f. j. p e. s -p m. a. vilariÑoarchive.numdam.org/article/m2an_1998__32_2_131_0.pdfmembrane locking...

MODÉLISATION MATHÉMATIQUE ET ANALYSE NUMÉRIQUE

D. CHOI

F. J. PALMA

E. SANCHEZ-PALENCIA

M. A. VILARIÑOMembrane locking in the finite element computationof very thin elastic shellsModélisation mathématique et analyse numérique, tome 32, no 2 (1998), p. 131-152<http://www.numdam.org/item?id=M2AN_1998__32_2_131_0>

© SMAI, EDP Sciences, 1998, tous droits réservés.

L’accès aux archives de la revue « Modélisation mathématique et analyse numé-rique » (http://www.esaim-m2an.org/) implique l’accord avec les conditions géné-rales d’utilisation (http://www.numdam.org/conditions). Toute utilisation commer-ciale ou impression systématique est constitutive d’une infraction pénale. Toutecopie ou impression de ce fichier doit contenir la présente mention de copyright.

Article numérisé dans le cadre du programmeNumérisation de documents anciens mathématiques

http://www.numdam.org/

http://www.numdam.org/item?id=M2AN_1998__32_2_131_0

http://www.esaim-m2an.org/

http://www.numdam.org/conditions



MATHEMATICA!. MODELLING AND NUMERICAL ANALYSISMODELISATION MATHÉMATIQUE ET ANALYSE NUMÉRIQUE

(Vol 32, n° 2, 1998, p 131 à 152)

MEMBRANE LOCKING IN THE FINITE ELEMENTCOMPUTATION OF VERY THIN ELASTIC SHELLS (*)

by D. CHOI 0), F. J. PALMA (2), E. SANCHEZ-PALENCIA (a), M. A. VILARINO (2)

Abstract — The membrane locking phenomenon anses in cases when the middle surface of the shell admits "pure bendings " satisfyingthe kinematic boundary conditions It then appears that the discrete approximations by finite éléments is unsuited to descnbe such purebendings, which are the limit configuration of solutions as the thickness tends to zero This phenomenon is descnbed in terms of lack ofrobustness (i e lack of uniformity of the finite element convergence h N 0 with respect to the thickness of the shell 2e ) We prove that anyfinite element scheme consisting of in piecewise polynomial functions necessarüy exhibits locking for certain shells (and probably for almostany shell admiting pure bendings) Numencal experiments are done for a hyperbolic paraboloid The superionty of scheme s involving highorder polynomials (Ganev-Argyns in particular) is shown It is also seen that reduced intégration have very little influence on membranelocking

Résumé. —Le phénomène de verrouillage membranaire intervient lorsque la surface moyenne de la coque admet des «flexions pures »satisfaisant aux conditions aux limites cinématiques Dans ce cas les approximations discrètes par éléments finis sont inappropriées pourdécrire ces flexions pures, qui sont les configurations limites des solutions lorsque V'épaisseur tend vers zéro Le phénomène est décrit enterme de manque de robustesse (ou manque d'uniformité de la convergence des éléments finis h N 0 par rapport à l'épaisseur de la coque2 £ ) Nous démontrons que tout schéma éléments finis qui consiste dans des fonctions polynomiales par morceaux, présente nécessairementle phénomène de verrouillage pour certaines surfaces (et probablement pour presque toutes les surfaces admettant des flexions pures). Ona fait des exemples numériques pour le paraboloide hyperbolique, montrant la supériorité des schémas qui font intervenir des polynômesde degré élevé (Ganev-Argyris en particulier) On voit aussi que l'intégration réduite a une très petite influence sur le verrouillagemembranaire

1. INTRODUCTION

It is known that the phenomenon of membrane locking in numerical computation of thin shells consists in aninadequacy of the finite éléments to describe the very peculiar déformations of a shell [15], [10], [6], [9]. Actually,the natural trend of a thin shell is to perform pure bendings, i.e. inextensional displacements of the middle surface(which do not modify the intrinsic metrics of the middle surface), but the explicit description of the phenomenonand the above mentioned inadequacy of the finite element approximation usually lack in the littérature.

The asymptotic study of the behavior of elastic shells as the thickness 2 s tends to zero [15], [16], [17] showstwo very different asymptotic behaviors in the cases when the medium surface with the kinematic boundaryconditions admits or no inextensional displacements. In the former case, the shell is called to be "non-inhibited"(for "with non-inhibited pure bendings") and in the later it is said to be "inhibited" (for "with inhibited purebendings"). It should be noticed that the concept of "inhibited" coincides with that of "geometrieally rigid", butwe adopt the term "inhibited", as "rigidity" is a different concept in mechanics of continua.

It then appears that the membrane locking only occurs in non-inhibited shells. This elementary assertion isnevertheless useful to introducé a little order in the comments on locking which are usually encountered in theliterature. For instance in [3], p. 238-239 are considered three examples of shells which are taken as benchmarks

(*) Manuscript received September 16, 1996C1) Laboratoire de Modélisation en Mécanique, Université Pierre et Mane Cune, 4 place Jussieu, 75252 Pans Cedex 05 (France)(2) Departamento de Analisis Matematico, Facultad de Ciencias, Campus de Teatinos, 29071 Malaga (Spain).This work is part of the "Action intégrée franco-espagnole PICASSO 95-96" Trus work is also part of the of the Human Capital and

Mobüity Program "Shells Mathematical Modellmg and Analysis, Scientific Computing" of the Commission of the European communities(contract ERBCHRXCT 940536).

M2 AN Modélisation mathématique et Analyse numérique 0764-583X798/02/$ 7.00Mathematical Modellmg and Numencal Analysis © Elsevier, Pans

132 D CHOI, F J PALMA, E SANCHEZ-PALENCIA, M A VILARINO

for testing the adequacy of finite element methods but only the third one (semi-sphere with free boundary) isnon-inhibited; the others are not concerned with membrane locking. Moreover, no tests are given for very thinshells, when membrane locking is an asymptotic phenomenon for small thickness. It should also be mentionedthat arches, which are the two-dimensional counterparts of shells are always non-inhibited, and the correspondinglocking is simpler and better known [6].

It is a remarkable f act that a shell is, by définition, thin, but nevertheless, the équations of classical theory ofshells (see, for instance [13], [4]) contain two terms corresponding to the déformation énergies of membrane strainand flexion. The corresponding coefficients are in a ratio e" as the membrane rigidity is higher than the flexionone. Consequently we may study the asymptotic behavior of a thin shell as e \ i 0 ([15], [16], [17] alreadymentioned). The classical convergence of the finite element approximation as the mesh diameter h tends to zero[4] holds true for fixed e, but the constants appearing in the convergence estimâtes depend on e. Babuska and Suridefine the locking [1] as the non-uniformity of the convergence with respect to e. They also call "robust" anapproximation method when the convergence h \ 0 is uniform with respect to a parameter s. Locking is in factlack of robustness. This définition, which we adopt in this paper, seems to be quite fitted as an approximationmethod in theory of shells, which are naturally thin, should convergence quite well for any value of the smallthickness.

Otherwise one must change the mesh diameter when changing the thickness. We shall see that this is the genera!situation, and the values of h which are necessary to get a good approximation with small e are impracticable.The paper [10] by Kamoulakos is quite interesting in this context. After giving a review of the work accomplisedmainly by L.S.D. Morley, he considers a spécifie type of quadrangle finite éléments and shows that the appropriateh should be of order e or even e2 in certain directions! He also exhibits finite éléments computations of shellstresses with errors going from 2 % to 46 000 %!

We also notice that the locking phenomenon seems also be present in a variety of approximation methods. Seefor instance [9], [6], [11] for the approximation of arches by straight beam éléments and [1] for approximationof shells by plane facets.

In this paper we adopt a view point very close to that of Chenais and Paunüer [6] but we consider the caseof shells instead of that of arches. Our main resuit (section 4) is that any finite element internai approximationmade of functions which are piecewise polynomials leads to locking for certain surfaces such as the hyperbolicparaboloid and the straight heücoid. The proof relies on incompatibilities involving explicit expressions of theéquations for these surfaces. We also show in a les s rigorous way (begining of section 4) that generally speaking,thereis incompatibility between the pure bendings and the finite dimensional space Vh of the finite elementapproximation from which the gênerai character of the locking phenomenon follows.

The previous considérations are then extended (section 5) to a widely used kind of non-conforming finiteéléments, the so-called DKT (= Discrete Kirchhoff Triangle) approximation.

Section 6 is devoted to numerical experiments. We chose a non-inhibited shell (a hyperbolic paraboloid withappropriate boundary conditions) which is not the case in a part of the usual benchmarks. Computations were donefor two finite element schemes: Ganev-Argyris (conformai, using high order polynomials) and Sander (non-conformal, with lower oder polynomials). For each one, several different intégration schemes were considered.Numerical results show the superiority of higher order polynomials. Indeed, fairly good computations wereobtained with an Apollo station for e > 10" 3 (Ganev-Argyris) and e > 10" 2 (Sander), where 2 e = thickness.Otherwise, very little influence of the intégration schemes was observed, contrary to the wide-spread opinion thatreduced intégration diminishes locking effects.

2. SETTING OF THE PROBLEM. ROBUSTNESS AND LOCKING

Let us first state the shell problem in the framework of Koiter theory (see for instance [4]). Let E3 be theEuclidean space referred to an orthonormed frame (e1; e2, e3) and let Q be a bounded open set of R2 withboundary F. The middle surface of the shell (denoted by S) is the image in E3 of Q by the map

(2.1) (p : ( / , ƒ ) E Ï2-*<p(y)e E3

M2 AN Modélisation mathématique et Analyse numériqueMathematical Modelhng and Numerical Analysis

MEMBRANE LOCKING IN THE FINITE ELEMENT COMPUTATION 133

At each point of 5 we consider the two tangent vectors

(2.2) aa = <p a = d<çfdyŒ ^ öaq> ( a = 1, 2 )

(greek indices run from 1 to 2, and latin ones from 1 to 3. Note also the different notations used for partialdifferentiation). The unit normal vector to S is

(2.3) a3 = a 1 xa 2 / | a 1 xa 2 |

The map <p, and then S are supposed to be smooth; we then consider, in a neighbourhood of S the "normalcurvilinear coordinates" y1, y2, y3 where y3 is the distance to S along the normal. For the sake of simplicity weonly consider the case of constant thickness 2 s. The shell is then the set

(2.4) C={MeE3, M = <?(y\y2) +y3 a3, ( y V ) e Û , |y3| < e}

Let u(y , y2) be the displacement vector of S when the shell is submited to forces e3 f by unit surface. We onlyconsider linearized theory for small u. The Koiter theory is then described in terms of the linear term of themembrane strain tensor

(2-5) >W(u) = («

and of the linear terms of the change of curvature tensor

(2-6)

In the previous expressions, aa^(resp. an) and b /resp. b «) dénote the coefficients of the first and secondfundamental forms of S before (resp. after) déformation, Le.

where, a dénotes differentiation with respect to y". Correspondingly, the symbol | a will be used for covariantdifferentiation.

The contravariant basis a" is defined by

(2.9) a B V = ai

where ö dénotes the Kronecker symbol. The contra variant components of the me trie tensor are

(2.10) fl^ = a a . a^ .

They are used, as well as aap to pass from covariant to contravariant components of vectors and tensors in theusual way. The tensors y and p are then:

(2.H) y « / u ) = \ <X|/? + w/?|J - *v u3

vol. 32, n° 2, 1998


where

(2.13)J À J X J-lX J V J-rV jXbP\a = bp, a + * av bp ~ l $a bv

u -u -rx uU3\a.p~ M 3, dp £ aft U2>, X

and F are the Christoffel symbols of S:

(2.14) /X = />a<\a^ = a<\a^.

We then define the bilinear forms of membrane and flexion energy:

(2.15) ao(u, v) = f af" y (u) y„/v) dS

af"(2.16) ax( u, v ) = j s af" / y ( u ) pafi( v ) dS

where the coefficients of a0 and al are elasticity coefficients satisfying the classical properties of symmetry andpositivity

(2.17) a^"a = a"aap = a""™1

(2-18) a^"" ^ , ^ c | | ^ | 2 V ^ symmetrie .

Moreover, the shell is supposed to be clamped by a part Fo of the boundary, simply supported by another oneFx and free by the remainder part. The kinematic boundary conditions to be prescribed are

fu = O, du3/dv onf0

( u = 0 onT,

where v dénotes the normal to the boundary. Let V be the space of the kinematically admissible displacements,defined by

(2.20) V={(vvv2,v3) e Hl x Hl x H2 ; v satisfy (2.19)}

The boundary conditions are supposed such that

(2.21) [ a o (v ,v )+^(v ,v ) ] 1 / 2

is a norm on V equivalent to the classical one. This is the case if Q is connected and Fo is not empty [4], [5].The shell problem is then:

Find ue e V such that(2.22)

Vv e

M2 AN Modélisation mathématique et Analyse numériqueMathematical Modelling and Numencal Analysis


where f G V", e e (0, e0) and the duality product is given by

(2.23) (f,v)yv= fvtdS.Js

Clearly, by virtue of the hypothesis that (2.21) is a norm equivalent to that of V, the shell problem (2.22) has aunique solution for each fixed value of £ G (0, £0), in the framework of the Lax-Milgram theorem.

Our aim is to study the asymptotic behavior of u£ as s tends to zero and the corresponding Galerkinapproximations.

Let Vh (with h \ 0) be a family of finite-dimensional subspaces of V such that Vh approaches V, i.e. such that

(2.24) Vv G V, üm inf || v - v j = 0

or, in other words:

(Vv G V and h \ 0

there exists \h G Vh such that

v G v strongly in V( h \ 0 ) .

We then consider the approximate problem associated with (2.22):

{Find u^ G Vh such that1

-2Û0(uh, V) + fl^U^, V) = (f, V) v v Vv G Vh.

Moreover we shall admit that

( 2 " 2 7 ) | uh -» u£ strongly inV(h \ 0, fixed e ) .

where u* and uÊ are respectively the solution of (2.26) and (2.22). We note that (2.27) holds true for a certainnumber of finite element approximations of the shell problem (see for instance [4], section II. 1).

DÉFINITION 2.1: We say that Vh implies a robust approximation (2.26) of the problem (2.22) if the convergence(2.27) is uniform with respect to e, i.e. if f or given ö there exists y such that

(2.28) \\uh~ue\\v<ôforhe (0 , y) and £ G ( 0 , £ 0 )

We also say that the approximation Vh locks when it is not robust.Remark 2.2: Property (2.28) ensures that an accurate approximation of the solution may be obtained with a

sufficiently small mesh diameter h, independently of the thickness. Another possible définition of robutnessinvolves uniformity of the convergence with respect to s and to f (for f belonging to the unit bail of V\ forinstance), but we shall adopt in this paper the définition 2.1. H

In practice if we have a locking approximation, we must fix y in a suited form for each value of £. In particular,changing the values of s with a fixed mesh may spoil the quality of the approximation.

The following theorem will be the main tool for the study of locking in the next sections.

vol. 32, n° 2, 1998

136 E>. CHOI, F. J. PALMA, E. SANCHEZ-PALENCIA, M. A. VILARINO

THEOREM 2.3: Under the previous gênerai hypotheses, let us consider the reiterated limits:

(2.29) lim lim u' strongly in V

(2.30) lim lim u' strongly in V

and let us suppose that one the two hypotheses a) and b) hereafter is satisfieda) the limits (2.29), (2.30) exist and are different,b) the limit (2.29) exists but (230) does not.

Then the approximation Vh is not robust.

Remark 2.4: Let us point out that, denoting the solutions of (2.22) and (2.26) by ue and u* respectively, thelimits

i) lim ue strongly in Vii) lim u£

h strongly in V, for fixed halways exist, as we shall prove later on in Theorem 3.2. In particular, from (2.27) and i) it follows that thereiterated limit (2.29) always exist, whereas (2.30) may exist or not.H

Proof of Theorem 2.3: It is a classical corollary of the theorem on uniform convergence of continuons functions.For the sake of completeness, let us recall theorem 66 of [20], vol. 1, p. 150: "Let E and F be two metric spaces,A a part of E and f09fv ...,ƒ„ ... a séquence of mappings of A into F converging uniformly to ƒ. Let a be anaccumulation point of A in E. If for each n, fn(x) has a limit when for x tending to a by points in A and if Fis complete, then f(x) has a limit for x tending to a by points in A and moreover"

(2.31) Jimy(x) = nlim |.lim fh(x)A Lx€A J

Let us apply this theorem taking E = IR, F = V (which is a complete space), x = e, A = (0, £0)f(x) = u£, fh(x) = ue

h (with h = hn, a séquence tending to 0 as n tends to infinity, a = 0. Let us suppose thatthe convergence u^ —» u£ is uniform with respect to e G (0, £0) ; then (2.31) holds true, taking the form

(2.32) lim u£ = lim [ lim u ' l

which is in contraction with a) and b). •

3. LIMIT PROCESSES AND CONSEQUENCES ON LOCKING

Let us define the closed subspace G of V formed by the pure bendings by:

(3.1) G - { v e V;a0(v,v) = 0}

= {V<E V;aQ(\, w) = 0 Vwe V}

We note that the three spaces appearing in the right side of (3.1) are the same. The equality of the first and thethird ones follows from the définition of the form a0 (2.15) and from the positivity of the coefficients (2.18). Theequality of the first and the second spaces follows from the Cauchy-Schwarz inequality for the seminorm

(3.2) | a o ( v ' w ) | ^ao(v,v)1/2ao(w,w)1/2.

M2 AN Modélisation mathématique et Analyse numériqueMathematical Modelling and Numerical Analysis


The fact that G is closed follows immediately from (3.1) using for instance the second expression in the right side.Consequently G is itselfa Hubert space for the topology induced by that of V, and we may state the limit problemfor e ^ 0 :

Find u° E G such that

Vve G

We note that, as a0 vanishes on G, the left side of (3.3) is a form continuous and coercive on G, and consequentlyu° exists and is unique in the classical framework of Lax-Milgram.

Remark 3.1: Let us consider the operators

(3.4) Ae e J2?( V, V) , i e j£?( G, G')

defined by

(3.5)

(3.6)

Vu, v e V

(Au, v)G,G = fll(u, v) Vu, v G G .

It classically follows from the Lax-Milgram theorem that Ae and Â define isomorphisms of V onto V' and of Gonto G' respectively. Moreover, V' may be decomposed as a direct sum:

(3.7) V'=G'®GP

where

GP = {ÎG V',<f>v>vv=0 VVG G}

is the polar set of G. We also note that the necessary and sufficient condition for u° ^ 0 is that the componentof f on G' does not vanish. •

In the same way we define for each h:

(3.8)

Let u£ be the solution of

We then have:

G, = G n Vh = {v e Vh ; aQ( v, v) = 0} .

{Find u? e G, such thath h

a1(u2,v) = <f,v>vv

THEOREM 3.2: Let ue, u^, u°, u^ be the solutions of (2.22), (2.26), (3.3) and (3.9) respectively. Then,

(3.10)

(3.11)

vol. 32, n° 2, 1998

uÊ -» u° strongly in V( s \ 0 )

-> u£ strongly in Vh(s \ 0, ûxedh)

138 D. CHOI, F. J. PALMA, E. SANCHEZ-PALENCIA, M. A. VILARINO

Proof: It is classical as (3.10) and (3.11) are penalty limit processes. Let us prove for instance (3.11). Takingv = u^ in (2.26) it follows from the coerciveness of a0 + ax :

(3.12)

G

for two non-vanishing constants c and C. We then see that the ush remain in a bail of the finite-dimensional space

Vh. After extraction of a subsequence (in fact it is the whole séquence because we shall see later that the limitis unique)

(3.13) u£ -» u* weakly and strongly in Vh(s \ 0, fixed h) .

It follows from (3.12) that

(3.14) ao(u%ueh)

m^Ke.

On the other hand, for any fixed v e Vh, we have from (3.13):

(3.15) ao<X>v) - > « o ( u * ' v ) O ^ 0, fixed/*),

and, using (3.14) and an estimate analogous to (3.2) we see that the left side of (3.15) converges to zero inmodulus, and consequently <zo(u*, v) = 0. Taking v = u* we see that u* e Gh. On the other hand, taking in(2.26) v G Gh we have:

(3.16) ax(u% v) = <f, v ) v v Vv e Gh .

and letting £ \ 0, we have, according to (3.13):

(3.17) a1(u*,v) = <f,v>vv V V E G , .

As we saw that u* G Gh, u* = nh is the only solution of (3.9) and (3.11) is proven.The proof of (3.10) follows exactly the same steps, showing the weak convergence in V. The strong

convergence in V needs and ulterior reasoning. As we pointed out before, the proof is classical, and may be seenfor instance in [6], Theorem 1. •The result (3.10) constitutes the fundamental property of thin shells. When the thickness tends to zero, thesolutions converge to a solution which is an inextensional displacement, i.e. a pure bending. This property ismeaningfull in the case when the shell is "non-inhïbited\ i.e. S admits non zero pure bendings, or

(3.18) G * {0} .

In the inhibited case, i.e. G = {0}, (3.10) always hold true, but evidently u° = 0. In order to have a betterdescription of the asymptotic behavior of ue we write ue = e2 \e and we study the limit behavior of Ve. Thisasymptotic process, which has nothing to do with the membrane locking may be seen in [15] or [16], [17].

According to (3.11) an analogous behavior holds for the approximate problem in the space Vh. The membranelocking occurs when Gh does not approach G. More precisely:

PROPOSITION 3.3: Let us admit the gênerai hypotheses of this paper, inparticular Vh approaches V, (see (2.25)).Moreover, let us assume that the subspaces Gh — G r\Vh defined in (3.8) do not approach G, defined in (3.1)(in other words, there exists at least a non-zero element v of G which cannot be approached in the form (2.25),with G and Gh instead of V and Vh, respectively).


MEMBRANE LOCKING IN THE FINITE ELEMENT COMPILATION 139

Then the approximation of problem (2.22) by Vh is not robust (locking holds at least for certain î,for instancefor f = Â\ where A dénotes the operator defined in (3.6)).

Proof: We consider problems (2.22) and (2.26) with f = Â\. We are showing that we then are in the hypothesesof theorem 2.3. From (2.27) we see that the limit / î \ 0 o f u ^ exists and is ue. Next, the limit £ \ 0 of u£ existsby (3.10) and is equal to u°, which is in fact v for the considered force f (see (3.3) and (3.6) if necessary). It thenfollows that the reiterated limit (2.29) exists and is equal to v. On the other hand, the reiterated limit (2.30)becomes by virtue of (3.11)

(3.19) lim u? strongly in V

but u° e Ghcz G, and the convergence in (3.19) is in fact "strongly in G". This limit may exist or not exist, butin any case by hypothesis, it is not equal to v, and theorem 2.3 applies. •

Remark 3.4: Locking holds in particular when G =É {0} but Gh = G r\ Vh~ {Q} for any h. In that case wehave a total locking, for any f with non-zero component in G'. •

Remark 3.5: We shall see in the next sections examples where v e Gh implies v3 = 0, but G contain élémentswith non-zero third component. Clearly in that case the Gh do not approach G (see (2.25) with G and Gh insteadof V and Vh, respectively) and locking holds. •

4. ANY PIECEWISE POLYNOMIAL INTERNAL APPROXIMATION IS NECESSARILY NOT ROBUST (OR LOCKS)

In this section we are proving that any finite element internai approximation of the shell problem such that theéléments of Vh are in each triangle of the mesh polynomial functions (i.e. practically any finite element method)locks for certain shapes of the middle surface. Clearly we may wonder if there is no locking for other surfaces.This is the reason why we first give some heuristic reasons showing that locking is probably a genericphenomenon, holding for almost any surface, with a very few exceptions.

As we saw in proposition 3.3, locking appears when the spaces Gh = G n Vh do not approach G where

(4.1) Gr{ve^

In order to get an intuitive idea of the nature of this problem, let us consider to simplify and fix ideas, the casewhen the b^ are constant and the Christoffel symbols vanish. The constraints appearing in (4.1) are then:

(4.2)

[ \ <<d2 Vl + dl V2) ~ bl2 V3 = 0

Let us admit that one of the coefficients bn and b22 does not vanish. This condition is generally satisfied, andcertainly it does if the surface is either elliptic or parabolic. Let bn = 0. The first condition (4.2) shows that

(4.3) ve Gh^v3 = ±d1vl

Let us consider the restriction of (4.3) to a triangle of the mesh. Then, as v l is a polynomial of some degree m,v~ is necessarily a polynomial of degree m — 1. This is perfectly incoherent with the fact that the components

1 9

va (a = 1, 2) belong to H but t>3 e H : any finite element scheme classically used in shell theory (cf. [4] forinstance) uses polynomials of higher order in the discretization of v3 than in that of vv v2. In these conditions,the convergence of the approximation is seriously compromised.

It should be noticed that the above reasoning is very close to that of [6], theorem 4, where it is shown that inthe case of a circular arch, there is a robust non-classical approximation by finite éléments where va

vol. 32, n° 2, 1998

140 D CHOI, F J PALMA, E SANCHEZ PALENCIA, M A VILARINO

(a — 1, 2) and v3 by polynomials of orders 4 and 3 respectively But we should not conclude that discretizmgi?3 by polynomials of lower order than v a is a gênerai way for avoiding locking We are showing that any internaiapproximation which is piecewise polynomial locks for certain surfaces

Our proof is mainly based on Remarks 3 4 and 3 5 For certain spécifie surfaces, we shall show thatv e Gh implies either v = 0 or i?3 = 0 The cases considered are the straight helicoid, and the hyperbolic andelliptic paraboloids Clearly the applications to locking only concern non-inhibited surfaces, an this implies someconditions on the fixation of the boundanes In particular, a part of them (or even the whole) must be freeExamples may be seen in [18] and [8], but we shall see later (section 6) an example of non-mhibited hyperbolicparaboloid But the following proofs are essentially very simple and are not concerned with the boundaryconditions According to (4 1), taking v e Gh and taking the restriction to an element of the mesh, it is sufficientto show that if each component of v is a polynomial then

(4 4) ya/?(v) = O(a , /?= 1,2) => v = 0 (ormerely v3 = 0)

Case of the hyperbolic paraboloid

It is the surface defined by (notation of sect 2)

(4 5)

We obtain easily

(4 6) a i = (l,0,/), a2=(0,

where

(4 7) a = l +

Moreover,

(4 8) bu = b22 =

V = a~ \ l + ( y 1 ) 2 , - y 1 / , / ) ,

(4 9)

(4 10)

The system (4 4) becomes

(4 11)

and consequently

(4 12) 2 am u3 = a{ d2ux + dxu2 ) - 2 / u2 - 2 / u2

which on account of (4 7) and of the fact that ut are polynomials, implies u3 = 0 As we mentioned above, thisimplies locking in the context of Remark 3 5 Moreover, we may show that ux - u2 - 0 To this end, we note


M1 + dxu2 = 2a~ m u3 -f 2 a 1 y2 ux + 2 a 1 yl u

MEMBRANE LOCKING IN THE FINTTE ELEMENT COMPUTATION 141

that whatever the considérée triangle be, polynomials are analytic functions, and it is sufficient to check that theyvanish in a neighbourhood of the origin. Equation (4.12) with u3 = 0 and (4.11) becomes

(4.13) [l + ( / ) 2 + ( / ) 2 ] K ( y 2 ) + « î ( y l ) ] = 2 / « 1 ( y ? ) + 2y1ii2(y1)

where' dénotes differentiation. For y1 = 0 this gives

(4.14) [1 + ( ƒ )2] [u\(y2) + const.] = 2y2 u^y2)

and it is easily seen that ux is of the form

(4.15) M l = C [ l + ( / ) 2 ] .

Obviously, u2(yl) has an analogous expression and replacing them in (4.13) we see that ux and u2 vanish

identically. Finally we are in the context of Remark 3.4 and we have total locking. •

Case of the elliptic parafooloidThe calculations are essentially analogous to those of the previous case. We have:

(4.16) v(y\y2) = {y\y\\i(yl? + (y2)2] )

(4.17) ^ = (1 ,0 , / ) , a2=(0,l ,y2) , a3 = a 1 /2(- y\ - y\ 1 )

with

(4.18) fl=i + (yx)2 + ( / ) 2

(4.19) bn = b22 = cTl/\ bl2 = 0

(4.20) rln = rl

22 = a-ly\ r\x = r\2 = a l

y\ r«2 = o

and the system (4.4) becomes

{ 1 2 -,d,U, = d~W9 = y~ M1 + 2- M -{ — M

1 1 2 2 a 1 a 2 fll/2 3Ö2Mj + 0^2 = 0

Multiplying the first équation (4.21) by a, we get

(4.22) abxux = ad2u2 = ylux + y2u2 + amu3

which, on account of (4.18) implies u3 = 0 as the ut are polynomials. It is also possible in this case to provethat Mj = u2 = 0, but the proof is more involved than in the previous examples. We shall not give it hère, butin any case we have locking in the framework or Remark 3.5. •

Case of the straight helicoidIn that case, we have

(4.22) <?(y\y2) = <p(0, r) - (rcos 0, r sin 0, kO)

(aj = ( - r sin 0, r cos 0, k ) ,

a2= (cos0, sin 0,0) ,

a3 = a m(-k sin 0, k cos 0, - r)

vol. 32, n° 2, 1998

142 D CHOI, F J PALMA, E SANCHEZ PALENCIA, M A VILARINO

where

(4 24) a = r2 + k2

Moreover

(4 25) bn = b22 = 0, bn = ka m

(4 26) r i 2 = r ln = f \ 2 = Q, r 2

n = - r , r \ 2 = a x r

(dlul = — ru2

d2u2 = 0

d2ul 4- 8jM2 = 2 a~ ru1 + 2 a~ l ku3

and from the tmrd équation of (4 27) we get

(4 28) 2 kamu3 = a{ d2ux + dxu2) - 2 rux

It then follows from (4 24) that u3 = 0, and we have locking again in the framework of Remark 3 5The previous considérations show the very gênerai character of membrane lockmg in non-mhibited shells

Moreover, it appears that membrane locking is concerned with the spécifie local structure of the équations, withouttaking into account the boundary conditions, this situation is in contrast with the classical locking of H1 fimteelement approximations in incompressible fluid mechanics, where v e Gh implies v = 0 only on account ofboundary conditions and interaction between the various fimte éléments

Remark 4 1 Obviously a plate is a particular case of shell When shell fimte éléments are used in plates,membrane locking do not appear This is easily seen from system (4 2) in the case of a plate, b^ = 0, and (4 2)shows that the éléments of G are such that (vv v2) is a ngid displacement m the plane (usually vamshingaccording to the boundary conditions) whereas v3 is arbitrary Gk has an analogous définition, and Gk approachesG provided that the approximation of V by Vh is independent for the three components as obviously happens forany usual fimte element approximation •

5. LOCKING IN THE NON-CONFORMING DKT APPROXIMATION

In this section we prove that the previous considérations on membrane locking also hold true for the widelyused DKT (= "Discrete Kirchhoff Triangle") which is non-conforming (i e Vh is non containted in V) Thisapproximation is desenbed in detail in [4] chap II 2

In prmciple the DKT approximation involves displacements and rotations Nevertheless, there exists anequivalent formulation involving only displacements, after eliminating the rotations (cf [4] formula (2 25),p 118) In fact, this formulation is exactly analogous to (2 26) with a certain space Vh not contained in V, but suchthat v e Vh implies that vt is a polynomial in each firute element We then define Gh by

(5 1) G„ = { v e Vh, y a / v ) = 0 , « = 1 , 2 }

and we obtain the convergence (3 11) in Vh exactly as in theorem 3 2 Moreover, theorem 2 3 holds true, providedthat the strong convergence in V be replaced by convergence in a space contammg V and Vh, for instance(L2(Q) )3 Let us also remark that the convergence (2 27) also holds true m this space (cf [4] theorem 2 2 6,p 122), we recall in this concern that the hypothesis that the shell is clamped all along lts boundary may be

M2 AN Modélisation mathématique et Analyse numériqueMathematical Modelhng and Numencal Analysis

MEMBRANE LOCKING IN THE FINITE ELEMENT COMPILATION 143

weakened according the remarks of [4] p 120 The définitions of "Vh approaches" and "Gh approaches G" are

analogous to (2 24) and (2 25) but the norms are agam in ( L ( £? ) ) In these conditions, Proposition 3 3 and Remarks 3 4 and 3 5

hold true, as well as the considération of section 4 and the locking of the DKT approximation follows

6. NUMERICAL EXPERIMENTS

In this section we display some numencal results concerning locking phenomenon, we compare the perfor-mance of two finite element methods (conforming and non-conforming) and the behaviour with respect to differentreduced intégration schemes

We consider the hyperbolic paraboloid problem (see the fi g 1 for the physical data of the shell) The thicknessof the shell is 2 e = 10" ", with n = 1, 2, 3, 4, 5 , we remark that for n = 4, 5 we have really a "very" thm shell,perhaps without physical interprétation

Q= [0,3] x [0,1]

£=2,85*104

v = 03

f = 8 £ 3 a 3

Boundary conditions:u = 5on(0,0)(0,l)

u symmetrie on (0, 0) (3, 0)(corresponding to a physical problem

on [0,3] x [- 1,+ 1]with symmetrie loading)

Figure 1. — Hyperbolic paraboloid.

We consider m the followmg two different finite element methods• the Ganev-Argyns' method this is a well known conforming method which uses the P4 -Ganev triangle for

the approximation of the tangent components of the displacement, and the P5 -Argyns triangle for the normalcomponent, there are 15 + 15 + 21 degrees of freedom in each triangle,

• the Sander's delinquent method this is a non-conforming method which uses the P2 -Lagrange-edge trianglefor the tangent components and the Sander's delinquent triangle for the normal component now there are6 + 6 + 12 degrees of freedom in each triangleThe définition of these finite éléments is desenbed m figures 2 and 3 using the classic notation of [7]

For the approximation properties of the above-mentioned éléments, we refer to [4] (p 110 and 114) Hère, werecall the main results

• the Ganev-Argyns' method is a high-degree approximation because the order of convergence is O(h4)(h is the diameter of the triangulation), this order is preserved if a reduced intégration scheme exact for thepolynomials of degree 6 is used,

• the order of convergence of the Sander's methods is O( h2) and the intégration scheme have to be exact forthe polynomials of degree 4

vol 32, n° 2, 1998


p — p

P — Pr KA ~ r 5

P:

Ganev triangle

(K) (dimPKG= 15)

Argyris triangle

.P^V'.i

WO.

= 1,2,3}

Figure 2. — The Ganev-Argyris method.

, -Lagrange - edge triangle

K2 = P2(K) (dimPK^ = \

a , ) . f Pdy:i=h2,3

Sander's delinquent triangle

PKS = Ps(K^ © ( ^ l ^ ^ ^ U 1,2,3 (din

-: i= 1,2,

Figure 3. — The Sander's delinquent method

M2 AN Modélisation mathématique et Analyse numénqueMathematical Modelling and Numerical Analysis


For each method we have made use of 6 intégration schemes (see [12]):1. Scheme exact for P2 with 3 nodes;2. Scheme exact for P3 with 7 nodes;3. Scheme exact for P4 with 6 nodes;4. Scheme exact for P6 with 12 nodes;5. Scheme exact for P8 with 16 nodes;6. Scheme exact for P9 with 22 nodes;

Remark 6.1: Actually, we tested 13 reduced intégration schemes, but the results obtained were similar to theabove mentioned ones to which we reduce our présentation. •

Clearly a scheme is optimal when it assures the convergence of the finite element method with a minimalnumber of nodes; for the Ganev-Argyris' method the optimal scheme is 4 and for the Sander's method is 3. Wenote that for the low précision schemes the order of convergence of the method is not assured (it is the case of1 and 2 for the Sander's method and 2 and 3 for the Ganev-Argyris' method) or even the matrix of the discreteproblem may be singular (it is the case of 1 for Ganev-Argyris). These schemes may be considered as "reducedintégration schemes".

The meshes follow the topological disposition of figure 1. The step h is reduced in the usual way. Computationwere performed on a HP/Apollo computer using the MODULEF Code.

The results are displayed for the normal displacement w3 at the point A = <p(3, 1 ). The results obtained forthe different meshes, intégration schemes, thickness and finite éléments methods are summarized in the figure 4aand 4b for the value a = 0.2 of the parameter (see fig. 1).

Remark 6.2: The same problem was also considered for a different geometry, a = 0.1 (see fig. 7), but the maintrends of the numerical results are the same as for a = 0.2.

In figures 5c, d, e, f we show (in logarithmic scale) the relation between the thickness and w3( A ) for two valuesof h (attention to the v-scale).

F.E.M.

Ganev - ArgyrisSander

Coarse

h =h =

Figures 5c, d,

mesh

1/21/4

e, f.

Fine

h =h =

mesh

1/101/123

In figure 6(g-j) we show the convergence of the methods for two values of the thickness when dim Vh —> <*> ( w e

recall that dim Vh —> °° is equivalent to h —» 0).We recall that membrane locking amounts to non-uniformity (with respect to e ) of the convergence h \x 0.

Equivalently, as our example is one of total locking (see (3.11), Remark 3.4 and section 4) the limit of u^ ase \ 0 with fixed h is zero. All these features are apparent in the previous numerical results.

Our computations were roughly done up to the reasonable capability of an Apollo station. It then appears thatwith the Ganev-Argyris scheme locking is only evident with very small e : the lack of convergence only appearsin our example for s ^ 10" 3 (we recall that 2 e = thickness). On the contrary, for the Sander method the lackingis clear (attention must be payed to the v-scale in fig. 5). Even for s = 10~ the convergence is hard and fore = 10" 3 the results are completely different from the Ganev-Argyris ones.

The better behavior of the Ganev-Argyris scheme may be understood as a better ability of the high degreepolynomials for approximating the very peculiar functions of the limit space G. Equivalently, the rate ofconvergence in h (instead of h for the Sander scheme) shows the higher efïicacity of the / i \ 0 process onconvergence. See also [14] in this concern.

vol. 32, n° 2, 1998

146 D CHOI, F J PALMA, E. SANCHEZ-PALENCIA, M A VILARINO

IIII

90

11

^ j

il

II

CN

!|

||

r—H1

br—iIICO

ness

ick

entiTtT-H

ON\Oi-H

mooONen

en

ti00

|

Ttr—i

CDO

en

tiCMCNOr—i

VO0000T-H

oTt

en

tiSCM

00CMCMOTt

en

entiT-H

r—irH00rtCN

enTtCM

en

ti00oONenr—i"*$

en"TtCM

en

ti

inTten

9CN

en

tiT-HONT-H

minen

en

GEN

oNOHT-H

TtenTtCM

en

tiNO00eninON

entirH

oNOT 1

r-^

TtenTtCN

en

WNOeninON

en enenenCM CM

en en

ti00inr—i

CM

NO

en

.43

CM

en

tiCM

CMr-H

NOO

W

ST-HCN[*-•«.

NOen

.43

CM

en

tiCM

enoN0oen en

en en

CM

en

ti woCN\OONON

in

ONen

en

SoenONCMenenTt

TtCM

en

tiinONenONenO

enTt

CM CN

en en

titi0N""00inONenoCM"*

C/3

OG

enPH

for

act

ON

TtCMi-H

enN0mTtCM

E/}

OGNO

Tt

OH

for

oXW

WooCN

3rHCNTtCM

cSoG

entiON

ONmT-HT-H

TtenTtCM

en

tiTtenenmONenenCM

en

wONsCM

en

.43

CM

en

ti1Soen

en en

CM

en

tiS00enenOrH

enTtCM

en

ti

''tCM

en

ti00

oenenenOrH

enTtCM

en

tiin ooNO enenNOCNNOOCMtCM

CA

OG

CM NOT H

NO

fi

for

act

XW

rH

00

for

act

X

enenNOOCMtCN

ce

OGCNCM

ON

ri

for

act

XW W

II

II

||

]|

||

CM•^

||

il

"

,4*J

cNi

brîII

ickness

cj

E"1

entienO

entit>

V3ON t>CMenNO00Tf

inONNO00rH

en en

tiONr-r—ir—i

inTtenOin

W

SinCN

entien•<^t

1—(

ooNO00T-H

enti

\oTtT-H

00NO00i-H

en en

tiWen ooenrHr—i

OON C--SO00f—(

m en

titir—i

NO

SNO NOin ONON NOCM

»n

en

tioCMin

ONNO

00rH

en

tio00CM

vo00T—i

en

tiCMONr-00ON

ON00

8NO00

enWCNON00f-»*ON

OOrH

enti

en00 TTO inNO

00 00in T-H

en en

ti00

CM

ONt-H

in

00

en

tiCMNOr—(T-H

Ot

TtCM

1OG

enfi

for

act

XW

WCMinCMrHCMOCMON

en

00

00r-H

en

enti

\oT (

ooNOOOrH

en

min00

£3

NO00i-H

en

tiONONrONNO00T-H

entitiVOON

Tt

00t-H

OTt

00 OOin in

00 00rH

entiCMCM

omNOO

^rtitiCM ON00 Tt

inON00oorH

rH

SoGNO

Tt

fi

for

act

Xm

OrH

oenCMON

-gGCNrH

NO

fi

for

act

XW

r—i

enti

645

TtTt

O

tiONT—i[-«.

00r—(

entiTtinNOTtCN

Ot>rH

47E-

inin CMinoONrH

ON

un

*OGNOT—i

00

fi

for

act

[T]

rHCMON00t-H

ON

S"oGCNCM

ON

PH

for

act

X

Os

^S

11

II^

entitiCMCMrH

TteninO

en

tiT-H

°p in

II

VU

ri)||

CM\

||

*">||

en

brHII

ickness

rH

<tenT-Ho00OrH

en

tiNO

ONNOT-H

OenHr-i

en

tiONT-H

NO

00

TtCM

enti

CMenCMCMOT-H

en

ti00CMONON

§CM

en

"oGt-

enfi

for

act

Xm

oON00

ON

inON00

2CM00

o0000

T-H

NOON

oo

S00r-*

NOON00

tioHt>

NONOONOO00

tiCMenO

00enONNOrH

oo

tiT-H

ooTtenT f

rHT—i

inm

Tt

WTt

tiON ONON 000000r—i

CMenrHTt

ON ON

oo oo

tirH<<^tCM

en00en

ti00

orHen

wSONenr—Ï

ON00

ti00NOoen

00 00enON ON

00

tiTt

ONenenCM

enONOO 00

tiTt

CM00• t

CM

00 00

tiCM

en

P-lrHON

pqsoCM00mONCM

00

Ti-

WON

ONNO\O ON

ONNO

NO00ON

00 00

timm00CMrH

NO

NONO

Tt

tiin00mOrH

ON

CM

«D

GVO

Tt

fi

for

oX

GCMr—i

NO

fi

for

act

XW

WCNT-H

CMON00ino00N0N0

W

068

SCM

CMNOCM

S'SG

1—f

00

fi

for

oXW

ON

ONNO00

W

T-H

3enr**N0NO

SenNOinONONenrHNOCM

c«

'SGCMCM

ON

&

for

act

xm

«\U

II

op* H||

i S

||

-s;

M

tN

||

II

l

br—iIIco

ickness

f-H

^* Tt

WWCM CM

Tt

WCNmtioooo

Tt

tiCMT-H

00 \D 00 O ONm in en en ooON NOen t—*en ON

SrHenOCNm00f-H

m oo ON ON ONin r-ON t^

ttW W

N0

CM en r-m ONen ooo ooo rÔN C^

ON C^

ON r-

titiNO t-ON enen OCM OON mON inNO NO

r—i £-»

en Tt

titiT-H VO

in ND

Tt '~HrH ONCM «n

en Tt

titiO 00r- orH OO rHCM inTt rH00 rHrH CM

TT

CMON0

NOt>

siCMONr-H

NO

Tt

tienCMTt

en ooenCMen

en enm

tioT-HTtmmi>

oT-H

Tt

rit-H

ON

SCNin

in

ti

in

tirHON0< -inrHON

.08

ti

00

ON

CMenin

Tt

tiinTfr—if"-*

p;

.08

ttienenos

NO mCM ONON 00in

Zti

f—1 T-HON CMênenO

ON 00r—i00

TtC^ 00 ON ONr—( en

en Tt

titim enen Ttoo enen NO

in

wCMinin

in

rH

m

ti

inin

tWoenONT-H

tÈ>TtONi-H

in

tiCM rH

en CMt-CMT-H

00 00 SO N000 00 r-r-O ON N0 ONTt CMC M en

<D <o

G Gr^ NO

en Tt

BkPÛ

ëëO ÜCv CZ

X XW W

NOi-H

1OG

TtrH

OO

OG

CM OH

vo

fi

for

act

Xm

T 1

oofi

for

act

XW

ONCMCMN000TtrH

1OGCMCM

ON

fi

for

act

XW

..n^

||

||

||

CM•3||

•JE;

II

«ni

bT-HII

ickness

[S

TtWa\NO

CMOCM

TtON

TT

tiCMenoONr—i00inSO00

Tt

tiON

Tt

ti• « ^

TT

SON enH

Tt

tiCMenen \D

O CM en00

oen8HON 00m T-H

enCMeneni-H

VO VO NO

tim00 O00 00

ooooenrHONON

ON in

en Tt

tiONCMN0

g

tir-i

OoONenen NO

CM t>

Si-H

en

ti00ON

NOenrHCMrH

en

tiin

r*»00enT-HTt00T-H

T-H

en

ti8oON

.77

tiON

NO

00CM

tmsomenenmTt

tir—!

WrHr—i

in00r-CMON

\OCMr-iT-H

enT-HVO

tiSenOin

in «nCMin

ti

CMin

Tt

tiNO m

ON O00>nmCMCM

.83

rH

Tt

.73

en en

w mr- ONTt inCMOT-HON in

enc*-CM

tivomNO0000NONONO00Tt

Tt

tir—i

enH00i-H

in

tiinoT-H

ooOenH

tiCN

NO NOr-HrH

CM00

ON N000 NOOTtCM

1OG

enfi

for

act

enTtrH

Q

1G\Q

Tt

fi

for

act

XW

CMNOCMCM

soGCMrH' w

NO

fi

for

act

XW

m

r—i

N0CMenCMON

.72

en

Tt

tio00enON ON

Tt rHTt<n eninr*-i-H

in

tiCMONinTt

CMCM

VO

tii-H

in

rH

in

tiT-H

en

enin00T-H

CMT-H

NO

tisO O NCMONin

£"».00CM

oo r»00 00T-H i—i

CN CM

1GN0T-H

'NW

00

10GCM' W

ON

fi PH

a

act

XW

for

act

M2 AN Modélisation mathématique et Analyse numénqueMathematical Modellmg and Numerical Analysis

Thickness EExact for P-2Exact for P-3Exact for P-4Exact for P-6Exact for P-8Exact for P-9

Thickness eExact for P-2Exact for P-3Exact for P-4Exact for P-6Exact for P-8Exact for P-9




= KT-1( 3 nodes)( 7 nodes)( 6 nodes)(12 nodes)(16 nodes)(22 nodes)

- l(T-2( 3 nodes)( 7 nodes)( 6 nodes)(12 nodes)(16 nodes)(22 nodes)

= l(T-3( 3 nodes)( 7 nodes)( 6 nodes)(12 nodes)(16 nodes)(22 nodes)

10A-4( 3 nodes)( 7 nodes)( 6 nodes)(12 nodes)(16 nodes)(22 nodes)

= ior-5( 3 nodes)( 7 nodes)( 6 nodes)(12 nodes)(16 nodes)(22 nodes)

h = 1/22.604991677E-32.306362717E-32.337713829E-32.337514121E-32.337514936E-32.337514275E-3

h = 1/24.006403018E-42.557792631E-42.999151400E-42.946492212E-42.946732281E-42.946724486E-4

h = 1/23.425484973E-57.236455110E-69.991200830E-67.104299501E-67.063661296E-67.063132749E-6

h = 1/22.944222034E-53.108962517E-6L716207731E-61.013441921E-79.631274995E-89.624801170E-8

h = 1/22.939394693E-53.066515034E-61.053361544E-61.022903144E-99.70062571E-109.693837900E-10

h = 1/42.488607496E-32.425501070E-32.428090433E-32.428085197E-32.428085213E-32.428085073E-3

h = 1/41.034478714E-38.552603124E-49.208149792E-49.202305087E-49.202313689E-49.202308294E-4

h = 1/46.535143521E-54.142733068E-55.197426694E-54.915126875E-54.916514927E-54.916494355E-5

h = 1/41.518837151E-51.911301184E-63.017522135E-61.008553067E-69.834400428E-79.833584537E-7

h = 1/41.465421827E-51.402741081E-67.307525941E-71.125287218E-81.078406109E-81.078261173E-8

h = 1/62.455951795E-32.432592325E-32.433187100E-32.433186366E-32.433186368E-32.433186309E-3

h = 1/61.165957105E-31.094175541E-31.117549289E-31.117480942E-31.117481054E-31.117480847E-3

h = 1/61.935798087E-41.328236774E-41.570063645E-41.540205174E-41.540368915E-41.540365947E-4

h = 1/61.349536986E-53.061244999E-65.362611729E-63.596620896E-63.576981011E-63.576930653E-6

h = 1/61.112837818E-58.127623607E-71.255049310E-64.931667438E-84.742161118E-84.741882419E-8

h = 1/82.445000073E-32.433789026E-32.434020027E-32.434019809E-32.434019807E-32.434019776E-3

h = 1/É1.189041148E-31.155219754E-3L163683370E-31.163669553E-31.163669586E-3L163669497E-3

h = 1/$3.855129820E-42.740887643E-43.200204855E-43.180255053E-43.180350765E-43.180346490E-4

h = 1/81.653739757E-56.711542106E-6L034847966E-58.667831467E-68.666838135E-68.666794753E-6

h = 1/39.622897451E-65.751547558E-7L665202089E-61.426724375E-71.376452068E-71.376404249E-7

h = 1/102.440402073E-32.434127369E-32.434249062E-32.434248964E-32.434248967E-32.434248936E-3

h = 1/101.192921624E-31.173755289E-3L177384480E-31.177380464E-31.177380474E-31.177380428E-3

h = 1/105.641766002E-44.350489440E-44.934630723E-44.925310702E-44.925346880E-44.925342162E-4

h = 1/102.461805902E-51.363281000E-51.896680004E-51.720887586E-51.721663575E-51.721657990E-5

h = 1/108.854582211E-65.243351400E-71.977574500E-63.123004157E-73.131926566E-73.131852676E-7

h = 1/122.438144762E-32.434259878E-32.434336964E-32.434336902E-32.434336908E-32.434336900E-3

h = 1/121.192815721E-31.180648349E-31.182439635E-31.182438140E-31.182438146E-31.182438121E-3

h = 1/126.893603019E-45.781561943E-46.327551341E-46.323765285E-46.323775751E-46.323771854E-4

h = 1/123.891121794E-52.478821174E-53.205862858E-53.014994335E-53.016118686E-53.016112046E-5

h = 1/128.476264421E-66.128394617E-72.286600818E-66.199536193E-76.047369685E-76.047266524E-7

Figure 4b. — P2-Lagrange edge/Sander.


10 ' 2

10-%

10-%

10-5_

u_3CA3

/

10~6 10~

1

f"1 ' ' L 3 ' V '

1 »h

1 lh l l( j • ' '""

c k n e s s

EXACT P_3

EXACT P_«

E^CT P_6

EXACT P_8

EXACT P_9

Figure 5c. — Ganev-Argyris, h = 1/2.

10-a | u . e ( A )

10'3J

10

10"6 10~5 10' 4 10"3 10"3 10"1 100

t h i c k n e s s

O-e-0

EXACT

EXACT

EXACT

EXACT

EXACT

_3i 4

'_*

' - 81 9

7 NODES

e i

12

16

22

4ODES

NODES

NOOES

NODES

Figure 5d. — Ganev-Argyris, A = 1/10.

M2 AN Modélisation mathématique et Analyse numénqueMathematical Modelling and Numencal Analysis

MEMBRANE LOCKÏNG IN THE FINITE ELEMENT COMPUTATION 149

10-a 1 u_3CA)

10'

10 10 10 10 10 10 10{ h i c h e s s

EXACT I

EXACT I

EXACT I

EXACT I

EXACT I

EXACT 1

Jt, 3 NODES

_3 7 NODES

_4, 6 NODES

'„6, 12 NODES

>_B. 16 NODES

' 9 . 22 NODES

Figure 5e. — Sander, h = 1/4.

10-2 u _3C AD

10 _

10-5J

10-GJ

10"I 1 I I l l l l | 1 I I ITII I j 1 I 1 11jll{ 1 I TTTTTI| 1 I I Illll| 1 I M Mil

10"G 10" 5 10~A 10~ 3 10" 2 10" 1 10'

i h i c k n e s s [

Geo

EXACT P_2,

EXACT P_3,

EXACT P_4,

EXACT P_6,

EXACT P_8,

EXACT P J ,

) MODES

J MODES

5 NODES

L2 NODES

L6 NOOBS

22 MODES

Figure 5f. — Sander, h = 1/12.

vol. 32, n° 2, 1998

150 D. CHOI, F. J. PALMA, E. SANCHEZ-PALENCIA, M A. VILARINO

0. 005

0. 004

0. 003 _

0. 002 _

0. 001

\V

0

u _3C A3

1 '3000 G000 9000

di m V_h

G-e-e

EXACT P_3

EXACT P_d

EXACT P_6

EXACT P_8

EXACT P 9

, 7 NODES

6 NODES

, 12 NODES

, 16 NODES

. 22 NODES

Figure 6g. — Ganev-Argyris, £ = 10

0. 0025

0. 0020 _

0. 0015 _

0. 0010 _

0. 0005 _

0 0000 _

u_3CA) |

t

Vf0 3000

1 h

6000 9000

| d. m V_h

EXACT

EXACT

Q Q Q EXACT

EXACT

EXACT

P_3 ? MOOBS

P_4. 6 NODBS

P_6. 12 NODBS

P_S 16 HODES

P_9. 22 HODES

Figure 6h. — Ganev-Argyris, e = 10 .


MEMBRANE LOCKEMG IN THE FTNÏTE ELEMENT COMPUTATION 151

0 005

0 004 __

0 003 _

0 002 _

0 001 _

u_3CA)

0 3000• i

G000 9000

. m V_h

EXACT P_9 22 HODES

Figure 6i. — Sander, e = 10

0 00004

0 00003 _

0 00002 _

0 00001 _

0 00000

3000I

G000 3000

di m V_h

EXACT I

EXACT I

EXACT I

EXACT I

EXACT I

EXACT I

_ 2 3

_ 3 7

_ 4 6

>_6 1

' 8 1

>_9 2

.6 NODES

!2 HODES

Figure 6j. — Sander, e = 10

vol 32, n° 2, 1998


We also point out that the différences due to the reduced intégration schemes are insignificant and only fors "very small" we fïnd some différences; consequently we conclude that the membrane locking is not eliminatedwhen we use a reduced intégration scheme (we observe that some low précision schemes eliminate the lockingbut the corresponding results are wrong since the convergence is not reached, see the scheme exact for P3 inGanev- Argyris).

REFERENCES

[1] J.L. AKIAN et E. SANCHEZ-PALENCIA, "Approximation des coques élastiques minces par facettes palnes. Phénomène deblocage membranaire", Compt Rend. Acad. Sci. Paris, 315, p. 363-369 (1992).

[2] I. BABUSKA and M. SURI, "On locking and robustness in the finite element method", SIAM Jour. Numer. Anal, 29,

p. 1261-1293 (1992).

[3] T. BELYTSCHKO, H. STOLARSKI, W.K. LIU, N. CARPENTER and J.S.J. ONG, "Stress projection for membrane and shear

locking in shell finite éléments", Comp. Meth. Appl Mech. Engng., 51, p. 221-258 (1985).

[4] M. BERNADOU, "Méthodes d'éléments finis pour les problèmes de coques minces", Masson, Paris (1994).

[5] M. BERNADOU and RG. ClARLET, "Sur l'ellipticité du modèle linéaire de coques de W.T. Koiter" in Computing methodsin Sciences and Engineering, R. Glowinski, J.L. Lions éd., p. 89-136, Lecture Notes in Econom. and Math. Systems,vol. 134 (1976).

[6] D. CHENAIS, J. C. PAUMIER, "On the locking phenomenon for a class of elliptic problems", Num. Math., 67, p. 427-440

(1994).

[7] RG. ClARLET, "The finite element method for elliptic problems", North-Holland, Amsterdam (1978).

[8] G. GEYMONAT and E. SANCHEZ-PALENCIA, "On the rigidity of certain surfaces with edges and application to shell

theory", Arch. Rat Mech. Anal, 129, p. 11-45 (1995).

[9] A. HABBAL and D. CHENAIS, "Détérioration of a finite element method for arch structures when thickness goes to zero",

Num. Math., 62, p. 321-341 (1992).

[10] A. KALMOULAKOS, "A Catenoidal patch test for the inextensional bending of thin shell finite éléments", Comput. Meth.

Appl. Mech. Engng., 92, p. 1-32 (1991).

[11] F. KJKUCHI, "Accuracy for some finite element models for arch problems", Comput. Meth. Appl. Mech. Engng., 35,p. 315-345 (1982).

[12] J.N. Lyess and D. Jespersen, "Moderate degree symmetrie quadrature rules for the triangle", Jour. Inst. Math. Appl, 15,p. 19-32 (1975).

[13] F. NlORDSON, "Shell Theory", North-Holland, Amsterdam (1985).

[14] J. PlTKARANTA, "The problem of membrane locking in finite element analysis of cylindrical shells", Num. Math., 61,

p. 523-542 (1992).

[15] E. SANCHEZ-PALENCIA, "Statique et dynamique des coques minces. I: Cas de flexion pure non inhibée", Compt Rend.

Acad. Sci. Paris, sériel, 309, p. 411-417 (1989).

[16] E. SANCHEZ-PALENCIA, "Statique et dynamique des coques minces. II: Cas de flexion pure inhibée, approximation

membranaire", Compt. Rend. Acad. Sci. Paris, sériel, 309, p. 531-537 (1989).

[17] E. SANCHEZ-PALENCIA, "Asymptotic and spectral properties of a class of singular-stiff problems", J. Math. Pures Appl,

71, p. 379-406 (1992).

[18] E. SANCHEZ-PALENCIA, "On the membrane approximation for thin elastic shells in the hyperbolic case", Rev. Mat Uni.

Comp. Mad, 6, p. 311-321 (1993).

[19] E. SANCHEZ-PALENCIA, "Surfaces et coques élastiques minces: problèmes et défis", La Vie des Sciences, 12, p. 239-258

(1995).

[20] L. SCHWARTZ, "Cours d'Analyse, vol T\ Hermann, Paris (1981).


f. j. p e. s -p m. a. vilariÑoarchive.numdam.org/article/m2an_1998__32_2_131_0.pdfmembrane locking...

Documents