You are on page 1of 13

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/227314312

The Use of Curved Elements in the Finite Element


Approximation of Thin Plates by High Order p and hp
Methods

Article  in  Journal of Scientific Computing · June 2006


DOI: 10.1007/s10915-005-9053-9 · Source: DBLP

CITATIONS READS

0 85

1 author:

Christos Xenophontos
University of Cyprus
76 PUBLICATIONS   705 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Christos Xenophontos on 06 June 2014.

The user has requested enhancement of the downloaded file.


Journal of Scientific Computing, Vol. 27, Nos. 1–3, June 2006 (© 2005)
DOI: 10.1007/s10915-005-9053-9

The Use of Curved Elements in the Finite Element


Approximation of Thin Plates by High Order p and hp
Methods
Christos Xenophontos1

Received September 29, 2004; accepted (in revised form) February 20, 2005; Published online December 23, 2005

In this article we focus on the approximation of the Reissner–Mindlin (R–M)


plate model by high-order p and hp versions of the Finite Element Method
(FEM), in the case when the mesh includes curved elements; this is of particu-
lar interest when plates with smooth boundary are considered. Our main goal
is to verify that an appropriately defined Mixed Interpolated Tensorial Compo-
nents (MITC) FEM performs well, in the presence of curved elements, when
the energy norm is used as an error measure, as well as when quantities of
engineering interest are computed. Comments on the performance of the stan-
dard FEM formulation are also made.

KEY WORDS: Reissner-Mindlin plate model; p and hp finite elements; MITC


method; shear locking; boundary layers.

1. THE R–M PLATE MODEL


The Reissner-Mindlin (R–M) plate model is a widely used system of partial
differential equations, which describes the deformation of a thin plate sub-
ject to transverse loading. This two-dimensional model often replaces the
full three-dimensional elasticity problem, when the thickness of the plate
is small. To present the R–M model, we consider the bending of a homo-
geneous isotropic plate of thickness t > 0, occupying the region R = Ω ×
(−t/2, t/2), where Ω ⊂ R2 represents the midplane of the plate, under nor-
mal load density given by t 3 g(x, y), where g is independent of t. The equa-
tions of equilibrium for the rotation φ, and transverse displacement w are

1 Department of Mathematical Sciences, Loyola College, 4501 N. Charles Street, Baltimore,


MD 21210, USA. E-Mail: cxenophontos@loyola.edu

465
0885-7474/06/0600-0465/0 © 2005 Springer Science+Business Media, Inc.
466 Xenophontos

D   

 − φ = 0,
 · φ − Gκt −2 ∇w
−  φ + (1 + ν)∇∇
(1 − ν)∆ (1)
2  
−Gκt −2 ∇ · ∇w  − φ = g, (2)

where E is Young’s modulus, ν is Poisson’s ratio, G = E/2(1 + ν), D =


E/12(1 − ν 2 ), and κ is the shear correction factor (often chosen as 5/6).
It is readily visible that the system (1)–(2) is singularly perturbed, hence
its solution will, in general, contain boundary layers (in the rotation, but not
in the transverse displacement). The accurate approximation of the bound-
ary layers is extremely important and can be achieved by using the p/ hp
Finite Element Method (FEM) on an appropriately designed mesh [8]. In
addition to containing boundary layers, it is well known that the solution to
(1)–(2) converges to the solution of the fourth-order Biharmonic equation
as t → 0, i.e., the solution to the R–M plate satisfies Kirchhoff’s constraint

∇w 
 − φ = 0. (3)

This leads to (additional) difficulty in approximating the solution to (1)–(2)


for very thin plates, mainly due to the inability of certain approximating
spaces to enforce the constraint (3). This phenomenon is called locking, and
currently there are two ways to deal with it: (i) enforce Kirchhoff’s con-
straint exactly (by using, e.g., the high-order p/ hp versions of the standard
FEM), or (ii) enforce Kirchhoff’s constraint weakly, by using a modified var-
iational formulation. When curved elements are present, however, locking in
the standard FEM is more pronounced, as was observed computationally in
[10], even though as p → ∞ the method appears to be locking-free. Indeed,
these observations were verified recently [6], and the amount of pre-asymp-
totic locking was identified and shown to depend on the mapping of the
(curved) elements—see Eq. (11) ahead. In addition to this shortcoming, the
standard formulation fails to produce satisfactory results when quantities of
engineering interest are computed, such as the stress and moment resultants,
unless some kind of post-processing is used (cf. [7]). We will illustrate this
in Sec. 4 ahead by considering, for concreteness, the approximation of the
shear force, defined as
 
 = Qx , Qy = −Gκt −2 (∇w
Q  − φ).
 (4)

Modified variational formulations, on the other hand, do not suffer


from locking and yield very good results independently of the error measure
used [1, 9]. One such widely used method is the so-called Mixed Interpolated
Tensorial Components (MITC) FEM, which was originally introduced in [3],
Use of Curved Elements in the Finite Element Approximation of Thin Plates 467

and analyzed in the context of the hp version in [9]. In this article we investi-
gate an extension of this method which includes curved elements, and verify
that the (original) definition of the hp MITC elements from [9] indeed works
in practice when one deals with curvilinear domains.
The rest of the paper is organized as follows: in Sec. 2 we present the
discretization of the R–M equations by the standard formulation. Section 3
presents an hp MITC method for curved elements and Sec. 4 contains the
results of numerical computations for a model problem. Our conclusions
are presented in Sec. 5. In what follows, the usual notation H k (Ω) will be
used for spaces containing functions on the domain Ω ⊂ R2 with boundary
∂Ω smooth, having k generalized derivatives in L2 (Ω). The norm on H k (Ω)
will be denoted by  · k,Ω . Finally, the condensed notation
 w)2
(φ,  2 + w2 = φ1 2 + φ2 2 + w2 ,
= φ
r,s,Ω r,Ω s,Ω r,Ω r,Ω s,Ω

defines a norm on the product space [H r (Ω)]2 × H s (Ω).

2. THE STANDARD FORMULATION


Without loss of generality, we will restrict our description of the
method to hard clamped plates, where the displacement and rotation are
zero on ∂Ω. Our numerical results in Sec. 4 ahead will demonstrate the
applicability of the method to other boundary conditions. The variational
formulation for the R–M equations reads: find (φ,  w) ∈ [H 1 (Ω)]2 × H 1 (Ω)
0 0
such that

D
 θ) + Gκt −2 b(φ,
a(φ,  w; θ, ζ ) = gζ dA (5)
2 Ω
for all (θ, ζ ) ∈ [H01 (Ω)]2 × H01 (Ω), where
  
 
a(φ, θ ) =  1 · ∇θ
(1 − ν)(∇φ  1 + ∇φ
 2 · ∇θ
 2 ) + (1 + ν)(∇ · φ)(∇
  dA
· θ)

and
    
 w; θ, ζ ) =
b(φ,  − φ · ∇ζ
∇w  − θ dA.

The standard discretization of (5) consists of constructing a pair of
finite-dimensional subspaces VN (Ω) ⊂ [H01 (Ω)]2 , WN (Ω) ⊂ H01 (Ω) of com-
bined dimension N (the total number of degrees of freedom), and solving
the problem: find (φN , wN ) ∈ VN (Ω) × WN (Ω) such that

D
  −2  
a(φN , θ ) + Gκt b(φN , wN ; θ , ζ ) = gζ dA (6)
2 Ω
468 Xenophontos

for all (θ, ζ ) ∈ VN (Ω) × WN (Ω). The global spaces VN (Ω) and WN (Ω)
are constructed by first partitioning the domain Ω into a mesh M of
curvilinear quadrilateral and/or triangular elements Ωk , each of which is
the image of a reference element Ω  under an invertible element map-

ping Fk : Ω → Ωk . The reference element Ω  is chosen as either the unit

square S = [−1, 1] or the reference triangle T = {(ξ, η) ∈ [0, 1]2 : η  1 − ξ }.
2

Then the global spaces VN (Ω), WN (Ω) are defined piecewise in the follow-
ing way: polynomial spaces V  and W
p1 (Ω)  are chosen on the refer-
p2 (Ω)
   
ence elements Ω = S or T , among Qp,q (Ω) = span{ξ i ηj : 0  i  p, 0  j 
 = Qp,p (Ω),
q}, Qp (Ω)  or Pp (Ω) = span{ξ i ηj : 0  i + j  p}. The reference
spaces are then mapped onto each element to create the spaces
 
Vp1 (Ωk ) = φp = φ
◦ F −1 : φ
k
∈ V  ,
p1 (Ω) (7)
 
Wp2 (Ωk ) = wq = w  ◦ Fk−1 : w
∈W  .
p2 (Ω) (8)

Finally, the global spaces are defined by


 
VN (Ω) = φN ∈ H01 (Ω) : φN |Ωk ∈ Vp1 (Ωk ) ∀Ωk ∈ M , (9)
 
WN (Ω) = wN ∈ H01 (Ω) : wN |Ωk ∈ Wp2 (Ωk ) ∀Ωk ∈ M . (10)

As already mentioned, the standard discretization (6) is highly sensi-


tive to the plate thickness t. It is well known that the standard h version
exhibits complete locking, unless polynomials of degree greater than 3 are
used for the approximation [11]. Unlike the h version, the high-order p
and hp versions are free of locking as p → ∞, when the error in the energy
norm is of interest (see [10] and [11] for more details). Although it has
been shown that standard p and hp methods are free of locking, rigor-
ous analyses have only been completed for meshes consisting of rectangu-
lar and straight sided triangular elements [11]. The analysis of the p/ hp
version for curvilinear meshes remains open, even though numerical evi-
dence suggests that these methods are indeed asymptotically locking-free,
even when certain curvilinear elements are used [5, 6, 8]. To our knowl-
edge, the only analytical results in this direction are those in [6], where the
case of the p version on a single element with four curved sides, each of
which is mapped from the reference element via a polynomial mapping,
was studied. In particular, it was shown that if the solution to (5) satis-


2
fies u = φ , w ∈ H r−1 (Ωk ) × H r (Ωk ) , r ≥ 2, i.e. u is sufficiently smooth
with no boundary layers present, and the invertible mapping Fk is polyno-
2

q , with V
mial, i.e. Fk ∈ V q = Qq Ω  or Pq Ω  , then there exists a con-
stant C ∈ R independent of t, p and α, such that
Use of Curved Elements in the Finite Element Approximation of Thin Plates 469

u − uN 1,1,Ωk  C (p − α)−r+1 ur,r+1,Ωk , (11)

where

q = Qq Ω
3q − 2 for V 
α=
, (12)

3q − 4 for Vq = Pq Ω
 


with uN = φN , wN ∈ [V p (Ωk )]2 × V p+q (Ωk ) denoting the solution to (6).
Practically, the above result describes the amount of pre-asymptotic locking
and, in addition, shows that indeed the method is free of locking as p → ∞,
even when certain curved elements are used for the discretization. (See [6]
for more details, including a more “practical” version of this result).
However, standard methods do not yield satisfactory results when the
moment and/or stress resultants are of interest. In [7] this problem was
somewhat alleviated through the use of the p version FEM along with a
post-processing scheme for computing the resultants, equivalent to using
the equilibrium (as opposed to the constitutive) equation. This will be
illustrated in the numerical results of Sec. 4 ahead.

3. THE MITC FORMULATION


One of the main advantages of MITC methods is that the need for
post-processing is eliminated, since these methods approximate both the
solution and the resultants well, without any additional computational
effort.
To define the MITC method, we start with a known stable1 space VN
for the rotation, and then project it, using a reduction operator Π N , onto a
space of polynomials (see [9] for several choices of such spaces and reduc-
tion operators). These global spaces are defined using the usual reference
spaces; however, as was shown in [4], curved elements require some “spe-
cial” treatment. In particular, the basis for the reference space for the rota-
tions is “split” into two disjoint subsets corresponding to the internal and
external basis functions. The external basis functions are those which are
non-zero along (at least one portion of) the boundary, while the internal
basis functions are zero along the boundary and non-zero in the interior
(see e.g., Ch. 6 in [12]). The space spanned by the external basis functions
is mapped using the usual mapping in order to ensure inter-element conti-
nuity. The space spanned by the internal basis functions is mapped using a
kind of Piola transform (see below and [4] where this idea, referred to as
Method 3, was successfully used for elasticity problems).

1 The term stable means that the spaces satisfy the inf-sup condition.
470 Xenophontos

Consequently, the rotation space VN is defined for a mesh M com-


posed of, e.g., curvilinear quadrilaterals as follows: the reference space
p (
V S) on the reference square is taken to be
p (
V S) = [Qp+1 ( p0 (
S)]2 = V pe (
S) ⊕ V S),
where the superscripts 0 and e are used to denote the subspace of inter-
nal and external functions, respectively. The element space Vp (Sk ) is then
defined by
Vp (Sk ) = Vp0 (Sk ) ⊕ Vpe (Sk )
   
= φ = Jk−T φ ◦ F −1 : φ p0 (
∈ V ◦ F −1 : φ
S) ⊕ φ = φ pe (
∈ V S) ,
k k

where Sk = Fk (
S) and Jk−T is the inverse transpose of the derivative of the
element mapping Fk . Finally, the global space VN is defined by
 
 Sk ∈ Vp (Sk ), ∀Sk ∈ M
VN = φ ∈ [H01 (Ω)]2 : φ|

and the midplane displacement space WN (Ω) is obtained in the usual


manner using (8) and (10) with W 
 = Qp (Ω).
p (Ω)
The space VN is projected by a reduction operator Π N , defined ele-
mentwise by

−T Π  ◦ F −1  Sk ∈ Vp0 (Sk ),
for φ|
(Π  Sk = J (Π p φ )−1
ΠN φ)| k
) ◦ F
Πp φ
(Π  Sk ∈ Vpe (Sk ),
for φ|
k

where the reference projection Π p is a strategically chosen projection onto


a space of polynomials. For concreteness, we choose the Raviart–Thomas
spaces Qp−1,p (S) × Qp,p−1 ( S), even though other choices are possible, e.g.
the BDFM spaces (cf. [9]). (The specific choices for the spaces used here
correspond to Method 4 of [9].) Specifically, the conditions defining Π p
are

− φ
Πp φ
((Π ) · t)
v = 0, for all   for every edge E
v ∈ Pp−1 (E)  of 
S,

E 
− φ
Πp φ
(Π ) · r ∈ Qp−1,p−2 (
r = 0, for all  S) × Qp−2,p−1 (
S).

S

The resulting discrete problem is: find (φN , wN ) ∈ VN (Ω) × WN (Ω)
such that

D
ΠN φN , wN ; Π N θ,
a(φN , θ) + Gκt −2 b(Π  ζ)= gζ dA (13)
2 Ω
for all (θ, ζ ) ∈ VN (Ω) × WN (Ω). (For implentational details see [5].)
Use of Curved Elements in the Finite Element Approximation of Thin Plates 471

4. NUMERICAL EXAMPLE
In this section we present the results of numerical computations for a
model problem with a known exact solution [2]; for additional numerical
results see [13]. The results presented here are for a soft-simply-supported,2
unit circular plate, with Young’s modulus E = 1, Poisson ratio ν = 0.3, shear
correction factor κ = 1, and transverse load density given in polar coordi-
nates by g(r, θ ) = cos(θ ). The mesh is shown in Fig. 1 and it includes thin
elements of width pt along the boundary of the domain in order for the
boundary layer to be uniformly approximated [8]. Here p is the degree of
the approximating polynomial, which is increased from p = 1 to p = 8 for our
computations, and t is the plate thickness which was chosen as t = 10−j , j =
2, 3—other choices for t yield similar results. We note that no approxima-
tion to the boundary of the plate is made, but rather the curved elements
are mapped exactly, using the blending map technique (cf. [12], pp. 107–108)
to construct the element mappings Fk .

Fig. 1. 9-element unit-circular mesh with boundary refinement.

2 Forthis choice of boundary conditions the boundary layer is strong, hence we are truly test-
ing the methods under consideration.
472 Xenophontos

We will be plotting the error measured in the energy norm

D
 w)2E =
(φ,  φ)
a(φ,  w; φ,
 + Gκt −2 b(φ,  w) (14)
2

versus the number of degrees of freedom N, in a semi-log scale, as calcu-


lated using the MITC and the standard finite element method. In addition
to (14), we are also interested in the pointwise error in the shear force. For
the MITC formulation the shear force will be computed using

 = −Gκt −2 (∇w
Q  −Π 
ΠN φ), (15)

while for the standard formulation the shear force will be computed using
the constitutive equation (4), as well as the equilibrium equation, which
basically amounts to a post-processing scheme—the computations for this
last case were performed using the commercial finite element code Stress-
Check (E.S.R.D., St. Louis, MO). Since Q ∈/ L2 (Ω) as t → 0, one can-
not expect pointwise approximations to have any accuracy uniformly in
t, especially near the boundary. For this reason we will compute Q  suffi-
ciently away from the boundary; in particular we will be measuring the
first component of the shear force Qx (x, 0) for 0  x  1 − pt, with p = 8
(the highest polynomial degree).
Figure 2 shows the error measured in the energy norm, as computed
by both the standard and MITC formulations. As this figure indicates,
both methods perform well, independently of the thickness t, and near
exponential convergence rates are observed, with the relative error being
reduced down to 1%.
Figure 3 shows the shear force distribution, as well as the error in
the shear force, for the standard formulation (with and without post-pro-
cessing) and for the MITC formulation. First, we note that for t = 0.01
most of the error comes from the interior of the plate where the num-
ber of elements is minimal—this would not be the case if more elements
were used. Second, we confirm that the standard formulation performs
very poorly unless post-processing is used; in this case the standard FEM
performs as well as the MITC method, even in the presence of curved
elements.
To compare the latter two approaches, we show in Fig. 4 the percent-
age relative pointwise error in the shear force at the point (0, 0), for t =
0.01 (for other values of t the results were almost identical). We see that
both methods converge at a near exponential rate, with the MITC formu-
lation having a slight advantage.
Use of Curved Elements in the Finite Element Approximation of Thin Plates 473

2
Soft-Simply-Supported plate, t=0.01
10
Standard Method
Percentage Relative Error in Energy Norm

hp-MITC Method

1
10

0
10

-1
10
0 500 1000 1500 2000 2500
Degrees of Freedom, N

2 Soft-Simply-Supported plate, t=0.001


10
Standard Method
Percentage Relative Error in Energy Norm

hp-MITC Method

1
10

0
10

-1
10
0 500 1000 1500 2000 2500
Degrees of Freedom, N
Fig. 2. Energy norm convergence, t = 0.01, 0.001.
474 Xenophontos

Shear force calculations for the S-S-S plate, t = 0.01


Exact
0.3 Stand. FEM
Stand. FEM w/ post-processing
0.2 hp-MITC
Qx(x,0)

0.1
0
-0.1
-0.2

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9


x

0.05
|Qx(x,0)-QxFEM(x,0)|

Stand. FEM
Stand. FEM w/ post-processing
0.04 hp-MITC

0.03

0.02

0.01

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
x

Shear force calculations for the S-S-S plate, t = 0.001


0.4

0.2
Qx(x,0)

Exact
-0.2 Stand. FEM
Stand. FEM w/ post-processing
hp-MITC
-0.4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
x

Stand. FEM
|Qx(x,0)-QxFEM(x,0)|

Stand. FEM w/ post-processing


hp-MITC
0.1

0.05

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
x

Fig. 3. Shear force distribution, t = 0.01, 0.001.


Use of Curved Elements in the Finite Element Approximation of Thin Plates 475

3 S-S-S plate, thickness = 0.01


10
Stand. FEM w/ post-processing
hp-MITC
100 × |Qx(0,0)-QxFEM(0,0)|/|Qx(0,0)|

2
10

1
10

0
10
1 2 3 4 5 6 7 8
Polynomial degree p
Fig. 4. Convergence of the shear force at (0,0) for t = 0.01.

5. CONCLUSIONS
We studied the approximation of the R–M plate model by the p/ hp
FEM, and commented on the performance of the standard and MITC for-
mulations in the presence of curved elements. The latter formulation, origi-
nally defined in [9], was implemented using the ideas of [4] on how to handle
curved elements using mixed formulations. We conclude that when the error
in the energy norm is used as an error measure, both formulations perform
well, assuming the appropriate mesh design is used to capture the bound-
ary layers. When quantities of engineering interest, such as the shear force,
are computed, the standard formulation fails to produce good results unless
some kind of post-processing is used, in which case its performance is signifi-
cantly improved and becomes comparable to that of the MITC formulation
(with the latter having a slight advantage).

REFERENCES
1. Ainsworth, M., and Pinchedez, K. (2002). hp-MITC finite element method for the Reiss-
ner-Mindlin plate problem. J. Comp. Appl. Math. 148, 429–462.
476 Xenophontos

2. Arnold, D., and Falk, S. (1989). Edge effects in the Reissner-Mindlin plate theory. In
Noor, A. K., Belytschko T., and Simo, J. C. (eds.), Analytic and Computational Models of
Shells, A.S.M.E., New York, pp. 71–90.
3. Brezzi, F., Bathe, K. J., and Fortin, M. (1989). Mixed-interpolated elements for Reissner-
Mindlin plates. Int. J. Numer. Methods Engrg. 28, 1787–1801.
4. Chilton, L., and Suri, M. (2000). On the construction of stable curvilinear p version ele-
ments for mixed formulations of elasticity and Stokes flow. Numer. Math. 86, 29–48.
5. Kurtz, J. (2002). A p-version Finite Element Method for Shear Force Computation in Re-
issner-Mindlin Plates with Curved Boundary, M.Sc. Thesis, Department of Mathematics
and Computer Science, Clarkson University, Potsdam, NY.
6. Kurtz, J., and Xenophontos, C. (2003). On the effects of using curved elements in the
approximation of the Reissner-Mindlin plate by the p version of the finite element
method. Appl. Num. Math. 46, 231–246.
7. Rank, E., Krause, R., and Preusch, K. (1998). On the accuracy of p-version elements for
the Reissner-Mindlin plate model. Int. J. Numer. Meth. Engrg. 43, 51–67.
8. Schwab, Ch., Suri, M., and Xenophontos, C. (1998). The hp finite element method for
problems in mechanics with boundary layers. Comput. Methods Appl. Mech. Engrg. 157,
311–333.
9. Stenberg, R., and Suri, M. (1997). An hp error analysis of MITC plate elements. SIAM
J. Num. Anal. 34, 544–568.
10. Suri, M. (1996). Analytic and computational assessment of locking in the hp finite ele-
ment method. Comput. Meth. Appl. Mech. Engrg. 133, 347–371.
11. Suri, M., Babuška, I., and Schwab, C. (1995). Locking effects in the finite element
approximation of plate models. Math. Comp. 64, 461–482.
12. Szabó, B., and Babuška, I. (1991). Finite Element Analysis, Wiley.
13. Xenophontos, C., Kurtz, J., and Fulton, S. R. (2003). A hp MITC finite element method
for Reissner-Mindlin plates with smooth boundaries, Technical Report 2003–01, Depart-
ment of Mathematical and Computer Science, Clarkson University.

View publication stats

You might also like