Professional Documents
Culture Documents
Numerical analysis of junctions between thin shells, Part 2 : Approximation by nite element methods
Michel Bernadou , Annie Cubier
N 2922
Juin 1996
` THEME 4
apport de recherche
ISSN 0249-6399
Numerical analysis of junctions between thin shells, Part 2 : Approximation by nite element methods
Michel Bernadou*, Annie Cubier**
Theme 4 | Simulation et optimisation de systemes complexes Projet MODULEF Rapport de recherche n 2922 | Juin 1996 | 36 pages
between thin shells. We describe the approximation by a \pseudo-conforming" nite element method associated with the Argyris triangle and taking into account the numerical integration. Under suitable hypothesis on the integration schemes and on the data, we prove the convergence of this method and we derive a priori error estimates. Key-words: Thin shells. Elastic junction. Rigid junction. Argyris Triangle. Numerical integration. Error estimate. Numerical simulations
Abstract: The purpose of this work is to study the numerical analysis of junctions
(Resume : tsvp)
Analyse numerique de jonctions de coques minces, Partie 2 : Approximations par des methodes d'elements nis
tions de coques minces. Nous decrivons une methode d'approximation "pseudoconforme" utilisant l'element ni d'Argyris et prenant en compte les phenomenes lies a l'integration numerique. Sous des hypotheses convenables sur les schemas d'integration numeriques et sur les donnees, nous montrons la convergence de cette methode et nous donnons des estimations d'erreur a priori tant pour le probleme de jonction elastique que pour le probleme de jonction rigide. Mots-cle : Coques minces. Jonction elastique. Jonction rigide. Triangle d'Argyris. Integration numerique. Estimations d'erreur. Simulations numeriques.
1 Introduction
The numerical analysis of junctions between thin plates was studied by BernadouFayolle-Lene (1989) while Fayolle (1987) thoroughly described the corresponding approximation by conforming nite element methods. Such an approximation is much more complicated in case of junctions between two general thin shells. Indeed the transmission conditions which appear in the de nition of the admissible spaces cannot be exactly satis ed in the associated discrete spaces. Thus, we consider pseudo-conforming nite element methods which are conforming everywhere except along the junction. In the rst part of this work (Bernadou-Cubier, to appear), we analyse the continuous problems of junctions between two thin shells associated with an elastic or a rigid behaviour of the hinge. We start by giving the equilibrium equations of these problems and the corresponding variational formulations. We study the numerical properties of theses equations and show the existence and uniqueness of the solution. We also prove that the solution of the elastic junction problem converges to the solution of the rigid junction problem when the coe cient of elastic sti ness of the hinge becomes very large. In this second part of the work, we start in Section 2 by rewriting the continuous problems in terms of matrices and vectors, which are well adapted to the approximation by nite element methods. In Section 3, we build the discrete spaces which are associated with the Argyris triangle, and which contain the discrete junction conditions. The discretization of the transmission conditions is based on the results obtained by Zenisek (1981) for nonhomogeneous boundary conditions. From Section 4, we restrict our attention to the elastic junction problem ; we give the main results concerning the rigid junction problem in Section 6. Then, we state the rst discrete problem taking into account the nite element approximation and the nonconformity of the method along the hinge. We introduce in this problem an additional linear form which is void for continuous problem and which takes into account the nonconformity of the method. Then we prove the existence and uniqueness of the solution. This result is based on the transmission of a clamped condition by the discrete junction conditions described in Section 3. Afterwards, we give an abstract error estimate which is reduced to the usual interpolation error ; this is a direct consequence of the de nition of the discrete problem by using an additional linear form as mentionned before. In Section 5, we study the second discrete problem which takes into account the additional e ect of the numerical integration. We prove the uniform ellipticity of the bilinear forms and thus the existence and uniqueness of the solution of this problem. Next, we give criteria on the integration schemes and required regularity conditions on the data so that the nite element method converges. These conditions lead to asymptotic error estimates of the same order than the interpolation error. These results are based on local error estimates studied by Bernadou (1996) and Cubier (1994). In Section 7, we illustrate previous results by some numerical tests on rigid and elastic junctions between a cylinder and a spherical end cap. In order to validate our results, we consider a corresponding
RR n 2922
test using the three dimensional elasticity model instead of a shell model and which represents a good approximation of a rigid junction. The result of this test is very closed to ours obtained for rigid junction. Moreover, we nd again numerically that the elastic junction becomes almost rigid when the coe cient of elastic sti ness of the hinge becomes very large. Notations and references : In this second part, we make many references to the notations and to the results of Part 1, just by adding "Part 1" before each reference.
a (u ; u ); (v ; v )] =
where the column matrix T V] (2.2) 1 12 = v1 v1;1 v1;2 v2 v2;1 v2;2 v3 v3;1 v3;2 v3;11 v3;12 v3;22]; and where the symmetrical 12 12 matrix A] (respectively A]) depends only on the shell thickness e, on the mechanical characteristics of the shell and on the rst, second and third partial derivatives of the application : ! S (resp. : ! S) which maps (resp. ) onto the middle surface S (resp. S). Subsequently, we assume that 2 (C 3( ))3 and 2 (C 3 ( ))3. The second bilinear form b :; :] which appears in the variational formulation of elastic junction between shells (Part 1, (3.10)) is associated with the strain energy of the hinge. We de ne a 24 24 matrix C ] which only depends on the geometry of the hinge, and a column vector V V ]24 1 which collects the vectors V ] and V ] de ned by relation (2.2). Thus, we have :
(2.1)
b (u ; u ); (v ; v )] =
Moreover, we introduce a new parameterization of the hinge ?, as the image of a one-dimensional interval ! =]0; 1 by a mapping , i.e., : ! ! ?. Now, we can substitute this application into relation (2.3) to obtain
U U ] C] V V ]ds:
(2.3)
b (u ; u ); (v ; v )] =
where the underlined quantities are obtained by composition with the mapping and are de ned on the interval ! . The element d! is associated to the line element 3 X ds along the hinge ? through the mapping = i (t)ei :
U U ] C] V V ]d!;
(2.4)
i=1
i1=2
;
INRIA
5
i1=2
so that
d! = ( 01(t))2 + ( 02 (t))2 + ( 03(t))2 dt The linear form ` :] which represents the work of the external loads can be written
T T
P ] V ]d
1d 2 +
Z
T
Ls] V ]d
R
1
Ls ] V ]d ;
(2.5) (2.6)
where and
T
P ]1
12 =
pa p1 0 0 p2 0 0 p3 0 0 0 0 0]
0 0 N 2 + b2 M 0 0
9 > = > ;
Ls ] 1
1 1 12 = a (g )0(g )0 N + b M
0 0 0]: In the above equations, pi and N i denote respectively the covariant components of the body force resultant, of the resultant and of the resultant moment of the surface loads while b and a are the second fundamental form and the determinant of the rst fundamental form ; all these quantities are refered to the middle surface S and, by de nition, M = M a a3 . The vectors P ]; L s ] are obtained by analogy.
N3
M1
M2
(2.7)
hK K (3.1) and 8 K 2 Th ; K ; 8 K 2 Th; hK where hK = diam(K ); hK = diam(K ), K = supfdiam(S ); S is a ball contained in K g and K = supfdiam(S); S is a ball contained in K g. ii) Let h be a real number de ned by (3.2) h = supfK max h ; max h ; max h g; 2T K K 2T K K 2T K where hK = diam(K ). Then, we assume that h ! 0: (3.3)
h h h
RR n 2922
?0
?1 ? ?1
?1
e3 e1 e2
2 0 1
a0 a1 a2 a3 1 !
a0 a1 a2 a3 1 1
s0 = 0 s1
s2
s3
=1
and Xh constructed from the Argyris triangle (Argyris-Fried-Scharpf (1968)), whose de nition is recalled in Figure 2, and the spaces Vh and Vh :
Fig. 1: Discretization of the hinge images With the triangulations Th and Th , we associate the nite element spaces Xh
Vh = Vh1 Vh1 Vh2 ; Vh = (Xh)3
where
Vh V and Vh V:
(3.4)
INRIA
a1 c3 b2 c2
b3
a2
b1
c1
a3
PK = P5(K ) ; dim PK = 21
K = fp(ai ); Dp(ai )(ai?1 ? ai ); Dp(ai )(ai+1 ? ai ); 1 i 3;
In order to build the discrete admissible spaces, we have to discretize the junction conditions ( equality of the displacements and equality of the tangential components of the rotations along the hinge), i.e., we have to express these conditions in terms of the degrees of freedom. This is a delicate step in the approximation of the continuous problems and it leads to the non-conformity of the method for the approximation of the transmission conditions along the hinge. The equality of the displacements along the hinge is a condition which appears for the elastic or the rigid junction problems as well. Thus we begin by studying the discretization of this condition.
3.1 The discrete admissible space for the elastic junction problem u(
First, let us recall the condition of continuity of the displacement along the hinge (Part 1, (2.31)1) : ) = u ( ); 8 2 ; 8 2 such that ( ) = ( ): (3.5) The equation (3.5) is vectorial. For its discretization, we have to use components of displacements. The vectors u and u are expressed upon the contravariant bases fa1; a2; a3g and fa1; a2; a3g which di er along the hinge. Therefore we have to write relation (3.5) upon one of these bases, for example fai g :
u i ( ) = Aj i ( ; )uj ( );
where Aj i ( ; ) = ai( )
aj (
).
RR n 2922
In addition, we introduce the two mappings F : ! = 0; 1] ! and F : ! ! which are assumed to be regular. The closed interval ! is subdivided into n+1 segments sp ; sp+1]; for p = 0; ::; n with s0 = 0 and sn+1 = 1. Thus F (sp ) = ap and F (sp) = a p where fapg; fa pg are the vertices of the triangles of Th and Th located along and . By analogy with Zenisek (1981) who considered the approximation of nonhomogeneous boundary conditions, the approximation through Argyris triangle leads naturally to impose the following conditions :
F (s` );
d Aj u F (s ); (s` ) = ds ` i hj 2 j j for ` = 0; ::; n+1 and where we have set for clarity Aj i (s` ) = Ai (a` ; a `) = Ai (F (s` ); F (s` )).
Remark 3.1.1 : Components uhi and u hj are piecewise two dimensional ve degree polynomials. Since we have supposed that and are rectilinear, the mappings F and F are a ne. Thus, the composed mappings uhi F and u hi F are piecewise one dimensional ve degree polynomials.
9 > > > > > > > > = > > > > > > > > ;
(3.6)
Now, we have to rewrite (3.6) in terms of degrees of freedom of Argyris triangle and thus to express these conditions on the reference domains and . There is no problem for (3.6)1 which can be directly written on the boundaries or . For (3.6)2 we use the following equation :
d ds (uhi F ) (s`) = Duhi (F (s` )) DF (s` ) A unit tangent vector to is given by : j DF (s` ) j (s` ) = DF (s` ) so that by setting uhi; (a` ) = Duhi (a` ) , we obtain d (u F ) (s ) =j DF (s ) j u (a ): (3.7) ` ` hi; ` ds hi
With similar arguments and since F is a ne, we have:
d2 (u F ) (s ) =j DF (s ) j2 u (a ): ` ` hi; ` ds2 hi
(3.8)
INRIA
Thus, the discrete junction conditions for elastic problem are obtained for ` = 0; :::; n + 1 by substituting (3.7) and (3.8) into relation (3.6) :
u hi(a `)
u hi; (a `) = u hi; (a `) =
(3.9) Thus, the admissible discrete displacement space for the elastic junction problem is de ned by :
9 > > > > > > > > > > > > > > > = > > > > > > > > > > > > > > > ;
Whel = (vh ; v h ) 2 Vh Vh; such that relations (3.9) are veri ed at the
corresponding vertices fa` g and fa ` g located on and
o
(3.10)
We proceed by similarity for the rigid junction problem which amounts to discretize the second condition (Part 1, (2.30)2) related to the equality of the rotations. The approximation through Argyris triangle leads to impose the following conditions : 9 > n (uh3; + b uh )] F (s` ) = (t t )n (u h3; + b u h )] F (s` ) > >
3.2 The discrete admissible space for the rigid junction problem
d f n (u + b u )] F g(s ) = d f (t t )n (u + b u )] F g(s ) h3; h3; h ` h ` ds ds (3.11) for ` = 0; ::; n +1 and | = 0; :::; n; where n = n a is the outward unit normal vector to the junction ? in the tangent plane to S and q| is the midpoint of s` ; s`+1 ].
In order to obtain normal and tangential derivatives, we use the relation
> > > > = > > > > > > > ;
RR n 2922
10
where and are respectively the unit tangent vector and the outward unit normal vector to the triangle of Th which has a side on . Relations (3.11) give
9 > > > > > > > > u h )(b |) > > > > > > > > = > > > > > > > > > > > > > > > > ;
(3.12) for ` = 0; ::; n + 1 and | = 0; ::; n ; are the components of the unit tangent vector used in (3.7) and b| is the midpoint of a` ; a`+1 ]. Thus, the admissible discrete displacement space for the rigid junction problem is de ned by :
9 > = > ;
Whrig = f(vh ; v h ) 2 Vh Vh; such that relations (3.9) and (3.12) are veri ed
at the corresponding vertices fa` g; fa `g; fb`g and fb `g located on and g (3.13) Remark 3.2.1. This discretization of junction conditions involves the nonconformity of the approximation, i.e.,
INRIA
11
9 > > > > > > = > > > > > > ;
5.
k Find (u k h ; u h ) 2 Whel such that k k k a (u k h ; u h ); (vh ; v h )] + kb (uh ; u h ); (vh ; v h )] = ` (vh ; v h )] + f (vh ; v h )];
F, v
k h
=v
k h
By comparison with (Part 1, (3.10)) we have introduced a new linear form f :] which takes into account the non-conformity of the approximation along the hinge, i.e.,
f (v ; v )] = fN v ? N v gd? = N vd +
?
N vd ;
where N and N are the resultants of the surface load. This form is identically zero when (v ; v ) 2 Wel while it is generally di erent from zero when applied to elements (vh ; v h ) 2 Whel . In that case, we rewrite in matrix form
f (vh ; v h )] =
where the column vectors
T
N ] V h]d
N ] V h]d ;
(4.2)
(and a similar expression for N ]). The introduction of this linear form f :] in (4.1) leads to a simpli cation in the abstract error estimate ( see Paragraph 4.3).
N ]1
V h ] and V h ] are de
12
= N 1 0 0 N 2 0 0 N 3 0 0 0 0 0]
In this paragraph, we prove the existence and uniqueness of a solution for problem (4.1). That leads to show the uniform Whel -ellipticity with respect to h of the bilinear form a :; :] + kb :; :]. First, let us recall some de nitions introduced in Part 1. Let space E be
E = (H 1( ))2 H 2( ) (H 1( ))2 H 2( )
equipped with the norm
2 2 2 2 2 1=2 k(v ; v )kE = fkv1k2 1; + kv2k1; + kv3 k2; + kv 1k1; + kv 2k1; + kv 3k2; g : The space Whel , de ned in (3.10), is a linear subspace of E and the above mapping is a norm on Whel .
RR n 2922
12
Proof :
: k(vh ; v h)kW = 0 ) (vh ; v h ) = (0; 0) in The assumption k(vh ; v h )kW = 0 immediately involves : Z e2 (v ) (v )]pad 1 d 2 = 0 so that the bouni) eE (vh ) (vh )+ 12 h h dary condition on 0 gives vh = 0 in ;
hel hel
ve degree polynomials upon each triangle side located on , relations (4.3) imply trv hi = 0 on . Moreover, (Part 1, (2.15), (3.12)) and b (vh ; v h ); (vh ; v h )] = 0 leads to v h3;n = 0 on ? where n = n a is the unit outward normal vector to ? located in the tangent plane to S. Here, our purpose is to obtain clamped condition on , i.e., v h3; = 0 on . We point out that v h3;n and v h3; are not the same quantities. Indeed, n is de ned along ? whereas is de ned along . The relation between these two quantities is : v h3;n = n v h3; = n ( v h3; + v h3; ) (4.4) Moreover, note that v h3 = 0 on implies v h3; = 0 on . Thus, by using in addition relation (4.4) and the assumption v h3;n = 0 on ? and noticing that the quantity n is di erent from zero, we obtain the required clamped condition v h3; = 0 on . Then, we obtain v h = 0 in .
v hi(a ` ) = v hi; (a ` ) = v hi; (a ` ) = 0 (4.3) for ` = 0; :::; n+1; where fa ` g is the set of vertices located on : Since v hi are
Lemma 4.2.2 : Upon the space Whel, the norms k(v ; v )kE and k(v ; v )kW are
uniformly equivalent with respect to h.
hel
Proof :
The proof of this lemma is based on the same arguments to those of Part 1, Theorem 3.1.1. The uniform-ellipticity comes from the inclusions (3.4) which allow us to choose the same constants than for the continuous problem (Part 1, (3.10)).
INRIA
13
Theorem 4.2.1 : Problem (4.1) has one and only one solution.
Proof :
Since the bilinear form a :; :]+ kb :; :] is uniformly Whel -elliptic and uniformly continuous with respect to h, and since the linear form ` :] + f :] is clearly uniformly continuous, we have just to apply the Lax-Milgram lemma to conclude.
The abstract error estimate is used in practice to obtain asymptotic error estimate. In the following theorem, the estimation is restricted to the usual approximation theory term inf k(uk ; uk ) ? (vh ; v h)kE which is known as soon as a nite (v ;v )2W element approximation is chosen.
h h hel
Theorem 4.3.1 : Let us consider the discrete problem (4.1) for which the bilinear form a :; :] + kb :; :] is uniformly Whel -elliptic, i.e., there exists a constant > 0, independent of h, such that :
a (vh ; v h ); (vh ; v h )] + kb (vh ; v h ); (vh ; v h )]
(4.5) We suppose, moreover, that there exists a constant M > 0, independent of h, such that ) j a (v ; v ); (w ; w )] + kb (v ; v ); (w ; w )] j M k(v ; v )kE k(w ; w )kE (4.6) 8(v ; v ) 2 Wel + Whel; 8(w ; w ) 2 Wel + Whel :
Then, there exists a constant C , independent of h, such that
k k(uk ; uk ) ? (uk k(uk ; uk ) ? (vh ; v h )kE inf h ; u h )kE C (v ;v )2W
h h hel
k where (uk ; uk ) (resp. (u k h ; u h )) denotes the solution of the continuous problem (Part 1, (3.10)) (resp. of the discrete problem (4.1)).
Proof :
Lemma 4.2.2 involves that relation (4.5) is veri ed. Likewise, relation (4.6) is a consequence of continuity properties of the bilinear forms a :; :] and b :; :]. Let (vh ; v h ) be any element of the space Whel ; we can write by using relations (4.1) and (4.5)
k k k 2 k k k(uk h ; u h ) ? (vh ; v h )kE a (u ; u ) ? (vh ; v h ); (uh ; u h ) ? (v h ; v h )] k + kb (uk ; u k ) ? (vh ; v h ); (uk h ; u h ) ? (v h ; v h )] k k k k k ?a (uk ; uk ); (uk h ; u h ) ? (vh ; v h )] ? kb (u ; u ); (uh ; u h ) ? (vh ; v h )] k k k +` (u k h ; u h ) ? (v h ; v h )] + f (u h ; u h ) ? (vh ; v h )]
RR n 2922
14
w ;w
h
sup
h
)2Whel
j a (uk ; uk ); (wh ; w h )] + kb (uk ; u k ); (wh ; w h)] ? ` (wh ; w h)] ? f (wh ; w h)] j : k(wh ; w h)kE
Now, let us obtain a new expression for a (uk ; uk ); (wh ; w h )]+kb (uk ; u k ); (wh ; w h )]. For that we come back to the equilibrium equations of the junction problem given in (Part 1, (2.27)-(2.28)). Making the product of these equations by test functions (wh ; w h ) 2 Whel and using Green's formula, we nally get with notations introduced in (4.1) 9 > a (uk ; uk ); (wh ; w h )] + kb (uk ; u k ); (wh ; w h )] > > = ` (wh ; w h )] +
Z
N wh d
N w hd
= ` (wh ; w h )] + f (wh ; w h )]; so that the second term in the above estimation disappears. To conclude, it remains to use the triangular inegality and to take the minimum with respect to (vh ; v h ) 2 Whel .
Remark 4.3.1 : The relation (4.7) allows us to cancel the consistency term that we usually nd in the abstract error estimate associated with nonconforming nite element method. Here, the nonconformity only appears along the hinge ? ; its e ect is circumvent by the introduction of the linear form f :] in (4.1). Another discrete problem could also be considered by dropping term f :] in (4.1). It should be di erent from problem (4.1) and would lead to a more classical abstract error estimate including a consistency term based on the linear form f :]. The solution of such a problem would be really closed to that of problem (4.1).
> > > > > = > > > > > > > > ;
(4.7)
INRIA
15
All the integrals appearing in the expressions of a :; :] and ` :] are of the form ^ K (x)dx. We use the usual correspondence between and through the a ne invertible mapping ^ ! FK (^ FK : x ^2K x) = BK x ^ + bK 2 K;
where BK is an invertible matrix, bK is a vector of IR2 such that FK (^ ai ) = a i ; i = ^ 1; 2; 3, where a ^i ; ai are the vertices of the triangles K and K . Then, the numerical ^ automatically induces a numerical integration integration scheme over the triangle K scheme over K , namely
Z
K
(x)dx
L X `=1
!`;K (b`;K ); `
Z
K
L. Moreover, we de ne the
( )d ?
L X `=1
We de ne similar functions for the domain and for the triangles K 2 Th . Finally, ^ 0 located upon we introduce a numerical integration scheme over each triangle side K the boundaries ; ; 1; 1 ; ! :
Z
^0 K
^ ( ^): EK ( ) = detBK E
^0 (^ s)ds ^
L0 X b0 ): ! ^ 0 ^0 (^ `=1 ` `
Thus, we use the correspondence between ^0 and 0 through the a ne invertible mapping ^ 0 ! GK 0 (^ s) = s ^ + 2 K 0; ^2 K GK 0 : s s ) = s ; = 1; 2 where s ^ ; s are the corresponding where 6= 0. We have GK 0 (^ 0 0 ^ vertices of the sides K and K . Then we have a numerical integration scheme over the segment K 0 0
Z
K0
0(s)ds
L X `=1
!`;K 0 0 (b`;K 0 );
b0`); 1 ` L0 and the associated error function^`0 and b`;K 0 = GK 0 (^ with !`;K 0 = ! nal
^ 0( ^0) = E so that
Z
K
X 0 0 0 0 0 ( 0) = ^0 (^ s ) d s ^ ? ! ^` ^ (^ b`); EK ^0
`=1
L0
Z
K0
0(s)ds ? X !
`=1
L0
`;K 0
0 (b
`;K 0 );
(5.2)
0 0 ( 0) = E ^ 0( ^0): EK
RR n 2922
16
Now, we can give the expression of the second discrete problem which takes into account the additional e ect of the numerical integration.
k Find (uk h ; u h ) 2 Whel such that k k k ah (uk h ; u h ); (vh ; v h )] + kbh (uh ; u h ); (vh ; v h )] = `h (vh ; v h )] + fh (vh ; v h )]
9 > > > > > = > > > > > ;
(5.3)
9 > > > > h ]g(b`;K ) > > > > = > > > > > > > > ;
ah (uh ; u h ); (vh ; v h )] =
+
L X X K 2T `=1
h
L X X K 2T `=1
h
!`;K f U h ] A] V
T
!`;K f T U h] A] V h]g(b`;K )
L0 X X
(5.4)
K 2T `=1
h
!`;K f P V
(5.6) where G1 and G1 denote the sets of the sides of triangles which are located upon 1 and 1 . Moreover (compare with (4.2)):
2G1 `=1
!`;K0 f T Ls ] V
9 > T ] > !`;K f T ] h ]g(b`;K ) > > h ]g(b`;K ) + > > > K 2T `=1 > = > > L0 > X X > > T > 0 0 0 ) ) + ] g ( b ] g ( b ] f ! > h `;K `;K s h `;K > ;
PV
K0
2G 1 `=1
L V
fh (vh ; v h )] =
L0 X X K 0 2G `=1
!`;K 0 f N ] V h ]g(b`;K0 ) +
T
L0 X X K 0 2G `=1
!`;K0 f T N ] V h ]g(b`;K0 )
(5.7) where G and G denote again the sets of the sides of triangles located on and .
Theorem 5.2.1 : Let us consider a family of discrete problems (5.3) for which the
ah (vh ; v h ); (vh ; v h )]+kbh (vh ; v h); (vh ; v h )]
bilinear forms ah :; :] + kbh :; :] are Whel -elliptic, uniformly with respect to h, i.e., there exists a constant > 0, independent of h, such that :
INRIA
Junctions between thin shells Then, there exists a constant C , independent of h, such that
k k(uk ; uk ) ? (uk h ; u h )kE C (v ;vinf )2W
h h
17
hel
w ;w
h
sup
h
)2Whel
+C
(5.8) k )) denotes the solution of the continuous problem (Part where (uk ; uk ) (resp. (uk ; u h h 1, (3.10)) (resp. of the discrete problem (5.3)).
h
w ;w
h
sup
)2Whel
j a (uk ; uk ); (wh ; w h )] + kb (uk ; u k ); (wh ; w h)] ? `h (wh ; w h)] ? fh (wh ; w h)] j k(wh ; w h)kE
Proof :
The assumption of Whel -ellipticity involves the existence and uniqueness k of a solution (uk h ; u h ) for the discrete problem (5.3). Then let (vh ; v h ) be any element of the space Whel ; we can write :
k 2 k k k k k(uk h ; u h ) ? (vh ; v h )kE ah (uh ; u h ) ? (vh ; v h ); (uh ; uh ) ? (vh ; v h )] k k k + kbh (uk h ; u h ) ? (vh ; v h ); (uh ; u h ) ? (vh ; v h )] k = a (uk ; uk ) ? (vh ; v h ); (uk h ; uh ) ? (vh ; v h )] k + kb (uk ; u k ) ? (vh ; v h ); (uk h ; u h ) ? (vh ; v h )] k k k + a (vh ; v h ); (uk h ; u h ) ? (vh ; v h )] ? ah (vh ; v h ); (uh ; u h ) ? (v h ; v h )] k k k + kb (vh ; v h ); (uk h ; u h ) ? (vh ; v h )] ? kbh (vh ; v h ); (uh ; u h ) ? (vh ; v h )] k k k k k ? a (uk ; uk ); (uk h ; uh ) ? (vh ; v h )] ? kb (u ; u ); (uh ; u h ) ? (vh ; v h )] k k k + `h (uk h ; u h ) ? (vh ; v h )] + fh (uh ; u h ) ? (vh ; v h )]
RR n 2922
18
w ;w
h
sup
h
j a (vh ; v h); (wh ; w h)] ? ah (vh ; v h); (wh ; w h)] j k(wh ; w h)kE
w ;w
h
sup
h
)2Whel
To conclude, it remains to use the triangular inegality and to take the minimum with respect to (vh ; v h ) 2 Whel .
In the estimate (5.8), in addition to the usual approximation theory term infk(uk ; uk )? (vh ; v h )k, we nd two additional terms which measure the consistency error between the bilinear forms a :; :] and ah :; :], b :; :] and bh :; :] ; they take into account the error due to the numerical integration. Finally, the last term combines the error due to both approximations, i.e., nonconforming approximation along the hinge and use of the numerical integration techniques.
The uniform Whel -ellipticity is based on the local error estimate theorems given by Bernadou (1996, p. 53-61) for a triangle K and by Cubier (1994, p. 76-87) for a triangle side K 0. These theorems give a general result of error estimate ; they specify criteria on the choice of numerical integration schemes in order to obtain the same order of asymptotic error estimate than for exact integration.
and satisfying properties (3.1) to (3.3). Let (K; PK ; K ) and (K ; PK ; K ) be two almost a ne families of nite elements associated with the Argyris triangle. Thus we have
^ (' ^ ). ii) E ^) = 0; 8' ^ 2 P8 (K ^ 0 veri es Likewise, the integration scheme on the reference segment K ^ 0 ; 8 ` = 1; ::; L0 ; iii) the integration nodes ^ b0` 2 K
^ satis es Moreover, assume that the integration scheme on the reference triangle K the following conditions : ^ 8 ` = 1; ::; L ; i) the integration nodes ^ b` 2 K;
INRIA
19
^ 0(' ^ 0) : iv) E ^) = 0; 8' ^ 2 P 8 (K Then, for any given AIJ 2 W 1;1 ( ); AIJ 2 W 1;1 ( ); I; J = 1; ::; 12 and C IJ 2 W 1;1 (!); I; J = 1; ::; 24, there exist constants > 0 and h1 > 0, independent of h, such that for all h 2]0; h1], we have ah (vh ; v h ); (vh ; v h )]+kbh (vh ; v h); (vh ; v h )] k(vh ; v h )k2 E ; 8(vh ; v h ) 2 Whel : (5.9)
Proof :
For any (vh ; v h ) 2 Whel Vh Vh, the inclusions (3.4) allow us to write 9 ah (vh ; v h ); (vh ; v h )] + kbh (vh ; v h ); (vh ; v h )] > > > = a (vh ; v h ); (vh ; v h )] + kb (vh ; v h ); (vh ; v h )] + ah (vh ; v h ); (vh ; v h )] ? a (vh ; v h ); (vh ; v h )] + kbh (vh ; v h ); (vh ; v h )] ? kb (vh ; v h ); (vh ; v h )];
> > > > > > > > = > > > > > > > > > > > ;
(5.10)
where a :; :] and b :; :] are de ned by relations (2.1) and (2.4). According to the proof of Lemma 4.2.2, there exists a constant > 0, independent of h, such that for any (vh ; v h ) 2 Whel a (vh ; v h ); (vh ; v h )] + kb (vh ; v h ); (vh ; v h )] k(vh ; v h )k2 (5.11) E: This result and the assumptions of Theorem 5.3.1 allow us to apply Theorem 1.3.3 of Bernadou (1996, p. 57) to the di erent types of terms which occur in the second hand member of the following inequality j a (vh ; v h ); (vh ; v h)] ? ah (vh ; v h); (vh ; v h)] j
12 X X
K 2T I;J =1
h
j EK (AIJ V hI V hJ ) j +
12 X X
K 2T I;J =1
h
j EK (AIJ V hI V hJ ) j :
Thus, there exists a constant C > 0, independent of h, which can change from an inequality to the next and such that 9 j a (vh ; v h ); (vh ; v h)] ? ah (vh ; v h); (vh ; v h )] j > > >
C(
X
K 2T
h
hK (
12 X
I;J =1
X
K 2T
h
hK (
12 X
I;J =1
> > > > > > > > 2 > k ) h V (K ) > > > > > = > > > > > > > > > > > > > > > > > ;
12 X
I;J =1
kA
9 = k IJ 1;1; ;
Chk(vh ; v h )k2 E:
(5.12)
RR n 2922
20
Now, we have to estimate the term b (vh ; v h ); (vh ; v h )] ? bh (vh ; v h ); (vh ; v h )]. By using de nition (5.2) of the error functional , we obtain
24 X X
K 2T I;J =1
h
0 (C V V ]I V V ]J ) j : j EK IJ h h h h
I1 = f1; ::; 12g f1; ::; 12g; I2 = f13; ::; 24g f13; ::; 24g; I3 = f1; ::; 12g f13; ::; 24g; I4 = f13; ::; 24g f1; ::; 12g;
K 2T I;J =1
h
0 (C j EK IJ
V h V h]I V h V h]J ) j =
X
h
X
h
K 2T I;J 2I1
0 (C j EK IJ
V h ]I V h ]J ) j
X
h
(5.13) With the hypotheses made in the statement of the theorem, we can apply to these one dimensional integration terms of the relation (5.13) the same kind of technique and we obtain (see Cubier (1994, p.83) for details) X X 0 (C V ]I V ]J ) j Ch X ( X kC k1;1;K )kvh k2 ; j EK IJ h h IJ V (K )
K 2T I;J 2I1
h
K 2T I;J 2I2
0 (C V ]I V ]J ) j + j EK IJ h h
K 2T I;J 2I3 I4
0 (C V ]I V ]J ) j j EK IJ h h
9 > > > > > > = > > > > > > ;
K 2T
I;J 2I1
X
h
X X
h
K 2T I;J 2I2
0 (C V ]I V ]J ) j j EK IJ h h 0 (C V ]I V ]J ) j j EK IJ h h
Ch
X
K 2T
h
I;J 2I2
K 2T I;J 2I3 I4
Ch
X
K 2T
h
I;J 2I3 I4
so that the substitution of these inequalities into relation (5.13) proves the existence of a constant C , independent of h, such that j b (vh ; v h); (vh ; v h )] ? bh (vh ; v h ); (vh ; v h)] j 9 > > >
Ch(
24 X
By substituting relations (5.11), (5.12), (5.14) into relation (5.10), we get ah (vh ; v h ); (vh ; v h )] + kbh (vh ; v h ); (vh ; v h)] ( ? Ch)k(vh ; v h )k2 E and it su ces to take = 2 and h1 = 2C to obtain the expected estimate (5.9). Remark 5.3.1 : Numerical integration schemes exact for polynomial of degree eight can be found in Dunavant (1985, p. 1140) for the triangle and in Zienkiewicz-Taylor (1989, p.173) for an interval.
I;J =1
(5.14)
INRIA
21
Now, we are able to evaluate the di erent terms of the abstract error estimate (5.8) and to derive an asymptotic error estimate k(u ; u ) ? (uh ; u h )k between the solution (u ; u ) of the continuous problem (Part 1, (3.10)) and the solutions of the discrete problems (5.3). Before giving the asymptotic error estimate theorem, let us specify some notations :
;q
=f =f
3 X
i=1
kpikq 4;q; +
3 X
i=1
1 ))12
(W 5;s (
1 ))12
12 X
I =1
kLsI ks 5;s; 1 +
12 X
I =1
1.
domains and satisfying the properties (3.1) to (3.3). Let (K; PK ; K ) and (K ; PK ; K ) be two almost a ne families of nite elements associated with the Argyris triangle. Moreover, we assume that the numerical integration scheme on the ^ satis es the following conditions : reference triangle K ^ 8 ` = 1; ::; L ; i) the integration nodes ^ b` 2 K;
^ 0 veri es Likewise, the numerical integration scheme on the reference interval K ^ 0 ; 8 ` = 1; ::; L0; iii) the integration nodes ^ b0` 2 K ^ 0(' ^ 0) : iv) E ^) = 0; 8' ^ 2 P 8 (K
Assume that
AIJ 2 W 4;1( ); AIJ 2 W 4;1( ) for I; J = 1; ::; 12; vii) C IJ 2 W 4;1 (! ) for I; J = 1; ::; 24;
vi) ix) N i 2 W 5;s ( ); N i 2 W 5;s ( ) for i = 1; ::; 3;
v) the solution (uk ; uk ) 2 Wel of the continuous problem (Part 1, (3.10)) belongs to the space K ( ) K ( ) = (H 5( ))2 H 6( ) (H 5( ))2 H 6 ( );
1)
RR n 2922
22
where q; s are integer numbers 2. Then, there exist constants C > 0 and h1 > 0, independent of h, such that for any h 2 ]0; h1], we have :
k 4 k(uk ; uk )k k(uk ; uk ) ? (uk K( h ; uh )kE Ch
) K( ) +
k(p ; p )k(W 4
;s
;q
;s
+ k(Ls ; L s )k(W 5
( 1 ))12 (W 5;s (
(5.15)
The conditions for applying Theorem 5.3.1 are satis ed. Hence, the condition of the uniform Whel -ellipticity is veri ed and it is possible to apply Theorem 5.2.1. Therefore, we are going to evaluate the di erent terms of the second hand member of the inequality (5.8). The proof takes ve steps.
Step 1 : Estimate of
(
Let
v ; v ) = ( h v; h v )
where h and h are the associated interpolation operators on Vh and Vh . Then, we obtain (Ciarlet (1978 p. 124))
(
v ;v
h
inf
h
)2Whel
u u
(5.16)
Step 2 : Estimate of
w ;w
h
sup
h
)2Whel
uk ; uk ); (wh ; w h)] j
12 X X
(5.17) We restrict our attention to the rst term of the second hand member of inequality (5.17). The hypotheses of Bernadou (1996, Theorem 1.3.3, p. 57) are veri ed. Thus, we obtain the existence of a constant C > 0, independent of h, such that
h
K 2T I;J =1
h
K 2T I;J =1
12 X X
K 2T I;J =1
h
j EK (AIJ hU k ]I W J ) j
X
K 2T
h
h4 K(
12 X
I;J =1
INRIA
23
where
h
u=
h u3;
h u1 ;
( h u1 );1; ( h u1);2 ;
h u2 ;
( h u2);1; ( h u2 );2;
k huik5;K kuik5;K + kui ? hui k5;K C kuik5;K ; i = 1; 2; 3: j EK (AIJ hU k ]I W ]hJ ) j C K max (h )4( 2T K
h
12 X
K 2T I;J =1
h
I;J =1
K 2T I;J =1
h
12 X
I;J =1
h(
uk ; uk ); (wh ; w h)] j
)
so that
(
w ;w
h
sup
h
)2Whel
12 X
I;J =1
kA
9 = k IJ k4;1; ; k( ;
u u
(5.20)
j b h(uk ; u k ); (wh ; w h )] ? bh h(uk ; u k ); (wh ; w h )] j Step 3 : Estimate of sup k(wh ; w h)kE (w ;w )2W The restriction of an Argyris triangle to one of its side is a P5 -one-dimensional
h h hel
uk ; u k ) = ( huk j F ;
uk j F ):
By using the matrix expressions (2.4) and (5.5) of the bilinear forms b :; :] and bh :; :] , a similar one-dimensional study to the previous ones gives (for details, see Cubier
RR n 2922
j b h(uk ; u k ); (wh ; w h)] ? bh h(uk ; u k ); (wh ; w h)] j 24 h X Ch4 ( kC IJ k4;1;K ) kuk kK( )kwh kV ( ) + kuk kK ( )kw hkV(
I;J =1
so that, we obtain
I;J =1
kC IJ k4;1;! )k(u u
k ; k )k
K ( ) K ( ):
9 > > > > > > = > > > > > > ;
(5.21)
Step 4 : Estimate of
(
w ;w
h
sup
h
)2Whel
j a (uk ; uk ); (wh ; w h )] + kb (uk ; u k ); (wh ; w h)] ? `h (wh ; w h )] ? fh (wh ; w h)] j k(wh ; w h)kE
w ;w
h h
sup
h
)2Whel
j a (uk ; uk ); (wh ; w h )] + kb (uk ; u k ); (wh ; w h)] ? `h (wh ; w h )] ? fh (wh ; w h)] j k(wh ; w h)kE
hel h h hel
bines the errors due to the nonconformity of the nite element method and to the use of numerical integration. Thus, it remains to study the following estimates j ` (wh ; w h )] ? `h (wh ; w h)] j Estimate of sup k(wh ; w h)kE (w ;w )2W We can rewrite the linear forms ` :] and `h :] by using the matrix decompositions introduced in Section 2 ; thus we obtain
h h hel
j ` (wh ; w h)] ? `h (wh ; w h)] j + j f (wh ; w h)] ? fh (wh ; w h )] j sup k(wh ; w h )kE k(wh ; w h )kE (w ;w )2W (w ;w )2W In this inequality, we nd the term j f (wh ; w h )] ? fh (wh ; w h )] j which comsup
h
12 X X
K 2T I =1
h
j EK ( P ]I W h]I ) j +
12 X X
12 X X
K 2T I =1
h
j EK ( P ]I W h]I ) j
K 0 2G1 I =1
0 0 ( Ls ]I W h ]I ) j + j EK
K 0 2G1 I =1
0 0 ( L s ]I W h ]I ) j j EK
9 > > > > > > > > = > > > > > > > > ;
(5.22)
INRIA
25
where the sets G1 and G1 are de ned in (5.6). The rst two terms of the second hand member of (5.22) correspond to body force resultants ; the last two terms correspond to surface load resultants. In order to estimate these terms, we can use Bernadou (1996, Theorem 1.3.2, p. 54) and their one dimensional counterparts by Cubier (1994, Theorem 5.2.3, p.79). That leads to :
j ` (wh ; w h )] ? `h (wh ; w h )] j C
X
X
K 2T
h
1=2?1=q( h4 K (mes(K ))
12 X
I =1
w hkV(K )
+C
X
K 0 2G1
=2 0 1=2?1=s( h9 K (mes(K ))
12 X
I =1
1=s kL sI ks 5;s;K 0 ) kw h kV (K );
where K (resp. K ) denote the triangles belonging to Th (resp. Th ) which count the sides K 0 (resp. K 0 ) located on 1 (resp. 1 ) among their own sides. Figure 3 summarizes these notations.
K K K0
GK 0
K0 K K
K
1
(F K )j K ^0
(F K )j K ^0
FK
FK
^ K ^0 K
RR n 2922
26
I =1
kP I kq 4;q;K =
3 X
i=1
12 X
I =1
kP I kq 4;q;K =
X
K
3 X
i=1
kpikq 4;q;K :
j aK bK cK j (
j aK j )1= (
X
K
j bK j )1= (
j cK j )1=
which is valid for any real numbers ; ; 1 satisfying 1 + 1 + 1 = 1. Here, we 1 1 take 1 = 1 = q et = 2 for body force terms and 1 = 1 = s et 2 ? q; 2 ? s; = 2 for surface load terms, for any q 2 and s 2. Hence 9 n j ` ( w h ; w h )] ? `h (wh ; w h )] j 4 > j (p ; p )k(W 4 ( ))3 (W 4 ( ))3 > > Ch sup > = ) k k ( w ; w E h (w ;w )2W h
;q ;q h h hel
+ h1=2k(Ls ; L s )k(W 5
;s
1 ))12
(W 5;s (
1 ))12
(5.23)
j f (wh ; w h )] ? fh (wh ; w h )] j sup k(wh ; w h )kE (w ;w )2W The nal result is obtained by taking into account the matrix expression of f :] and fh :] de ned by relation (5.7) ; a proof similar to the previous ones ( see Cubier (1994, Theorem 5.2.3, p.79)) would lead to
Estimate of
h h hel
w ;w
h
sup
h
)2Whel
;s
(5.24)
We can apply Theorem 5.2.1 and we get the estimate (5.15) by substitution of the inequalities (5.16), (5.20), (5.21) , (5.23) and (5.24) into the abstract error estimate (5.8) written for (v h ; v h ) = h (uk ; uk ).
Find (uhrig ; u hrig ) 2 Whrig such that for any (vh ; v h ) 2 Whrig ah (uhrig ; u hrig ); (vh ; v h )] = `h (vh ; v h)] + fh (vh ; v h)]
9 > = > ;
(6.1)
INRIA
27
where the space Whrig , the bilinear form ah :; :] and the form `h :] are respectively given by relations (3.13), (5.4), (5.6) and, concerning the approximation along the hinge :
fh (vh ; v h )] =
L0 X X K 0 2G `=1
!`;K0 f Ls ] V h ]g(b`;K 0 )+
T
L0 X X K 0 2G `=1
!`;K 0 f T Ls ] V h]g(b`;K0 )
where Ls and L s are de ned in (2.7) and sets G and G are given in relation (5.7).
Theorem 6.1 (Abstract error estimate) : Let us consider a family of discrete prok(vh ; v h)k2 E ; 8(v h ; v h ) 2 Whrig : Then, there exists a constant C > 0, independent of h, such that
ah (vh ; v h ); (vh ; v h )]
blems (6.1). We suppose that the bilinear forms ah :; :] are Whrig -elliptic, uniformly with respect to h, i.e., there exists a constant > 0, independent of h, such that
hrig
+ +C
(
w ;w
h
sup
h
)2Whrig
w ;w )2W where (urig ; u rig ) and (uhrig ; u hrig ) denote the solutions of the continuous and
h h hrig
sup
Proof : The proof is similar to that of Theorem 5.2.1 for the elastic hinge problem ; it su ces to note that b :; :] = bh :; :] = 0.
of triangulations of the domains and satisfying properties (3.1) to (3.3). Let (K; PK ; K ) and (K ; PK ; K ) be two almost a ne families of nite elements associated with the Argyris triangle. Moreover, we assume that the numerical integration ^ satis es the following conditions : scheme on the reference triangle K ^ 8 ` = 1; ::; L ; i) the integration nodes ^ b` 2 K;
Theorem 6.2 (Asymptotic error estimate) : Let Th and Th be two regular families
^ 0 veri es Likewise, the numerical integration scheme on the reference interval K ^ 0 ; 8 ` = 1; ::; L0 ; iii) the integration nodes ^ b0` 2 K ^ 0(' ^ 0) : iv) E ^) = 0; 8' ^ 2 P 8 (K
RR n 2922
28
Assume that v) the solution (urig ; u rig ) 2 Wrig of the continuous problem ( Part 1,(3.24)) belongs to the space K ( ) K( ) = (H 5( ))2 H 6( ) (H 5( ))2 H 6( );
) K( )
1
+ h1=2k(Ls ; L s )k(W 5
;s
))12 (W 5;s (
))12 ]
7 Numerical Results
We consider the example given in (Part 1, Section 2.4), i.e., a structure constituted by a circular cylinder with a spherical end cap subjected to a vertical body force. The cylinder is clamped along its foundations. Figure 4 and Table 1 give the notations and the data used for this benchmark. The axisymmetry of this structure leads to simpli cations, in particular if we consider axisymmetrical loading. Dimensions L = 130:9m r = 95:1m R = 100m e = e = 1m 2 0= 5 = 10 Material constants Forces = ?be3 b = 10N=m2
E = E = 105N=m2 = = 0:33
p=p
INRIA
29
z
?1
2
p
?
2
0 1
S
1
e3
S
x
e1 e2 r
e=e
y
?0
x = r cos y = r sin z= 1
2 2
Here, we consider the junction between the sphere and the cylinder ; the above mappings are such that ?1 (?) = ?1 (?) = . Thus we do not need to introduce the third mapping and the interval ! (see (2.3) and (2.4)). In this benchmark the bilinear form b :; :] is given by
1 u )(v ? v ? 1 v )d b (u ; u ); (v ; v )] = (u3;1 ? u 3;1 ? R 3;1 R 1 1 3;1
RR n 2922
30
As a reference deformed con guration of the structure we use the approximated solution of the three-dimensional linearized elasticity model. The axisymmetry allow us to restrict the problem to a two-dimensional axisymmetrical elasticity problem set on a radial section of the structure. We approximate this problem by a nite element method using P1 -Lagrange nite elements. In order to have a good approximation of the deformed con guration, we have used a mesh including 17408 triangles. Figure 5 gives the result that we have obtained with the Modulef code (see Bernadou & al. (1988)).
Fig. 5: Reference deformed con guration for rigid junction obtained from a 2d
INRIA
31
For this test, we use the Argyris triangle described in Figure 2 and the mesh given in Figure 6.
a2 a 2
a1 a 1
Fig. 6 : Mesh
Because of the axisymmetry of the problem, we restrict the reference domains to an axial strip of width `b = =16 and we impose symmetry conditions on the boundaries 1 and 2 (Figure 6). The rigid junction conditions along are obtained from equations (3.9) and (3.12). They consist of linear relations between the degrees of freedom located on . The equality of the displacements along the junction is j j given by u i = Aj i uj with Ai = a i a . In our example, we have : where is de ned in Table 1 and Figure 4. Conditions (3.9) can be written as:
8 > < u h1(a u h2(a > :
) = cos uh1;2 (a ) + sin uh3;2(a ) ) = uh2;2(a ) u h3;2(a ) = ? sin uh1;2(a ) + cos uh3;2(a ) 8 > < u h1;22(a ) = cos uh1;22(a ) + sin uh3;22 (a ) u h2;22(a ) = uh2;22(a ) > : u h3;22(a ) = ? sin uh1;22(a ) + cos uh3;22(a ) where a ; a are displayed on Figure 6, = 1; 2. The equality of rotations is given by the conditions (see (3.12)):
1 u (a ) = 0 uh3;1 (a ) ? u h3;1(a ) ? R h1
RR n 2922
32
interpolating function for Argyris triangle (see for example Bernadou (1996, (1.5.1), p. 99) give 1 u (a ) + u (a )] + 5 Du (a )(a ? a ) + Du (a )(a ? a )]+ u h1(b ) = 2 2 h1 2 1 h1 2 h1 1 32 h1 1 2 1 1 D2u (a )(a ? a )2 + D2u (a )(a ? a )2] 1 1 h1 2 2 h1 1 2 64 5 1 2 = 1 u h1(a 1) + u h1 (a 2)] + `b u h1;2(a 1) ? u h1;2(a 2)] + `b u h1;22(a 1) + u h1;22(a 2)] 2 32 64 where `b = =16 is the width of the radial strip into consideration in this test. The numerical results are given in Figure 7. They are very closed to the reference solution considered in Paragraph 7.2. This is not surprising since the computation of the reference solution through 2D elasticity equations does not introduce e ects of elastic hinge including a jump of rotation. Thus this reference solution is really very closed to the rigid hinge solution.
1u uh3;12(a ) ? u h3;12(a ) ? R h1;2 (a ) = 0 1 u (b ) = 0 uh3; (b) ? u h3; (b ) ? R h1 where b; b are the midpoints of a1; a2] and a 1; a 2]. In the above relation, term u h1 (b ) is not a degree of freedom of Argyris triangle, but the expression of the
INRIA
RR n 2922
33
34
In order to simulate an elastic junction, we have implemented in the Modulef code a junction nite element which is built from two Argyris triangles which joint along one of their sides (see Figure 8) and whose rigidity matrix includes the bilinear form bh :; :]. We use again the mesh given in Figure 6, the junction element is represented in grey. We have seen in Part 1 that the elasticity of the junction depends of a coe cient k. When k is small (about 10), the junction is elastic and is very distorded. When k is very large (about 107), the junction becomes almost rigid and the angle between the two shells remains unchanged during the deformation (see Figure 7).
a3 a6 b3 b5 K b4 a2 a4 a5
a1
K b2 b6 b1
8 Final remarks
In this study we have approximated the thin shell junction problem by using the high accuracy nite element method based on the Argyris element. Another possibility is to use the delinquent nite element introduced by Sander (1985). This element seems promising since it combines high accuracy (an error estimate in O(h2 ) in energy norm would seem reasonable) with a choice of degrees of freedom well adapted to the junction problem. These elements are shown schematically in Figure 9. A study along these lines is in progress by Bernadou, Palma, Trouve and Vilari~ no (to be published). Another interesting study would be the study of the junction between two shells modelized by Reissner-Mindlin equations. In this way, the transverse shear e ects and the boundary layers phenomena have to be considered.
INRIA
35
References
Argyris, J.H., Fried, I., Scharpf, D.W. (1968) : The TUBA family of plate elements for the matrix displacement method, Aero. J. Royal Aeronautical Society, 72, pp. 701-709 Bernadou M. (1996): Finite Element Methods for Thin Shell Problems, J. Wiley & Sons, Chichester. Bernadou M., Cubier A. (to appear) Numerical analysis of junctions between thin shells, Part 1 : Continuous problems. Bernadou M., Fayolle S., Lene F. (1989): Numerical analysis of junctions between plates, Comp. Methods App. Mech. Engrg. 74, pp. 307-326. Bernadou M., George P.L., Hassim A., Joly P., Laug P., Muller B., Perronet A., Saltel E., Steer D., Vanderbork G., Vidrascu M. (1988): Modulef : Une Bibliotheque Modulaire d'Elements Finis, I.N.R.I.A Bernadou M., Palma F., Trouve P., Vilari~ no M.A. (to appear) Ciarlet P.G. (1978) The Finite Element Method for Elliptic Problems, NorthHolland, Amsterdam. Cubier A. (1994) Analyse et Simulation Numeriques de Jonction de Coques Minces, These de l'Universite P. et M. Curie. Dunavant D.A. (1985) High degree e cient symmetrical gaussian quadrature rules for the triangle, Internat. J. Numer. 13, pp. 1129-1148.
RR n 2922
36
Fayolle S. (1987) Sur l'Analyse Numerique de Raccords de Poutres et de Plaques, These de l'Universite P. et M. Curie. Sander G. (1985) Modeles d'elements nis delinquants, Groupe de travail \Methodes Numeriques pour les Problemes de Coques", 17 Janvier 1985, Clamart. Zenisek A. (1981): Nonhomogeneous boundary conditions and curved triangular nite element, Apl. Mat. 26, pp. 121-140. Zienkiewicz O.C., Taylor R.L. (1989) The Finite Element Method Fourth Edition Vol. 1, Basic Formulation and Linear Problems, MacGraw-Hill, New York.
INRIA
Unit e de recherche INRIA Lorraine, Technop ole de Nancy-Brabois, Campus scientique, ` NANCY 615 rue du Jardin Botanique, BP 101, 54600 VILLERS LES Unit e de recherche INRIA Rennes, Irisa, Campus universitaire de Beaulieu, 35042 RENNES Cedex Unit e de recherche INRIA Rh one-Alpes, 46 avenue F elix Viallet, 38031 GRENOBLE Cedex 1 Unit e de recherche INRIA Rocquencourt, Domaine de Voluceau, Rocquencourt, BP 105, 78153 LE CHESNAY Cedex Unit e de recherche INRIA Sophia-Antipolis, 2004 route des Lucioles, BP 93, 06902 SOPHIA-ANTIPOLIS Cedex
Editeur INRIA, Domaine de Voluceau, Rocquencourt, BP 105, 78153 LE CHESNAY Cedex (France) ISSN 0249-6399