Sie sind auf Seite 1von 216

Universit e catholique de Louvain

Ecole polytechnique de Louvain


D epartement de M ecanique
Center for Systems Engineering and Applied Mechanics
Modelling of texture and hardening
of TWIP steel
-
Advanced nite element representation of
polycrystalline aggregates
Maxime Melchior
Th`ese soutenue
en vue de lobtention du grade
de Docteur en Sciences de lIngenieur
Membres du Jury:
Prof. G. Winckelmans, President (UCL-Departement de Mecanique)
Prof. P. Jacques, (UCL-Departement des Sciences des Materiaux et des Procedes)
Prof. A.M. Habraken, (ULg-Departement ArGEnCo)
Prof. V. Legat, (UCL-Departement de Mecanique)
Prof. J.F. Remacle, (UCL-Departement de Genie Civil)
Dr. F. Roters, (Max-Planck-Institut f ur Eisenforschung, D usseldorf)
Prof. P. Van Houtte, (KULeuven-Department of Metallurgy and Materials
Engineering)
Prof. L. Delannay, Promoteur (UCL-Departement de Mecanique)
November 13, 2009
i
Acknowledgments
First of all, I would like to thank my advisor Laurent Delannay for
his invaluable support and help, for providing me with so much useful
advice, many ideas and for the huge amount of time he devoted to me,
especially during the rst two years of my Ph.D.. I am also very grateful
for his careful reading of the thesis and for the numerous improvements
that he suggested.
I would also like to thank all the members of my thesis committee for
their questions and feedback, which have helped me greatly in improving
this thesis. I would like to thank all the people involved in the Winnomat
project for the discussions and their advices. In particular, I would like
to thank Alain Schmitz, Mathieu Iker, Krystel Renard, Sophie Ryelandt
and Pascal Jacques for the experimental measurements and their insights
on the modelling assumptions.
Working in the Euler building has been a pleasure, and I would like
to thank all the people who made these years enjoyable. In particular,
I am grateful to Evelyne, Bertin and Bilel with whom I have shared my
oce and to Christophe, Olivier P., Laurent W., Olivier G., Olivier L.,
Damien, Pierre, Richard and Laurence who have become friends more
than colleague. I would like to thank all my chess opponents who have
transformed the lunch time in an escape to another thinking world. I
would like to thank Christophe, Pierre, Damien and Richard for all their
help and advice on the use of Linux and Latex.
I would like to thank Kateline Ondel for her careful reading of the
thesis, which helped to improve the overall quality of the English even
though the subject of the thesis was totally unknown to her.
It would have been very dicult to accomplish this work without the
support of my parents, family and friends, who I would like to thank all.
In particular, I would like to thank my younger brothers who motivated
me throughout my university cursus to be the best possible exemple. Fi-
nally, I would like to thank Christophe, Serge, Julie, Xavier and Melanie
for their friendship and for increasing my self-condence by showing they
hold me in high esteem.
Pecunias nervus belli : This thesis has been funded by the Winnomat
Program of the Region Wallonne (Contract 415660).
ii
Abstract
Steels with enhanced properties are required by the industry for example
in order to produce cars that are safer, lighter and more ecological. These
steels can only be used once their behaviour is fully understood by taking
into account the specic characteristics of these steels in models.
In this thesis, we have mainly focused our attention on the prediction
of the behaviour of the TWIP steel where TWIP is the acronym for
TWinning Induced Plasticity. The combination of high strength and
ductility is obtained because the presence of twins that act as obstacles
to the dislocations. Besides these particular hardening properties, TWIP
steel shows specic textures, especially a brass-type texture after large
deformation by cold rolling. Our modelling was aimed to succesfully
predict the texture and the hardening of TWIP steel in tension and in
plane strain compression. In order to reach this goal I have performed
conjugate eorts at both the micro and the macro scale.
At the micro scale, I have proposed and implemented a new mod-
elling of twinning in a crystal plasticity code. This new modelling re-
lies on experimental measurements available in the literature as well
as new measurements done at UCL. Particularities of our modelling
are the subdivision of grains into untwinned and twinned subregions, a
new technique for the computation of the twin orientation and a criti-
cal comparison of various micro-macro transition schemes. Finally, new
descriptions of the hardening of the slip and twin systems are proposed.
Texture predictions obtained during this thesis are in accordance
with the expectations of Leers in his survey about the prediction of
the brass-type texture [47]. Our modelling improves the quality of the
prediction of rolling textures. Moreover the same set of parameters
provides valid prediction of the texture after a tensile test and of the
hardening.
At the macro scale, I have rstly designed a novel algorithm for the
sampling of crystallographic textures into discrete lattice orientations.
Contrary to previous discretisation techniques, the new method is valid
also when the model microstructure consists of grains with non-uniform
size. The accuracy of the texture representation has been assessed in
the case of cold rolled IF-steel. I have then elaborated a generator
of periodic microstructures containing either one or two phases coupled
with a mesh coarsening procedure. Each microstructure generated fulls
iii
experimental statistics measured on a real microstructure such as the
volume fraction of each phase and the ratio between the mean sizes of
grains of each phase. The particularity of the desired mesh coarsening
is that we preserve grain geometries and, of course, the periodicity of
the mesh. The size eld prescribed to the coarsening procedure has the
specicity that it enforces a larger characteristic length at the centre of
the grains.
As a consequence, we are able to reduce the number of elements
by two while preserving the minimum quality imposed to the elements.
Predictions of the planar anisotropy of steel sheets have shown that
the new type of RVE allows to substancially reduces the number of
grains required in simulations. Results of an extensive study on mesh
renement lead to the advice of performing nite elements simulations
with second-order tetrahedral elements and a minimum of 2000 gauss
points per grain for macroscopic trends. We have used heterogeneous
model microstructures in dierent applications.
iv
Glossary
Crystal plasticity
L
p
: plastic velocity gradient
M
s
: the Schmid tensor of a given crystallographic plane and di-
rection

s
: shear rate due to dislocation slip along a given crystallo-
graphic plane and direction

0
: the reference slip rate

s
c
: critical resolved shear stress inducing slip

s
c0
: initial value of
s
c
Twinning modelling

t
c
: critical resolved shear stress inducing twinning

t
c0
: initial value of
t
c
SFE: stacking fault energy

tw
: shear induced by part of a crystalline which twins
f
t
: volume fraction of twins created along a given crystallographic
plane and direction
TV F: twin volume fraction

t
: shear rate due to twinning along a given crystallographic plane
and direction
: dislocation density
Contents
1 Introduction 1
2 Theoretical and experimental framework 5
2.1 Context of the study . . . . . . . . . . . . . . . . . . . . . 5
2.2 Ultra high-strength steels . . . . . . . . . . . . . . . . . . 7
2.3 Mathematical treatment of plasticity under nite strains . 9
2.4 Finite element simulations . . . . . . . . . . . . . . . . . . 10
2.5 Relevant length-scales from micro to macro . . . . . . . . 11
2.6 Denition of texture . . . . . . . . . . . . . . . . . . . . . 14
2.7 Representation of texture . . . . . . . . . . . . . . . . . . 15
2.8 Modelling of the micro-macro transition . . . . . . . . . . 18
I Modelling of the mechanical behaviour of TWIP steel
23
3 Model description 25
3.1 Physical basis of the model . . . . . . . . . . . . . . . . . 25
3.2 Theory of crystal plasticity in metals deforming by dislo-
cation slip . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
v
vi CONTENTS
3.3 Theory of crystal plasticity adapted to metals deforming
by dislocation slip and by twinning . . . . . . . . . . . . . 38
3.3.1 Generalised Schmid law extended to twinning . . . 38
3.3.2 Characteristics of newly-formed twins . . . . . . . 39
3.3.3 Hardening of the slip systems . . . . . . . . . . . . 40
3.3.4 Evolution of the CRSS for twinning . . . . . . . . 43
3.3.5 Onset of twinning . . . . . . . . . . . . . . . . . . 44
3.3.6 Initial state of the newly created twin . . . . . . . 46
3.4 Numerical implementation . . . . . . . . . . . . . . . . . . 47
3.4.1 Fully-implicit time integration of crystal plasticity
theory in absence of twinning . . . . . . . . . . . . 51
3.4.2 Partly-explicit time integration of crystal plastic-
ity theory in presence of twinning . . . . . . . . . . 52
3.4.3 Explicit versus implicit integration of the slip sys-
tem hardening . . . . . . . . . . . . . . . . . . . . 54
4 Assessment of the model 57
4.1 Inuence of twinning on texture . . . . . . . . . . . . . . . 59
4.2 Inuence of grain interaction on texture . . . . . . . . . . 66
4.3 Eect of anisotropic hardening on texture . . . . . . . . . 73
4.3.1 Dierence between the twinning and hardening rates 73
4.3.2 Twin-twin interaction . . . . . . . . . . . . . . . . 78
4.3.3 Slip-twin interaction . . . . . . . . . . . . . . . . . 80
4.3.4 Texture prediction with our most elaborated parametri-
sation . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.4 Hardening prediction . . . . . . . . . . . . . . . . . . . . . 87
4.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
CONTENTS vii
II Advanced CPFEM in polycrystalline aggregates 97
5 Construction of model polycrystals 99
5.1 A novel algorithm for texture discretisation . . . . . . . . 99
5.1.1 State of the art . . . . . . . . . . . . . . . . . . . . 100
5.1.2 Description of the algorithm . . . . . . . . . . . . . 102
5.1.3 Comparison with the Toth and Van Houtte method105
5.2 Mesh generation algorithm . . . . . . . . . . . . . . . . . 107
5.2.1 State of the art . . . . . . . . . . . . . . . . . . . . 107
5.2.2 Description of the algorithm . . . . . . . . . . . . . 109
5.2.3 Mesh coarsening in a periodic RVE . . . . . . . . . 113
5.2.4 Parameterisation of the size eld . . . . . . . . . . 117
5.2.5 Post processing operation . . . . . . . . . . . . . . 118
5.2.6 Parametric study of the coarsening procedure . . . 121
6 Assessment of the new CPFEM 127
6.1 Optimal size of the RVE . . . . . . . . . . . . . . . . . . . 127
6.2 Convergence of CPFEM with mesh renement . . . . . . 134
6.2.1 Description of the reference simulation . . . . . . . 134
6.2.2 Mesh renement adapted to the simulation goal . . 136
6.2.3 Inuence of the polycrystal size . . . . . . . . . . . 142
6.2.4 Inuence of the type of nite elements . . . . . . . 142
6.2.5 Inuence of the crystal symmetry (BCC vs. HCP) 147
6.2.6 Inuence of the positioning and shape of grains . . 148
6.3 The benet of selective mesh coarsening . . . . . . . . . . 152
viii CONTENTS
6.3.1 Trends in the displacement eld . . . . . . . . . . 153
6.3.2 Trends in the strain eld . . . . . . . . . . . . . . . 156
6.3.3 Summary . . . . . . . . . . . . . . . . . . . . . . . 159
6.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . 162
7 Use of CPFEM in practical studies 163
7.1 Strain heterogeneity in dual phase steel . . . . . . . . . . 163
7.2 Grain shape inuence on planar anisotropy . . . . . . . . 168
7.3 Application of the advanced CPFEM to TWIP steel . . . 170
8 Conclusion 175
A In depth comparison of various predictions obtained dur-
ing this thesis 183
B Quaternion manipulation 193
C k-d Tree code 195
Chapter 1
Introduction
There is an ongoing race among materials engineers to produce steels
with ever more complex microstructures in order to improve mechanical
properties. One of the consequences is a better combination of strength
and ductility than in many other types of alloys. This is required in
the automotive industry where the objective is to produce safer, lighter
and more ecological cars. However, relying on such advanced steels
implies improving the modelling in order to tackle the complexity of
the microstructure. Indeed, the full potential of these steels can only
be obtained if their behaviour is completely understood. The latter is
possible only in polycrystalline models accounting for phenomena that
occur at the microscopic scale.
The steel grade investigated in this thesis deforms by dislocation slip
and by twinning. Experimental measurements show that, probably due
to the formation of very thin twin lamellae acting as obstacles against
dislocation slip, the combination of strength and ductility of this alloy
is excellent. Another particularity of this steel is the texture produced
during cold rolling which aects the anisotropy of the sheets. The only
way to predict both the texture development and the hardening capac-
ity is obviously to incorporate twinning in the crystal plasticity code
previously available in UCL.
The rst part of my thesis is dedicated to the modelling of TWIP
steel mostly at the grain level. I rstly present observations that have
been made on various steels which deform by slip and by twinning and
I emphasise which observations are used as guidelines for the modelling
1
2 CHAPTER 1. Introduction
and how they are incorporated in the crystal plasticity theory. I then
present our mathematical modelling and highlight the dierences with
other existing models. One of the originalities of our approach is the
simplied modelling of the interaction of adjacent grains. In the last
chapter (Chap. 4) of the rst part, as an introduction to the second part
of the thesis, predictions of the texture development and of the hardening
obtained with this modelling are evaluated. The inuence of the various
parameters of the model is studied with special attention to the inuence
of the micro-macro transition on the prediction. An optimal combination
of parameters is then identied based on a comparison with experimental
data in the case of a tensile test as well as cold rolling where we consider
two values of the thickness reduction ratio.
In the second part of the thesis, we attempt to improve the model
predictions using nite element modelling at the grain scale and relying
on an advanced representation of the alloy microstructures. In common
practice, modellers typically consider idealised microstructures with uni-
form grain shape and size whereas real microstructures are highly het-
erogeneous. Furthermore, when several phases are present, the mean
volume of a grain can be quite dierent from one phase to the other. In
the case of FE simulations, the accuracy of the result hinges on a su-
cient mesh renement and there is no thorough study of the number of
elements required in each grain in order to obtain a converged solution.
The second part of the thesis is organised as follows. I rstly present
a novel algorithm for texture discretisation valid when grains have non
uniform sizes. I briey recall existing techniques and then explain the
novel algorithm in detail based on a simple example. The inuence of our
technique on the number of grains required in a model microstructure is
highlighted. I then present the generator of model microstructures that
has been developed during my thesis and compare the outcome with the
model microstructures used in the literature. I also introduce a tech-
nique of mesh coarsening that helps reducing the number of elements
used in the simulation while maintaining periodic boundary conditions
and without signicant decrease of the prediction quality. The inuence
of various parameters of the coarsening procedure on the quality of the
prediction is studied. The number of integration points required per
grain is shown to depend on the material characteristics. A compari-
son between predictions obtained with heterogeneous and homogeneous
model microstructures is also presented. The inuence of the microstruc-
ture and particularly of the grain shape on the Lankford coecient is
studied in detail. Our model microstructure is used for the prediction
3
of strain heterogeneity in dual phase steels. Finally, in order to link the
two parts of my thesis and to close the loop, CPFEM simulations with
our optimal combination of parameters have been performed. The goal
is to study the inuence of the mesh renement on the possibility for
each grain to achieve the deformation with a single slip system.
4 CHAPTER 1. Introduction
Chapter 2
Theoretical and
experimental framework
In this chapter, I introduce all the concepts that will be used throughout
the work.
Which mathematical formalism do we use to perform metal form-
ing simulations?
Which length scales are involved in our simulation?
What kind of model microstructures do we use in our simulations?
How do we visualise the texture?
How do we model scale transition?
Which mathematical tools do we use in order to obtain these pre-
dictions?
How do we incorporate experimental observations in our model?
2.1 Context of the study
Nowadays, large amounts of steel are used in the automotive industry.
A steel sheet delivered to an automotive constructor must pass through
5
6 CHAPTER 2. Theoretical and experimental framework
a metal forming process in order to obtain the required shape (Fig. 2.1).
People in charge of metal forming processes rely on computer models to
Figure 2.1: Result of the simulation of a normalised deep drawing test
with a cross-shaped tool [1]. One quarter of the sample is shown.
determine whether a given alloy will sustain the strains involved in the
forming operation without undergoing plastic localisation or damage.
They also need to know the loads to be applied to the tools and the
exact shape and thickness of the steel at the end of the process, which
depend not only on the shape of the tools but also on elastic springback,
that is : on the stress state within the part before removal of the tools.
Residual stresses and the hardening state are important to control as
they inuence the response of the part in service, for example the ability
of a car body frame to absorb energy in case of a crash. Simulations are
less time consuming and can be used more eectively than experiments
in order to optimise forming operations.
Simulations of metal forming operations imply the solution of equi-
librium equations over a domain (the work piece) subjected to specic
2.2. Ultra high-strength steels 7
boundary conditions, such as the contact with the tools. This in turn
requires being able to determine the stress corresponding to any given
strain history that a material point may undergo. Updating the stress at
a given time and position is done by time-integration of the constitutive
law. In elastic-plastic materials, the stress update depends not only on
the current total strain but also on the strain history reected in the
values of a set of state variables. The displacement eld which is a
solution of the problem and the associated stress and state variables can
almost never be determined analytically. Most problems of interest are
solved numerically and we choose to adopt the nite element technique
although the work done in this thesis could be extended to other numer-
ical schemes. When such nite element simulations are carried out in
industry, most of the reported shortcomings are believed to be the result
of insuciently accurate constitutive laws. Advanced metallic alloys do
not obey classical plasticity theories, especially regarding anisotropy and
hardening (both isotropic and kinematic).
The present work is based on the assumption that the reliability of
the predictions could be improved through the use of a micro-macro
modelling scheme. In this case, the macroscopic response is computed
as the average response obtained on a microstructure that is statistically
representative of the real microstructure.
2.2 Ultra high-strength steels
In order to full more and more severe performance criteria, the industry
is developing new steel grades which lead to lighter and more ecologi-
cal nal products, e.g. car body frames. In this context, multiphase
materials with high resistance are developed (Fig. 2.2). For most mate-
rials, high ductility implies low strength whereas energy absorption and
formability require a high ductility combined with a high strength.
Dual phase steels have a high strength and a low ductility. They are
composed of two phases, martensite, the hard phase, coupled with ferrite
or bainite, the soft phase. Martensite inclusions increase the strength
but they are generally brittle. Ductility, up to 10% uniform elongation,
is provided by the soft matrix. Ductility is improved in the case of TRIP
steel, where TRIP is the acronym for TRansformation Induced Plastic-
ity. These multiphase-steels are composed of ferrite, bainite and retained
8 CHAPTER 2. Theoretical and experimental framework
Figure 2.2: Illustration of the combination of strength and ductility for
some ultra high-strength steel. TRIP is the acronym for TRansforma-
tion Induced Plasticity. TWIP is the acronym for TWinning Induced
Plasticity. F represents ferrite, B, bainite and M, martensite
austenite. During deformation, the austenite transforms progressively
into martensite. This increases the strain hardening capacity, postpon-
ing necking occurrence. The principal alloying elements are manganese
and carbon. When the volume fraction of manganese is raised, one
produces TWIP steel which is a second family of ultra high-strength
steels. TWIP is the acronym for TWinning Induced Plasticity. In my
thesis, the alloy used in the experiments is Fe-22Mn-1.3C. TWIP steels
are fully austenitic. During deformation, the mean free path of disloca-
tions is reduced due to the appearance of twins. The reduction of the
mean free path is one of the explanations for the combination of very
high strength and ductility reached by these steels. One of the motiva-
tions for my thesis is related to the very complex and evolutive structure
of TWIP steel. It is the reason why simplied models do not lead to
accurate predictions.
2.3. Mathematical treatment of plasticity under nite strains 9
2.3 Mathematical treatment of plasticity under
nite strains
Metal forming processes lead to very large deformation and a non-linear
behaviour of the material. As plasticity is strain-path dependent, con-
stitutive equations are written in a rate form. For the purpose of the
simulation, the prescribed deformation is thus subdivided into small
increments. The size of the increments is chosen so as to ensure con-
vergence of the solution of the non-linear systems of equations involved
and also to capture non-monotonic loading conditions. At the end of
each increment, the new state of the material is computed in function
of the previous state of the material and of the prescribed increment of
deformation.
When the material undergoes large deformation, the reference and
the current conguration cannot be considered equivalent anymore. Strain
and stress tensors take dierent values depending on whether they are
computed with respect to one conguration or the other. New deni-
tions are hence required. Each material point P has a position

X in the
reference conguration that becomes a position

x in the current cong-


uration. There is a biunivocal relation between these two vectors. Each
property of a material can be described in either type of coordinates.
For example, the displacement can be expressed as

u(

X, t) =

x(

X, t)

X (2.1)
or

u(

x, t) =

X(

x, t). (2.2)
The relation between an innitesimal distance

dX in the reference
and

dx in the current conguration is given by the deformation gradient


tensor F:

dx =

dX = F

dX. (2.3)
Based on Eq. 2.1, the relation between the displacement and F is
F = I +

X
(2.4)
10 CHAPTER 2. Theoretical and experimental framework
The velocity gradient tensor L which is dened as

dv = L

dx (2.5)
is equal to
L =


u

x
=


x

x
=


x

x
=

FF
1
. (2.6)
This tensor can be decomposed into its symmetric and antisymmetric
parts:
L = D+W =
1
2
(L+L
T
) +
1
2
(LL
T
) (2.7)
where D is the strain rate and W is the spin tensor. The internal
deformation energy may be computed by integrating the product of the
Cauchy stress with the strain rate:
E(t) =
_
t
t
0
: Ddt (2.8)
The Cauchy stress is a common, rather intuitive denition of stress in
the nite strain regime. It has one leg in the initial conguration and
one leg in the deformed conguration: if we name

n the normal of an
innitesimal area in the deformed conguration, and

f the force applied


on the same surface in the reference conguration, then: f
i
=
ij
n
j
.
The spin W does not contribute to the deformation energy. It will
however induce or inuence the rotation of the stress tensor. Properly
accounting for such rotation is required in order to ensure objectivity
of the model, i.e. ensure that the material response is independent of
the observer referential. In plastically deforming materials, the rate of
rotation of the stress tensor is obtained by subtracting from W the
plastic spin denoted W
p
. In crystalline materials, elastic strain are
propotional to lattice plane spacings. Hence, the stress tensor rotates
with the crystal and one writes:

= W W
p
where

is the rate
of lattice rotation with regard to the reference conguration. A proper
expression for W
p
is provided in Section 3.4.1.
2.4 Finite element simulations
In the present work, all FE simulations are performed with the ABAQUS
software. The nite element mesh is generated using dedicated software
2.5. Relevant length-scales from micro to macro 11
(gmsh) when the geometry is elaborated and it is generated by hand
in the case of structured polycrystals. The material response is im-
plemented using a UMAT, i.e. a user subroutine which describes the
constitutive behaviour and performs time-integrations of the selected
law.
When the simulation involves nite deformations, ABAQUS ensures
objectivity by relying on the Hughes and Winget algorithm [2]. As
thoroughly justied in [3], this implies deriving the velocity gradient
tensor L at time t +
t
2
from the deformation gradient tensor that is
known at time t and estimated at time t + t based on the iteratively
determined displacement eld:
L
t+
t
2
=
1
t
_
1
2
(F
t
+F
t+t
)
_
1
(F
t+t
F
t
) (2.9)
The velocity gradient tensor is considered constant and equal to L
t+
t
2
during the time step, which is a valid assumption if one controls the
time step so as to impede strain increments larger than, say, 5% and
rotation increments larger than a few degrees. In any case, beyond
such thresholds, the FE computation is slowed down because solving the
very large set of non-linear equations leading to the displacement eld
becomes more complicated. The latter iterative FE solution requires
that the UMAT routine provides not only the stress tensor and the
state variables at time t +t but also the consistent (or algorithmic)
tangent operator dened as follows:
C
alg
=
1
t

t+t
D
t+
t
2
(2.10)
Here D refers again to the symmetric part of L.
In the output of the UMAT that we have developed, all tensors are
expressed in a xed reference frame that is common to all integration
points across the FE mesh.
2.5 Relevant length-scales from micro to macro
Various length scales (Fig. 2.3) must be dened before we start reviewing
the literature and developing the model.
12 CHAPTER 2. Theoretical and experimental framework
Figure 2.3: The various length scales considered in a micro-macro sim-
ulation of metal forming.
The rst one is the atomistic scale at which the ordering of the
atoms in the crystal lattice and the eventual presence of alloying
elements may be described. This scale is not used in the model
but it denes the material that we are dealing with.
The second one is the scale of individual dislocations. These crys-
talline defects should at the same time be seen as the vector of
plastic deformation and as the obstacles responsible for hardening.
2.5. Relevant length-scales from micro to macro 13
In this work, dislocations will never be considered individually but
one will account for the crystal directions and planes along which
dislocations slip across the lattice, i.e. the slip planes. Moreo-
ever, in a phenomenological way, the hardening laws account for
dislocation interactions.
The third one, the scale of individual twins, is particular to our
material. As explained later in the text, twins are lamella-shaped
micro-regions which are delimited with regard to the surrounding
crystal by a particular arrangement of dislocations. The size of
twins and their thickness-to-width ratio depend on the alloy under
consideration.
The fourth one is the scale of slip systems and of twinning systems.
This scale is not linked to real features of the microstructure. It
is the scale at which crystal plasticity theory is assumed to ap-
ply and also the smallest at which continuum mechanics makes
sense in crystalline materials. At this scale, the eect of individ-
ual dislocation loops and the eect of occurrence of new twins can
be averaged out into, respectively, a rate of dislocation slip along
each slip system and a twinning rate. It is also the scale at which
the generalised Schmid law applies: the resistance against dislo-
cation slip is averaged out over a crystalline volume. When the
alloy is subjected to twinning, the lattice orientation is no longer
unique at this scale: a dierent lattice orientation is associated to
each twinning system.
The fth scale is the grain scale. This scale is clearly identied in
the real microstructure. In the modelling, it is sometimes used, not
systematically however. Indeed, some models - sometimes referred
to as Taylor-type models - do not dierentiate this scale and the
previous one (the fourth scale) whereas in more advanced mod-
els, such as CPFEM, the fourth scale typically applies to volumes
much smaller than the grain
The sixth scale is the scale of clusters of adjacent grains (or sub-
grains) that may be involved in micro-macro transition schemes.
This scale is dened for modelling purposes only. As discussed
later on, in certain models, crystal plasticity equations are solved
simultaneously over regions comprising more than a single grain.
The seventh scale is the scale of the polycrystalline aggregate. In
the literature, this scale is commonly called the Representative Vol-
ume Element (RVE) because it provides a statistically meaningful
14 CHAPTER 2. Theoretical and experimental framework
description of the microstructure (and the texture). Depending
on the alloy under consideration and on the model used, the poly-
crystalline aggregate should contain between 500 and 5000 grains.
The last scale is the macro scale. It is the scale at which strain
gradients must be captured in real-scale simulations of metal form-
ing processes, i.e. the volume assigned to an individual integration
point if the nite element technique is used.
Note that the last two scales may be interchanged. In nite element
simulations of real-scale forming processes, the polycrystalline aggregate
(i.e. the RVE) may be reproduced completely at every integration point
of the mesh, or it may be distributed over several adjacent elements
using the CPU-saving strategy described in [4]. Other strategies have
been proposed to perform micro-macro simulations of real scale forming
processes [5, 6, 7].
2.6 Denition of texture
Crystallographic texture is an important property of metallic alloys that
changes during plastic deformation as well as heat treatments (due to
recrystallisation or phase transformation). A material is said to be tex-
tured if certain lattice orientations are more frequent (or probable) than
others across the microstructure. In this work, we are mostly inter-
ested in deformation-induced textures. In this case, texture components
(i.e. preferential lattice orientations) occur because they correspond
to stable orientations of the lattice under the imposed deformation
mode. This notion of stable orientation is discussed in Section 3.2, where
crystal plasticity is presented. The more a polycrystalline aggregate is
deformed, the greater the proportion of stable lattice orientation with
respect to other orientations, and the stronger the texture.
It is important to obtain accurate predictions of deformation tex-
ture because there is a link between texture and the anisotropy of the
formed part. Texture, when it is present, makes the anisotropy of indi-
vidual grains visible at the macroscopic scale. Although there are a few
exceptions, such as zinc alloys, the anisotropy of the plastic deformation
is generally much larger than the elastic anisotropy. In rolled sheets, it
is common to probe the planar anisotropy by performing tensile tests,
2.7. Representation of texture 15
during which the Lankford coecient is computed as the ratio of the
plastic deformation rates in the two transverse directions,

D
p
22

D
p
33
where

x
3
is the sheet plane normal. The importance of planar anisotropy
appears clearly after deep-drawing of a cylindrical cup [8].The relative
height of the ears and the largest height of the cup that can be drawn
without failure (linked to the limiting drawing ratio) depend strongly
on the texture of the sheet.
In polycrystalline materials, it is the texture which mainly denes
the size of the RVE. Indeed reproducing a texture, i.e. an orientation
distribution function (ODF), requires a much larger number of grains
than, for instance, reproducing a distribution of grain sizes or shapes.
Texture is important both as an input and as an output of crystal plas-
ticity simulations. The polycrystalline aggregate that is used as RVE
should
respect the volume fraction of the various phases.
contain a set of lattice orientations which is statistically represen-
tative of the texture inside each phase.
have a geometric representation of the grain which is statistically
representative of the average grain shape.
In the reminder of the thesis, we will consider that the lattice disorien-
tation of adjacent grains is random and that grain shape and size is not
correlated with lattice orientation. These notions are sometimes referred
to as microtexture as opposed to the macrotexture which we here
simply call texture.
Figure 2.4 represents a nite element mesh representation of an RVE
in which the average grain shape is equi-axed and the initial experimen-
tal texture is discretised into 97 grain orientations.
2.7 Representation of texture
Texture is mathematically represented by an orientation distribution
function, ODF, which denes the probability f(g) that an elementary
volume dv presents the lattice orientation g. Throughout this work, we
16 CHAPTER 2. Theoretical and experimental framework
Figure 2.4: Example of a model microstructure statistically representa-
tive of a real microstructure.
make use of the MTM-FHM software for texture analysis [9] in which the
ODF is expressed in the form of a truncated development of generalised
spherical harmonic functions. The latter allow, for example, to eciently
convert a set of discrete lattice orientations into a continuous ODF based
on a superposition of gaussian distributions [10]. We use the Bunge
convention in order to dene individual lattice orientations based on
Euler angles
1
, and
2
which represent three successive rotations that
must be applied to bring the referential of the lattice in correspondence
with the macroscopic referential.
The ODF can be obtained experimentally. It is either directly ob-
tained from electron back-scatter diraction (EBSD) measurements (based
on the volume fraction of each orientation in the mapped area) or can
2.7. Representation of texture 17
be derived from pole gures produced from X-ray or neutron diraction.
There are many ways to visualise a texture or part of a texture.
The rst one is to plot the ODF intensity across a given section of the
orientation space, where one of the Euler angles keeps a constant value.
For example, in gure 2.5, the
2
= 45 section of an experimental
texture measured after cold rolling is plotted. Contour lines separate
orientations that have an ODF intensity between 1 and 2, 2 and 4 etc..
The angle
1
evolves increasingly from left to right and the angle
evolves decreasingly from top to bottom.
PAGE
0 90
PHI2= 45
Figure 2.5: Example of a section
2
= 45 of an ODF. Contour lines :
1-2-4-8.
90

0
0
1
90
Another option is to visualise a pole gure using a stereographic pro-
jection. Pole gures are partial representations of the texture (because
several pole gures are required to compute the ODF) but they are of-
ten easy to interpret. For example, in the < 100 > pole gure, each
lattice orientation is represented by three points in a pole gure. Each
point corresponds to the stereographic projection of one direction of the
lattice referential. Contour lines are then used to represent the densities
of the pole (Fig. 2.6).
A third way to visualise a part of the texture is the skeleton line. It
is the evolution of ODF intensities along an orientation bre. A bre
is a path of high ODF intensities between two main components of the
texture. A skeleton line is fully represented by three graphs (Fig. 2.7),
one for the ODF intensity along the path and two others for the position
of the bre in Euler space. In gure 2.7,
2
evolves by step of 5 from
90 to 0. Therefore, there is a plot for the value of
1
and a plot for
the value of along the path.
18 CHAPTER 2. Theoretical and experimental framework
Figure 2.6: Example of pole gure 111 and 100. Contour lines : 1-1.4-
1.7-2.1-2.4.
A last way to visualise the texture is the computation of the volume
fractions of the main components of the texture. These volume frac-
tions are obtained by integrating the ODF on a box centred on each
component.
There are various typical components in cold rolling textures. In this
thesis, we will deal mainly with face centred cubic crystal symmetry and
the components are then Copper, Brass, S, Goss and Cube. In the case of
TWIP steel, the rolling texture is mainly composed of the Brass and the
Goss components (Fig. 2.8). The Copper component is also present but
not with the same intensity as for fcc metals deforming by dislocation
slip only.
2.8 Modelling of the micro-macro transition
We intend to predict the macroscopic mechanical response of metallic
alloys, such as TWIP steels, based on crystal plasticity theory. This
means that time-integration of the crystalline constitutive law has to
be performed at dierent places throughout the model microstructure,
i.e. at the fourth scale presented above, where the local deformation is
likely to dier from the macroscopically applied deformation. Indeed,
the assumption of a uniform strain across the polycrystal due to Taylor
[11] leads to an upper bound of the deformation energy. In reality,
the polycrystal is subjected to a heterogeneous strain eld where softer
2.8. Modelling of the micro-macro transition 19
(a) Intensity
(b) Position of
(c) Position of
1
Figure 2.7: Example of skeleton line along the bre.
20 CHAPTER 2. Theoretical and experimental framework
PAGE
0 90
PHI2= 45
Figure 2.8: Example of a section
2
= 45 of an ODF. Contour lines :
1-2-4-8

Copper

Brass
d
ds
Goss
regions are more deformed than harder regions. However, the stress
developed in the softer regions remains lower than the stress in harder
regions since the iso-stress assumption [12] is a lower-bound estimate of
reality.
We make use of the velocity gradient tensor L dened in Section
2.3. We impose to all homogenisation schemes to be tested that <
L >

=

L where < L >

represents the average of L over the volume


of microstructure at the borders of which the macroscopic velocity
gradient

L is prescribed. The macroscopic Cauchy stress is then
computed as < >

.
Full eld solutions of the stress and strain distributions across the
model microstructure can be applied with numerical techniques such as
the FE method. Based on variational approaches, such homogenisation
schemes rely on a minimisation of the overall deformation energy while
allowing interactions between neighbouring grains as well as in-grain het-
erogeneity of strain. Boundary conditions ensure that the macroscopic
velocity gradient is achieved on average over the volume. However, as
illustrated during this thesis, FEM simulations have huge computational
cost. Therefore, cheaper micro-macro transitions based on simplifying
assumptions need to be developed. The latter models do not exactly
satisfy stress and strain compatibility in each material point but their
solution always lies between the upper and lower bounds.
The ALAMEL model [13] allows interactions only between a pair of
neighbouring grains (Fig. 2.9a), assuming that micromechanics is dic-
2.8. Modelling of the micro-macro transition 21
tated by such short range interactions. Local velocity gradients resulting
from the interactions are computed at the interface between the grains.
Many pairs of grains are used in order to represent a polycrystalline
aggregate (i.e. a RVE) (Fig. 2.9b) and the orientations of the normals
to the interfaces,

n, are statistically representative of the average grain
shape. For this purpose, the initial grain shape is represented by an
ellipsoid and the interface of each stack of two regions is dened as the
plane tangent to the ellipsoid, at the intersection of a random direc-
tion,

x with the ellipsoid. If the principal diameters of the ellipsoid are
denoted a, b, and c, the initial normal direction is expressed by :

n =
_
x
1
a
2
,
x
2
b
2
,
x
3
c
2
_
/

_
x
1
a
2
_
2
+
_
x
2
b
2
_
2
+
_
x
3
c
2
_
2
. (2.11)
As the material deforms, the interface normal is updated according to
the macroscopic deformation gradient :

n =
_
F
1
_
T
:

n.
(a)
(b)
Figure 2.9: Illustration of the ALAMEL model. a) Types of interaction
between a pair of neighbouring grains. b) Many pairs of grains are used
in order to represent a real grain and his neighbourhood.
22 CHAPTER 2. Theoretical and experimental framework
The velocity gradient is assumed to be uniform within individual
grains. However, it diers from the macroscopic value. In supplement
to the latter

L, the two regions undergo opposite shear parallel to the
grain boundary. Hence, the velocity gradient tensors of the two grains
are expressed by :
L
I
=

FF
1
+
(1)
rlx
(

d
(1)

n) +
(2)
rlx
(

d
(2)

n),
L
II
=

FF
1

(1)
rlx
(

d
(1)

n)
(2)
rlx
(

d
(2)

n),
(2.12)
where

d
(1)
and

d
(2)
are arbitrary directions in the plane of the inter-
face. The local heterogeneity of strain is thus cancelled out over every
stack of two regions. In other words: the macroscopic velocity gradi-
ent is achieved on average over the two regions and therefore also over
the whole polycrystal. The amplitude of shear is determined by the
two scalars,
(1)
rlx
and
(2)
rlx
, which are computed by minimising the de-
formation energy of the two regions. This is approximately fullled by
requiring that
irlx
is the solution of

I
: D
I
+
II
: D
II
_

irlx
= 0 for irlx = 1, 2. (2.13)
This is equivalent to requiring that :
(
I

II
) : (d
(1)
n) = (
I

II
) : (d
(2)
n) = 0, (2.14)
where
I
and
II
represent the Cauchy stresses of the two regions.
In the viscoplastic self-consistent (VSPC) formalism [14], the RVE
is represented by a collection of weighted, ellipsoidal grains. Each of
these grains represents the average behaviour of all grains that have
a particular crystallographic orientation and morphology, but dierent
environments. Each grain is treated as an ellipsoidal viscoplastic inclu-
sion embedded in an eective viscoplastic medium. Both, inclusion and
medium, have anisotropic properties. It is therefore a model which takes
long range interactions into account.
In the Relaxed Constrained model [15], which is only valid for rolled
materials, the 13 and 23 components of the velocity gradient tensor are
free if

x
3
represents the normal to the rolling plane. In these materials,
grains are attened. Therefore a large proportion of the grain boundaries
are parallel to the rolling plane.
Part I
Modelling of the
mechanical behaviour of
TWIP steel
23
Chapter 3
Model description
3.1 Physical basis of the model
In this section, we review experimental observations of the deformation
mechanisms in TWIP steel that will be used in the design of our model.
The section is organised in function of the scale where these observations
are made, progressing from the nano to the macroscale.
Crystalline defects : In TWIP steel, the crystal symmetry is face-
centred cubic (FCC) (Fig. 3.1). The crystal contains a lot of line defects
which are called dislocations (Fig. 3.2). Plastic deformation is achieved
mainly by dislocations with a < 110 > Burgers vector and by partial
dislocations with a < 112 > Burgers vector. Partial dislocations are for
example obtained by splitting edge dislocations and have a Burgers vec-
tor which does not link two atoms. The sum of the Burgers vectors of
the two partial dislocations is the original Burgers vector. As shown on
gure 3.3, the glide of a partial dislocation brings about a stacking fault,
i.e. a fault in the normal packing of close packed planes of the fcc crys-
tal. Whereas dislocations are one-dimensional defects, stacking faults
are two-dimensional defects carrying a certain stacking fault energy
(SFE). Dislocations form loops which may grow and sweep the crystal
under an applied stress. The relative alignment of the dislocation line
and the Burgers vector gives to the dislocation mixed edge and screw
character.
25
26 CHAPTER 3. Model description
Figure 3.1: Representation of the FCC crystal structure and of the or-
ganisation of the lattice of an FCC material.
Figure 3.2: Schematic representation of a dislocation
Slip systems : When many dislocations have moved across the lattice,
the material is sheared. Such shear occurs along well-dened (densely-
packed) planes and directions. In the case of FCC crystals, one has
to consider the 111 plane family and the < 1

10 > direction family.


There are four 111 planes and three < 1

10 > directions in every one


of them, thus 12 combinations, also called slip systems. It is observed
that activated slip systems are those which are oriented so that the
externally applied stress results in the largest resolved shear stress. The
resolved shear stress is the shear stress acting parallel to the slip plane
in the direction of the Burgers vector.
Description of twins in FCC metals : A deformation twin is a de-
fect resulting from plastic deformation of some materials characterised
by low SFE. Due to this low SFE, dislocations tend to dissociated them-
selves into partial dislocations. Glide of partial dislocations on adja-
cent planes as illustrated on gure 3.3, then leads to the formation of
a twinned region. The lattice of a twin is a mirror image of the lattice
3.1. Physical basis of the model 27
Figure 3.3: Illustration of the crystal lattice of partial dislocations and
twins.
of the parent grain (Fig. 3.4a). A deformation twin can thus be con-
sidered as a lenticular subregion which has undergone a discrete shear
and in which the lattice is rotated relative to the surrounding by 180
around the normal to the twinning plane. Heavily twinned grains lead
to subdivision of grains in a large number of small untwinned regions
(Fig. 3.4b).
There is no consensus about the exact formation mechanism of twins.
It seems that this mechanism diers from one material to another. Ex-
perimental observations tend to show that twins are formed by particular
reactions between dislocations bringing about particular sessile disloca-
tions. Twins can present dierent thickness evolution. For example in
hcp material, twins thicken with plastic deformation [17] whereas in our
material, the evolution of the twin volume fraction is only governed by
the apparition of new thin twins which never thicken. Twins appear on
well-dened planes and and are sheared in well-dened directions. In
the case of the TWIP steel, these planes are the 111 plane family and
the < 1

21 > direction family. The number of combinations is therefore


12. Contrarily to dislocation glide, twinning is unidirectional : only a
positive resolved shear stress may create a twin.
28 CHAPTER 3. Model description
(a)
(b)
Figure 3.4: a) Representation of the mirror symmetry between the twin
lattice and the lattice of the untwinned region [16]. b) EBSD image of
a twinned grain.
3.1. Physical basis of the model 29
For the purpose of the modelling, we will combine all twin lamellae of
the same type into a - so called - twin system. In fcc metals, such twins
are grouped in bundles. Each bundle is composed of twins separated by
small fractions of untwinned regions. It must be noticed that twins or
twin bundles of a same system may have slightly dierent lattice orien-
tations if they did not appear at the same level of deformation. Indeed,
we have seen that the twin orientation is function of the orientation of
the parent grain and the latter evolves with the deformation. There is
therefore a question arising about the number of subgrains to be used
in order to represent the whole set of orientations created by the twins.
Onset of twinning : Twinning does not appear at the same level
of deformation in each material. In an austenitic stainless steel (AISI
316L), Chowdbury et al. [18] did not see twinning after a 20% reduction
by cold rolling but well after a 40% reduction and even more after a 60%
reduction. In the case of TWIP-steel (Fe-30Mn-3Al-3Si), Vercammen
et al. [19] observed the presence of twins in more than 50% of the
grains after an eective strain of 0.1 equivant to a 8% reduction by cold
rolling. After an eective strain of 0.21, equivalent to a 17% reduction,
they observed that nearly all grains contain twins. Asgari et al. [20]
observed deformation twins at a strain level of 0.09 after a simple
compression test on brass and MP35N. The latter is an alloy mainly
composed of nickel, cobalt, chromium and molybdenum. Karaman et al.
[21] observed twins on a micrograph of a [

111] oriented single crystal of


Hadeld steel after 3% uniaxial elongation. Allain et al. [22] observed
twins in a majority of grains after a tensile deformation of 18% on an
austenitic steel with high manganese content Fe 22Mn0.6C. In our
material, rst twins can be observed by EBSD after a deformation of
10% during uniaxial elongation. The prediction of the deformation level
of the twinning onset will help to validate the model. The quality of this
prediction is the rst factor inuencing the quality of the prediction of
the twin volume fraction.
Material properties that inuence twinning : It has been ob-
served that twin formation is highly dependent on the stacking fault
energy (SFE). The lower the SFE, the higher the twin volume fraction
in steel. Karaman [21] has reported evidence that the orientation of sin-
gle crystals and the yield strength have an inuence on the appearance
of twinning. He also observed that twinning does not appear when a
single crystal does not exhibit multiple slip even if its SFE is very low.
30 CHAPTER 3. Model description
El Danaf et al. [23] observed that the grain size has an inuence on
twinning. They compared the same composition for three grain sizes
and it appears that the twin volume fraction was close to zero when the
grain size was approximately 5nm and very important for a grain size
of 200nm. These material properties have to be used in our modelling
of the twin activation if we want to have the most accurate prediction
of the onset of twinning both macroscopically and in individual grains.
Number of twin systems activated per grain : Observations
made in UCL have indicated that one twin system is activated in the
majority of the grains and there is a maximum of two twin systems ac-
tivated per grain (Fig. 3.5). Similar observations are reported in the
literature [19, 20, 21, 22]. This second twin system is observed after
35% deformation under tension by Karaman [21], after 33% deforma-
tion under tension by Allain [22], after a true strain of 0.31 in simple
compression by Asgari [20], after a deformation of 0.21 after cold rolling
by Vercammen [19]. Our model must reproduce the fact that there is a
maximum of two twin systems activated per grain with a dominant one.
Slip-slip and slip-twin interaction : It is well known that interac-
tions among dislocations increase the resistance to dislocation motion
in a grain. The stress eld around a dislocation acts as an obstacle to
the motion of other dislocations. It is also observed that the grain size
inuences the hardening of the slip systems because grain boundaries
act as strong barriers to dislocation slip. Hence, the smaller the grain
size, the higher the slip resistance. Twin boundaries also act as strong
barriers to dislocation motion. This can be compared to a subdivision of
grain in smaller subgrains. Therefore, the higher the twin volume frac-
tion, the higher the slip resistance. This slip-twin interaction is mainly
anisotropic because only dislocations gliding on a slip plane which is not
coplanar to the twin plane activated have a reduction of their mean free
path (Fig. 3.6) [24, 25]. In order to predict accurately the hardening,
this slip-twin interaction must be implemented in our model.
Evolution of the twin volume fraction : In our material, twins
are so thin that they are dicult to observe separately. EBSD images
reveal reoriented lamellae which probably constitute twin bundles. We
have to look at more detailed images (TEM) in order to observe separate
3.1. Physical basis of the model 31
Figure 3.5: EBSD image of the microstructure of a TWIP steel.
twins. In this case, the number of grains covered is too small to obtain a
statistical measure of the twin volume fraction. Nevertheless, a careful
study of the twin volume fraction (TVF) has been done by Renard [26]
in our TWIP steel leading to a measure of the twin volume fraction at
32 CHAPTER 3. Model description
Figure 3.6: Representation of the slip-twin interaction. On the left, dis-
locations moving on slip system which are non-coplanar with the twin
system of twins present in the material have their mean free path re-
duced. On the right, dislocations moving on slip systems which are
coplanar with the twin system of twins present in the material have the
same mean free path as without twins.
necking after a tensile test. At the moment we have only two other
measures of the TVF taken after a true strain of 0.08 and 0.18 in a
tensile test. Twins are not seen before a true strain of 0.05 in the sample.
We dot not have measures of the TVF for a plane strain compression
experiment. Adler and al. [27] have measured the evolution of the
TVF in an Hadeld steel thanks to point counting on microstructures
obtained after a tensile test at 293K. They observed a TVF close to 25%
after a plastic strain of 0.3. Shuen and al. [28] have measured the TVF
after 40% strains in tension for two steels, a Fe-Mn-C and a Fe-Mn-Al-
C, for various temperature in the range going from 193 to 493K. They
observed a TVF decreasing with the increase of the temperature from
0.6 at 193K to 0.3 at 493K for the rst steel and from 0.45 to 0.15 for
the second steel. The decrease of TVF with the increasing temperature
is also observed on our steel. A summary of all TVF measurements is
shown in gure 3.7.
The two main observations about the evolution of the twin volume
fraction are that the twin volume fraction evolves continuously, Allain et
al. [22], but that its rate of formation tends to decrease after a threshold
deformation, e.g. a strain of 0.2 in simple compression in the case of
Asgari et al. [20]. Allain [29] computes a twin volume fraction f
i
for
a tensile test from measurement of the mean free path of dislocation d
i
and from the stereological relation of Fullman :
1
d
i
=
f
i
n
i
e
i
(1 F)
(3.1)
3.1. Physical basis of the model 33
Figure 3.7: Summary of the measurements of the twin volume fraction
on various FCC steels
where F is the total fraction of twins, n
i
the mean number of twins in a
bundle and e
i
the twin thickness. He obtains a twin volume fraction of
12% after a true strain of 0.4. We can see (Fig. 3.7) that the rate of twin
nucleation is decreasing after a true strain of 0.15. This decrease of the
rate of twin formation can be due to the decrease of the space available
for new twins due to previous twins. We should take this observation
into account in the modelling of twin-twin interactions. The quality of
the prediction of the decrease of the twinning rate is the second factor
inuencing the quality of the prediction of the twin volume fraction.
Shear banding : The twin-slip interaction seems to lead to the ap-
pearance of micro shear bands in steels which deform by slip and twin-
ning. Those are reported in several cases after large deformation. Ver-
cammen [19] observed that the deformation becomes inhomogeneous as
shear bands become visible at very high strain levels (1.05 true strain
after cold rolling). These bands are oriented at +35 and 35 to the
rolling direction. Chowdbury [18] observed shear bands after 90% re-
34 CHAPTER 3. Model description
duction by cold rolling. The length of the bands remains within the level
of grain scale in majority but can extend to other grains in a few cases.
El-Danaf [30] observed micro-shear bands after a true strain of 0.5 after
plane strain compression. The onset of these bands is correlated with
a decrease of the volume fraction of the Copper component and the S
component.
Copper-to-Brass texture transition : Twinning in TWIP steel re-
sults in the creation of a brass-type texture after rolling. The intensity
of the Brass orientation becomes much more important than the inten-
sity of the Copper orientation. This is partly due to the fact that the
twin orientation of a grain having a Copper orientation has the Goss
orientation. With further deformation, Goss orientation stabilises near
the Brass orientation. This can be observed through the various mea-
sures performed on our TWIP steel. Vercammen [19] observed that
the rolling texture of the TWIP steel is a Brass component 110 112)
with a spread towards the Goss component 110 001). As the rolling
strain increases, the intensity of the brass orientation increases gradu-
ally resulting at = 2.30 in a total volume fraction of about 35% (using
a spread of 16.5 around the ideal orientation). On the exact copper
position, the intensity slightly increases, reaching a maximum value at
= 0.43 and decreasing at higher strain levels : > 0.5. The prediction
of the brass-type texture in the case of steel which deforms by twinning
will be one of our major goals in this thesis.
3.2 Theory of crystal plasticity in metals de-
forming by dislocation slip
This section is essentially extracted from a published article [31]. Exper-
imental observations underlying the theory of crystal plasticity in metals
deforming solely by dislocation slip can be summarised as follows:
Dislocation slip is a stress-driven phenomenon which occurs along
well-dened crystallographic planes and directions.
As dislocations sweep through the crystal lattice, the material is
sheared, the volume remains constant, and the crystal lattice is
globally unaected.
3.2. Theory of crystal plasticity in metals deforming by dislocation slip 35
Lattice deformations are elastic and their amplitude is two orders
of magnitude lower than plastic strains.
In addition to elastic stretches, the crystal lattice rotates so as to
full local kinematical constraints.
These observations have convinced modellers to distinguish elastic and
plastic strains, based on a multiplicative decomposition of the deforma-
tion gradient tensor :
F = F

F
p
= R

U
el
F
p
. (3.2)
In this expression, F
p
transforms the initial conguration into a cti-
tious intermediate conguration corresponding to a crystal deformed by
dislocation slip only (Fig. 3.8). The subsequent transformation may in
turn be decomposed into a symmetric tensor U
el
, representing elastic
stretch, and an orthogonal tensor R

representing lattice rotation. The


latter rotation is a state variable of the model. It is important on the
one hand because it determines the macroscopic texture development;
on the other hand, because the stress tensor is attached to the crystal
lattice, and a proper account of the lattice rotation ensures objectivity
of the theory. As nicely explained by Aifantis [32] and explicited below,
the lattice rotation is constitutively prescribed in crystalline materials
whereas there is no consensus about a suitable denition of the material
spin in macroscopic theories.
The velocity gradient tensor is equal to :
L=

FF
1
=

R

R
T
+R

_

U
el
U
el
1
_
R
T
+R

U
el
_

F
p
F
p
1
_
U
el
1
R
T
(3.3)
As F
p
represents a purely plastic deformation, the elastic strain ten-
sor E may be calculated with regard to the intermediate rather than
the initial conguration :
E =
1
2
_
_
FF
p
1
_
T
FF
p
1
1
_
=
1
2
_
U
el
T
U
el
1
_
. (3.4)
This expression of the strain is Greens strain. It is work-conjugate to
the second Piola-Kirchho stress T that is related to the Cauchy stress
36 CHAPTER 3. Model description
Figure 3.8: Decomposition of a crystals deformation into the elastic and
the plastic components.
through:
T = det
_
FF
p
1
__
FF
p
1
_
1

_
FF
p
1
_
T
(3.5)
or
T = det
_
U
el
_
U
el
1
R

T
R

U
el
1
(3.6)
The fourth-order elasticity operator C sets the proportionality of T with
regard to E:
T = C : E. (3.7)
In all metals, thus including TWIP steels, elastic strains are innites-
imal. This means that we may write U
el
I + with symmetric and
such that || << 1. One then derives U
el
1
I , det
_
U
el
_
1 and
E . By introducing these simplications into Eq.3.3, one comes up
with the following equation.
L

R

R
T
+R

( +L
p
) R
T
(3.8)
where
L
p
=

F
p
F
p
1
=

, (3.9)
3.2. Theory of crystal plasticity in metals deforming by dislocation slip 37
and M

= n

is the Schmid tensor of the slip system labelled , n

is the slip plane normal and b

, the slip direction. The skew symmetric


part of 3.8 dictates lattice rotation, it may be written as :
W =

+W
p
(3.10)
where

is the lattice spin and W


p
is the plastic spin. Under a given
F, leading to a given velocity gradient L, one obtains for certain lattice
orientations W = W
p
and, hence,

= 0. These orientations are


called stable orientations. As discussed in Section 2.6, the increasing
proportions of stable orientations among all grains of a microstructure
explains the formation of a texture.
The crystal plasticity theory is completed by relating the slip rates

to the stress T. The rst link is made through the generalised Schmid
law: slip occurs when the resolved shear stress

reaches a threshold
value:

= T : M

c
. (3.11)
The second link is made by minimising the plastic work rate dissipated
internally by friction

W =

c
[

[ (3.12)
under the condition that the equation 3.9 is satised. If the strain rate
sensitivity of the material is considered negligible, it has been shown by
Van Houtte [15] that, in most cases, there is more than one solution to
the constrained minimisation, the problem is therefore undetermined.
This is called the Taylor ambiguity. In order to avoid this ambiguity, we
have adopted the viscoplastic expression introduced by Hutchinson [33]:

=
0

1/m
sign(

). (3.13)
In this expression, the reference slip rate
0
and the sensitivity expo-
nent m are material parameters. However, when m is very low (and
this the case in TWIP steel at room temperature), the solution is rather
insensitive to these parameters and the generalised Schmid law is closely
approximated. Strain hardening is introduced by evolving the critical
resolved shear stress (CRSS)

c
in function of the accumulated disloca-
tion slip. This is discussed later.
38 CHAPTER 3. Model description
3.3 Theory of crystal plasticity adapted to met-
als deforming by dislocation slip and by twin-
ning
The above part of the modelling has not implied personal choices. De-
formation by dislocation slip was already present in the model available
in UCL. My task was to consider twinning as a supplementary and uni-
directional mode of deformation. Several questions are raised at this
stage:
How does twinning participate to the plastic deformation?
What are the characteristics of newly formed twins?
How do we model slip system hardening?
How do we model twin system hardening?
How do we model the onset of twinning?
How do we take the twin-twin interaction into account?
All these questions have already been addressed by previous re-
searchers. A short litterature survey summarizes the state of the art.
It shows that some questions remain open and that there is no obvious
choice among existing models. The model that we have developed is
partly inspired from existing models and partly new. It is rather simpli-
ed in some regard. Indeed, the aim was to rely on as few parameters
as possible in order to study the eect of the interaction of adjacent
grains on the combined prediction of texture development, hardening
and twinning kinetics.
3.3.1 Generalised Schmid law extended to twinning
Like most of my predecessors [34, 35, 36], I have accounted for the fact
that twinning, like dislocation slip, induces simple shear along the twin
systems with the highest resolved shear stress. Therefore, the plastic
velocity gradient L
p
is computed as the sum of the shear rates on all
3.3. Theory of crystal plasticity adapted to metals deforming by
dislocation slip and by twinning 39
slip and twin systems :
L
p
=

M
s,

s,
+

M
t,

t,
. (3.14)
The Schmid tensors for the twin systems M
t,
are equal to the dyadic
product between b
t,
, the unit vector of the twin shear direction (pro-
portional to the associated Burgers vector) and n
t,
, the unit vector
normal to the twin plane. As crystallographic slip does not distort the
lattice, M
s,
and M
t,
have the same expression in the initial and in
the intermediate conguration (cf. Fig. 3.8).
s,
and
t,
are scalars
representing the rate of dislocation slip and the rate of twinning on the
slip system and the twin system . The viscoplastic expression due
to Hutchinson ([33], Eq. 3.13) is used in order to predict the activation
of the slip and twin systems. The volume fraction of twins created on
a twin system is proportional to the simple shear induced on this twin
system :

=

t,

0
(3.15)
where
0
, equal to
1

2
in FCC steel, is the discrete shear that undergoes
a part of the crystal which twins.
3.3.2 Characteristics of newly-formed twins
An originality of the present work with regard to state-of-the-art litera-
ture is to take into account the fact that the orientation of newly created
twins is dierent from the orientation of previously created twins within
the same twin system.
If we want to model separately each new orientation created, twin-
ning is likely to create an innite number of grains. In order to avoid
this, Van Houtte [34] proposed to replace the original orientation by
the twin orientation when the twin volume fraction becomes suciently
large. After each time step, the probability that the twin orientation is
introduced is equal to the twin volume fraction inside the grain. Another
solution proposed by Tome et al. [37] is to work with a large number of
orientations covering the whole Euler space. After each step, new unit
volumes created by twinning are transferred in the appropriate boxes
in Euler space. Such scheme is not applicable to homogenisation theories
involving neighbour interaction.
40 CHAPTER 3. Model description
Kalidindi [35] was the rst to consider the interaction of the parent
grain and the twins. He assumed that the parent-to-twin orientation
relation was preserved throughout the simulation. There is therefore one
orientation representing the whole set of twin orientations belonging to
the same twin system. This approximation does not have an important
impact on the prediction since twin orientations are in practice very
close to each other. In supplement, Kalidindi assumed iso-deformation
of the twinned and the non-twinned region. This assumption is very
strong since as the twin are very thin, it is expected that they are very
hard. However, Gil Sevillano [38] agreed with the iso-strain hypothesis,
considering that it is favoured by the lamellar structure of the twins and
helps explaining the huge Bauschinger eect in these alloys.
I have chosen to reproduce the Kalidindis procedure since this is
the only one that allows interactions between the twin regions and the
parent grain region. However, the lattice orientation of the twin system
is computed in a slightly dierent manner that will be presented in the
next section.
Regarding the interaction of the twin regions and the parent grain
region, I have tested dierent modelling assumptions and nally decided
to adopt the iso-strain assumption which is the simplest and gives rea-
sonable predictions.
3.3.3 Hardening of the slip systems
Many expressions have been proposed in the literature for the hardening
of the slip systems which is represented by the evolution of
s,
c
. Ka-
lidindi [39] modelled this interaction by relating the saturation value of
each slip system
s,
sat
to the volume fraction of twins created on non-
coplanar twin systems :

s,
sat
=
so
+
pr
_

knc
f
k
_
0.5
(3.16)
where nc is the set of non coplanar twin systems with ,
so
is the
saturation value in absence of deformation twinning and
pr
is a mate-
rial constants, which controls the contribution of deformation twinning
to the strain hardening response. This saturation value is used in the
3.3. Theory of crystal plasticity adapted to metals deforming by
dislocation slip and by twinning 41
evolution law of the critical resolved shear stress :

s,
c
= h
s
_
1

s,
c

s,
sat
_

kN
s

s,k
. (3.17)
h
s
represents the initial strain hardening rate and N
s
the set of slip
systems.
Staroselsky and Anand [36] introduced the slip-twin hardening di-
rectly in the evolution of the critical resolved shear stress :

s,
c
=

k
h
k

k
+h

sltw

tw
(3.18)
where
tw
is the shearing rate due to twinning on the major twin system,
h
k
is a latent hardening matrix and h

sltw
a slip-twin hardening. The
slip-twin hardening is governed by a simple saturation-type law :
h

sltw
= h
sltw
0
_
1

s,
c

sltw
_
a
sltw
(3.19)
where h
sltw
0
,
sltw
and a
sltw
are material constants.
Karaman [40] directly introduced the twin impact in the computation
of the evolution of the dislocation density :
=

k
_
1
b
_
K

0
d
1
+
K

0
d
2
_
+k
1

k
2

_
[
k
[ (3.20)
where d
1
is the distance between twins and d
2
the grain size, b is the
Burgers vector, k
1
and k
2
are constants and the two K
0
are geomet-
ric constants which represent the dierent strength of twin and grain
boundaries. The rst term represents an empirical geometric storage
term due to twin boundaries as being a barrier to dislocation motion.
The distance d
1
between neighbouring twins is evaluated directly from
Fullmans volumetric analysis (Eq. 3.1).
Allain [41] modelled this interaction by taking into account the fact
that the mean free path of dislocations,
1

i
, is reduced with the increase
of the twin volume fraction :
1

i
=
1
D
+
1
t
i
+k
_
A
ij

j
. (3.21)
42 CHAPTER 3. Model description
where D is the grain size and t
i
, the mean free path between microtwins
secant to the slip systems, is computed as
t
i
= B
ij
1
d
j
. (3.22)
B is a symmetric slip-twin matrix with B
ij
= 0 if the slip i and the twin
system j are coplanar and B
ij
= 1 if they are secant. d
j
, the distance
between coplanar microtwin stacks of a given system is computed using
a modied Fullmans stereological relationship (Eq. 3.1).
The hardening law of the CRSS for slip
s
c
of Allain ([41]) has been
implemented in order to evaluate the inuence of physical parameters
which are included in this expression. This has helped us in order to
design our expression of the hardening which is a function of the activity
on the slip and twin systems :

s,
c
=
s
c0
_
1 +

s,
tot

0
_
n
(3.23)
and

s,
tot

0
=
_
t
0
_
_
1

k

s,k
+
1

st

st

t,
_
_
dt (3.24)
where
st
is a constant which weights the inuence of the twin activity
and A
st
is a slip-twin interaction matrix. I only allow two combinations
of this matrix, one where all entries are equal to 1 and another one where
the entry (, ) is equal to a
st
if the slip system and the twin system
are coplanar and b
st
if they are non-coplanar where a
st
and b
st
are two
user dened values. I expect to use a very low value for
st
and a high
value for b
st
so that the twin activity on twin systems non-coplanar with
the slip system becomes the main factor of the slip systems hardening.
Our expression has the same goal as the four other modellings presented
below, which is mainly to take the anisotropic inuence of the twin on
the slip systems into account. The inuence of the twins is a reduction
of the mean free path of the dislocation. I have chosen this expression
because it is an enhancement of the expression used in the model without
twinning. The advantage compared to Kalidindis expression is that it
is easier to see the inuence of the parameters included in our twin-slip
interaction on the slip system hardening. Our expression is more generic
compared to Staroselky and Anands expression because they only have
interactions between slip systems and the major twin system activated
in their grain.
3.3. Theory of crystal plasticity adapted to metals deforming by
dislocation slip and by twinning 43
3.3.4 Evolution of the CRSS for twinning
There are few expressions which have been proposed in the literature for
the hardening of the twin systems. Staroselsky and Anand [36] proposed
a hardening law with a saturation of twin systems with a volume
fraction superior to 3%:

t,
c
= h
tw

t
(3.25)
where
t
is the shearing rate due to twinning on the major twin system
and
h
tw
= h
tw
0
_
1

t,
c

tw
_
a
tw
(3.26)
where h
tw
0
,
tw
and a
tw
are material constants. Other twin systems have
their CRSS for twinning arbitrarily increased to
t,
c
= 5

c0
. If none
of the twin systems has created a twin volume fraction superior to 3%,
they do not harden.
Kalidindi [39] proposed a hardening law for twin systems that
depends on weighted contribution of the activity on twin systems which
are coplanar with and of the activity of twin systems which are non-
coplanar with :

t,
c
= h
nc
_

k
f
k
_
b

knc

t,k
+h
cp
_

k
f
k
_

kcp

t,k
(3.27)
where nc is the set of twin systems non coplanar with , cp is the set
of twin systems coplanar with , and b, h
cp
and h
nc
are three material
constants to be determined.
We have implemented Kalidindis expression in order to design our
expression of the hardening of twin systems which have the same form
as the one for the hardening of slip systems :

t,
c
=
t
c0
_
1 +

t,
tot

0
_
n
. (3.28)
You can see that parameters n and
0
are used in both hardening laws of
slip and twin systems. The reason is that we want to limit the number of
parameters in the model. The value of

t,
tot

0
is one of the particularities
of our modelling. In order to model various slip-twin and twin-twin
44 CHAPTER 3. Model description
interactions, I choose a generalised law, which is function of the weighted
activity on the various slip and twin systems :

t,
tot

0
=
_
t
0
_
1

ts

ts

s,
+
1

tt

k
A
k
tt

t,k
_
dt. (3.29)

ts
and
tt
are two constants which weight the inuence of the slip and
twin activity. If the inuence of the slip activity on the hardening of the
twin systems is neglected,
1

ts
is put to 0. A
ts
is a twin-slip interaction
matrix. Typically, we will only use two combinations of this matrix, one
where all entries are equal to 1 and another one where the entry (, ) is
equal to a
ts
if the slip system and the twin system are coplanar and
b
ts
if they are non-coplanar where a
ts
and b
ts
are two user dened values.
A
tt
is a twin-twin interaction matrix. We only allow two combinations of
this matrix, one where all entries are equal to 1 and another one where
the entry (k, ) is equal to a
tt
if the twin system k and the twin system
are coplanar and b
tt
if they are non-coplanar where a
tt
and b
tt
are two
user dened values.
3.3.5 Onset of twinning
The appearance of the rst twins is mainly function of the value of
t
c0
.
We have seen in the previous chapter that the onset of twinning is a
function of the SFE, the grain size and other factors. Many authors
have presented their own expression for
t
c0
.
Allain [22] introduces a CRSS for twinning that corresponds to the
stress required for nucleation, which can be expressed as :

c0
=
SFE
b
+
loop
(3.30)
where
SFE
b
is the stress required to extend a stacking fault and
loop
the
critical shear stress to expand a dislocation loop.
Karaman and al. [42] compute the CRSS for twinning from a simple
force equilibrium on the Shockley partials. They take into account the
eect of the applied stress on the eective stacking fault energy and the
additional frictional stress on dislocation movement due to interstitial
3.3. Theory of crystal plasticity adapted to metals deforming by
dislocation slip and by twinning 45
atoms :

friction
(solute atoms) +(stacking fault)
+
c
(interaction w/forest dislocations)
=
t
c0
+(pile up)
F
max
b
2

+
9SFE
eff
4

3b
+

3Gb
3l
=
t
c0
+

2
2
_
L
r
c
_
2

s
c0
(3.31)
where F
max
is the maximum friction force for a dislocation loop, b
2
is
the burgers vector for the leading Shockley partial, is the distance
between solute atoms, SFE
eff
is the eective SFE, b the Burgers vector
of a perfect dislocation, l is the mean distance of forest dislocations,
is related to the stress necessary to cut through the forest dislocations,
G is the shear modulus, L the width of the pile-up, and r
c
is the critical
size of the twin nucleus. The eective SFE is given by the following
relation :
SFE
eff
= SFE
0
+
m
i
_
n
(2)
j
n
(1)
j
_
2

ij
b (3.32)
where SFE
0
is the initial SFE, m
i
is the slip plane normal, n
(2)
and n
(1)
are the direction of Burgers vector of the trailing and leading Shockley
partials, respectively, and
ij
is the applied stress.
Meyers and al. [43] have developed an analytical description of the
CRSS for twinning which takes into account the temperature, the grain
size, the stacking fault energy but not the eect of texture and stress
state :

T
= M
T

t
c0
(3.33)
where M
T
is the orientation factor and
T
the critical normal stress
dened as

T
= K
1/m+1
exp
_
Q
(m+ 1)RT
_
(3.34)
where
K = M
T
_
nlE
MA
0
_
(1/m+1)
. (3.35)
Q is the activation energy, T is the temperature, A
0
and m are constants,
l is the distance between source and barrier and n is the number of
dislocations which pile up in a time t. The inuence of the grain size is
obtained through the value of l which is function of the grain size.
As there is no consensus about the twinning activation, I did not
choose between these expressions. In our model, the value of
t
c0
is
46 CHAPTER 3. Model description
chosen by the user. He can either choose this value to be proportional
to
s
c0
or as the result of one of the expressions presented above. The
choice of the expression of
t
c0
is independent of the hardening law of the
CRSS for twinning.
3.3.6 Initial state of the newly created twin
The initial stress and hardening state of a new twin is something which
is seldom discussed in the literature. It appears that many models use
the state of the parent grain as the initial state of twin regions. Clausen
and Tome [44] recently proposed an alternative. They rst dened a
local coordinate system with the axis 1 along the shear direction of
the twin system and the 3 direction along the twin plane normal, i.e.,
with the 1 and 2 axes within the twin plane. They then enforced the
following continuity conditions across the twin plane for the elastic strain
components :

el,parent
11
=
el,twin
11
,
el,parent
22
=
el,twin
22
and
el,parent
12
=
el,twin
12
(3.36)
and for the stress components:

parent
33
=
twin
33
,
parent
23
=
twin
23
and
parent
13
=
twin
13
. (3.37)
They chose to assign the current plastic strain in the parent grain as the
initial plastic strain in the twin. The initial stress state in the twin is set
equal to the back stress and the initial elastic strain component is set
equal to the back strain. The back-stress and back-strain are the stress
and strain that appears in the grain in order to accommodate the plastic
shear strain of the twin nucleation. They estimated the back-stress by
assuming that the imposed plastic shear strain
0
results in an equal
and opposite elastic shear strain :

0,t
ij
= m
t
kl

0
(3.38)
where
0,t
ij
is the elastic back-strain in global coordinates, and m
t
kl
is
the Schmid tensor for the twin system. They computed the back-stress
through this simple relation :

0,t
ij
= C
twin
ijkl

0,t
kl
(3.39)
where C
twin
ijkl
is the elastic stiness tensor of the twin. For the parent
grain, the back-stress is introduced as a correction of the current stress
3.4. Numerical implementation 47
through the following relation :

parent
=

parent
+

0,t
f
0
1

t
f
0
(3.40)
where f
0
is the initial relative size of the twin.
3.4 Numerical implementation
All the predictions presented in this thesis have been obtained with the
same FORTRAN routine. The latter may thus be used in conjunc-
tion with a variety of modelling assumptions about the micro-to-macro
scale transition. The routine is versatile and it is programmed in mod-
ules. This allows to rigourously compare dierent assumptions about
the interaction of adjacent grains while using exactly the same crystal
plasticity theory at the local scale (the scale of slip systems and twin
systems described in Section 2.5), and vice-versa.
The general modelling framework is presented in gure 3.9. The
simulation may (but does not always) involve nite element modelling,
in which case the routine is implemented as a UMAT in the ABAQUS
nite element code. If we perform macroscopic FEM simulations, an
upper loop on the integration point of the mesh is added to gure 3.9.
If FEM simulations are performed on a RVE, the code is organised as
presented in gure 3.10. The core of the algorithm, which is represented
by call to BOX2 in both gures, may be presented as follows. It
implies Newton-Raphson solutions of non-linear equation sets at various
levels. Two of them are described in BOX1 and BOX 2, respectively.
Whatever the type of simulations, the following input must be pro-
vided by the user:
History of load applied to the polycrystalline aggregate which may
be controlled simultaneously by strain and by stress (e.g. trans-
verse directions in a uniaxial tensile test)
Discretised texture = set of lattice orientations
Modelling assumptions about next-neighbour interactions (Taylor,
LAMEL or ALAMEL)
48 CHAPTER 3. Model description
Material parameters describing the slip and twinning systems (incl.
hardening) as well as the elastic modulus (Table 3.1)
C
11
, C
12
and C
44
m,
0
,
0
, n

s,grain
c0
,
t,grain
c0
,
s,twin
c0

st
,
ts
,
tt
b
st
,b
ts
, b
tt
Table 3.1: Material parameters
Figure 3.9: Scheme of the code in the case of a simulation with Taylor
or ALAMEL model
3.4. Numerical implementation 49
Figure 3.10: Scheme of the code in the case of a CPFEM simulation
50 CHAPTER 3. Model description
BOX1 : Compute plastic velocity gradient (L
p
)
and twinning rates (

f

) based on
the current best estimate of the stress

T
t+dt
1. Compute resolved shear stresses (Eq. 3.11)
2. First estimate of the CRSS : set
c
t+dt
=
c
t
3. Compute slip and twinning rates using Eq. 3.13
4. Compute new estimate of CRSS :


c
t+dt
using Eq. 3.23 or
3.28
5. If implicit integration, check |
c


c
| < tol. Compute new
best estimates of CRSS,
c
, using Newton-Raphson and go
back to 2 until convergence
6. New estimate of L
p
and

f

using Eq. 3.14 and 3.15


7. New estimate of the lattice rotation R

t+dt
using Eq. 3.43
8. New estimate of the stress

T
t+dt
using Eq. 3.41
BOX2 : Compute stress within a grain, a grain region,
or a pair of adjacent grain regions (ALAMEL)
1. First estimate of the stress in the parent grain and in the
twins : set

T
t+dt
= T
t
2. If ALAMEL, rst estimate of the relaxation shears : set

rlx
t+dt
=
rlx
t
3. New estimate of stress in parent grain,

T
t+dt
, using BOX 1
4. New estimate of stress in existing twins,

t+dt
, using BOX
1 without (2nd generation) twinning
5. Compute volume average of the stress in the grain region
6. If ALAMEL, repeat 3-5 for the second grain region
7. Check whether |

T| < tol within all grain regions and


twins and, if ALAMEL, check whether Eq 2.14 is satised.
8. If not yet converged, perform corrections on

T
t+dt
and
rlx
t+dt
using Newton-Raphson and go back to 3
3.4. Numerical implementation 51
Based on such data, the computer simulation is run and the output
automatically comprises the following items:
History of stress throughout the deformation
Evolution of the state variables throughout the deformation
Statistics about the activity of slip and twinning systems (e.g.
Taylor factor)
Consistent tangent operator allowing rapid convergence of the -
nite element solution
3.4.1 Fully-implicit time integration of crystal plasticity
theory in absence of twinning
By combining Eqs. 3.4, 3.6, 3.7, 3.8, 3.9, and by relying on fully-implicit
time integration, one obtains a system of six non-linear equations allow-
ing to compute the components of the Second Piola-Kirchho stress
tensor T[
t+t
in function of the applied strain increment. Denoting

D
and D
p
the symmetric parts of

L and L
p
, respectively, we obtain:
T[
t+t
T[
t
= C :
_
R
T
[
t+t

DR

[
t+t
D
p
[
t+t
_
t (3.41)
The lattice spin is computed at every iteration of the solution based
on the skew-symmetric part of Eq. 3.8. Denoting

W and W
p
the skew-
symmetric parts of

L and L
p
, respectively, we have:

=

R

R
T
[
t+t
= R
T
[
t+t

WR

[
t+t
W
p
[
t+t
(3.42)
The lattice rotation R

[
t+t
is updated based on the lattice spin and the
exponential map [45]. As the strain increments are small and R

[
t+t
defers from R

[
t
only by a few degrees (cf. Section 2.4), it is convenient,
from a numerical viewpoint to simplify Eq. 3.41 by approximating:
R
T
[
t+t

DR

[
t+t
= R
T
[
t

DR

[
t
+
_


D
_
t (3.43)
After that the iterative solution has converged, the Cauchy stress is
computed as:
[
t+t
= R

[
t+t
T[
t+t
R
T
[
t+t
(3.44)
52 CHAPTER 3. Model description
and a consistent tangent operator, useful e.g. in nite element simula-
tions, may be derived as explained elsewhere [46].
In the multisite model (Section 2.8), the deformation applied to each
subregion is a priori unknown. It is dependent on the response of the two
crystals. The system of equations to be solved then contains a total of 14
equations, six equations prescribing the achievement of the deformation
for each grain (cf. Eq. 3.9) and the last two equations ensuring that
the relaxation relative to the macroscopically prescribed deformation
minimises the deformation energy of the two grains (cf. Eq. 2.14).
3.4.2 Partly-explicit time integration of crystal plasticity
theory in presence of twinning
The extension of the crystal plasticity theory to treat metals deform-
ing by twinning has been presented in Section 3.3. The FORTRAN
routine has been modied accordingly. For the sake of simplicity, time-
integration of the constitutive law has become partly explicit. Indeed, as
demonstrated in the next subsection, for the various types of simulations
performed in this thesis, the dierence relative to a fully-implicit time
integration is minor. Full scale simulations of forming processes, if they
are undertaken in the future, would probably justify that the routine is
adapted to allow fully-implicit time integration.
My numerical implementation of the twinning model relies on the
following strategy.
Supplementary state variables : The size of the array containing
state variables has been increased. Next to the lattice orientation, the
stress tensor and the hardening state

c
of the parent grain, we have
foreseen storage of the lattice orientation, the stress tensor and the hard-
ening state of each twin system that is activated. Based on experimen-
tal observations, we allow a maximum of four simultaneously activated
twinning systems (a maximum of two is observed in practice) and we
forbid second generation twinning, i.e. twinning within twin subre-
gions. It will be shown that proper tuning of the material parameters
leads indeed to the activation of two twinning systems and not more.
3.4. Numerical implementation 53
Volume transfer : The array of state variables also contains the local
volume fraction of each twin system. This variable is updated explicitly,
at the end of each increment. The volume fraction is updated by adding
the volume fraction of the parent grain which has twinned on the stored
twin system times the volume fraction of the parent grain
f
t
(t +t) = f
t
(t) +

f
t,
f
0
(t)t (3.45)
where f
0
(t) is the volume fraction of the parent grain at the end of the
previous increment and

f
t,
is the volume fraction of twins created on
the twin system .

f
t,
is linked to the shear rate on this twin system

t,
through the equation 3.15 The value of f
0
(t +t) is then computed
so that f
0
+

t
f
t
= 1.
Stress update : As we do not aim to model non-monotonic or cyclic
loading in this thesis, we do not account for the probable backstress due
to the twin formation. Hence, the initial stress state of the new twin is
equal to the stress state of the untwinned part of the grain. The stress
tensor assigned to the twin system is computed as the volume average
of the previous stress of this system and the stress of the newly-formed
twin. This allows avoiding that explicit update of the twin volume leads
to stress discontinuities. The stress of a grain is computed as the volu-
metric average of the stress of the dierent subregion of the grain.
Lattice orientation update : The twin orientation is also modied
at the end of each time increment. The quaternion of a twin orientation
is computed by multiplying the quaternion of the orientation of the
untwinned part of the grain with the quaternion corresponding to a
rotation of 180 around the 111 plane of the twin systems. Due to the
lattice rotation during plastic deformation, each new volume fraction of
twin has a dierent orientation. Therefore, the twin system orientation,
stored at the end of each increment, is computed as the volumetric
average of the previous twin system orientation and the orientation of
the newly created twin.
Strength of the twin : The initial value of the critical resolved shear
stress for dislocation slip within the twin is prescribed in the user ma-
terial le. Once a twin system becomes active, this value is raised with
further plastic deformation inside the twins.
54 CHAPTER 3. Model description
3.4.3 Explicit versus implicit integration of the slip sys-
tem hardening
In the scope of this thesis, variables are determined by explicit in-
tegration of Eq. 3.13 whereas this was done by implicit integration in
the original model. The value of
c
is only updated at the end of the
increment. Therefore, during a time increment

c

is equal to 0. This
has been done, in rst approximation, in order to bypass the more com-
plicated expression of

c

due to the interaction between the activity on
the slip and twin systems. Prior to this choice, we have checked that the
predictions obtained for steels which deform by dislocation slip only are
very close if variables are determined by implicit or explicit integra-
tion. Dislocations move along 111 < 100 > slip systems. Value of the
parameters of Eq. 3.23 are presented in table 3.2. The results of this
verication are presented below.
Elastics constants [GPa] Hardening parameters
C11 C12 C44
s
c0

0
n
231.4 134 48.7 120MPa 0.2 0.6
Table 3.2: Components of the anisotropic elastic modulus and parame-
ters of the hardening law
In the case of a tensile test with a true strain of = 0.3, we have
compared the prediction of the Taylor model for the two types of inte-
gration and for two numbers of increments. When 30 steps are used to
perform the simulation, the two predictions are close to each other, the
dierence is less than 1% (Table 3.3). As expected, the simulation is
quicker with the explicit integration : 91 seconds compared to the 133
seconds with implicit integration. When 200 steps are used to perform
the simulation, the two predictions are very close to each other, the dif-
ference is less than 0.1% (Table 3.4). The explicit simulation took 457
seconds and the implicit one 706 seconds.
3.4. Numerical implementation 55

11

12

13

23
Implicit 0.82E+03 -0.70E+01 0.80E+01 0.58E+00
Explicit 0.81E+03 -0.70E+01 0.79E+01 0.58E+00
Table 3.3: Value of the stress components after a tensile test with a true
strain of
11
= 0.3. 30 steps are used in the simulation. The steel is
deforming by dislocation slip only.

11

12

13

23
Implicit 0.80E+03 -0.74E+01 0.75E+01 0.52E+00
Explicit 0.80E+03 -0.74E+01 0.75E+01 0.52E+00
Table 3.4: Value of the stress components after a tensile test with a
true strain
11
= 0.3. 200 steps are used in the simulation. The steel is
deforming by dislocation slip only.
56 CHAPTER 3. Model description
Chapter 4
Assessment of the model
The general objective of the crystal plasticity modelling of TWIP steel
is to simulate various loading modes and obtain, in each case and for a
xed value of all material parameters, a valid combined prediction of:
the crystallographic texture evolution,
the rate of twinning,
the strain hardening,
the induced anisotropy.
The results presented in this chapter demonstrate that even the sole
prediction of texture evolution is not an easy task. This, in spite of
the fact that texture predictions are inuenced by (i) the rate of twin-
ning, (ii) the assumptions formulated about grain interaction, and the
hardening resulting from (iii) twin-slip and (iv) twin-twin interaction.
The latter four trends are presented in the next sections whereas the
conditions ensuring the best overall predictions are discussed in section
4.5 For the sake of conciseness, the acronym TVF will be used for twin
volume fraction in this chapter. The gure representing skeleton lines
of the ber of the texture measured after large deformation by cold
rolling will show that that the position of Copper does not correspond
to 112 < 111 > (the exact Copper) but is rather in between Copper
and Cube (001 < 100 >). This position is closer to the Cube com-
ponent. This is due to the fact that the Copper component has partly
57
58 CHAPTER 4. Assessment of the model
disappeared in this texture. Nevertheless, in this chapter we made a
simplication in our language by considering that we are dealing with
the Copper component.
Available experimental data: Some samples have been cold rolled
and others have undergone uniaxial tensile tests. The latter were car-
ried out up to the onset of diuse necking (true strain of 0.3). The
cold rolled samples were investigated along their mid-thickness section
where plane strain deformation applies. Two thickness reductions were
considered corresponding to EFFECTIVE strains,
eq
, of 0.3 and 0.9, re-
spectively. Textures measured at the end of the experiment are present
in gure 4.1 and 4.2. Stress-strain curve and incremental hardening
Figure 4.1: Experimental textures measured after a tensile test. Contour
lines : 1-1.4-1.7-2.1-2.4.
PAGE
0 90
PHI2= 45
(a) Rolling (
eq
= 0.3)
PAGE
0 90
PHI2= 45
(b) Rolling (
eq
= 0.9)
Figure 4.2: Experimental textures measured after plane strain compres-
sion. ODF contour lines : 1-2-4-8.
4.1. Inuence of twinning on texture 59
curve measured at the end of the tensile test are presented in gure 4.3.
Throughout the chapter, two initial textures are used, one measured
before the tensile test and one before cold rolling reduction. Although
the alloy preparation was very similar, the samples subjected to cold
rolling and uniaxial tension, respectively, did not have exactly the same
initial texture. The two experimental textures have been discretised in
10000 equally weighted orientations. The rst one (Fig. 4.4a) is used
only for simulation of a tensile test and the second one (Fig. 4.4b) for
the simulation of plane strain compression.
Main modelling assumptions: Dislocations move along 111 <
100 > slip systems in twin and parent grain regions and twinning arises
on 111 < 211 > twin systems in the parent grain region. The following
parameters of slip hardening (Eq. 3.23 and 3.24) remain constant in
all simulations :
s
c0
,
0
and n. Value used for these parameters are
presented on table 4.1. The only parameter of equations 3.28 and 3.29
(resistance against twinning) that remains constant is n and it has the
value 0.6 as for the slip systems equation. I have chosen a value of
s
c0
for the slip systems in twin regions which is ve times larger than in
parent grain regions in order to model the fact that twin regions are
harder than the parent grain.
Elastics constants [GPa] Hardening parameters
C11 C12 C44
s
c0

0
n
Untwinned region 231.4 134 48.7 120MPa 0.2 0.6
Twinned region 231.4 134 48.7 650MPa 0.2 0.6
Table 4.1: Components of the anisotropic elastic modulus and parame-
ters of the hardening law
4.1 Inuence of twinning on texture
In this section, we evaluate the inuence of twinning on the texture pre-
dictions when the crystal hardens isotropically (no consideration of the
slip-twin and twin-twin interaction) and when the polycrystalline aggre-
gate deforms homogeneously (Taylor model). This preliminary study
should help identifying later the exact inuence of the other parameters
of the modelling.
60 CHAPTER 4. Assessment of the model
(a) Hardening
(b) Incremental hardening
Figure 4.3: Hardening curves measured after a tensile test (
eq
= 0.3).
4.1. Inuence of twinning on texture 61
(a)
PAGE
0 90
PHI2= 45
(b)
Figure 4.4: a) Initial texture used for tensile test simulation. b) Initial
texture used for simple place compression simulation. These experimen-
tal textures have been discretised in 10000 equally weighted orientations.
Pole gure contour lines : 1-1.4-1.7-2.1-2.4. ODF contour lines : 1-2-4-8.
At this stage, all parameters of the anisotropic hardening, i.e. param-
eters of equations 3.24 and 3.29, are chosen such that the hardening is
isotropic, i.e.
st
=
ts
=
tt
=
0
and A
st
= A
ts
= A
tt
= E, where E is
a matrix where each entry is equal to 1. Therefore,

s,
tot

0
=

t,
tot

0
=

tot

0
for
all slip systems and all twin systems (Eq. 3.23 and 3.28). However,
the value of the CRSS is dierent for dislocation slip and for twinning.
Hardening parameters are chosen to ensure a constant ratio:
t
c0
=
s
c0
.
Therefore, Eq. 3.23 and Eq. 3.28 becomes respectively :

s,
c
=
s
c0
_
1 +

tot

0
_
n
(4.1)

t,
c
=
s
c0
_
1 +

tot

0
_
n
(4.2)
62 CHAPTER 4. Assessment of the model
where the expression of

tot

0
is given by

tot

0
=
1

0
_
t
0
_
_

k

s,k
+


t,
_
_
dt. (4.3)
Predictions of textures obtained with various values of the parame-
ter are presented for the three mechanical tests after which experimental
textures are available :
Tensile test, true strain of 0.3 : A characteristic of the experimental
pole gure (Fig. 4.5a) is that there is an equal volume fraction of grains
having respectively their < 111 > or < 100 > axis aligned with the
tensile direction. Let us call R
T
the ratio between the volume fraction
of grains having their < 111 > axis aligned with the tensile direction and
the volume fraction of grains having their < 100 > axis aligned with the
tensile direction. Comparisons between the experimental and predicted
textures are based on the value of R
T
. When 1.21, we do not predict
any twinning (Table 4.2) and the ratio R
T
is much higher than 1 (Fig.
4.5b). When we decrease the value of , we increase the TVF. This
increase of the TVF induces a decrease of R
T
. We nd a value of R
T
close to 1 when = 1.17. In this case, a TVF of approximately 8% is
predicted.
Exp 1.21 = 1.17 = 1.14
Tensile test (
eq
= 0.3) 18 0 8 30
Rolling (
eq
= 0.3) ? 0 3.5 15
Rolling (
eq
= 0.9) ? 0 11 40
Table 4.2: Twin volume fraction (%) predicted for various values of ,
the ratio between the CRSS for twinning and the CRSS for slip.
Plane strain compression (
eq
= 0.3) : We can see on the
2
= 45
section of the experimental ODF (Fig. 4.6a) that the intensity of the
Brass component is larger than the intensity of the Copper component
already after such a small deformation. Let us call R
L
the ratio between
the ODF intensity of the Brass component and the ODF intensity of the
Copper component. In the experimental texture, its value is equal to
2.0 (Table 4.3). When 1.21, we do not predict any twinning (Table
4.1. Inuence of twinning on texture 63
(a) Experiment
(b) 1.21
(c) = 1.17
(d) = 1.14
Figure 4.5: 111) and 100) pole gures obtained after tensile test for a
true strain of 0.3 and for various values of , the ratio between the CRSS
for twinning and the CRSS for slip. Contour lines : 1-1.4-1.7-2.1-2.4.
64 CHAPTER 4. Assessment of the model
4.2) but the texture is close to the experimental one (Fig. 4.6b). The
value of R
L
is equal to 1.8. When the value of decreases, the TVF
increases. This increase of the TVF leads to an increase of the ratio R
L
.
We nd a R
L
value close to the experimental value when = 1.17. In
this case, a TVF of 3.6% is predicted.
PAGE
0 90
PHI2= 45
(a) Experiment
PAGE
0 90
PHI2= 45
(b) 1.21
PAGE
0 90
PHI2= 45
(c) = 1.17
PAGE
0 90
PHI2= 45
(d) = 1.14
Figure 4.6:
2
= 45 ODF section after plane strain compression (
eq
=
0.3) for various values of , the ratio between the CRSS for twinning
and the CRSS for slip. Contour lines : 1-2-4-8.
Exp 1.21 = 1.17 = 1.14
Rolling (
eq
= 0.3) 2.0 1.8 2.1 4.0
Rolling (
eq
= 0.9) 4.7 0.7 1.0 3.5
Table 4.3: R
L
ratio between the ODF intensity of the Brass and the
Copper component after plane strain compression with various values
of , the ratio between the CRSS for twinning and the CRSS for slip.
Contour lines : 1-2-4-8-12.
4.1. Inuence of twinning on texture 65
Plane strain compression (
eq
= 0.9) : We can see on the
2
= 45
section of the experimental ODF (Fig. 4.7a) that the Copper compo-
nent has nearly completely disappeared whereas the Brass component
is very dominant. The Goss component is also strongly present. The
spread around these two orientations is very important compared with
the texture measured after small deformation. The value of R
L
is equal
to 4.7 (Table 4.3). When 1.21, we do not predict any twinning
(Table 4.2) and as already known, the texture predicted is far from the
experimental one (Fig. 4.7b). The value of R
L
is equal to 0.7. When the
value of decreases, the TVF increases. This increase of the TVF leads
to an increase of the ratio R
L
. We nd a value of R
L
equal to 3.5 when
= 1.14. This value is lower than the experimental one but the inten-
sity of the Brass orientation is much higher than in the experimental
texture. The predicted TVF is equal to 40%.
PAGE
0 90
PHI2= 45
(a) Experiment
PAGE
0 90
PHI2= 45
(b) 1.21
PAGE
0 90
PHI2= 45
(c) = 1.17
PAGE
0 90
PHI2= 45
(d) = 1.14
Figure 4.7:
2
= 45 ODF section after plane strain compression (
eq
=
0.9) for various values of , the ratio between the CRSS for twinning
and the CRSS for slip. Contour lines : 1-1.4-1.7-2.1-2.4.
66 CHAPTER 4. Assessment of the model
Referring to literature (e.g. [47]), the preliminary study tends to
demonstrate that twinning has correctly been incorporated in the model.
Our predictions captured the main well-known trends about the inu-
ence of twinning on texture predictions. (Many other sanity checks have
of course been carried out.) In supplement, it appears that the threshold
value of giving rise to twinning is roughly identical in tension and in
rolling. If there is no twinning in tension for a given value of , there is
no twinning in rolling for the same value of . However, the rate of twin
formation is higher in tension than in plane strain compression. This
leads to a higher twin volume fraction in tension after a given eective
strain. This observation was also made experimentally [48].
Even with such simplied modelling, predictions are fairly close to
the experimental texture after small deformation. However, after a
larger deformation, none of the values of leads to valid texture pre-
diction. Either the intensity of the copper orientation is too high or the
intensity of the brass orientation is too high. Hopefully (e.g. [47]), the
use of a grain interaction model and of anisotropic hardening improves
texture predictions.
4.2 Inuence of grain interaction on texture
Three models of grain interactions are compared in this thesis : Tay-
lor, ALAMEL and FE. Similar to the previous section, the comparison
is done with isotropic hardening of the crystal. This reveals the eect
of grain interaction separately from the potential, supplementary eect
of anisotropic hardening. For each mode of deformation and for each
model, texture predictions are presented with and without twinning.
Exceptionally, the polycrystalline aggregate is composed of only 1458
grains in order to have a relatively cheap FE simulation. In this com-
parison, the FE mesh represents an idealised polycrystal wherein grains
are shaped as truncated octahedrons (Fig. 4.8). Each grain is thus dis-
cretised into 16 rst-order hexahedral nite elements. The same initial
texture is used for the three models of grain interactions.
Tensile test, true strain of 0.3 : The three models are compared
for a same value of = 1.8. We can see on the gure 4.9 that tex-
ture predictions are quite dierent with and without twinning. Without
4.2. Inuence of grain interaction on texture 67
Figure 4.8: Idealised microstructure wherein grains are shaped as trun-
cated octahedrons.
twinning, the model predicts the same R
T
value when associated to the
Taylor and ALAMEL scale transitions but a very slightly lower value
during the FE simulation. The volume fraction of grains having their
< 111 > axis aligned with the tensile direction is more important with
the Taylor and ALAMEL model than with FE. The texture obtained
with the Taylor model is the sharpest and the texture obtained with FE
element is the softest. But texture predictions are quite similar when
twinning is accounted for. The value of R
T
is approximately equal to 1
for the FE simulation and slightly larger for the two other modellings.
The volume fraction of grains having their < 100 > axis aligned with the
tensile direction is the same for the three models. None of the textures
is softer or sharper. These predictions are fairly close to the experi-
mental texture. They only overestimate the volume fraction of grains
having their < 111 > or < 100 > axis aligned with the tensile direction.
The main dierence between predictions is the TVF (Table 4.4). The
twin volume fraction decreases when strain heterogeneity increases in
the model of grain interactions.
Exp Taylor ALAMEL FE
Tensile test (
eq
= 0.3) 18 4.3 3.1 3.0
Rolling (
eq
= 0.3) ? 3.5 2.7 2.7
Rolling (
eq
= 0.9) ? 40 36 29
Table 4.4: Twin volume fraction (%) predicted for various models of
grain interactions.
68 CHAPTER 4. Assessment of the model
(a) Exp
(b) Taylor without twinning (c) Taylor with twinning
(d) ALAMEL without twinning (e) ALAMEL with twinning
(f) FE without twinning (g) FE with twinning
Figure 4.9: 111) and 100) pole gures obtained after a tensile test
for a true strain of 0.3. Simulations are performed with three dierent
models of grain interactions with and without twinning. Contour lines
: 1-1.4-1.7-2.1-2.4.
4.2. Inuence of grain interaction on texture 69
Plane strain compression (
eq
= 0.3) : Figure 4.10 shows texture
predictions obtained with the value of = 1.17 for all models. When
simulations are performed without twinning, predictions are quite close
to each other (Table 4.5). Predictions obtained with the Taylor model
Exp Taylor ALAMEL FE

eq
= 0.3 w/o twinning 1.7 1.6 2.1

eq
= 0.3 with twinning 2.0 2.0 2.2 2.4

eq
= 0.9 w/o twinning 0.7 0.7 0.9

eq
= 0.9 with twinning 4.7 3.3 2.4 3.2
Table 4.5: R
L
ratio between the ODF intensity of the Brass and the
Copper component after plane strain compression. Simulations are per-
formed with three dierent models of grain interaction with and without
twinning.
have the highest intensity of the copper component. The FE simula-
tion leads to the lowest intensity of the copper component and to two
peaks next to the brass component. The lowest intensity of the brass
component is obtained with the ALAMEL simulation. When simula-
tions are performed with twinning, predictions remain quite close to
each other. Nevertheless, the FE simulation is the only one leading to a
spread around the brass component which is close to the experimental
texture. For each one of the micro-macro models, the intensity of the
copper component decreases while the intensity of the brass component
increases. Texture predictions are closer to the experimental texture
when we take account of twinning. On table 4.4, we can see that these
texture predictions are obtained for a very low volume fraction of twins.
Plane strain compression (
eq
= 0.9) : On gure 4.11, we can see
texture predictions obtained with the value of = 1.14 for all models.
When simulations are performed without twinning, predictions are not
close to each other. Observations made for the true strain of 0.3 can
also be made here with higher dierences between the intensities of the
Copper component. When simulations are performed with twinning,
there is practically no more dierence in the intensity of the Copper
component. But, the intensity of the brass component is now dierent.
The highest intensity is obtained with the Taylor model. For a same
parameterisation, the FE simulation predicts a much lower TVF when
the value of R
L
is equivalent to the Taylor model prediction(Table 4.5).
In the case of the ALAMEL model, the TVF and the value of R
L
is
70 CHAPTER 4. Assessment of the model
PAGE
0 90
PHI2= 45
(a) Experiment
PAGE
0 90
PHI2= 45
(b) Taylor without
twinning
PAGE
0 90
PHI2= 45
(c) Taylor with twin-
ning
PAGE
0 90
PHI2= 45
(d) ALAMEL with-
out twinning
PAGE
0 90
PHI2= 45
(e) ALAMEL with
twinning
PAGE
0 90
PHI2= 45
(f) FE without twin-
ning
PAGE
0 90
PHI2= 45
(g) FE with twinning
Figure 4.10:
2
= 45 ODF section after plane strain compression
(
eq
= 0.3). Simulations are performed with three dierent models of
grain interactions with and without twinning. Contour lines : 1-2-4-8.
4.2. Inuence of grain interaction on texture 71
PAGE
0 90
PHI2= 45
(a) Experiment
PAGE
0 90
PHI2= 45
(b) Taylor without
twinning
PAGE
0 90
PHI2= 45
(c) Taylor with twin-
ning
PAGE
0 90
PHI2= 45
(d) ALAMEL with-
out twinning
PAGE
0 90
PHI2= 45
(e) ALAMEL with
twinning
PAGE
0 90
PHI2= 45
(f) FE without twin-
ning
PAGE
0 90
PHI2= 45
(g) FE with twinning
Figure 4.11:
2
= 45 ODF section after plane strain compression
(
eq
= 0.9). Simulations are performed with three dierent models of
grain interactions with and without twinning. Contour lines : 1-2-4-8.
72 CHAPTER 4. Assessment of the model
lower.
In order to highlight the dierence between the models, I have per-
formed new simulations with the Taylor and the Alamel model leading
to a prediction of the TVF which is equal to the one obtained with the
FE simulation. This implies using a dierent value in each model. We
can see that, in this case, the ALAMEL model predicts approximately
the same value of R
L
as the Taylor model prediction but for a softer tex-
ture (Table 4.6). If we compare predictions based on the value of R
L
,
these two predictions are far from the FE prediction. But all models
predict at the same time too large intensities of both the Copper and
the Brass components. None of the models predicts the disappearance
of the Copper component and the wide spread around the Brass and the
Goss components.
Exp Taylor ALAMEL FE
Brass 6.8 12.2 9.2 11.3
Copper 1.4 5.9 5.2 3.5
R
L
4.7 2.06 1.8 3.2
Table 4.6: Strength of the Brass and Copper texture components after
plane strain compression (
eq
= 0.9). Simulations are performed with
dierent grain interaction models with and without twinning.
In conclusion : These comparisons show that the modelling of grain
interactions does not signicantly improve texture prediction for small
deformation. The major factors that inuence texture predictions for
small deformation are the initial texture and a qualitatively good pre-
diction of the volume fraction of twinning. But, modelling of grain in-
teractions leads to predictions of the TVF which are lower than without
grain interactions. The inuence of grain interaction on the texture pre-
diction appears when simulations of larger deformation are performed.
The use of FE more signicantly improves predictions with twinning
than without twinning. The ALAMEL model also improves predictions
compared to the Taylor model but improvements are weaker. The fact
that the ALAMEL model predictions are not so close to the FE pre-
dictions can be explained by the fact that FE allows grain interactions
of all types whereas ALAMEL allows grain interactions only according
to two relaxation modes. The two relaxations allowed by the ALAMEL
model seem no longer sucient in order to reproduce the local strain
heterogeneity. Nevertheless, FE tends to underestimate the amplitude
4.3. Eect of anisotropic hardening on texture 73
of strain heterogeneity when the mesh is insuciently rened whereas
no such constraints applies to the ALAMEL model.
4.3 Eect of anisotropic hardening on texture
Until now, we have considered isotropic hardening of individual crystals.
The CRSS for dislocation slip,
,s
c
, was identical on all slip systems
and the CRSS for twinning,
,t
c
, was identical on all twin systems .
Moreover, the ratio between
t
c
and
s
c
was kept constant throughout
the simulation. In reality, the hardening of individual crystals is cer-
tainly anisotropic, i.e. if a yield locus was computed for a plastically
deforming crystal, not only the size of the yield locus but also its shape
should evolve with plastic strain. We will see that anisotropic harden-
ing of individual crystals cannot be avoided when it comes to predicting
macroscopic hardening properly (Section 4.4). Hopefully, the texture
predictions too can be improved by accounting for the various origins of
such anisotropic hardening. All simulations carried out in this section
rely on the ALAMEL modelling of grain interaction which seems to oer
a fair compromise between accuracy and computational cost.
4.3.1 Dierence between the twinning and hardening rates
There are various hints in experimental data indicating that the rate
of twin formation is dierent from the hardening rate and that the
ratio should not necessarily be kept constant. For example, twinning
seems to be absent from the rst percents of plastic deformation during
which the hardening rate is the largest. In the simulations presented so
far, twinning occurs either at the onset of plastic deformation or never,
depending on the constant value. After large strains, the saturation
of the twin volume fraction is explained more easily by twin-twin inter-
action than by slip-twin interaction. Indeed, we have seen in Section
3.1 that above a threshold value of the TVF, the rate of twinning de-
creases. One explanation is that previously created twins reduce the
space available for the creation of new twins.
An advanced modelling of the evolution of
s,
c
and
t,
c
is presented
in Sections 3.3.3 and 3.3.4. However, at this stage, we will not use the
model in its most general form but rather consider a unique
s
c
for all slip
74 CHAPTER 4. Assessment of the model
systems and a unique
t,
c
for all twin systems. The main dierence with
previous sections is that the ratio is varied. The shape of the single
crystal yield locus is therefore evolving, reecting anisotropic hardening.
We further consider that the evolution of
t
c
is governed solely by the
TVF. In Eq. 3.29, we set
1

ts
= 0 or very close to 0 and we t
t
c0
and
1

tt
in order to reproduce the experimental evolution of the TVF during the
uniaxial tensile test. According to the latter data (Fig. 3.7), the rst
twins appear between 5 and 10% elongation whereas the TVF 33% at
the onset of necking (after 45% elongation).
Three simulations of the uniaxial tensile test have been carried out.
The rst one, already presented in the previous section, relies on isotropic
hardening (constant ratio=1.17). The second and third simulations
assume
1

ts
= 0 and the
t
c0
and
1

tt
values are such that twinning is
triggered after, respectively, 5% and 10% elongation. The value of
t
c0
is
respectively equal to 190 and 230MPa while the value of

0

tt
is respec-
tively equal to 5 and 2.5. Texture predictions are compared to experi-
mental data in Fig. 4.12, revealing that the various parameterisations
do not inuence texture predictions so much. All of them lead to a value
of R
T
1. Hence, the texture predicted after uniaxial tension depends
on the nal TVF (see Section 4.1) but not on the rate at which twins
are formed. In the present simulations, the TVF 8% at the end of all
simulations of uniaxial tension (Table 4.7) whereas the evolution of the
TVF with plastic strain is very dierent (Fig. 4.13). As expected, in
the case of isotropic hardening (constant ratio), twinning begins at the
start of the plastic deformation and there is no reduction of the rate of
twin formation after large strains. The parametrisation leading to the
onset of twinning after 10% elongation is used in subsequent sections
because it seems most realistic.
Exp 0% 5% 10%
Tensile test (
eq
= 0.3) 18 7.9 8.0 8.1
Rolling (
eq
= 0.3) ? 3.6 6.4 6.0
Rolling (
eq
= 0.9) ? 11 22 26.5
Table 4.7: Twin volume fraction (%) predicted for various parameteri-
sations leading to the appearance of the rst twins after 0, 5 and 10%
elongation.
The same three parametrisations have been used in plane strain com-
pression simulations. We can see that for an eective strain of 0.3 (Fig.
4.14), all parameterisations lead to texture predictions which are close
4.3. Eect of anisotropic hardening on texture 75
(a) Exp
(b) Constant ratio
(c) Twinning after 5%
(d) Twinning after 10%
Figure 4.12: 111) and 100) pole gures obtained after tensile test for
a true strain of 0.3. Simulations are performed with parameterisations
leading to the appearance of the rst twins after 0, 5 and 10% elongation.
Contour lines : 1-1.4-1.7-2.1-2.4.
76 CHAPTER 4. Assessment of the model
(a) TVF
(b) Rate of twinning
Figure 4.13: Evolution of the TVF and of the rate of twin formation for
a tensile test. Simulations are performed with parameterisations leading
to the appearance of the rst twins after 0, 5 and 10% elongation.
4.3. Eect of anisotropic hardening on texture 77
to the experimental texture. The value of R
L
are very close to each
other (Table 4.8).
PAGE
0 90
PHI2= 45
(a) Exp
PAGE
0 90
PHI2= 45
(b) 0%
PAGE
0 90
PHI2= 45
(c) 5%
PAGE
0 90
PHI2= 45
(d) 10%
Figure 4.14:
2
= 45 ODF section after plane strain compression
(
eq
= 0.3). Simulations are performed with parameterisations leading
to the appearance of the rst twins after 0, 5 and 10% elongation in
uniaxial tension. Contour lines : 1-2-4-8.
Exp 0% 5% 10%
Rolling (
eq
= 0.3) 2.0 2.1 2.1 2.1
Rolling (
eq
= 0.9) 4.7 1.0 1.2 1.4
Table 4.8: R
L
ratio between the ODF intensity of the Brass and the
Copper component after plane strain compression for parameterisations
leading to the appearance of the rst twins after 0, 5 and 10% elongation
in uniaxial tension.
The evolution of the TVF and the rate of twin formation are presented
in Fig. 4.15. The level of deformation where the rst twins appear is
78 CHAPTER 4. Assessment of the model
the same as in tension. However, the rate of twin formation is dierent.
It is approximately constant at an eective strain of 25% whereas it is
already decreasing for a same strain amplitude in tension. Although
the three parameterisations lead to similar TVF after 30% elongation
in uniaxial tension, the TVF predicted in plane strain compression are
completely dierent. The TVF is much larger (and more realistic in
view of qualitative experimental data) when the ratio is not constant
(Table 4.7).
We can see in Fig. 4.16 that for an eective strain of 0.9, none of the
parameterisations gives good texture predictions. The intensity of the
Brass component is slightly too high and the intensity of the Copper
component is too low. But we can observe that the intensity of the
Copper component decreases and the R
L
value increases with the strain
amplitude at which the rst twin appears (Table 4.8). This is in fact
rather correlated to the larger TVF at large strains (Table 4.7).
4.3.2 Twin-twin interaction
Experimental observations indicate that twinning occurs predominantly
along a single twinning system within each grain or within each sub-
region of a grain. This is most probably due to the fact that existing
twin lamellae constitute obstacles to the formation of new twin lamellae
along intersecting crystalline planes. This eect can be implemented
into the model by assigning a dierent CRSS,
t,
c
, to the dierent twin-
ning systems instead of considering a unique (isotropic)
t
c
value. In
particular, the anisotropy is introduced by imposing a matrix A
tt
in
equation 3.29 where the entry (k, ) is equal to 1 if the twin system k
and the twin system are coplanar and b
tt
1 if they are non-coplanar.
Texture predictions obtained after plane strain compression for an ef-
fective strain of 0.9 for various values of b
tt
are compared to each other.
When b
tt
= 1, there is no anisotropy of the CRSS of the twin systems.
The value of
1

tt
(Eq. 3.29) is modied in each simulation in order to
predict the same nal TVF, equal to 23.7%. The inuence of twin-twin
anisotropy on the intensities of the Brass and Copper components can
be observed in Table 4.9. The fraction of twins created on the rst twin
system activated is also given. The value of b
tt
has only a small inuence
on the intensity of the Brass component. There is a slight increase with
the increase of b
tt
. The inuence is more signicant on the intensity
of the Copper component. Increasing b
tt
reduces the Copper intensity,
4.3. Eect of anisotropic hardening on texture 79
(a) TVF
(b) Rate of twinning
Figure 4.15: Evolution of the TVF and of the rate of twinning during
plane strain compression. Simulations are performed with parameteri-
sations leading to the appearance of the rst twins after 0, 5 and 10%
elongation in uniaxial tension.
80 CHAPTER 4. Assessment of the model
PAGE
0 90
PHI2= 45
(a) Experiment
PAGE
0 90
PHI2= 45
(b) 0%
PAGE
0 90
PHI2= 45
(c) 5%
PAGE
0 90
PHI2= 45
(d) 10%
Figure 4.16:
2
= 45 ODF section after plane strain compression
(
eq
= 0.9). Simulations are performed with parameterisations leading
to the appearance of the rst twins after 0, 5 and 10% elongation in
uniaxial tension. Contour lines : 1-2-4-8.
leading to a lower value of R
L
. When b
tt
= 100, 98% of the twins are
produced on the rst twin system activated and the texture prediction
is closest to experimental data. Therefore, this value is used again in
the next section.
4.3.3 Slip-twin interaction
Twin lamellae act as strong obstacles to dislocation glide. As a con-
sequence, grains containing twins are probably harder than grains free
of any twins. The inuence of such twinning-induced hardening at the
crystal level can be assessed by varying the value of the =

0

st
ratio
(Eq. 3.29) leading to
1

st
=
1

0
. We have performed simulations with
4.3. Eect of anisotropic hardening on texture 81
Exp b
tt
= 1 b
tt
= 5 b
tt
= 20 b
tt
= 100

tt
- 4.6 4 3.8 3.75
Brass Intensity 6.8 7.7 7.7 7.8 7.9
Copper Intensity 1.4 6.2 5.7 5.6 5.5
%TVF on 1
st
TS ? 70 90 96 98
Table 4.9: Inuence of the twin-twin interaction on the ODF intensities
of the Brass and Copper components and on the percentage of twins cre-
ated on the rst twin system activated in a grain grain after plane strain
compression (
eq
= 0.9). Simulations are performed with ALAMEL and
with parameterisations leading to the appearance of the rst twins after
10%.
values of ranging from 1 to 10000. In the latter case, once rst twins
have appeared in a grain, hardening of the slip systems is only governed
by the activity on the twin systems. In order to obtain the same TVF at
the end of the simulation, the value of
1

tt
must grow with the value of
1

st
. The ODF intensity of the Copper and the Brass components after
plane strain compression and an eective strain of 0.9 is given in Table
4.10. The intensity of the Brass component is mainly not aected by
the value of except when the latter is very large, i.e. larger than 1000.
However, the intensity of the Copper component decreases strongly with
the increase of this value.
The evolution of the TVF on the rst twin system and the evolution
of the position of the copper component in the orientation space are
shown in Table 4.10. The volume fraction of twins created on the rst
twin system slightly increases with the increase of the . The position
of the copper component increases with the increase of and the texture
obtained with = 10000 is closest to the experimental texture. However,
this value of leads to a very high hardening of the slip systems. We
will see in the next section whether this is conciliable with experimental
data.
So far we have considered isotropic hardening of the slip systems:
the CRSS was identical along all slip systems, i.e.
s,
c
=
s
c
. In
reality, when a family of twin lamellae is present, one may expect that
slip systems intersecting the twin plane are hardened more than co-
planar slip systems, i.e. dislocation glide along crystal planes parallel to
the twin lamellae is not hindered so much. The anisotropy is introduced
82 CHAPTER 4. Assessment of the model
Exp 1 4 20 100 1000 10000

tt
- 3.75 4.35 10.7 43.2 337 2100
Brass Intensity 6.8 7.9 7.9 7.9 7.9 7.8 7.2
Copper Intensity 1.4 5.5 5.6 5.5 4.7 3.5 1.6
position of C (

) 37.5 26 26 26 27 28 29
%TVF on 1
st
TS ? 98 99 99 99.5 100 100
Table 4.10: Intensity of the Brass and Copper components, position of
Copper and percentage of twins created on the rst twin system after
plane strain compression (
eq
= 0.9). Simulations are performed with
ALAMEL, with parameterisations leading to the appearance of the rst
twins after 10%, with b
tt
= 100 and for various values of , the weight
of the twin contribution to the hardening of slip systems.
by imposing a matrix A
st
in equation 3.24 where the entry (k, ) is
equal to 1 if the twin system k and the twin system are coplanar and
b
st
1 if they are non-coplanar. The ODF intensity of the Copper and
the Brass components obtained after plane strain compression and an
eective strain of 0.9 is given in Table 4.11 for various values of b
st
.
Exp 1 4 20 100

tt
- 3.75 3.5 3.15 2.95
Brass Intensity 6.8 7.9 7.6 6.8 4.9
Copper Intensity 1.4 5.5 5.7 5.3 3.9
Table 4.11: Intensity of the Brass and Copper components after
plane strain compression (
eq
= 0.9). Simulation are performed with
ALAMEL, with parameterisations leading to the appearance of the rst
twins after 10%, with b
tt
= 100 and for various values of b
st
, the param-
eter which models the slip-twin interaction.
The intensity of the Copper component decreases with the increase of
b
st
. However, one observes a more important decrease of the Brass
component. This does not conform to experimental data.
4.3. Eect of anisotropic hardening on texture 83
4.3.4 Texture prediction with our most elaborated parametri-
sation
In order to conclude this section, we compare textures predicted with
ALAMEL and our most elaborated parametrisation (Table 4.1 and 4.12)
with experimental textures (Fig. 4.17). I want to emphasise that our
most elaborated parametrisation remains simplied compared to the
physics reality. The most important parameters are b
st
and b
tt
which
Values

t
c0
230MPa

st

0
1
b
st
100

ts

0
0

tt
2.95
b
tt
100
Table 4.12: Value of the parameters specic to our modelling of twinning
model the slip-twin and the twin-twin interaction. The value of
t
c0
and

ts

0
= 0 is also important in order to predict the onset of twinning at
the right strain level. Finally, the value of

0

tt
is a tting parameter reg-
ulating the produced twin volume fraction. It can be observed that the
texture predicted after a tensile test with the most elaborated parametri-
sation is very close to the one presented when the crystal hardens isotrop-
ically (Fig. 4.9). The texture is only globally slightly smoother. The
latter conclusion can also be drawn for the texture predicted after sim-
ple plane compression (
eq
= 0.9) (Fig. 4.18). One can observe that
the spread around the Brass component is larger. It appears that the
enhancement of our modelling towards a more physical modelling of the
hardening of slip and twin systems leads to a texture prediction after
large deformation by simple plane compression which is way better than
the texture without twinning for a TVF of 24%. This conclusion is
dierent from the one obtained by Kalidindi [39] for an equivalent de-
formation and TVF who sees only a very minor inuence of twinning
on the texture. Nevertheless, as stated by Leers [47], we have seen that
the volume fraction can not be the sole explanation to the creation of
the brass-type texture.
It should be emphasized that our physical parameterisations allows
84 CHAPTER 4. Assessment of the model
(a) Tensile test (
eq
= 0.3) : Experimental
(b) Tensile test (
eq
= 0.3) : Prediction
PAGE
0 90
PHI2= 45
(c) Rolling (
eq
= 0.9) : Ex-
perimental
PAGE
0 90
PHI2= 45
(d) Rolling (
eq
= 0.9) :
Prediction
Figure 4.17: Comparison between experimental textures and predicted
texture after simulation of a tensile test (contour lines : 1-1.4-1.7-2.1-2.4)
and a simple plane compression (contour lines : 1-2-4-8).
4.3. Eect of anisotropic hardening on texture 85
(a) Intensity
(b) Position of
(c) Position of
1
Figure 4.18: Comparison between experimental textures and predicted
texture after simulation of simple plane compression.
86 CHAPTER 4. Assessment of the model
a good prediction for small deformation in tension
a good predictions for small deformation in simple plane compres-
sion
an appearance of the rst twins at a level of deformation which is
close to the appearance of the rst twins in our material
a concentration of the twins on a single twin system in a grain
an evolution of the twinning rate which is close to experimental
observation, particularly, the decrease of this rate at high level of
deformation.
The fact that our parameterisation does not predict the brass-type
texture can be explained by the fact that the ALAMEL model does not
allow sucient strain heterogeneity. Leers [47] pointed out that twin-
ning leads to a concentration of the slip in a grain on a single slip plane
parallel to the twinning plane. The ALAMEL model does not allow
that. One solution could be to use a tuned version of the viscoplastic
self-consistent (VPSC) model for which it has been shown that it allows
localisation of the deformation on one slip system. In the future, clusters
of grains of the ALAMEL model could be used as inclusion in the VPSC
model.
In the same review, Leers mentionned that the concentration of
the slip on a single slip system must be correlated with the presence
of a non-negligeable amount of non-twinned grains which can deform
by multi-slip leading to a composite deformation pattern. Our model
tends to conrm this analysis since the best texture prediction (Table
4.10) is coupled with a volume fraction of grains having a TVF lower
than 10% which is twice as large as for other parameterisations (for the
full detail of this study, see Appendix A).
Ane other way to improve predictions could be the modelling of
shear banding. It as not been incorporated in our model because they
are expected to appear after larger deformations than the ones studied in
this thesis. Furthermore, they have not been observed in our material
for equivalent strain up to 0.9. And their appearance is dicult to
reproduce even with FE simulations.
4.4. Hardening prediction 87
4.4 Hardening prediction
In this thesis, I have mainly focused my eorts on texture prediction.
The idea was rst to introduce in our model mechanical phenomena
which occur in the real material in order to improve texture predictions.
The second part of the plan is to validate our modelling based on the
prediction of the hardening. The goal is to check if the parameterisa-
tion which allows good prediction of the texture leads to an accurate
description of the macroscopic hardening : the stress-strain curve, the
incremental hardening n =

and the derivative of the hardening


both in tension and in plane strain compression.


In this section, I will check separately the inuence of each modelling
choice on the prediction of the hardening. On each gure, the harden-
ing curve obtained without twinning is presented as a reference. The
stress-strain curve obtained without twinning has been tted before all
texture predictions in order to reproduce the stress-strain curve without
twinning predicted in the work of Allain [41].
Isotropic hardening : The value of the CRSS is dierent for disloca-
tion slip and for twinning. Hardening parameters are chosen to ensure a
constant ratio:
t
c0
=
s
c0
(see Section 4.1 for details of the parametrisa-
tion). The prediction of the hardening obtained with = 1.17 leading to
the best prediction of the texture after a tensile test is presented in g-
ure 4.19 with the red line. The simulation has been performed with the
Taylor model. The overall strength is close to the experimental stress-
strain curve. However, when we look at the incremental hardening and
at the derivative of the hardening (Fig. 4.20), one can observe that
we do not predict the change of slope that appears after approximately
a true strain of 10% in both gures. Twinning leads to an increase
of 200MPa. These results must be compared to the one obtained
by Allain [41] and Shiekhelsouk [49] that also observed a dierence of
200MPa between the stress with and without twinning after a true
strain of 0.3.
88 CHAPTER 4. Assessment of the model
Figure 4.19: Stress-strain curve obtained after a tensile test for a true
strain of 0.3. Simulations are performed with parameterisations leading
to appearance of the rst twins after 0, 5 and 10% of elongation.
Postponing the onset of twinning (Fig. 4.19): The evolution of

t
c
is governed solely by the TVF
1
leading to an appearance of the
rst twins after 5% or 10%
2
and leading to good texture prediction
after a tensile test. Simulations have been performed with the Taylor
model whereas all simulations presented in the remainder of the section
have been performed with ALAMEL. Macroscopically, both predictions
seem close to the experimental stress-strain curve. When we look at the
incremental hardening and at the derivative of the hardening (Fig. 4.20),
one can observe that a change in the slope appears after approximately
a true strain of 10% in both gures of the experimental curves. Our
modelling leads to a change of slope at a level of deformation which is
1
Reminder : in Eq. 3.29, we set
1

ts
= 0 or very close to 0 and we t
t
c0
and
1

tt
in order to reproduce the experimental evolution of the TVF during the uniaxial
tensile test.
2
The value of
t
c0
is respectively equal to 190 and 230MPa while the value of

0

tt
is respectively equal to 5 and 2.5.
4.4. Hardening prediction 89
(a) Incremental Hardening
(b) Derivative Hardening
Figure 4.20: a) Incremental hardening and b) Derivative of the hard-
ening obtained after a tensile test for a true strain of 0.3. Simulations
are performed with the Taylor model and parameterisations leading to
appearance of the rst twins after 0, 5 and 10% of elongation.
90 CHAPTER 4. Assessment of the model
equivalent to the level of deformation where the rst twins appear.
Favoring a single twin system : Hardening predictions obtained
after introduction of twin-twin interactions
3
are presented on gure
4.21a and 4.22. We have found that the best texture predictions are
obtained for b
tt
= 1000. The value of

0

tt
= 3.75. Macroscopically,
the prediction is close to the experimental stress-strain curve. The gap
between the two curves is important but this must be due to the use of
ALAMEL instead of Taylor. With this parametrisation the reduction
of the incremental hardening exponent is less important than without
twin-twin interaction but is not constant as observed experimentally.
Twinning induced hardening : Hardening predictions obtained af-
ter introduction of isotropic slip-twin interaction
4
is presented on the
gure 4.21b and 4.23. We have found that the value of = 10000 leads
to the best texture predictions. The value of

0

tt
= 2100. It is clear that
hardening predictions are bad. This parametrisation leads to a very high
hardening of the slip systems in grains that contain a lot of twins.
Anisotropic obstacles against dislocation glide : Finally, harden-
ing predictions obtained after introduction of anisotropic slip-twin hard-
ening
5
are presented on gure 4.21. The value of b
st
= 100. The value
of

0

tt
= 2.95. The introduction of anisotropic slip-twin hardening leads,
as expected, to an increase of the hardening. But, it was not expected
that the increase of the hardening be so important when the rst twin
appears. And even less expected that a second change of slope appears
in the incremental hardening and in the derivative of the hardening.
The reason must be our too naive modelling of the slip-twin interaction.
Indeed, our slip-twin interaction remains constant with the deformation
whereas the more there are twins in the grain the more the slip-twin
interaction must be important.
3
Reminder : the anisotropy is introduced by imposing a matrix A
tt
in equation
3.29 where the entry (k, ) is equal to 1 if the twin system k and the twin system
are coplanar and b
tt
1 if they are non-coplanar.
4
Reminder : the inuence of such twinning-induced hardening at the crystal level
can be assessed by varying the value of the =

0

st
ratio (Eq. 3.29) leading to
1

st
=
1

0
.
5
Reminder : the anisotropy is introduced by imposing a matrix A
st
in equation
3.24 where the entry (k, ) is equal to 1 if the twin system k and the twin system
are coplanar and b
st
1 if they are non-coplanar.
4.4. Hardening prediction 91
(a) Stress-strain curve
(b) Stress-strain curve
Figure 4.21: Stress-strain curve obtained during a tensile test. Simu-
lations are performed with ALAMEL and a) two parameterisations of
anisotropic hardening, b) the parameterisation leading to the best tex-
ture predictions.
92 CHAPTER 4. Assessment of the model
(a) Incremental Hardening
(b) Derivative Hardening
Figure 4.22: a) Incremental hardening and b) Derivative of the hard-
ening obtained during a tensile test. Simulations are performed with
ALAMEL and two parameterisations of anisotropic hardening.
4.4. Hardening prediction 93
(a) Incremental Hardening
(b) Derivative Hardening
Figure 4.23: a) Incremental hardening and b) Derivative of the hard-
ening obtained during a tensile test. Simulations are performed with
ALAMEL and the parameterisation leading to the best texture predic-
tions.
94 CHAPTER 4. Assessment of the model
4.5 Discussion
Our rst goal in this chapter was to study the inuence of the various pa-
rameters of our modelling of the hardening of the slip and twin systems.
We started from the simplest parameterisation to the most elaborated
one. Our second goal was to study the inuence of the various micro-
macro transition schemes on the predictions. And nally our last goal
was to see if our most elaborated parameterisation was able to correctly
predict all the experimental observations available.
The results of our simulations show that a simple incorporation of
twinning is sucient to obtain good predictions of the texture after
small deformation in both tension and rolling. But, as expected, it does
not allow good texture predictions after large deformation in rolling. It
allows us to obtain an overall good prediction of the stress-strain curve
but it does not predict the changes of slope which are observed in the
experimental incremental hardening.
The introduction of anisotropic hardening rst permits to control
the onset of twinning. Indeed, when working with a constant =

t
c0

s
c0
ratio, twinning is either triggered at the onset of plastic deformation, or
never. On the other hand, by varying , prior activity on the slip system
is needed in order to activate twin systems, as observed experimentally.
This change has allowed the improvement of the texture prediction in
rolling after large deformation for a same quality of texture predictions
after small deformation. Furthermore, it allows the prediction of the
change of slope in the incremental hardening and in the derivative of
the hardening as observed experimentally. It seems that the best over-
all predictions are obtained when the rst twin appears after 10%, as
observed experimentally.
The introduction of twin-twin interaction allows the improvement of
the texture prediction after large deformation in rolling without aecting
the other predictions. It leads to a concentration of the twins on the
rst twin system activated.
The most elaborated parametrisation with anisotropic slip-twin in-
teraction does not improve predictions in large deformation for both the
texture and the hardening. It leads to a decrease of the intensity of both
the Copper component and the Brass component. It leads to an increase
of the hardening but this increase is too important when the rst twins
4.5. Discussion 95
appear. A second change in the slope of the incremental hardening ap-
pears after 15% whereas experimentally, the stress-strain curve is close
to a straight line up to a deformation of 0.3.
It must be noticed that the hardening behaviour could be predicted
by using other initial values of the resistance against twinning (i.e. other
values of
t,
c0
). But, in the case of larger (resp. lower) value of
t,
c0
, the
same hardening behaviour would be obtained for a lower (resp. higher)
twin volume fraction and therefore for a dierent quality of the texture
prediction. In this thesis, we have reproduced the main trends of the
hardening behaviour for a parameterisation that leads to a good texture
prediction.
Whereas it was expected that replacing the Taylor model by a more
physical micro-macro modelling would be a solution to reach the brass
type-texture, we found out that CPFEM and ALAMEL only slighty
improve texture predictions. However a question remains about the re-
nement of the nite element mesh used. As simulations already take
a long time, I dit not try to perform simulations with a highly rened
mesh. This might improve the texture prediction, especially for large de-
formation in rolling where local deformation can be very heterogeneous.
Whereas Kalidindi [39] only observed a very minor eect of the
twinning on the texture prediction for a reduction of 58% in plane strain
compression and for a TVF of 20%, our most physical parametrisation
coupled with ALAMEL leads to a texture prediction which is much bet-
ter than the texture without twinning for a TVF of 24% after a reduction
of 54%. The intensity of the Copper component is lower than the inten-
sity of the Brass component. Nevertheless, as stated by Leers [47], the
volume eect associated with the creation of new lattice orientations
within the twins does not fully explain the Brass-type texture.
Our study shows that even with anisotropic slip-twin interaction
combined with the use of a grain-interaction model or CPFEM, it is
dicult to reach the Brass-type texture. An explanation is that even
with this combination, our model does not predict localisation of the
deformation on a single slip system as expected by Leers [47]. Another
explanation is the use of the Schmid law for the activation of the twin
systems that could be too simple. An enhancement to this law is to take
preferial sites for twin nucleation into account [50]. A last explanation
is the choice of a linear relation for the anisotropic slip-twin interaction.
Indeed, with this choice, each new fraction of twin created leads to the
96 CHAPTER 4. Assessment of the model
same increase of the hardening of the slip systems whereas experimen-
tally, the later twins are created, the more they reduce the mean free
path of dislocations in the untwinned part of the grains and the more
the slip-twin interaction is important. The linear relation could easily be
replaced by a more physical law inspired from earlier works ([36, 40, 41]).
However, this has not been done in the context of this thesis.
It is also expected that more realistic and more rened represen-
tations of the microstructures in CPFEM could help to improve the
prediction for highly heterogeneous material, which is the case of TWIP
steel. Whereas, in this part of the work, we have worked on a very
complicated material where a lot of experimental measurements must
still be made, I have chosen to work for the second part on simpler al-
loys in order to avoid adding up diculties. Nevertheless, all the work
presented in the next part also applies to TWIP steel. I had initially
intended to perform simulations with highly rened meshes on a rep-
resentative volume element at the end of my thesis in order to study
the inuence of the renement on texture predictions. This turns out
to require excessive computational power. Therefore, the simulations to
be presented at the end of the second part (Section 7.3) involves only
10 grains.
Part II
Advanced CPFEM in
polycrystalline aggregates
97
Chapter 5
Construction of model
polycrystals
Crystal plasticity based nite element modelling (CPFEM) provides an
accurate prediction of strain heterogeneity throughout polycrystalline
aggregates on the condition that (i) the microstructure representation is
realistic and (ii) mesh renement is sucient. The second part of this
thesis is aimed at reaching these two conditions.The present chapter is
organised as follows. In the rst section, we assume that the RVE is
known and we propose a mathematical algorithm allowing us to assign
a lattice orientation to each grain while ensuring a statistically represen-
tative reproduction of the global texture. Such texture discretisation
had never been applied to RVEs with heterogeneous grain sizes in the
past. The second section is devoted to the generation and the meshing
of the model polycrystal. There, the originality is to perform selective
mesh coarsening in the bulk of the grains while leaving grain boundary
regions highly rened and while maintaining periodicity of the FE mesh.
5.1 A novel algorithm for texture discretisation
This part of the work has already been published in an international
journal [51]. Some parts of the text in the present section come from
this publication.
99
100 CHAPTER 5. Construction of model polycrystals
5.1.1 State of the art
The orientation distribution function (ODF) denes the probability f(g)
that an elementary volume dv presents the lattice orientation g. A con-
tinuous description of the ODF can be derived from diraction experi-
ments [10, 52], but such a continuous description is not suited for crystal
plasticity simulations that must rely on a sampling of discrete orienta-
tions.
Prior to any crystal plasticity modelling, a discrete sampling of
lattice orientations that is statistically representative of the crystallo-
graphic texture measured in the undeformed sample must be generated.
The discretisation method is unbiased if, for any orientation depen-
dent property H(g), the following relation holds:
lim
N

N
i=1
v
i
H(g
i
)

N
i=1
v
i
=
_
f(g)H(g)dg (5.1)
wherein v
i
is the weight of orientation g
i
in the discrete sampling.
So far, relatively few methods have been proposed in order to dis-
cretise crystallographic textures initially represented by a continuous
ODF. Kocks et al. [52] and Helming and al. [53] discretise the tex-
ture by constructing a grid of orientations that are nearly equi-distant
in orientation space. Each orientation on the grid corresponds to one
grain. The volume of each grain is not exactly equal to the ODF value
for that orientation; it is determined iteratively by tting recalculated
pole gures on the experimental data. The total number of grains is
reduced subsequently by eliminating those with a weight lower than a
threshold value. Nowadays, new methods are also proposed as the one
of Eisenlohr [54].
A dierent discretisation technique has been proposed by Toth and
Van Houtte [55], which produces an equally-weighted set of orientations.
Toth and Van Houtte [55] demonstrate that their method is unbiased.
They use the Bunge convention and the input texture is represented by
ODF intensities calculated over a grid in which the Euler angles evolve in
steps of 5.
1
, and
2
range, respectively, from 0 to 360, 0 to 90,
and 0 to 90. The orientation space is subdivided into boxes centred
on the grid points and the ODF intensity is integrated over each box:
f
i
=
_
box
i
f(g)dg where dg =
1
8
2
sind
1
dd
2
. (5.2)
5.1. A novel algorithm for texture discretisation 101
Figure 5.1: a) Cumulative probability function calculated by integration
of the experimental ODF along a path in Euler space, b) close-up on the
stair-case function
If N is the total number of grid points, from the denition of f(g) [10],
we have
N

i=1
f
i
= 1. (5.3)
Then, they dene a path in Euler space so that every grid-point is en-
countered only once. At any orientation j along the path, they compute
a cumulative probability function (Fig. 5.1) having the following expres-
sion :
F(j) =
j

i=1
f
i
, with F(N) = 1. (5.4)
Finally, for each elementary volume in the model microstructure, they
draw x randomly between 0 and 1 and assign, to the selected elementary
102 CHAPTER 5. Construction of model polycrystals
volume, the discrete orientation i so that
F(i) x < F(i + 1). (5.5)
After a large number of draws, the proportion of elementary volumes
having orientation i is close to f
i
. The accuracy of the texture repro-
duction increases with the number of elementary volumes.
Our method is unbiased and it is more exible than any of the pro-
cedures available so far since it permits constructing model microstruc-
tures in which the texture and the grain-size distribution can be pre-
scribed independently. The only limitation of the new method is that
it does not account for any correlation between the grain size and the
grain orientation. Such a correlation could exist in real polycrystalline
aggregates, for example, after recrystallisation.
Attention should be drawn on the fact that such procedures, de-
signed to generate discrete orientation samplings, must not be confused
with another class of algorithms that are aimed at characterising crys-
tallographic textures by identifying the orientation, the strength and the
shape of predominant texture components [56, 57, 58, 59]. The latter
techniques do not provide suitable input for crystal plasticity modelling,
with the exception of [59].
5.1.2 Description of the algorithm
The objective of the algorithm is to generate a discrete sampling of
orientations associated to a list of weight so that the weighted list of
orientations is statistically representative of the experimental texture.
The two inputs of our algorithm are therefore a continuous description of
the experimental texture and a list of weights. The output is a weighted
list of orientations. The prescribed weight can be considered as the
volume of the grain of a heterogeneous microstructure.
The basic idea of our algorithm is to discretise each grain as a sum
of elementary volumes. The volume of the elementary volume is pre-
scribed as V
u
. Once we have determined the total number of elementary
volumes, N
u
, we can generate N
u
orientations with the method of Toth
and Van Houtte. Finally, we distribute the elementary volumes into the
grains so that all elementary volumes of a grain have a low disorienta-
tion with regard to one another. The term disorientation refers to
5.1. A novel algorithm for texture discretisation 103
the smallest rotation angle bringing one crystal lattice onto the other.
As shown in Appendix B, disorientations are eciently computed us-
ing quaternions [60]. The threshold disorientation relative to the mean
orientation of the box is denoted .
As an illustration of the procedure that we use to regroup the elemen-
tary volumes into the grains, let us consider an example microstructure
containing twelve elementary volumes labelled with letters A to L. The
elementary volumes must be sorted so as to form three grains with pre-
scribed sizes : N
1
= 5, N
2
= 4, N
3
= 3. Table 5.1 is a symmetric binary
table containing a 1 in position (i, j) if the lattice disorientation between
orientations of elementary volumes i and j is less than . The last col-
umn of row i indicates the largest number of elementary volumes that
can possibly be allocated to a box centred on i.
We proceed from the largest to the smallest grain. According to
Table 5.1, the box representing the rst grain may be centred either on
A, B, D, E or J, respectively. After a random selection, we choose to
centre the rst box on D. As the required grain size is ve, the rst grain
is dened with ve of the six elementary volumes appearing in row D
of Table 5.1. These elementary volumes A, B, C, D, E are withdrawn
and a new table (Table 5.2) is created. The sixth elementary volume,
J, is not included in the rst grain because this elementary volume,
appearing in four rows of Table 5.2, can easily be allocated to another
grain.
The next largest grain must be constituted of four elementary vol-
umes. Only a box centred on G and containing F, G, H, I is suitable.
Finally, three elementary volumes are left which are disoriented by less
than . Hence, the last grain is formed by J, K, L. If a suciently
large box could not be dened, the number of elementary volumes re-
quired would be decreased to the size of the largest box that can be
created. At the end of this operation, one computes the mean orien-
tation within every box (Appendix B), and this orientation is assigned
to the corresponding grain.
In practical situations, the total number of grains is much larger than
three and the number of elementary volumes is much larger than twelve.
The construction of the binary table becomes the most costly operation
of the algorithm. Indeed, the computation time of the naive algorithm,
which consists in the computation of the disorientation between each
pair of elementary volumes, grows with the square of the elementary
104 CHAPTER 5. Construction of model polycrystals
A B C D E F G H I J K L Total
A 1 1 1 1 1 1 0 0 0 1 0 0 7
B 1 1 1 1 0 1 0 0 0 0 0 0 5
C 1 1 1 1 0 0 0 0 0 0 0 0 4
D 1 1 1 1 1 0 0 0 0 1 0 0 6
E 1 0 0 1 1 0 0 0 0 1 1 1 6
F 1 1 0 0 0 1 1 0 0 0 0 0 4
G 0 0 0 0 0 1 1 1 1 0 0 0 4
H 0 0 0 0 0 0 1 1 1 0 0 0 3
I 0 0 0 0 0 0 1 1 1 0 0 0 3
J 1 0 0 1 1 0 0 0 0 1 1 1 6
K 0 0 0 0 1 0 0 0 0 1 1 1 4
L 0 0 0 0 1 0 0 0 0 1 1 1 4
Table 5.1: Preparation of the sorting of 12 elementary volumes into 3
grains. The elementary volumes, which have similar lattice orientations,
are highlighted with a 1.
F G H I J K L Total
F 1 1 0 0 0 0 0 2
G 1 1 1 1 0 0 0 4
H 0 1 1 1 0 0 0 3
I 0 1 1 1 0 0 0 3
J 0 0 0 0 1 1 1 3
K 0 0 0 0 1 1 1 3
L 0 0 0 0 1 1 1 3
Table 5.2: Reproduction of Table 1 after removal of the rst grain
5.1. A novel algorithm for texture discretisation 105
volume number. This construction has therefore been optimised in order
to have a computational cost that is proportional to the number of
elementary volumes. The quaternion space is subdivided into 102020
boxes b where the centre of the box b
ijk
is q
center
= (q
0
, q
1
, q
2
, q
3
) =
(
i+0.5
10
,
j10+0.5
10
,
k10+0.5
10
, 1.0q
2
0
q
2
1
q
2
2
). One can verify that with this
subdivision of the quaternion space, each quaternion, or their symmetric
value, belongs to a box.
Thanks to this subdivision, we can avoid computing the disorienta-
tion of pairs of quaternions that belong to remote boxes. Two boxes
are nearby if each of them contains at least one quaternion with a low
disorientation relative to the other box. Nearby boxes are recorded only
once at the rst execution of the program.
5.1.3 Comparison with the Toth and Van Houtte method
The accuracy of the new discretisation technique is assessed in the case of
three samples of respectively, 100, 500 and 1000 orientations. For these
three cases, the discretisation is done with our method and with the
method of Toth and Van Houtte. The discretised experimental texture
(Fig. 5.2a)) is an IF steel with a cold-rolling texture [61]. The parameter
values of our technique used for this evaluation are : = 7 and N
u
=
10000 (in the case with 100 grains) or N
u
= 100000 (in the case with
512 and 1000 grains). The weight distribution is a Gaussian one.
After Texture discretisation, continuous ODFs are recalculated by
replacing each discrete orientation by a spherical gaussian distribution
with a half-scatter width of 5. The series expansion method is used for
this purpose [10]. Pole gures are represented in Figure 5.2 for the two
techniques and the three numbers of grains. For each discrete sampling,
the dierence between the recalculated ODF and the experimental ODF
which is equal to:
Misfit =
_
(f
exp
(g) f
reconstructed
(g))
2
dg
_
(f
exp
(g))
2
dg
(5.6)
is given in Table 5.3.
These results demonstrate that the discretisation of a texture re-
quires fewer individual orientations if weights (i.e. grain volumes) are
106 CHAPTER 5. Construction of model polycrystals
Figure 5.2: a) Representation of the experimental texture with pole
gures. b-d) Recalculated pole gures. Weights of discrete orientations
are homogeneous in b)and c), and heterogeneous in d). The number of
orientations is 100 in b) and d), whereas it is 1000 in c). Pole Figure
contours : 1/2/3/4/5/6/7/8/9/10
5.2. Mesh generation algorithm 107
Homogeneous grain size Heterogeneous grain size
100 512 1000 10000 100 512 1000
Mist 1.787 0.494 0.403 0.246 0.280 0.1451 0.1384
Table 5.3: Mist between the experimental texture and the discrete
textures for dierent numbers of grains and dierent grain size distribu-
tions.
non-uniform than if all weights are equal. This is not a surprising con-
clusion in the case of severely textured metals, such as the cold rolled
steel sheet considered here (the maximum ODF intensity is 11). In-
deed, when all weights are equal, strong texture components must be
represented by many grains with identical or very similar crystal orien-
tation. The accuracy of the texture reproduction is the same if strong
texture components are represented by one or two grains having a much
larger weight than the other grains. This dierence in the eciency of
the two discretisation techniques vanishes if the experimental texture is
very weak. Actually, there would be no dierence at all if lattice orienta-
tions were perfectly random. However, crystal plasticity modelling is less
justied in such samples having a quasi isotropic mechanical response.
5.2 Mesh generation algorithm
The mesh generation algorithm proposed in this section has already been
used, and sometimes also described, in various publications [62, 63, 64].
5.2.1 State of the art
Mechanical properties of polycrystalline aggregates are the result of
physical phenomena occurring at the grain scale. Multi-scale models link
microstructure characteristics and macroscopic mechanical response, but
the accuracy of their predictions hinges on a realistic representation of
the polycrystalline microstructure. A realistic representation has to take
account of the texture, the grain shape and the topological arrangement
of all phases composing the multiphase steel.
Various micro-macro transition theories available in the literature
108 CHAPTER 5. Construction of model polycrystals
have been presented in the rst chapter of the thesis. One of the most ac-
curate micro-macro transitions is obtained with crystal-plasticity-based
FE modelling (CPFEM). The microstructure is discretised over a mesh
in order to model the interaction between each grain and its neighbour-
hood. Most of such numerical analyses have so far been performed with
homogeneous microstructures, so that all grains are represented by a
single nite element or by the same set of elements (Fig. 5.3a and 4.8).
This kind of RVE shows its limits when the alloy is composed of more
than one phase such as dual-phase steel (Fig. 5.3b). It has been shown
[46, 65] that too simplied RVEs do not allow a valid prediction of the
plastic strain heterogeneity between the phases. It has also been shown
[46, 66] that predictions of CPFEM simulations are strongly dependent
on the topological arrangement of the second phase.
(a) (b)
Figure 5.3: a) Example of an homogeneous microstructure where grains
are represented by a cube divided in 24 tetrahedra, b) Example of the
microstructure of a dual-phase.
Another kind of simulations is based on image reconstruction of real
microstructures [67]. Simulations are run on a mesh which reproduces a
mapped area of the real microstructure. The sampling is generally small
in order to have cheap simulations (Fig. 5.4). The disadvantage of this
kind of simulations is that it predicts the behaviour of the small sample
of the material but not necessarily the overall behaviour of the material.
These microstructures are not an RVE since the number of grains is too
low.
5.2. Mesh generation algorithm 109
(a) (b)
Figure 5.4: Example of a real microstructure and of the mesh obtained
after reconstruction by pattern recognition [67].
Recently, new model microstructures for single phase materials have
been generated where grains are created as Voronoi cells [68, 69]. Diard
et al. [68] distribute the seed positions with a Poisson process. They
then map the microstructure on a quadratic element mesh (voxels) so
that each integration point of the mesh is assigned to a grain. The latter
technique is also used by Barbe [69].
More recently, Quey [70] developed a new meshing technique with
tetrahedra. Grain surfaces are meshed and nally the volume of each
grain is meshed separately from the mesh surfaces. The main advantages
of this kind on unstructured meshes are a better representation of grain
boundaries and the use of more realistic grain shapes.
We have developed a new generator of model microstructures based
on unstructured FE meshes and with the ability of performing selective
renement.
5.2.2 Description of the algorithm
The algorithm creates a periodic microstructure which is representative
of a multiphase polycrystal. A cube of volume V is discretised in tetra-
hedra, having approximately the same edge length l and such that each
tetrahedra face that lies on the periphery has a periodic equivalent on
the opposite cube face. The user prescribes the total number of grains
N, the volume fraction of the second phase f
2
and the ratio R between
the mean diameter of the grains in either phase.
110 CHAPTER 5. Construction of model polycrystals
The variables describing the microstructure are denoted by the fol-
lowing symbols:
V , the volume of the cube
L, the length of cube vertices
l, the mean length of a tetrahedron edge
D
1
, the mean diameter of the grains of the primary phase
D
2
, the mean diameter of the grains of the secondary phase
f
1
, the volume fraction of the rst phase
f
2
, the volume fraction of the second phase
V
1
, the mean volume of the grains of the rst phase
V
2
, the mean volume of the grains of the second phase
N
1
, the number of grains of the rst phase
N
2
, the number of grains of the second phase
N, the total number of grains
R, the ratio between D
1
and D
2
D
min
, the average nearest-neighbour distance in the secondary
phase.
R
D
, the ratio between D
min
and D
2
The following trivial relations link the V
1
, V
2
, N
1
, N
2
, D
1
, D
2
vari-
ables to the input variables :
N
1
+N
2
= N (5.7)
N
1
V
1
+N
2
V
2
= V (5.8)
N
2
V
2
V
= f
2
(5.9)
D
1
D
2
= R (5.10)
V
1
=
4
3
(
D
1
2
)
3
(5.11)
5.2. Mesh generation algorithm 111
V
2
=
4
3
(
D
2
2
)
3
(5.12)
The algorithm is divided in two steps, one for each phase. For the
generation of the rst phase, a seed is positioned randomly for each
grain. In order to obtain a quasi-uniform distribution of grain volumes,
a minimum distance between the seeds is imposed, which is propor-
tional to the grain diameter. Grains are then generated as Voronoi cells
around the seeds. The latter is obtained with the following algorithm :
one computes the distance d
tg
between the centre of each tetrahedron
t and the seed position of grain g. Then, one assigns tetrahedron t to
grain g

so that d
tg
< d
tg
g (Fig. 5.5). In general, d
tg
is computed
in the Euclidian metric in order to create equiaxed grains. But we also
allow the computation of d
tg
in other metrics in order to create grains
with various aspect ratios. Finally, a smoothing procedure suppresses
eventual excessive roughness of the grain boundaries. All tetrahedra be-
longing to a grain A while being neighbours to three tetraedra belonging
to a grain B are assigned to the latter.
(a) Whole mesh (b) Triangle on grain boundaries
Figure 5.5: Periodic microstructure generated with our algorithm before
the introduction of grains of the second phase.
For the secondary phase, seed positions are selected among the nodes
that are at the intersection of four grains of the primary phase. This
takes into account the fact that grains of the second phase appear mainly
at triple junctions in EBSD images of real microstructures (Fig. 5.3b).
The position of the rst seed is selected randomly and then new seeds
are dened in such a way that the minimum distance between all pairs of
112 CHAPTER 5. Construction of model polycrystals
seeds is maximised. Grains of the second phase are then allowed to grow
with the addition of elements that are neighbours of elements already
composing the grain. This growth ends when the volume fraction of the
second phase reaches the prescribed value (Fig. 5.6).
(a)
(b) (c)
Figure 5.6: a) Periodic microstructure generated by our algorithm after
the introduction of grains of the second phase. b) Cut in the mesh before
the introduction of the grain of the second phase. c) Cut in the mesh
after the introduction of the grain of the second phase at triple junction
point.
Specic care is taken in order to create a periodic mesh. Already the
rst subdivision of the cubic RVE into tetraedra does not automatically
5.2. Mesh generation algorithm 113
produce a periodic mesh [71]. Therefore one starts by meshing three
faces of the cube with triangles (Fig. 5.7a). Then, the three faces
are reproduced seven times by applying mirror symmetry in the three
directions of space. This results in a fully periodic mesh of the surfaces
of the cube (Fig. 5.7b). The volume of this cube is eight times larger
than the volume of the cube initially meshed. The interior of the cube
is then meshed with tetrahedra with the surface mesh prescribed. In
order to ensure the periodicity of the rst phase, all distances between
the elements and the seeds have been computed in such a way that
they satisfy the following property : The distance between two points is
computed as the minimum value of the distances between the position of
the rst point and the 27 duplicated periodic positions of the second point.
In order to ensure a periodic grain growth of grains of the second phase,
a pre-processing is performed. The latter adds periodic neighbours to
elements that are at the boundaries of the mesh.
Microstructures are generated starting from an initial mesh in which
all tetrahedra have approximately the same edge length. As you can
see in Figure 5.8, the edge length is far from homogeneous in an initial
mesh produced by GMSH [71]. This was expected as the mesh generator
optimises the element quality instead of satisfying the prescribed edge
length throughout the mesh. The edge length should be chosen in order
to have a sucient mesh renement in the smallest grains. If R, the
ratio between average grain radii of both phases, is large, grains of the
rst phase hence comprise a very large number of elements. In order
to reduce the computational time, the mesh is coarsened in the bulk of
large grains.
5.2.3 Mesh coarsening in a periodic RVE
The part of the work presented in this subsection has been carried out
by Cecile Dobrinzky. This is presented in the thesis for the sake of
consistency with the remainder of this section.
The particularity of the desired mesh coarsening is that we want to
preserve grain geometry and of course the periodicity of the mesh. In
order to ensure that the grain geometry is not altered, all triangles lying
on grain boundaries are constrained (Fig. 5.5b). It is more complicated
to coarsen the mesh while preserving periodicity. The main diculty of
periodic mesh adaptation is the management of the mesh entities that
114 CHAPTER 5. Construction of model polycrystals
(a)
(b)
Figure 5.7: a) Mesh of three adjacent faces of a cube. b) Periodic mesh
of the surface of a cube obtained after copy and translation of the three
faces of a)
5.2. Mesh generation algorithm 115
Figure 5.8: Nodal eld of characteristic length in a mesh before the
coarsening procedure.
are on periodic boundaries. For instance, if an edge has to be split,
swapped or collapsed on a face, all its periodic counterparts have to
be modied as well. As a rst approach, we could modify every basic
mesh operation in order to take periodicity into account : this includes
point insertion, edge swap, edge collapse, edge split, face collapse... This
represents a lot of eort and often leads to slowing down the non-periodic
mesher.
We have used another way to adapt this kind of meshes. Our method
is based on the specicity of a periodic mesh. A periodic domain is in-
nite in the three directions of space: it may be duplicated along each
such direction. We are only interested in cases for which the periodic
domain is discretised with a periodic mesh. One consequence of period-
icity is that the bounding box of the grid can change without modifying
the geometry. In other words, if we have a periodic mesh representing
a periodic domain, we can displace elements touching periodic bound-
aries toward the opposite matching boundary and the resulting mesh
116 CHAPTER 5. Construction of model polycrystals
will characterise the same domain.
Based on this characterisation of a periodic mesh, we use a two-step
iterative procedure:
Adapt the mesh, leaving periodic entities unchanged,
Update periodic boundaries by migrating elements.
Consequently, some kind of mesh migration procedure has to be added
to the algorithm. The advantage of this approach is that mesh migration
is completely independent of mesh adaptation, i.e. those two building
blocks are completely separated in their implementation.
The process begins with a periodic mesh (representing a periodic
domain) and then proceeds iteratively. In order to modify the mesh,
we call a general mesher/remesher. The only constraint of the mesh
generator is to leave unchanged the elements on the periodic boundaries
and the constrained elements on the grain boundaries, i.e. not to touch
the triangles which are constrained. The main part of the algorithm is
the displacement of the periodic boundaries towards the mesh interior in
order to be able to adapt them later on. As stated before, this does not
aect the intrinsic geometry of a periodic microstructure. Starting with
the observation that all triangles on the periodic boundaries separate
pairs of elements, we move one element from one side to the other so as
to regroup the pair. Then the general mesher/remesher is called again
unless a stop criterion decides that mesh adaptation has converged. The
process is illustrated in gure 5.9 for a two-dimensional example (which
eases visualisation). In this example, one creates a periodic uniform
mesh for a domain which is periodic in one direction: the horizontal
direction.
The rst key point on the migration process is to move all the ele-
ments containing a periodic entity towards the same destination. This
assertion can appear obvious. However, when several boundaries are
periodic, several destinations are possible for the entities and the desti-
nation choice should not be made randomly. The second thing to do is
to classify the edges according to their agreement with the desired sizes.
Lets choose e an edge of the mesh and l(e) its length on the prescribed
metric. We dene the ratio:
r(e) = [1 l(e)[.
5.2. Mesh generation algorithm 117
Figure 5.9: Illustration of the coarsening procedure on a two dimensional
example which is periodic in one direction : the horizontal direction.
The aim of mesh adaptation procedure is to generate a unit mesh, which
means that we minimise the maximum of r(e). Thus we classify the mesh
edges in function of their r(e) value and we rst decide the migration
destination of the edge which has the maximum r(e). Coupled with
the process described in the previous paragraph, the advantage of this
method is to be sure that at least the worst edge (i.e. the edge with the
largest r(e)) will be treated by the (re)mesher.
The aim of the algorithm is to adapt the mesh to a prescribed metric.
Consequently, we want to stop it when we obtain a good mesh, i.e. a
mesh as close as possible to a unit mesh. The two main characteristics
are the quality of the elements and the length of their edges.
5.2.4 Parameterisation of the size eld
A size eld is required in order to coarsen the mesh selectively at the
grain centres. The size eld is the nodal eld of prescribed edge lengths.
Coarsening is performed at the centres of the grains because the strain
eld is expected to be more heterogeneous near grain boundaries [72, 73]
due to grain interactions. In order to capture this strain heterogeneity,
it seems important to have a higher renement of the region close to grain
boundaries. Therefore, the size eld is decomposed into two subregions.
The rst region is a layer containing the nodes that are close to grain
boundaries while the second region is a core containing the nodes
that are near grain centres. Each node n is assigned to one of the two
118 CHAPTER 5. Construction of model polycrystals
subregions depending on its distance to the closest grain boundary. The
latter is computed as the minimum over all the distances between the
node n and a grain boundary node. This distance is then compared to
a threshold distance which describes the layer thickness.
There are various possible choices for the characteristic edge lengths
imposed in the two subregions. For the nodes in the layer close to grain
boundaries, we can either assign a characteristic length which is equal
to the mean characteristic length of the initial mesh or we can assign
the mean length of all edges adjacent to the node. I chose the second
option because a remesher will mainly try to respect the prescribed size
eld rather than trying to preserve the mesh quality. As an homoge-
neous characteristic length in the layer will not improve the prediction,
I prefer to keep the discretisation given by the initial mesher which en-
sures high element quality. For the nodes near the grain centres, we can
either impose a unique characteristic length or impose a characteristic
length which increases with the increase of the distance to the closest
grain boundary. I chose the second option in order to have smoother
transitions in the size eld. Indeed, I want to impose a value of the
characteristic length at the centre of the grain which is up to 200 times
larger than the initial characteristic length. This might lead to elements
with very bad quality at the boundary between the two subregions. And
elements of very bad quality result in an inaccurate solution. Therefore
the characteristic length assigned to a node n near the grain centre is
equal to a weighted contribution of L
GB
, the mean characteristic length
of the initial mesh, and L
GI
, the maximum characteristic length which
is proportional to L
GB
.
Lc
n
= L
GB
weight +L
GI
(1 weight) (5.13)
where the weight has a value of 0 if the distance is equal to the layer
width and 1 if it is equal to the radius of the largest grain in the mi-
crostructure. Typically, the maximum value of the ratio
L
GB
L
GI
is equal to
200. The weight evolves linearly between L
GB
and L
GI
. An example of
the size eld is shown in Figure 5.10.
5.2.5 Post processing operation
The coarsening algorithm leads to a mesh envelope which is far from
the initial cube (Fig. 5.11). It was expected due to the process of element
migration during the algorithm. Such mesh is however inconvenient for
5.2. Mesh generation algorithm 119
(a) layerwitdh= 0 (b) layerwitdh=
1
20
of the cube length
Figure 5.10: Example of size eld imposed to the microstructure for two
widths of the grain boundary layer.
the simulation because the number of equations needed to ensure the
periodicity of the simulation is increased. Moreover, from a practical
point of view, some comparisons between coarsened and non-coarsened
meshes will be dicult if the mesh has such an highly irregular shape.
Post processing is therefore applied after the coarsening procedure.
A cube is delimited in such a way that the position of the centres of all
coarsened grains is the same as the position of the grain centres before
coarsening. One also veries that the eight corners of the initial cube,
which are not eliminated in the coarsening procedure, have exactly the
same position as in the initial mesh. Then, if the centre of a tetrahedron
is outside the cube, we translate the tetrahedron into the cube. This
operation requires checking whether each node is already present in the
mesh. This can be time consuming with meshes containing more than
one hundred thousand nodes. A k-d Tree (Appendix C) is therefore
constructed in order to quickly check the existence of a node. Each
entry of the k-d Tree is composed of the node index and of the three
node coordinates. The shape of the reconstructed microstructure is not
exactly a cube but very close to it (Fig. 5.12). Other node sets than
the six facets of the cube must be constructed in order to impose the
periodic boundary conditions. There are twelve such supplementary sets
that correspond to the twelve edges of the cube. These sets are composed
of nodes that have a single periodically equivalent neighbour lying at
the opposite, along a diagonal of the cube.
120 CHAPTER 5. Construction of model polycrystals
Figure 5.11: Mesh after the coarsening procedure and before the post
processing.
(a) Initial mesh before coarsening (b) Mesh after coarsening and post
processing
Figure 5.12: Meshes before and after the coarsening procedure and the
post processing.
A second post processing of the mesh transforms rst-order elements
into second-order elements. This post processing method has been de-
5.2. Mesh generation algorithm 121
signed in order to work with a large range of meshes. It takes on input
several .inp les where .inp is the format of the input le of Abaqus. A
rst le must contain all the node positions. A second one must con-
tain the connectivity of all elements. And the last one must contain the
node sets used to prescribe boundary conditions. We iterate through
the six edges of all tetrahedra and we add second-order nodes to all the
edges that are visited for the rst time. If the two nodes of an edge
belong to a node set, the second-order node is added to the node set.
All second-order nodes that belong to an edge for which the two nodes
do not belong to the same set are recorded. At the end, for each of these
recorded nodes, we iterate through all the nodes in order to nd their
periodic equivalents and then to add the pair in the appropriate node
sets.
5.2.6 Parametric study of the coarsening procedure
Whereas all parameters of the microstructure generation, except the
total number of grains forming an RVE which is discussed in the next
section, are obtained from experimental measurements, the coarsening
procedure introduces purely numerical parameters. In this section, let us
study the inuence of these parameters on the reduction of the number of
elements and on the mesh quality. The parameters that will be studied
are mainly L
prop
and the minimum quality required for an element of
the mesh. I will also show the inuence of the layer thickness and of the
initial number of elements per grain.
The initial mesh contains ten grains and approximately 50000 ele-
ments. 30% of the nodes belong to a grain boundary. Five selectively
rened meshes have then been created from the same initial mesh with
values of the ratio
L
GI
L
GB
ranging from 2 to 200. The ve meshes are
obtained with a layer thickness equal to
1
20
of the length of the cube.
If we consider that the average grain shape is a sphere of radius R,
the layer thickness is equal to approximately
1
6
R. The number of el-
ements obtained after the coarsening procedure and a measure of the
mesh quality are given in Figure 5.13 with the black line. One observes
that the number of elements begins to decrease very slowly when the ra-
tio
L
GI
L
GB
40. We cannot eliminate more than one half of the elements
with the coarsening procedure for this specic initial mesh with the cho-
sen layer thickness. It appears that the mesh quality steadily decreases
with the increase of the ratio
L
GI
L
GB
. This decrease seems proportional to
122 CHAPTER 5. Construction of model polycrystals
the decrease of the number of elements.
If the mesh quality needs to be improved, the only tuneable param-
eter is Q
min
, the minimum quality of all elements which is computed
as:
Q
min
= 15552
V
2
(

i
l(e
i
)
2
)
3
(5.14)
where V is the volume of the tetrahedron and l(e
i
), the length of the edge
i. The value 15552 ensures that Q
min
ranges between 0 and 1. I have
created a new set of rened meshes obtained from the same initial mesh
with Q
min
= 0.15, the Q
min
value in the initial mesh. The evolutions
of the number and of the quality of elements are shown in Figure 5.13
with the blue line. With this higher value of Q
min
, selective coarsening
reduces only by 20% the total number of elements. The mean element
quality also evolves proportionally to the decrease of the number of
elements. The factor of proportionality seems to be the same as for the
black line.
In order to check this, three meshes obtained from the same initial
mesh but for three values of Q
min
have been created by choosing dierent
values of the ratio
L
GI
L
GB
so that the same number of elements is obtained
after coarsening. It can be observed in Table 5.4 that the value of the
minimum quality of the elements does not aect the global quality of
the mesh for a similar total number of elements. In fact, it only aects
the distribution of the element quality values.
Q
min
.01 0.1 0.2
NbElement 45697 45740 46010
AverageQuality 0.68 0.665 0.66
L
GI
L
GB
3.7 5 40
Table 5.4: Comparison of the mesh quality for three dierent values
of Q
min
, the minimum quality of all elements. The three meshes are
obtained from the same initial mesh and with values of the ratio
L
GI
L
GB
leading to approximately 46000 elements. L
GI
and L
GB
are the average
element size, respectively, in the grain interior and in the grain boundary
region.
Finally, we have created a set of coarsened meshes where the layer
thickness is equal to 0. The nal number of elements and the mesh
quality are evaluated in Figure 5.13 with the red line. As expected,
5.2. Mesh generation algorithm 123
(a) Number of elements evolution
(b) Mesh quality evolution
Figure 5.13: Evaluation of the meshes created by selective coarsening
of the grain interiors. a) Reduction of the number of elements in the
mesh. b) Lowering of the average element quality. L
GI
and L
GB
are
the average element size, respectively, in the grain interior and in the
grain boundary region. LT is the layer thickness and Q
min
the minimum
quality imposed for an element.
124 CHAPTER 5. Construction of model polycrystals
when the layer thickness is nil, coarsening is facilitated and 10% more
elements can be eliminated. This is however correlated with a decrease
of the mesh quality.
A much more rened initial mesh has been created in order to study
the inuence of the initial renement on the coarsening procedure. The
new mesh has approximately 230000 elements and 20% of the nodes
lie on a grain boundary. The proportion of nodes on grain boundaries
thus decreases with the increased renement and one expects a greater
reduction of the number of elements e.g. when the layer thickness is
equal to 0. The quality of the mesh and the nal number of elements
are presented in gure 5.14. It appears that the nal number of elements
evolves approximately in the same way for the two initial meshes except
for the largest values of the ratio
L
GI
L
GB
. However, the evolution of the
mean quality of the elements is completely dierent. The most rened
mesh, with a ratio
L
GB
D
equal to 0.05, has a much larger mean quality,
especially for the largest values of the ratio
L
GI
L
GB
.
These statistics show that the mesh quality is a direct function of the
required size eld, i.e. the value of the ratio
L
GI
L
GB
, and of the boundary
proportion which is the ratio between the initial number of nodes in a
region that must be coarsened and the initial number of nodes in the
mesh. The larger the prescribed maximum characteristic length, the
lower the mesh quality. The larger the boundary ratio, the lower the
mesh quality. Therefore, in order to improve the quality of the mesh for
a given number of elements, we aim to reach this number of elements
for the lowest value of the ratio
L
GI
L
GB
. When the width of the layer is
equal to 0, the boundary ratio is directly function of the number of
nodes on grain boundaries. Therefore, smoothing of grain boundaries
performed after the microstructure generation is important for the nal
quality of the coarsened mesh.
With the proposed coarsening procedure, it is possible to selectively
rene meshes. A reduction of 50% of the number of elements can be
reached. But the decrease of the number of elements is correlated with
a decrease of the mesh quality. Increasing the minimum quality of ele-
ments does not change the average mesh quality but changes the distri-
bution of the element quality values. It also decreases the eciency of
the coarsening procedure.
5.2. Mesh generation algorithm 125
(a) Number of elements evolution
(b) Mesh quality evolution
Figure 5.14: Evolution of the mesh quality and the remaining number
of elements in function of the ratio
L
GI
L
GB
for two initial meshes having
respectively approximately 50000 and 230000 elements leading to a ratio
L
GB
D
equal to respectively 0.08 and 0.05. L
GI
and L
GB
are the average
element size, respectively, in the grain interior and in the grain boundary
region.
126 CHAPTER 5. Construction of model polycrystals
Chapter 6
Assessment of the new
CPFEM
In this chapter, we investigate the inuence of the model microstructure
on the prediction of the mechanical response. The inuence of the mi-
crostructure representation and the renement of FE mesh are studied
separately. The various parameters involved in the mesh generation are
the type and the interpolation order of elements, the number of gauss
points and the coarsening of selected regions. We hope to nd a com-
bination of parameters that allows accurate nite element simulations
with a moderate computational cost.
6.1 Optimal size of the RVE
One question that arises when we generate a model microstructure is
: how many grains are needed in order to create a statistically rep-
resentative volume element? In this section, we try to answer to this
question and show that it depends on the discretisation technique and
on the micro-macro scheme. Somes parts of the text present in this
section come from a publication [51].
The planar anisotropy of the IF steel sheet, having the texture pre-
sented in gure 5.2a, is simulated using two kinds of model microstruc-
tures, a regular and an irregular one. Predicted R-values are recorded
127
128 CHAPTER 6. Assessment of the new CPFEM
after 5% uniaxial elongation. In order to assess the extent to which
model predictions are aected by the sampling of initial grain orienta-
tions, each simulation is repeated ten times. The ten simulations all
rely on the same model but the lattice orientation of individual grains
is dierent each time. The texture discretisation algorithm ensures that
the global texture is always nearly the same (Fig. 5.2).
A rst set of ten simulations relies on the FE mesh shown in Fig.
6.1c. The idealised microstructure consists of 8 8 8 = 512 grains.
Each grain is a cube subdivided into 24 tetrahedral nite elements with
quadratic shape functions (four integration points). Initial grain orien-
tation samplings are generated by applying the discretisation algorithm
of Toth and Van Houtte (Section 5.1.1) ten times. Then, a second set
of ten simulations is performed, considering irregular microstructures
such as the FE meshes shown in Fig. 6.1a,b. The number of grains is
unchanged, 512, but the grains have a non-uniform shape and size. An
average of 24 tetrahedral nite elements is used within every grain. The
new algorithm for texture discretisation is used for assigning initial lat-
tice orientations based on a list of the grain volumes in each simulation.
Material parameters have identical values in all simulations. Plastic
deformation is achieved by dislocations slip along 110 < 111 > and
112 < 111 > slip systems. The initial value of the critical resolved
shear stress (CRSS),
c
, is unique over all grains, but becomes dierent
with the deformation, and a Voce hardening law is used :

c
= H
0

sat

sat

c0

[ (6.1)
The values of the hardening parameters are :
c0
= 190,
sat
= 380,
H
0
= 38. Elastic constants are set as follows: C
11
= 231.4GPa, C
12
=
134GPa and C
44
= 116GPa. Dislocation slip rates are calculated based
on a viscoplastic exponential law [33] where
0
= 0.001/s and m = 100.
The upper and lower bounds of the R-values predicted in the two
sets of 10 simulations are presented in gure 6.2. One observes that the
spread between the bounds is about three times less when the 512 grains
have heterogeneous grain sizes.
The twenty samplings of initial grain orientations are used a second
time in simulations based on the Taylor model. The upper and lower
bounds of the predicted R-values are shown in gure 6.3. This time, the
6.1. Optimal size of the RVE 129
Figure 6.1: Mesh representing model microstructures with a-b) irregular
grain shape and size and with c) a regular brick shape. The total number
of nite elements is always close to 24000.
130 CHAPTER 6. Assessment of the new CPFEM
0 15 30 45 60 75 90
1.5
1.7
1.9
2.1
2.3
2.5
Angle of the tensile direction relative to the rolling direction ()
R
v
a
l
u
e
Max Bound HE
Min Bound HE
Max Bound HO
Min Bound HO
Figure 6.2: Illustration of the variability of the R-Values predicted in ten
simulations relying on dierent initial orientation samplings. The nite
element meshes consist of 512 grains with regular (HO) or irregular (HE)
grain shape and size
512 grains undergo uniform deformation. At each time step, a Newton-
Raphson scheme is used in order to identify the values of transverse
strain rates
22
and
33
that correspond to uniaxial tension along x
1
,
i.e. <
22
>=<
33
>=0 where < x > means polycrystalline average
of x. The spread between the upper and lower bounds is again much
less when the 512 grains have distinct volumes. In order to reduce the
spread between the bounds of predictions assuming a uniform grain size,
the number of grains included in the RVE must be raised to 10000 (Fig.
6.4). Supplementary to the spread, the average R-Value predicted in
each direction is identical in the two cases.
When simulations rely on the Taylor model, the stress developed
within each grain depends solely on the lattice orientation of the grain.
This is not true in FE simulations due to the interaction of adjacent
grains. The eect of grain interaction on predicted R-values is investi-
gated using the mesh with cuboidal grains (Fig. 6.1c). In gure 6.5,
the upper and lower bounds result from ten simulations corresponding
to ten permutations of the same set of 512 initial orientations. Hence,
the grains have dierent neighbours in each simulation but the global
texture is exactly the same. As a consequence of grain interactions,
6.1. Optimal size of the RVE 131
0 15 30 45 60 75 90
1.5
1.7
1.9
2.1
2.3
2.5
2.7
Angle of the tensile direction relative to the rolling direction ()
R
v
a
l
u
e
Max Bound HE
Min Bound HE
Max Bound HO
Min Bound HO
Figure 6.3: Illustration of the variability of the R-Values obtained by
Taylor simulation for homogeneous (HO) and heterogeneous (HE) grain
size distribution. 512 grains are used in each simulation.
0 15 30 45 60 75 90
1.5
1.7
1.9
2.1
2.3
2.5
2.7
Angle of the tensile direction relative to the rolling direction ()
R
v
a
l
u
e
Max Bound 512
Min Bound 512
Max Bound HO 10000
Min Bound HO 10000
Figure 6.4: Illustration of the variability of the R-Values obtained by
Taylor simulation for 10000 grains of homogeneous size (HO). Results
already presented in gure 6.3 (HE) are shown for comparison.
132 CHAPTER 6. Assessment of the new CPFEM
predicted R-values vary between bounds that are independent of the
accuracy of the texture discretisation procedure.
0 15 30 45 60 75 90
1.5
1.7
1.9
2.1
2.3
2.5
Angle of the tensile direction relative to the rolling direction ()
R
v
a
l
u
e
Max Bound HE
Min Bound HE
Max Bound Permutation
Min Bound Permutation
Figure 6.5: Illustration of the variability of the R-Values obtained after
FE simulations on ten topological permutations of the same orientation
sampling. Results already presented in gure 6.2 (HE) are shown in
comparison.
In gure 6.6, the average R-values predicted in each set of ten sim-
ulations is compared to experimental measurements. Except for the
tensile test performed at 90 relative to the rolling direction, FE mod-
elling yields more accurate predictions of the R-values than the Taylor
model. Applying FE modelling to regular and irregular microstructures,
respectively, leads to very similar predictions of the R-values.
Based on the nite element simulations presented in the present sec-
tion, it appears that steel sheets with a cold rolling texture are repre-
sented satisfactorily by a RVE comprising 512 grains with heterogeneous
size. The generation of a FE mesh containing twice as many grains does
not help reducing the variability of the R-values predicted in successive
runs. Indeed, as shown in gure 6.5, the spread between the upper and
lower bounds of the predictions is mostly due to the topological arrange-
ment of the grains, not to an inaccurate texture discretisation. In order
6.1. Optimal size of the RVE 133
0 15 30 45 60 75 90
1.5
1.7
1.9
2.1
2.3
2.5
2.7
Angle of the tensile direction relative to the rolling direction ()
R
v
a
l
u
e
Experimental
TaylorHE
FE (HE)
FE (HO)
Figure 6.6: Evaluation of the model predictions by comparison to the
experimental R-Values [61]. Numerically predicted curves correspond to
the average of the simulations.
to lower the sensitivity of the macroscopic response to local grain inter-
actions, the size of the RVE should be much larger than what is strictly
needed in order to reproduce texture. Nevertheless, if one accepts 5%
variability of predicted R-values, it is sucient to consider 512 grains.
In this case, each nite simulation of the uniaxial tensile test takes 30
minutes when running ABAQUS on a Pentium IV - 3GHz.
As illustrated in gure 6.6, a discrepancy remains between the predic-
tions and the experimental R-values. One probable reason for this gap
is the assumption of isotropic hardening within each grain. According to
microstructural observations of deformed samples, dislocation walls are
created within the grains, and dislocation slip is hindered across them.
Several authors simulate this phenomenon based on an anisotropic hard-
ening of the slip systems [74, 75, 76].
134 CHAPTER 6. Assessment of the new CPFEM
6.2 Convergence of CPFEM with mesh rene-
ment
Unfortunately, in typical CPFEM study, a negligible improvement in the
accuracy of the predictions implies a non-negligible increase of the com-
putational cost. The results obtained by FEM simulations are improved
when the mesh is rened. They tend to converge asymptotically towards
an optimal prediction. Once we have found the converged results, we
dene an optimal number of gauss points per grain, which allow us to
have a suciently accurate prediction in a moderate computational
time.
6.2.1 Description of the reference simulation
The same reference simulation is performed throughout this section.
This simulation is a 5% elongation uniaxial tensile test on a sheet sample
with bcc crystal symmetry. Dislocations move along 110 < 111 > and
112 < 111 > slip systems. The critical resolved shear stress (CRSS),

c
is identical on all slip systems (isotropic hardening) and its initial
value is equal to 200MPa. A Swift law is used for the hardening.
0
is equal to 0.15 and n to 0.203. When planar anisotropy is studied,
the tensile test is performed along various directions in the plane of the
sheet, i.e. for dierent angles relative to the rolling direction.
The nite element mesh is deformed under periodic boundary condi-
tions. Both regular and irregular grain shapes are considered. The mesh
renement study is conducted while maintaining the geometry of the in-
creasingly rened microstructures as constant as possible. In irregular
model microstructures, this implies controlling the seeds positioning and
not allowing grain boundary smoothing.
The rst aggregate treated contains ten equiaxed grains. Thirteen
unstructured meshes with uniform (but dierent) characteristic lengths
are generated to reproduce the unique microstructure. The number of
elements in these meshes is respectively equal to 1792, 5040, 8344, 11632,
16720, 22608, 34440, 54648, 106288, 159604, 251168, 431696 and 837064
(Fig. 6.7). Grain volumes are very similar despite the large dierences
in the number of elements (Table 6.1).
6.2. Convergence of CPFEM with mesh renement 135
Figure 6.7: Meshes with 1792 and 54648 elements of the same periodic
arrangement of 10 grains.
N V1 V2 V3 V4 V5 V6 V7 V8 V9 V10
1792 10.3 9.8 8.8 11.7 10.2 10.0 8.7 11.0 10.6 8.7
5040 10.3 9.9 8.9 11.6 10.1 10.0 8.6 10.7 10.9 8.9
8344 10.3 9.9 9.0 11.5 10.2 10.0 8.6 10.7 10.9 8.9
11632 10.3 9.9 8.9 11.6 10.2 9.8 8.6 10.7 11.0 9.0
16720 10.4 9.8 9.0 11.6 10.0 9.8 8.6 10.8 10.9 9.0
22608 10.3 9.8 9.1 11.6 10.0 9.9 8.6 10.8 11.0 9.0
34440 10.5 9.9 8.9 11.6 10.0 9.8 8.6 10.8 10.9 9.0
54648 10.4 9.8 9.0 11.6 10.0 9.9 8.6 10.8 10.9 9.0
106288 10.4 9.8 8.9 11.6 10.1 9.9 8.6 10.8 10.9 9.0
159064 10.4 9.9 9.0 11.6 10.1 9.9 8.6 10.8 10.9 9.0
251168 10.4 9.8 9.0 11.6 10.1 9.9 8.6 10.8 10.9 9.0
431696 10.4 9.8 9.0 11.6 10.1 9.9 8.6 10.8 10.9 9.0
837064 10.4 9.8 9.0 11.6 10.1 9.9 8.6 10.8 10.9 9.0
Table 6.1: Grain volume fractions in the thirteen FE representations of
the microstructure.
The reference simulation is performed two times on each microstruc-
ture, once with rst-order elements and once with second-order elements.
As we can see in gure 6.8, the number of nodes and the computation
time of simulations using the iterative FE solver evolve linearly with the
number of gauss points. The computation time of simulations using the
direct FE solver evolves quadratically with the number of gauss points.
So, from now on, all the simulations are performed with the iterative
136 CHAPTER 6. Assessment of the new CPFEM
FE solver.
The iterative solver becomes faster than the direct solver for a lower
total number of gauss points which is smaller in the case of second-order
simulations than in the case of rst-order simulations. One explanation
is the fact that the bandwidth, which is the number of non-zero en-
tries in a row, of the matrix constructed in the nite element problem
is smaller in rst-order simulations. Indeed, in the second-order sim-
ulations, each rst-order and second-order node of an element appear
in the same equation. Another explanation is that the preconditioner
used for the iterative solver is more ecient for second-order simulations.
We can see in gure 6.9 that a same number of gauss points leads to
a larger number of nodes in second-order simulations. It was expected
as the number of gauss points is multiplied by four when second-order
elements are used instead of rst-order elements. As the number of
edges, and therefore the number of second-order nodes added, is greater
than three, the number of nodes is multiplied by more than four. For a
similar total number of gauss points, the computation time is faster for
simulations with second-order elements than with rst-order elements
even with the larger number of nodes.
6.2.2 Mesh renement adapted to the simulation goal
As illustrated in this section, convergence of CPFEM results with mesh
renement is faster for certain mechanical properties than others. Five
such properties are considered here. At the end of each simulation, we
record : the variance of
c
, the average in-grain misorientation angle,
which is the average angle between any integration point and the mean
orientation of a grain, the Lankford coecient, the macroscopic value of

11
and W. The latter is the dierence between external and internal
works where the external work is equal to the product of the force applied
on the master node of the microstructure and the displacement of this
node, and the internal work is equal to the weighted sum of
_
d on all
gauss points.
We want to know the optimal number of gauss points per grain for
each of these properties. If the main goal of a simulation is to compute
the stored energy in each grain, the emphasis is put on an accurate pre-
diction of the evolution of
c
within the grains. In a similar way, the
emphasis will be put on the misorientation angle when one characterises
6.2. Convergence of CPFEM with mesh renement 137
(a)
(b)
Figure 6.8: Evolution of the number of nodes and of the computation
time in function of the total number of gauss points for (a) rst- and (b)
second-order elements.
138 CHAPTER 6. Assessment of the new CPFEM
Figure 6.9: Evolution of the number of nodes and of the computation
time using the iterative solver in function of the total number of gauss
points. Comparison of rst- and second-order simulations.
grain subdivision, on the Lankford coecient (also called R-value) when
one aims to predict the height of deep drawing ears and on the macro-
scopic value of
11
when one predicts the stress-strain curve. The dif-
ference between external and internal work is a common measure of the
quality of the nite element prediction.
Predictions of these ve properties for the two sets of simulations
(rst- and second-order) are shown in gure 6.10, 6.11, and 6.12. It
appears that even for a very large number of gauss points per grain,
convergence is not reached for simulations with rst-order elements. The
convergence is nearly obtained for the macroscopic value of
11
and
the dierence between the external and the internal work. Microscopic
properties, such as the variance of
c
and the misorientation angle, are far
from a converged value. These observations were expected. The Abaqus
documentation warns the user that rst-order tetrahedral elements are
overly sti and exhibit slow convergence with mesh renement and must
therefore be avoided.
From now, when tetrahedral elements are used, second-order inter-
polation is relied upon. In such case, we observe that 2000 gauss points
6.2. Convergence of CPFEM with mesh renement 139
(a)
(b)
Figure 6.10: In (a) (resp. (b)), the average of
11
(resp. W) is divided
by its converged value obtained on the nest mesh. The stellar and
circled line correspond, respectively, to rst- and second-order elements.
The squared line is obtained by dividing the rst-order prediction by
the converged value of second-order simulation.
140 CHAPTER 6. Assessment of the new CPFEM
(a)
(b)
Figure 6.11: In (a) (resp; (b)), the average in-grain misorientation angle
(resp. the variance of
c
) is divided by its converged value obtained on
the nest mesh. The stellar and circled line correspond, respectively,
to rst- and second-order elements. The squared line is obtained by
dividing the rst-order prediction by the converged value of second-
order simulation.
6.2. Convergence of CPFEM with mesh renement 141
Figure 6.12: Lankford coecient divided by its converged value obtained
on the nest mesh. The stellar and circled line correspond, respectively,
to rst- and second-order elements. The squared line is obtained by
dividing the rst-order prediction by the converged value of second-
order simulation.
per grain, which corresponds to approximately 800 nodes per grain, are
required in order to have predictions of the Lankford coecient and
the macroscopic value
11
which are very close to the converged value.
For the more microscopic properties, the same quality of prediction is
obtained if more than 8000 gauss points per grain, which corresponds
to approximately 3200 nodes per grain, are used. It was expected that
the convergence of macroscopic properties requires a lower renement.
Indeed, as macroscopic properties are computed as an average of local
eld, local errors have only a second-order inuence. On the opposite,
as microscopic properties are computed as a variance over a local eld,
local errors have a rst-order inuence. In the remainder of the section,
we focus on predictions of Lankford coecients, the optimal renement
of which is further discussed.
142 CHAPTER 6. Assessment of the new CPFEM
6.2.3 Inuence of the polycrystal size
It must be emphasised that ten grains do not constitute a valid RVE
(following the discussion of Section 6.1) whatever the texture under con-
sideration. Such reduced grain sampling was used in order to allow ex-
tremely rened meshes with aordable computational cost. However,
the convergence study turns out to be dependent on the number of
grains included in the model microstructure. Indeed, averaging out lo-
cal errors is more ecient when the number of grains increases. To
illustrate this, the reference simulation is performed a third time on ve
microstructures, composed again of ten grains and discretised into, re-
spectively, 1792, 5040, 16720, 34440 and 54648 elements. We can see
in gure 6.13a that the optimal number of gauss points per grain is ap-
proximately equal to 14000, which corresponds to 5000 nodes per grain.
The quality of the prediction is dierent from one angle to the other,
especially for meshes with low renement. A mesh with approximately
7000 gauss points per grain allows us to obtain a prediction which is
1% from the converged solution. We have then created a new set of mi-
crostructures containing 50 grains. Indeed, 50 grains allow a qualitative
reproduction of the main components of typical textures. In order to
have approximately the same number of gauss points per grain as the
microstructure with 10 grains, three microstructures containing respec-
tively 667, 1808 and 6277 gauss points per grain have been created. In
gure 6.13b, the convergence of simulations relying on respectively 10
and 50 grains are compared. The number of gauss points per grain was
controlled with care: respectively, 717, 2016 and 6688 in the simulations
with 10 grains. Hence, for roughly the same number of gauss points,
predictions are approximately three times more accurate with 50 grains.
As the dierence between the predictions obtained with 700 gauss points
and the converged results is not larger than 2%, such renement seems
optimal with 50 grains.
6.2.4 Inuence of the type of nite elements
In the literature, it is common to nd CPFEM studies relying on hex-
ahedral nite elements instead of tetrahedral elements [31]. The use
of hexahedral elements is, however, rather complicated when the grain
topology is irregular (unless one relies on rened voxels [68]). In order
to compare the two types of elements, we consider structured meshes
of model microstructures with uniform grain shape and size. The mi-
6.2. Convergence of CPFEM with mesh renement 143
(a)
(b)
Figure 6.13: Comparison between the prediction of the Lankford coef-
cient for various meshes and various angles relative to the rolling
direction. All predictions are divided by their converged value. In (a),
the microstructure is composed of 10 grains. In (b), results shown with
the dotted line are obtained with 50 grains whereas the other ones are
obtained with 10 grains.
144 CHAPTER 6. Assessment of the new CPFEM
crostructures are composed of 8 brick-shaped grains and each grain is
represented either by a set of hexahedral elements (Fig. 6.14a) or sub-
divided in 40 tetrahedral elements (Fig. 6.14b).
Figure 6.14: Elementary meshes composed of 8 grains. On the left, each
grain is represented by a single hexahedral element. On the right, each
grain is represented by 40 tetrahedral elements.
When rst-order hexahedral elements are used, the number of gauss
points per grain of the various meshes is respectively equal to 216, 1000,
4128, 8000 and 17576. We can see in gure 6.15a that the optimal num-
ber of gauss points is approximately equal to 8000, which corresponds
approximately to 1150 nodes per grain. When second-order hexahedral
elements are used, the number of gauss points per grain of the various
meshes is respectively equal to 729, 1728, 5832, 9261 and 19683. We can
see in gure 6.15b that the optimal number of gauss points is approxi-
mately equal to 9000, which corresponds approximately to 1150 nodes
per grain.
When bricks are subdivided into 40 second-order tetrahedral ele-
ments, the number of gauss points per grain of the various meshes is
respectively equal to 160, 1269, 4320, 10240 and 20000. We can see
in Fig. 6.16a that the optimal number of gauss points is approximately
equal to 4000, which corresponds approximately to 2000 nodes per grain.
The inuence of the number of grains on the optimal number of gauss
points is lower for structured meshes (Fig. 6.16b) than for unstructured
meshes (Section 6.2.3). This may be explained by the fact that grains
6.2. Convergence of CPFEM with mesh renement 145
(a)
(b)
Figure 6.15: Comparison between the prediction of the Lankford coe-
cient obtained for various mesh renements and various angles relative
to the rolling direction. All predictions are divided by their converged
value. Elements used for the simulation are (a) rst- and (b) second-
order hexahedral elements. Microstructures are composed of 8 grains
with BCC slip.
146 CHAPTER 6. Assessment of the new CPFEM
(a)
(b)
Figure 6.16: Comparison between the predictions of the Lankford coe-
cient obtained for various mesh renements and various angles relative
to the rolling direction. All predictions are divided by the prediction of
the Lankford coecient obtained on the most rened mesh which con-
tains 4320 gauss points per grain. The number of grains is (a) 8 and (b)
64.
6.2. Convergence of CPFEM with mesh renement 147
volumes are constant, and thus independent of the renement, in the
case of structured meshes. As the dierence between the prediction ob-
tained with 1260 gauss points per grain and the converged results is less
than 5% in the case of second-order tetrahedral elements 6.16b, such
renement seems appropriate.
It is observed in gure 6.17 that the converged results obtained with
the dierent types of elements are roughly equivalent. The slight dier-
ence between rst-order hexahedral elements and other elements can be
explained by the slightly lower number of gauss points per grain used
in this case. As the choice of elements inuences the results, we may
rightfully use second-order tetrahedral elements in the remainder of the
study.
Figure 6.17: Comparison between the converged prediction of the Lank-
ford coecient for various angles relative the rolling direction and for
three types of elements.
6.2.5 Inuence of the crystal symmetry (BCC vs. HCP)
It is well-known that crystals with hexagonal close packed (HCP) sym-
metry show a greater anisotropy than crystals with BCC symmetry (e.g.
148 CHAPTER 6. Assessment of the new CPFEM
[78]). Simulations have been performed to assess how this might inu-
ence the convergence of CPFEM with mesh renement. For the HCP
simulation, plastic deformation is achieved by dislocations slip along
prismatic slip (1010 < 1210 >), basal slip (0001< 2110 >)
and pyramidal < c + a > slip (10 11 < 1 123 >). The critical
resolved shear stress,
c
is unique within all grains and a Swift law is
used (Eq. 3.23). The elastic constants have the following value [GPa]
: C
11
= 162.6, C
12
= 0.0256, C
13
= 50.8, C
33
= 62.3 and C
44
= 18.9.
Dislocation slip rates are calculated based on a viscoplastic exponential
law [33]. The ratio between the CRSS of prismatic slip and basal slip is
equal to 15, the ratio between the CRSS of pyramidal slip and basal slip
is equal to 10. The value of
0
and n are the same for each slip systems
and are respectively equal to 0.1 and 0.2.
Relying on second-order tetrahedral structured meshes, we can see in
gure 6.18a that the optimal number of gauss points per grain is higher
than 10000. The fact that hexagonal metals require more rened meshes
than BCC metals in order to reach convergence was expected. Indeed,
due to the greater anisotropy of individual grains, hexagonal materi-
als exhibit larger intra-grain heterogeneity than BCC materials. Each
grain must therefore be discretised in a larger number of elements. The
computation time is also compared for the two materials (Fig. 6.18b).
We can see that the computation time is approximately 8 times larger
for hexagonal materials with exactly the same microstructures. This is
explained in part by the fact that integration of the crystal plasticity
law is slightly slower and that the FE solution requires more iterations
(due to the heterogeneity of the strain eld).
6.2.6 Inuence of the positioning and shape of grains
Thanks to the new generator of unstructured meshes, grain centres can
be positioned in a more or less regular way. In order to study the inu-
ence of the regularity of the seeds positioning, small random variations
are added to the positions of the grain centres. Structured and unstruc-
tured meshes are compared too, using the same lattice orientations in
all microstructures. All meshes are thus slightly dierent discretisations
of the same reference microstructure. Replacing the structured mesh
(Fig. 6.19a) by an unstructured mesh (Fig. 6.19b) induces a small vari-
ance of the grain volumes (3.710 5). The biggest dierence with the
structured mesh is the grain geometry, which is not smooth. When the
6.2. Convergence of CPFEM with mesh renement 149
(a)
(b)
Figure 6.18: (a) Comparison between the predictions of the Lankford
coecient obtained for BCC and for HCP crystal symmetry and various
angles relative to the rolling direction. All predictions are divided by
their converged value. The elements used for the simulation are second-
order tetrahedral elements. (b) Comparison between the computation
times with BCC and HCP crystal symmetry.
150 CHAPTER 6. Assessment of the new CPFEM
Figure 6.19: Microstructures composed of 27 grains having the same
orientations. a) Regular microstructure. b) Irregular microstructure
where the positions of the seeds are equal to the positions of the grain
centres in a). c) Irregular microstructure where the positions of the
centres are obtained by a perturbation of the positions of the grain
centres in a).
6.2. Convergence of CPFEM with mesh renement 151
seeds positions are slightly shued, the variance of the volumes is larger
(1.310 3) Finally, other non-regular microstructures are generated by
imposing a seed positioning, which is obtained as a weighted average
of the seed positioning of the regular microstructure and the irregular
microstructure, to the generator.
We can see in gure 6.20 that for a similar seed positioning, the dif-
ference in the mesh has only a negligible inuence on the prediction. The
inuence of perturbations of the seed positioning is approximately ten
times bigger. As expected, the maximum dierence between predictions
decreases when the amplitude of the perturbation is decreased.
Figure 6.20: Comparison between various discretisations of the same
microstructure. Seed positioning of microstructures constructed on the
unstructured mesh is a weighted average between the positioning of grain
centres of the regular microstructure and a perturbation of this position-
ing. The weight of the perturbation is equal to 0 for RegularUnstruc-
tured and to 1 for IrregularUnstructured
Conclusion of section 6.2 We have shown that rst-order tetrahe-
dral elements converge towards a result, which is not equal to the con-
verged results obtained by second-order tetrahedral elements. We have
152 CHAPTER 6. Assessment of the new CPFEM
found that the optimal number of gauss points per grain is equal to
14000 for irregular microstructures and to 4000 for regular microstruc-
tures when a low number of grains is used. Our explanation of this
dierence is that in the case of irregular microstructure, the conver-
gence of the geometry is added to the convergence of the discretisation
of this geometry. We have shown that the convergence is much better for
unstructured meshes when the number of grains is increased. In the case
of structured meshes, the convergence is only slightly better. We have
shown that second-order tetrahedral elements, rst- and second-order
hexahedral elements converged towards the same result on a structured
mesh. We have shown that the main dierence between the predictions
of structured and unstructured grains does not come from the discreti-
sation of the grain geometry but from the dierence in the volumes of
the grain.
6.3 The benet of selective mesh coarsening
The goal of selective mesh coarsening is to reduce the computational
cost of the simulation with an as limited as possible detrimental eect
on the accuracy of the predictions. We will assess this, on the one hand,
based on the volume average of the error on the displacement eld;
on the other hand, based on the predicted variance of the strain eld.
Indeed, as already mentioned in Section 6.2.3, predictions of stored en-
ergy throughout the aggregate or of grain subdivision require accurate
simulation of not only the average macroscopic trends but also the het-
erogeneity of displacement (or rather strain) within individual grains.
This is estimated here based on the overall variance of the strain tensor
components. The latter tensor is dened as:
LE
ij
=
_
t
t
0
D
ij
dt (6.2)
The reference case study is identical to that of the previous section: a
uniaxial tensile test with 5% total elongation on a polycrystal with BCC
symmetry. The tensile direction is x
1
. Most of the meshes used in this
section have been presented in Section 5.2.6. They represent a periodic
arrangement of 10 grains. The most rened mesh, containing approxi-
mately 50000 elements, is used as the reference solution throughout the
section.
6.3. The benet of selective mesh coarsening 153
6.3.1 Trends in the displacement eld
Meshes with homogeneous characteristic length The average
displacement error is the mean value, computed over all nodes, of the
dierence between the displacement of a node in the reference mesh
and the displacement interpolated at the same coordinates in the less
rened mesh. For each node of the reference mesh, we look for the
elements of the compared mesh which contain the node coordinates.
Then node coordinates are computed in the parent space of this element
and the displacement is interpolated from the displacement of the nodes
of this element. Finally, the dierence between the interpolated value
and the value predicted on the reference mesh is normalised. We can
see in gure 6.21a that the error of the displacement eld decreases with
mesh renement. The error on the geometry, i.e. the volume fraction of
elements of the rened mesh which no longer belong to the right grain,
decreases accordingly (Fig. 6.21b).
Selectively coarsened meshes The computation of the displace-
ment error for coarsened meshes requires a more complex algorithm than
in the case of uniformly rened meshes. Indeed, due to the periodicity
requirement during mesh adaptation, some elements of the coarsened
mesh are partly outside the initial cube. There are also some voids in
the cube geometry which are lled with elements lying on the opposite
periodic boundary. Therefore, some nodes of the reference mesh are not
automatically within an element of the selectively coarsened mesh. All
27 periodic equivalents of the coordinates are computed until we nd
an element which contains these coordinates. Then, the displacement
of the node on the reference mesh and the displacement interpolated in
the mesh selectively rened cannot be compared directly. The relative
displacement between every boundary and its periodic counterpart must
be taken into account.
Comparisons are made for meshes containing one half less elements
than the initial mesh (Fig. 6.22a). Therefore new uniformly coarsened
meshes, also containing 10 grains, have been generated in order to have
a number of elements which is close to the number of elements of the se-
lectively coarsened meshes, respectively equal to 6335, 12158 and 26089.
The error on the geometry is also presented (Fig. 6.22b). Each uniformly
rened mesh is used as initial mesh for the next coarsening. We can see
that the error decreases when the mesh is rened even with selectively
154 CHAPTER 6. Assessment of the new CPFEM
0 1 2 3 4 5 6 7 8
x 10
4
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
x 10
3
Number of nodes
E
r
r
o
r

o
n

t
h
e

d
i
s
p
l
a
c
e
m
e
n
t

f
i
e
l
d
(a) Error on the displacement eld
0 1 2 3 4 5 6 7 8
x 10
4
0
0.05
0.1
0.15
0.2
0.25
Number of nodes
E
r
r
o
r

o
n

t
h
e

g
e
o
m
e
t
r
y
(b) Error on the grain geometry
Figure 6.21: Evolution of the error on the displacement eld and the er-
ror on the grain geometry (as dened in the text) in function of the num-
ber of elements of meshes with an homogeneous characteristic length.
6.3. The benet of selective mesh coarsening 155
0.5 1 1.5 2 2.5 3
x 10
4
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
x 10
3
Number of elements
E
r
r
o
r

o
n

d
i
s
p
l
a
c
e
m
e
n
t


Coarsened meshes
Uncoarsened meshes
(a)
0.5 1 1.5 2 2.5 3
x 10
4
0
0.02
0.04
0.06
0.08
0.1
0.12
Number of elements
E
r
r
o
r

o
n

t
h
e

g
e
o
m
e
t
r
y


Coarsened meshes
Uncoarsened meshes
(b)
Figure 6.22: a) Evolution of the average displacement error in function of
the number of elements for selectively and uniformly coarsened meshes.
b) Evolution of the volume fraction of elements which are not in the
right grain for selectively and uniformly coarsened meshes.
156 CHAPTER 6. Assessment of the new CPFEM
coarsened meshes. This decrease is correlated with the accuracy in the
reproduction of the grain geometry. In the case of selectively rened
meshes, the average displacement error is also correlated with the mesh
quality. Comparisons show that selectively rened meshes allow a more
accurate prediction of the displacement eld.
6.3.2 Trends in the strain eld
For all results presented in this subsection, shear components have been
weighted by the volumes associated to their integration point.
Meshes with homogeneous characteristic length Distributions
of the shear components of the strain tensor are represented in gure
6.23. The more rened the mesh, the more we allow strain localisation
and the more we nd higher volume fractions of the RVE which bear
increasingly large shear strains as compared to the macroscopic average
which is enforced to be nil. Whereas only 4% of the volume of the RVE
has a shear component LE
23
larger than 0.02 in the less rened mesh,
approximately 22% of the volume of the RVE belong to this category in
the most rened mesh. These results are similar to those obtained by
Cuiti no [77].
Selectively coarsened meshes Since the distributions are very close
to each other, we represent only the dierence between the strain dis-
tribution in selectively rened meshes and the strain distribution in the
initial meshes (Fig. 6.24). The volume fraction of the RVE having a
LE
12
shear component between 0.0 and 0.1 is larger than in the refer-
ence mesh but the dierence decreases with the increase of the number
of elements. For the other intervals, the volume fraction of the RVE
which has a component LE
12
inside the interval is lower and the dier-
ence also decreases with the increase of the number of elements. Similar
conclusions can be drawn in the case of the other two strain components.
Hence, Cuitinhos observations are also reproduced for selectively rened
meshes.
I have also compared uniformly and heterogeneously rened meshes
based on the distribution of the shear components. The dierence be-
tween the distribution of each mesh and the distribution of the reference
6.3. The benet of selective mesh coarsening 157
(a) LE
12
(b) LE
13
(c) LE
23
Figure 6.23: Distribution of the shear components of strain in function
of the number of gauss points per grain in meshes with a homogeneous
characteristic length.
158 CHAPTER 6. Assessment of the new CPFEM
(a) LE
12
(b) LE
13
(c) LE
23
Figure 6.24: Comparison between the distribution of the shear compo-
nents for selectively rened meshes and the same distribution obtained
for the reference mesh in function of the number of gauss points per
grain. V F is the volume fraction of the RVE of the coarsened meshes
and V F
ref
is the volume fraction of the RVE of the initial mesh.
6.3. The benet of selective mesh coarsening 159
mesh is shown in gure 6.25. For the three shear components it appears
clearly that for meshes having approximately 26000 elements, the dif-
ference between the distribution of the uniformly rened mesh and the
reference mesh is lower than the dierence between the distribution of
the selectively rened mesh and the reference mesh. This is less clear
in meshes having approximately 30000 and 36000 elements. But, on
average, uniformly rened meshes lead to predictions which are closer
to the reference mesh prediction. For the most rened mesh, there is
only little dierence between the predictions obtained with the rened
meshes and those obtained with the reference mesh. Neither of the two
rened meshes gives better prediction than the other. These results tend
to show that mesh quality has a greater inuence on the prediction of
strain localisation than on the average error in the displacement eld.
In order to evaluate this inuence, I have compared (Fig. 6.26) the
predicted distributions of the shear components for the three meshes
obtained with a dierent value of the minimum element quality presented
in Section 5.2.6. For all shear components it appears that the lowest
dierence between the distribution of a selectively rened mesh and the
reference mesh is obtained when the imposed minimum element quality
is equal to 0.2. One can easily observe that these predictions are much
better than the predictions obtained with an imposed minimum quality
of 0.01. Therefore, we can conclude that elements which have a very
poor quality aect the quality of the prediction.
6.3.3 Summary
I have shown that both uniformly rened and selectively rened meshes
reproduce Cuiti nos observations. I have shown that selectively rened
meshes allow improved predictions of mean-like statistics compared to
uniformly rened meshes. But in the case of variance-like statistics, the
quality of the prediction is decreased with selectively rened meshes.
The main reason seems to be the too large number of elements having a
poor quality. When the imposed minimum element quality is increased,
the quality of the predictions of variance-like statistics is increased. The
problem is that when we increase the minimum element quality, we
decrease the eciency of the coarsening procedure. With a minimum
element quality of 0.2, we can only reduce the number of elements by
10% whereas with 0.01, we can reduce the number of elements by 50%.
160 CHAPTER 6. Assessment of the new CPFEM
(a) LE
12
(b) LE
13
(c) LE
23
Figure 6.25: Comparison between the distribution of the shear compo-
nents for selectively and uniformly rened meshes and the same distri-
bution obtained for the reference mesh for microstructures with various
numbers of elements. V F is the volume fraction of the RVE of the coars-
ened meshes and V F
ref
is the volume fraction of the RVE of the initial
mesh.
6.3. The benet of selective mesh coarsening 161
(a) LE
12
(b) LE
13
(c) LE
23
Figure 6.26: Comparison between the distribution of the shear compo-
nents for selectively rened meshes obtained for dierent values of the
minimum quality of elements and the same distribution obtained for the
reference mesh. V F is the volume fraction of the RVE of the coarsened
meshes and V F
ref
is the volume fraction of the RVE of the initial mesh.
162 CHAPTER 6. Assessment of the new CPFEM
6.4 Conclusion
As a rule of the thumb, a model microstructure must be composed of
500 grains and created on a mesh composed of second order tetrahedral
elements. Simulation must be run with an iterative solver. In the case of
a BCC materials, I advise to use 2000 gauss points per grain for macro-
scopic values and 8000 gauss points per grain for microscopic values for
small sets of grains. It must be noticed that the larger the number
of grains, the lower the number of gauss points per grain required for
macroscopic values. I advise a minimum of 500 gauss points per grain
for large sets of grains. In the case of HCP materials, the number of
gauss points must be 5 times larger than in the case of a BCC material.
Chapter 7
Use of CPFEM in practical
studies
In this chapter, I will present some of the various practical studies in
which our generator of model microstructures has been applied dur-
ing my thesis. Three of them have been published [62, 63, 78]. I will
present in detail the prediction of the strain heterogeneity of a dual phase
steel since I have performed the whole study by myself. I will present
the study on the grain shape inuence on the Lankford coecient only
briey since I have only contributed to a small part of the analysis. I will
present in detail a preliminary study on the use of heterogeneous model
microstructures on the prediction of the texture of TWIP steel. I will
just mention here that I have also generated 2D model microstructures
for an investigation of shear band apparition [78].
7.1 Strain heterogeneity in dual phase steel
Finite element representations of irregular microstructures are particu-
larly interesting in the case of multiphase materials where there typi-
cally is a large dierence of grain size among the phases. Usually, small
grains belong to the reinforcement phase. This is not taken into account
in many of the published analyses relying on regular model microstruc-
tures with a uniform grain size. In this section, we want to see if the
use of more realistic microstructures can improve the prediction of the
163
164 CHAPTER 7. Use of CPFEM in practical studies
strain heterogeneity of a dual phase alloy.
The steel sheet considered here has undergone extensive experimen-
tal investigation [79, 80]: neutron diraction, EBSD, Mosbauer spec-
trometry, deformation mapping by image correlation in the SEM, and
mechanical testing. The heat treatment, performed prior to forming
operations, creates a nely dispersed multiphase microstructure. The
metal consists of 45% ferrite, 43% bainite and 12% meta-stable austen-
ite. Transformation induced plasticity (TRIP) is observed in biaxial
stretching, but austenite is fully retained during the uniaxial tensile test
considered here [79, 80]. Previous numerical analysis on this steel has
only been performed with micro-macro transition schemes considering
homogeneous microstructures so that all grains are represented by the
same element or set of elements. It has been shown [46, 65] that these
RVE do not allow a valid prediction of the plastic strain heterogeneity
between the phases. It has also been shown [46, 66] that predictions
of CPFEM simulations are strongly dependent of the topological ar-
rangement of the second phase. As mentioned already, this was the
main motivation for the development of the generator of dual phase
microstructures.
The objective of this study is to check whether a heterogeneous mi-
crostructure helps to improve the prediction of lattice strains and plastic
strain heterogeneity. Generated RVE are statistically representative of
the real microstructure. Because ferrite and bainite cannot be distin-
guished in neutron diraction experiments, they are here modelled as a
single BCC phase interacting with the harder austenite phase (FCC).
The tensile test is performed up to 15% elongation on a specimen cut
parallel to the rolling direction (RD). Lattice strains are measured in
situ [79].
All simulations (Taylor FC, ALAMEL, CPFEM) rely on the same
routine to integrate the constitutive law of single crystals [65, 46]. In
BCC grains, dislocations move along 110 < 111 > and 112 < 110 >
slip systems, whereas 111 < 110 > slip systems are active in FCC
grains. The critical resolved shear stress (CRSS),
c
is identical on all
slip systems. Hardening is implemented as follows :

c
= H
0

sat

sat

c0

. (7.1)
In Eq. 7.1,
c0
and
sat
are, respectively, the initial and the saturation
values of the CRSS. Hardening parameters have been determined only
7.1. Strain heterogeneity in dual phase steel 165
Elastics constants [GPa] Hardening parameters [MPa]
C11 C12 C44
c0

sat
H
0
b.c.c phase 231.4 134 116 216 0.015 0.203
f.c.c phase 256 110 73 425 0.037 0.129
Table 7.1: Components of the anisotropic elastic modulus and parame-
ters of the hardening law
once by tting the lattice strains measured in the tensile direction in the
case of simulation with truncated octahedra (Fig. 4.8). Elastic constants
and average parameter values are listed in Table 7.1.
All FE simulations are performed with periodic boundary conditions
on ABAQUS. Model microstructures are representative of the multi-
phase steel presented in the introduction. The assignment of initial
orientation has been done thanks to our new discretisation technique.
The total volume of the cube V is the same in all simulations. The
volume fraction of the second phase f
2
is also xed. In this case, it is
xed to the value 0.12 as in the real microstructure.
Predictions of the lattice strains and the plastic strain heterogene-
ity obtained with irregular RVE are compared to available predictions
obtained with regular RVE. Regular RVE used for the comparison are
labelled Taylor FC, BRCK1, BRCK3 and OCTA. In BRCK1, each grain
is represented by a single hexahedral elements. In BRCK3, each grain
is represented by a set of 3 3 3 hexahedral elements. In OCTA, each
grain is represented by a truncated octahedron composed of 16 hexahe-
dral elements (Fig. 4.8). Two sets of irregular RVE are generated with
dierent values of R, the ratio between the mean diameter of grains of
the two phases, which is respectively equal to 1.45 and 2.5 (Fig. 7.1).
This corresponds to a value of R
D
, the ratio between the mean of the
minimal distance between grains of the second phase (D
min
) and the
mean diameter of the grains of the second phase (D
2
), equal to 1.4 and
1.16. In the case of regular microstructure, R is equal to one. R
D
is
equal to 1.6 in BRCK1 and BRCK3 and to 1.75 in OCTA.
All predictions are also compared to experimental measurements.
In this case, the matrix phase is composed of ferritic grains and the
inclusion phase of austenitic grains. For each set, 21 microstructures
containing 50 ferritic grains have been generated. There are therefore
23 austenitic grains in the microstructures of the rst set and 114 in the
second.
166 CHAPTER 7. Use of CPFEM in practical studies
(a) 1.45
(b) 2.5
Figure 7.1: a) Representation of the grains of the second phase of a
microstructure with a ratio between the mean grain diameters of the
two phases equal to 1.45, b) Representation of the grains of the second
phase of a microstructure with a ratio between the mean grain diameters
of the two phases equal to 2.5.[62]
7.1. Strain heterogeneity in dual phase steel 167
Figure 7.2: a) Average plastic strain of austenite grains in the tensile
direction for two values of R, the ratio between mean volume of grains
of the two phases b) Elastic lattice strains in the tensile direction. The
dotted upper curve corresponds to < 112 > diraction in f.c.c. and the
lower one to < 133 > diraction in b.c.c.[62]
168 CHAPTER 7. Use of CPFEM in practical studies
The average value on 21 simulations of plastic strain heterogeneity
and lattice strains is presented in gure 7.2 for several RVE. We can
see that the highest value of R
D
allows the best prediction of the strain
heterogeneity. This value of R
D
also allows a better prediction than
BRCK3. Only the OCTA simulation reproduces the lattice strains cor-
rectly. All other simulations overestimate the stress heterogeneity in the
two phases. An even more uniform (and, hence, less valid) strain eld
would be predicted by all microstructures if the strength of austenite
grains was reduced so as to t elastic lattice strains. The best predic-
tion is hence still obtained with OCTA, which has the largest R
D
ratio.
Surprisingly, the value of R is larger than 5 so that R
D
would be close
to 1. Regular ordering of the inclusion induces a strong anisotropy of
the macroscopic response. This is avoided with the new microstructure
generator due to the randomised positioning of the second phase.
7.2 Grain shape inuence on planar anisotropy
In general, model microstructures are represented by a set of equiaxed
grains whereas real microstructures are composed of grains with dierent
shapes and sizes. In the particular case of microstructures obtained after
rolling deformation, the grains are elongated in the rolling direction. I
have compared predictions of the Lankford coecient for two types of
model microstructures representative of a single-phase material.
In both cases, the RVE contains 500 grains which are subdivided into
tetrahedral elements with second-order interpolation. The rst one is a
3D microstructure like all heterogeneous microstructures previously used
in this thesis (Fig. 7.3a). The grains are equiaxed and there are 1200
integration points per grain. The second one is a 2D microstructure in
the sense that there is only one grain layer in the rolling direction (Fig.
7.3b). With the use of periodic boundary conditions, this represents
grains that are innitely long in the rolling direction. In the latter model
microstructure, each grain contains 600 integration points. For each of
the two types of microstructure, the tensile test is simulated ve times.
Each simulation corresponds to a dierent positioning of the seeds of
the Vorono tessellation. The 500 grains of each model microstructure
are assigned lattice orientations which are statistically representative of
the global texture.
7.2. Grain shape inuence on planar anisotropy 169
(a)
(b)
Figure 7.3: Two model microstructures of a single-phase material [63]. a)
Grains are equiaxed, b) Grains are innitely long in the rolling direction
(RD).
170 CHAPTER 7. Use of CPFEM in practical studies
The simulations, performed for two dierent textures labelled Mat
B and Mat C, illustrate that the planar anisotropy is very sensitive to
small changes of the texture (Fig. 7.4 and 7.5). The eect of the grain
shape is less important but it is not negligible.
7.3 Application of the advanced CPFEM to TWIP
steel
In order to link the two parts of my thesis and to close the loop, I have
performed CPFEM simulations with our most elaborated parametrisa-
tion (Table 4.1 and 4.12) obtained at the end of the rst part of the
manuscript. However, when the mesh is highly rened, it is not possible
to include a number of grains suciently large to produce statistically
meaningful texture predictions. The simulations presented below con-
sider only 10 grains. The goal is to study the inuence of the mesh
renement on the possibility for each grain ot achieve the deformation
with a single slip system. As stated in the discussion of the rst part of
the thesis, the development of a Brass-type texture is favoured by a lo-
calisation of dislocation slip on a single slip system ([47]). The tendency
to localise slip on a single slip system is evaluated through the following
measure:
Loc =
max[
,s
[

s

,s
. (7.2)
If Loc = 1, the slip activity is localised on a single slip system. CPFEM
predictions are compared to the predictions of the ALAMEL and the
Taylor model.
Simulations are performed on two of the model microstructures al-
ready used in this thesis. The most rened one which contains approxi-
mately 50000 elements has been used throughout the section 6.2 and the
less rened one which contains approximately 6000 elements has been
used in the section 6.3.1. They represent the same periodic microstruc-
ture of 10 equiaxed grains. Plane strain compression is simulated up to
an equivalent strain of 0.3.
Predictions are presented in table 7.2. It appears that the more we
allow heterogeneity throughout the model microstructure, i.e. the more
we rene the mesh, the more we predict a concentration of the slip on
a single slip system. The proportion of slip realised by the predominant
7.3. Application of the advanced CPFEM to TWIP steel 171
Figure 7.4: a)
2
= 45 section of the ODF of Mat B. b-c) Planar
anisotropy of Mat B under uniaxial tension, revealed by the Lankford
coecient (b) and the yield strength (c) calculated at dierent angles
relative to RD. In the legend, a, b and c designate the principal axes of
the ellipsoidal grain shape.[63]
172 CHAPTER 7. Use of CPFEM in practical studies
Figure 7.5: a)
2
= 45 section of the ODF of Mat C. b-c) Planar
anisotropy of Mat C under uniaxial tension, revealed by the Lankford
coecient (b) and the yield strength (c) calculated at dierent angles
relative to RD. In the legend, a, b and c designate the principal axes of
the ellipsoidal grain shape.[63]
7.3. Application of the advanced CPFEM to TWIP steel 173
TV F(%) Loc
Taylor 6.5 45.2
ALAMEL 5.7 52.3
FE6000 4.4 56
FE50000 4.1 60
Table 7.2: Twin volume fraction (%) and measure of the localisation of
the deformation on a single slip system predicted for various meshes of
the same model microstructure with our most elaborated modelling of
TWIP steel.
slip system is 6% higher with 50000 elements than with 6000 elements.
In the most rened CPFEM, the proportion of slip realised by the pre-
dominant slip system is 20% larger than when the ALAMEL model is
applied to the same ten grains. Rening the mesh even more than the
50000 elements used so far would allow concentrating dislocation even
more on a unique slip system. However, the computation time becomes
prohibitive. Indeed, the simulation with 50000 elements as already taken
two days. Convergence of the predicted strain eld with mesh rene-
ment is nowadays not possible in the context of crystal plasticity. This
partially explains the inaccuracy of the texture predictions when each
grain is represented by a truncated octahedron containing 16 rst-order
hexaedral elements (Section 4.2). The two relaxation modes allowed by
ALAMEL are no more sucient in this case. These simulations have
also conrm that the TVF predicted decreases with the increase of the
heterogeneity allowed.
174 CHAPTER 7. Use of CPFEM in practical studies
Chapter 8
Conclusion
In this thesis, I have tried to improve various components of metal form-
ing simulations in order to obtain accurate prediction of the behaviour
of TWIP steel. As more complex steels are developed, models must be
improved by taking account of new phenomena that occur at the micro
scale. The long-term goal is to perform simulations of real scale metal
forming processes and thus to be able to predict the behaviour of the
steel for various deformation paths. In order to reach this goal, I have
proposed and implemented new modelisations at two dierent scales :
the micron scale where I have designed a new modelling of twinning
implemented in a crystal plasticity code previously available in UCL
and the polycristalline aggregate scale where I have elaborated a new
technique of texture discretisation and designed new representations of
polycrystalline aggregates. All implementations have been done in order
to ensure numerical eciency. The tools elaborated have been validated
a rst time in this thesis but they can be used for other types of study
and for other single and multi-phase alloys.
New crystal plasticity model : The crystal plasticity model that
we have developed relies on the following selection of observations :
The rst twinning system activated in a grain has the highest
Schmid factor.
There is often one twin system activated per grain.
175
176 CHAPTER 8. Conclusion
There are at the most two twin sytems activated per grain.
The rst twins are observed during a tensile test after a true strain
of 0.05 to 0.1.
Twins act as obstacles to dislocation gliding on slip systems which
are non-coplanar to the twinning plane.
Our development can be summarised as follows :
adding twinning as a deformation mode complementary to dislo-
cation glide but unidirectional,
subdividing grains in twin regions and parent grain region,
representing all the twins created on a same twin system by a
single region,
adding slip-twin interaction in the hardening of the slip systems,
designing the evolution of the CRSS for the twin systems.
Twin volume fraction is an output of our model which is computed from
the activity on the twin systems. Subdividing a grain into subregions
leads to a supplementary modelling choice of the relation between stress
and strain in the grain and in the subregions. We have decided to assume
iso-deformation of the twins and the parent grain. Stress of the grain is
computed as the volumetric average of stress in the subregions. Special
care has been taken in order to add the lowest number of parameters.
This is implemented in a UMAT that allows the use of our model for
other materials than TWIP steel. Our model is implemented in such an
ecient way that a sole input le has to be modied in order to perform
simulations on other alloys.
Critical evaluation of several micro-macro transition schemes
: An originality of the approach is the comparison of various micro-macro
transition schemes relying on exactly the same single crystal constitu-
tive law. The Taylor model, ALAMEL and CPFEM have been com-
pared based on texture predictions after a tensile test and plane strain
compression. It has appeared that texture predictions with elaborated
177
micro-macro transition schemes are close to texture predictions obtained
with the Taylor model in the frame of a steel deforming by twinning.
As it is the case without twinning, ALAMEL remains a good and cheap
alternative to nite elements.
The evalution of our model combined with the most physical parametri-
sation leads to a non-minor inuence of twinning for a thickness reduc-
tion close to 50%. Our modelisation leads to a twin volume fraction
of 24% and a larger intensity of the Brass component than the Copper
component for a reduction of 54%.
Our model does better than previous models regarding the combined
predictions of the hardening and the texture. Our modelisation leads to
the appearance of the rst twins after 10%, predicts the global trends of
the hardening and of the incremental hardening and a valid prediction
of the texture for both small and large deformation.
The parameterisation study performed during this thesis provided
an extensive set of data that has been stored and analysed. Not all of
them have been compared with experimental measurements since the
latter were missing but comparisons could be made in the future when
they are available.
Analysis : Our predictions are in accordance with the expectations
of Leers in his survey about the prediction of the brass-type texture
[47]. Nevertheless, it should be mentioned that sensitivity of texture to
the use of elaborated micro-macro transition schemes is not as high as
expected. The prediction of the brass-type texture is a problem which is
open since more than 30 years in the texture community. Our modelling
allows the improvement of the quality of the prediction of this texture
but we dot not suceed to accurately predict it.
We have listed several reasons for the non-success of the prediction of
the brass-type texture. The micro-macro transition schemes used in this
thesis do not predict localisation of the deformation on a single slip sys-
tem as expected by Leers [47]. This may be explained by the fact that
the ALAMEL model does not allow sucient strain heterogeneity. The
Schmid law for the activation of the twin systems could be too simple.
The linear relation used for the anisotropic slip-twin interaction does
not reproduce the fact that, the later twins are created, the more they
reduce the mean free path of dislocations in the untwinned part of the
178 CHAPTER 8. Conclusion
grains. One solution that could help to improve the prediction is to use
more realistic and more rened representations of the microstructures
in CPFEM. This was the goal of the second part of my thesis.
Advanced CPFEM in polycrystalline aggregates A new tool for
the representation of polycrystalline aggregate has been designed during
this thesis. This tool is composed of three dierent programs.
The rst one discretises experimentally measured crystallographic
textures in a sampling of discrete orientations which is representa-
tive of the experimental texture. This technique is an enhancement
of the technique proposed by Toth and Van Houtte. Contrary to
any discretisation method available so far, the algorithm allows
the generation of model microstructures in which the texture and
the grain-size distribution may be prescribed independently of one
another.
The second one creates periodic microstructures containing either
one or two phases. The microstructure is generated according to a
Vorono tesselation. Seeds are positioned randomly in space with
a minimum distance between them which is proportional to the
average grain diameter. For the secondary phase, seed positions
are selected among the nodes that are at the intersection of four
grains of the primary phase in order to take into account the fact
that grains of the second phase appear mainly at triple junctions
in EBSD images of real microstructures. Growth of grains of the
second phase ends when the volume fraction reaches the prescribed
value.
The third one coarsens meshes at the centre of the grains. The
particularity of the mesh coarsening is that we preserve grain ge-
ometry and of course the periodicity of the mesh.
These three tools form a package, they can be used successively or
independently. The program for texture discretisation requires an ex-
perimental texture and a list of weights and gives a list of orientations
which has the form of an input le for our simulation. The genera-
tor of periodic microstructures requires a FE mesh created by the open
source software gmsh and data describing the real microstructure and
gives a list of weights containing the grain volumes for each phase (that
179
can be used as input for the discretisation technique) and input les
for Abaqus simulation describing the microstructure mesh and periodic
bounday conditions. The coarsening program takes a FE mesh created
by gmsh and a size eld and gives a description of the coarsened meshes
in the gmsh format. A post-processing program is then used to create
input les for Abaqus simulations that describe the coarsened mesh and
periodic boundary conditions. These tools have been extensively vali-
dated and are ready to be used in a large variety of studies even very
dierent from the one performed in this thesis.
During this validation, it has appeared that the choice of hexahe-
dral or tetrahedral elements does not inuence the prediction for a same
discretisation of the microstructure and that rst-order tetrahedra of
Abaqus do not allow us to reach a converged solution. We have there-
fore chosen to generate our microstructure with second-order tetrahedral
elements. Advantages of the use of tetrahedral elements are that : it
allows us to mesh all grain shapes and it eases the coarsening and the
positioning of second phase grains.
Assessment of the new CPFEM We have shown that the new tech-
nique for texture discretisation coupled with heterogeneous microstruc-
ture allows the reduction of the number of grains required in a RVE [51].
Indeed, in the case of a simulation with the Taylor model, the variance
predicted with 512 heterogeneous grains is of the same order as that the
variance with 10000 homogeneous grains in a simulation of the planar
anisotropy of an IF steel while the mean value of the Lankford coecient
is not signicantly aected.
We have shown that selectively rened meshes, like uniformly rened
meshes, allow the reproduction of the results presented by Cuiti no [77]
on the evolution of local shear strain intensity with the renement [64].
We have then shown that selectively rened meshes allow the reduction
of the computation time and also that predictions of mean-like statistics
are improved in comparison with predictions of uniformly rened meshes
for a similar number of elements.
We have shown that heterogeneous microstructures improve the pre-
diction of the strain heterogeneity compared to microstructures where
each grain is represented by the same brick [62]. We have shown that
grain shape has an inuence on the prediction of the Lankford coe-
cient but not as important as the texture inuence [78]. We have shown
180 CHAPTER 8. Conclusion
that the more we allow heterogeneity throughout a model microstruc-
ture containing 10 grains, i.e. the more we rene the mesh, the more we
predict a concentration of the slip on a single slip system.
Future developments : In the present thesis, I have mainly devel-
oped various tools that aim at the same goal : perform the most ac-
curate simulation of a TWIP steel. It is hoped that, in the future,
computational power will be sucient enough in order to completely
close the loop by performing nite element simulations with selectively
rened heterogeneous microstructures containing thousands of grains in
the case of TWIP steel.
It would be interesting to see if the improved crystal plasticity model
and the new model microstructures lead to good predictions of the
Lankford coecient of TWIP steel (and more generally of the plastic
anisotropy). The inuence of the mesh renement on the texture pre-
dicted after cold rolling simulations is a study that helps to evaluate
the real quality of the actual modelling of the TWIP steel. Each one
of the tools can also still be improved. So far simulations have only
been performed on equiaxed or innitely elongated grains. The inu-
ence of intermediate grain shape with moderately elongated grains must
be tested in the future.
An improvement that I clearly have in mind regarding the modelling
of the TWIP steel is the improvement of the twin activation. Currently,
twin systems are activated in function of a Schmid law. I think that the
use of the Schmid law is a good rst-order hypothesis but that in order to
improve the prediction, a more elaborated activation law must be used.
Indeed, as we think that slip-twin interaction has an important inuence
on the texture prediction, it is important to predict twinning activation
on the right twin systems. The more we improve the prediction of the
twin activation, the more we will improve the prediction of the inuence
of the slip-twin interaction.
Another way to improve our modelling of the TWIP steel is to nd a
more elaborate model of the slip-twin interaction. Currently, this inter-
action is only modelled with a linear function of the activity on a twin
system representing the slip-twin interaction in the hardening of the slip
systems. A simple way to improve this modelling is to replace the linear
relation by a relation of higher order. Another one is to take account
of the accumulation of the twins in this relation. We can also improve
181
the modelling of this interaction by removing the assumption of iso-
deformation of the twin and the grain. During this thesis, various other
assumptions have been tested (but not mentioned in this text) but none
of them has led to much improved texture prediction. One modelisation
that could be tested is to use the viscoplastic self-consistent model where
inclusions are clusters of grains having ALAMEL type interaction. The
nal goal is to nd a modelisation that leads to a concentration of the
slip activity on a single slip system as stated by Leers [47].
182 CHAPTER 8. Conclusion
Appendix A
In depth comparison of
various predictions
obtained during this thesis
I will now compare some of the predictions obtained in these sections to
each other. The goal is to nd a correlation between the evolution of var-
ious characteristics of the predictions. The characteristics investigated
here are
Intensity and position of the peak of the orientation component
along the bre.
Distribution of orientations in function of their TVF
Texture of the grains having less than 10% of twins
Texture of the grains having more than 10% of twins
Prediction of the texture after a tensile test
The measurements of the experimental texture are also compared when
they are available. The four predictions are obtained for dierent parametri-
sations :
isotropic hardening of the slip systems, isotropic hardening of the
twin systems.
183
184
CHAPTER A. In depth comparison of various predictions obtained during
this thesis
isotropic hardening of the slip systems, anistropic hardening of the
twin systems (b
tt
= 100).
twin acts as barrier to dislocation motion in an isotropic way (
1

st
=
10000), anistropic hardening of the twin systems (b
tt
= 100).
twin acts as barrier to dislocation motion in an anisotropic way
(b
st
= 100), anistropic hardening of the twin systems (b
tt
= 100).
They all have an expression of
t,
c
which is only function of the twinning
activity.
Intensity and position of the bre: The skeleton line along the
bre, going from Brass to Copper, is represented on gure A.1. It
Figure A.1: Intensity of orientations along the bre after plane strain
compression (
eq
= 0.9) for various hardening parameters. Simulations
are performed with ALAMEL and with parameterisations leading to the
appearance of the rst twins after 10%.
can be observed that the best texture prediction is very close to the
185
experimental texture along the whole bre. All other textures have
a ratio between the intensity of the Brass and the Copper component
sligthly lower than one. The texture obtained with anisotropic slip-twin
interaction is the lowest.
The simulation with isotropic slip-twin interaction also gives the best
prediction of the position of
1
for all orientations along the bre
A.2. All other predictions are very close to each other. None of the
predictions leads to a good prediction of the position of the peak of the
Copper component.
Figure A.2: Position (
1
) of orientations along the bre after plane
strain compression (
eq
= 0.9) for various hardening parameters. Simu-
lations are performed with ALAMEL and with parameterisations leading
to the appearance of the rst twins after 10%.
The same conclusion can be drawn for the position of (Figure
A.3). A gap remains between the best prediction and the experimental
measurement of the position of the peak of the Copper component. This
can be explained by the very low intensity of the component which
lead to large uncertainties on the measurement. Predictions show that,
186
CHAPTER A. In depth comparison of various predictions obtained during
this thesis
Figure A.3: Position () of orientations along the bre after plane
strain compression (
eq
= 0.9) for various hardening parameters. Simu-
lations are performed with ALAMEL and with parameterisations leading
to the appearance of the rst twins after 10%.
in general, the decrease of the intensity of the copper orientation is
correlated with an improvement of the prediction of the peak of the
copper component.
In order to study the exact inuence of the hardening parameter on
texture predictions, I have separated each texture into two textures. The
rst one is the texture obtained without twin orientations (Fig. A.4a)
and the second one is the texture obtained only with twins orientations
(Fig. A.4b). It appears that for the parameterisation leading to the
best texture prediction, the texture without twins is already close to the
experimental texture. Twin orientations only help to slightly increase
the intensity of the Brass component in the overall texture. The same
conclusion can be made for all other predictions. The texture prediction
without twins obtained with the anisotropic slip-twin interaction has a
particular trend. All the components of the bre approximately have
the same intensity. Anisotropic twin-twin interaction helps to reduce the
187
(a)
(b)
Figure A.4: a) Intensity of orientations along the bre taken on the tex-
ture without twin orientations after plane strain compression (
eq
= 0.9)
for various hardening parameters. b) Intensity of orientations along the
bre taken for the texture with only twin orientations after plane strain
compression (
eq
= 0.9) for various hardening parameters. Simulations
are performed with ALAMEL and with parameterisations leading to the
appearance of the rst twins after 10%.
188
CHAPTER A. In depth comparison of various predictions obtained during
this thesis
intensity of the peak of the copper component for the texture without
twin orientations. The dierences observed in the texture with only
the twin-twin interaction concern the intensity of the peak of the brass
component. It should be noticed that these observations have only a
minor impact on the overall texture since the TVF is approximately
equal to 24%. Nevertheless, it was not expected that the modeling of
the hardening of the slip and twin systems leads to such dierence on
the twinning texture.
Distribution of orientations in function of their TVF: Whereas,
on average over the whole set of orientations, each parameterisation leads
to the same TV F, TVF predicted for individual orientations are dier-
ent from one parameterisation to another. This results is shown on table
A.1 where we present the distribution of orientations in function of their
TVF. It can be observed that the distributions are all dierent from
1 2 3 4
TV F < 10% 10 19 28 15
10 < TV F < 20% 13 18 15 25
20 < TV F < 30% 51 20 19 20
30 < TV F < 40% 25 35 14 29
40 < TV F < 50% 0 7 17 10
50 < TV F < 60% 0 0 7 1
60 < TV F < 70% 0 0 1 0
70 < TV F < 80% 0 0 0 0
Table A.1: 1 : Isotropic, 2 : b
tt
= 100, 3 : A
st
= 10000b
tt
= 100,
4 : b
st
= 100b
tt
= 100. Distribution of orientations in function of their
TVF obtained after plane strain compression (
eq
= 0.9). Simulations
are performed with ALAMEL and with parameterisations leading to
appearance of the rst twins after 10%.
each other. Whereas a majority of orientations have a TVF close to
the average TVF for the isotropic simulation, this is not the case for
all other simulations. The introduction of twin-twin interaction leads
to a more uniform distribution and particularly increases the number of
orientations having a TVF between 30% and 40%. The simulation lead-
ing to the best texture prediction globally has a uniform distribution of
orientations having a TVF between 0 and 50%. Nevertheless, there is
a larger number of orientations having a TVF lower than 10%. Some
189
orientations also have a very large TVF. Compared to the latter sim-
ulation, the simulation with anisotropic slip-twin hardening predicts a
lower number of orientations having a TVF lower than 10% and a lower
number of orientations having a TVF larger than 40%.
Texture of the orientations having less than 10% of twins: We
have seen that the texture without twin orientations and the distribution
of the TVF are strongly inuenced by the hardening parameters. Let
us see if there is a correlation between these two characteristics. First, I
look at the texture, without twins, of the orientations that have a TVF
lower than 10% (Figure A.5). Globally these textures are quite equiv-
PAGE
0 90
PHI2= 45
(a)
PAGE
0 90
PHI2= 45
(b)
PAGE
0 90
PHI2= 45
(c)
PAGE
0 90
PHI2= 45
(d)
Figure A.5: a) Isotropic, b) b
tt
= 100, c) A
st
= 10000b
tt
= 100, d) b
st
=
100b
tt
= 100.
2
= 45 ODF section obtained from the orientations
having a TVF lower than 10%. TVF and orientations are obtained
after simulations of plane strain compression (
eq
= 0.9) performed with
ALAMEL and with parameterisations leading to the appearance of the
rst twins after 10%.
190
CHAPTER A. In depth comparison of various predictions obtained during
this thesis
alent but the texture of the isotropic simulation is mainly composed of
the Goss orientation with some Brass orientations. All other simula-
tions have at least some Copper orientations and the peak of the Brass
component is closer to the usual position of Brass than to the position
of the Goss component. The introduction of the anisotropic twin-twin
interaction already allows the obtention of these two trends.
Surprisingly, the texture obtained after the simulation with isotropic
hardening of the slip-twin interaction and the one obtained after sim-
ulation with anisotropic hardening of slip-twin interaction lead to two
very close textures whereas in average over the whole set of orientations,
these textures are completely dierent. Therefore, we can conclude that
one explanation for the dierence between both overall textures with
slip-twin interaction is due to the distribution of orientations having a
low TVF. When the interaction is isotropic, 27% of orientations have
a low TVF and that governs the overall texture. When the interaction
is anisotropic, only 15% of orientations have a low TVF. However, this
cannot be the sole explanation.
Texture of the orientations having more than 10% of twins: In
order to nd other explanations for the dierence in the texture without
twin orientations, I have looked at the texture created from the orien-
tations having a TVF between 10% and 20%, a TVF between 20% and
30% and a TVF higher than 30% (Fig. A.6). I have compared only
the textures obtained with parametrisations accounting for slip-twin in-
teraction since they are the most elaborated. The overall trend is the
same for both parameterisations. The more important the TVF, the less
important the Brass intensity, whereas the Copper intensity is globally
the same. Therefore when the TVF is higher than 30%, there is more
Copper than Brass but it is the opposite when the TVF is between 10%
and 20%.
Nevertheless, dierences appear clearly between both textures. The
position of the peak of the Copper and the Brass components also
slightly evolves from one texture to another. When there is isotropic
slip-twin interaction, the Copper component is very weak. The peak
of the Brass component is not at its usual position but towards larger
values of
1
except when the TVF is higher than 30% where the Brass
component is inexistant. This shift towards larger values of
1
is not
observed in texture predicted with anisotropic slip-twin interaction. If
we look at the experimental texture (Fig. 4.2), we can see that there are
191
PAGE
0 90
PHI2= 45
(a) 10% < TV F < 20%
PAGE
0 90
PHI2= 45
(b) 20% < TV F < 30%
PAGE
0 90
PHI2= 45
(c) TV F > 30%
PAGE
0 90
PHI2= 45
(d) 10% < TV F < 20%
PAGE
0 90
PHI2= 45
(e) 20% < TV F < 30%
PAGE
0 90
PHI2= 45
(f) TV F > 30%
Figure A.6: First line : A
st
= 10000b
tt
= 100, second line : b
st
=
100b
tt
= 100.
2
= 45 ODF section obtained from the orientations
having a given TVF. TVF and orientations are obtained after simula-
tions of plane strain compression (
eq
= 0.9) performed with ALAMEL
and with parameterisations leading to the appearance of the rst twins
after 10%.
indeed some of the orientations of the Brass component that are located
towards these larger values of
1
.
When there is anisotropic slip-twin interaction, the decrease of the
intensity of the peak of Brass component is less obvious and the inten-
sity of the peak of the Copper component remains high. It therefore
appears that whereas the anisotropic twin-twin interaction only inu-
ences the selection of twin activation and not the texture of non-twin
orientatiosn, the slip-twin interaction inuences both the twin activation
(leading to a dierent distribution of the orientations in function of the
TVF) and the texture of non-twin orientations for orientations having a
TVF larger than 10%. With isotropic slip-twin interaction, the Copper
component is nearly completely destroyed and the peak of the Brass
component is shifted towards larger values of
1
whereas for isotropic
192
CHAPTER A. In depth comparison of various predictions obtained during
this thesis
slip-twin interaction the intensities of both the Brass and the Copper
component are decreased and their positions remain unchanged. This is
my second explanation for the dierence between the texture with the
isotropic slip-twin interaction and the one with the anisotropic slip-twin
interaction.
Appendix B
Quaternion manipulation
Denition of quaternion based on Euler angles :
q
0
= cos(

2
)cos(

1
+
2
2
) (B.1)
q
1
= sin(

2
)cos(

2
2
) (B.2)
q
2
= sin(

2
)sin(

2
2
) (B.3)
q
3
= cos(

2
)sin(

1
+
2
2
) (B.4)
Expression of a rotation matrix based on a quaternion :
R = 2
_
_
q
2
0
+q
2
1

1
2
q
1
q
2
+q
0
q
3
q
1
q
3
q
0
q
2
q
1
q
2
q
0
q
3
q
2
0
+q
2
2

1
2
q
2
q
3
+q
0
q
1
q
1
q
3
+q
0
q
2
q
2
q
3
q
0
q
1
q
2
0
+q
2
3

1
2
_
_
(B.5)
The product of two orthogonal matrices, R
c
= R
a
R
b
(R
c
ij
= R
a
ik
R
b
kj
with
Einsteins convention for the summation over repeated indices), may be
computed eciently using quaternions: q
c
= q
a
q
b
where:
q
c
0
= q
a
0
q
b
0
q
a
1
q
b
1
q
a
2
q
b
2
q
a
3
q
b
3
(B.6)
q
c
1
= q
a
0
q
b
1
+q
a
1
q
b
0
q
a
2
q
b
3
+q
a
3
q
b
2
(B.7)
q
c
2
= q
a
0
q
b
2
+q
a
2
q
b
0
q
a
3
q
b
1
+q
a
1
q
b
3
(B.8)
q
c
3
= q
a
0
q
b
3
+q
a
3
q
b
0
q
a
1
q
b
2
+q
a
2
q
b
1
(B.9)
193
194 CHAPTER B. Quaternion manipulation
The orientation of a crystal lattice relative to a macroscopic reference
frame can be characterised using a quaternion. If one orientation is
represented by q
a
and another by q
b
, the angle of the rotation bringing
one lattice onto the other is computed as :
= 2 acos([q
a
0
q
b
0
+q
a
1
q
b
1
+q
a
2
q
b
2
+q
a
3
q
b
3
[) (B.10)
The orientation of a lattice with cubic symmetry can be represented by
24 quaternions. The disorientation between two crystal lattices a and
b is the smallest rotation angle bringing one lattice onto the other.
The mean orientation within a grain is dened as the lattice orienta-
tion with the lowest average disorientation with regard to all elementary
volumes constituting the grains. As shown by Humbert et al. [81], this
mean orientation may be approximated by:
m =
a
[a[
=

N
i=1
q
i

N
i=1
q
i

. (B.11)
In applying this formula, one should be careful that all quaternions
pertain to a same zone in the orientation space. In other words, one
should make the right choice among the 24 quaternions representing
each elementary volume.
Appendix C
k-d Tree code
A k-d tree is a data structure composed as a tree which allows multidi-
mensionnal searches for entities in k dimensional space. The C++ code
that I have used is presented below. The goal is to nd as quickly as
possible if a node is already present in a mesh. The entity is therefore
a node and its dimension is four : three reals that dene its position
in the 3D-space and its index. When I want to create a new node in a
given position, I call the giveNumber function that recursively searches
into the three if there is already a node at this position. If the answer
is yes, the result is the node index, if the answer is no, the result is 0.
This code has been found on internet and modied in order to deal with
nodes as entities. It requires two les that are given after the code :
vector.h and vector.c.
KdTree.c
#include "vector.h"
#include <map>
using namespace std;
template <class Comparable>
class KdTree
{
public:
KdTree( ) : root( NULL ) { }
void insert( const vector<Comparable> & x, const int d )
{
195
196 CHAPTER C. k-d Tree code
insert( x, root, 0, d );
}
/**
* Print items satisfying
* low[ 0 ] <= x[ 0 ] <= high[ 0 ] and
* low[ 1 ] <= x[ 1 ] <= high[ 1 ]
*/
void printRange( const vector<Comparable> & low,
const vector<Comparable> & high ) const
{
printRange( low, high, root, 0 );
}
int giveNumber( const vector<Comparable> & pos,
const int d ) const
{
int a=giveNumber( pos, root, 0,d);
return a;
}
private:
struct KdNode
{
vector<Comparable> data;
KdNode *left;
KdNode *right;
KdNode( const vector<Comparable> & item )
: data( item ), left( NULL ), right( NULL ) { }
};
KdNode *root;
void insert( const vector<Comparable> & x, KdNode * & t,
int level, const int d )
{
int newlevel;
newlevel=level+1;
if (newlevel==d) newlevel=0;
if( t == NULL )
t = new KdNode( x );
else if( (x[ level ] < t->data[ level ]) &&
(fabs(x[level]-t->data[level])>1.e-10))
insert( x, t->left, newlevel, d );
else
197
insert( x, t->right, newlevel, d );
}
void printRange( const vector<Comparable> & low,
const vector<Comparable> & high,
KdNode *t, int level ) const
{
if( t != NULL )
{
if( low[ 0 ] <= t->data[ 0 ] &&
high[ 0 ] >= t->data[ 0 ] &&
low[ 1 ] <= t->data[ 1 ] &&
high[ 1 ] >= t->data[ 1 ] )
cout << "(" << t->data[ 0 ] << ","
<< t->data[ 1 ] << ")" << endl;
if( low[ level ] <= t->data[ level ] )
printRange( low, high, t->left, 1 - level );
if( high[ level ] >= t->data[ level ] )
printRange( low, high, t->right, 1 - level );
}
}
int giveNumber( const vector<Comparable> & pos,
KdNode *t, int level, const int d ) const
{
int newlevel;
int a;
newlevel=level+1;
if (newlevel==d) newlevel=0;
if( t != NULL )
{
if( (fabs(pos[ 0 ]-t->data[ 0 ])<1.e-10) &&
(fabs(pos[ 1 ]- t->data[ 1 ])<1.e-10) &&
(fabs(pos[ 2 ]- t->data [ 2 ])<1.e-10))
{
return (int) t->data[3];
}else{
if( (pos[ level ] < t->data[ level ]) &&
(fabs(pos[level]-t->data[level])>1.e-10)){
a=giveNumber( pos, t->left, newlevel, d );
}else{
a=giveNumber( pos, t->right, newlevel, d );
}
return a;
}
198 CHAPTER C. k-d Tree code
}else{
return 0;
}
}
};
Bibliography
[1] L. Delannay, P.J. Jacques, K. Dejaegher, S-S. Sablin (2008) Evalu-
ation of two multiscale models for the simulation of cross-die form-
ing of DP steel. In: Proc. Esaform (Lyon, Apr. 2008) International
Journal of Material Forming DOI: 10.1007/s12289-008-0045-9
[2] T.J.R. Hughes and J. Winget, Finite rotations eects in numerical
integration of rate constitutive equations arising in large deforma-
tions analysis, Int. J. Num. Meth. Eng. 15, (1980), 1862-1867
[3] ABAQUS Theory Manual, Version 6.7, Hibbit, Karslon and
Sorensen, (2008)
[4] L. Delannay, M. Beringhier, Y. Chastel, R.E. Loge, Simulation of
cup-drawing based on crystal plasticity applied to reduced grain
samplings, Mat. Sci. Forum 495-497, (2005), 1639-1644
[5] D. Raabe, F. Roters, Using texture components in crystal plastic-
ity nite element simulations, Int. J. Plast. 20, (2004), 339-361
[6] A.M. Habraken, L. Duchene, Anisotropic elasto-plastic nite ele-
ment analysis using a stressstrain interpolation method based on a
polycrystalline model, Int. J. Plast. 20, (2004), 1525-1560
[7] P. Van Houtte, A. Van Bael, Convex plastic potentials of fourth
and sixth rank for anisotropic materials, Int. J. Plast. 20, (2004),
1505-1524
[8] J. Winters, PhD thesis, MTM department, KULeuven, (1996)
[9] MTM-FHM software, MTM department, KULeuven, Belgium,
(2001)
[10] H.J. Bunge, Texture Analysis in Materials Science, Butterworths
Publishers, London (1982)
199
200 BIBLIOGRAPHY
[11] G.I. Taylor, Plastic strain in metals, J. Inst. Metals 62, (1938),
307-324
[12] E. Sachs, Z. Ver. Dt. Ing. 72, (1928), 734
[13] P. Van Houtte, S. Li, M. Seefeldt, L. Delannay, Deformation tex-
ture prediction: from the Taylor model to the advanced Lamel
model, Int. J. Plast. 21, (2005), 589-624
[14] R.A. Lebensohn, C.N. Tome, A self consistent anisotropic ap-
proach for the simulation of plastic deformation and texture de-
velopment of polycrystals : application to zirconium alloys, Acta
Metall. Mater. 41, (1993), 2611-2624
[15] P. Van Houtte, A comprehensive mathematical formulation of an
extended Taylor-Bishop-Hill model featuring relaxed constraints,
the Renouard-Wintenberger theory and a strain rate sensitivity
model, Text. Microstruct. 8-9, (1988), 313-350
[16] F.D. Fischer, T. Schaden, F. Appel, H. Clemens, Mechanical twins,
their development and growth, Eur. J. of Mech. A/Solids 22,
(2003), 709-726
[17] L. Capolungo, I.J. Beyerlein and C.N. Tome, Slip-assisted twin
growth in hexagonal close-packed metals, Scripta Mat. 60, (2009),
32-35
[18] S.G. Chowdhury, S. Das, P.K. De, Cold rolling behaviour and
textural evolution in AISI 316L austenitic stainless steel, Acta Mat
53, (2005), 3951-3959
[19] S. Vercammen, B. Blanpain, B.C. De Cooman, P. Wollants, Cold
rolling behaviour of an austenitic FE-30Mn-3Al-3Si TWIP-steel :
the importance of deformation twinning, Acta Mat. 52, (2004),
2005-2012
[20] S. Asgari, E. El-Danaf, S. Kalidindi and R. Doherty, Strain harden-
ing regmies and microstructural evolution during large strain com-
pression of low stacking fault energy FCC alloys that form defor-
mation twins, Met. Mat. Trans. A 28a, (1997), 1781-1795
[21] I. Karaman, H. Sehitoglu, A.J. Beaudoin, Y. Chumlyakov,
H.J. Maier and C.N. Tome, Modeling the deformation behavior
of hadeld steel single and polycrystals due to twinning and slip,
Acta Mat. 48, (2000), 2031-2047
BIBLIOGRAPHY 201
[22] S. Allain, J.P. Chateau, D. Dahmoun, O. Bouaziz, Modeling of
mechanical twinning in a high manganese content austenitic steel,
Mat. Sci. Eng A 387-389, (2004), 272-276
[23] E. El-Danaf, S. Kalidindi and R. Doherty, Inuence of grain size
and stacking-fault energy on deformation twinning in FCC metals,
Met. Mat. Trans. A 30A, (1999), 1223-1233
[24] S. Mahajan and G.Y. Chin, Twin-slip, twin-twin and slip-twin
interactions in Co-S wt.% Fe alloy single crystals, Acta Metall.
21, (1973), 173-179
[25] M. Dao, L. Lu, Y.F. Shen, S. Suresh, Strength, strain-rate sensi-
tivity and ductility of copper with nanoscale twins, Acta Mat. 54,
(2006), 5421-5432
[26] K. Renard, Caracterisation et analyse mecanique et mi-
cromecanique du maclage et des mecanismes decrouissage dans
les aciers `a haute teneur en mangan`ese - Alloy design et opti-
misation des relations composition-elaboration-proprietes.,Unite
dIngenierie des Materiaux et des Procedes, Ph. D. Thesis, Uni-
versite catholique de Louvain (UCL), Belgium, (2011).
[27] P.H. Adler, G.B Olson, W.S. Owen, Strain hardening of hadeld
maganese steels, Metall. Trans A 17A, (1986), 1725-1737
[28] T. Shun, C.M. Wan and J.G. Byrne, A study of work hardening
in austenitic Fe-Mc-C and Fe-Mn-Al-C alloys, Acta metall. mater.
40, (1992), 3407-3412
[29] S. Allain, Caracterisation et modelisation thermomecaniques
multi-echelles des mecanismes de deformation et decrouissage
daciers austenitiques `a haute teneur en mangan`ese - Application `a
leet TWIP, Ph. D. Thesis, Ecole des Mines de Nancy, (2004).
[30] E. El-Danaf, S.R. Kalidindi, R.D. Doherty and C. Necker, Defor-
mation texture and transition in brass : critical role of micro-scale
shear bands, Acta Mater 48, (2000), 2665-2673
[31] L. Delannay, P.J. Jacques, S.R. Kalidindi, Finite element modeling
of crystal plasticity with grains shaped as truncated octahedrons,
Int. J. Plast. 22, (2005), 1879-1898
[32] E.C. Aifantis, The physics of plastic deformation, Int. J. Plast.
3, (1987), 211-247
202 BIBLIOGRAPHY
[33] J.W Hutchinson, Creep and plasticity of hexagonal polycristals as
related to single crystal slip, Metall. Trans A 8, (1969), 1465-1469
[34] P. Van Houtte, Simulation of the rolling and shear texture of brass
by the taylor theory adapted for mechanical twinning, Acta Metall.
26, (1978), 591-604
[35] S.R. Kalidindi, Incorporation of deformation twinning in crystal
plasticity models, J. Mech. Phys. Solides 46, (1998), 267-290
[36] A. Staroselsky and L. Anand, Inelastic deformation of polycrys-
talline face centered cubic materials by slip and twinning, J. Mech.
Phys. Solids 46, (1998), 671-696
[37] C.N. Tome, R.A. Lebensohn, U.F. Kocks, A Model for Texture
Development Dominated by Deformation Twinning : Application
to Zirconium Alloys, Acta Metal. et Mat. 39, (1991), 2667-2680
[38] J. Gil Sevillano, An alternative model for the strain hardening
of FCC alloys that twin, validated for twinning-induced plasticity
steel, Scripta Mat. 60, (2009), 336-339
[39] S.R. Kalidindi, Modeling anisotropic strain hardening and defor-
mation textures in low stacking fault energy FCC metals, Int. J.
Plast. 17, (2001), 837-860
[40] I. Karaman, H. Sehitoglu, H.J. Maier and Y.L. Chumlyakov, Com-
peting mechanisms and modeling of deformation in austenitic stain-
less steel single crystals with and without nitrogen, Acta mater.
49, (2001), 3919-3933
[41] S. Allain, J.P. Chateau, O. Bouaziz, A physical model of the
twinning-induced plasticity eect in a high manganese austenitic
steel, Mat. Sci. Eng. A 387-389, (2004), 143-147
[42] I. Karaman, H. Sehitoglu, A.J. Beaudoin, Y.I Chumlyakov,
H.J. Maier and C.N. Tome, Modeling the deformation behavior
of hadeld steel single and polycrystals due to twinning and slip,
Acta Mater. 48, (2000), 2031-2047
[43] M.A. Meyers, O. Vohringer and V.A. Lubarda, The onset of
twinning in metals : A constitutive description, Acta Mater. 49,
(2001), 4025-4039
[44] B. Clausen, C.N. Tome, D.W. Brown, S.R. Agnew, Reorientation
and stress relaxation due to twinning : Modeling and experimental
characterization for Mg, Acta Mat. 56, (2008), 2456-2468
BIBLIOGRAPHY 203
[45] J.C. Simo, T.J.R. Hughes, Computational inelasticity, Springer
Verlag, New-York, (1998)
[46] L. Delannay, P.J. Jacques, S.R. Kalidindi,Finite element modeling
of crystal plasticity with grains shaped as truncated octaedrons,
I nt. J. Plast. 22, 18791898, 2006
[47] T. Leers, R.K. Kay, The brass-type texture and its deviation
from the copper-type texture, Prog. Mat. Sci. 54, (2009), 351-396
[48] Private communication.
[49] M.N. Shiekhelsouk, V. Favier, K. Inal, M. Cherkaoui, Modelling
the behaviour of polycristalline austenitic steel with twinning-
induced plasticity eect, Int. J. Plast. 25, (2009), 105-133
[50] L.A. Hantcherli, P. Eisenlohr, F. Roters, On the role of mechani-
cal twinning in microstructure evolution of high manganese steels :
experiments and modelling, 15 th International Conference on the
Texture of Materials (ICOTOM 15), Pittsburgh, (2008)
[51] M.A. Melchior, L. Delannay, A texture discretization technique
adapted to polycristalline aggregates with non-uniform grain size,
Comp. Mat. Sci. 37, (2006), 557-564
[52] U.F. Kocks, J.S. Kallend, A.C. Biondo, Accurate representations
of general textures by a set of weighted grains,Textures and Mi-
crostructures, Gordon and Breach Science Publishers SA, United
Kingdom, 14-18, (1991), 199204.
[53] K. Helming, R. Tamm, B. Fels, An automated component
method, Materials Science Forum, 273-275, (1998), 119124
[54] P. Eisenlohr, F. Roters, Selecting a set of discrete orientations
for accurate texture reconstruction, Comp. Mat. Sci 42, (2008),
670-678.
[55] L.S. Toth, P. Van Houtte, Discretization technique for orientation
distribution functions, Textures and Microstructures, 19 (1992),
229244
[56] J. Pospiech, J. Jura, G. Gottstein, Statistical analysis of single
grain orientation data generated from model textures, Materials
Science Forum, 157-162 (1994), 407412
[57] K. Helming, Texture approximations by model components, Ma-
terials Science Forums, 273-275 (1998), 125132
204 BIBLIOGRAPHY
[58] L. Delannay, P. Van Houtte, A. Van Bael, D. Vanderschueren, Ap-
plication of a texture parameter model to study planar anisotropy
of rolled steel sheets, Modelling Simul. Mater. Sci. Eng., 8 (2000),
413422
[59] F. Roters, H.S. Jeon-Haurand, D. Raabe, A texture evolution
study using the texture component crystal plasticity FEM, Ma-
terials Science Forum, 495-497 (2005), 934-943
[60] S. Altmann, Rotations, Quaternions, and Double Groups,
Clarendon Press, Oxford (1986)
[61] T. Kuwabara, A. Van Bael, E. Iizuka, Measurement and analysis
of yield locus and work hardening characteristics of steel sheets with
dierent r-values, Acta Materialia, 50, (2002), 37173729
[62] M.A. Melchior, J.-F. Remacle, L. Delannay, Crystal-plasticity
based FE modelling of a dual-phase microstructure in which grains
have non-uniform shape and size, In : NUMIFORM 07, Materi-
als Processing and Design: Modeling, Simulation and Applications,
American Institute of Physics, 2007, pp.381-386.
[63] L. Delannay, M.A. Melchior, J.-F. Remacle, J.W. Signorelli, Inu-
ence of grain shape on the planar anisotropy of rolled steel sheets -
evaluation of three models, Comp. Mat. Sci. 45, (2009), 739-743
[64] C. Dobrzynski, M.A. Melchior, L. Delannay, J.-F. Remacle, A
simple mesh adaptation procedure for periodic domains, Int. J.
Num. Meth. Eng., to be published.
[65] L. Delannay, M.A. Melchior, P.J. Jacques, P. Van Houtte, Simula-
tion of the lattice strains developed during a tensile test on a multi-
phase steel, Materials Science Forum 495-497, (2005), 1627-1632
[66] J. Segurado, J. Llorca, Computational micromechanics of com-
posites : The eect of particle spatial distribution, Mechanics of
materials 38, 873883, 2006
[67] A. Bhattacharyya, E. El-Danaf, S.R. Kalidindi, R.D. Doherty.,
Evolution of grain-scale microstructure during large strain sim-
ple compression of polycrystalline aluminum with quasi-columnar
grains : OIM measurements and numerical simulations, Int. J.
Plast. 17, (2001), 861-883
BIBLIOGRAPHY 205
[68] O. Diard, S. Leclerq, G. Rousselier,G. Cailletaud, Evaluation of -
nite elements based analysis of 3D multicrystalline aggregates plas-
ticity. Application to crystal plasticity model identication and the
study of stress and strain els near grain boundaries, Int. J. Plast.
21, 691722,2005
[69] F. Barbe, L. Decker, D. Jeulin, G. Cailletaud, Intergranular and
intragranular behavior of polycristalline aggregates. Part 1: F.E.
model, Int. J. Plast. 17, 513536, 2001
[70] R. Quey, F. Barbe, Free meshing of microstructures based on mod-
ied Voronoi Tesselations, In: 17th Int. Workshop Computational
Mechanics of Materials, Paris (France), 22-24 August, (2007).
[71] C. Geuzaine, J.-F. Remacle, GMSH: A 3-D nite element mesh
generator with built-in pre- and post-processing facilities, Int. J.
Num. Meth. Eng. 79, (2009), 1309-1331
[72] S.K. Mishra, P. Pant, K. Narasimhan, A.D. Rollett and I. Samaj-
dar, On the widths of orientation gradient zones adjacent to grain
boundaries, Scripta Mat. 61, 273-276, 2009
[73] H. Resk, L. Delannay, M. Bernacki, T. Coupez and R. Loge, Adap-
tive mesh renement and automatic remeshing in crystal plasticity
nite element simulations, to be published.
[74] P. Franciosi, The concepts of latent hardening and strain harden-
ing in metallic single crystals, Acta Metall. 33 (1985), 16011612
[75] G. Winther, D. Juul Jensen, N. Hansen, Modeling ow stress
anisotropy caused by deformation induced dislocation boundaries,
Acta Materialla 45 (1997), 2455-2465
[76] B. Peeters, S.R. Kalidindi, C. Teodosiu, P. Van Houtte,
E. Aernoudt, A theoretical investigation of the inuence of dis-
location sheets on evolution of yield surfaces in single-phase BCC
polycrystals, J. Mech. Phys. Met. 50 (2002), 783807
[77] Z. Zhao, S. Kuchnicki, R. Radovitzky, A. Cuiti no, Inuence of in-
grain mesh resolution on the prediction of deformation textures in
fcc polycrystals by crystal plasticity FEM, Acta Mat. 55, (2007),
2361-2373
[78] L. Delannay, M.A. Melchior, A.K Kanjarla, P. Van Houtte,
J.W. Signorelli, CPFEM investigation of the eect of grain shape
206 BIBLIOGRAPHY
on the planar anisotropy and the shear banding of textured metal
sheets, In : Proceedings of Icotom 15, (2008), pp.paper S14-5, 1-12
[79] Q. Furnemont, The micromechanics of TRIP-assisted multiphase
steels, Ph. D. Thesis, Universite catholique de Louvain (UCL),
Belgium, (2003).
[80] P.J. Jacques, Q. Furnemont, S. Godet, T. Pardoen, K.T. Conlon,
F. Delannay, Crystallograpic micromechanical characterization of
TRIP-assisted multiphase steels by in situ neutron diraction,
Phil. Mag. A 86, 23712392, 2006
[81] M. Humbert, N. Gey, J. Muller, C. Esling, Determination of a
mean orientation from a cloud of orientations. Application to elec-
tron back-scattering pattern measurements, J. Appl. Cryst, 29,
(1996), 662-666

Das könnte Ihnen auch gefallen