Contact Methods in Finite Element Simulations
Contact Methods in Finite Element Simulations
voorzitter en secretaris:
Prof. dr. ir. H.J. Grootenboer Universiteit Twente
promotor:
Prof. dr. ir. J. Huétink Universiteit Twente
leden:
Dr. R.M.J. van Damme Universiteit Twente
Prof. dr. ir. J.B. Jonker Universiteit Twente
Prof. dr. ir. A. van Keulen Technische Universiteit Delft
Prof. dr. J. Molenaar Technische Universiteit Eindhoven &
Universiteit Twente
Prof. dr. ir. D.J. Schipper Universiteit Twente
Kloosterman, Gertjan
PROEFSCHRIFT
door
Gertjan Kloosterman
Summary ix
Samenvatting xi
1 Introduction 1
1.1 Contact in forming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 The finite element method . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Aim of this thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Outline of this thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Contact Mechanics 5
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Notational conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2.1 The reference configuration . . . . . . . . . . . . . . . . . . . . . 6
2.2.2 The current configuration . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.3 Differentiation and variations . . . . . . . . . . . . . . . . . . . . 8
2.3 Kinematics of the contact surface . . . . . . . . . . . . . . . . . . . . . . . 9
2.3.1 Parametrisation of the contact surface . . . . . . . . . . . . . . . . 9
2.3.2 The local basis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.3 Normal distance and interface velocity . . . . . . . . . . . . . . . . 11
2.4 Contact in elasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.5 The contact constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.5.1 Stresses and tractions . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.5.2 The impenetrability constraint . . . . . . . . . . . . . . . . . . . . 16
2.5.3 Frictional constraints . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.6 The weak form of contact . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.6.1 Forming the contact integrals . . . . . . . . . . . . . . . . . . . . . 17
2.6.2 Equilibrium of tractions . . . . . . . . . . . . . . . . . . . . . . . 18
2.6.3 Weak form with virtual displacements . . . . . . . . . . . . . . . . 20
2.6.4 Weak form with virtual velocities . . . . . . . . . . . . . . . . . . 23
2.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
vi Contents
3 Regularisation Methods 25
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2 Discretisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.2.1 Restricting the solution space . . . . . . . . . . . . . . . . . . . . 26
3.2.2 The discrete elasticity problem . . . . . . . . . . . . . . . . . . . . 28
3.2.3 The variational problem . . . . . . . . . . . . . . . . . . . . . . . 30
3.3 Non regularisation methods . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3.1 The mixed method . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3.2 The constraint method . . . . . . . . . . . . . . . . . . . . . . . . 32
3.4 The regularisation framework . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.4.1 The penalty method . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.4.2 The method of augmented Lagrangians . . . . . . . . . . . . . . . 37
3.4.3 The modified barrier method . . . . . . . . . . . . . . . . . . . . . 39
3.5 Regularising Friction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.5.1 Penalty approach . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.5.2 The method of augmented Lagrangians . . . . . . . . . . . . . . . 47
3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4 Distance Functions 49
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.2 Discretisation of the contact integral . . . . . . . . . . . . . . . . . . . . . 51
4.2.1 The discrete integral formulation . . . . . . . . . . . . . . . . . . . 51
4.2.2 The appearance of geometric incompatibilities . . . . . . . . . . . 51
4.2.3 Dealing with geometric incompatibilities . . . . . . . . . . . . . . 53
4.2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.3 Evaluating the discretised contact integral . . . . . . . . . . . . . . . . . . 54
4.3.1 Numerical integration schemes . . . . . . . . . . . . . . . . . . . . 54
4.3.2 Integrating complex master boundaries . . . . . . . . . . . . . . . 56
4.3.3 Piecewise linear boundaries: problems in sliding . . . . . . . . . . 58
4.3.4 Piecewise linear boundaries: problems in computing . . . . . . . . 59
4.3.5 Piecewise linear boundaries: projecting . . . . . . . . . . . . . . . 60
4.3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.4 Smoothing the master boundary . . . . . . . . . . . . . . . . . . . . . . . 62
4.4.1 Goals to attain when smoothing . . . . . . . . . . . . . . . . . . . 62
4.4.2 Smoothing curves . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.4.3 Smoothing 3-dimensional surfaces . . . . . . . . . . . . . . . . . . 66
4.4.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.5 Efficient projection finding . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.5.1 The capture box problem . . . . . . . . . . . . . . . . . . . . . . . 73
4.5.2 Building the data structure . . . . . . . . . . . . . . . . . . . . . . 74
4.5.3 Querying the data structure . . . . . . . . . . . . . . . . . . . . . . 76
4.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
Contents vii
Bibliography 107
Glossary 115
Acknowledgements 119
The application of finite element methods in forming processes has gained a lot of ground
in recent years, especially due to the spectacular advances in the chip industry. Simulating
the process with the finite element method offers a producer the advantage of spotting
possible production problems early in the development cycle, the ability to optimise the
forming process for superior final results and an increased insight into the shaping process
in general.
A typical forming process employs one or more tools with which a work piece is de-
formed into a desired shape, from which we can conclude that there is interaction between
different solid components within the process. To model the interactions between the vari-
ous components, a contact algorithm is required. It is this algorithm which forms the topic
of this thesis.
The popularity of the method has sparked a flurry of activity within the finite element
community to develop increasingly accurate material models and perform simulations with
ever decreasing granularities. The rate of convergence of the simulations has been erratic,
which has been often blamed on the specifics of the contact algorithm employed. This
brings about the aim of this work: Study the contact problem applied to metal forming
simulations, see where it originates, formulate the equations and constraints, discuss the
various ways in which to enforce those constraints, and finally how to incorporate it all
into a finite element framework. By scrutinising the complete path of development of the
equations to the implementation, we intend to identify all the possible pitfalls occurring
when contact methods are applied.
The first part of this thesis studies the purely theoretical part, and develops the contact
equations in forming, as well as discusses most popular methods to solve it with a sufficient
degree of accuracy, stability and efficiency. By retaining a high level of abstractness it is
shown that all the popular methods arise from a common form.
The second part of this thesis deals with the practicalities of the methods. This involves
computing distances accurately and efficiently and using those to perform the integrations
that are required in the contact methods developed in the first part. As it turns out it is this
part which is most responsible for the ill behaviour of the algorithm, whereas most of the
developments in the literature have addressed the discussion in the first part.
To overcome these difficulties, different integration algorithms are discussed as well as
x Summary
verantwoordelijk is voor het slechte convergentie gedrag van het algoritme, en dat terwijl
het meeste werk in de literatuur wordt besteed aan de discussie in het eerste deel.
Om deze problemen de baas te worden, worden verschillende integratie algoritmes be-
schouwd alswel de mogelijkheid om de convergentieproblemen op te lossen door het glad-
maken van de contactoppervlakken. Bovendien ontwikkelen we een zeer efficiënt projec-
tie algoritme om de toepasbaarheid van contact algoritmes verder te vergroten. Tenslotte
worden de theorie en de implementatie geı̈llustreerd aan de hand van een verzameling nu-
merieke experimenten.
Chapter 1
I NTRODUCTION
punch
blankholder
die
flat or pre-shaped sheet material called the blank is placed between a blank holder and a
die. The blank holder is loaded with a force to prevent wrinkling and to control the amount
of material that flows into the die. After that the punch is moved downwards, pushing
into the blank, which then partly stretches, and partly pulls material from under the blank
holder into the die. The result of the process is a product which contains both the shape
of the punch and that of the die. The flow of the material into the die is essential for the
formability of the product and its properties.
When the blank holder force is very high, the material cannot draw in from under the
blank holder into the die, and consequently the blank only stretches. When the process is
applied thus, it is called stretch forming.
Another process is rolling. The schematic for this process is illustrated in Figure 1.2.
In the rolling process a slab of material, is deformed between two rolls which deform the
top roll
bottom roll
Figure 1.2: Schematic of the rolling process.
slab. Rolling is used to create sheet material, which can subsequently be used in a deep
drawing process. Alternative uses are to give the rolls a specific shape across their width.
Using the rolling process in this way, is called profile rolling.
All the three aforementioned methods have one thing in common: the forming is
achieved trough the interaction between the work piece and the tools. The interaction
forces are then the contact forces which consist of normal components which prevent the
objects from interpenetrating and tangential components, which are also known as the fric-
tional components. If the forming process hinges on the transfer of these forces from the
work piece to the tool, then we need to determine these forces as accurately as possible if
we wish to simulate the problem correctly.
modelled by dividing it into more simple elementary parts. The resulting model is loaded
on the boundary of the mesh, and due to this loading a certain response is observed from the
model. Part of the response is the deformation of the mesh. Due to the deformation, parts
of the workpiece may come into contact with the tool, or conversely may loose contact.
This change of contact status, changes the loading that is applied on the boundary of the
mesh. The method that tracks the locations of contact of the workpiece with the tools
during the simulation, and accounts for the change in boundary conditions is called the
contact algorithm.
C ONTACT M ECHANICS
2.1 Introduction
In this chapter, the equations which characterise the contact problem in forming simulations
are deduced. Upon consultation of the literature, one finds that most authors prefer to skip
the derivation. This is perfectly reasonable, since the derivation is quite involved and can
obscure the ideas that are being proposed. However, in a thesis dedicated to the solution of
contact problems in finite element simulations it can not be omitted.
The derivation presented in this chapter is rather rigorous, in the spirit of Laursen and
Simo (1993)1. The rigor is necessary to underline the validity of the methods employed to
solve the contact problem. Apart from correctness of methods, something else is gained by
following the rigorous derivation: We can see the common ground in all the ideas which
are proposed in the literature to solve the problem. Moreover, several seemingly different
ideas can be cast in a single guise by retaining sufficient abstractness.
The purpose of this chapter is to find structure in the description of contact. This struc-
ture turns out to follow from work conjugate pair associations. Let us clarify this some-
what. On the one hand there are the spatial quantities, such as displacements, penetration,
and sliding distance. Whereas on the other hand there are the force quantities such as
stresses and tractions. Typical work conjugate pairs as found in structural mechanics are
the different stress-strain pairs. The Cauchy stress tensor is for example work conjugate to
the rate of deformation tensor, and equivalently so to the Euler-Almansi strain tensor. The
work conjugate pairs are multiplied in the equations to give the dimension of virtual work
(or virtual power), hence the phrase ‘work conjugate’.
Through this relationship, the force quantity is characterised by its work conjugate
spatial quantity. The characterisation is often called a constitutive relationship. As an
example: the stress tensor is assumed to be a function of the rate of deformation. By
1 Unfortunately the article contains an error, which is discussed in Appendix A.
6 Contact Mechanics
applying the same ideas to contact, we intend to pair up the contact normal traction with
the penetration distance as well as pair up the frictional traction with the slip velocity. In
the next chapter the contact tractions are assumed to be a function of their work-conjugate
spatial quantities. This approach can be seen as one which regularises the discontinuous
nature of the problem.
The derivation requires a background knowledge in continuum mechanics, see for ex-
ample Malvern (1969); Bonet and Wood (1997). Notational conventions differ among au-
thors. For clarity, we set them at the beginning of the chapter. Apart from continuum mech-
anics, some differential geometry is employed. The concepts used, are mostly straightfor-
ward. For more background information, refer to do Carmo (1976). An important topic
is the computation of variational derivatives. For the sake of completeness we discuss the
manner in which these computations are performed.
The contact equations are derived twice: once using the principle of virtual work, and
once using the principle of virtual power. It turns out that the interpretation of the weighing
functions results in different derivation paths, but the results are equivalent. This should
obviously hold since the starting point for both derivations is equilibrium of forces. Al-
though only one derivation is sufficient to be able to read the remainder of this thesis, both
of the derivations are presented for sake of completeness.
The layout of this chapter follows the line of reasoning above. Before anything can
be discussed a notational convention is required, as well as a number of mathematical
preliminaries, this is done in Section 2.2. In Section 2.3 the contact kinematics are dis-
cussed. Concepts that are introduced comprise the normal distance and interface velocity.
In Section 2.4 the contact problem in elasticity is discussed, as a simplified but illustrative
example of the general contact problem. In Section 2.5 the characterising constraints on
the contact tractions and frictional responses are discussed. Finally, in Section 2.6 the weak
form is derived and manipulated to obtain the form in which the force quantities are paired
up with their work conjugate spatial quantities. Once with virtual displacements, and once
with virtual velocities. In the next chapters algorithms are proposed to solve the resulting
set of equations. Finally, in Section 2.7 we present the conclusions.
ϕ (1)
Reference configuration
(1) γ (1)
(1) X(1)
x(2)
x(1)
(2)
X(2) γ (2)
(2)
capital Roman and Greek letters. In the reference configuration the set of points that make
up the i -th body is denoted by (i) . The points are elements of Rnd , where n d denotes the
number of space dimensions, which is either 2 or 3.
A typical element of (i) is denoted by X(i) . Objects containing multiple components,
such as vectors and tensors are printed bold, whereas scalar quantities, and the associated
length of a vector are not. As an example if a is a vector, then “a” is the length of that
vector (in the 2-norm), and a is a scalar, which has got nothing do to with the vector a.
The boundary of (i) is denoted by ∂(i) . The portions of ∂(1) and ∂(2) where
contact may occur are assigned to the sets (1) and (2) respectively.
Properties in the current configuration are generally written with lowercase Roman and
Greek letters. A point in the current configuration is denoted by x(i) . It is assumed that
there exists an invertible map ϕ (i) that maps each point in the reference configuration onto
a point in the current configuration. By assigning a subscript t to these functions, a map-
ping for each time can be obtained. Or equivalently, we could extend the function ϕ (i)
(1)
to take an additional argument. Hence, ϕ t (X(1) ) for t ∈ [0, T ] gives the trajectory of a
(i)
material point X(1) in (1). For a particular time t, the notations ϕ t (X(i) ), ϕ (i) (X(i) , t)
(i) (i)
and ϕ (X ) can be used to denote the same thing: The current position of a particular
material point. Here it is assumed that the time is fixed and known. In an identical manner,
the boundary of either body in the current configuration can be found by applying ϕ (i) to
(i) (i)
(i) . The boundary in the current configuration is denoted by γt = ϕ t ( (i) ). In the
previous notation, the subscript t can again be omitted if the time is known.
8 Contact Mechanics
Partial derivatives of a function f are denoted by f ,i or f,α . The former notation using
Roman letters indicates that i = 1, . . . , n d , whereas the latter notation, using Greek letters,
indicates that α = 1, . . . , n d − 1. As an example, the contact surface in 3D is a function
of two independent parameters. Consequently, each coordinate function of that surface has
only two partial derivatives, one less than the number of space dimensions.
This type of indexing is also assumed on indexing of other functions. Thus, writing
down X i , using a Roman letter indicates that i runs over 1, . . . , n d . Whereas, writing ξ α ,
assumes that α = 1, . . . , n d − 1. Also note that the superscript does not indicate taking ξ
to the power α, but merely the α-th component in the (contravariant) vector ξ .
A functional is a function that takes functions as an argument. As an example of a
functional consider the potential energy due to gravitation of a body :
P(ϕ) = ρgϕ(X) · nz d, (2.2)
where nz is the unit normal pointing upward. The variational derivative of a functional is
a directional derivative. It is computed by adding a small variation to the argument of the
functional and then taking the derivative (hence, the name variational derivative). As an
example, consider a fixed function ϕ and an arbitrary function ϕ̂. A perturbation of ϕ in
the direction of ϕ̂ is denoted by:
ϕ = ϕ + ϕ̂. (2.3)
Entering the perturbed function into a functional f yields a functional dependent on .
Differentiating this functional with respect to results in the variational derivative in the
direction of ϕ̂ at ϕ.
Let Da [g(x)] denote the derivative of g in the direction of a at the point x. Then taking
the variational derivative of a functional f , can be denoted by:
d
d
Dϕ̂ [ f (ϕ)] = f (ϕ ) = f (ϕ + ϕ̂) . (2.4)
d =0 d =0
As can be concluded from the above equation the variational derivative of a functional
is again a functional, this time taking two functions as arguments. In an equivalent manner
it is possible to take the directional derivative of a function itself:
d
d
Dϕ̂ [ϕ] = ϕ = ϕ + ϕ̂ = ϕ̂. (2.5)
d =0 d =0
Further property notations are introduced in the places where they are required. In the
next section a start is made with the description of properties on the (potential) contact
surface.
2.3 Kinematics of the contact surface 9
(2)
0 (2)
t
Parameter space
A(2) ξ (2)
The dimension of the contact surface (i) is one lower than the number of space di-
mensions. Thus, for a 2D problem, (i) is a curve, and for a 3D problem (i) is a surface.
Furthermore, without loss of generality we assume that (i) is simply connected. As a
consequence of the connectivity, all the points in (i) can be continuously mapped onto a
unique point in Rnd −1 and vice versa. The mapping in itself is not unique.
To cast the above reasoning in a formula: call A(i) the connected set of points mapped
to in Rnd −1 . Using this notation there exists a continuous and invertible map
(i)
0 :A
(i)
⊂ Rnd −1 −→ Rnd , (2.6)
such that
(i) = (i) (i)
0 (A ). (2.7)
(i)
Furthermore, if (i) is sufficiently smooth, 0 can be chosen differentiable. We denote a
typical element of A by ξ .
10 Contact Mechanics
The concepts are illustrated in Figure 2.2. The mapping discussed is known as a para-
metrisation of the contact surface, where ξ is the parameter.
A parametrisation of the contact surface at any later time, can be obtained through
applying ϕ (i) (i) (i)
t after 0 . This composition of the two functions is given a new name, t .
Thus
(i) (i) (i)
γ (i) = ϕ t ( (i) ) = ϕ t 0 (A(i) )
= ϕ (i)
t ◦ 0
(i)
(A(i) ) = (i) (i)
t (A ). (2.8)
= F(i) (i)
t · Tα . (2.13c)
(i) (i)
A remark has to be made about the vectors Tα and τ α : They are not necessarily
orthogonal with respect to one another, nor will they in general have unit length. Hence,
care has to be taken in how quantities are expressed with regard to the local basis.
As an additional remark, if n d = 3, the unit normal vector can be computed by:
(i) (i)
τ1 × τ2
n(i) = . (2.14)
τ (i) (i)
1 × τ2
Note that the outer product is dependent on the order of the vectors that appear in it, since
a × b = −b × a. As a result, we have to take care in the selection of the orientation of the
parametrisation, so that the normal is always pointing outward.
The point of projection is thus a function of the point to be projected, and is denoted by
x(2) . Hence, x(2) ∈ γ (2) is that point, which gives the minimum distance
Unfortunately, the point is not necessarily unique. If that is the case, the distance between
the bodies is necessarily non zero. We are then free to choose one of the possible projection
points.
From now on quantities defined on the master body defined through the slave body are
overlined. For a clarification of concepts, refer to Figure 2.3.
The location of x(2) is the map of a point ξ (2) ∈ A(2). This point then depends on
the deformations of both bodies and ξ (1) . Sticking to the notation, where properties on the
master boundary that depend on the slave boundary are overlined, we introduce:
(2) (2)
ξ =ξ (ξ (1), ϕ (1) (2)
t , ϕ t ). (2.17)
12 Contact Mechanics
γ (1)
x(1)
dN
n(2)
γ (2)
x(2)
Thus, we can say that the contact surface of interest is completely parametrised through
ξ (1) . Parameters ξ (2) that are not addressed through the previous projection mapping can
never be in contact with ∂(1), and as a result can be ignored for this master-slave pair.
The normal distance vector, is expressed using this mapping:
(1)
d N (x(1)) = x(1) − x(2) , (2.18)
where (2)
(2)
x(2) = t ξ . (2.19)
The signed normal distance is found by projecting this vector along the outward unit normal
vector of the master body:
d(1)
N = n(2)
· x (1)
− x (2)
. (2.20)
Using this definition, a description for the slip velocity can be found. First it is noted
that there can only be a slip velocity at a certain point ξ (1) if the normal distance remains
zero, which is expressed through the relationship:
d (1)
x − x(2) = 0. (2.21)
dt
Let us analyse this expression somewhat further. First, the total time derivative of x(1) is:
Secondly, the total time derivative of x(2) is to be computed, which is somewhat more
complex:
d (2) d (2) (2) (2)
x = ϕ t ( 0 (ξ ))
dt dt
∂ϕ (2) d (2) (2)
= v(2)t + · 0 (ξ )
∂X(2) dt
= v(2) + F · ξ̄˙ (2)α
(2) (2)
t t 0,α
=
(2)
+ τ (2) ˙ (2)α .
vt α ξ̄ (2.23)
Using the previous result in (2.21), we find for the tangential slip velocity which is the
difference in velocity between the contacting points:
where the bar on the properties indicate that their properties are with respect to the master
body.
To simplify the introduction of the contact problem in general deformation problems,
we first present a simpler case: Contact in elasticity.
∂(i) on which it is possible to prescribe tractions (forces per area). For now, we assume
that this set is empty. We note that the distances only need to be checked for the master
body, the other boundary is completely characterised through points on the master body.
For brevity of notation, and to simplify the problem somewhat, we assume that all func-
tions ϕ (i) , which are considered satisfy the prescribed boundary conditions. Furthermore,
we introduce the composition of functions:
ϕ (1) (X) if X ∈ (1)
ϕ(X) = . (2.26)
ϕ (2) (X) if X ∈ (2)
In this case, P (1) ϕ (1) = P (1) (ϕ), since the dependency is now implicit, the same holds
for P (2) . Writing P = P (1) + P (2) , we arrive at the following simplified form:
min P(ϕ) (2.27a)
ϕ
s.t. d N (ϕ) ≥ 0 on . (2.27b)
In this equation we dropped the superscript from the contact boundary. The problem ex-
pressed in (2.27a) has the following Lagrangian:
L(ϕ, λ N ) = P(ϕ) + λ N d N (ϕ) d. (2.28)
The solution of (2.27a) is a saddle point of L. The conditions for this saddle point are that:
Dδϕ [L(ϕ, λ N )] = 0 for all δϕ, (2.29a)
λ Nd N = 0, (2.29b)
λN ≤ 0, (2.29c)
dN ≥ 0. (2.29d)
The first equation (2.29a) is telling us that the final solution needs to be a critical point with
respect to ϕ. Thus that all directional derivatives are zero. The conditions (2.29b)–(2.29d)
are known as the Karush-Kuhn-Tucker conditions. They follow from the minimisation
problem directly. It is now interesting to note that we obtain the KKT conditions from the
optimisation problem (2.25a-2.25d), whereas later, we have to put them in explicitly to get
to the variational problem. The meaning of these constraints are discussed in Section 2.5.
Expanding (2.29a) leads to the following equation:
Dδϕ [P(ϕ)] + λ N Dδϕ [d N(ϕ)] d = 0, (2.30)
or equivalently, rewriting the integral to the current configuration:
Dδϕ [P(ϕ)] + λ N Dδϕ [d N (ϕ)] dγ = 0. (2.31)
γ
Here the determinant of the Jacobian, which appears when the integration parameter is
changed, is multiplied directly into λ N which results in a (different) Lagrange multiplier
λ N.
In the next section the contact constraints are introduced, which are inserted into the
general equations in Section 2.6. From these general equations, we then attempt to regain
some of the structure which is evident in (2.31).
2.5 The contact constraints 15
f df
t = lim = . (2.32)
A→0 A dA
A unique tensor σ exists, that relates the normal of a surface element to the traction vector
for that surface element:
t = σ · n. (2.33)
The tensor σ is called the natural or Cauchy stress tensor. It is fully defined on the current
configuration, and is therefore physically meaningful.
where
(i)
tN = t(i) · n(i) n(i) ,
(i)
tT = P(i) · t(i) .
The magnitudes of the respective vectors are denoted by the same symbols, but by using a
(i)
regular character, such as t N and v N , instead of bold.
The first condition (2.37a) states that no penetration may occur. Hence, this is the form in
which the impenetrability constraint is cast. Using this, the normal traction can be char-
acterised. The second condition (2.37b) states that the contact normal traction should be
compressive. Finally, the third condition (2.37c) states a complementarity condition. If
there is no contact, then no compressive tractions can occur. Alternatively: If there are no
compressive stresses, then the distance must be positive.
These conditions are known as the Karush-Kuhn-Tucker conditions for optimality. For
the general case, they are required to be introduced into the weak form of equilibrium.
However, if the equilibrium conditions can be derived from a minimum energy principle
(as in elasticity), then we get the contact constraints from just demanding impenetrability.
As with the Karush-Kuhn-Tucker constraints for plasticity one additional equation can
be added to the previous set, which is the persistency condition, see Simo and Hughes
(1998),. If the point remains in contact, then:
(i) (i)
t N · ḋ N = 0. (2.38)
2.6 The weak form of contact 17
The first condition (2.39a) states the (Coulomb) friction condition. In the case that µ is
allowed to be a function of velocity and/or pressure, more general frictional laws can be in-
troduced. The second condition (2.39b), together with the third condition (2.39c) constrains
the tangential traction to work opposite to the direction of slip. Finally the fourth condition
states another complementarity condition: There is no slip if the tangential traction has not
reached its (local) maximum. And if there is slip, then it has reached its maximum.
It can be seen here that the friction tractions are complementary to the slip velocity.
Most finite element methods are concerned with displacement based methods. How this
can be dealt with is discussed in the next section.
In the equilibrium equation σ (i) is the Cauchy stress tensor in body i and f(i) represents the
body forces acting on body i . There are furthermore boundary conditions prescribed on
(i)
γσ , which is that part of the boundary on which tractions are prescribed. The prescribed
(i)
tractions are t(i) . Also boundary conditions are prescribed on γϕ , which is that part of the
boundary on which displacements are prescribed. The prescribed displacements are given
by ϕ (i) .
To see that this is true, choose for w(i) consecutively e1 , e2 and e3 to regain the equilibrium
equations.
Integrating the above equation yields:
∇ · σ (i) + f(i) · w(i) dω = 0. (2.42)
ω
Integrating the previous equation by parts by applying the divergence theorem, yields:
(i) (i) (i) (i)
σ : ∇w −f ·w dω − t(i) · w(i) dγ
(i)
ω γσ
(i) (i) (i)
− σ ·n · w dγ − t(i) · w(i) dγ = 0. (2.43)
γϕ(i) γ (i)
(i)
Introducing the compatibility condition w(i) = 0 on γϕ , does not change the solution of
the original problem. It will however cause the integral over the prescribed displacement
boundary to vanish. By using the following shorthand:
(i) (i) (i) (i) (i) (i) (i)
G ϕ ,w = σ : ∇w − f · w dω − t(i) · w(i) dγ ,
ω γσ(i)
(i) (i) (i)
Gc ϕ , w = − t(i) · w(i) dγ ,
γ (i)
Up until this point, nothing was really done to obtain the work conjugate pairs as was
suggested in the introduction. To obtain these, an interpretation is to be made on the
meaning of w(i) . On the one hand, w(i) are often interpreted as virtual displacements. If this
approach is selected, the work of Laursen and Simo (1993) is followed. The interpretation
of the above weak equations is then that of virtual work. On the other hand w(i) can be
interpreted as virtual velocities. In that case one arrives at the principle of virtual power,
discussed in Bonet and Wood (1997)2. To show the effects on the resulting derivation, both
interpretations are used.
In the next section, we discuss the equilibrium of tractions and join the contact in-
tegrals. In the two sections following that, the derivation is finalised using first virtual
displacements, and then virtual velocities.
As it turns out, this one equation is just as powerful as all the original equations. This is due
to the fact that w(1) and w(2) can be chosen independently. To see this, filling in w(2) = 0
as a test function yields the equation for i = 1. An identical results holds if w(1) = 0 is
substituted, in which case the equation for i = 2 reappears.
The last equation is rather lengthy, so as in Section 2.4, we introduce a composition of
functions to state the same more briefly. The composition is made in the following way:
It is the second term in the previous equation which holds our interest. Expanding this term
for i = 1, 2 leads to a new form for the combined contact integral G c :
(1) (1)
G c (ϕ, w) = − t · w dγ − t(2) · w(2) dγ
γ (1) γ (2)
= − t(1) · w(1) − w(2) dγ . (2.52)
γ (1)
In the combination of the two integrals we used (2.46) As before w(2) is the value of w(2)
at the projection point of x(1).
20 Contact Mechanics
In the following two sections, an interpretation is made for the weighing functions w.
With an interpretation, we can find some additional structure in the problem. Depending on
which interpretation is chosen, different derivations appear. As it turns out, we can recreate
the same form for the contact normal distances paired with the contact normal tractions as
it appeared in (2.31).
The generality follows from the fact that the perturbations are arbitrary. In particular,
for problems that do not include contact we have that:
which states that the fields ϕ (2) and ϕ (1) are independent.
However, upon using contact, things are different. Most notably, ϕ (2) is dependent on
(1)
ϕ on the contact boundary. The variations of both functions need to be coupled as a
result of this dependency.
Using the product rule on the right hand side, (plus using the the symmetry of the inner-
product) leads to:
Dδϕ [d N ] = d N · Dδϕ [n(2) ] + n(2) · Dδϕ [d N ]. (2.57)
By definition of normals we have that n(2) · n(2) = 1. Taking derivatives left and right
of this identity, and again using symmetry of the inner-product leads to:
Now expanding d N = d N n(2) in (2.57), and applying (2.58) leads to the following deriva-
tion for the directional derivative of the normal distance:
Dδϕ [d N ] = d N n(2) · Dδϕ [n(2) ] + n(2) · Dδϕ [d N ]
= d N n(2) · Dδϕ [n(2) ] + n(2) · Dδϕ [d N ]
= n(2) · Dδϕ [d N ]. (2.59)
Since, x(1) is a boundary point, there exists a ξ (1), such that x(1) = (1) (1)
t (ξ ). The point
(2) (2)
x(2) is then known through the mapping t (ξ ). If we interpret d N as a function, which
gives for each boundary point the distance to the master then this function can also be
written as:
Dδϕ [d N ] = Dδϕ [ϕ (1) ] − Dδϕ [ϕ (2) ]. (2.61)
The first partial derivative is easy:
Dδϕ [ϕ (1) ] = δϕ (1). (2.62)
This follows straight from the Definition (2.5).
The computation of the second partial derivative is somewhat more involved, and we
refer back to the aforementioned definition to compute it.
(2)
Dδϕ ϕ (2) = Dδϕ ϕ (2) X
(2) (2)
d
= ϕ (2) X + δϕ (2) X
d =0
(2) (2)
= F · Dδϕ [X ] + δϕ (2) . (2.63)
(2)
In the above we used for the directional derivative of ϕ (2) (X ) the same reasoning as in
(2.12) and (2.23).
(2) (2)
Remains the computation of X , this follows again straight from the definition of X :
(2)
(2)
Dδϕ [X ] = Dδϕ [ (2) 0 ξ ]
(2) (2)α
= Tα Dδϕ [ξ ]. (2.64)
Combining (2.63) and (2.64) results in:
(2)α
Dδϕ ϕ (2) = τ (2)
α Dδϕ [ξ ] + δϕ (2) . (2.65)
Substituting (2.62) and (2.65) in (2.61) yields the following expression for the directional
derivative of the normal distance vector:
(2)α
Dδϕ [d N ] = δϕ (1) − δϕ (2) − τ (2)
α Dδϕ [ξ ]. (2.66)
22 Contact Mechanics
We now have to insert the result that was found for the normal distance vector back into the
contact integral. Let us first again state (2.52) for the case where w is interpreted as virtual
displacements:
G c (ϕ, δϕ) = − t(1) · δϕ (1) − δϕ (2) dγ . (2.67)
γ (1)
In it we recognise immediately the term δϕ (1) − δϕ (2) , which is also present in the
derivative of the normal distance vector. Making the appropriate substitution of (2.66) in
(2.67) results in:
(2)α
G c (ϕ, δϕ) = − t(1) · Dδϕ [d N ] + τ (2)
α Dδϕ [ξ ] dγ . (2.68)
γ (1)
(1) (1)
t(1) = tN + tT
(1) (1)
= t N n(1) + t T
= −t(1)
N n
(2)
+ t(1)
T . (2.69)
The last step in this equation holds, because only when the bodies are in contact is t N non-
zero, and do the normals have opposite directions. Substituting the result (2.69) into (2.68)
and using (2.59) gives:
(1) (1) (2)α
G c (ϕ, δϕ) = − −t N n(2) + t T · n(2) Dδϕ [d N] + τ (2)
α D δϕ [ξ ] dγ
γ (1)
(1) (1) (2)α
= t N Dδϕ [d N] − t T · τ (2)
α Dδϕ [ξ ] dγ . (2.70)
γ (1)
In which we used the orthogonality of the tangential vectors and normal vectors. Defining
t(1)
T,alpha
= −t(1) (2)
T · τ α , finally results in:
(1) (2)α
G c (ϕ, δϕ) = t N Dδϕ [d N] dγ + t(1)
T,α Dδϕ [ξ ] dγ . (2.71)
γ (1) γ (1)
Or dropping the superscripts and introducing δd N = Dδϕ [d N] and for the sliding paramet-
(2)α (2)α
ers δξ = Dδϕ [ξ ], we obtain the more familiar:
(2)α
G c (ϕ, δϕ) = t N δd N dγ + t T,α δξ dγ . (2.72)
γ (1) γ (1)
Notice the equivalence of the above contact integral for the normal displacements with the
result we had for the elastic potential in Section 2.4.
2.6 The weak form of contact 23
Applying (2.24) and the definition of normal velocity, the previous equation can also
be written as:
v N + v T = −v N n(2) + ξ̄˙ (2)α τ (2)
α . (2.75)
Taking the directional derivative in the direction of δv of the previous results in:
The derivatives with respect to the velocity for the normal and tangential vectors are 0,
since they do not depend on the velocity.
Differentiating (2.74) in an identical manner yields
Using (2.69) in (2.52) results in the following contact integral part for the virtual velo-
cities case:
G c (ϕ, δv) = t N Dδv [v N] dγ + t T,α Dδv [ξ̄˙ (2)α ] dγ , (2.79)
γ γ
By doing some renaming:
δv N = Dδv [v N ], (2.80)
δv T = D [ξ̄˙ (2)α ] τ (2) .
δv α (2.81)
The discussion of the remainder of this thesis is limited to the method based on the
virtual displacement. The reasons we have given the virtual velocity based discussion are
twofold:
24 Contact Mechanics
• In a lot of articles, authors prefer the use of virtual velocity based schemes. Therefore
it is useful to show that the same type of structure arises even when using virtual
velocity schemes.
• To illustrate the impact of interpretation on a mathematical structure. Obviously the
complete derivation of the structure is different even though the original equations
were completely equivalent. This realisation is important when dealing with inter-
pretation of symbols in any type theory forming.
2.7 Conclusions
In this chapter, the derivation was presented for the weak form of equilibrium including
contact. The equivalence of the necessary conditions of the contact problem in elasticity
and the final one for the virtual displacement case is shown. From this it can be seen
that there is a strong coupling between variational problems and the minimisation of func-
tionals. However, the incorporation of Coulomb frictional constraints into the potential
minimisation based scheme is theoretically not possible, since minimisation problems al-
ways yield a symmetric variational operator. The Coulomb frictional constraints depend
on the normal traction, whereas the normal traction does not dependent on the result of the
friction. It is for this reason the derivations for the contact problem need to be made in the
variational notation.
An advantage of the insight gained by the fact that there exists an association between
the variation and minimisation problems, is that we can use methods that solve the contact
problem for the minimisation problem and use it for the variation problem. There is a
rich body of literature for the former problem, and several algorithms are discussed in the
following section.
To solve the contact equations, we need to solve the following variational problem:
The first functional in this equation can be computed in the usual sense by a finite element
method. The second functional contains the contribution of contact. This latter functional
can be expanded by interpreting w either as virtual displacements or as virtual velocities.
For virtual displacements, the result is:
(2)α
G c (ϕ, δϕ) = t N δd N dγ + t T,α δξ dγ . (2.84)
γ (1) γ (1)
It is these two integrals that need to be computed to again have a fully defined finite element
procedure. In these tow equations, only the values of the tractions are unknown. The
computation of these tractions is the topic of the next Chapter.
Chapter 3
R EGULARISATION M ETHODS
3.1 Introduction
In this chapter, various solution methods are proposed and discussed that can be used to
enforce the impenetrability constraint in the finite element framework. We note that the
impenetrability constraint is an inequality constraint: the signed normal distance between
the two contacting boundaries must remain non-negative throughout the simulation.
Most methods employed in the simulation of contact are based upon enforcing equality
constraints, in which the contacting points are constrained to have exactly zero distance.
Since the contact constraints are inequality constraints, a method is required which selects
those constraints that are active. An inequality constraint is said to be active, if it is on its
bound in the solution of the problem.
Unfortunately, there is no simple procedure which is known to always make the correct
selection of which constraints are active for a general inequality constrained problem. An
active set selection strategy can portend extremely long simulation times if no care is taken,
and possibly incorrect final results otherwise. What may go wrong is discussed in more
detail later in this chapter.
Though the active set selection problem can be cumbersome at some point in the finite
element simulation, in general it is somewhat mollified by taking small time increments.
The latter procedure ensures that no large changes in contact status occur during an incre-
ment. This in turn means that the active set of constraints will not change too much within
an increment.
After application of an active set selection scheme, we obtain an equality constrained
problem which can be solved directly by using a mixed method or a constraint elimination
method. An alternative procedure is to hide the explicit active set selection process by
attempting to estimate the value of the normal traction. In this approach the normal traction,
which in fact is a Lagrange multiplier, is assumed to be a function of the value of the
26 Regularisation Methods
signed normal distance. The latter approach is based upon the strong coupling between
optimisation problems and variational problems. This type of approach is known as a
regularisation, and methods that employ it are discussed in greater detail in this chapter.
This chapter is organised as follows: First the discretisation of smooth contact problems
is given. This discretisation is required to explain some of the methods, and also to give an
insight as into where problems might occur. In Section 3.3 a short overview is given of the
different classes of methods that are currently available to solve the contact problem. Next
in Section 3.4, the framework of the regularisation methods is discussed. Subsequently in
Sections 3.4.1 till 3.4.3, the penalty method, the method of augmented Lagrangians and the
modified barrier method are discussed. Next in Section 3.5 the regularisation of friction is
discussed. Finally, in Section 3.6, we give a summary and present some conclusions.
3.2 Discretisation
In this section the discretisation is discussed of a single slave body with a non-discretised
undeformable master body having a smooth boundary description. The more general case
of a discretised slave body with a discretised and possibly deformable master body is post-
poned until Chapter 4. The reason for introducing this discretisation now, is that some
of the advantages and disadvantages of several methods are best discussed in the discrete
setting. This is especially true for the mixed and constraint methods. The latter method is
even only presented in the discretised setting. The fact that we do not immediately pose
the most general discrete form is that it unnecessarily draws the attention to problems and
details which are not related to the solution method employed. The result of choosing
the discrete problem in the proposed setting, is that the distance functions have the same
smoothness as in the continuum setting. In the next chapter we see what happens when this
is not the case. As it turns out, this may introduce a bigger problem than the selection of a
constraint satisfaction method.
h
the location of its corner points. These corner points are consequently the nodes for the
quadrilaterals. Typically, neighbouring elements share nodes, so that their boundary shapes
are compatible.
In the mesh the boundary is no longer tracked exactly. It is in itself approximated
through the discretised boundary h . Hence, not all original reference points X can be
addressed, instead we will refer to points in the approximated reference configuration h
as Xh .
The members ϕ h of S h are assumed to take the form:
m
ϕ h (Xh , t) = xi (t)Ni (Xh ), (3.1)
i=1
where xi (t) is the nodal displacement of node i at time t, m is the number of nodes in the
discretisation and Ni is the composition of element shape functions (or interpolation func-
tions) for node i . For a more complete explanation refer to Hughes (1987); Belytschko et al.
(2000). From this equation, we see that the functions ϕ ht tracks the evolution of discretised
material point Xh , just as ϕ did this for X. Moreover, it is assumed that ϕ h (Xh , 0) = Xh .
The functions ϕ h can also be written in the more usual displacement based formulation
1
as :
m
ϕ h (Xh , t) = xi (t)Ni (Xh )
i=1
m
m
= xi (0)Ni (Xh ) + ui (t)Ni (Xh )
i=1 i=1
m
= Xh + ui (t)Ni (Xh ). (3.2)
i=1
x = [x1 x2 . . . xm ]T . (3.3)
Each ϕ h is completely defined through this column vector. Creating an identical type of
stacking for the interpolation functions Ni into a matrix N and omitting arguments, we can
write:
xh = ϕ h (Xh ) = N x. (3.4)
in which
N1 0 0 N2 0 0 Nm 0 0
N = 0 N1 0 0 N2 0 ... 0 Nm 0 . (3.5)
0 0 N1 0 0 N2 0 0 Nm
By inserting the discretisation ϕ h into the above equation, and using the fact that x is now
the real independent argument, we arrive at:
In the above, we also changed the impenetrability condition as a constraint over the ori-
ginal boundary to one over the discretised boundary. There are only a finite number of
nodes lying on the boundary, so we can replace the infinite number of constraints with a
finite subset of them. The idea being, that if the constraints are satisfied at a number of
sufficiently densely distributed points, that they are satisfied (sufficiently accurate) every-
where. The latter property follows from the dependency of the complete boundary on only
a finite number of nodes. Each of the selected constraints is given a label i , and all labels
are collected into a set I . Which constraints are chosen is decided by the selection of a
boundary integration algorithm: Each integration point corresponds to a constraint. Integ-
ration points as constraints is more extensively discussed in Chapter 4. For now, it suffices
to say that the number of constraints is reduced to a finite number. Using this observation
with (3.7), we arrive at the discretised version of our initial optimisation problem.
This is the same equation as (3.7), with the slight difference that there are now only a finite
number of constraints. The minimum in (3.8) is characterised through the following first
order conditions, that are also known as the Karush-Kuhn-Tucker or KKT-conditions:
∇x P h (x) + λ N,i ∇x d N,i = 0, (3.9a)
i∈I
λ N,i d N,i = 0, (3.9b)
λ N,i ≤ 0, (3.9c)
d N,i ≥ 0. (3.9d)
The above conditions specify a saddle-point in the (x, λ N ) space. For a more extensive
survey of these properties the reader is referred to Luenberger (1973) or Bazaraa and Shetty
(1979).
We note that the fact that one of the bodies was deformable or not was not used in
obtaining the first order conditions in (3.9). Hence, the manner of discretisation is valid for
an arbitrary (elastic) problem.
The only influence that the non-deformability has on the formulation as it is stated
in (3.7) is on the properties of the distance functions d N,i . If the non-discretised master
boundary is sufficiently smooth, the radius of curvature is not too small and if the con-
strained node is not too far from the boundary, then the distance function is continuously
differentiable. Continuous differentiable distance functions are required for the correct the-
oretical discussion of the methods which we review in this chapter. Later we can show that
some relaxation of this assumption is possible.
A C
Distance
B
Boundary
C B
0 Particle Path D
D
Particle Path
G(ϕ h , w) + G c (ϕ h , w) = 0, (3.10)
In this equation ei represents the vector of all zeros, except for a 1 at the i -th location.
Furthermore, the sum runs over all i -s that are not lying on the prescribed displacement
boundary. An identical result holds for G c . Thus, we end up with the following equation:
wi G(ϕ h , Nei ) + G c (ϕ h , Nei ) = 0. (3.14)
i∈Ifree
Since this is to hold for each wi , we end up with the following vector equation:
Suppose we know the set of active constraints Ia ⊂ I . In that case, the solution (x, λ N )
of this set of equations is a saddle point of the reduced Lagrangian L. The reduction means
in this case that the inactive constraints are omitted. The first order conditions for a saddle
point are then:
∇x L = 0
. (3.17)
∇λaN L = 0
This results in the following equations for this particular situation:
∇ P+
x i∈Ia λ Ni ∇x d N,i (x) = 0
. (3.18)
d (x) = 0
Ni
From (3.18), one can see some structure arising which is typical of mixed methods:
loss of positive definiteness; The result not surprising, because the solution to the original
problem is a saddle point for the Lagrangian: the solution is located in such a point that
it is (locally) a minimum for variations in the displacements, and (locally) a maximum for
32 Regularisation Methods
variations in the Lagrange multiplier. The latter property is retained after discretisation.
When we apply a Newton scheme to solve the set of equations, we end up with a stiffness
matrix that is in part positive definite and in part negative definite. This type of mixed
matrix structure can cause stability problems upon inversion of a discretised step.
Another drawback is that the method introduces a large number of additional unknowns
in the discretised setting. Especially in plate forming simulations, where the number of
degrees of freedom nearly doubles upon employing a mixed scheme.
Moreover, since the method is based on active inequality constraints, an active set
method is required. The selection of the active set can be troublesome and time-consuming,
since it is not evident at the beginning of an increment which constraints are going to be
active at the end of the increment.
The aforementioned three reasons indicate that employing a mixed method does not
seem a wise choice if one wants to find the solution efficiently and in a stable manner.
needs to be repeated several times. This then is known as sequential quadratic program-
ming (SQP). This approach is very efficient and accurate, for equality constrained prob-
lems. Remains the problem of the selecting the active set method, and how it interacts with
friction problems.
The application of this method to the general form is completely equivalent to the
above. But in that case the matrix K is assumed to be the stiffness matrix, and the vec-
tor F is the residual vector without contact.
The solution of the SQP problem can be achieved in various ways, one could employ
pivoting methods such as Lemke’s method, see Chabrand et al. (2001). Or the constraints
can be immediately eliminated from the system to leave a reduced system. The latter is the
method used in the finite element code M ARC, also by Farahani et al. (2000, 2001).
After having dealt with the mixed method and the constraint method, as well as hav-
ing glimpsed the connection between optimisation problems and variational problems yet
again, we set out in the next sections with a discussion on regularisation methods.
minϕ P(ϕ)
(3.19)
s.t. d N (ϕ(X)) ≥ 0 for all X ∈ .
If we would only have a problem of the form min f (ϕ), then (the discretised version)
could be solved by employing an unconstrained local search strategy such as a Newton-
Raphson or a quasi-Newton method. Under the assumption that we can solve unconstrained
problems, we convert (3.19) into a sequence of unconstrained optimisation problems whose
limit is the solution ϕ ∗ of the original problem.
34 Regularisation Methods
The form we are seeking can be obtained by appending to the objective P a function
which penalises the (near) violation of a constraint in a single point. Typically such
a function employs a set of parameters q at each constrained point, which modifies the
function in such a way as to obtain an exact penalisation. If an exact penalisation could
be obtained, then the solution of the extended potential would yield the solution of (3.19).
Exact interpretations of how these parameters are supposed to work seem quite abstract
now, but each of the methods discussed in the following sections gives an example which
will make things more clear.
The extended unconstrained problem now takes the form:
min P (ϕ) + (d N (ϕ(X); q(X)) d. (3.20)
ϕ
In the following sections, we introduce various methods from the optimisation liter-
ature which have found their way into applications in contact mechanics. The penalty
method and method of augmented Lagrangians are the most widely applied. We propose
the modified barrier and smooth penalty methods to overcome some of the problems that
are occurring when equality constrained based methods such as the penalty method are
used. Another method which was proposed in Zavarise et al. (1998), which also falls into
this class of methods, is not further discussed, since no further applications of it are known
in the literature, but it serves to illustrate that the framework can lead to a plethora of
methods.
3.4 The regularisation framework 35
0 if x < 0,
x = (3.24)
x if x ≥ 0.
x ≥ 0 ⇐⇒ −x = 0. (3.25)
The above conversion comes at a cost, however, since the Macaulay bracket is not dif-
ferentiable at x = 0. The non-differentiability can have serious consequences for the
convergence of a Newton process that is used to solve the discretised problem.
The function for the penalty problem in contact mechanics can now be defined as:
p
(d N ; { p}) = −d N 2 , (3.26)
2
in which the only parameter used is a penalty value p. This is supposed to be a large num-
ber, such that when the constraint value is violated it gives a significant contribution to the
extended unconstrained problem. Consequently the larger the penalty value is, the smaller
the constraint value ought to be for the solution of the problem (3.21). The derivative of
with respect to d N is:
∂ p
= 2−d N H (−d N) · (−1) = − p−d N . (3.27)
∂d N 2
In the equation above H (x) represents the Heaviside function, which is defined as:
0 if x < 0,
H (x) = (3.28)
1 if x ≥ 0.
The result now follows from using the chain rule and the fact that H (x) · x = x.
The result for the normal traction in this case is:
∂(d N ; { p})
t N (d N , p) = = − p−d N . (3.29)
∂d N
36 Regularisation Methods
2 2
1.5
1.5 1
functional-value
functional-value
0.5
1 0
-0.5
0.5 -1
-1.5
0 -2
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
x-value x-value
For an example take a look at Figure 3.3. In this figure the penalty functional is given
for the following problem:
1
min (x + 1)2 (3.30a)
x 2
s.t. x ≥ 0. (3.30b)
The extended penalty functional for the above problem is:
1 p
(x + 1)2 + −x2 . (3.31)
2 2
In Figure 3.3 the penalty is chosen as 1. Also drawn in the figure is the traction estimate
function, which is the derivative of with respect to d N . The minimum can be seen to lie
at x = − 12 , with an estimate for the Lagrange multiplier of λ = − 12 . In Table 3.4.1 a set of
penalty values is presented versus the convergence of x for the example.
Table 3.1: Convergence of the example problem with the penalty method.
Although the method is conceptually simple, there are a number of drawbacks to it.
First of all there is the notable introduction of a non-differentiability upon contact. The
second problem is that only for the penalty p going to infinity can the constraint violations
be eliminated. The final problem is that the penalty p can not be too large because this
worsens the condition number of the matrix that is obtained in the numerical solution of
the problem. The last two problems counteract, and for a successful application some
fiddling with the penalty parameter has to undertaken.
3.4 The regularisation framework 37
In the next section, the method of augmented Lagrangians is discussed, which attempts
to solve the contact problem accurately, while keeping the penalty parameter small.
The solution which satisfies (3.32) is denoted by (ϕ ∗ , λ∗N ). Here λ∗N is the Lagrange multi-
plier that is equivalent to the normal traction at the solution point.
From the penalty method, employing (3.22), we already learned that − p−d N at the
solution of the regularised problem, gives an estimate for t N from (3.29). If we could
somehow improve our estimate of the Lagrange multiplier using this previous value, then
it would improve the accuracy in a subsequent minimisation problem.
This can be done by choosing to contain part of the actual Lagrangian, but keeping
the Lagrange multipliers fixed. The function now takes the following form:
1
(d N ; { p, λ N }) = −(λ N + pd N)2 , (3.33)
2p
where we stress that in this case the λ N is not a variable, but a constant. The complexity
of the form of the regularisation in (3.33) arises from the inequality-equality constraint
conversion. However, it is not that different from the penalty functional, which could have
been written as 2p −d N 2 = 21p − pd N 2 . From this equation we can see that in the penalty
method the Lagrange multiplier is approximated as 0.
The resulting traction can be expressed as the derivative with respect to d N of the pre-
vious equation:
t N = − −(λ N + pd N ) . (3.35)
If the estimate of λ N would be the correct Lagrange multiplier λ∗N , then the result after
minimisation of (3.20) would yield the solution of (3.6). However, in general we do not
have the correct values of λ N . After solving (3.20) with the values of λ N set to 0, will
result in an estimate to the solution. Let us call this result ϕ 1 (not to be confused with the
field ϕ restricted to (1) , which is denoted by ϕ (1) ). After completion of the computation,
we also have an approximation t N (d N (ϕ 1 )) to the correct multipliers λ∗N . Let us name this
approximation λ1N . So upon completion we have an approximation (ϕ 1 , λ1N ) to the correct
solution of (3.6).
The values λ1N are better estimates to the optimal Lagrange multipliers than the values
λ N , which were set to 0. That this is so, can be deduced as follows. If the distance for
0
a particular point X is positive, then from (3.34), we can see that t N remains 0. Thus in
38 Regularisation Methods
2 2
1.5
1.5 1
functional-value
functional-value
0.5
1 0
-0.5
0.5 -1
-1.5
0 -2
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
x-value x-value
that case, λ1N = λ0N = 0. However, if there is penetration then obviously the Lagrange
multiplier is chosen too small. At the converged solution, the Lagrange multiplier is larger
in magnitude, but still too small, since there will still be penetration. Hence, the result of
restarting the minimisation process, only now with using λ1N as the estimating constants
will yield an better approximation. Name the new results (ϕ 2 , λ2N ). After finishing this, λ2N
would again form a better value, etc. The process of updating the λ N with better estimates
is called augmentation. We augment the estimation to the Lagrangian multipliers, which
explains the name. Indeed for equality constrained problems, the procedure converges to
the exact solution, see for example Bazaraa and Shetty (1979); Bertsekas (1982). There
are many articles written on the application of the augmented Lagrangian method in the
finite element method, among them are Laursen and Simo (1992, 1993); Simo and Laursen
(1992); Zavarise et al. (1995); Refaat and Meguid (1997); Pietrzak and Curnier (1999).
An illustration of the process is given in Figure 3.4. Again the problem that is used for
the illustration is given by:
1
min (x + 1)2 (3.36a)
x 2
s.t. x ≥ 0. (3.36b)
1 1
(x + 1)2 + −(λ + px)2 . (3.37)
2 2p
Using again a penalty p of 1, and the estimate for the Lagrange multiplier that was obtained
from the penalty method as − 12 , the result as in Figure 3.4 is obtained as the functional to
minimise. From the picture one can see that the location of the minimum of the extended
potential is a better estimation of the actual minimum which is lying at 0 than the initial
value of − 12 that was obtained from the penalty method. The value that is obtained after
3.4 The regularisation framework 39
one augmentation is x = − 14 . On the right in Figure 3.4 the estimate for the augmented
Lagrangian method is shown for λ = − 12 . In Table 3.4.2, the convergence of the example
problem is given for several different penalties. In each case, the initial value for the Lag-
range multiplier λ(0) is 12 .
Table 3.2: Convergence of the example problem with the method of augmented Lagrangians.
The method of augmented Lagrangians is designed for use with equality based prob-
lems. This means, that the non-differentiability introduced by the Macaulay bracket hinders
the convergence of an augmented Lagrangian procedure in the same way as it hinders con-
vergence in the penalty method. There are some advantages though: The penalties can be
chosen a lot smaller than with the penalty method, since the augmentation procedure will
help to converge to the correct solution. This helps the stability of the numerical method
employed. Moreover, it is possible to get to the actual solution, which was not the case with
the penalty method. An additional advantage is that if we increase the penalty parameter
amidst augmentations, we gain superlinear convergence. A feat, which was also sought in
Zavarise and Wriggers (1999) by an acceleration scheme.
In the next section, we look at a method that attempts to overcome the problems of
the non differentiability and tries to retain the advantageous properties for the method of
augmented Lagrangians.
2 2
1.5
1.5 1
functional-value
functional-value
0.5
1 0
-0.5
0.5 -1
-1.5
0 -2
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
x-value x-value
The intent of this addition of a penalising function is clear: For large positive values of
d N , the contribution to the extended potential is negative. For small values near zero the
contribution is big and positive. In effect, a constraint encounters a barrier in its range of
allowable values.
There are some drawbacks to this approach, however. First of all, as it is the case with
the penalty method, the barrier method is only accurate for large penalties, and generates
poorly conditioned problems for penalties that are too large2. What can be considered even
worse, is that the problem is only defined for feasible solutions. Hence, no constraint may
be violated anywhere during the computation. This may pose a problem in some optimisa-
tion problems where finding a feasible initial configuration is sometimes already as hard
as finding an optimal solution. This is not necessarily true for finite element simulations
with contact, however, although small violations due to meshing may occur in the initial
configuration.
To overcome the above problems, both an augmentation method is required, as well as
a solution to the problem of the penalisation function ’s existence for negative values of
the constraint functions. This is achieved by setting to the following functional form for
feasible solutions:
1
(d N ; { p, λ N }) = λ N log (1 + p d N ) . (3.39)
p
In this equation, λ N is a fixed estimate for the Lagrange multiplier, as it was in the augmen-
ted Lagrangian procedure. The inverse penalty 1p is known as the barrier parameter.
The concept again can be explained by the following reasoning: As d N approaches
− 1p , the logarithm adds a significant penalty to the the extended unconstrained potential as
defined in (3.20). Whereas, for d N large, the function offers a negative contribution and is
preferred by the global minimisation algorithm.
2 For linear optimisation problems, there is a strong interest nowadays for barrier methods as an interior point
approach. By using the barrier method, solutions to the linear optimisation problem can be found in polynomial
time.
3.4 The regularisation framework 41
Table 3.3: Convergence of the example problem with the modified barrier method.
The numerical approach that is followed is now roughly equivalent to the augmented
Lagrangian approach for equality constrained problems. A value for p is chosen, as well
as a fixed value for the estimation of the Lagrange multiplier. The minimisation for the
extended potential is then solved using these parameters. After that, using the minimiser ϕ,
the estimate for the Lagrange multipliers are improved, p is increased and we start again. It
was proven in Polyak (1992) that under certain conditions this approach converges linearly
to a solution of (3.20).
To improve the rate of convergence, the constraints are scaled. This can be useful if for
example the constraints appear with different orders of magnitude. The necessity of this
is discussed in Breitfeld and Shanno (1994). If all constraints are of the same type scaling
42 Regularisation Methods
is not necessary, but for sake of completeness they are included in the discussion. To this
end, a scaling factor si for each of the constrains is introduced. The value of the constraint
is then scaled to be d N,i (x)/si . Combining this with the unconstrained subproblem as in
(3.41) results in:
# $
1 d N,i (x)
F(x, p, λ N , s) = P(x) + λ N,i log 1 + p
p si
i∈I
# $
1 1
= P(x) + λ N,i log si + p d N,i (x) (3.42)
p si
i∈I
1 1
= P(x) + λ N,i log si + p d N,i (x) − λ N,i log (si ) .
p p
i∈I i∈I
Notice that the last term in (3.42) contains only terms in p, λ N and s. Thus, if we take
the minimum with respect to x for set values of these parameters, the last term does not
influence the value of the optimal x. Consequently, minimising (3.42) is equal to:
! "
1
min F(x, µ, λ N , s) = min P(x) + λ N,i log si + p d N,i (x) . (3.43)
x x p
i∈I
An additional limitation that may occur, is that the function F(x, p, λ, s) is not defined
for all possible values for d N,i (x). This is due to requiring that d N,i (x) > − spi , since
otherwise the logarithm is not defined. To overcome this limitation we introduce a quadratic
continuation of the logarithm, from somewhere in the admissible region. This continuation
location is defined through the parameter β, which is some relative point. If d N,i (x) ≥
− βsp i , we use the logarithm. If d N,i (x) < − βsp i , we use the quadratic continuation. The
parameter β should lie within the range (0, 1). The problem as it is defined in (3.43)
becomes:
F(x, p, λ, s, β) = P(x) + (d N,i (x), λ N,i , p, si , β). (3.44)
i∈I
In the previous formula is the combination of the logarithmic term and the quadratic
extrapolation term. It is defined as:
λ N,i
p log(si + p d N,i (x)) if d N,i (x) ≥ − βsp i ,
(d N,i (x), λ N,i , p, si , β) = 1
a 2 if d N,i (x) < − βsp i .
2 i d N,i (x) + bi d N,i (x) + ci
(3.45)
In this equation ai , bi and ci are chosen such that the continuation and the logarithmic
part attach twice continuously differentiable. From this conditions, it follows that they are
defined as: 2
(k) − p(k)
ai = (k) ,
(si (1 − β))2
(k) p(k) (1 − 2β)
bi = (k) , (3.46)
si (1 − β)2
(k) β(2 − 3β) (k)
ci = + log(si (1 − β)).
2(1 − β)2
3.4 The regularisation framework 43
The λ N,i are initialised to 1. Their correct value is computed through application of an
augmentation scheme. Using this augmentation scheme and an updating strategy for the
barrier parameter 1p , it was shown by Breitfeld and Shanno (1994) that a subsequence of
points (x k , λkN ) generated by the algorithm converges to a Karush–Kuhn–Tucker point of
(3.40).
In the next section we give a complete presentation of the algorithm that is used to solve
(3.40).
The Algorithm
We now present a precise description of the barrier algorithm, such as employed in the non-
linear optimisation framework for elastic problems only. The adaptation of this algorithm
for use in non-linear finite element simulations is straightforward.
• Choose x (0) ∈ R N . This is the starting point for the computation. In an incre-
mental computation, x (0) is best chosen as the result of the previous increment.
• Choose τ > 0, the outer loop termination criterion. This termination criterion
is approximately equivalent to the unbalance termination criterion. 10−3 is a
reasonable choice. It is a dimensionless parameter.
• Select a sequence of barrier parameters { p(k) }k≥0 , which is ascending. In Breit-
feld and Shanno (1994) p(k+1) is chosen as 10 p(k).
• Select a sequence { (k) }k≥0 , the inner loop termination criterion. A possible
selection is to choose this sequence equal to the inverted p sequence.
(0)
• Select λ N,i = 1, for i ∈ I .
• Choose 0 ≤ β ≤ βu < 1, the relative extrapolation point.
(0)
• Choose the scaling terms si = min{max{1, −d N,i (x (0))}, su } for i ∈ I , where
su is some upper bound for the scaling terms.
• Compute the extrapolation coefficients for the i -th barrier function ai(0), bi(0)
and ci(0) by (3.46). These coefficients are completely determined by demanding
to be twice continuously differentiable.
• Set k = 0.
Step 2. Check convergence: We now test whether the convergence criterion is met. First
we define v (k+1) according to:
(k)
(k+1)
| λ N,i d N,i (x (k+1) )|
i∈I
v = % . (3.52)
(k) 2
i∈I λ N,i
If v (k+1) < τ then stop. The parameter measures the satisfaction of the Kuhn–Tucker
complementarity condition. The satisfaction of the unbalance criterion is already
established at the end of Step 1.
Step 3. Update parameters: We now update the scaling parameters, by scaling with re-
spect to the new point:
(k+1)
si = min{max{1, −di (x (k+1) )}, su }. (3.53)
λ(k+1)
N,i
= # (d N,i (x (k+1) ), λ(k)
N,i
, p(k) , si(k) , β), (3.54)
3.5 Regularising Friction 45
where # denotes the derivative of with respect to the first variable only.
The motivation for choosing λ N in this manner comes from (3.49). If we expand this
equation we find:
d (k) (k)
∇ P(x (k+1) ) + (d N,i (x (k+1) ), λ N,i , p(k) , si , β)∇d N,i (x (k+1) ) = 0.
dd N,i
i∈I
(3.55)
(k+1)
Thus using the solution of (3.49), we get that selecting λi in the proposed way
we have: (k+1)
∇ P(x (k+1) ) + λ N,i ∇d N,i (x (k+1) ) = 0. (3.56)
i∈I
In this we can recognize part of the conventional variational form of the equilibrium
equations including contact. It is, however, not completely the same, since the λ N,i
are held constant.
Step 4. Update coefficients: What remains to be done is to update the extrapolation coef-
ficients used in the barrier functions . These coefficients are fully determined by
enforcing to be twice continuously differentiable, and are presented in (3.46).
To achieve fast convergence some fine tuning is required for the barrier sequence 1p .
Regardless of the fine tuning, Breitfeld and Shanno (1994) showed that the algorithm has
at least a convergent subsequence for any such sequence. However, in this proof it was
assumed that functions P and d N,i where twice differentiable functions. This is for the
discretised finite element method not generally the case, but can be enforced if a good
integration scheme is used. We return to this topic in Chapter 4.
The way in which the tangential tractions are computed when slipping is by employ-
ing a regularisation method as with the normal contact tractions, and then apply a return-
mapping scheme (See also Laursen and Simo, 1992; Giannakopoulos, 1989) In fact the
return mapping is just a scaling of the frictional tractions to satisfy the Coulomb friction
constraint.
The lay-out of this section is as follows: In Section 3.5.1, the penalty method in friction
is introduced. Next in Section 3.5.2 the method of augmented Lagrangians in the case of
friction is discussed.
First let us name tζ = ζ . Also, we define pd T as the predictor of the frictional
p
traction t T . The frictional tractions can now be computed by:
p p
tT if t T ≤ |t N |,
tT = tT
p
(3.60)
µ|t |N p
t T
otherwise.
3.6 Conclusions
In this chapter, we discussed several methods that can be employed to solve the contact
problem in finite element methods. First we discussed the mixed method, which is com-
putationally expensive, and the constraint method, that is very efficient but can present
problems for example in plate forming simulations were the detection of contact is very
important. Moreover, there may be too many constraints in the plate forming simulation to
use an effective active set method.
After that, we discussed the general form of regularisation problems, and noted that all
of them stem from a basic mold. Everything is really decided by selecting the form of a
48 Regularisation Methods
In the same fashion, two methods were discussed that enforce friction, they rely on
a predictor-corrector scheme. Only the predictor is determined differently for the two
schemes that were considered. These two schemes are:
• The penalty method for friction, in which case the predictor is set to:
p
t T = pd T . (3.66)
• The method of augmented Lagrangians, in which case the predictor is set to:
p
t T = λ T + pd T . (3.67)
From the predictor, the correct frictional traction is computed, by using either the fixed
estimate to the normal contact traction λ N or the current estimate t N . The frictional traction
now follows from: p p
tT if t T ≤ |t N |,
tT = t
p
(3.68)
µ|t N | t pT otherwise.
T
Chapter 4
D ISTANCE F UNCTIONS
4.1 Introduction
In Chapter 2 the equations were formed that characterise the contact problem. In Chapter 3
several methods were discussed to convert the contact integral to a displacement based form
only. In this chapter, the final piece of machinery is developed and discussed to simulate
contact in forming: The actual evaluation of the contact integral.
In this chapter, it is assumed that a regularisation method is chosen to approximate the
tractions. In that case there is only one fundamental unknown left, which is the computation
of the distance function.
There are three main issues which are paramount to the success of a contact algorithm,
these are stability, accuracy and efficiency. Each of these three issues is also important in
the selection of the regularisation method, and this is the topic of most of the articles that
appear on contact. However, even when such a regularisation method is already chosen,
which does not cause instabilities, can satisfy the contact constraints accurately, and is rel-
atively efficient, still stability problems can arise if the integral is not properly considered.
Moreover the overall accuracy and efficiency depends on other factors than the regularisa-
tion method alone.
Upon the introduction of a discretisation for both the master and slave boundary, mesh
incompatibilities can be introduced between the master and slave body. By mesh incom-
patibilities we mean that the element sizes of master and slave may not fit together when
meshed independently. An alternative would be to mesh the master and slave together, to
form an (initially) compatible contact topology.
Let us consider the first key factor: stability. When there is large sliding contact
between the incompatible meshes, the gradient of the distance function can change ab-
ruptly. This abrupt change can cause stability issues. A careful examination of when these
changes occur, can catch the instabilities before they distort the convergence of the finite
50 Distance Functions
The two aspects are not completely independent. The selection of the most appropriate in-
tegration method depends strongly on how the discretisation is performed. The aforemen-
tioned mesh incompatibilities can cause inaccurate results. Meshing the master and slave
boundaries simultaneously and employing contact elements does not solve this problem:
Due to sliding during the simulation, the contact elements can become severely distorted,
and their results are then again inaccurate.
The third point in our discussion is efficiency: The computation of the distance func-
tions should not take too much time. Having rejected contact elements due to possible
accuracy problems, the projection location of the integration points is no longer fixed.
Consequently, we need an algorithm to find the projection of the integration points on the
master surface. Assuming that the master is meshed using linear elements, this means that
the nearest segment on the master surface needs to be found. Various algorithms have been
proposed. A hierarchical based search method is proposed by Zhong and Nilsson (1988)
and Zhong and Nilsson (1990). The pinball algorithm was proposed by Belytschko and
Neal (1991). An algorithm incorporated in the dynamic finite element package DYNA 3D
is by Oldenburg and Nilsson (1994). The method originally employed in the finite ele-
ment package D IEKA is the block search method, discussed in Atzema (1994) and Car-
leer (1997). Further alternatives are discussed in Wang and Nakamachi (1997); Wang and
Makinouchi (2000) and Wang et al. (2001). To illustrate the efficiency that can be gained by
exploiting structure, the reader is referred to Munjiza and Andrews (1998). In this thesis we
present a new search method, based on computational geometric data structures discussed
in Berg et al. (1997).
The discussion in this chapter follows roughly the reasoning as presented above. In
Section 4.2 the discretisation of the boundary is introduced, and the results for the distance
function. From it a slightly adjusted formulation for the contact integrals is obtained. In
Section 4.3 various integration methods are discussed to approximate the discretised integ-
rals. In Section 4.4, the alternative of smoothing the master boundary is considered. In
Section 4.5 we present a new method to quickly locate the projection locations for each of
the integration points on a piecewise linear boundary. Finally in Section 4.6 the conclusions
are presented.
4.2 Discretisation of the contact integral 51
And since ϕ h (X) = N(X) · x and δϕ h (X) = N(X) · δx, we can convert the directional
derivative in the integral in the following manner:
In the latter equation the specific dependencies of d N on the field variables were omitted
for notational clarity. What remains to be done is to compute the integral between the
parenthesis.
contact boundary, it was argued, does not matter in the computation. Upon discretisation,
however, there may not be a proper contacting boundary left. In most of the situations,
there will no longer be a contacting boundary left, only some contacting points.
Let us clarify the statement with some qualitative examples. The first example presents
the contact between a discretised plate and a non-discretised roll. The roll is modelled
using a circle. The second example is primarily the same as the first, only now the roll is
discretised.
The examples are illustrated in Figure 4.1. The left picture in this figure illustrates
Figure 4.1: Contact between roll and plate, both smooth and discrete.
the contact without penetration in the case the plate is discretised and the roll is smoothly
modelled with a circle. As can be seen, to have a non-penetrating solution, there is only
contact between the roll and the plate at two distinct locations. In fact, each of the surface
segments can have at most one point of contact with the circle in an allowable solution.
This is an essential geometric property of contact between a line segment and a circle. As
a consequence, refinement of the mesh may increase the number of contacting points, but
the number of these points will remain finite. This implies that there is no proper boundary
to integrate over, merely a set of contacting points.
It is clear that such point wise constraints do not generally occur within the actual
problem that is being modelled. In fact, in the actual problem, there will be a region of
contact, not a set of points. The discretisation operation has rendered the two boundaries
geometrically incompatible.
The situation does not improve upon discretisation of both the boundaries as is shown
in Figure 4.1 in the picture to the right. However, the situation does not really seem to
deteriorate either. There are still two contacting points.
To overcome the boundary incompatibilities, one may choose to employ contact ele-
ments, whereby both bodies are meshed simultaneously. However, when there is sliding
occurring in the simulations, the contact elements are no longer guaranteed to have prop-
erly facing boundary segments. The situation then reverses to that illustrated in Figure
4.1 to the right, with possible penetration occurring. This problem can be avoided by em-
ploying an Arbitrary Lagrangian Eulerian scheme (see Huétink, 1986; van der Lugt, 1988).
This type of procedure can keep the contact elements facing each other properly during the
simulation. The result is geometrically compatible, but this comes at the price of having
to perform a Eulerian step at each iteration. Also within some simulations it is nearly im-
possible to keep the contacting nodes at their proper locations and still have an acceptable
mesh to continue the computation. Other attempts to overcome the inaccuracy arising due
4.2 Discretisation of the contact integral 53
4 1
3.5
0.8
3
y-location
2.5
distance
0.6
2
1.5 0.4
1
0.2
0.5
0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
x-location x-location
On the left side of this figure, part of a roll is modelled by a circle. Below it, two
line segments are given, which could be assumed to belong to a discretised body lying
underneath. In the right side of the figure, the distance with respect to the roll is given as a
function of the x-coordinate. From this figure, it can immediately be seen that the distance
function is not smooth with respect to this coordinate.
This type of non-smoothness may cause the gradient to be discontinuous. To make sure
that the evaluation will be accurate, the integral has to be split into its piece-wise smooth
parts. It is assumed here that the master body is smooth, thus as long as the slave body
is smooth, then the distance function will be smooth. The jumps only occur at element
boundaries, thus the integral is to be computed by:
t N ∇x d N d = t N ∇x d N d. (4.5)
h e eh
This all seems still quite trivial. However, assume that X is a material point lying on the
boundary of both 1h and 2h . If the master boundary is continuous, then we will still have
that d N,1 (x; X) = d N,2 (x; X). The gradients at these points may now differ. This has no
54 Distance Functions
influence on the results of the integral if an analytical scheme was used. However, it can
influence the numerical integration scheme.
We wish to make one remark on the analysis: Although the distance function is not
smooth with respect to the boundary discretisation it can still be smooth with respect to
the unknowns (displacements) in the vector x. It is the non-smoothness with respect to x
which influences the convergence characteristics of the Newton-Raphson process. The non
smoothness with respect to the boundary parametrisation influences the final accuracy of
the problem if not taken into account during the integration process.
4.2.4 Conclusions
We can conclude that in general problems that do not employ an ALE method will have
to live with the geometric incompatibilities. In the case that both bodies are to be meshed,
this means that a method which does not employ contact elements is to be preferred. The
contact regions will in that case generally be limited to a number of points. The integral is
furthermore to be split into smooth parts for an accurate result.
Several different methods are shown in Table 4.1. The parameter η̄ is a point in the
interval [−1, 1].
One look at this table, suggests that one would use a one-point Gaussian integration
scheme for data which needs to be second order accurate and a two point Gauss rule if
more accuracy is required. Clearly, the nodal (i.e. trapezoidal) integration scheme is the
least efficient, since it requires two points evaluated for only second order accuracy. Still it
is the most widely applied rule in contact mechanics.
Node
Integration point
What went wrong, is that the integration points do not sufficiently suppress the geo-
metrical freedom of the slave boundary. The situation is comparable to that in incompress-
ibility when zero energy modes arise. So although the contact integral may be computed
within a second order accurate result using the 1D Gauss rule, it does so with respect to
the configuration. And as can be seen from the example, the configuration is not always
acceptable.
The location of the integration points determines the locations where contact is en-
forced. Employing a 2-points Gauss rule, improves the situation, as depicted in Figure
4.3, but does not completely solve it: It is still possible to have penetration without the
algorithm noticing it. The only algorithm that enforces the contact constraints correctly for
this example is nodal integration.
Constraint counting
There is an additional possible problem with the 2 points Gaussian rule that we wish to
mention, and that is the problem of normal traction oscillations. Each integration point
introduces a constraint as was already mentioned. In turn, each boundary node introduces
a degree of freedom. When there are as many degrees of freedom as there are constraints,
then the situation is stable, and sufficiently accurate. For n boundary segments, there are n+
1 degrees of freedom (in the x-direction). For a 1-point Gauss rule there are n constraints
(one integration point for each boundary segment). Consequently, there is one degree of
freedom left, which can cause the zero energy mode.
The 2-points Gauss rule introduces 2n constraints. Which are n − 1 constraints too
many. This results in the property that the contact constraints can divide the normal pres-
sures in an arbitrary way to come up with a valid solution for the aggregated normal trac-
tions in the node. This can lead to normal pressure oscillations. The trapezoidal rule has
exactly as many degrees of freedom as there are constraints.
The example in the previous section plus the constraint counting argument leads us to
select the nodal integration rule.
section, we discuss what additional measures have to be taken if more complex geometries
are involved.
Smooth geometries
When dealing with smooth geometries, we nearly automatically no longer have geometric
compatibility. An additional problem is that by selecting an integration rule a-priori, it
becomes nearly impossible to avoid penetration from the smooth master body into the
slave. As an example consider again the contact of a roll with a plate, and no matter which
integration rule is selected, there is always possible penetration of the roll into the element.
This does not necessarily poses a real problem though. Even if penetration is possible,
this will always be within the interior of a boundary segment with the nodal integration
case. Since the linear boundary segment is only an approximation of the actual smooth
boundary curve, the overlap can be either smaller or larger than observed in the simulation.
The important point here is that the most accurate locations in the simulation, the nodes,
are at least correctly constrained.
Non-smooth geometries
The same problem of geometric incompatibility and the a-priori selection of constraint
location holds for non-smooth geometries. The interpenetrations can also be larger.
Sometimes a two-pass procedure is proposed to overcome the problem (See Chabrand
et al., 2001). In a two-pass procedure, first boundary A is assumed to be the master and
boundary B the slave, and next boundary A is assumed to be the slave and boundary B the
master. The average of the two integrals is then set to be the value of the contact integral.
Although this can be geometrically advantageous in some situations, in others, it can
actually be worse. A definite drawback, however, is again the possible oscillation of the
contact normal pressure over the contact boundary. This oscillation arises from the arbitrary
distribution of pressures over the two integrals, since they are in large parts clearly inter-
dependent. An example of a geometric advantageous situation for a two-pass procedure
versus a one-pass procedure is illustrated in Figure 4.4.
On the left hand side of the figure the two-pass result is illustrated. On the right hand
side the one-pass result or direct evaluation of the contact integral is given. Due to the more
expensive evaluation time of a two-pass procedure, and the fact that it may improve the
geometric configuration, but may deteriorate the quality of the normal pressure, a decision
is made to stick to the one-pass nodal integration rule.
An important remark to be made on the example that is shown in Figure 4.4 is that
the discretisation is really poor. If one would want to approximate the sliding of a block,
the discretisation should be much finer near the corner point. In that case the computa-
tional differences between a one-pass and two-pass procedure are really only limited to
one element.
Other methods claiming better results by employing a separate discretisation of the
interface are the mortar finite element methods, such as discussed in Belgacem et al. (1998),
Hild et al. (1998) and McDevitt and Laursen (2000).
58 Distance Functions
Figure 4.4: Example of a geometric configuration where a 2-pass procedure is better than a 1-
pass procedure.
10 2.8
2.6
8
2.4
Normal Distance
y-location
6 2.2
Particle Path
2
4 1.8
1.6
2
1.4
0 1.2
0 2 4 6 8 10 0 2 4 6 8 10
x-location Particle Path
In the figure two distinct transition points can be noted. The first transition point is
when the projection of the slave node changes from the first to the second segment. This
change of projection is abrupt, and has a discontinuous gradient associated with it. The
second transition point is when the projection of the slave node changes from the second
to the third segment, and this transition appears smooth. For the three-dimensional case
likewise transitions appear along edges.
The abrupt changes in the gradient do not necessarily pose a problem in the compu-
tation of the gradient as was already discussed in Section 4.2.3. It can, however, distort
the convergence of the Newton method used to iteratively solve the simulation. Due to
the sudden change of projection the approximations can either oscillate around the actual
solution, or convergence can be painfully slow. Using line-search can overcome the first
problem, but the global convergence can remain very slow.
When dealing with smooth boundaries, the situation improves when approaching the
contact boundary. Due to finite curvature, there is a distance after which the distance func-
tion is smooth with respect to a trajectory parametrisation, as long as the trajectory stays
within that distance. However, in a discretisation, we do not have a finite curvature at the
node locations of the master. No matter how close the node approaches the master curve,
the gradient will remain discontinuous. This property holds for all the convex sections in
the master boundary
Interestingly, the convex parts seem to be smooth. This is only so in appearance, how-
ever. The closer the slave node moves to the master node, the shorter the transition length
will become. Upon touching the master node, it will even be 0. At that point the gradient is
discontinuous, and remains so after penetration. (As was to be expected, since the convex
and concave parts are complementary).
To stabilise the global simulation procedure, a projection will only be made once per
step: at the beginning of the step. This will solve the non-smoothness with respect to
variables in x.
where use was made of the trapezoidal rule. In this, e1 is assumed to be one of the two
end-points of the segment, and e2 is assumed to be the other end point.
60 Distance Functions
In the smooth case, the trapezoidal rule can be chained for obtaining a more efficient
evaluation, since the end-points of the different integrals coincide. However, due to the
non-smoothness with respect to the boundary parametrisation this is no longer necessarily
so. For an illustration consider Figure 4.6.
Figure 4.6: The corner problem with nodal integration and only one projection
If a chained trapezoidal integration rule would be employed, then only one projection
for the lower left corner of the white box in the figure can be chosen. In the left of the figure
the possible projection segments are given by a fat line. In the picture to the top right, the
final solution is presented if the lower segment is chosen as the projection location. And in
the picture to the bottom right the final solution is presented if the left segment is chosen as
the projection location. Both selections do not deliver the correct final solution.
The reason for this is, that the lower left corner of the white box needs to project both
on the lower as well as on the left segment. The necessity is driven by the non-smoothness
of the distance function with respect to the boundary parametrisation.
111111111111111111111111111
000000000000000000000000000
000000000000000000000000000
111111111111111111111111111 q5
000000000000000000000000000
111111111111111111111111111
000000000000000000000000000
111111111111111111111111111
000000000000000000000000000
111111111111111111111111111
q1 000000000000000000000000000
111111111111111111111111111
slave
000000000000000000000000000
111111111111111111111111111
000000000000000000000000000
111111111111111111111111111
000000000000000000000000000
111111111111111111111111111
000000000000000000000000000
111111111111111111111111111
000000000000000000000000000
111111111111111111111111111
000000000000000000000000000
111111111111111111111111111
q3 q4
000000000000000000000000000
111111111111111111111111111
000000000000000000000000000
111111111111111111111111111
000000000000000000000000000
111111111111111111111111111
v3
000000000000000000000000000
111111111111111111111111111
n
00000000000000000000000000
11111111111111111111111111
000000000000000000000000000
111111111111111111111111111
00000000000000000000000000
11111111111111111111111111
v2
000000000000000000000000000
111111111111111111111111111
00000000000000000000000000
11111111111111111111111111
000000000000000000000000000
111111111111111111111111111
00000000000000000000000000
11111111111111111111111111
q2 00000000000000000000000000
11111111111111111111111111
00000000000000000000000000
11111111111111111111111111
00000000000000000000000000
11111111111111111111111111
00000000000000000000000000
11111111111111111111111111
master
00000000000000000000000000
11111111111111111111111111
v4 11111111111111111111111111
00000000000000000000000000 v1
00000000000000000000000000
11111111111111111111111111
00000000000000000000000000
11111111111111111111111111
boundary are shown. We show how the node q3 of the slave boundary is to be projected
on the master boundary. Note that the node q3 is part of both the edge (q2 , q3 ) as well as
the edge (q3 , q4 ). Both cases are to be considered separately. First the nearest projection
location on the master boundary is to be found which has an opposing normal. From the
picture it can quickly be seen that this is the node v3 for both cases. Now what is the
projection of q3 if it is the integration point for the edge (q3 , q4 )? For this the normals need
to be checked. The normal for (q3 , q4 ) is drawn in the figure marked as n. The maximum
opposing normal for n of the adjacent edges of v3 is that of (v2 , v3 ). Hence, the projection
segment of q3 for (q3 , q4 ) is (v2 , v3 ). A similar argumentation leads to the conclusion that
q3 for (q2 , q3 ) projects on (v3 , v4 ).
This method can also be used for self-contact. The method automatically rejects all
faces that have normals pointing in the same direction.
Most integration points will turn out to project on the same master segment, irrespective
of the segments they are attached to. In this case, we can join the integration points to
save computation time. The method does require additional time in making the global
projections. A complete discussion on how to find the projections efficiently is presented
in Section 4.5.
4.3.6 Conclusions
From the above discussion, it can be seen that the trapezoidal integration rule is best for
use in simulations where both bodies are approximated with piece-wise linear surfaces.
The piece-wise linear boundaries introduces problems when sliding occurs and to stabilise
this, projections should only be made once per step and remain fixed. Additionally, to be
able to perform an accurate computation in the case of non-smoothness with respect to
62 Distance Functions
the boundary parametrisation, more work has to be performed to find all the projection
locations.
To summarise, the reason for the problems in Section 4.2 and this section, is the non-
smoothness of the distance function. By smoothing the master boundary, all these problems
might be overcome. Smoothing the master boundary is the topic of the next section.
a(w) if 0 ≤ w < 1,
c(w) = (4.10)
b(w − 1) if 1 ≤ w ≤ 2.
4.4 Smoothing the master boundary 63
dc dc da db
lim = lim ⇐⇒ (1) = (0). (4.11)
w↑1 dw w↓1 dw du dv
(w, w) if 0 ≤ w < 1,
c(1) (w) = (4.12)
(2w − 1, 2w − 1) if 1 ≤ w ≤ 2.
which is a straight line and is smooth. However, when checking the derivatives at w =
1, one can observe that the curve is not continuously differentiable with respect to the
parameter w. This can easily be overcome by changing the parametrisation.
Since, the distance is a geometric invariant, the surface parametrisation should not mat-
ter. To this end a different kind of smoothness is required, one that does not depend on the
parametrisation. This is achieved through the concept of geometric continuity. A curve is
geometrically continuous of degree, or G 0 if it is connected. A curve is geometrically con-
tinuous of degree 1, or G 1 if the curve tangents have the same direction, but not necessarily
the same magnitude. Another property of a G 1 curve is that it can be made C 1 by selecting
an appropriate parametrisation. In an equivalent manner, higher orders of smoothness can
be defined.
Hence, what we require is a surface that is at least G 1 . A surface that is G 2 would
be even better, since that would mean that the tangent matrix would even be continuous.
However, constructing G 2 surfaces is considerably more expensive than constructing G 1
surfaces, and gaining time was our primary objective. From this we can set up the following
demands:
• The construction time of the smoothing surface should be low, unless we are dealing
with a fixed tool, in which case the smoothing surface needs only to be constructed
once.
• The evaluation of the distance function, as well as its derivatives needs to be not too
complex. Otherwise the time required in computing it, may become too great an
overhead.
• Projecting a slave node on the smoothed master surface should not be too complex
for the same reasons as mentioned for the evaluation of the distance function.
In the following sections these demands will guide the decisions in selecting the way
in which the piecewise linear surfaces are to be smoothed.
64 Distance Functions
Splines
The boundary is assumed to be a polygon, with vertices v1 , v2 , . . . , v N . Each of the vertices
has an a coordinate, which is labeled similarly as x1 , x2 , . . . , x N . The i -th segment of this
boundary is then a line-segment which runs from Xi to Xi+1 . Each segment of the boundary
is going to be replaced by a polynomial curve.
All the polynomial curves pasted together form a piecewise polynomial curve known
as a spline. The global smoothness of the spline is the minimum smoothness over each of
the connections of adjacent polynomial pieces. The maximum degree of polynomial pieces
is called the degree of the spline. A spline with degree p and smoothness q is an element
p p p p−1 p
of the spline space Sq . It is trivial to note that Sq ⊂ Sq−1 , and Sq ⊂ Sq . Also note that
the original polygonal boundary is already a spline, namely one from S01 .
Bézier curves
A Bézier curve is defined in terms of the Bernstein polynomials. The i -th Bernstein poly-
nomial of degree n is defined through:
# $
n n i
Bi (ζ ) = ζ (1 − ζ )n−i . (4.13)
i
n
C(ζ ) = Pi Bin (ζ ). (4.14)
i=0
Bézier curves posses some nice properties that make them a natural choice for model-
ling surfaces (see Piegl and Tiller (1997)):
• Convex hull property: The curves are contained in the convex hulls of their defining
control points: (the Pi ). This property is of importance in Section 4.5, that is why it
is mentioned here.
• End point interpolation: C(0) = P0 , C(1) = Pn .
Using the last two properties, we can thus produce a globally order one smooth cubic
interpolating Bézier spline, which has only local support. The spline is thus from the space
S13 .
The local polynomial curve for the k-th segment takes the form:
3
(k)
Ck (ζ ) = Pi Bi3 (ζ ), (4.15)
i=0
where ζ is a local coordinate which runs from 0 to 1, and the global curve C(ξ ) is the
union of all these curves. To form the k-th local curve segment, we require the points xk
and xk+1 , as well as the average outward unit normals in these points nk and nk+1 . The
normals are computed by averaging the normals to the adjacent line segments of the point.
For the begin and end point no averaging is required.
P1
n3
P2
n2
1111111111111111111111111111111111
0000000000000000000000000000000000
0000000000000000000000000000000000
1111111111111111111111111111111111
0000000000000000000000000000000000
1111111111111111111111111111111111
P3 = x 3 P0 = x 2
0000000000000000000000000000000000
1111111111111111111111111111111111
0000000000000000000000000000000000
1111111111111111111111111111111111
0000000000000000000000000000000000
1111111111111111111111111111111111
v4 0000000000000000000000000000000000
1111111111111111111111111111111111
0000000000000000000000000000000000
1111111111111111111111111111111111
0000000000000000000000000000000000
1111111111111111111111111111111111
0000000000000000000000000000000000
1111111111111111111111111111111111
0000000000000000000000000000000000
1111111111111111111111111111111111
0000000000000000000000000000000000
1111111111111111111111111111111111
0000000000000000000000000000000000
1111111111111111111111111111111111
0000000000000000000000000000000000
1111111111111111111111111111111111
0000000000000000000000000000000000
1111111111111111111111111111111111 v1
0000000000000000000000000000000000
1111111111111111111111111111111111
Figure 4.8: Illustration of Bézier approximation of line segment
From the construction it can be seen that the curve is continuous, since each curve piece
interpolates the nodes at its end points. Furthermore, we can see that the derivatives of
adjacent curve pieces lie along the same direction by the derivative property. Specifically,
66 Distance Functions
the derivative of the local curve segments take the following form:
dCk (k)
2
(ζ ) = 3 Pi+1 − P(k)
i Bi2 (ζ ). (4.16)
dζ
i=0
Conclusions
As can be seen, the construction of a spline, using Bézier polynomials is relatively easy.
By using cubic basis functions the description is even completely local, which makes for
efficient evaluation. This type of efficiency is absolutely necessary if both curves are the
boundary of a body that is modelled as possibly deformable. In the next section we see
how smoothing can be performed on three-dimensional triangulated boundaries.
Splines in 3D
The boundary in the three-dimensional case is assumed to consist of triangles. A triangu-
lation is constructed from vertices v1 , v2 , . . . , v N . Each of the vertices has a coordinate,
which is labeled as x1 , x2 , . . . , x N as in the two-dimensional case. A triangle T in the
triangulation consists of three different vertices v1T , v2T , v3T . Each of the triangles in the
triangulation is to be replaced by a smooth polynomial surface patch.
Again, pasting all the patches together, gives rise to a piecewise polynomial surface,
which is also called a spline. The same smoothness requirements hold as in the two-
dimensional case. In this manner, we can see that a proper triangulation, is a spline from
p
S01 , and our aim is to replace this spline by one from S1 .
Barycentric coordinates
A topic that was not discussed in Section 4.4.2, but used nonetheless is that of barycentric
coordinates. Barycentric coordinates are interior coordinates on a simplex. In the three-
dimensional case, a simplex is a triangle. Points on the triangle can be either given by their
4.4 Smoothing the master boundary 67
From the above, it can be concluded, that although there are three distinct numbers given
for the local coordinates, there are only two really independent ones. This will cause all
points described with barycentric coordinates to lie in the plane of the triangle.
Comparing this with what we did for the two-dimensional coordinate ζ running from
0 to 1 is the second barycentric coordinate for the two-dimensional simplex which is a line
segment.
Bézier patches
A Bézier patch is given with respect to the triangular Bézier basis function. The triangu-
lar Bézier basis functions are given with respect to the barycentric coordinates over the
triangle. The (i, j, k)-th triangular basis function of degree n is given by:
n! i j k
Binj k (λ) = λ λ λ , (4.18)
i ! j !k! 1 2 3
where i + j + k = n, and i, j and k are all 0 or larger.
The local patch is now given by:
S(λ) = Pi j k Binj k (λ). (4.19)
i+ j +k=n
The coefficients Pi j k are called the control points, and are elements of R3 . The triangulation
of the control points in 3D forms the control net. The control net is an approximation to
the patch given in (4.19).
As with their 2D counterparts, Bézier patches posses some nice properties which makes
their use convenient:
• Convex hull property: The surface is contained in the convex hull of the defining
control points. This property is used in Section 4.5.
• Vertex interpolation:
S(1, 0, 0) = Pn00 , (4.20a)
S(0, 1, 0) = P0n0 , (4.20b)
S(0, 0, 1) = P00n . (4.20c)
Furthermore, the boundary curve over an edge only depends on the control points
defined over that edge.
68 Distance Functions
• The tangent planes over the boundaries are coplanar with the triangles of the Bézier
net that are edge incident with the boundaries.
These properties are more easily explained, when illustrated with a picture. In Figure
4.9 the control point locations for a 5th degree Bézier patch are shown when projected
onto the parent triangle. Above each of these locations a control point is specified. The
005
104 014
Figure 4.9: The control point locations for the 5th degree Bézier basis functions.
boundary curve along the left edge of the illustrated triangle only depends on the control
points specified by locations 500, 401, 302, 203, 104 and 005. In an identical fashion, the
boundary curve over the lower edge of the triangle only depends on the control points for
locations 050, 140, 230, 320, 410 and 500. Thus if two patches are to connect continuously,
then the control points over their shared edge should coincide.
The continuity conditions are illustrated for two adjacent degree 4 Bézier patches in
Figure 4.10. In this figure, we see two triangles drawn. The triangle with as vertices
v1 , v2 , v3 has a Bézier net with coefficients Pi j k and is of degree 4. The triangle with ver-
tices v4 , v2 , v3 also has a smoothing polynomial of degree 4 defined on it, with coefficients
Qi j k .
The smoothness conditions from Lai (1997) now tell us that for the patches to join
G 1 continuously, the control net triangles that are edge incident with the joining edge of
the parent triangle need to be coplanar. The exact conditions for the triangle presented in
Figure 4.10 are then for G 0 continuity:
P040 = Q004 ,
P031 = Q013 ,
P022 = Q022 ,
P013 = Q031 ,
P004 = Q040 .
4.4 Smoothing the master boundary 69
v2
P004 Q040
P400 P310 P220 P130 P040 Q004 Q103 Q202 Q301 Q400
v1
v3 v4
We now note that due to the smoothness conditions, the control points at 410, 320 and
311 need to be coplanar with the triangle attached to the displayed one along the edge
(v1 , v2 ). Also, 401, 302 and 311 need to be coplanar with the triangle attached along the
edge (v3 , v1 ). Both the constraints share the control point located at 311. This means that
the system is no longer local, unless some scheme is fixed for choosing the control point
at 311. The scheme we propose is to fix the control points at 320, 311 and 302 such that
the 410, 320 and 311 is also coplanar with 500, 410 and 401. Symmetrically, 401, 311 and
302 are also selected coplanar with 500, 410 and 401. Applying the same scheme later for
the triangle below and to the left, will render the transition smooth of the required degree.
Applying the same scheme in the other corners leaves us with only three control points to
compute: 221, 122 and 212.
The remaining 3 control points are selected by demanding smoothness over the edge.
The approximation for the normal to the edge tangent plane is computed by averaging the
normals of the two connected triangles along that edge. Subsequently, the projection of the
normal onto the boundary edge is subtracted to form a new normal which is orthogonal to
the edge. The control point at 212 is now chosen such that the Bézier net triangle 302, 212,
203 is orthogonal to the corrected normal. This scheme yields a method that, when applied
to each individual triangle, produces a spline surface that is G 1 .
Example
An example of the scheme applied to a data set describing a human scalp, is shown in
Figure 4.11. In the left figure, the original triangulation is shown for the head. In the right
figure a refinement of the triangulation is given, where the refined points are located on
the smoothing surface. Sharp corners, where adjacent normals have an inner product of
less than 0.7 are omitted from the averaging process. This leaves the sharp corners at the
bottom of the head and on the nose. As can be seen, the nose remains angular, and from
this we can conclude that the original mesh was too rough.
4.4.4 Conclusions
The two dimensional smoothing problem turns out to be (relatively) efficient in terms of
finding the projection point and constructing the curve.
It can be seen that in comparison with the two-dimensional case, the three-dimensional
smoothing procedure is costly. This is so, despite the fact that the procedure is completely
local. The reason is that a quite high degree polynomial is required to be able to keep the
construction local. If a lower degree polynomial is used, a global system of equations needs
to be solved, which makes things even more complex. This is not a problem when the body
being smoothed is rigid. However, if the body is deformable, then smoothing needs to be
performed at each iteration of the simulation, which is then prohibitively costly.
Apart from the construction cost, the localisation of the projection point is also signi-
ficantly more expensive for the three-dimensional case than it is for the two-dimensional
one. The reason for this is that a full fledged constrained optimisation problem has to be
solved for each integration point. This makes things even more costly. After some simple
tests, it turns out that the additional stability of the method does not outweigh the addi-
tional iterations required for the non-smoothed case. What was observed, however, is that
the time required in projecting the integration points is a significant with respect to the total
simulation time.
X
absolute
projected
n
111111111111111111111111111111
000000000000000000000000000000
000000000000000000000000000000
111111111111111111111111111111
000000000000000000000000000000
111111111111111111111111111111
000000000000000000000000000000
111111111111111111111111111111
000000000000000000000000000000
111111111111111111111111111111
000000000000000000000000000000
111111111111111111111111111111
000000000000000000000000000000
111111111111111111111111111111
000000000000000000000000000000
111111111111111111111111111111
000000000000000000000000000000
111111111111111111111111111111
000000000000000000000000000000
111111111111111111111111111111
Returning our attention to the single point projection problem: Assume that there are
N master faces. In that case we require an amount of operations to find the projection
face which lies in the order of magnitude of N. Increasing N by a factor of two, would
approximately increase the amount of computation time by a factor of 2. For this we denote
that the single point projection algorithm costs order N operations, or more concisely O(N)
operations.
Applying the one point projection algorithm to a set of M points leads to an algorithm
that requires O(M N) operations. If M lies in the order of magnitude of N, then we could
also write O(N 2 ). For typical large scale problems, the matrix solution time for a direct
solver lies in the order of O(Nk 2 ) where k is the bandwidth of the problem. Since, N
clearly dominates k and even k 2 , this would mean that the total simulation time is domin-
ated by the time which is required to find the projections. This is unacceptable. To improve
things, we can set up a search data structure on the faces. If the cost of creating this data
structure is less than O(N 2 ), and if it would bring down the search time for a single query to
less than O(N), then the time necessary to find all the projections using this data structure
is less than O(N 2 ).
Due to the possible time requirements of the projection problem, a small body of liter-
ature for solving it has been established. Each article arrives at a solution procedure from
first principles. Instead of attempting to solve the problem directly, one could also look for
problems that are similar in structure and have an efficient algorithm to solve them. Modi-
fying the projection problem slightly so that it resembles an already solved problem then
becomes our goal. Similar looking problems can be found in the field of computational
geometry.
To apply the available geometric algorithms requires the split of the problem into two
stages: first a global search is made using modified computational geometrical algorithms.
The result of this global search is a set of faces which lie in the vicinity of the point to
project. Subsequently a local search is made within this set of candidate faces to determine
which is the actual projection face.
Depending on the structure of the problem, the result of the global search may in fact
contain (a significant portion of) all the segments. This can occur in problems in which
there is a large difference in magnitude between the smallest and largest element, or when
the step-size is very large with respect to the average element size. But usually, the method
is very efficient.
4.5 Efficient projection finding 73
The rationale behind the choice of this size of box is as follows. If the point is inside
the capture box, it may come into contact with the face that corresponds to the capture box.
If the point is outside the capture box, it can never enter the bounding box during the step.
Thus, it can never come into contact with the face that is contained within the bounding
box.
When given a point, we wish to select those capture boxes that contain it. These capture
boxes contain the candidate projection faces. Which one of these candidate faces is the
actual projection face is to be determined by a local search strategy. A good candidate for
the local search is the brute force approach, since the total number of candidates is small.
74 Distance Functions
In case the master surface is non-deformable, but is free to move, then this movement
can be incorporated into the size of the capture box. Additionally, if the master surface
is even free to deform, then we add twice the maximum displacement to the capture box.
For accurate forming simulations, the maximum displacements do not usually exceed the
size of the elements. Thus, the boxes are typically no more than three times the size of the
original elements.
• There are boxes which lie fully to the left of the cutting plane specified through the
median. These form the set Ileft .
• There are boxes which lie fully to the right of the cutting plane, these fall into the set
Iright.
• There are some boxes which intersect the cutting plane, these are placed in Imid .
The three sets are obviously disjoint, and together they again form the set I.
The selection situation is shown in Figure 4.15. The boxes in bold belong to the set
Imid .
We are building a tree on the complete set of boxes I. The root partitions the set into
the three mentioned sets. The left child is to further contain a data structure of the boxes in
Ileft . Equivalently the right child is to contain a further data structure on the boxes in Iright.
At the node itself, we store the set Imid in a so-called associated data structure.
The left set and right set of boxes do not overlap each other in space, and are unstruc-
tured sets, just like the set I was originally. Hence, we can again repeat the process on the
left and right child to obtain sets Ileftleft , Ileftmid and Ileftright . In an identical manner the set
for the right child is subdivided. As an illustration consider Figures 4.16 and 4.17 the tree
is shown after the first and second subdivision step on the first coordinate direction.
As can be seen the data structure is a recursive one. Each node contains a set of boxes
that are split on the first dimension. What remains is to structure all the middle sets that
were created. The boxes in the middle set are already partitioned on the first coordinate
dimension, and no additional separation is to be expected on the first coordinate direction.
Instead, we have a second coordinate direction which is not yet taken into account.
As boxes in the original set I are ordered on the first coordinate direction, all the boxes
in Imid are ordered on the second coordinate direction. The approach for splitting I in the
first coordinate direction, can be repeated for Imid for the second coordinate direction. In
the case of three-dimensional problems, such as sheet-forming, a further subdivision can
be made on the third coordinate direction.
After we have changed dimension 3 times, each middle-set can no longer be subdivided
in the proposed manner with any further separation of boxes to be expected. The maximum
number of boxes left in any middle set after the third subdivision is called K . This number
is the maximum number of boxes overlapping a certain point in space. For a typical sim-
ulation in two dimensions, K is 2, since the capture boxes spanned by the faces attached
76 Distance Functions
to a single node will overlap. For a simulation in three dimensions using triangles the av-
erage number of elements attached to a node is 6, and thus K is in a regular simulation
approximately 6.
We finish this section by making some remarks: The worst case search time for the
algorithm presented is widely overestimated. In fact the pre-factor starts to be very import-
ant if log N is not too large. After a longer analysis, it turns out that the pre-factor can be
decreased to at least K /4.
The worst case bound is quite strict, because it assumes a certain substructuring in
which half of the boxes fall into the left set, half of the boxes fall into the middle set and
the right set is empty.
4.6 Conclusions
In this chapter, the integration of the contact integral was discussed. We saw that quite a
few problems remain, even when a good regularisation method was already selected. One
of the key factors that arises is that of geometric incompatibility when the problem under
consideration is discretised. The geometric incompatibility between the slave and master
surface can cause the traction approximations to be inaccurate, and when the master surface
is non-smooth can deteriorate convergence.
The geometric incompatibility also limits the choice of integration algorithms. By se-
lecting a nodal integration algorithm, the accuracy problems can be alleviated. The in-
fluence of the arising stability problems can be diminished by employing a line search
procedure. Alternatively, we can smooth the master boundary. The procedure for smooth-
ing general 2D and 3D boundaries was shown to be effective for the 2D problem, but rather
cumbersome for the 3D problem.
A third factor apart from accuracy and stability is efficiency. In the last part of this
chapter, a method was presented to improve the efficiency of finding the projection loca-
tions for the integration points.
With this, the theoretical description of how to deal with contact in simulations is com-
plete, and in the next chapter we take a look at several numerical experiments and applica-
tions to compare the different methodologies.
Chapter 5
5.1 Introduction
In this chapter the theory that was developed in this thesis is illustrated by several examples.
In Section 5.2 we compare the different traction regularisations. In Section 5.3 the accuracy
of the contact normal tractions are tested. In Section 5.4, the accuracy of the implementa-
tion of the friction algorithm is tested. In Section 5.5, a non-unique projection problem is
illustrated. In Section 5.6 an example is given of a large sliding problem, with deformable
tools. In Section 5.7 an industrial application is presented. We end the chapter in Section
5.8 with the conclusions.
The top slab is the punch for which the displacements are prescribed. Again a contact
constraint is set for the interface for the punch with the block. The problem is simulated
using the penalty method, the method of augmented Lagrangians, and the barrier method.
For the penalty parameter for normal contact in all the penalty and augmented Lag-
rangian scheme we use 10000. The penalty for the modified barrier method was set to
100. A distinction between the two penalties is made, since the contribution to the stiffness
matrix for the penalty based methods and the barrier methods are different. For the penalty
method the penalty could be interpreted as force per unit area per unit penetration depth
[N · m−3 ], whereas in the barrier methods, the interpretation would be: per unit penetration
depth [m−1 ]. Choosing them to be the same would cause completely different convergence
characteristics. The penalty parameter for the frictional part is set to 10000, and the friction
coefficient is chosen as 0.1. For the barrier method and method of augmented Lagrangi-
ans, we set a maximum allowable penetration of 0.01. The default projection method as
presented in Chapter 4 is chosen.
The results are illustrated in Figure 5.2. As can be seen the penalty method causes large
penetrations for this example. Increasing the penalty causes the method to have slightly
poorer convergence, but solves the problem of the large penetrations. This result illustrates
the problems with convergence of the penalty method that were discussed in Chapter 3.4.1:
If the penalty is chosen too small there is unacceptable overlap, if it is chosen too large, one
gets convergence problems. From the solutions of the augmented Lagrangian and modified
barrier method it can be seen that these methods deliver a geometrically good solution with
the specified penalties.
There is furthermore a large geometric overlap for each of the methods for the last
element sliding away under the punch. This overlap arises from the application of the same
basic projection scheme for each of the methods. What occurs is the same situation as
the one that was depicted in Figure 4.4. In this figure an element is shown to slide of a
corner. At this point a choice is to be made: Either the complete element is in contact,
or the complete element is not in contact. Either selection is incorrect, since the actual
situation is that the element is only partly in contact. One might be tempted by preferring
the non-overlapping solution, since it is geometrically more pleasing.
A non-overlapping solution for the upsetting problem can be obtained by applying a
two-pass procedure as was proposed in Chapter 4.3.2. Upon evaluation of the results in
Figures 5.2 we could decide to use either the modified barrier method or the method of
5.2 Efficiency of traction regularisations 81
augmented Lagrangians as the method to solve the two pass problem. Interestingly, the
method of augmented Lagrangians does not converge within a reasonable number of itera-
tions anymore. The modified barrier method does converge. An explanation could be that
the modified barrier method is more non-linear and can predict larger changes in contact
tractions with smaller changes in displacements. This additional non-linearity, however,
does slow down the rate of convergence.
The final result is illustrated in Figure 5.2. The solution now seems to be geometrically
more correct. The actual solution, however, lies somewhere between the results of the one-
pass and two-pass solution. This can easily be understood by realising that the line segment
which is lying partly under the block should be split into two parts: One part which lies
under the punch, and a part that does not. In the case of the one-pass procedure the line-
segment loses contact early. In the case of the two-pass procedure, the line-segment loses
contact too late. Thus the one and two-pass procedures form a lower and upper bound on
the actual solution. These bounds can be improved by mesh-refinement, which brings us
back to our initial observation that the large overlap is really due to a discretisation error,
and is not an artifact of the application of the contact method.
As a conclusion of the above simulation we can state that the accuracy of the penalty
method is strongly dependent on the magnitude of the penalty, and thus after a simulation a
check should be run to see if there are no unacceptable violations. The modified barrier and
augmented Lagrangian method do not have this problem. Moreover, the modified barrier
method is somewhat more non-linear than the method of augmented Lagrangians, which
can help in some situations as with the solution of the two-pass procedure in which the
local material non-linearities are also stronger. If this is not the case, then probably the
method of augmented Lagrangians is the best choice: it gives the highest accuracy with the
least local increase of non-linearities.
The combined radius is computed by using the radii of the two cylinders R1 and R2 as
follows:
R1 R2
R∗ = . (5.5)
R1 + R2
To validate the implementation using these formulas we set up a contacting interface
between a cylinder and a plate. We assume the cylinder and plate to be infinitely high, so
we can model the problem in a two-dimensional setting. The plate can be considered as a
cylinder with an infinite radius. For the combined radius we then find:
R1 R2
R∗ = lim = R1 . (5.6)
R2 −→∞ R1 + R2
Furthermore we assume that the plate is undeformable. To achieve this, we make the plate
much stiffer than the cylinder. For the combined elasticity radius this means:
2E 1 E 2 2E 1
E∗ = lim = . (5.7)
E 2 −→∞ E 2 (1 − ν ) + E 1 (1 − ν 2 )
2 1 − ν12
1 2
In the simulation the radius of the cylinder was set to 50 mm. The elasticity modulus
for the cylinder was set to 210000 MPa and the Poisson ratio was selected as 0.31. Two dif-
ferent meshes composed of triangular elements were used to compute the resulting normal
pressure. They are given in Figure 5.4. The second mesh contains twice as many elements
along the circular part as the first. The top of the cylinder is suppressed in the y-direction
and the sides of the plate are suppressed in the x-direction. A total force of 35000 N is ap-
plied to the plate in the y direction. To achieve an accurate result, the method of augmented
Lagrangians was employed to obtain a maximum penetration of 1.0 · 10−5 .
The results for the normal tractions are presented in Figure 5.5. The value for the half
contact width is from the Hertzian formulas: a = 3.10 mm. The value for the maximum
pressure is: pmax = 7195 N. The results are presented in Table 5.1. As can be seen from
the table, the correct half contact width lies nicely within the upper and lower bounds of
the left and right contact widths. And the pressure is approximated more accurately if the
84 Numerical Experiments and Applications
Figure 5.4: The two meshes employed in the simulation of Hertzian contact.
(a) Contact pressure for mesh 1 (b) Contact pressure for mesh2
Figure 5.5: The two meshes employed in the simulation of Hertzian contact.
number of elements is increased. What is interesting, is the fact that the maximum contact
normal pressure is overestimated by the algorithm, whereas the original results by ter Haar
(1996) showed an underestimation, and a wider contact area. The latter is probably due
to using a penalty method with a too low penalty value. The result of that simulation will
show more overlap, consequently it has a larger contact with and hence on average lower
contact normal tractions.
Another thing that is observed from the simulations is that the maximum pressure is
overestimated. For the first mesh, the overestimation is approximately 6%. The second
mesh that has twice as fine a boundary discretisation has an overestimation of the maximum
pressure of approximately 3.5%. Since local inaccuracies can play an important role in
global instabilities, we can conclude that a high number of elements is required to be able
to accurately simulate the contact behaviour.
L hl The L-parameter for the transition of mixed lubrication to hydro dynamic lubrication.
0.13 0.13
0.12
0.125
0.11
0.12
0.1
0.09 0.115
µ
µ
0.08
0.11
0.07
0.105
0.06
0.05 0.1
1e-08 1e-06 0.0001 0.01 1 0.0001 0.001
log10 L log10 L
2
c = − L bl
, (5.10)
log L hl
'
L0 = L bl · L hl . (5.11)
We have performed a sliding simulation using a Stribeck friction algorithm. The lub-
ricant viscosity was set to 1 Ns/m2 , the surface roughness was assumed to be 1 µm. The
sum velocity and pressure are obtained from the simulation. The friction parameters that
were used are as follows: L hl = 4.3 · 10−3 , L bl = 2.7 · 10−4 , µhl = 0.05 and µbl = 0.13.
A graph of the corresponding Stribeck curve is given in Figure 5.6(a) The simulation per-
forms 5 steps with a displacement of 0.1 mm in the positive x-direction each 0.1 s. The
other parameters required are identical to that in Section 5.3.
At the fifth step, all the nodes are sliding, and the L-parameters and coefficients of
friction can be extracted from the simulation. The result is given in Figure 5.6(b). From
this figure we can see that most of the nodes are in the mixed lubrication regime.
The distribution of the coefficients of friction over the contact area are given in Figure
5.7(a). From it can be seen that the higher pressure on the nodes in the centre of the
contacting area is clearly causing higher coefficients of friction. In Figure 5.7(b) a graph is
presented of the total amount of sliding at the end of the fifth step. From this it can be seen
that the largest amount of sliding is done by the trailing nodes.
5.4 Accuracy of friction 87
slided distance
0.13 0.47
0.125 0.46
0.12 0.45
0.115 0.44
µ
0.11 0.43
0.105 0.42
0.1 0.41
47 48 49 50 51 52 53 54 47 48 49 50 51 52 53 54
x-location x-location
(a) Distribution of friction parameters over (b) Slided distance of the contact area.
the contact area.
Figure 5.7: Graphs of the distribution of sliding and friction coefficient over the contact area.
Figure 5.8: The initial and final configuration of the pulling problem.
200
Required Force (N)
150
100
50
0
0 5 10 15 20 25 30
Leftward motion of strip (mm)
Figure 5.9: The computed pull-out force on the leftmost edge of the strip.
5.4 Accuracy of friction 89
Figure 5.10: The initial and final configuration of the pulling problem.
250 250
200 200
Required Force (N)
100 100
50 50
0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Downward motion of strip (mm) Downward motion of strip (mm)
(a) Rough die mesh results (b) Finer die mesh results
Figure 5.11: The computed pull-out force on the leftmost edge of the strip.
FN
T1 T0
θ
not accelerated, thus there is equilibrium of forces. The force normal to the line segment
exerted by T0 is T0 21 θ . Equivalently so, the force exerted by T1 exerted normal to the line
segment is T1 21 θ . The total force exerted by the plate on the cylinder is thus:
1
θ (T0 + T1 ) . (5.12)
2
This force is in equilibrium with the force exerted by the cylinder on the plate, which we
name F N . Taking into account Coulomb’s law, to let the plate slip to the left, a force has
to be exerted in the left direction, which is µF N larger than the force exerted to the right.
Putting this all in formula form leads to:
T1 − T0 = µF N ⇐⇒
1
T = µ θ (T0 + T1 ) . (5.13)
2
Taking the limit θ → 0, leads to the differential equation:
dT
= µT. (5.14)
dθ
The solution of this differential equation is:
In our case T (0) = 100 N according to the results in Section 5.4.1. The total angle
traversed is a quarter circle, which means that the resulting force should be: T ( π2 ) =
100 exp 0.1 π2 = 117 N.
Comparing this value with the results that were obtained by the simulations we see that
both the rough and the fine approximation of the die agree well with the analytic value of
117 N. We do not encounter any parasitic stiffnesses discussed by Vreede (1992). From
the flat pull, and the pull along a cylinder, we can conclude that the friction algorithm is
implemented correctly.
2
10
outer convergence
inner convergence
1
10
18
16
0
10
14
Error norm
12
−1
10
y−location
10
8 −2
10
−3
4 10
−4
0 10
0 5 10 15 20 0 1 2 3 4 5 6 7 8 9
x−location Newton iteration
A distinction can be made between inner iterations, in which for a fixed penalty parameter
and fixed estimations to the normal tractions an equilibrium solution is computed, and outer
iterations in which the normal contact tractions are updated, and the penalty parameter can
be adjusted.
The convergence results are plotted in Figure 5.13 on the right side, where the inner
convergence shows the progress within each subproblem loop, i.e. for set values of the
penalty parameter p and the estimations of the contact normal traction λ. The outer con-
vergence shows the progress of the total convergence. Just as with augmented Lagrangian
schemes, the total scheme converges linearly. Each inner problem converges quadratically,
when sufficiently close to the optimum for that subproblem. The zigzag pattern is due to
fact that when we update the Lagrange multipliers and the value of the penalty parameter
p, the error in our initial estimate increases.
From the simulation we can see that the projection algorithm as discussed in Chapter
4.3 handles the non-unique projection problem well. It furthermore shows that the modified
barrier method converges linearly to the solution in one iteration per augmentation, which
makes it a competitor for the method of choice in contact simulations.
0.7 -1 0.7
-1
y-location
0.4 -1.6 0.4
-1.6
0.2 -2 0.2
-2
0 0
-0.4 -0.2 0 0.2 0.4 0.6 0.8 1 1.2 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 1.2
x-location x-location
(a) Equivalent plastic strain in the initial (b) Equivalent plastic strain in the final state.
state.
Figure 5.14: The initial and final state of the sliding problem.
used in conjunction with a Ludwik–Nadai hardening law is used, this takes the form σY =
C(ε0 + ε p )n . In this formula σY is the yield stress and ε p is the equivalent plastic strain.
The parameter setting that was used is given in Table 5.2. The indenter remains elastic
throughout the simulation and thus no plastic property values for the indenter are listed.
The results of the computation are shown in Figure 5.14. The equivalent plastic strain
is shown in the deformed mesh. The top row of nodes on the indenter is prescribed. The
bottom nodes on the block are all suppressed in both directions. The indenter has pre-
scribed movement on the top nodes. First a 0.1 downward movement is prescribed on
these nodes. The x-displacement on the indenter top nodes are then suppressed. After that
a 1.0 displacement to the right is prescribed on the top nodes, as well as suppressing the
y-displacement. The downward part of the motion is performed in 8 steps, whereas the
right displacement is performed in 80 steps. This corresponds to steps of approximately
25% of the element size used. The time discretisation used is an Euler-backward scheme.
In the computation of both the modified barrier scheme and augmented Lagrangians
algorithm, we can distinguish between inner and outer iterations. The inner iterations are
the normal Newton steps, where we keep the estimations for the tractions fixed. Whereas
94 Numerical Experiments and Applications
in the outer iterations we augment the estimations for the tractions. The complexity of the
problem can be expressed in the cumulative amount of inner iterations required to achieve
convergence of the contact traction estimations. The measure of convergence we use is the
following:
|λi ||di |
e= i . (5.16)
1 + λ
For the convergence of the inner iterations we find that it does not have to exceed the
convergence of the outer iteration, since that will be wasted effort. This does not hold for
the step before the first augmentation, since we do not know the accuracy of the traction
estimates then.
In this e represents the error. The vector λ represents the estimation for the contact
tractions in the boundary nodes, and di is the distance at node i . The division is done by
the given term to have no conflicts when λ is equal to 0. Hence, the error approximates
the average penetration. We have set our convergence goal to e = 1.0 · 10−5 . A plot of
this parameter for the Augmented Lagrangian algorithm and the modified barrier method
is shown in Figure 5.15.
From this picture it can be seen that the modified barrier method requires less augment-
ations to converge to the required accuracy. The question is, however, how many more
inner iterations will this take? Since, each inner iteration requires the assembly of a tangent
stiffness matrix, and the solution of a linear system of equations.
The convergence results can be seen in Tables 5.3 and 5.4. Here we see that the mod-
ified barrier method requires more inner iteration per outer iteration than the method of
augmented Lagrangians. However, the modified barrier method requires two less augment-
ations to converge. This results in an equivalent number of inner iterations required to
converge.
σ y (ε p ) = C(ε0 + ε p )n . (5.17)
5.7 An industrial application 95
10−3
mod. barrier
aug. Lagrangians
10−4
e
10−5
10−6
1 2 3 4 5 6
augmentation number
In this equation σ y is the yield stress for a certain equivalent plastic strain ε p . The point
ε0 is the initial yield strain, and this is set at 1.2 · 10−3 . The hardening coefficient C is set
to 500 MPa, and the parameter n is chosen as 0.2. Combining this all leads to an initial
yield stress of 150 MPa. The last parameter important in the simulation is the coefficient of
friction, which is chosen successively as 0.1 and 0.2.
The sheet is modelled with discrete Kirchhoff triangular elements. For the clamping
mechanism ideal clamping is assumed, so that a displacement on the boundary nodes can
be prescribed. The sheet is meshed with 1200 elements and has a refinement along the top.
Since the simulation is fully symmetric, we only model half of the plate. The contact model
that was employed to obtain the solution is the method of augmented Lagrangians.
The results are shown in Figures 5.17 and 5.18. In Figure 5.17 the thickness of the final
plate is shown after using a coefficient of friction of 0.1. The result in Figure 5.18 shows
the thickness after using a coefficient of friction of 0.2.
From the pictures a clear dependence on the friction coefficient can be seen. The value
of the frictional coefficient can be influenced by the roughness of the die and the type of
lubricant that is employed. It is this type of analysis which provides insight into the process
of stretch forming.
5.8 Conclusions
From the examples we can see that the way contact is modelled has a large influence on the
rate of convergence. The results from the upsetting problem in Section 5.2 show us that the
augmented Lagrangian method and the modified barrier method give good, geometrically
acceptable solutions. In the same section it was shown that having a fully conforming
geometrical solution does not necessarily mean that the simulation was performed very
accurately. This in fact depends strongly on the discretisation.
The result of the Hertzian contact problem in Section 5.3 shows us that the contact
normal tractions can be computed accurately within the error of the discretisation.
The result of the pull-out problem in Section 5.4 shows us that the friction algorithm
was implemented correctly and can be simulated accurately using the method of augmented
Lagrangians for both the contact normal tractions as well as the tangential tractions. It
furthermore shows that there are no parasitic tangential stiffnesses.
The corner contact problem in Section 5.5 illustrates that the correct projection seg-
ments are obtained by the projection algorithm.
The large sliding problem in Section 5.6 further illustrates the capabilities of both the
augmented Lagrangian as well as the modified barrier method as traction estimation meth-
ods.
Finally from the stretch forming problem discussed in Section 5.7 it follows that the
contact model is an essential component in simulating a forming process. In this section
one can see the influence of the coefficient of friction has on the final thicknesses on the
plate. This type of insight can then be used to adjust the forming process itself to obtain a
better product.
98 Numerical Experiments and Applications
C ONCLUSIONS AND
R ECOMMENDATIONS
As it was stated in the the introduction of this thesis, our aim is to study the contact problem
applied to metal forming simulations. To see where it originates, formulate the equations
and constraints, discuss the various ways in which to enforce those constraints, and finally
how to incorporate it all into a finite element framework. By scrutinising the complete
path of development of the equations to the implementation, we intend to identify all the
possible pitfalls occurring when contact methods are applied.
The first part of our aim was to formulate the equations and constraints. This part of
the work is demonstrated in Chapter 2. The derivation is presented for the weak form of
equilibrium including contact. The equivalence of the necessary conditions of the contact
problem in elasticity and the final one for the virtual displacement case is shown. From
this it can be seen that there is a strong coupling between variational problems and the
minimisation of functionals.
An advantage of the insight gained by the fact that there exists an association between
the variation and minimisation problems, is that we can use methods that solve the contact
problem for the minimisation problem and use it for the variation problems.
The following equation in general form represents the weak form with contact:
G(ϕ, w) + G c (ϕ, w) = 0 (6.1)
The first functional in this equation can be computed in the usual sense by a finite element
method. The second functional contains the contribution of contact. This latter functional
can be expanded by interpreting w either as virtual displacements or as virtual velocities.
For virtual displacements, the result is:
(2)α
G c (ϕ, δϕ) = t N δd N dγ + tTα δξ dγ (6.2)
γ (1) γ (1)
100 Conclusions and Recommendations
We can conclude that what needs to be done is to compute the contact tractions, which en-
force the impenetrability constraints. Several ways in which this can be done are discussed
in Chapter 3.
First we discussed the mixed method, which is computationally expensive, and the
constraint method, which is very efficient but can present problems for example in plate
forming simulations where the detection of contact is very important. Moreover, there
may be too many constraints in the plate forming simulation to use an effective active set
method.
The methods that were dealt with more in depthly are methods that regularise the con-
tact traction by estimating them through allowing small violations of the impenetrability
condition. By taking this viewpoint, all the different methods can be seen to stem from a
basic mold. The method is specified by assuming the form of a specific penalisation func-
tion . This penalisation function is introduced in the optimisation setting For the more
general variational setting, one just fills in the results for the first order conditions of the
optimisation problem. The different methods discussed are:
• The penalty method, for which:
p
(d N , { p}) = −d N 2 , (6.3a)
2
t N (d N , { p}) = − p−d N. (6.3b)
Using the same reasoning but immediately in the variational setting, the contact frictional
tractions can be regularised. These methods rely on a predictor-corrector scheme. Only the
predictor is determined differently for the two schemes that were considered. These two
schemes are:
• The penalty method for friction, in which case the predictor is set to:
p
t T = pd T . (6.6)
• The method of augmented Lagrangians, in which case the predictor is set to:
p
t T = λ T + pd T . (6.7)
101
From the predictor, the correct frictional traction is computed, by using either the fixed
estimate to the normal contact traction λ N or the current estimate t N . The frictional traction
now follows from: p p
tT if t T ≤ |t N |,
tT = tT
p
(6.8)
µ|t N | t p otherwise.
T
In this work we only considered regularisation methods and shortly discussed the con-
straint method. It might be interesting to see these two methods compared more in depth.
Additionally one could solve the intermediate problems that arise during an SQP-step (see
Section 3.3.2) with either the augmented Lagrangian or modified barrier method, instead
of performing direct eliminations.
Next in Chapter 4, the integration of the contact integral is discussed. As it turns out,
a lot of the convergence problems observed with contact arise from geometric incompatib-
ilities. The geometric incompatibility between the slave and master surface can cause the
traction approximations to be inaccurate, and when the master surface is non-smooth it can
deteriorate convergence.
The geometric incompatibility also limits the choice of integration algorithms. By se-
lecting a nodal integration algorithm, the accuracy problems can be alleviated. The in-
fluence of the arising stability problems can be diminished by employing a line search
procedure. Alternatively, we can smooth the master boundary. The procedure for smooth-
ing general 2D and 3D boundaries is shown to be effective for the 2D problem, but rather
cumbersome for the 3D problem. It is therefore recommended that additional research is
performed on finding more efficient local smoothing methods for 3D problems.
A third factor apart from accuracy and stability is efficiency. In the last part of Chapter
4, a method is presented to improve the efficiency of finding the projection locations for
the integration points. The total time that is required for the simulation can increase signi-
ficantly if the projections are performed in a straightforward manner, it is therefore recom-
mended to always use a more sophisticated search algorithm.
Finally in Chapter 5, several numerical problems are studied to illustrate the operation
of the different parts of the contact algorithm. These are an upsetting problem to illustrate
the difference between the several regularisation methods. A pull-out problem shows that
the methods implemented do not contain parasitic stiffnesses that may be encountered when
using contact elements, and it furthermore shows that the friction algorithm is implemen-
ted correctly. The corner contact problem illustrates that the correct projection segments
are obtained by the projection algorithm. The large sliding problem further illustrates the
capabilities of both the augmented Lagrangian as well as the modified barrier method as
traction estimation methods. Finally a stretch forming problem is analysed, from which we
can conclude that the contact model has an influence on the final product geometry.
A chain is only as strong as its weakest link, so far the contact problem has been held
responsible for the failure of a lot of simulations. It is with this work that many of the
stability, accuracy and efficiency issues that arise due to contact are addressed, so that
research may again focus on other areas which require attention for the correct simulation
of forming processes.
Appendix A
In this appendix, we discuss the small mistake that was made in the much referred paper
Laursen and Simo (1993). The paper goes into great depth in describing the continuum
mechanics framework for contact. One of the claims made, is that the contact slip velocity
can be made frame-indifferent when one would just use a convected description expressed
in the dual basis. In fact, it is argued in this article, that the description of the velocity in
this basis is a constitutive decision. The first author even goes into greater detail in Laursen
(1994) to clarify the choices made. The author is hoping to provide a general and simple
theory to solve a very complex class of problems. Unfortunately, this is not so.
To see this, we need to introduce some theory concerning covariant and contravariant
vectors. Using this theory, we can quickly locate the error in both articles.
A.1 Vectors
By the term vector, two things can be meant. In linear algebra, a vector is usually con-
sidered to be just a bunch of numbers. These numbers can be stacked in a column vector,
or put sequentially in a row vector. In the physical world, a vector is a quantity which is
defined to have a direction and a length. The number of space dimensions that the vector
“lives in”, is usually 2 or 3. We only concern ourselves with 3 dimensional space. A vector,
being a physical quantity, is independent of the representation which is used, it just is.
To work with arbitrary vectors, we usually represent them with respect to a basis. The
most preferred basis is the orthonormal Cartesian basis {e1 , e2 , e3 }. A vector q is represen-
ted with respect to this Cartesian basis as:
q = q i ei , (A.1)
104 Mistakes in the Convected Description
where we used the summation convention. If we are well aware of what basis is being used,
the basis vectors are often omitted:
q = qi (A.2)
An interesting thing to note is that using a Cartesian basis still depends on the location
of an observer.
Now, we consider a set of three independent vectors (v1 , v2 , v3 ), which are not neces-
sarily orthogonal, nor of unit length. In order to represent an arbitrary vector a in terms of
these vectors, we need to find components (a 1 , a 2 , a 3 ) such that
a = a 1 v1 + a 2 v2 + a 3 v3 . (A.3)
One way of doing this is by forming a reciprocal basis (v1 , v2 , v3 ). The reciprocal basis
vectors satisfy the following relationship:
j 1 if i = j
v i · v j = δi = . (A.4)
0 otherwise
Using the reciprocal basis, the components a i are easily found by:
j
a · v j = a i v i · v j = a i δi = a j . (A.5)
In an identical fashion, the components of a with respect to the reciprocal basis can be
found. In that case, a takes the form:
a = ajvj, (A.6)
and the components can be found by
ai = a · vi . (A.7)
The components with respect to the original basis vectors vi , are called the contravari-
ant components of a. The components with respect to the reciprocal basis vectors v j are
called the covariant components of a. Note that the numbers ai and a i may differ, but that
in combination with the basis vectors, they represent the exact same vector.
An additional simple relationship between the components exists, which is given now.
First introduce the numbers gi j and g i j as
gi j = vi · v j . (A.8)
In an identical fashion:
g i j = vi · v j . (A.9)
From the properties of the inner product, we can immediately conclude that gi j = g j i and
g i j = g j i . The numbers gi j are called the metric components of the space and the numbers
g i j are called the conjugate metric components of the space. Using this notation, we can
quickly find that
ai = gi j a j , (A.10)
and
ai = gi j a j . (A.11)
A.2 The mistake 105
τ α = F · Tα , (A.14)
where
∂ϕ
F = (A.15)
∂X
Tα = 0,α (A.16)
vT = vαT τ α . (A.17)
VT = F −1 · vT (A.18)
−1
= F vαT τ α (A.19)
= F −1 vαT F · Tα (A.20)
= vαT Tα (A.21)
but this is inconsequential for the discussion, and we stick to the usage of vαT .
The metric components (as discussed in the previous section) for the space in the ori-
ginal configuration on the contact surface can be expressed as
Mαβ = Tα · Tβ . (A.23)
Using the metric, the currently contravariantly represented vector VT can also be written
using the covariant basis:
VT = vαT Tα = Mαβ vαT Tβ . (A.24)
106 Mistakes in the Convected Description
The distinction in representation in the article is made by attaching a “flat” sign to the
symbol, thus a distinction is made between
VT = vαT Tα (A.25)
and
#
VT = Mαβ vαT Tβ . (A.26)
#
But we stress here that the distinction is only in representation! The vectors VT and VT are
in fact totally identical. It is only the selection of the basis, which changes the components.
In an identical manner as we constructed a reciprocal basis in the undeformed config-
uration, we can also construct one in the deformed configuration. This gives rise to a basis
τ β , with a metric
m αβ = τ α · τ β . (A.27)
Representation of the original vector vT in this basis would give rise to:
vT = m αβ vαT τ β . (A.28)
The “claim to fame” in the articles now arises from the statement that pushing forward
#
the vector VT gives rise to an invariant object to describe the friction with. The reasoning
#
goes as follows: The covariantly represented vector VT is given by:
#
VT = Mαβ vαT Tβ . (A.29)
Pushing the vector forward, would yield according to the article (but incorrectly so):
#
vT = Mαβ vαT τ β . (A.30)
Unfortunately, the push forward of Tβ is not τ β . The correct computation of the push-
forward, will require the reversal of the coordinate transformation to the reciprocal basis,
obtaining the contravariant tensor. The contravariant tensor can then be pushed-forward.
Transforming the push-forward again to covariant form yields (A.28).
The description (A.30) is no longer a description of the velocity. The mistake made lies
in the incorrect transformation of a covariant tensorial quantity, which needs to be pushed
forward in a different manner than contravariant tensorial quantities, which is attempted in
both articles. It seems the author, must have had some feeling there was something wrong,
by his elaborate derivation using contravariant tensorial quantities in Laursen (1994), upon
which follows the sentence: “One may readily conclude from these facts that the same
relationship holds for the dual basis,...”, which is obviously not so.
A.3 Conclusions
The conclusion we can draw here is that using the push-forward of the covariant description
in the undeformed configuration of the velocity, does not yield the result as proposed in
Laursen and Simo (1993) and Laursen (1994).
Bibliography
Atzema, E. (1994, February). Formability of Sheet Metal and Sandwhich Laminates. Ph.
D. thesis, Universiteit Twente, 7500 AE, Enschede.
Barlam, D. and E. Zahavi (1999). The reliability of solutions in contact problems. Com-
puters and Structures 70, 35–45.
Bathe, K. J. and P. A. Bouzinov (1997). On the constraint function method for contact
problems. Computers and Structures 64, 1069–1085.
Belgacem, F. B., P. Hild, and P. Laborde (1998). The mortar finite element method for
contact problems. Mathematical and Computer Modelling 28, 263–271.
Belytschko, T., W. K. Liu, and B. Moran (2000). Nonlinear Finite Elements for Continua
and Structures. Wiley.
Belytschko, T. and M. O. Neal (1991). Contact-impact by the pinball algorithm with pen-
alty and lagrangian methods. International Journal for Numerical Methods in Engineer-
ing 31, 547–572.
Bonet, J. and R. D. Wood (1997). Nonlinear Continuum Mechanics for Finite Element
Analysis. Cambridge University Press.
Hild, P., F. B. Belgacem, and P. Laborde (1998). Unilateral contact for non-matching finite
element meshes: the global contact condition. In S. Idelsohn, E. O. nate, and E. Dvorkin
(Eds.), Computation Mechanics, New Trends and Applications.
Hughes, T. J. R. (1987). The Finite Element Method: Linear Static and Dynamic Analysis.
Prentice-Hall.
Jean, M. (1999). The non-smooth contact dynamics method. Computer Methods in Applied
Mechanics and Engineering 177, 235–257.
Kane, C., E. A. Repetto, M. Ortiz, and J. E. Marsden (1999). Finite element analysis of
nonsmooth contact. Computer Methods in Applied Mechanics and Engineering 180,
1–26.
Kloosterman, G., R. M. J. van Damme, A. H. van den Boogaard, and J. Huétink (2001). A
geometrical based contact algorithm using a barrier method. International Journal for
Numerical methods in Engineering 51, 865–882.
Lei, X. (2001). Contact friction analysis with a simple interface element. Computer Meth-
ods in Applied Mechanics and Engineering 190, 1955–1965.
Leung, A. Y. T., C. Guoqing, and C. Wanji (1998). Smoothing newton method for solv-
ing two- and three-dimensional frictional contact problems. International Journal for
Numerical Methods in Engineering 41, 1001–1027.
Lie-heng, W. (1998). On the duality methods for the contact problem in elasticity. Com-
puter Methods in Applied Mechanics and Engineering 167, 275–282.
Luenberger, D. G. (1973). Introduction to Linear and Nonlinear Programming. Addison-
Wesley.
Malvern, L. E. (1969). Introduction to the Mechanics of a Continuous Medium. Prentice-
Hall.
McCormick, J. P. (1983). Nonlinear Programming. John Wiley & Sons.
McDevitt, T. W. and T. A. Laursen (2000). A mortar-finite element formulation for fric-
tional contact problems. International Journal for Numerical Methods in Engineer-
ing 48, 1525–1547.
Munjiza, A. and K. R. F. Andrews (1998). Nbs contact detection algorithms for bodies of
similar size. International Journal for Numerical Methods in Engineering 43, 131–149.
Oldenburg, M. and L. Nilsson (1994). The position code algorithm for contact searching.
International Journal for Numerical Methods in Engineering 37, 359–386.
Pang, J. S. (1990). Newton’s method for b–differentiable equations. Mathematics of Oper-
ations Research 15, 311–341.
Papadopoulos, P. and J. M. Solberg (1998). A lagrange multiplier method for the finite
element solution of frictionless contact problems. Mathematical and Computer Model-
ling 28, 373–384.
Parisch, H. and C. Lübbing (1997). A formulation of arbitrarily shaped surface elements
for three-dimensional large deformation contact with friction. International Journal for
Numerical Methods in Engineering 40, 3359–3383.
Piegl, L. and W. Tiller (1997). The NURBS book. Springer.
Pietrzak, G. and A. Curnier (1999). Large deformation frictional contact mechanics: Con-
tinuum formulation and augmented lagrangian treatment. Computer Methods in Applied
Mechanics and Engineering 177, 351–381.
Polyak, R. (1992). Modified barrier functions (theory and methods). Mathematical Pro-
gramming 54, 177–222.
Refaat, M. H. and S. A. Meguid (1997). Updated lagrangian formulation of contact prob-
lems using variational inequalities. International Journal for Numerical Methods in
Engineering 40, 2975–2993.
112 Bibliography
Refaat, M. H. and S. A. Meguid (1998). A new strategy for the solution of frictional contact
problems. International Journal for Numerical Methods in Engineering 43, 1053–1068.
Rieger, A. and P. Wriggers (2001). Adaptive methods for frictionless contact problems.
Computers and Structures 79, 2197–2208.
Shimizu, T. and T. Sano (1995). An application of a penalty method contact and friction
algorithm to a 3-dimensional tool surface expressed by a b-spline patch. Journal of
Materials Processing Technology 28, 207–213.
ter Haar, R. (1996, May). Friction in Sheet Metal Forming. Ph. D. thesis, Universiteit
Twente, 7500 AE, Enschede.
van der Lugt, J. (1988, October). A Finite Element Method for the Simulation of Thermo-
Mechanical Contact Problems in Forming Processes. Ph. D. thesis, Universiteit Twente,
7500AE, Enschede.
Wang, F., J. Cheng, and Z. Yao (2001). Ffs contact searching algorithm for dynamic finite
element analysis. International Journal for Numerical Methods in Engineering 52, 655–
672.
Bibliography 113
Wang, S. and A. Makinouchi (2000). Contact search strategies for fem simulation of the
blow molding process. International Journal for Numerical Methods in Engineering 48,
501–521.
Wang, S. P. and E. Nakamachi (1997). The inside-outside contact search algorithm for
finite element analysis. International Journal for Numerical Methods in Engineering 40,
3665–3685.
Westeneng, A. (2001, Mar). Modelling of Contact and Friction in Deep Drawing Pro-
cesses. Ph. D. thesis, Universiteit Twente, 7500 AE, Enschede.
Wriggers, P., L. Krstulovic-Opara, and J. Korelc (2001). Smooth c1 -interpolations for two-
dimensional frictional contact problems. International Journal for Numerical Methods
in Engineering 51, 1469–1495.
Zavarise, G. and P. Wriggers (1998). A segment-to-segment contact strategy. Mathematical
and Computer Modelling 28, 497–515.
Zavarise, G. and P. Wriggers (1999). A superlinear convergent augmented lagrangian pro-
cedure for contact problems. Engineering Computations 16, 88–119.
Zavarise, G., P. Wriggers, and B. A. Schrefler (1995). On augmented lagrangian algorithms
for thermomechanical contact problems with friction. International Journal for Numer-
ical Methods in Engineering 38, 2929–2949.
Zavarise, G., P. Wriggers, and B. A. Schrefler (1998). A method for solving contact prob-
lems. International Journal for Numerical Methods in Engineering 42, 473–498.
Zhong, Z. H. and L. Nilsson (1988). A contact searching algorithm for general contact
problems. Computers & Structures 33, 197–209.
Zhong, Z. H. and L. Nilsson (1990). A contact searching algorithm for general 3d contact-
impact problems. Computers & Structures 34, 327–335.
List of Symbols
C A curve
g Gravitational acceleration
L A Lagrangian functional
λ A barycentric coordinate
116 Bibliography
w A weighing function
ϕ The mapping operator of points from the reference configuration to the current
configuration
µ Friction coefficient
ρ Density
A penalisation functional
118 Bibliography
This project has been made possible by the financial support of the Netherlands Institute
for Metals Research, and the scientific support of the group of Technical Mechanics of
the University of Twente, their essential input is gratefully acknowledged. An additional
acknowledgement goes to Fokker aerostructures for allowing me the use of a toolshape for
the stretch forming simulation in this thesis.
The first person I want to thank is Ton van den Boogaard, my supervisor during the
project. Our different backgrounds would cause quite a few discussions to result in agreeing
that we had disagreed over something we both agreed on. Off course also thanks to Han
Huétink for having me work in his group in the first place, and for being so enthusiastic
in discussing technical difficulties. Your “ooh, I have a thesis that contains something
somewhere in a chapter on this,” has certainly increased my stack of booklets. Yu Yuhong
is thanked for bearing with me, when my enthusiasm on understanding something would
mean listening to yet another story on contact. Also she has made sure I can intrigue
chinese restaurant attendants by ordering szeh chuan chicken in chinese.
A big thank you also goes to Ruud van Damme who has set me on my tracks to go
and do this work, and who has been a big help ever since he supervised me through my
Masters degree. How quickly you understand stuff is incredible, as well as your ability to
find corrections as if the page was initially written in pidgin English.
Thanks to Herman van Corbach for the tech-talk, at least there is someone working here
with a passion for gadgets and Linux, and together with Nico van Vliet for taking care of
the computer thing in general. Annemarie Teunissen, Debbie Vrieze, Tanja Gerritsen and
Jacqueline Emmerich are thanked for helping with the office practicalities, where it needs
to be remarked that office practicalities comprise everything from handing out paperclips
to offering moral support. Also thanks to all the other colleagues in the applied mechanics
group for providing such a relaxed working environment.
Apart from colleagues and work related things, I also want to thank the people who
have made my stay in Enschede so much more enjoyable. In one particular order there
is Gert Brendel, with whom I suffered through a whole year of organising fun activities.
Then there is Mark Arends, with whom I have spent quite a few nights watching movies
and doing a Waldorf and Stadler. And finally there is Sander Korthouwer, who is just too
incorrigible for his own good. The latter two have even agreed to being my paranimfs. And
120 Acknowledgements
to make sure I do not forget anyone: a thank you to all my friends for providing the ideal
background for doing a PhD.
A special note of appreciation goes to my parents for pushing me to go and study, and
making sure I had every opportunity to do so. And the last thank you goes to Daniëlle for
giving the best support anyone could want.
A BOUT THE AUTHOR
was performed as an employee of the Netherlands Institute for Metals Research (NIMR)
and was performed at the faculty of Engineering Technology in the group of Applied Mech-
anics and Composites of the University of Twente.