Kirchhoff Plates and Large Deformations

Modelling and C1-continuous Discretization

Jens Ruckert

Dissertation submitted to the

Department of Mathematics
at
Chemnitz University of Technology
in accordance with the requirements for the degree Dr. rer. nat.

Advisor: Prof. Dr. Arnd Meyer

Chemnitz, 24th April 2013

Contents

1 Introduction 1

2 The deformation problem in the three-dimensional space 5

2.1 General differential geometry of deformation in the three-dimensional
space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Equilibrium of forces . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Material laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4 The weak formulation . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3 Newtons method 19
3.1 The modified Newton algorithm . . . . . . . . . . . . . . . . . . . . . 19
3.2 Second linearization of the energy functional . . . . . . . . . . . . . . 20

4 Differential geometry of shells 23

4.1 The initial mid surface . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.2 The initial shell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.3 The plate as an exception of a shell . . . . . . . . . . . . . . . . . . . 25
4.4 Kirchhoff assumption and the deformed shell . . . . . . . . . . . . . . 25
4.4.1 Differential geometry of the deformed shell . . . . . . . . . . . 25
4.4.2 The Lagrangian strain tensor of the deformed plate . . . . . . 27

5 Shell energy and boundary conditions 31

5.1 The resulting Kirchhoff deformation energy . . . . . . . . . . . . . . . 31
5.2 Boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.3 The resulting weak formulation . . . . . . . . . . . . . . . . . . . . . 36

6 Newtons method and implementation 37

6.1 Newton algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
6.2 Finite Element Method (FEM) . . . . . . . . . . . . . . . . . . . . . 41
6.2.1 Bogner-Fox-Schmidt (BFS) elements . . . . . . . . . . . . . . 43
6.2.2 Hsiegh-Clough-Tocher (HCT) elements . . . . . . . . . . . . . 46
6.3 Efficient solution of the linear systems of equation . . . . . . . . . . . 47
6.4 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6.5 Newtons method and mesh refinement . . . . . . . . . . . . . . . . . 52

i
Contents

7 Numerical examples 55
7.1 Plate deflection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
7.1.1 Approximation with FEM using BFS-elements . . . . . . . . . 56
7.1.2 Approximation with FEM using reduced HCT-elements . . . . 63
7.2 Bending-dominated deformation . . . . . . . . . . . . . . . . . . . . . 66
7.2.1 Approximation with FEM using BFS-elements . . . . . . . . . 66
7.2.1.1 1st example: Cylinder . . . . . . . . . . . . . . . . . 66
7.2.1.2 2nd example: Cylinder with further rotated edge
normals . . . . . . . . . . . . . . . . . . . . . . . . . 71
7.2.1.3 3rd example: Mobiusstrip . . . . . . . . . . . . . . . 74
7.2.1.4 4th example: Plate with twisted edge . . . . . . . . . 77
7.2.2 Approximation with FEM using reduced HCT-elements . . . . 80
7.2.2.1 1st example: Partly divided annular octagonal plate 80
7.2.2.2 2nd example: Divided annulus with rotated edge
normals . . . . . . . . . . . . . . . . . . . . . . . . . 84

8 Outlook and open questions 89

Bibliography 91

Notation 97

Theses 99
List of Figures 101

List of Tables 103

ii
Acknowledgement
Financial support
The research I present in this thesis was financially supported by the German Federal
Ministry for Education and Research (BMBF) within the joint research project
GeoMec (grant 05M10OCC). Furthermore I am grateful for the employment at the
Department of Mathematics at Chemnitz University of Technology. Thus I had the
ability to held tutorials and to support lectures.

Personal thanks
First and foremost I like to express my deep gratitude to my supervisor Prof. Dr.
Arnd Meyer. He aroused my interest first for the FEM and later for plate deform-
ations and he was always willing to give me advice and help. He offered me the
opportunity to further my personal development and expertise. In addition, due
to him, I was able to participate in different scientific conferences in Germany and
abroad.
During my years at the university he was the most important person. Without
him I would not have been able to develop this dissertation. Thank you, very much.
Furthermore I am much obliged to Prof. Dr. Bernd Luderer and to Dipl.-Math.
Martin Stocker for their efforts and their time spent in the revision of this thesis
and for their advice.
Although they are not named separately, I thank all my colleagues at the Depart-
ment of Mathematics for the inspiring discussions and explanations, by what I got
an better overview of the mechanical background.
Moreover I owe a debt of gratitude to the members of staff of Fraunhofer ITWM
in Kaiserslautern, who were partners in the mentioned BMBF project GeoMec.
Particularly, I am indebted to Dr. Joachim Linn, whose suggestions and advice
broadened my prospectives concerning my research.
Last but not least, I would express my sincerest gratitude to my whole family
and my friends, who all gave me the support which I needed to author this work.
Especially, I thank my parents Ines and Willy R uckert, which are the best parents I
could imagine. Without them, I would not be the man, who I am, today. In addition
I want to mention my parents-in-law, who gave me the opportunity to shorten my
duty stroke enormously in the last months. Finally, my special thanks I extend to
my wife Nadja, for her patience, for her continuing support and not at least for her
great, infinite and unconditional love. At the end, my warmest words
Acknowledgement

are dedicated thankfully to my little daughter Arja, who brings unlimited pleasure
and happiness into my life.

iv
1 Introduction
In this thesis we present our research as a result of the joint project GeoMec (grant
05M10OCC) supported by the German Federal Ministry of Education and Research
(BMBF). In this project, large deformations of plates and shells are investigated from
different points of views. In that way, one project partner was examining a mathe-
matical theory, where large deformations can be calculated very fast, possibly with
low accuracy. Such methods are used, for example, in the animation of movements,
which becomes an integral part in modern animation movies or in the computer
game industry. Here, the main focus is on the simulation of movements in a way
that it looks good but not necessarily expressing the full mechanical truth.
By contrast to that we pursue the goal of mechanical accuracy as best as pos-
sible. For this approach much more aspects have to be taken into account. So, the
deformation is based on the minimization of the deformation energy. This espe-
cially concerns to some industrial project partners in automotive engineering. Here,
for instance, the bending of car headliners during the manufacturing process is of
interest.
For elasticity and deformations of plates or shells a very large number of literature
exists. The roots of linear plate theory go back far into the past. First approaches
have been made in the 17th century, but with some shortcomings (see [3]). Then
in 1850 Kirchhoff has published his linear plate theory and is considered as the
founder of modern plate theory. Quite nearly a century later, some extensions have
been made by Reissner (1944/45) and Mindlin (1951) to include transverse shear.
In the late seventies and eighties of the last century first non-linear plate theories
were established for example by Ciarlet (see [9]). For that only linear material laws
and so only small deformations were considered. A first approach for the calcula-
tion of large deformation of plates was the von Karman plate. Due to the rapid
development in the computer industry for the last 20 years, it is possible to com-
pute more complex problems considering non-linear theories and non-linear material
laws. Several approaches are found in [7]. In [14] the differential geometry is intro-
duced in a similar way as here, following by a restriction to linearized strain. Here,
the well known split of the total energy into a change of metric part and a change
of curvature part is used. Furthermore, the resulting theory is only adopted to a
cylindrical shell. Another approach one can find in [2]. Here, instead of the normal
of the shell mid surface an arbitrary direction vector is considered first and a non-
linear system of equations representing the equilibrium of forces is established. By
the identification of the direction vector as the normal of the deformed mid surface
the Kirchhoff hypothesis is assumed again. In the end the theory is constrained to
linearized strain.

1
1 Introduction

For our approach the Kirchhoff hypothesis is considered, too. By contrast with
the theories in the cited articles, we do not linearize the strain tensor and maintain
these non-linearities for its derivatives. Hence, the usual Kirchhoff assumption is
the one and only restriction. No other simplifications are done.
Therefore, at the beginning we consider the well-established 3D-theory for large
deformations. We discuss the inherent differential geometry and define the basic
principles like the deformation gradient, the right Cauchy-Green deformation tensor
or the Lagrangian strain tensor.
The equilibrium of forces corresponds to a minimum of an appropriate energy
functional, which is considered here. Its first derivative leads to the weak formulation
of the non-linear boundary value problem. For its solution we have to apply some
linearizations such as Newtons method.
In the third chapter Newtons method is explained and the incremental Newton
algorithm is given for the approximation of a resulting deformation. In this algorithm
the second derivative of the energy functional plays an important role. For the full
3D case it can be calculated comparatively easy.
We discuss the differential geometry for a shell and its simplification to a plate
in the fourth chapter. Generally, in the shell theory the displacement of the mid
surface is considered and then amplified to the whole plate. In our research the
Kirchhoff assumption is used for that. As a result a strain tensor of lower rank
(rank 2) occurs (mapping the tangential space into itself).
Furthermore, the emerging displacement vector depends on the displacement vec-
tor of the mid surface. Therefore, in Chapter 5 we establish the resulting energy
functional, which only depends on this mid surface displacement vector. The bound-
ary conditions are considered for hard and soft clamped shells.
For examining the weak formulation of the plate deformation problem the first
derivative of the energy functional is determined, which is now more complicated
than in the full 3D-theory.
For the resulting energy functional the thickness dependence is not eliminable as
usually in using linear strain, but the unknown vector function U only depends on
( 1 , 2 ), thus not on the thickness parameter . Therefore, we can use the 2D-FEM
for its approximation on the mid surface and we have to integrate numerically over
thickness direction . So, in Chapter 6 we deduce the second derivative of the energy
functional and take a closer examination of the difficulties in its calculation. With
this second derivative the Newton algorithm is established, which is integrated into
the F.E.-solution process. The method and both kinds of considered C1 -elements,
the BFS- and the reduced HCT-element, are explained. A very short excursion in
the field of efficient solvers follows, before we go into details of the implementation.
A short discussion about the use of Newtons method in the approximation process
brings Chapter 6 to an end.
After that we present two deflection examples, one using the BFS-elements and
one the reduced HCT-elements. Subsequently we take a closer look to bending
problems, which was the main focus of this work. For that, in the seventh chapter
four examples of bending-dominated large plate deformations are presented, using

2
 1 Introduction

BFS-elements. These are followed by further two bending-dominated examples

using reduced HCT-elements, simulating really large deformations.
This thesis ends up with a reflection of the main results of our work and a prospect
of further key aspects of research.

3
2 The deformation problem in the
three-dimensional space
In the following we consider vectors or vector fields written as bold capital or lower
case letters, respectively. Furthermore a vector is also called a first-order tensor.
Then we understand second-order tensors as linear maps of first-order tensors onto
first-order tensors, again. For the beginning a pair of first-order tensors such as
A1 A2 is understood as a second-order tensor. In general a second-order tensor is
any linear combination of such pairs.
Below, some linear operations for the tensor calculus are defined:

(1) The mapping properties, applying a second-order tensor onto a (three-dimen-

sional) vector function U or the other way around, respectively:

(A1 A2 ) U = A1 (A2 U )
U (A1 A2 ) = A2 (A1 U ).

(2) The tensor multiplication as a linear mapping of a second-order tensor onto a

second-order tensor, again:

(A1 A2 ) (A3 A4 ) = (A2 A3 )(A1 A4 ).

(3) The trace of a second-order tensor is a scalar function

tr(A1 A2 ) = A1 A2 .

(4) The transpose of a second-order tensor is given by

(A1 A2 )> = (A2 A1 ).

(5) The double contraction of second-order tensors is a scalar function

 
(A1 A2 ) : (A3 A4 ) = tr (A1 A2 ) (A3 A4 )
= (A2 A3 )(A1 A4 ).

In general, the definitions in (1) to (5) are true for every second-order tensor, because
all five definitions above are linear operations.

5
2 The deformation problem in the three-dimensional space

In the same way we use fourth-order tensors mapping second-order tensors onto
second-order tensors via the double contraction (5).
Note that in general a fourth-order tensor is a linear combination of such four-
tuple of vectors or of pairs of second-order tensors.
Throughout this paper capital letters are used for second-order tensors, displayed
in a flowing font. Vectors are written in bold and matrices in underlined letters.
Due to the fact that we consider large deformations, we have to distinguish between
the undeformed and the deformed configuration of the considered body using capital
and lower case letters, respectively. This is mainly used for the basis vectors.

2.1 General differential geometry of deformation in

the three-dimensional space
In this section basic facts of differential geometry for the three-dimensional Euclidean
space are presented, following [19].
A domain 0 in the Euclidean space is parametrized as

0 = {X() : P R3 }, (2.1)

where X denotes the position of a material point in the undeformed body under
consideration. Herein, = ( 1 , 2 , 3 ) is a given, possibly curvilinear coordinate
system. 1 , 2 and 3 are called the coordinates of the point X.
Note that

X = ( 1 , 2 , 3 ) (2.2)

is a unique map between the position of a point and its coordinates.

Thus, for every point X in the domain 0 (2.1) the vectors

Gi = i X, i = 1, . . . , 3, (2.3)

form the co-variant tensor basis.
The elements of the matrix

G = (Gij )3i,j=1 , Gij = Gi Gj , i, j = 1, . . . , 3, (2.4)

are called metric coefficients of the tensor basis.

At any co-variant tensor basis Gi there exists a contra-variant tensor basis Gj
defined in such a way that

Gi Gj = ij , i, j = 1, . . . , 3.

ij describes the so-called Kronecker-symbol. Let Gij = Gi Gj , then the equality

G1 = (Gij )3i,j=1 , is true.

6
 2.1 General differential geometry of deformation in the three-dimensional space

From now on we use the Einstein summation convention. For this, the indices
in this chapter run from 1 to 3. Later on, when the shell theory is considered, the
indices will run from 1 to 2, only.
So, the gradient operator is defined by

Grad = Gi i . (2.5)

If we apply the gradient operator (2.5) to a given function (), we obtain a first-
order tensor Grad = Gi i . In the same way, we have a second-order tensor

i
Grad U () = G i U = Gi U ,i , if we apply (2.5) to the vector function U () 0 .

Here and in the following we use the abbreviation U ,i for the partial derivatives
U , i = 1, . . . , 3.
i
Let a deformation be described by the isomorphism

: 0 7 , x = (X). (2.6)

Then we introduce a vector field U of displacement vectors. The position of a

material point in the deformed body can now be identified as

x = X + U : 0 7 . (2.7)

Here denotes the deformed domain.

The displacement vector U is a vector function at the point X out of the unde-
formed domain 0 (2.1). Due to the dependency of the point X on , the function
U depends on as well. So x() is a parametrization of the deformed domain .
Again, we assume

x = ( 1 , 2 , 3 ) (2.8)

to be a unique mapping between the material point of the deformed body and its
coordinates. Therefore, with the parametrization (2.7) of , we can denote a co-
variant and a contra-variant tensor basis again:
Let

g i = i x, i = 1, . . . , 3,

be the basis vectors belonging to the co-variant tensor basis in . Then, there exists
a contra-variant tensor basis g j , j = 1, . . . , 3, such that g i g j = ij , i, j = 1, . . . , 3.
The co-variant tensor basis of the deformed domain is connected to the co-
variant tensor basis of the initial domain 0 via the displacement vector field U :

g i = i x = i (X + U ) = Gi + i U = Gi + U ,i , (2.9)

7
2 The deformation problem in the three-dimensional space

and the gradient operator in the deformed domain is defined analogously to (2.5):

grad = g i i .

Obviously a Taylor expansion of the isomorphism (2.6) is supposed to be such
that

(X + V ) = (X) + F V + O(kV k2 )

with a linear operator F (a second-order tensor) applied to any (small) direction V .

Furthermore from definition (2.5) we have

(X + V ) = (X) + V Grad () + O(kV k2 ).

Hence,

F = (Grad )> = (Grad (X + U ))>

= [Gi Gi + Gi U ,i ]>
= I + (Grad (U ))>

denotes the so called deformation gradient.

(I is the identity map, the second-order identity tensor.)

With (2.9) we have

F Gk = Gk + U ,i Gi Gk
= Gk + U ,k = g k (2.10)
>
= Gk F .

Hence F = g i Gi F > = Gi gi and F 1 = Gi g i .

The right Cauchy-Green deformation tensor is defined as

C = F > F = I + Grad (U ) + Grad (U )> + Grad (U ) Grad (U )> (2.11)

and the Lagrangian strain tensor as

1
E = (F > F I) (2.12)
2
1 
= Grad (U ) + Grad (U )> + Grad (U ) Grad (U )> .
2
Note that both C and E are symmetric tensors.
Remark 2.1 From (2.10) we have

[g 1 , g 2 , g 3 ] = [F G1 , F G2 , F G3 ] (2.13)
= det F[G1 , G2 , G3 ].

8
 2.1 General differential geometry of deformation in the three-dimensional space

On the strength of (2.2) and of (2.8), respectively, we have

[G1 , G2 , G3 ] 6= 0

and

[g 1 , g 2 , g 3 ] 6= 0.

Hence, with the constants c and c

0 < c det F c P (2.14)

is the main assumption on our deformation.

Remark 2.2 We consider vector fields V () as functions over 0 as well as over

. Such a vector field can be considered to be in both (L2 (0 ))3 and (L2 ())3 ,
respectively, due to (2.14) together with the volume elements

d0 = [G1 , G2 , G3 ] d 1 d 2 d 3

and

d = [g 1 , g 2 , g 3 ] d 1 d 2 d 3 .

Obviously, with (2.13) the volume element of the deformed domain, d, can be re-
presented by the product of the determinant of F with the volume element of the
undeformed domain, d0 .

(1) Hence
 3  3
V () L2 () V () L2 (0 ) , (2.15)

because
Z Z
V V d = V V [g 1 , g 2 , g 3 ] d 1 d 2 d 3

Z
= V V det F [G1 , G2 , G3 ] d 1 d 2 d 3 .
0

With (2.14) we have

kV k2L2 () kV k2L2 (0 )

and both function spaces have the same set of vector functions V ().

9
2 The deformation problem in the three-dimensional space

(2) Furthermore the same is true for H1 vector fields:

 3  3
V () H1 (0 ) V () H1 () .

For the half norm we have

Z
(grad V )> : (grad V ) d

Z
= (F > Grad V )> : (F > Grad V ) det F d0
0
Z  
= tr (Grad V )> (F 1 F > ) (Grad V ) det F d0
0
Z  
= tr (Grad V )> C 1 (Grad V ) det F d0 .
0

Due to equation (2.14), there exist two constants 0 < < such that in sense
of positive definiteness

I F > F = C I

and therefore

1 1
I F 1 F > = C 1 I

are true.
Hence, together with (2.15) we obtain

kV k2H1 () kV k2H1 (0 )

and both function spaces have the same set of vector functions V () again.

2.2 Equilibrium of forces

We assume all force interactions being a result of the volume load and of the surface
load. In the equilibrium of forces all these forces and moments compensate each
other. Hence, after a deformation all forces and moments have to be balanced in
the deformed domain as well. Cauchys theorem guarantees the existence of the
so-called Cauchy stress tensor . The volume load is displayed by the volume force
density p() in such a way that the equation

div (x) + p(x) = 0 x (2.16)

10
 2.2 Equilibrium of forces

is true. Generally, the divergence operator div is defined as

div = g i i .

The proof of Cauchys theorem and some more information can be found, for ex-
ample, in [11].
To set up the weak formulation for the Finite Element Method (FEM) we have
to multiply equation (2.16) with any vector function V and to integrate over the
deformed domain .
Hence, for a special function space
 3 
H := H1 (0 ) with additional boundary conditions
we have to solve the integral equation
Z
div V + p V d = 0 V () H.

Then integration by parts yields
Z Z Z
: grad V > d = p V d + n V dS. (2.17)

Here describes the boundary of the deformed domain and dS the surface
element.
For the usual pullback into the initial configuration 0 we need one of both tensors,
1
the first Piola-Kirchhoff stress tensor T or the second Piola-Kirchhoff stress tensor
2
T with
1 2
(det F) = F T = F T F > .
1 2
While T is an unsymmetrical second-order tensor, the second-order tensor T is
symmetric, which gives some advantages in the tensor analysis.
Let us now take a look at each of the three integrals in (2.17) separately, following
the explanations in [19].
Z Z
: grad V > d = J : grad V > d0
0
Z 1
= (F T ) : (Grad V > F 1 ) d0
0
Z 1
= T : Grad V > d0 (2.18)
0
Z 2
= T : (F > Grad V > ) d0
0
Z 2
= T : (I + Grad U ) Grad V > d0 V () H.
0

11
2 The deformation problem in the three-dimensional space

2
Because of the symmetry of T and from equation (2.18) we have
Z 2
T : (I + Grad U ) Grad V > d0
0
Z 2
= T : [(I + Grad U ) Grad V > ]> d0 (2.19)
0
Z 2
= T : E 0 (U ; V ) d0 V () H.
0

Obviously E 0 (U ; V ) is the Frechet derivative of E(U ) applied to V :

2E(U ) = F > F I = (I + Grad U )> (I + Grad U ) I.

Therefore,

2E(U + V ) = 2E(U ) + Grad V + Grad V > + Grad U Grad V > +

+ Grad V Grad U > + Grad V Grad V >
= 2E(U ) + 2E 0 (U ; V ) + O(kV k2 )

with

2E 0 (U ; V ) = Grad V + Grad V > + Grad U Grad V > +

(2.20)
+ Grad V Grad U > V () H.

For the first integral of the right-hand side of (2.17) we obtain

Z Z
p V d = (det F)p V d0
0
Z (2.21)
= 0 p V d0 V () H.
0

Due to the mass conservation in 0 and , the equation (det F) = 0 holds, where
0 is the density in 0 .

Remark 2.3 Usually the outer acceleration field p is considered to be constant.

Therefore the right-hand side of equation (2.17) is a linear functional over V .
In case that p depends on the position x = X + U we get an additional non-
linearity in the right-hand side. To avoid this extra non-linearity, in the following,
we only use a constant acceleration field.

For the second integral of the right-hand side we have to consider the boundary
conditions.

12
 2.2 Equilibrium of forces

Let
0 = 0,N D
and
= N D
with U (X) = 0 X D . Hence x and X coincide on D (fixed boundary). In
that way, we define
  3 
1
H = V H (0 ) : V = 0 on D .

Then
Z Z
n V dS = n V dS V () H.
N

By fixing one of the coordinates, for instance, 3 = c, we obtain (at least a part of)
the boundary N or 0,N , respectively, with
n o
N = x( 1 , 2 , c) : ( 1 , 2 ) P N .
Here, the vectors g 1 and g 2 of the appropriate co-variant tensor basis act as tangen-
tial vectors of this surface and with
dS = kg 1 g 2 k d 1 d 2
we obtain the surface element. In that way, the normal vector of the deformed mid
surface is given by
g1 g2
n= .
kg 1 g 2 k
Therefore
Z Z
[g 1 , g 2 , V ]
n V dS = kg 1 g 2 k d 1 d 2
kg 1 g 2 k
N PN
Z
= [g 1 , g 2 , V ] d 1 d 2
PN
Z
1 1
= [F G1 , F G2 , F T V ] d 1 d 2 (2.22)
det F
PN
Z 1
= [G1 , G2 , T V ] d 1 d 2
PN
Z 1
= N T V dS 0
0,N

13
2 The deformation problem in the three-dimensional space

with dS 0 = kG1 G2 k d1 d2 being the appropriate surface element and N =

G1 G2 the external normal vector in the original domain at all points
0
kG1 G2 k
X 0,N .
Now we can deduce the weak formulation from equations (2.17), (2.19), (2.21)
and (2.22) as follows:
For a vector function U with U = 0 on D solve
Z 2 Z Z
0
T : E (U ; V ) d0 = 0 p V d0 + ~g V dS 0 V H, (2.23)
0 0 0,N

1
where ~g = N T are the given traction on 0,N .
We are still not able to solve this non-linear equation system in (2.23), due to the
2
fact that, until yet, the Piola-Kirchhoff stress tensor T is unknown. In literature
these equations are called not to be closed. By consideration of the material laws
this closure is done in the next section.

2.3 Material laws

In this thesis only those material laws are considered that map strains E(U ) onto
2
stresses T . For convenience, we utilize non-linear elastic material behaviour, where
2
T is a function of E (or of C). Hence, material laws for devices with plastic behaviour
are excluded, because for them only the change of stresses depends on the change
of strain.
In order to get the weak formulation, we define the energy functional
Z
(U ) = (C(U ))d0 f (U ). (2.24)
0

Herein (C(U )) denotes the energy density depending on the invariants of C. These
invariants we get from the characteristic polynomial
det(C I) = 3 + IC 2 IIC + IIIC .
Here IC , IIC , IIIC are called the principal invariants of C and if 1 , 2 , 3 are the
three eigenvalues of the tensor C, these invariants can be worked out as
IC = tr(C) = 1 + 2 + 3 ,
1h i
IIC = (trC)2 tr(C 2 ) = 1 2 + 2 3 + 3 1 ,
2
IIIC = det C = 1 2 3 .
For convenience, we introduce the functions
1
ak = tr(C k ) (k = 1, . . . , 3) (2.25)
k

14
 2.3 Material laws

and use these ak instead of the principal invariants of C to express material laws,
i.e.

IC = a1 ,
1
IIC = (a1 )2 a2 , (2.26)
2
IIIC = (a31 6a1 a2 + 6a3 )/6.

Minimizing the energy functional (2.24) leads to

0 (U ; V ) = 0 V H. (2.27)

We can calculate (2.27) by the derivation of :

(U + V ) = (U ) + 0 (U ; V ) + O(kV k2 ) V H. (2.28)

Furthermore,
 
(C(U + V )) = C(U ) + C 0 (U ; V ) + O(kCk2 ) V H. (2.29)

Because the strain tensor E is calculated by the Cauchy-Green deformation tensor

C from the equations (2.11) and (2.12), we get

C = 2E + I, (2.30)

where I is the unity gain tensor in the first-order tensor space.

Let us now study

C(U + V ) = C(U ) + C 0 (U ; V ) + O(kV k2 ) V H.

With (2.30) we obviously have

C 0 (U ; V ) = 2E 0 (U ; V ) V H.

Hence, for the energy density in (2.29) we get

(C(U + V )) = (C(U )) + 0 (C(U ); 2E 0 (U ; V )) + O(kEk2 )

2 (2.31)
= (C(U ))+ L: 2E 0 (U ; V ) V H.
2
In mechanical literature the second-order tensor L is usually written as
2
L= .
C
Because of (2.25) and (2.26) we can write this tensor as follows:
3 3
2 X ai X i1
L= = C .
i=1 ai C i=1 ai

15
2 The deformation problem in the three-dimensional space

With the equations (2.28) and (2.31) we get

3
!
Z
i1
0
: E 0 (U ; V ) d0 f (V ) V H.
X
(U ; V ) = 2 C (2.32)
i=1 a i
0

Together with (2.23) the second Piola-Kirchhoff stress tensor is identified as

3
2 X i1
T =2 C (2.33)
i=1 ai

and, therefore, it can be calculated for X 0 (which is needed for the calculation
at the integration points of a FE-discretization, later on).

The materials with linear elastic behaviour represent a special case. For these
materials we assume that there exists a linear map between the Right-Cauchy-Green
deformation tensor and the second Piola-Kirchhoff stress tensor. So the density
function can be calculated separately by (2.33). From Hooks law we have
2
T = 2E + tr(E)I
with the Lame coefficients and . Using (2.30) we get
3
2 1 3 ! i1
T = C 1 + C 0 ( + a1 ) = 2
X
C .
2 2 i=1 ai

Integration by parts leads to

1 3 1 1
(a1 , a2 , a3 ) = a1 ( + ) + a21 + a2 + c
2 4 8 2
with a certain constant c.
If we claim zero energy for zero strain (e.g. C = I), the constant c is fixed and
1 1 3 1 3
 
(a1 , a2 , a3 ) = a21 + a1 + a2 + (2 + 3)
8 2 4 2 8
follows.
As mentioned in the introduction, we want to consider non-linear elastic materials,
but as we have seen, this theory includes linear elastic behaviour as well.
However, there are many non-linear elastic material laws discussed in mechanical
literature, i.e. the Neo-Hook material and some of its generalizations. For the Neo-
Hooke material the density function is specified by
(a1 , a2 , a3 ) = c10 (a1 ln(IIIC ) 3) + D2 ln2 (IIIC ), (2.34)
where c10 and D2 are constants depending on the material itself.
Overall, the material law consists of the definition of a energy density function
2
(a1 , a2 , a3 ), which leads to the second Piola-Kirchhoff stress tensor T (2.33) de-
pending on C (or E).

16
 2.4 The weak formulation

2.4 The weak formulation

In the sections before we have seen that with the equilibrium of forces a energy
functional (2.24) and its derivative (2.32) can be calculated by defining an energy
density and therefore defining the second Piola-Kirchhoff stress tensor (2.33).
Then, with the abbreviations
Z 2
a(U ; V ) = T : E 0 (U ; V ) d0
0

and
Z Z
f (V ) = 0 p V d0 + ~g V dS 0 V () H (2.35)
0 0,N

we can write the well-known weak formulation as:

Find the vector function U with U = 0 X D with

a(U ; V ) = f (V ) V H. (2.36)

Obviously, equation (2.36) coincides with (2.23) and

0 (U ; V ) = a(U ; V ) f (V ) V H (2.37)

holds.

17
3 Newtons method
The purpose of this chapter is to solve the non-linear equation system (2.36) using
Newtons method. Therefore, the second derivative of the energy functional and
the first derivative of the second Piola Kirchhoff stress tensor are needed. Due to
the fact that Newtons method converges in a small ball around the unknown exact
solution only, it is essential to choose a sufficiently well starting approximation. For
small deformations U = 0 as the initial approximation was adequate, but in the case
of large deformations the displacement vector differs very much from zero. Hence,
for Newtons method the well known incrementation of the forces f (U ) is used. In
that way, we get a series of Newton iterations with an incremental increase of the
outer forces f (U ). The proof of convergence for an incremental solving of equation
systems and some more information about are given in, for example, [11].

3.1 The modified Newton algorithm

Newtons algorithm is about an iteration procedure solving non-linear equations or
non-linear equation systems. Hence, computing the solution of the equation system

D(U ) = 0

where D displays an operator, depending on the vector U , means choosing a start

vector U (0) and using repeatedly the iteration code
 1  
U = D0 (U (k) ) D(U (k) )
U (k+1) = U (k) + U
k : = k + 1.

Here, D0 is the Frechet derivative of D.  1

Practically, we do not calculate the inverse D0 (U (k) but consider the lineari-
zation of the operator equation

D0 (U ; U ) = D(U )

in every Newton step.

Generally, the iteration ends if the stopping criterion
kU k
<
kU k

19
3 Newtons method

is fulfilled.
Applying Newtons method to the problem (2.36) means solving (2.37). Then, in
every Newton step we have to calculate U from

00 (U ; U , V ) = 0 (U ; V ) V H.

Herein, with 00 (U ; U , V ) we depict the second linearization of the energy func-

tional (U ) as linearization of 0 (U ; V ):

0 (U + U , V ) = 0 (U ; V ) + 00 (U ; U , V ) + O(kU k2 ).

Newtons method converges quadratically, only in the neighbourhood of the root.

Hence, for ensuring a well-suited starting vector we have implemented an incremental
algorithm, whereby a multiplier t [0, 1] is applied to the force vector f (U ). This
multiplier has to be increased appropriately.

Algorithm 3.1 (Incremental Newton solution process)

(1) Start: U = 0, t := t

(2) for U solve 00 (U ; U , V ) = tf (V ) a(U ; V ) V H

(3) update U := U + U

(4) if (kU k / kU k < ): t := t + t until t = 1

(5) goto (2).

With the algorithm above, a series of Newton methods is applied, due to the
incrementation of the force vector via t.

3.2 Second linearization of the energy functional

We achieve the second derivative of the energy functional by applying the chain rule
to its derivative 0 (U ; V ) from (2.32):
Z 2 Z 2
00 (U ; U , V ) = T 0 : E 0 (U ; V ) d0 + T : E 00 (U ; U , V ) d0 .
0 0

2
For computing the derivative T 0 we have to linearize the second Piola-Kirchhoff
stress tensor itself:
2 2
T (C + C) = T (C) + C : C + O(kCk2 ) (3.1)

with a fourth-order tensor C.

20
 3.2 Second linearization of the energy functional

Furthermore, we consider
2   2  
T C(U + U ) = T C(U ) + 2E 0 (U ; U ) + O(kU k2 ) .

Therefore, (3.1) in conjunction with (2.33) yields

2   2  
T C(U + U ) = T C(U ) + C(U ) : E 0 (U ; U ) + O(kCk2 ).

Obviously, we achieve
2
T 0 = C(U ) : E 0 (U ; U ).

Thereby, the fourth-order tensor C arises from the equations (2.33, 3.1) as
3 X
3
C(U ) = 4
X 2 hC i1 C j1 i + 4 J + 4 C
b (3.2)
ai aj a2 a3
i=1 j=1

with the fourth-order unity tensor J, mapping any second order tensor to itself, and
the fourth-order tensor C
b with the property

b :X =CX +X C
C

for all symmetric second-order tensors X . C b emerges from the expansion of the
square of the right Cauchy Green deformation tensor C 2 (U ).
The second derivative of the Lagrangian strain tensor is easily obtained by the
linearization of its first derivative (2.20):

2E 00 (U ; U , V ) = Grad U Grad V > + Grad V Grad U > V () H.(3.3)

To sum up, due to the symmetry of the fourth-order tensor C and of the second-
order tensor E 0 (U ; U ), the second derivative of the energy functional can be written
as
Z Z 2
0 0
00
(U ; U , V ) = E (U ; U ) : C : E (U ; V ) d0 + T : E 00 (U ; U , V ) d0 .(3.4)
0 0

In the literature of mechanics the tensor C(U ) is called the material tangent. With
(3.4) now the Algorithm 3.1 is complete.

21
4 Differential geometry of shells
In the geometrical characterization of shells we go along with the majority of mech-
anical literature. Therein, an initial shell 0 is characterized as a static, isotropic
three-dimensional object. One regards a mid surface m 3
0 R of the undeformed
shell and a thickness h diam(m 0 ). Now the deformation of the shell is approx-
imated by a deformation of the mid surface.
This chapter begins with the differential geometry of shells in general. Later on,
we consider the initial configuration as a plane shell, called plate. At the end of this
chapter the Kirchhoff assumption is introduced and than combined with the theory
for the plane shell.
In what follows all indices now run between 1 and 2. The content of this chapter
refers to the explanations in [21] and [24].

4.1 The initial mid surface

In this paragraph we illustrate the basic theory of the differential geometry in the
undeformed shell, the initial domain. In the initial configuration all vectors and
matrices, mainly the co- and contra-variant basis vectors and the matrices of the
first and the second fundamental forms, are typed in capital letters.
We denote
m 1 2 1 2 2
0 := {Y ( , ) : ( , ) P 2 R } R
3
(4.1)
as mid surface of the undeformed shell. Here, the points of the surface in the three-
dimensional space are denoted by Y and the coordinates ( 1 , 2 ) run throughout
the parameter domain P 2 . Therewith the tangential vectors are defined by

Ai = i Y , i = 1, 2,

and the surface normal vector is described by
A1 A2
A3 = A3 = . (4.2)
|A1 A2 |
The two tangential vectors together with the surface normal vector establish a co-
variant tensor basis in R3 .
The first fundamental forms A and the second fundamental forms B are written
as (2 2)-matrices:
A = (Aij )2i,j=1 , Aij = Ai Aj ,

23
4 Differential geometry of shells

and the second fundamental forms

!
2
B= (Bij )2i,j=1 , Bij = Y A3 = Ai,j A3 = Ai A3,j ,
i j
respectively. We define the corresponding contra-variant tensor basis from
Aj = Ajk Ak with Aj Ak = kj , where Ajk are the entries of A1 .
The surface element is defined by
1
dS = |A1 A2 | d 1 d 2 = (det A) 2 d 1 d 2 .
Therefore, the gradient operator on the tangential space, the surface gradient, is
specified as
Grad S = Ai i .

4.2 The initial shell

We understand the initial shell as a three-dimensional manifold
1
 
1 2 3 1 2 1 2
0 := X( , , = ) = Y ( , ) + h A3 , ( , ) P 2 , | | <
2
with constant thickness h and A3 from (4.2). Without loss of generality, we use
= 3 as a synonym for the thickness coordinate and consider the coordinates
i , i = 1, 2, providing the length of dimension in the corresponding direction. Then
the basis vectors Ai and Ai , i = 1, 2 have to be dimensionless, consistently, but
A3 = A3 has no dimension, anyway. The co-variant tensor basis (2.3) from chapter
2 can now be calculated as
Gi = i X = Ai + h A3,i i = 1, 2 (4.3)

and
G3 = hA3 .
Therefore the contra-variant tensor basis Gi , i = 1, 2 as well as G3 = h1 A3 =
h1 A3 and, moreover, the matrix of the metric coefficients (2.4) can be calculated.
Obviously,
det(G) = h2 det(G)b

with the (2 2)-matrix

b = (G )2
G ij i,j=1 , Gij = Gi Gj

= A(I h A1 B)2
 
= A I 2h A1 B + h2 2 (A1 B)(A1 B) (4.4)
= A 2h B + h2 2 BA1 B
= (A h B)A1 (A h B).

24
 4.3 The plate as an exception of a shell

Hence, the volume element of the shell can be restated as

1
dV = [G1 , G2 , G3 ] d 1 d 2 d = h det(G)
b 2 d 1 d 2 d.

4.3 The plate as an exception of a shell

As written before, we think of a plate as a shell which is initially plain. After
applying some forces to the plate, which do not cause rigid body motions only, a
shell arises.
Due to the plane initial state of a plate, we get some simplifications. So on the
initial mid surface m
0 in (4.1) the physical points Y are in the e1 -e2 -plane, for
instance

Y ( 1 , 2 ) = e1 1 + e2 2 , (4.5)

where the coordinates i , i = 1, 2, feature the length of the dimension in direction

of each unit vector ei , i = 1, 2. With (4.2) this simplification yields

A3 = e 3

for the normal vector of the undeformed mid surface independent of ( 1 , 2 ). To-
gether with the definition of the co-variant tensor basis (4.3) of the initial shell we
get the co-variant tensor basis with the normal vector for the initial plate as
1
Gi = Gi = ei , i = 1, 2, G3 = he3 , G3 = e3 . (4.6)
h
Because of the independence of the normal vector from the coordinates ( 1 , 2 ),
the second fundamental forms at the mid surface of the initial plate vanish:

B = O.

For simplicity, in the following we restrict ourselves to the plate.

4.4 Kirchhoff assumption and the deformed shell

4.4.1 Differential geometry of the deformed shell
The following assumption is one possibility to reduce the space dimension. The
idea is to describe the deformed shell by its mid surface only. With the Kirchhoff
hypothesis no change in thickness of the shell is allowed and the shell is assumed to
be shear rigid. This means that a certain straight fibre of points
1
 
1 2 1 2
Y ( , ) + h A3 ( , ) : | |
2

25
4 Differential geometry of shells

which is perpendicular to the undeformed mid surface m

0 (4.1) has to be straight
and perpendicular to the deformed mid surface
n o
m 1 2 1 2 1 2
t = y(1 , 2 ) = Y ( , ) + U ( , ) : ( , ) P 2 , (4.7)
after deformation as well. Here, U is the unknown displacement vector of the mid
surface, which is a function of ( 1 , 2 ) as well as of Y .
The Kirchhoff assumption defines the deformed shell as
n o
t = x( 1 , 2 , ) = y( 1 , 2 ) + h a3 , (4.8)
where a3 is the new surface normal vector of the deformed mid surface m
t following
its differential geometry: With
ai = i y = Ai + U ,i , i = 1, 2,

the tangential vectors of the mid surface after deformation, the surface normal vector
of the deformed mid surface m t can be calculated as
a1 a2
a3 = . (4.9)
|a1 a2 |
Consequently, we get
a = (aij )2i,j=1 with aij = ai aj , (4.10)
b = (bij )2i,j=1 with bij = ai,j a3 , (4.11)
the first and second fundamental forms in the deformed state.
Now the 3D co-variant basis is
g i = i x = ai + h a3,i = Ai + U ,i + h a3,i , i = 1, 2 (4.12)

g 3 = ha3 .
Analogously to G b = (gij )2i,j=1 with
b in (4.4), we can define the (2 2)-matrix g
gij = g i g j , yielding
gb = a(I h a1 b)2 = (a h b)a1 (a h b).
Hence, mathematically the Kirchhoff assumption means
x( 1 , 2 , ) : = y( 1 , 2 ) + h a3 ( 1 , 2 )
 
= Y ( 1 , 2 ) + U ( 1 , 2 ) + h a3 U ( 1 , 2 )

1




  (4.13)
+ U ( 1 , 2 ) + h (a3 U ( 1 , 2 ) e3 )

=
2




h
 
= X() + U ( 1 , 2 ) + h (a3 U ( 1 , 2 ) e3 ).

26
 4.4 Kirchhoff assumption and the deformed shell

Accordingly, the new 3D displacement vector is

U KH ( 1 , 2 , ) = U + h (a3 (U ) e3 ), (4.14)

with the vector field U depending on ( 1 , 2 ) only.

4.4.2 The Lagrangian strain tensor of the deformed plate

We deduce the 3D-deformation gradient from (4.3) and (4.12) in connection with
(4.6) as

Fb = g i Gi + a3 A3
= (a1 + h a3,1 )e1 + (a2 + h a3,2 )e2 + a3 e3 .

Then, the coefficients of the right Cauchy-Green deformation tensor (2.11) are given
by

Cb = Fb Fb = cij ei ej + e3 e3 (4.15)

with

cij = g i g j = gij (4.16)

= (ai + h a3,i ) (aj + h a3,j ), i, j = 1, 2.

In this way, the right Cauchy-Green deformation tensor is calculable as

Cb = gij ei ej + e3 e3 (4.17)
h i
= (ai + ha3,i ) (aj + ha3,j ) ei ej + e3 e3
h i
= (ei + U ,i + ha3,i ) (ej + U ,j + ha3,j ) ei ej + e3 e3
h
= ij ei ej + e3 e3 + (U ,i + ha3,i ) ej + ei (U ,j + ha3,j )
i
+ (U ,i + ha3,i ) (U ,j + ha3,j ) ei ej
h
= I + (U ,i + ha3,i ) ej + ei (U ,j + ha3,j )
i
+ (U ,i + ha3,i ) (U ,j + ha3,j ) ei ej .

Thus, with (2.12) we can calculate the Lagrangian strain tensor of the plate as

Eb = ij ei ej (4.18)

with

2ij = gij ij = (ai + h a3,i ) (aj + h a3,j ) ij , i, j = 1, 2.

27
4 Differential geometry of shells

Accordingly with (4.17) we can compute the Lagrangian strain tensor as

h
2E(U
b ) = (U ,i + ha3,i ) ej + ei (U ,j + ha3,j )
i
+ (U ,i + ha3,i ) (U ,j + ha3,j ) ei ej
= ei [(U ,i + ha3,i ) ej ]ej + ei [ei (U ,j + ha3,j )]ej
+ ei [(U ,i + ha3,i ) (U ,j + ha3,j )]ej
= GradS (U + ha3 ) ej ej + ei ei GradS (U + ha3 )>
+ GradS (U + ha3 ) GradS (U + ha3 )> .

In that,

GradS := Gi i , i = 1, 2

defines the surface gradient operator. Then, with the orthoprojector into the tan-
gential space A = ei ei the Lagrangian strain tensor is written as

2E(U
b ) = GradS (U + ha3 ) A + A GradS (U + ha3 )>
(4.19)
+ GradS (U + ha3 ) GradS (U + ha3 )> .

For using matrix syntax we define the matrices containing the coefficients of the
associated tensors in the chosen tensor basis.
In case of the right Cauchy-Green deformation tensor Cb given in (4.15) and (4.16),
the appropriate matrix is

c = (cij )2i,j=1
= [(ai + h a3,i ) (aj + h a3,j )]2i,j=1
= a + h [ai a3,j + a3,i aj ]2i,j=1 + (h )2 [a3,i a3,j ]2i,j=1 . (4.20)

From the definition (4.9) we have

y ,i a3 = ai a3 0, i = 1, 2.

Consequently,

(ai a3 ),j 0 i = 1, 2

is also true.
Hence, by using the product rule and the equation (4.11) we can deduce

0 = (ai a3 ),j = ai,j a3 + ai a3,j = bij + ai a3,j .

In this way, we get

ai a3,j = bij = bji = aj a3,i = a3,i aj , i = 1, 2. (4.21)

28
 4.4 Kirchhoff assumption and the deformed shell

With (4.21) the following equation is true:

a3,i aj = bij i = 1, 2.

Therefore

a3,i = bij aj
= bij aj (ak ak )
= bij (aj ak ) ak
= bij ajk ak .

Consequently, we achieve

a3,i a3,j = (bik akl al ) (bjm amn an )

= bik akl (al an ) anm bmj
= bik akl aln anm bmj
= nk bik anm bmj
= bin anm bmj i, j, k, m, n = 1, 2.

Obviously

(a3,i a3,j )2ij=1 = b a1 b (4.22)

and with (4.20), (4.21) and (4.22) the matrix of the right-Cauchy-Green deformation
tensor can be calculated:

c = a 2h b + (h )2 ba1 b
= (a h b)a1 (a h b).

Analogously, the matrix with the coefficients of the strain tensor (4.18) is

e = (ij )2i,j=1

with

2e = (a h b)a1 (a h b) I. (4.23)

29
5 Shell energy and boundary
conditions
In this chapter we combine the three-dimensional theory for large deformations,
described in the sections 2.2 and 2.3, with the new displacement vector U KH from
(4.14) as a result out of the Kirchhoff hypothesis. Consequently, we achieve a new
a-form used in the variational formulation.

5.1 The resulting Kirchhoff deformation energy

We consider the energy functional (2.24) but use the displacement vector U KH from
(4.14) achieved from the Kirchhoff hypothesis.
Then
Z  3
KH
(U )= (U KH ) dV f (U KH ) with U KH H10 (0 ) , (5.1)
0

where H10 is the Sobolev function space with functions vanishing at the Dirichlet
boundary D .
As seen in (2.35), in the right-hand side of this new energy functional (5.1), for
the boundary forces f (U KH ) we have to integrate over 0 . Within, the integration
over the thickness parameter leads to zero for the second part in U KH (4.14).
Consequently, for the right-hand side in (5.1) we obtain

f (U KH ) = f (U ).

Obviously we get the new energy functional

(U KH ) = (U + h (a3 (U ) e3 ))
Z
= (U + h (a3 (U ) e3 )) dV f (U )
0
Z (5.2)
= (U
b ) dV f (U )
0

=: (U
b ).

Remark 5.1 With the Kirchhoff hypothesis we achieve a three-dimensional dis-

placement vector U KH . This is the analogue to the displacement vector U in the

31
5 Shell energy and boundary conditions

common three-dimensional theory of large deformations given in chapter 2. In what

follows we consider the displacement vector of the mid surface U = U ( 1 , 2 ), not
to be mistaken for the displacement vector in the common three-dimensional theory.

Remark 5.2 With the displacement vector (4.14) from the Kirchhoff hypothesis and
remark 5.1 we have to consider the function space for the displacement vector U of
the mid surface and the arbitrary test functions V , respectively.
Therefore we define U H b where
 3
b := { H2 (m )
H with appropriate boundary conditions} (5.3)
0

is a special function space, due to the Kirchhoff assumption.

Later on in this chapter we consider the reason for using the Sobolev function
3
space (H2 ) and the appropriate boundary conditions.

Analogously to the 3D-displacement vector U in (2.7) out of the common three-

dimensional theory that minimizes the energy functional (2.24), the new displace-
ment vector U ( 1 , 2 ) here minimizes the energy functional b above in such a way
that the deformation energy (2.24) gets minimized over a non-linear set (4.14).
Hence, the first derivative b0 (U ; V ) of the energy functional should vanish for all
virtual displacements V H: b

!
b0 (U ; V ) = 0.

The derivative of this energy functional arises from the variation of ,

b

(U
b + V ) = (U
b ) + b0 (U ; V ) + h.o.t.
2
Here, the second Piola-Kirchhoff stress tensor T (U KH ) is double the derivative of
(U KH ) w.r.t. C. Hence
(U KH )
b0 (U , V ) = V
U ( 1 , 2 )
" 3 #
KH
 
X ak : CKH U
= ak C U U V
k=1
2  
= T (C) : E 0 U KH ; L(V ) V H
b

with
KH
L(V ) := UU V .
The symbol denotes the application of the second part of the chain rule to an
arbitrary vector function V in the way that L(V ) is a linear operator with respect
to V arising from the variation of U KH :
U KH (U + V ) = U KH (U ) + L(V ) + h.o.t. (5.4)

32
 5.1 The resulting Kirchhoff deformation energy

Obviously,

L(V ) = V + h a03 (U ; V ) V H.
b

Let us consider the set

n o
K = V KH : V KH = V + h(a3 (V ) e3 ); V H
b (5.5)

of vector functions being defined in the same way as the displacement vector (4.14),
resulting from the Kirchhoff hypothesis. Hence, with (4.18) the strain tensor E(U )
forall U K is of lower rank (rank = 2) without components of ei e3 or e3 e3 .
Now,
 (5.4) iswritten formally in a correct way but it seems to result in a tensor
0 KH
E U , L(V ) , which is of rank greater than two. Apparently, the derivative L(V )
of the vector function U KH (U ) in direction of a vector V is no longer in K (5.5) and
lest in the considered non-linear set (4.14), over which the energy functional (5.1)
should be minimized.
Contrary, if we consider the first derivative of the energy functional (2.32) directly
and insist on the vector functions to be from K (5.5), we first need the linearization
of the strain tensor E(U
b ) (4.19):
 
2E(U
b + V ) = GradS U + V + ha3 (U + V ) A
 >
+ A GradS U + V + ha3 (U + V )
   >
+ GradS U + V + ha3 (U + V ) GradS U + V + ha3 (U + V )
h     i
= GradS U + ha3 (U ) + GradS V + ha03 (U ; V ) + h.o.t. A
h     i>
+ A GradS U + ha3 (U ) + GradS V + ha03 (U ; V ) + h.o.t.
h     i
+ GradS U + ha3 (U ) + GradS V + ha03 (U ; V ) + h.o.t.
h     i>
GradS U + ha3 (U ) + GradS V + ha03 (U ; V ) + h.o.t.
   >
= 2E(U
b ) + GradS V + ha03 (U ; V ) A + A GradS V + ha03 (U ; V )
   >
+ GradS U + ha3 (U ) GradS V + ha03 (U ; V )
   >
+ GradS V + ha03 (U ; V ) GradS U + ha3 (U ) + h.o.t.

Hence
0  
2Eb (U ; V ) = GradS V + ha03 (U ; V ) A
 >
+ A GradS V + ha03 (U ; V )
   > (5.6)
+ GradS U + ha3 (U ) GradS V + ha03 (U ; V )
   >
+ GradS V + ha03 (U ; V ) GradS U + ha3 (U ) .

33
5 Shell energy and boundary conditions

This linearization of the strain tensor (5.6) is a second-order tensor of rank two with
a03 (U ; V ) being a differential operator, which can be retrieved from the variation of
a3 :

a3 (U + V ) = a3 (U ) + a03 (U ; V ) + h.o.t.
a1 (U + V ) a2 (U + V )
=
ka1 (U + V ) a2 (U + V )k
[a1 (U ) + V ,1 ] [a2 (U ) + V ,2 ]
=
k[a1 (U ) + V ,1 ] [a2 (U ) + V ,2 ]k
a1 a2 + V ,1 a2 + a1 V ,2 + h.o.t.
=
ka1 a2 + V ,1 a2 + a1 V ,2 + h.o.t.k

The abbreviations
b := a1 a2
(5.7)
:= V ,1 a2 + a1 V ,2
v

yield

b+v + h.o.t.
a3 (U + V ) = .
kb + v
+ h.o.t.k

The Taylor expansion for kb + v

k is calculated by
q q
kb + v
k = k2
kb + v kbk2 + 2b v
+ h.o.t
!
bv
kbk 1 +
kbk2

and in this way

!
1 1 bv
kb + v
k 1 .
kbk kbk2

Therefore, we can estimate

!
1 bv
a3 (U + V ) (b + v
+ h.o.t.) 1
kbk kbk2
b
v b(v b)
= + + h.o.t.
kbk kbk kbk3

Consequently
!
1 bb 1
a03 (U ; V )= I v
= (I a3 a3 ) v
(5.8)
kbk kbk2 ka1 a2 k

34
 5.2 Boundary conditions

with (I a3 a3 ) being the orthoprojector onto the tangential space of the deformed
mid surface. In mechanical literature this projector is also called the metric tensor
of the deformed mid surface.
Concluding, we can write the first derivative of the new energy functional as
Z 2 Z Z
0
b0 (U ; V ) = T (C) : Eb (U ; V )d0 0 p V d0 + ~g V dS 0 . (5.9)
0 0 0,N

Until now, we have not considered any boundary conditions of the function space
b Therefore we take a closer look onto this space in the following paragraph.
H.

5.2 Boundary conditions

As one can see in (4.14), U KH contains the normal of the deformed mid surface,
a3 , which is a non-linear differential operator, applied to U . In (5.9) differential
operators are applied to U KH and accordingly to a3 as well. Hence, we need the
3
second partial derivatives in U . Therefore, the Sobolev function space (H20 (m 0 )) is
used in (5.3).
We consider the normal a3 in (4.9) at the boundary of the shell. Hence, one of
the coordinates ( 1 , 2 ) has to be constant. Without loss of generality we consider
1 =h c with 2
i a fixed constant c, defining a boundary part (for all and for all
21 , 12 ). If this boundary part is classified as hard clamped with homogeneous
boundary conditions, then U = 0 and a3 = e3 as can be easily seen in the equation
for the undeformed mid surface of a plate (4.5) combined with the mid surface of
the deformed shell (4.7) and the definition of the deformed shell (4.8) in conjunction
with the definition of the normal vector (4.9). Generalized inhomogeneous bound-
ary conditions mean that U and a3 are appropriately given. Now we evolute the
contained cross product, omitting higher order parts of U :
(e1 +U ,1 )(e2 +U ,2 )
a3 = ka1 a2
a1 a2 k = k(e1 +U ,1 )(e2 +U ,2 )k

(e1 + U ,1 ) (e2 + U ,2 ) = a3 k(e1 + U ,1 ) (e2 + U ,2 )k

e3 + (e1 U ,2 ) + (U ,1 e2 ) = a3 ke3 + (e1 U ,2 ) + (U ,1 e2 )k

e3 + (e1 U ,2 ) (e2 U ,1 ) = a3 ke3 + (e1 U ,2 ) (e2 U ,1 )k .

For a hard clamped boundary part the condition a3 = e3 means


(3) (3)
U,1 U,1 0





(3) = (3) .

U,2


U,2
0



(2) (1) (2) (1)

1 + U,2 + U,1
1 + U,2 + U,1
1

35
5 Shell energy and boundary conditions

Hence,
(3) (3)
U,1 = U,2 = 0
and therefore
(3)
U,12 = 0
are the resulting boundary conditions for the displacement vector.
Consistently, the displacement vector has to be chosen out of the appropriate
function space
 3 
2 (3) (3) (3)
H
b = H (0 ) : U,1 = U,2 = U,12 = 0, X D , j = 1, 2 .

In case of considering a soft clamped boundary part, one out of the two partial
(3)
derivatives U,i , i = 1, 2, is free.

5.3 The resulting weak formulation

We have seen that in comparison with the deformation energy in 2.24 from the
common three-dimensional theory we have to accomplish a further operator L, lin-
ear applied to V , due to the Kirchhoff  assumption. This operator L includes a
differential operator a03 (U ; V ) = a

U V (5.8), applied to virtual displacements
3

V H.
b
In this way
Z 2
0
ab(U ; V ) = T (C) : Eb (U ; V ) dV, V H
b (5.10)
0

and
Z Z
f (V ) = 0 p V d0 + ~g V dS 0 V H.
b (5.11)
0 0,N

Hence, we can establish the new weak formulation:

Find the solution U in the appropriate function space H

b such that

ab(U ; V ) = f (V ) for all V H.

b (5.12)
Additionally, with both equations above, (5.10) and (5.11), for the first derivative
of the energy functional (5.9)
b0 (U ; V ) = ab(U ; V ) f (V ) (5.13)
is true. Therefore, solving (5.12) is equivalent to compute
b0 (U ; V ) = 0.

36
6 Newtons method and
implementation
6.1 Newton algorithm
In the Chapter 3 we have introduced Newtons method and we have defined the incre-
mental algorithm (3.1). Therein the displacement vector U of the three-dimensional
theory is used. We have replaced this displacement vector by the appropriate U KH in
(4.14), resulting from the Kirchhoff hypothesis (4.13). As a consequence, we get the
new energy functional (5.2) after recalculating the right Cauchy-Green deformation
tensor Cb from (4.15) and, therefore, the Lagrangian strain tensor Eb in (4.19). By
reason of this recalculation it takes much more effort to obtain the first linearization
0
Eb (U ; V ) ((5.6) in conjunction with (5.8)) with the vector V so, that U KH (U + V )
is from K, defined in (5.5).
In that way, we obtain the incremental Newton algorithm for this problem as

Algorithm 6.1 (Incremental Newton solution process)

(1) Start: U = 0, t := t

(2) for U solve b00 (U ; U , V ) = tf (V ) ab(U ; V )

(3) update U := U + U

(4) if (kU k / kU k < ): t := t + t until t = 1

(5) goto (2).

Herein the second linearization of the energy functional b00 (U ; U , V ) is needed,

which we obtain as the linearization of equation (5.13). The chain rule, used in
equation (5.10), implies
Z 2 2
0 00
b00 (U ; U , V ) = T 0 (U ; U ) : Eb (U ; V ) + T (C)
b :E
b (U ; U , V ) dV. (6.1)
0

Here
2 0 0 0
T 0 (U ; U ) : Eb (U ; V ) = Eb (U ; U ) : C(U ) : Eb (U ; V ) (6.2)

37
6 Newtons method and implementation

using Cb instead of C for the material tangent C(U ) (3.2).

Hence, for the calculation of the second linearization of the energy functional
b00 (U ; U , V ) we have to compute the second linearization of the strain tensor
00
Eb (U ; U , V ). Therefore, we have to linearize the first linearization of the strain
0
tensor Eb (U ; V ) (5.6) in connection with (5.8). For this purpose we get
0  
2Eb (U + U ; V ) = GradS V + ha03 (U + U ; V ) A
  >
+ A GradS V + ha03 (U + U ; V )
    >
+ GradS U + U + ha3 (U + U ) GradS V + ha03 (U + U ; V )
    >
+ GradS V + ha03 (U + U ; V ) GradS U + U + ha3 (U + U )
     
= GradS V + ha03 (U ; V ) + GradS ha003 (U ; U , V ) + h.o.t. A
     >
+ A GradS V + ha03 (U ; V ) + GradS ha003 (U ; U , V ) + h.o.t.
     
+ GradS U + ha3 (U ) + GradS U + ha03 (U ; U ) + h.o.t.
     >
GradS V + ha03 (U ; V ) + GradS ha003 (U ; U , V ) + h.o.t.
     
+ GradS V + ha03 (U ; V ) + GradS ha003 (U ; U , V ) + h.o.t.
     >
GradS U + ha3 (U ) + GradS U + ha03 (U ; U ) + h.o.t.
0     >
=2Eb (U ; V ) + GradS ha003 (U ; U , V ) A + A GradS ha003 (U ; U , V )
    >
+ GradS U + ha3 (U ) GradS ha003 (U ; U , V )
    >
+ GradS ha003 (U ; U , V ) GradS U + ha3 (U )
    >
+ GradS U + ha03 (U ; U ) GradS V + ha03 (U ; V )
    >
+ GradS V + ha03 (U ; V ) GradS U + ha03 (U ; U ) + h.o.t.

In that way
0  
2Eb (U ; U , V ) = GradS ha003 (U ; U , V ) A
  >
+ A GradS ha003 (U ; U , V )
    >
+ GradS U + ha3 (U ) GradS ha003 (U ; U , V )

38
 6.1 Newton algorithm

    >
+ GradS ha003 (U ; U , V ) GradS U + ha3 (U )
    >
+ GradS U + ha03 (U ; U ) GradS V + ha03 (U ; V )
    >
+ GradS V + ha03 (U ; V ) GradS U + ha03 (U ; U ) .
Obviously, for the completion of the calculation of the second derivative of the
energy functional b00 (U ; U , V ) the second linearization of the normal vector of the
deformed mid surface a003 (U ; U , V ) has to be calculated.
In addition to the abbreviations (5.7) we define
g := U a + a U .
U ,1 2 1 ,2

Furthermore, the first derivative of the normal of the deformed mid surface is already
given in (5.8). Then, the variation of the orthoprojector in (5.8) yields
 
I a3 (U + U )a3 (U + U )
= (I a3 a3 ) a3 (U )a03 (U ; U ) a03 (U ; U )a3 (U ) + h.o.t.
Hence, its linearization is
a3 (U )a03 (U ; U ) a03 (U ; U )a3 (U )
1 g 1 (I a a ) U
= a3 (I a3 a3 ) U 3 3
ga
3
kbk kbk
1 h g g i (6.3)
= a3 U + U a3 (a3 Ug )a a (U
3 3
g a )a a
3 3 3
kbk
1 h i
= 2(a3 U
g )a a a U
3 3 3
g U
ga .
3
kbk
Therefore, the variation of the first derivative of the normal of the deformed mid
surface a03 (U ; V ) is obtained as
I a3 (U + U )a3 (U + U )
a03 (U +U ; V ) =
ka1 (U + U ) a2 (U + U )k
(V ,1 a2 (U + U ) + a1 (U + U ) V ,2 )
1 
= (I a3 a3 ) a3 (U )a03 (U ; U )
b + U
g + h.o.t.

a03 (U ; U )a3 (U ) + h.o.t. (ve + V ,1 U ,2 + U ,1 V ,2 )
!
1 g b 
U
= 1 (I a3 a3 ) a3 (U )a03 (U ; U )
kbk kbk2

a03 (U ; U )a3 (U ) + h.o.t. (ve + V ,1 U ,2 + U ,1 V ,2 )
!
1 1
= (I a3 a3 ) ve + (I a3 a3 )
kbk kbk
1 
(V ,1 U ,2 + U ,1 V ,2 ) a3 (U )a03 (U ; U )
kbk

39
6 Newtons method and implementation

 g b
U
+ a03 (U ; U )a3 (U ) ve (I a3 a3 ) ve + h.o.t.
kbk3
= a03 (U ; V ) + a003 (U ; U , V ) + h.o.t.
In that way, the second linearization a003 (U ; U , V ) can be identified as
1
a003 (U ; U , V ) = (I a3 a3 ) (V ,1 U ,2 + U ,1 V ,2 )
kbk
1  
a3 (U )a03 (U ; U ) + a03 (U ; U )a3 (U ) ve
kbk
g b
U
(I a3 a3 ) ve
kbk3
1
= (V ,1 U ,2 + U ,1 V ,2 )
kbk
1
+ a3 a3 (U ,2 V ,1 + V ,2 U ,1 )
kbk
1 h i
(6.3) + 2 2(a3 U )a3 a3 a3 U U a3 v
g g g e
kbk
1 g  
2 (U a3 ) (I a3 a3 ) v
e
kbk
1
= (V ,1 U ,2 + U ,1 V ,2 )
kbk
1  
+ [U ,2 , V ,1 , a3 ] + [V ,2 , U ,1 , a3 ] a3
kbk
2 1 g
+ 2 (a3 U )(a3 v)a3 (U v)a
g e 3
kbk2
e
kbk
1 1 g
2 (v a3 )U
g (U a3 )ve
kbk2
e
kbk
1
+ (a3 U
g )(a v)a
3 e 3
kbk2
1
= (V ,1 U ,2 + U ,1 V ,2 )
kbk
1 1 g
2 (v a3 )U
g (U a3 )ve
kbk2
e
kbk
(6.4)
1 
+ [U ,2 , V ,1 , a3 ] + [V ,2 , U ,1 , a3 ]
kbk
3 1 g 
+ (a3 U
g )(a v)
3 e ( U v)
e a3 .
kbk kbk
With the linear functional
1
1 (U ) = U a3
kbk2

40
 6.2 Finite Element Method (FEM)

and the bilinear functional

1

2 (U ,V ) = [U ,2 , V ,1 , a3 ] + [V ,2 , U ,1 , a3 ]
kbk
3 1 g

+ (a3 U )(a3 v)
g e (U v) e
kbk kbk

equation (6.4) is transformed into

1
a003 (U ;U , V ) = (V ,1 U ,2 + U ,1 V ,2 )
kbk
1 (U f (V
g )V
1
f )U
g + (U , V )a .
2 3

In that way, equation (6.1) in conjunction with (6.2) is now calculable as

Z
0 0
b00 (U ; U , V ) = Eb (U ; U ) : C(U ) : Eb (U ; V ) dV
0
Z 2 (6.5)
00
+ T (C)
b :E
b (U ; U , V ) dV.
0

We have seen, that the calculation of the second derivative of the Lagrangian strain
00
tensor Eb (U ; U , V ) is much more labour-intensive as for the appropriate derivative
in the 3D theory (3.3).

6.2 Finite Element Method (FEM)

The incremental Newtons method is embedded into the Finite Element Method
(FEM), that approximates the infinite dimensional problem (5.12) above with a
finite dimensional sub-space Vh of piecewise low order polynomials pk . Therefore,
the solution of the weak formulation (5.12) is discretized by a finite dimensional
solution U h out of this sub-space. The ansatz functions pk are defined on a mesh
Th of non overlapping finite elements T . This mesh of elements completely covers
the given domain,

m
[
0 = T, Ti Tj = Ti , Tj Th , i6=j.
T Th

In general, finite elements are triangles or quadrilaterals for R2 and for example
tetrahedrons or hexahedrons for R3 , respectively.
A finite element mesh is called a conform mesh, if every side of a finite element
Ti Th is either part of the boundary or a side of another finite element Tj Th .
We speak of isotropic finite elements T , if there exists a constant c such that

hT cT

41
6 Newtons method and implementation

with the element diameter hT and T the diameter of the inner element circle or of
the inner element sphere, respectively. A finite element mesh is called shape regular,
if it only consists of isotropic finite elements.
Now, with h = maxT Th hT we consider the biggest element diameter of the mesh
Th . Then, we call a mesh quasi uniform, if the constants c1 and c2 exist such that

c1 h hT c2 h.

In that way, for example locally refined meshes are not quasi uniform.

FEM is used to approximate the solutions. Hence, we consider equation (6.5). In

the first integral we have the first derivative of the strain tensor, depending linearly
on U and once again linearly on V . The material tangent C depends only on U .
In the second integral of (6.5) the second derivative of the strain tensor depends
linearly on U and V at once.
Therefore, we now consider the sub-space Vh with its ansatz functions pi , forming
a basis of Vh . Then the displacement vector U as well as the directions V and U
are displayed in Vh by

3N 3N 3N
Uh =
P
pi ui , V h =
P
pi vi , U h =
P
pi dui . (6.6)
i=1 i=1 i=1

Here, the entire number of degrees of freedom in the sub-space Vh is denoted by

3N . Let du = (du1 , du2 , . . . , du3N )> and u = (u1 , u2 , . . . , u3N )> . Then (5.12) is
approximated in the finite dimensional sub-space Vh via Newtons method as

Algorithm 6.2 (Discretized Newton solution process for one fixed increment t)
 3N
(1) b = tf (pi ) a(U h , pi ) ,
i=1
" #3N
0 0 2 00
Eb (U Eb (U +T : Eb (U
R
(2) K(U h ) = 0 h ; pj ) :C: h ; pi ) h ; pj , pi ) d0
i,j=1

(3) solve K(U h ) du = b

(4) update u := u + du, until = kduk / kuk 108 .

As written before, the ansatz functions pi are piecewise polynomials in m 0 w.r.t.

Th . Obviously, the vector valued ansatz functions pi are pj e1 , pj e2 , pj e3 , (j =
1, . . . , N ) and the scalar functions pj are defined elementwise via the form functions
qk on T

pj |T =: qk,T .

42
 6.2 Finite Element Method (FEM)

With this form functions qk,T we can define the element stiffness matrix KT (U h )
and the element right hand side bT as

1
Z2 Z
0 0
K T (U h ) = Eb (U h ; qj,T ek ) : C : Eb (U h ; qi,T el )
12 T
NT
(6.7)
3
2 00
+T : Eb (U h ; qj,T ek , qi,T el ) dS 0 hd ,
i,j=1 k,l=1
 NT 3
bT = f (qi,T el ) a(U h , qi,T el ) .
i=1 l=1

3NT is the number of the local degrees of freedom in the element T . After calculating
the element stiffness matrix and the element right hand side in every element T Th
the stiffness matrix K(U h ) and the right hand side b are assembled in the common
way.
After calculating the solution on the actual mesh, a mesh refinement may possibly
improve the approximation. Then, the interpolated solution of the coarser mesh is
the starting solution for the calculation on the refined triangulation.

Due to the problem (5.12) with functions in H2 , the conformal FE-approach requires
C 1 -continuity of the ansatz functions pj as proved in [5].
For axis-parallel rectangular elements, the Bogner-Fox-Schmidt elements can be
used and are considered. For a more general shape of the plate some generalizations
using triangular meshes are known, such as the (reduced) Hsiegh-Clough-Tocher
element or similar.

6.2.1 Bogner-Fox-Schmidt (BFS) elements

Using the FEM in the ( 1 , 2 )-coordinate system, we apply BFS elements for our
numerical experiments, in that the mid surface of the plate is a polygon with its
boundaries parallel to the 1 - and to the 2 -axis. Due to quasi uniform mesh refine-
ment, the boundaries of all finite elements in the FEM-mesh remain parallel to the
axes, too.
One typical axis-parallel element T should be represented by the four nodes
! ! ! !
1 1 1 1
x1 = 12 ; x2 = 22 ; x3 = 32 ; x 4 = 42 .
1 2 3 4

As written in the section above, we use ansatz functions with a local support.
ab
For that purpose, in the BFS-element T 16 bi-cubic form functions qk,T , k =
1, . . . 4, a, b = 0, 1 have to be defined. To avoid the definition of these form func-
tions on every finite element we consider a reference element Tb [1, 1]2 R2 , also

43
6 Newtons method and implementation

called master element. We map Tb = [1, 1]2 to T simply by

1
!
1 h 0
1

x= = x

b + x0 , (6.8)
2

2
0 h2

with the master coordinates xb = ( 1 , 2 )> . Furthermore,

!
h1
:= x2 x1 = x4 x3 ,
0

!
0
:= x4 x1 = x3 x2 ,
h2
and
1
x0 = (x1 + x2 + x3 + x4 )
4
is the element mid point.
As proved in [5], for a C 1 -continuous FE-approach we need bi-cubic form functions
in the master element Tb . For this purpose, we consider the four cubic functions paL
and paR (a = 0, 1), displayed in Figure 6.1.

16
0.8
p0L (s) p0R (s)
0.6

0.4

p1L (s)
0.2

0 -
1 p1R (s) 1 s
0.2

1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1

Figure 6.1: third-order polynomials.

These functions are given by

1
p0L : [1, 1] 7 R; p0L (s) := (1 s)2 (2 + s),
4
1
p1L : [1, 1] 7 R; p1L (s) := (1 s)2 (1 + s), (6.9)
4
pR : [1, 1] 7 R; pR (s) := p0L (s),
0 0

p1R : [1, 1] 7 R; p1R (s) := p1L (s).

44
 6.2 Finite Element Method (FEM)

The derivatives of the cubic functions (6.9) are

 0 3
p0L (s) = (s + 1)(s 1),
4
 00 3
p0L (s) = s,
2
 0 1
p1L (s) = (3s + 1)(s 1),
4
 00 1
p1L (s) = (3s 1).
2
Note, that the values of these polynomials and their derivatives in (1) are
 0
p0L (1) = 1, p0L (1) = 0, p0L (1) = 0,
 0  0
p1L (1) = 0, p1L (1) = 1, p1L (1) = 0

Analogously, the values of paR (s), (a = 0, 1) at s = (1) as well as the values of

their derivatives at s = (1) can be calculated.
Now, we define the bi-cubic form functions on the master element Tb as
qb1ab ( 1 , 2 ) := paL ( 1 )pbL ( 2 ),
qb2ab ( 1 , 2 ) := paR ( 1 )pbL ( 2 ),
(6.10)
qb3ab ( 1 , 2 ) := paL ( 1 )pbR ( 2 ),
qb4ab ( 1 , 2 ) := paR ( 1 )pbR ( 2 ).
Finally, with the mapping (6.8) we define the form functions for the element T as
a b
1 1
   
qkab (x)
= h1 h2 qbkab (xb). (6.11)
2 2
Herein, the factors are needed, due to the transformation of the derivatives
2
i = bi .
hi
With (6.10) as well as the abbreviations

i := i , := (1 , 2 )> and i j := 2 (6.12)

i j
for these form functions (6.11) the following properties are obtained:

qk00 (xi ) = ik , qk00 (xi ) = (0, 0)> , 1 2 qk00 (xi ) = 0,

qk10 (xi ) = 0 , qk10 (xi ) = (1, 0)> ik , 1 2 qk10 (xi ) = 0,

qk01 (xi ) = 0 , qk01 (xi ) = (0, 1)> ik , 1 2 qk01 (xi ) = 0,

qk11 (xi ) = 0 , qk11 (xi ) = (0, 0)> , 1 2 qk11 (xi ) = ik .

45
6 Newtons method and implementation

Hence,

c d qkab (xl ) = kl ca db

for all k, l = 1, . . . , 4 and for all a, b, c, d = 1, 2.

For the rearrangement of the expansion coefficients u = (u1 , u2 , . . . , u3N ) of U h

from (6.6), we define the N = 4n vectors uab 3
k R as

uab a b
k = (1 ) (2 ) U h (xk ), k = 1, . . . , n; a, b = 0, 1, (6.13)

wherein n denotes the total number of nodes in the actual triangulation.

Then, in every element the displacement vector is approximated by
4 X
1
uab ab
X
U h (x)|T = k qk (x),
k=1 a,b=0

with uab ab ab ab
1 , u2 , u3 and u4 the nodal values of T after resorting their indices. Hence,
for the BFS-element we have 12 degrees of freedom per node and globally

3N = 3 4 n = 12n.

Let us now consider the approximation of U . Then, with the same rearrangement
(6.13) of the appropriate expansion coefficients du R3N on each element we have

4 X
1
duab ab
X
U h (x)|T = k qk (x).
k=1 a,b=0

6.2.2 Hsiegh-Clough-Tocher (HCT) elements

As written above, the BFS elements are used for example with edges, parallel to
the axis. For all other plate examples, especially for plates with curved boundary,
triangle elements are more suitable. An appropriate C 1 -continuous triangle element
is the HCT element. Here, we have three degrees of freedom at each node. These are
the values of the finite element function U and its partial derivatives. Furthermore
the normal derivative at the mid point of each edge of the element, which is a
quadratic polynomial, is specified. For the reduced HCT element we insist on the
normal derivative to be linear along the edges of the element. For our numerical
experiments we have implemented the reduced HCT element.

The form functions are not polynomials (as usual) in the whole element T , but
polynomial within three sub-triangles. Hence, we subdivide T into three sub-
triangles Tk . Therefore, an arbitrary internal point x0 T has to be connected

46
 6.3 Efficient solution of the linear systems of equation

with the vertices xj , (j = 1, . . . , 3) of the element T . In the literature the cen-

ter of gravity is preferred, due to the effect that the calculation of the Jacobian is
simplified. Hence,
3
1X
x0 = xj .
3 j=1

We define three form functions

qia , a = 0, 1, 2,

associated with the vertex xi , which are cubic inside of each sub-triangles fulfilling
the following properties:

qia are C 1 continuous over T ,

qia (xj ) = ij 0a , qia (xj ) = ij (1a 2a )> ,

q a is linear along each edge of T , with n the outer normal vector of the
n i
considered edge.

As result we get nine nodal base functions qia . For further information we refer to
[22].

6.3 Efficient solution of the linear systems of

equation
It is well known that the arising tangential stiffness matrix K(U h ) is very ill con-
ditioned even in the linear elastic case (which can be understood as K(0) in the
very first step of Newtons method). From the character of the bi-harmonic plate
equation together with the local basic functions, a condition number proportional
to h4 is expected.
Here, we solve these linear systems with a preconditioned conjugate gradient
method, where the appropriate preconditioner is the key for obtaining an efficient
procedure. It should be noted that we use a multilevel technique at least for the
construction of the mesh, that is given as a coarse mesh of some relatively large
rectangles, which are successively subdivided into 4 quarters from level to level.
This hierarchical mesh generation procedure is successfully used for preconditioning
the linear system by hierarchical techniques.
The ansatz functions of the BFS-elements as described above fulfil a refinement
formula, such that the 4-tuple of functions belonging to one node xk :
10 01 11 coarse
(p00
k , pk , pk , pk )

47
6 Newtons method and implementation

for a fixed level l of the mesh refinement, are linear combinations of the next finer
functions
10 01 11 f iner
(p00
i , pi , pi , pi )

for nodes xi around the coarse-mesh node xk of the next level (l + 1).
Based on this refinement formula a BPX-preconditioner [6] has been investigated
in [17],[27] and [28] for the case of a linear Koiter-shell equation. All the details of
the refinement and the resulting base transformation are found in [27].
Here, we have used a slightly simpler approach, a hierarchical basis preconditioner
due to [29], which is a simple matrix multiply C 1 = Q J Q> , where Q is three times
the (N N )-matrix of basis transformation of the hierarchical basis of the finest
FE-space into the given nodal basis. (This matrix multiplication is extremely cheep
from using the refinement formula above.)
The diagonal matrix J contains important scaling factors, which are different for
the 4 functions pab
k (a, b = 0, 1) on each node xk . These factors can be obtained dur-
ing the mesh refinement from the diagonals of the actual stiffness matrices together
with the nodal hierarchy used in the Q-multiply.
So, these ingredients would remain valid when our uniform mesh refinement is
replaced by an adaptive regime.

6.4 Implementation
In this section we want to give an overview of the ideas for the realization of the the-
oretical aspects from the last chapters. The basic principles of the implementation
for computing the element stiffness matrix will follow.
For simplicity we restrict ourselves to BFS-elements. For the reduced HCT-
element the implementation is nearly the same, but the number of form functions
and the dimension of the element stiffness matrix differs. The mathematical pro-
cedures are equal.
For the implementation of the element stiffness matrix K T from (6.7) the most
0
interesting second-order tensors are Eb (U ; V ) in the first part and E(U
b ; U , V ) in
the second part of the integral. Note, that these tensors are of rank 2. There-
fore, their Cartesian coefficients are displayed as the (2 2)-matrices e0 (U , V ) and
e00 (U ; U , V ), respectively. These matrices are obtained by applying the chain rule
to (4.23):
 2
e0 (U ; V ) = ij
i,j=1
   
= a0 (U ; V ) b0 (U ; V ) a1 (U ) a(U ) h b(U ) (I)
    (6.14)
a(U ) h b(U ) a1 (U )a0 (U ; V )a1 (U ) a(U ) h b(U ) (II)
   
+ a(U ) h b(U ) a1 (U ) a0 (U ; V ) b0 (U ; V ) . (III)

48
 6.4 Implementation

Obviously, the matrix of the first derivative of the strain tensor is a result of multi-
plications of the matrix of the first fundamental forms a(U ) from (4.10), the matrix
of the second fundamental forms b(U ) from (4.11) as well as the matrix of their lin-
earization a0 (U ; V ) and b0 (U ; V ), respectively. These linearizations are obtained by
utilizing the chain rule to (4.10) and (4.11), respectively. For the first fundamental
forms we achieve

a0ij (U ; V ) = ai (U ) V ,j + aj (U ) V ,i , i, j = 1, 2. (6.15)

For the second fundamental forms b we define

  1
2
D(U ) = det a

and obtain
0
b0ij (U ; V ) = D0 (U ; V )ebij (U ) + D(U )ebij (U ; V ) (i, j = 1, 2) (6.16)

with
0
b ij (U ; V ) = [a1 , V ,2 , U ,ij ] + [V ,1 , a2 , U ,ij ] + [a1 , a2 , V ,ij ], i, j = 1, 2
e

and D0 (U ; V ) a scalar function. With [, , ] the scalar triple product is displayed.

For a simple implementation and due to the fact that we have symmetric (2 2)-
matrices only, we introduce Voigts notation. With that a (2 2)-matrix is written
as a vector with three components. In that way, the vector notation of the matrix
of the first derivative of the strain tensor (6.14) is
 >
E0 (U ; V ) = 11 (U ; V ), 22 (U ; V ), 212 (U ; V ) .

Let us consider the (3 3)-matrices

2e>
1 z1



2U >
,1 z1



DA (z) =
2e>
2 z2

and DA (U , z) =
2U >
,2 z2
.
(6.17)


U> >

e> >
1 z2 + e2 z1 ,1 z2 + U ,2 z1

With these linear operators, applied to from (6.12), the first derivative of the first
fundamental forms can be computed in Voigts notation by

A0 (U ; V ) = DA ()V + DA (U , )V .

49
6 Newtons method and implementation

Now, let us consider the operators

> >


(a1 a2 ) z1 z1



(U ,11 aa ) zb



DB (z) =
>
(a1 a2 ) z2 z2
and DBa (U , zb ) =
,
(U ,22 aa )> zb (6.18)



(a1 a2 )> z1 z2 (U ,12 aa )> zb

which are (3 3)-matrices, too. Then, using the abbreviations from (6.12), the
0
vector notation of eb (U ; V ) is computable as
0
e (U ; V ) = D ()V + D (U , )V D (U , )V .
B B B1 2 B2 1

Hence, the vector notation of the first derivative of the strain tensor E0 (U ; V ) is a
linear combination of the components of the first derivative of the first fundamental
forms in vector notation A0 (U ; V ) and the one of the second fundamental forms
B0 (U ; V ), following the matrix multiplication rules, using A(U ) as the vector nota-
tion of the first fundamental forms and B(U ) the one of the second fundamental
forms, respectively.
In that way, the first part of the integral in (6.7)
Z
0
Eb (U ; V 1 ) : C(U ) : E(U
b ; V 2 ) dV
0

is the same as
Z
E0 (U ; V 1 )> C E0 (U ; V 2 ) dV (6.19)
0

with the corresponding (3 3)-matrix C = C(U ) of the material tangent C.

For the BFS element T the directions V i (i = 1, 2) in (6.19) have to pass through
the 48 form functions qkab (x)ei , k = 1, . . . , 4, a, b = 0, 1, i = 1, . . . , 3, with qkab (x)
from (6.11) and ei the unity vectors in the Euclidean space. Let Q be the row
vector of the 16 values of the form functions qkab (x) and I the (3 3)-unity matrix.
Furthermore, with

D(U , )V = E0 (U , V )

the derivation operator D(U , ) defines all the linear combinations of the derivation
operators (6.17) and (6.18) for calculating E 0 (U , ), following the matrix calculus.
Then, there exists the (3 48)-matrix

M = D(U , )(Q I)

50
 6.4 Implementation

with denoting the Kronecker product. With this matrix M above, the first part
of the element stiffness matrix (6.19), written in Voigts notation, yields
Z
M > C(U ) M dV.
0

As written before, for the second part of equation (6.7) for the element stiffness
matrix K T the matrix of the second linearization of the strain tensor is needed. The
reapplication of the chain rule to (6.14) yields
   
e00 (U ; U , V ) = a00 (U ; U , V ) b00 (U ; U , V ) a1 (U ) a(U ) h b(U )
   
a0 (U ; V ) b0 (U ; V ) a1 (U )a0 (U ; U )a1 (U ) a(U ) h b(U ) (I)
   
+ a0 (U ; V ) b0 (U ; V ) a1 (U ) a0 (U ; U ) b0 (U ; U )
   
a0 (U ; U ) b0 (U ; U ) a1 (U )a0 (U ; V )a1 (U ) a(U ) h b(U )
   
+ a(U ) h b(U ) a1 (U )a0 (U ; U )a1 (U )a0 (U ; V )a1 (U ) a(U ) h b(U )
   
a(U ) h b(U ) a1 (U )a00 (U ; U , V )a1 (U ) a(U ) h b(U ) (II)
   
+ a(U ) h b(U ) a1 (U )a0 (U ; V )a1 (U )a0 (U ; U )a1 (U ) a(U ) h b(U )
   
a(U ) h b(U ) a1 (U )a0 (U ; V )a1 (U ) a0 (U ; U ) b0 (U ; U )
   
+ a0 (U ; U ) b0 (U ; U ) a1 (U ) a0 (U ; V ) b0 (U ; V )
   
a(U ) h b(U ) a1 (U )a0 (U ; U )a1 (U ) a0 (U ; V ) b0 (U ; V ) (III)
   
+ a(U ) h b(U ) a1 (U ) a00 (U ; U , V ) b00 (U ; U , V ) .

Obviously, the matrix of the second linearization of the strain tensor is a result of
multiple matrix multiplications of the matrices of the first and second fundamental
forms from (4.10) and (4.11), respectively, with their first linearizations from (6.15)
and (6.16) as well as their second derivatives. These second derivatives we obtain
by applying the chain rule to their first derivatives, respectively. Hence,

a00ij (U ; U , V ) = U ,i V ,j + U ,j V ,i (i, j = 1, 2)

and
0
b00 (U ; U , V ) = D00 (U ; U , V )eb(U ) + D0 (U ; V )eb (U ; U )
0 00
+ D0 (U ; U )eb (U ; V ) + D(U )eb (U ; U , V )

with

00
b (U ; U , V
e ) = [V ,1 , U ,2 , U ,ij ] [V ,2 , U ,1 , U ,ij ] + [V ,2 , U ,ij , a1 ]

51
6 Newtons method and implementation

2
[V ,ij , U ,2 , a1 ] + [V ,ij , U ,1 , a2 ] [V ,1 , U ,ij , a2 ] .
i,j=1

D00 (U ; U , V ) is a scalar function arising from the linearization of D0 (U ; V ) in

direction U .
Thus, with the second integral of the element stiffness matrix for two fixed direc-
tions V and U we have to integrate over a matrix multiplication
2 2
T ij e00ji (U ; U , V ),
X

i,j=1

2 2  2 2
with T = T (U ) = T ij the coefficients of the second Piola-Kirchhoff stress
i,j=1
 2
tensor as well as e00 (U ; U , V ) = e00ij (U ; U , V ) .
i,j=1
For calculating the element stiffness matrix the vectors V and U now pass
through all 48 vector valued form functions of the BFS element. Therefore, every
component Hij = e00ij (U ; Q I, Q I) is understood as a (48 48)-matrix, itself.
Hence, the element stiffness matrix (6.7) is
Z Z 2 2
>
X
K T (U ) = M C(U ) M dV + T ij H ji dV. (6.20)
0 0 i,j=1

The integration over the undeformed plate 0 in (6.20) is realized in two steps.
First, the integration over the undeformed mid surface m 0 is accomplished by a
Gaussian integration with 4 4 points and weights. For the integration over the
thickness numerical experiments have shown that we achieve good results in a ac-
ceptable time for the Gaussian integration with 7 points and weights.
For each Gaussian point the matrix multiplications above, have to be done. Hence,
in the realization of the theory we reduce the calculations to the Kronecker products,
which are needed. After that, we multiply them with all scalars, which result from
the matrix multiplications. In that way, the performing time for the assembly of
the element stiffness matrices has been reduced to less than a fourth in contrast to
apply the full matrix multiplications in (6.20).

6.5 Newtons method and mesh refinement

As mentioned before, the calculation of our plate deformations require Newtons
Algorithm 6.2 for two reasons:
(1) As generally known, Newtons method converges only locally. Hence, for large
deformations an incrementation of the forces (possibly of the inhomogeneous
Dirichlet boundary conditions) is used in Algorithm 6.2, analogously to Al-
gorithm 6.1. In the case that the algorithm diverges, the increment has to be
decreased. For a well suited increment the method certainly converges. This
issue is proved, for example, in [11].

52
 6.5 Newtons method and mesh refinement

(2) After the forces are fully applied to the considered device on a relatively coarse
mesh, we may start mesh refinements to obtain an appropriate approximation
of displacements, stresses or strains. Then, in every refinement stage we apply
Newtons Algorithm 6.2 on the actual finite element mesh. Hence, on a mesh
of level l we achieve an approximated vector function U h . After the next
refinement is performed, on the finer mesh with the level (l + 1) we need an
initial vector function U h/2 (x) for the continuation of Newtons method. For
that, the vector function U h (x) is interpolated with the ansatz functions of
the finer mesh and we claim pointwise

U h/2 (xi ) U h (xi ) (6.21)

at all nodes xi of the finer mesh. Now, on this finer mesh Newtons Algorithm
6.2 starts with the interpolated

vector function U h/2 (x). Note, that the stiff-
ness matrix K U h/2 (x) now is of a higher dimension than it was on the
coarser mesh of level l, due to the increased number of elements.
For BFS-elements we achieve nested function spaces for the ansatz functions,

V1 V2 . . . Vl V(l+1) .

Therefore, the ansatz functions in the mesh of level l can be displayed as linear
combinations of the ansatz functions of the next finer mesh. For that purpose
the demand (6.21) is not only fulfilled pointwise at xi , but the vector functions
are obviously identical.
Unfortunately, the mesh refinement does not generate nested function spaces
in the case of HCT-elements. That means

Vl 6 V(l+1) . (6.22)

Here, the vector function U h/2 (x) defined by (6.21) in the finer mesh differs
from the vector function U h (x) in the mesh of level l. Therefore, the interpo-
lated vector function U h/2 (x) as a starting vector function in Algorithm 6.2 is
not as good as in the case of BFS-elements. Consistently, Newtons method
needs some more steps in the HCT-case.
Additionally from using shape regular triangles, the HCT-elements are more
flexible and may be used for approximating curved boundaries. In this case,
after the refinement the new boundary nodes are corrected to the true boun-
dary position (another reason for (6.22) and for more complications with New-
tons method).
There are some approaches in [13] or [26] for generalizing HCT-elements in a
way that nested function spaces for the ansatz functions arise. But for these
elements more degrees of freedom have to be regarded.

53
7 Numerical examples
In this chapter we present some numerical examples. For that, we consider plates
in different forms and sizes. The goal of the theory, which was presented in the
last chapters, is the calculation of large deformations for thin plates. Hence, for all
devices a thickness is chosen, which is one eightieth of the half of the sum of its
width and its length.
Furthermore, we use the Neo-Hooke material law for all plates. The density
function of this material was already defined in (2.34).
The examples are grouped by their kind of plate deformation as membrane-
dominated and bending-dominated examples. For the deflection of a plate we apply
a volume force, perpendicular to the mid surface. The second group consists of
examples with bending-dominated plate deformations, where one or more edges will
follow a prescribed movement.
The procedure of the simulation in nearly all examples is the same. First of all the
forces are applied to the plate incrementally. For each increment a Newton iteration
has to be done. The incremental stage is followed by a number of mesh refinements
to improve the approximation.
For all examples we present some pictures in different approximation stages of
the plates mid surface. With the exception of the pictures of the initial states, the
surfaces are coloured by the Frobenius norm of the mid surface strain tensor or by
one component of the displacement vector. The colour bar is adapted particularly to
each example. For this postprocessing the open source software ParaView, version
3.98.0 nightly, was used.

A word of caution on the graphical output: Note, that for the display
of the deformed mid surface we have chosen the following simplification.
Even though the obtained solutions are piecewise (bi)cubic functions,
the figures only display their linear interpolation with the nodal values
(additionally calculated nodal derivatives are ignored). Especially for
coarser meshes the reader should keep in mind this simplification for the
graphics only the calculated functions are correctly C1 -continuous!

7.1 Plate deflection

For this work the main focus of attention lies on the group of bending-dominated
examples. Nevertheless, for the other group of examples approximating membrane-
dominated deformations Newtons method converges very fast.

55
7 Numerical examples

However, a further word of caution on the results of these deflection examples

should be said. In addition to the error obtained by the FEM, which decreases
due to the mesh refinement, the error resulting from the Kirchhoff hypothesis has
a significant effect. Due to this assumption, the plate is assumed to be shear rigid
in thickness direction and its thickness does not change during the deformation
process. If you consider small deformations, it does not matter too much. But for
large strain the displacements are not as large approximated as they would be in
the reality. Hence, the plate is considered to be more inflexible then it truly would
be. For the deflection examples this is a serious problem, because here the in-plane-
stresses are much higher than they are for bending-dominated problems. Hence, the
displacement vector is approximated too short. The true value of this work can be
seen particularly in the application of the theory for bending-dominated examples.
However, we present one deflection example for each kind of finite elements to
show the excellent convergence of the Algorithm 6.1.

7.1.1 Approximation with FEM using BFS-elements

In the first example we consider a quadratic plate with sides of length 4. The
thickness is fixed to 0.05 over the full length of the plate. All four boundaries are
hard clamped. Figure 7.1 displays the FE-mesh of the undeformed mid surface
consisting of 16 uniform elements.
2 -axis

1-axis

Figure 7.1: Undeformed mid surface, 16 elements.

We apply a volume force vector in normal direction A3 to the plate. As written

before, this force vector is not applied at once but it is increased incrementally
1
(ten increments for t = 10 in Algorithm 6.1). The stopping criterion for each

56
 7.1 Plate deflection

Newton iteration is determined to = 108 . For this the Newton iteration ends,
if the calculated correction term U is smaller than 108 of the actual computed
displacement vector. Figure 7.2 shows the deformed mid surface after one Newton
iteration in the first increment.
strain
1.3045
1.2

0.8

0.4
2 -axis

1-axis

Figure 7.2: Deformed mid surface after applying the first increment, 16 elements.

Expectedly, the values in Figure 7.2 of each element are still small and do not
differ very much over the different elements.

strain
1.3045
1.2

0.8

0.4
2-axis

1 -axis

Figure 7.3: Deformed mid surface after applying the full force vector, 16 elements.

57
7 Numerical examples

Note that for the graphical outputs a constant colour bar for the values of the
norm of the strain tensor is defined to make the colouring for all pictures of one
example comparable.
Following the algorithm, the increment would be increased until the full force
vector is applied. The deformed mid surface, approximated in the coarse mesh, is
illustrated in Figure 7.3. As can be seen, the interpolated value of the strain in the
elements in the center of the surface increased a little bit but the resolution of the
mid surface is far to coarse for a good approximation.
The proceeding before, was necessary to execute the full force to the plate. For the
improvement of the approximation we adopt five mesh refinements. The following
Table 7.1 shows the number of Newton steps per iteration in the single stages.

1st increment, 16 elements : 15

2nd increment, 16 elements : 4

3rd increment, 16 elements : 4

4th increment, 16 elements : 3

..................

10th increment, 16 elements : 3

1st refinement, 64 elements : 5

2nd refinement, 256 elements : 4

3rd refinement, 1.024 elements : 4

4th refinement, 4.096 elements : 3

5th refinement, 16.384 elements : 3

Table 7.1: Example 7.1.1; number of Newton steps at the single periods.

The higher number of Newton steps in the first iteration is easily explained by
the used starting vector U 0 = 0. Obviously, it is not as close to the solution in the
first increment as later the approximated values from the Newton iteration of the
incremental step before, which are used as initial solution for the Newton iteration
of the next increment. The number of Newton steps in the single refinement stages
decreases, due to the fact, that the error for the approximation of the displacement
vector in the finer mesh is smaller than in the coarser triangulation.

58
 7.1 Plate deflection

Note, that the number of Newton steps in each iteration is small, by reason of the
fact that we get nested function spaces for BFS-elements.

strain
1.3045
1.2

0.8

0.4

0
3-axis

1 -axis

Figure 7.4: Deformed mid surface, 16.384 elements, full force applied.

strain
1.3045
1.2

0.8

0.4
2 -axis

1-axis

Figure 7.5: Deformed mid surface, 16.384 elements, full force applied, birds eye
view.

At this point, let us consider the deformed mid surface after five refinements,
approximated by 16.384 elements. As seen in Figure 7.4 and Figure 7.5 the highest

59
7 Numerical examples

strain occurs in the center of the surface boundaries. Due to the symmetry of the
example, the side view in Figure 7.4 looks identically for all four sides.
Apart from the performance test of Newtons method, a second thought for con-
sidering this example above is the verification of our code. For that, we consider the
third component of the displacement vector in Figure 7.6.

U^3
1.1756001
1
0.75
0.5
0.25
2 -axis

1 -axis

Figure 7.6: Deformed mid surface, 16.384 elements, third component of displace-
ment vector U (3) .

Here, a very small boundary layer is observable as it is typical for large deflection
examples.
In contrast to that let us consider a small deformation example. For that purpose
we apply a very small force vector to the plate. Due to that, we do not need to
apply the forces incrementally. Figure 7.7 shows the resulting deformed mid surface
after five refinements, which looks as similar as in the small deformation case.
As expected, in Figure 7.7 the values of the strain tensor as well as for the dis-
placement vector are very small compared to the appliance of large deformations
(Figure 7.4).
Another difference towards the results of large deformation are the boundary
layers. So, in Figure 7.4 they have a steeper slope than in Figure 7.7. This fact is
also seen directly in Figure 7.6 in comparison with Figure 7.8, which both display the
third component of the displacement vector in the case of large and of small strain,
respectively. In Figure 7.6 the coloured ranges get finer towards the boundary. On
the contrary, in Figure 7.8 the coloured ranges are thicker at the boundary, in order
to get finer in direction to the center of the plate, before they grow wider near this
center again.

60
 7.1 Plate deflection

strain
6.107e-5
6e-5

4e-5

2e-5

0
3 -axis

1-axis

Figure 7.7: Deformed mid surface, 16.384 elements, small strain.

U^3
0.0028221

0.002

0.001
2 -axis

1 -axis

Figure 7.8: Deformed mid surface, 16.384 elements, small strain, third component
of the displacement vector U (3) .

Last, we consider the first and the second component of the displacement vector
for small deformation, seen in Figure 7.9 and 7.10, respectively. In addition to the
symmetry of the displacements in the corresponding direction 1 or 2 , the values
are very close to zero.
In contrast to this results U 1 and U 2 are obviously zero in the small strain theory.

61
7 Numerical examples

With the use of our code, the first and the second component of the displacement
vector differ from zero, because of the non-linear parts, which are not discarded as
in the small deformation theory.

U^1
1.0382e-5
8e-6
4e-6
0
-4e-6
-8e-6
2 -axis

-1.038e-5

1 -axis

Figure 7.9: Deformed mid surface, 16.384 elements, small strain, first component of
the displacement vector U (1) .

U^2
1.0382e-5
8e-6
4e-6
0
-4e-6
-8e-6
2 -axis

-1.038e-5

1 -axis

Figure 7.10: Deformed mid surface, 16.384 elements, small strain, second compo-
nent of the displacement vector U (2) .

62
 7.1 Plate deflection

7.1.2 Approximation with FEM using reduced HCT-elements

For approximating the deflection using reduced HCT-elements, we consider a circular

slab with a diameter length of 2. The thickness of the plate is determined to 0.025
and we consider the boundary to be hard clamped. As in the example above, a
volume force perpendicular to the mid surface is applied using ten increments.

strain
0.43587

0.1

0.01
2-axis

0.005

1 -axis

Figure 7.11: Deformed mid surface after applying first increment, 24 elements,
birds eye view.

The deformed mid surface after the first Newton iteration is shown in Figure 7.11,
displayed with 24 elements. The values for the strain norm are very small. As before
in the corresponding mesh of the quadrilateral plate, the mesh is much to coarse for
a well suited approximation.
For comparison, Figure 7.12 depicts the view heading to the ( 1 , 3 ) - plane of
the deformed mid surface in the same mesh as seen in Figure 7.11. In that way, the
third component of the displacement vector U (3) is shown at the same time.
Following the algorithm, at this point the increments are increased and for every
increment a Newton iteration is computed. After 10 increments the full force vector
is now applied and Figure 7.13 displays the deformed mid surface in this stage.
For approximating the displacement vector good enough, the mesh is refined four
times, uniformly.

63
7 Numerical examples

strain
0.43587

0.1

0.01
0.005
3 -axis

1-axis

Figure 7.12: Deformed mid surface after applying the first increment, 24 elements.

strain
0.43587

0.1

0.01
0.005
3 -axis

1-axis

Figure 7.13: Deformed mid surface after applying the full force vector, 24 elements.

The deformed mid surface after all refinement steps is shown in Figure 7.14,
above. It consists of 6.144 elements. Just like in the example for the BFS-elements
the boundary layers are small, as it is typical for this kind of large deformations.
Figure 7.15 displays the birds eye view of the deformed mid surface in the finest
triangulation. In both figures we can see that the maximum of the norm of the
strain tensor is approximated near the boundary.

64
 7.1 Plate deflection

strain
0.43587

0.1

0.01
0.005
3 -axis

1-axis

Figure 7.14: Deformed mid surface, 6.144 elements, full force applied.

strain
0.43587

0.1

0.01
2-axis

0.005

1 -axis

Figure 7.15: Deformed mid surface, 6.144 elements, full force applied, birds eye
view.

In Table 7.2 the number of Newton steps per iteration is displayed in the respective
situation.

65
7 Numerical examples

1st increment, 24 elements : 7

2nd increment, 24 elements : 4

3rd increment, 24 elements : 4

4th increment, 24 elements : 3

.................

10th increment, 24 elements : 3

2nd refinement, 96 elements : 5

3rd refinement, 384 elements : 6

4th refinement, 1.536 elements : 4

5th refinement, 6.144 elements : 4

Table 7.2: Example 7.1.2; number of Newton steps at the single periods

The number of Newton steps per iteration in the single increment stages is simi-
larly small as in the example before. However, we observe a slightly increase of the
numbers in the individual refinement stages. This is due to the refinement technique
of the HCT-elements. Furthermore, we have a curved boundary. Nevertheless, the
algorithm converges, due to the fact that the boundary is hard clamped and only
the lateral adjustment has to be put to the mid point of each outer edge.

7.2 Bending-dominated deformation

As written at the beginning of the last subsection, the main goal of this work is a fast
approximation of large bending deformation problems. For this kind of problems
the error emerging from the Kirchhoff hypothesis is considered to be small.

7.2.1 Approximation with FEM using BFS-elements

7.2.1.1 1st example: Cylinder
For the first example for bending dominated problems we consider a plate with its
mid surface m 0 = [0, 1.5] [0, 1] and a thickness of 0.015625. The initial mesh
consists of 8 elements, as can be seen in Figure 7.16.

66
 7.2 Bending-dominated deformation

2-axis

1-axis

Figure 7.16: Undeformed mid surface, 8 elements.

As mentioned before, for this group of examples we do not apply any force vector
to the plate but simulate an incremental transport of a part of the boundary.
Here, both boundaries in 2 -direction at 1 = 0 and 1 = 1.5, respectively, are hard
clamped with inhomogeneous Dirichlet type boundary conditions. The remaining
edges are left free (do nothing condition).
For the left edge (0, 2 , 0)T we apply a prescribed displacement

U = (0.75(1 cos t), 0, 0.75 sin t)T (7.1)

and for the right edge (1.5, 2 , 0)T

U = (0.75(1 cos t), 0, 0.75 sin t)T , (7.2)

h i
respectively, with t 0, 2 . The outer normals of the edges will be rotated through-
out their transport and are defined by

U ,1 = ((1 cos 2t), 0, sin 2t)T ,

respectively.
The edge displacements are applied in 10 equidistant steps of t. The result of the
first increment is displayed by Figure 7.17.

67
7 Numerical examples

strain
0.051177

0.04

0.02
3 -axis

2 xi
s
1-axis -a

Figure 7.17: Deformed mid surface after applying the first increment, 8 elements.

Increasing the increments leads to a further bending, as we can see in Figure

7.18. Therein, the fully deformed mid surface, approximated with 8 elements is
displayed.

strain
0.051177

0.04

0.02
3 -axis

2 xi
s
1-axis -a

Figure 7.18: Deformed mid surface, 8 elements.

At that, as usual for the algorithm, we start the mesh refinement. After every
refinement a Newton iteration is carried out.

68
 7.2 Bending-dominated deformation

For the full simulation progress the number of Newton steps per Newton iteration
in every incremental stage and in the refinement states are displayed in Table 7.3.

1st increment, 8 elements : 14

2nd increment, 8 elements : 13

3rd increment, 8 elements : 8

4th increment, 8 elements : 8

5th increment, 8 elements : 7

6th increment, 8 elements : 6

7th increment, 8 elements : 5

8th increment, 8 elements : 4

9th increment, 8 elements : 6

10th increment, 8 elements : 5

1st refinement, 32 elements : 3

..................

5th refinement, 8.192 elements : 3

6th refinement, 32.768 elements : 2

Table 7.3: Example 7.2.1.1; number of Newton steps in the single periods.

Obviously, in contrast to the example in Section 7.1.1 the iteration numbers in-
creases a little bit for each individual incremental stage. This may depend on the
fact that for bending deformations the change of the displacement vector and of its
partial derivatives from one to the next incremental state is much higher than in
the deflection examples.
In the refinement stage the number of Newton steps per iteration are expectedly
small, due to the nested function spaces of the ansatz functions for the refined
triangulations.
As result of the six refinements we achieve a cylindrical shell. Its mid surface is
shown in Figure 7.19, which consists of 32.768 elements. Furthermore, the strain is

69
7 Numerical examples

much less than in the deflection examples.

strain
0.051177

0.04

0.02
3 -axis

2 xi
s
1-axis -a

Figure 7.19: Deformed mid surface, 32.768 elements.

For a graphical depiction of the approximation process the graphical output for
each incrementation and the one for the finest mesh is displayed in the following
figure (Figure 7.20).

strain
0.051177

0.04

0.02
3 -axis

2 xi
s
1-axis -a

Figure 7.20: Cylinder, deformation progress.

70
 7.2 Bending-dominated deformation

7.2.1.2 2nd example: Cylinder with further rotated edge normals

In the second example we consider the same plate as in example 7.2.1.1, before, with
equal boundary conditions. The transport of the two outer edges in 2 -direction
follows the same prescribed path, as given in (7.1) and (7.2).
In contrast to example 7.2.1.1 we apply a further rotation of the outer normals of
both edges, which now is defined by

U ,1 = ((1 cos 3t), 0, sin 3t)T ,

respectively, with the increment t [0, 2 ].

We apply the transport of the edges and their simultaneous rotation in ten incre-
ments. In Figure 7.21 the deformed mid surface, approximated with eight elements,
after the implementation of the first increment is seen.

strain
0.058776

0.04

0.02
3 -axis

2 x
is
1-axis -a

Figure 7.21: Deformed mid surface, 8 elements, first increment.

After applying further nine increments, the deformed mid surface looks like in
Figure 7.22, displayed still with a mesh consisting of eight elements.

71
7 Numerical examples

strain
0.058776

0.04

0.02
3 -axis

2 x
is
1-axis -a

Figure 7.22: Deformed mid surface, 8 elements, full force applied.

At this point of the simulation, the complete edge transport and the full edge
rotation is computed. For a useful approximation of the deformation the initial
mesh for the mid surface has to be refined. As in the example before, six refinement
steps are carried out. The graphical result can be seen beneath in Figure 7.23.

strain
0.058776

0.04

0.02
3 -axis

2 x
is
1-axis -a

Figure 7.23: Deformed mid surface, 32.768 elements.

Here, we have plotted the mid surface with 32.768 elements. The mid-surface
strain is slightly higher than in the example before, due to the further rotation

72
 7.2 Bending-dominated deformation

of the edge normals. But in contrast to the corresponding value in the deflection
example 7.1.1 the coloured value here is much smaller.
In the following Table 7.4 the number of Newton steps per iteration are displayed.

1st increment, 8 elements : 14

2nd increment, 8 elements : 7

3rd increment, 8 elements : 5

4th increment, 8 elements : 7

5th increment, 8 elements : 6

6th increment, 8 elements : 5

7th increment, 8 elements : 6

8th increment, 8 elements : 7

9th increment, 8 elements : 5

10th increment, 8 elements : 5

1st refinement, 32 elements : 11

2nd refinement, 128 elements : 6

3rd refinement, 512 elements : 3

..................

6th refinement, 32.768 elements : 3

Table 7.4: Example 7.2.1.2; number of Newton steps in the single periods.

The number of Newton steps in each single incremental situation does not differ
very much from the corresponding number of steps in example 7.2.1.1. Obviously,
the further edge rotation has only a small influence on the numbers of Newton
steps for these stages. By comparison to the example before, the numbers in the
refinement stages, respectively, increase a little bit, in particular for the first two
refinement states. This shows that the approximation of the displacement vector and
its partial derivatives in the coarser meshes is still far away from the approximations

73
7 Numerical examples

in the finer triangulations.

7.2.1.3 3rd example: M

obiusstrip

For the next example we consider a new, longer plate with a mid surface
m0 = [0, 64][2, 2] and a thickness of 0.425. The boundary conditions are compar-
able to the one in Example 7.2.1.1 above. In contrast to that example, additionally
to the boundary edge transport and the boundary edge normal rotation, a second
rotation with an angle () around the 1 -axis is defined to both boundary edges
of the mid surface in 2 -direction.
Therefore, a further inhomogeneous Dirichlet boundary condition arises.
The transport and the rotation of the two edges as well as the rotation of the outer
normals are simulated at once. Due to that, we have twice as much increments as
in all examples before, applying the full prescribed displacement vectors to the two
boundaries. Using 20 increments, for the mid surface of the plate after the first
Newton iteration we get the slightly deformed mid surface.
To guarantee the convergence of all Newton iterations during the computing pro-
cess, we apply two mesh refinements. So, in Figure 7.24 the slightly deformed mid
surface approximated with 256 elements is displayed.

strain
0.043872
0.04

0.03

0.02

0.01
2-axis

1-axis

Figure 7.24: Deformed mid surface, 256 elements, first increment.

After applying all increments we achieve the fully deformed mid surface, approx-
imated with 256 elements in Figure 7.25.

74
 7.2 Bending-dominated deformation

strain
0.043872
0.04

0.03

0.02

0.01
2-axis

1-axis

Figure 7.25: Deformed mid surface, 256 elements, full force applied.

Subsequently, three further refinements are carried out. In Figure 7.26 we can see
that the computed mid-surface strain is very small over the whole plate.

strain
0.043872
0.04

0.03

0.02

0.01
2-axis

1-axis

Figure 7.26: Deformed mid surface, 16.384 elements.

Although we apply the transport as well as both rotation angles for the edges
in 2 -direction at the same time, it seems that the simulation does not need more
Newton steps per Newton iteration prior reaching the stopping criterion as in the
examples before. Certainly, the smaller increments have a positive influence on the

75
7 Numerical examples

number of Newton steps per Newton iteration. Because we have chosen a mesh,
which was twice refined before the incrementation takes place, we have a very fast
convergence in the refinement stage as can also be seen in the following Table 7.5.

1st increment, 256 elements : 4

2nd increment, 256 elements : 7

3rd increment, 256 elements : 6

4th increment, 256 elements : 6

5th increment, 256 elements : 4

6th increment, 256 elements : 4

7th increment, 256 elements : 5

..................

15th increment, 256 elements : 5

16th increment, 256 elements : 4

17th increment, 256 elements : 7

18th increment, 256 elements : 3

19th increment, 256 elements : 6

20th increment, 256 elements : 11

3rd refinement, 1.024 elements : 3

4th refinement, 4.096 elements : 2

5th refinement, 16.384 elements : 2

Table 7.5: Example 7.2.1.3; number of Newton steps in the single periods.

At the end of this example we show the deformation progress in Figure 7.27, in
that the approximated mid surface in all increment stages as well as the most refined
mid surface is displayed.

76
 7.2 Bending-dominated deformation

strain
0.043872
0.04

0.03

0.02

0.01
2-axis

1-axis

Figure 7.27: M
obiusstrip, deformation progress.

7.2.1.4 4th example: Plate with twisted edge

This example considers the same initial plate as it was considered in Example 7.2.1.3,
before. But we change the boundary conditions. For the left edge of the mid surface
the values for the boundary displacement vector and its derivatives are fixed to zero,
except for its first component U (1) , which is left free.

strain
0.11832
3-axis

0.1
0.075
0.05
0.025

1-a
xis
s
xi
-a
2

Figure 7.28: Deformed mid surface, 256 elements, first increment.

77
7 Numerical examples

For the right boundary edge (64, 2 , 0) of the mid surface we apply inhomogeneous
Dirichlet boundary conditions so that a rotation around the 1 -axis until 4 is carried
out.
This rotation is applied with 100 increments, which is much more than in all
examples, above. The deformed mid surface approximated with again 256 elements
in the first increment is displayed in Figure 7.28.

strain
0.11832
3-axis

0.1
0.075
0.05
0.025

1-a
xis

s
xi
-a
2
Figure 7.29: Deformed mid surface, 256 elements, full rotation of 4 applied.

strain
0.11832
3-axis

0.1
0.075
0.05
0.025

1-a
xis
s
xi
-a
2

Figure 7.30: Deformed mid surface, 16.384 elements.

78
 7.2 Bending-dominated deformation

After the application of all increments the fully deformed mid surface, consisting
of 256 elements, can be seen in Figure 7.29.
After further three refinements the deformed mid surface now is approximated
with 16.384 elements. It is to be seen in Figure 7.30.
As before, we observe the norm of the strain tensor being small over the whole
plate. However, the maximum of the strain tensor is nearly twice the corresponding
value in the other bending-dominated examples. The largest strain is generated at
the free boundary edges in 1 -direction. Expectedly, the left edge in 2 -direction is
shifted around 1.5 in direction of the 1 -axis.

1st increment, 256 elements : 4

..................

41st increment, 256 elements : 4

42nd increment, 256 elements : 5

..................

66th increment, 256 elements : 5

67th increment, 256 elements : 4

..................

100th increment, 256 elements : 4

3rd refinement, 1.024 elements : 6

4th refinement, 4.096 elements : 5

5th refinement, 16.384 elements : 5

Table 7.6: Example 7.2.1.4; number of Newton steps in the single periods.

The number of Newton steps per Newton iteration is relatively constant 4 or 5

in the incremental stage, as can be seen in Table 7.6, due to the large number of
increments, we had to choose for applying the full rotation angle. The constant
numbers of Newton steps in each mesh refinement may are indicators for heavier
problems in computing a larger rotation of edges around a vector, which is not
directed along the considered edge. In contrast to that, we could observe a decreasing
number of Newton steps for at least the last two refinements to three or even two

79
7 Numerical examples

in all examples before.

7.2.2 Approximation with FEM using reduced HCT-elements

7.2.2.1 1st example: Partly divided annular octagonal plate
Starting point for the approximation of the deformation for this example is the
octagonal mid surface with a diameter of 2 in both, 1 - and 2 -direction. The initial
triangulation of the mid surface consists of 12 elements as can be seen in Figure
7.31.
2-axis

1 -axis

Figure 7.31: Undeformed mid surface, 12 elements.

The mid surface of the plate is cut along the edge from (0, 0.5, 0)> to (0, 1, 0)> ,
yielding two edges with inhomogeneous Dirichlet type boundary conditions. Only
the outer normals of these both boundaries are rotated throughout the simulation
by
 >
U ,1 = (1 cos t), 0, sin t (7.3)
h i
with the incremental parameter t 0, 2 . The full rotation angle is applied in 10
incremental steps. The remaining edges are left free.
After applying the first increment, we get a slightly deformed mid surface, in that
the outer normals of the emerged boundary edges have rotated a little bit. Due to
that, the outer boundary is bending upwards. We have applied one mesh refinement
in order to secure the convergence of Newtons method during the incremental stage.
The mid-surface strain is very small in this stage, which can be seen in Figure 7.32.

80
 7.2 Bending-dominated deformation

strain
0.026407

s
axi
0.01

1- 0.0026407
is
3 -ax

2 -axis

Figure 7.32: Deformed mid surface, 48 elements, 1 increment.

Now we have to increase the increment until the full rotation angle is applied. In
that way, we get a heavily deformed mid surface, which is displayed in Figure 7.33.

strain
0.026407
s
axi

0.01
1-

0.0026407
is
3 -ax

2 -axis

Figure 7.33: Deformed mid surface, 48 elements, full force applied.

After three consecutive refinements we obtain a subdivided mid surface with 3.072
elements. As seen in Figure 7.34 the norm of the strain tensor is bounded by about
0.02.

81
7 Numerical examples

strain
0.026407

s
axi
0.01

1-
0.0026407
is
3 -ax

2 -axis

Figure 7.34: Deformed mid surface, 3.072 elements, full force applied.

Another refinement step and a further solved Newton iteration yields the deformed
mid surface, displayed with 12.288 elements in Figure 7.35.

strain
0.026407
s
axi

0.01
1-

0.0026407
is
3 -ax

2 -axis

Figure 7.35: Deformed mid surface, 12.288 elements, full force applied.

Here, the maximum of the norm of the strain tensor is about 0.0264. This is much
less than in the deflection example in Section 7.1.2.
For the approximation progress the following numbers of Newton steps were
needed for the Newton iteration in every incremental and every refinement stage.

82
 7.2 Bending-dominated deformation

1st increment, 12 elements : 12

1st increment, 48 elements : 8

2nd increment, 48 elements : 7

3rd increment, 48 elements : 6

4th increment, 48 elements : 6

5th increment, 48 elements : 5

6th increment, 48 elements : 5

7th increment, 48 elements : 5

8th increment, 48 elements : 4

9th increment, 48 elements : 4

10th increment, 48 elements : 6

2nd refinement, 192 elements : 8

3rd refinement, 768 elements : 8

4th refinement, 3.072 elements : 4

5th refinement, 12.288 elements : 3

Table 7.7: Example 7.2.2.1; number of Newton steps in the single periods.

For this example, the number of Newton steps is similar to the numbers in the
second example for bending problems using BFS-elements, Example 7.2.1.2, al-
though the rotation angle here is only one third the rotation angle in the named
example, respectively.

Obviously, using HCT-elements may be more problematic than using BFS-elements,

due to the discussion in section 6.5, which is seen in the relatively high number of
Newton steps for the second and for the third refinement.

83
7 Numerical examples

7.2.2.2 2nd example: Divided annulus with rotated edge normals

For the last example we consider an annulus. This is similar to the plate in Example
7.2.2.1, above, but with a curved inner and outer boundary. The diameter is 2 again
and the annulus is cut at the same edge, as the octagonal plate, before. The two
new edges have inhomogeneous Dirichlet boundary conditions. All other edges are
left free.
The undeformed mid surface looks alike the respective one in the example above,
displayed in Figure 7.31.
The outer normals of the two arose straight edges are rotated in the same way as
in the example before, applying equation (7.3).
For this rotation a serious problem occurs. The incrementation was successfully
done for 80 or more increments. Unfortunately, after the full rotation angle was
applied, the Newton iteration fails for the first refinement, due to both, the special
refinement technique and the curved boundary.
A way out was changing the procedural method. So, first we apply four refine-
ments in the first incremental stage. After that, we get a slightly deformed mid
surface approximated with 3.072 elements, displayed in Figure 7.36.

strain
0.014533
s
axi

0.01
3-

0.0014533

1
-a
xi
s

2-axis

Figure 7.36: Deformed mid surface, 3.072 elements, 1 increment.

From the same reason as discussed in section 6.5, here the incrementation starts
at a more refined mesh than before. The result can be seen in Figure 7.37.

84
 7.2 Bending-dominated deformation

strain
0.014533

s
axi
0.01

3-

0.0014533
1
-a
xi
s

2-axis

Figure 7.37: Deformed mid surface, 3.072 elements, full force applied.

The maximum of the norm of the strain tensor is about 0.0145. This is a little
bit shorter than in Example 7.2.2.1 using 12.288 elements (Figure 7.35). It appears
in the same region as in the example, above. Unfortunately, the computing time is
much higher than in the example before, due to the large number of increments and
the change of the sequence of incrementation and refinement.

strain
0.014533
s
axi

0.01
3-

0.0014533

1
-a
xi
s

2-axis

Figure 7.38: Cut slab, deformation progress, 3.072 elements.

In Figure 7.38 the mid surface is displayed in the undeformed state using 12

85
7 Numerical examples

elements and in the first incremental stage as well as in the refined period with
3.072 elements, respectively, for the imagination of the deformation progress.
The number of Newton steps for each stage of the simulation is shown in the
following Table 7.8.

1st increment, 12 elements : 6

1st increment, 48 elements : 6

1st increment, 192 elements : 5

1st increment, 768 elements : 5

1st increment, 3.072 elements : 5

.................

4th increment, 3.072 elements : 5

5th increment, 3.072 elements : 4

.................

10th increment, 3.072 elements : 4

11th increment, 3.072 elements : 3

.................

80th increment, 3.072 elements : 3

Table 7.8: Example 7.2.2.2; number of Newton steps in the single periods.

Due to the problems discussed in section 6.5, here it was necessary to perform more
mesh refinements in the beginning, before the incrementation starts. Unfortunately,
this enlarges the total operating expense drastically. Nevertheless, the number of
Newton steps is not much higher for one Newton iteration as in the examples before.

In summary we can determine that in the examples for BFS-elements we have

a smaller number of Newton steps per Newton iteration than in the examples for
the reduced HCT-elements. This effect is observable, due to the nested function
spaces for the form functions. Unfortunately the BFS-elements are only usable for
examples with axis-parallel edges.

86
 7.2 Bending-dominated deformation

Contrary, the HCT-elements are more flexible and can also be used for curved
boundaries. But due to the discussion in section 6.5 there occur other problems to
overcome. Because of the many different demands to a finite element, one would
never find an element, which is the best for all the examples, at the same time.

87
8 Outlook and open questions
The main focus of this work is on the numerical simulation of large deformations
of thin plates. Therefore, we consider a non-linear plate theory with a non-linear
material law. Additional simplifications have been set aside. The displacement
vector function is derived from a non-linear system of equations, which emerges from
the equilibrium of forces. The reduction of dimension to the mid surface requires
a plate assumption. Here, we use the Kirchhoff hypothesis, where no change in
thickness direction is allowed during the deformation process. The basic principle for
the simulation is the minimization of the energy functional over the deformed plate.
This implies that the first derivative of this energy functional vanishes. For solving
this non-linear system of equations Newtons method is applied, which necessitates
the calculation of the second derivative of the energy functional. Its calculation and
the efficient use in the F.E.-simulation was one of the most difficult parts within
this work. As a result of each Newton iteration we obtain a correction term for the
displacement vector function. Some examples for plates of different sizes and for
different kinds of deformation are presented.

One starting point for additional research can be the weakening of the Kirchhoff
hypothesis. Therefore let us consider the deformed plate as

x() = Y ( 1 , 2 ) + U + h d

with an arbitrary direction vector d independent of U , which is no longer restricted

to be perpendicular to the deformed mid surface. In linear elasticity such a model
is known as Mindlin-Reissner plate. Additionally, the straight line h d can be
replaced by non-linear curves depending on , which is known as hierarchical shell
models. For both approaches more degrees of freedom have to be taken into account.
Unfortunately, new difficulties for the Finite-Element-practice arise.

Another research issue is the development of an adaptive approach for the FEM.
For that an error estimator is needed. In [20] a basis for an error functional in the
linear and in the non-linear case is provided for the full 3D theory. At present, the
development of an error estimator for shells and plates and linear problems is in
progress. In future, potential results have to be transferred to the non-linear case.

89
Bibliography
[1] S. S. Antman. Nonlinear Problems of Elasticity. Springer, Berlin, Heidelberg,
New York, 2nd edition, 2005.

[2] Y. Basar and W. Kratzig. A Consistant Shell Theory for Finite Deformations.
Acta Mechanica, 76:7387, 1989.

[3] M. Bischoff, W. A. Wall, K.-U. Bletzinger, and E. Ramm. Models and Finite
Elements for Thin-walled Structures. Encyclopedia of Computational Mechan-
ics, 2:59137, 2004.

[4] J.P. Boehler. Applications of Tensor Functions in Solid Mechanics. Interna-

tional Center for Mechanical Science (CISM Courses), 292, 1987.

[5] D. Braess. Finite Elements: Theory, Fast Solvers and Applications in Solid
Mechanics. Cambridge University Press, Cambridge, 3rd edition, 2007.

[6] I. H. Bramble, J.E. Pasciak, and J. Xu. Parallel Multilevel Preconditioners.

Mathematics of Computation, 55:122, 1990.

[7] D. Chapelle and K.-J. Bathe. The Finite Element Analysis of Shells - Funda-
mentals. Springer, Berlin, Heidelberg, 2003.

[8] P. G. Ciarlet. An Introduction to Differential Geometry with Application to

Elasticity. Journal of Elasticity, 78-79:1215, 2005.

[9] P. G. Ciarlet and P. Dstuynder. A Justification of a Nonlinear Model in Plate

Theory. Computer Methods in Applied Mechanics and Engineering, 17/18:227
258, 1979.

[10] P.G. Ciarlet. The Finite Element Method for Elliptic Problems. North-Holland,
Amsterdam, 1978.

[11] P.G. Ciarlet. Mathematical Elasticity Volume I: Three Dimensional Elasticity.

North-Holland/Elsevier, Amsterdam, 1988.

[12] P.G. Ciarlet. Mathematical Elasticity Volume II: Theory of Plates. North-
Holland/Elsevier, Amsterdam, 1997.

[13] W. Dahmen, P. Oswald, and Shi Xi-Quan. C1 -Hierarchical Bases. Journal of

Computational and Applied Mathematics, 51:3756, 1994.

91
Bibliography

[14] V. V. Eliseev and Y. M. Vetyukov. Finite Deformation of Thin Shells in the

Context of Analytical Mechanics of Material Surfaces. Acta Mechanica, 209:43
57, 2010.

[15] Ch. Gromann and H.-G. Roos. Numerical Treatment of Partial Differential
Equations. Springer, Berlin, Heidelberg, New York, 2007.

[16] M. Jung and U. Langer. Methode der Finiten Elemente f

ur Ingenieure Eine
Einf
uhrung in die Numerischen Grundlagen und Computersimulation. Teubner,
Stuttgart, Leipzig, Wiesbaden, 2001.

[17] H. Matthes. Die nicht

uberlappende Gebietszerlegungsmethode zur Parallelisie-
rung und Vorkonditionierung iterativer Verfahren zur Losung von Platten- und
Schalenproblemen. PhD thesis, TU Chemnitz, Chemnitz, 1996.

[18] A. Meyer. Efficient Preconditioners for Special Situations in Finite Element

Computations. Lecture Notes in Computational Science and Engeneering
(LNCSE), 52:6785, 2006.

[19] A. Meyer. Grundgleichungen und Adaptive Finite-Elemente-Simulation bei

Groen Deformationen. Chemnitz Scientific Computing Preprints, 02:16,

2007.

[20] A. Meyer. Error Estimators and the Adaptive Finite Element Method on Large
Strain Deformation Problems. Mathematical Methods in the Applied Sciences,
32:21482159, 2009.

[21] A. Meyer. The Koiter Shell Equation in a Coordinate Free Description. Chem-
nitz Scientific Computing Preprints, 02:14, 2012.

[22] A. Meyer. A Simplified Calculation of Reduced HCT-Basis Functions in a Finite

Element Context. Computational Methods in Applied Mathematics (CMAM),
12(04):486499, 2012.

[23] A. Meyer and J. Ruckert. C1 -continuous FEM for Kirchhoff Plates and Large
Deformation. Journal of Applied Mathematics and Mechanics, (ZAMM), sub-
mitted, 92, 2012.

[24] A. Meyer and J. Ruckert. Kirchhoff Plates and Large Deformation. Chemnitz
Scientific Computing Preprints, 01, 2012.

[25] A. Meyer and J. R

uckert. Large Deformation and Plates. Proceedings in Applied
Mathematics and Mechanics (PAMM), 12:669670, 2012.

[26] P. Oswald. Hierarchical Conforming Finite Element Methods for the Bihar-
monic Equation. Siam Journal Numerical Analysis, 29(6):16101625, 1992.

92
 Bibliography

[27] M. Thess. Parallel Multilevel Preconditioners for Thin Smooth Shell Finite Ele-
ment Analysis. Numerical Linear Algebra with Applications, 5:401440, 1998.

[28] M. Thess. Parallel Multilevel Preconditioners for Thin Smooth Shell Problems.
PhD thesis, TU Chemnitz, Chemnitz, 1998.

[29] H. Yserentant. Two Preconditioners Based on the Multilevel Splitting of Finite

Element Spaces. Numerical Mathematics, 58:163184, 1990.

93
Notation

= ( 1 , 2 , 3 ) coordinate system

i i-th coordinate

= 3 thickness coordinate

0 () domain of a undeformed (deformed) shell/plate

m m
0 ( ) domain of the mid surface of a undeformed (deformed)
shell/plate

0 () boundary of a undeformed (deformed) shell/plate

0,N Neumann boundary part of the undeformed shell/plate

D , N Dirichlet and Neumann boundary part of the deformed

shell/plate

X() (x()) material point of the undeformed (deformed) domain

Y ( 1 , 2 ) (y( 1 , 2 )) material point of the undeformed (deformed) mid sur-

face

Gi , Gi (g i , g i ) covariant and contravariant tensor base of the unde-

formed (deformed) domain

Grad (grad ) gradient operator in the undeformed (deformed) domain

div divergence operator

U , V , U KH , U vector functions

F deformation gradient

(C, c, ) C, Cb (matrix of the) right Cauchy-Green deformation tensor

95
Notation

IC , IIC , IIIC principal invariant of the right Cauchy-Green deforma-

tion tensor

ai , i = 1, 2, 3 functions of the principal invariants of the right Cauchy-

Green deformation tensor

(E, e, ) E, Eb (matrix of the) Lagrangian strain tensor

Ai (ai ) i = 1, 2 tangential vectors of the undeformed (deformed) mid

surface

A3 (a3 ) normal of the undeformed (deformed) mid surface in

thickness direction

A, B (a, b) matrix of the first end second fundamental forms in the

undeformed (deformed) domain
1 2
T ,T first and second Piola-Kirchhoff stress tensor
2
T matrix of the second Piola-Kirchhoff stress tensor

I, I identity tensor of 2nd and of 4th order

C 4th order tensor, material tangent

Cauchy stress tensor

material density

p outer acceleration

l0 (U ; V ) Frechet derivative of functional l(U ) in direction of V

l00 (U ; U , V ) Frechet derivative of functional l0 (U ; V ) in direction of

U

N (n) normal vector of the undeformed (deformed) domain in

the 3D-theory

a(U ; V ), f (V ) a-form and force vector for the variational formulation

L(V ) special defined differential operator applied on V

96
 Notation

Th FE-mesh of the mid surface

T finite element

Vh finite function space of the actual mesh Th

U h , V h , U h finite dimensional vector functions from Vh

u, v, u coefficients of the corresponding vector functions

97
 Theses

Theses
(1) We introduce the 3D-theory for large deformations and describe the common
non-linear shell theory. We restrict us to the plate, a shell, which is flat in the
initial state. Both theories combined with the Kirchhoff hypothesis lead to a
strain tensor E of lower rank 2 and to the non-linear weak formulation of the
boundary value problem.

(2) The numerical solution of the non-linear boundary value problem obtained
in the weak formulation requires both a Newton-linearization and a Finite
Element-approximation. For a correct Newton iteration the second derivative
of the energy functional has to be established. Its implementation in generating
the element matrices after the F.E.-discretization is the main difficulty for
obtaining an efficient simulation.

(3) For the F.E.-discretization the BFS-element as well as the reduced HCT-
element have been considered, which both satisfy the special requirements
for a C1 -continuous treatment.

(4) A collection of various examples illustrates the flexibility of the considered

approach. The plate deflection examples are given for the validation of the
code in comparing with small strain and for displaying the fast convergence of
the method. The individual bending-dominated examples were the principal
reason for this research and display the potential as well as the limits of the
method. Some limits of the finite elements used are discussed additionally.

(5) This work can be used as a starting point for additional research. For example,
the use of other C1 -continuous elements or a non-conformal approach could be
investigated. Additionally, the weakening of the Kirchhoff hypothesis would
require C0 -continuous elements only, but implies other numerical difficulties.
Furthermore, an error estimator would allow adaptive mesh refinement, which
increases the total efficiency.

99
List of Figures

6.1 third-order polynomials. . . . . . . . . . . . . . . . . . . . . . . . . . 44

7.1 Undeformed mid surface, 16 elements. . . . . . . . . . . . . . . . . . . 56

7.2 Deformed mid surface after applying the first increment, 16 elements. 57
7.3 Deformed mid surface after applying the full force vector, 16 elements. 57
7.4 Deformed mid surface, 16.384 elements, full force applied. . . . . . . . 59
7.5 Deformed mid surface, 16.384 elements, full force applied, birds eye
view. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
7.6 Deformed mid surface, 16.384 elements, third component of displace-
ment vector U (3) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
7.7 Deformed mid surface, 16.384 elements, small strain. . . . . . . . . . 61
7.8 Deformed mid surface, 16.384 elements, small strain, third component
of the displacement vector U (3) . . . . . . . . . . . . . . . . . . . . . . 61
7.9 Deformed mid surface, 16.384 elements, small strain, first component
of the displacement vector U (1) . . . . . . . . . . . . . . . . . . . . . . 62
7.10 Deformed mid surface, 16.384 elements, small strain, second compo-
nent of the displacement vector U (2) . . . . . . . . . . . . . . . . . . . 62
7.11 Deformed mid surface after applying first increment, 24 elements,
birds eye view. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.12 Deformed mid surface after applying the first increment, 24 elements. 64
7.13 Deformed mid surface after applying the full force vector, 24 elements. 64
7.14 Deformed mid surface, 6.144 elements, full force applied. . . . . . . . 65
7.15 Deformed mid surface, 6.144 elements, full force applied, birds eye
view. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
7.16 Undeformed mid surface, 8 elements. . . . . . . . . . . . . . . . . . . 67
7.17 Deformed mid surface after applying the first increment, 8 elements. . 68
7.18 Deformed mid surface, 8 elements. . . . . . . . . . . . . . . . . . . . . 68
7.19 Deformed mid surface, 32.768 elements. . . . . . . . . . . . . . . . . . 70
7.20 Cylinder, deformation progress. . . . . . . . . . . . . . . . . . . . . . 70
7.21 Deformed mid surface, 8 elements, first increment. . . . . . . . . . . . 71
7.22 Deformed mid surface, 8 elements, full force applied. . . . . . . . . . 72
7.23 Deformed mid surface, 32.768 elements. . . . . . . . . . . . . . . . . . 72
7.24 Deformed mid surface, 256 elements, first increment. . . . . . . . . . 74
7.25 Deformed mid surface, 256 elements, full force applied. . . . . . . . . 75
7.26 Deformed mid surface, 16.384 elements. . . . . . . . . . . . . . . . . . 75
7.27 Mobiusstrip, deformation progress. . . . . . . . . . . . . . . . . . . . 77

101
List of Figures

7.28 Deformed mid surface, 256 elements, first increment. . . . . . . . . . 77

7.29 Deformed mid surface, 256 elements, full rotation of 4 applied. . . . 78
7.30 Deformed mid surface, 16.384 elements. . . . . . . . . . . . . . . . . . 78
7.31 Undeformed mid surface, 12 elements. . . . . . . . . . . . . . . . . . . 80
7.32 Deformed mid surface, 48 elements, 1 increment. . . . . . . . . . . . . 81
7.33 Deformed mid surface, 48 elements, full force applied. . . . . . . . . . 81
7.34 Deformed mid surface, 3.072 elements, full force applied. . . . . . . . 82
7.35 Deformed mid surface, 12.288 elements, full force applied. . . . . . . . 82
7.36 Deformed mid surface, 3.072 elements, 1 increment. . . . . . . . . . . 84
7.37 Deformed mid surface, 3.072 elements, full force applied. . . . . . . . 85
7.38 Cut slab, deformation progress, 3.072 elements. . . . . . . . . . . . . 85

102
List of Tables
7.1 Example 7.1.1; number of Newton steps at the single periods. . . . . 58
7.2 Example 7.1.2; number of Newton steps at the single periods . . . . . 66
7.3 Example 7.2.1.1; number of Newton steps in the single periods. . . . . 69
7.4 Example 7.2.1.2; number of Newton steps in the single periods. . . . . 73
7.5 Example 7.2.1.3; number of Newton steps in the single periods. . . . . 76
7.6 Example 7.2.1.4; number of Newton steps in the single periods. . . . . 79
7.7 Example 7.2.2.1; number of Newton steps in the single periods. . . . . 83
7.8 Example 7.2.2.2; number of Newton steps in the single periods. . . . . 86

103