheme for plane and curved element domains.
Int. J. Num. Meth. Engng 3, 519--520(1971).
4. g, J. Melosh and P. V. Marcal, An energy basis for mesh
refinement of structuralcontinua. Int. f. Num. Meth. Engng
II, 1083-1091 (1977).
5. E. R. de Arantes e Oliveira,Optimization of finiteelement
solutions. Proc. 3rd Conf. on Matrix Meth. in Structural
Mech. Wr/ght-Patterson A/r Force Base, Ohio (1971),
19
20.
21.
22
23.
24.
25.
65