Beruflich Dokumente
Kultur Dokumente
Gutachter
Prof. Dr. Rene Pinnau (Technische Universitat Kaiserslautern)
Prof. Dr. Martin Brokate (Technische Universitat M
unchen)
Disputation
2. Oktober 2007
D 386
For my family
Contents
I
II
INTRODUCTION
THEORY
13
.
.
.
.
15
15
19
20
26
.
.
.
.
.
.
.
.
.
.
.
29
30
30
31
31
32
34
36
39
43
48
50
53
53
60
63
4 Two generalisations
4.1 Generalisation for linear kinematic plus isotropic hardening material . . . . .
4.2 Generalisation for the class of continuous functions with bounded variation .
71
71
82
5 The
5.1
5.2
5.3
5.4
85
85
92
93
97
2 The
2.1
2.2
2.3
2.4
2.5
2.6
2.7
2.8
3 The
3.1
3.2
3.3
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
BVPs
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
III
PRACTICE
99
material
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
101
102
104
107
107
109
111
111
113
115
116
122
129
131
131
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
137
138
138
141
142
144
145
149
150
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
IV
APPENDIX
197
A Proof of Krej
cs theorem
B Miscellaneous
B.1 Two useful lemmas . . . . . . . . . . . . . . . . . . . .
B.2 Remarks on linear operators and function spaces . . .
B.3 Remarks on the Doring-Vormwald constitutive model .
B.4 Open challenges and questions . . . . . . . . . . . . .
C Notations
199
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
203
203
205
207
211
213
Part I
INTRODUCTION
introduction
3
Motivation
In order to give a reliable lifetime or damage prediction for an elastoplastic metallic body,
which is subjected to exterior loads, good knowledge of the local stresses and strains over
time is necessary. If measurements are not possible (experimentally or financially), numerical
analysis may be applied.
The usual process of finite element based fatigue analysis in practical engineering is depicted in
figure A. In any case, the starting point of a durability analysis is a finite element discretisation
of the body.
Finite Element Analysis
Difference? Estimates?
(this thesis)
?
Elastic BVP
(linear, quasistatic, superposition)
?
Tensorial elastic stresses and
strains e(x, t), e(x, t)
?q
Scalar elastic stresses and
strains e q (x, t), eq (x, t)
@
?
R
@
Elastoplastic BVP
(nonlinear, quasistatic, transient)
?
Multiaxial correction
(this thesis)
?
No/Uniaxial correction
(e. g. Neuber, ESED)
?
?
Tensorial stresses and strains
?
?q
Scalar stresses and strains
q (x, t), q (x, t) resp. q (x, t), q (x, t)
?
Uniaxial damage models, Hysteresis counting algorithms
?
Damage D(x)
hcf
lcf
Figure A. Possible ways in fatigue analysis. Here q denote some one-dimensional scalar equivalent magnitude, e. g. a signed norm or a projection of the tensorial quantity, the tildedenotes the correction of a (scalar
or tensorial) elastic quantity.
For low cycle fatigue analysis (lcf), the numerically expensive transient elastoplastic boundary value problem with an appropriate elastoplastic constitutive material law is solved in
order to receive the exact elastoplastic stresses (x, t) and strains (x, t). This is the right
hand branch of figure A.
In high cycle fatigue analysis (hcf), the linear elastic boundary value problem with Hookes
constitutive material law is solved to receive the elastic stresses e(x, t) and strains e(x, t).
Here the linear superposition principle is essential for the low computational costs. If the
locations, where the elastic stress crosses the yield stress of the material, are small in volume
introduction
and almost the whole body behaves elastic (the effect of elastic support), local corrections
may be applied,
after projection of the tensorial quantities e(x, t), e(x, t) to a scalar quantity e q (x, t),
e (x, t). Here clearly multiaxiality gets lost. For uniaxial corrections, Neuber s method,
q
cf. Neuber [90, 91], and the equivalent strain energy density (ESED) method, cf.
Glinka/Molski [45], are well established in the meantime. This corresponds to the
very left hand branch in figure A.
before projection on scalar quantities. Here multiaxial Neuber approaches, cf. Glinka
et al. [18, 83, 44, 104] and Chu [23], yield surface approaches, cf. Barkey et al. [3],
ttgen et al. [72, 73]
and pseudo parameter approaches have been suggested, cf. Ko
and Hertel [53]. This corresponds to the centred branch in figure A.
An example for hcf. In engineering durability, the exterior load functions are typically
sampled with a frequency of about 400 sec1 . Half an hour on a test track, i. e. 30 min
= 1800 sec, consequently yield 720 000 data points. So it is and will be in the next years
impossible to solve the full transient nonlinear problem in an acceptable computational time
with sufficiently fine discretisation of the body. Nowadays, practicians cannot help choosing
the leftmost branch in figure A.
At the end of the fatigue analysis, which is common for all branches, the accumulated damage
is computed, where it is widespread consensus to use the uniaxial Rainflow counting method,
see e. g. Dressler et al. [4, 15, 35, 36, 37], Seeger [26], Rychlik [101] or Hack [48].
Thereby, for the damage quantification, a lot of damage parameters have been developed.
They assign a certain amount of damage to a given closed hysteresis loop. Some of the most
established ones are those of Manson, Coffin and Morrow [84, 28, 87] or Smith, Watson and Topper [105]. These methods are usually available in commercial software tools for
fatigue analysis nowadays.
Independently of which special kind of elastoplastic constitutive material law we consider, the
elastoplastic BVP model (nonlinear, hysteresis operator with memory structure
between stress and strain , energy dissipation, damage accumulation, irreversible process!)
is much closer to nature than the
elastic BVP model (linear, one-to-one operator between stress e and strain e, no
energy dissipation, no damage, fully reversible process!).
However, for stresses, which are below the yield stress, they coincide. (Cf. lemmas 2.3.5 and
4.1.3 as well.)
introduction
difference estimates exist so far. We answer this question in chapters 2 and chapters 4
for linear kinematic and linear kinematic plus isotropic hardening material. It turns out
that the difference of stresses, strains and displacements can be expressed as composition
of some linear operators and a play operator with respect to the exterior forces.
So far, no mathematical literature exists about those multiaxial correction schemes. We
give a first mathematical approach on this subject in chapters 3 and 5, where we try to
ttgen et al. [72, 73] and Hertel [53]
justify the multiaxial correction methods of Ko
in an analytical setting.
So far, computational results for the correction approaches [3, 18, 23, 44, 53, 72, 73,
83, 104] differ excessively from the numerical transient elastoplastic boundary value
problem solution. We suggest some improvements for the correction methods in [53, 72,
73] and apply large-scale parameter identification techniques, yielding in a significant
improvement of existing results. This is done in chapters 6, ..., 8 of this thesis.
Before starting, we have to give a more detailed exposition of the basic relations. An overview
about general notations is given in appendix C.
Mathematical setting
Let R3 be a bounded domain, representing the elastoplastic metallic body, that is
subjected to exterior loads. We assume, that its boundary is sufficiently smooth, e. g. of
class C 1 , and that it decomposes into = 0 1 , where 0 , 1 are relatively open
and disjoint, 0 1 = , satisfying Area(0 ) > 0.
each undergo a nonlinear multiaxial constitutive hysteretic material law of
The points x
the form
Z t
C ep (x, ), (x,
), v1 (x, ), . . . , vn (x, ) (x,
) d,
(1)
(x, t) = H[(x, )](t) =
0
where the elastoplastic tensors C ep R3333 depend on the whole history of the material
and its points, condensed in a set of internal state variables vi , representing the memory of
the material at each point x. Examples for H are the constitutive models of
Melan [86], Prager [95, 96], Ziegler [114], Besseling [5], Bower [7], Mroz [16, 88, 89,
13], Garud [43], Ohno-Wang [92, 93], Chu [22], Chaboche [12, 20, 21], ArmstrongFrederick [2], W. Jiang [68, 69], Y. Jiang-Sehitoglu [63, 64, 65, 66, 67], Tanaka
[106], Doring-Vormwald [33, 34, 54], and many more ...
(EPK)
(EPKI)
Linear kinematic plus isotropic hardening material [9, 14, 75, 76],
(EPJ)
introduction
Model (EPK) is as well called the model of Melan [86] and Prager [95, 96]. The model
(EPJ) of Y. Jiang and H. Sehitoglu is quite successful in practice and widely accepted
in the engineering community. We briefly name it Jiangs model. It reproduces many transient multiaxial stress-strain phenomena like cyclic hardening/softening or ratchetting. An
overview of all represented material effects may be found in table (6.36).
Let exterior surface forces g : 1 [0, T ] R3 and exterior volume forces f : [0, T ] R3
be given. The nonlinear transient elastoplastic initial boundary value problem for
the elastoplastic displacement u : [0, T ] R3 ,
the elastoplastic strain : [0, T ] R33
s ,
the elastoplastic stress : [0, T ] R33
s
with hysteresis is given by
u + DIV (I + u) + f = 0
= 21 u + ut + uut
= H[]
with the boundary conditions
u = 0 on 0
(I + u)n = g on 1
u(, 0) = u0
u(,
0) = u 0
on
in .
in
(2)
(3)
(4)
Here always DIV = DIVx and = x with respect to the space variable. H is a hysteresis
operator of the form (1), describing the constitutive elastoplastic material behaviour.
The unique solvability of the full nonlinear problem (2), ..., (4) is very hard in general. We
restrict ourselves to the problem with some simplified assumptions: We assume geometric linearity and neglection of dynamic and viscous effects. Then (2), ..., (4) reduce to the quasistatic
nonlinear transient elastoplastic boundary value problem with linearised geometry
DIV + f = 0
in
(5)
= 21 u + ut
= H[]
and the initial condition
u = 0
n = g
on 0
on 1
u(, 0) = u0
on
in .
(6)
(7)
So far, the only elastoplastic material laws, for which system (5), ..., (7) is known to be uniquely
solvable, are the following.
introduction
Absence of any hardening, i. e. perfect plasticity, cf. Temam [107], not considered in
this work.
ger [47], Zeid Linear kinematic hardening (EPK), cf. Han/Reddy [49, 50], Gro
ler [113] and theorem 2.5.1. The governing relations in operator form are given by
(2.7), ..., (2.11) later.
Linear kinematic hardening plus isotropic hardening (EPKI), cf. Han/Reddy [49, 50]
and theorem 4.1.4. The governing relations in operator form are given by (4.1), ..., (4.6)
later.
Nevertheless, there exist algorithms in finite element codes for materials, for which the
uniquely solvability of (5), ..., (7) is not guaranteed so far.
For the same exterior forces g : 1 [0, T ] R3 and f : [0, T ] R3 , the linear
quasistatic elastic boundary value problem for
the elastic displacement eu : [0, T ] R3 ,
the elastic strain e : [0, T ] R33
s ,
the elastic stress e : [0, T ] R33
s
is given by
DIV e + f = 0
e = 1 eu + eut
2
e = C e
eu
en
= 0
= g
on 0
on 1
in
on ,
(8)
(9)
for each fixed t [0, T ] being a static problem. C(x) = (cijkl (x)) R3333 is Hookes
tensor of elasticity with generally x-dependent components
cijkl (x) = (x) ij kl + (x) (ik jl + il jk ),
where
(x) =
E(x)(x)
,
1 + (x) 1 2(x)
(x) =
E(x)
,
1 + (x)
(10)
(11)
cf. Reismann [98] or section 2.3.4 in Leipholz [80]. The parameters are
E(x), (x) : Youngs modulus and Poissons ratio at x,
(x), (x) : Lames constants at x,
(x) :
density at x.
Korns inequality
Z
kv(x) + v(x)t k2 dx 2 () kvk2W 1,2 ()
v W01,2 (, 0 , R3 )
(12)
is well known and the basic constituent of the proof of existence and uniqueness of weak
solutions of system (8), (9), see e. g. Duvaut/Lions [38], section III.3.3, or Ciarlet et al.
[25, 24] for the static case. For the quasistatic elastic case, which is usually not mentioned in
the mathematical literature, we refer to this thesis:
introduction
Quasistatic linear elastic boundary value problem (E), cf. theorem 2.4.1. The governing
equations in operator form are given by (2.5), (2.6) later.
For a proof of (12) we refer to Duvaut/Lions [38] or section 62.15 in Zeidler [113].
Some comments about the terminology. Whenever the adjective elastic occurs, we mean
the eu, e and e of the quasistatic elastic boundary value problem with linearised geometry.
Elastic always means linear elastic.
Whenever the adjective elastoplastic occurs, we mean the u, and of the quasistatic
elastoplastic boundary value problem with linearised geometry.
If the adjective true is applied, we mean experimental strain or stress measurements tr, tr.
(This will be only used in chapter 8, where measurements will be benchmark- tests.)
Due to the linearity, the quasistatic elasticity problem decouples into time t and space x. This
means: If the exterior forces f , g are given by a linear superposition of finitely many static
exterior forces, the so-called load cases
fi L2 (, R3 ),
i. e.
f (x, t) =
g(x, t) =
Xi
i
gi L2 (1 , R3 ),
(x, t) [0, T ]
Li (t) fi (x),
(x, t) 1 [0, T ]
Li (t) gi (x),
ui : R 3 ,
i : R33
s ,
(13)
i : R33
s
denote the solution triples of the static problem (8), (9), that are corresponding to the fi , gi ,
then the elastic quasistatic stress and strains tensor and displacement fields are given by
X
X
X
e
e
e
(x, t) =
Li (t) e i (x),
(x, t) =
Li (t) ei (x),
u(x, t) =
Li (t) eui (x) (14)
i
for (x, t) [0, T ]. This is easy to verify: Inserting (13), (14) into (8) and (9), using the
linearity and the fact that for each i, the triple e i , ei , eui is a solution of the corresponding
static problem for the exterior forces gi , fi , we arrive at
X
X
1X
in
Li (t) eui (x) + euti (x)
Li (t) ei (x)
2 Xi
Xi
Li (t)e i (x) C
Li (t) ei (x)
i
and
on 0
on 1
on .
Here denotes equality of each summand. And it is this decoupling property, which makes
the elastic approach very cheap and attractive in engineering applications, compared to the
costs when solving the full transient problem with the exact nonlinear stress-strain relationship (1). In section 2.4, the quasistatic elastic problem is considered in purely space-time
decoupled form.
denote a fixed point.
We especially point out a special case. To this end, we let
introduction
en
First, it is possible to obtain the stresses e and from its deviator in an easy fashion,
cf. remark 5.1.2.
Second, in most practical examples, the stress maxima are located on the boundary, cf.
Neuber [90] or Wheeler [109].
Third, in most practical examples, the points of highest accumulated damage, where
cracks emerge and begin to grow, turn out to lie at the boundary, cf. Lemaitre/Cha ring [34].
boche [79] or Do
But there should be paid attention, since a general maximum principle even just for the static
linear elastic boundary value problem does not exist, see Polya [94] or Langhaar/Stippes
[78] or section 4 in Wheeler [109]. (Note that the problem is elliptic, but not scalar.)
If lies in an area, where the true stress is high, e. g. in a notched region, plasticity comes
into play, so almost all practical experiences indicate this
the elastic stress e overestimates the elastoplastic stress (and the true stress tr).
the elastic strain e underestimates the elastoplastic strain (and the true strain tr).
Typically these regions are small, and the overall material behaviour is elastic. This is usually
called the effect of elastic support. The goal of the correction model, presented in this thesis,
is to rectify e(, t), e(, t) to
(, t), (, t), from now on referred to as the corrected stress
resp. strain, by an appropriate modification/extension of the constitutive elastoplasticity
model in order to get closer to the elastoplastic solution (, t), (, t). Eventually, this is the
point, on which hysteresis counting may be set up.
10
introduction
The former part gives some new mathematical insights into the connection between the the
elastic and elastoplastic boundary value problems.
The latter part shall satisfy the everyday computing reader, who is confronted with engineering applications, and give impulses for further mathematical research. Here just small pieces
can be proven, but truly practical improvements are presented.
In more detail, this thesis is structured in the following way.
Chapter 1 resumes some standard Lipschitz estimates for the so-called stop and the play
operator, which are the basic tools for the description of elastoplastic constitutive material
behaviour. These estimates can be generalised in the sense that different input functions, different initial memories and different scalar products are admitted. They are the main devices
for the proof of the estimates of the next chapters.
Chapters 2, 3, 4 are the main body of this work. Mathematical analysis is done for the
elastic and elastoplastic boundary value problems. These chapters cover:
The differences e , e and euu of the solutions of the elastic (E) and elastoplastic
boundary value problems (EPK), (EPKI) can be written as a composition of some linear
operators and a play operator. Especially, the model solution differences may grow at
most linearly in the total variation of the applied exterior forces. See sections 2.7 and
4.1.
As a consequence, we prove that it is possible to get the elastic stress e out of the
elastoplastic stress , both for (EPK) and (EPKI) material, and vice versa.
We prove that it is possible to estimate the elastic stress e by a constant times the
elastoplastic stress , which is possible independently of the constitutive plastic material
law. This result may be considered as a kind of main lemma for hcf analysis, where
no correction is applied at all (in other words: if the correction is just the identity).
This is done in section 2.8.
We give new input and parameter robustness proofs for the elastoplastic boundary value
problems as an easy consequence of the generalised Lipschitz estimates for stop and play
of chapter 1. Both (EPK) and (EPKI) material is considered, cf. sections 2.6 and 4.1.
The wellposedness of the boundary value problems for (EPK) and (EPKI) material
is generalised for the function class W 1,q ([0, T ], ), for arbitrary 1 q , and
CBV ([0, T ], ), cf. section 2.5 and 4.2.
A global correction model for (EPK) material is introduced in chapter 3, for which error
estimates are given. Localisation under the assumption of the effect of elastic support
is done.
Chapter 5 transfers the correction idea to set up a pointwise correction model for (EPK)
material. Error estimates for the pointwise correction model are given. The analysis is performed here without the surrounding partial differential coupling.
introduction
11
In chapter 6, we transfer the pointwise correction model from chapter 5 to (EPJ) material.
We concentrate on Jiangs model, whose wellposedness is not yet established, because of the
following reasons.
For practical application purposes, it is indispensable to consider an elastoplastic material law, which contains enough complexity to reflect enough aspects of nonlinear,
transient multiaxial hysteretic material behaviour. From the engineers point of view,
there is too much mathematical/structural idealisation in the (EPK) and (EPKI) models. The (EPJ) model instead reproduces ratchetting, non-Masing behaviour, transient
cyclic hardening and softening, etc.
The (EPJ) model is formulated by differential equations. This is the most general kind of
formulation for constitutive elastoplastic material laws. Every constitutive elastoplastic
material law on page 5 can be written as such a system of (discontinuous) differential
equations, see Lemaitre/Chaboche [79].
We do not want to overkill the reader with full generality (1) and concentrate ourselves
on (EPJ) material. The translation to the other laws on page 5 is then straightforward.
We can prove a collection of lemmas for the (EPJ) model, which will hopefully give
rise to a final wellposedness proof, both for the stress-strain law in isolation and for
the associated boundary value problem. So far, no analytical literature exists that is
concerned with the (EPJ) model.
In chapter 7 we expose a fast numerical method to implement the pointwise correction model
of chapter 6, where we concentrate on (EPJ) material. The algorithm can straightforwardly
be transferred to all other constitutive laws, which are contained in the collection of page 5.
Our algorithm is very fast and designed for the case that the elastic stress tensor is piecewise
linear in time.
The results for the corrected stress can be significantly improved introducing Seegers relation as an additional constraint, cf. Hoffmann/Seeger [59, 60, 61]. A predictor-corrector
method for the compliance of such empirical relations is presented. The modified algorithm
is still very performant. As it yields very good approximations for all the three plane stress
tensor components, it can be used in connection with multiaxial Rainflow methods in high cycle fatigue analysis, especially with the so-called critical plane approach, see [48] for example.
This is a crucial improvement to the widespread uniaxial corrections like Neubers method or
the ESED method.
Chapter 8 gives numerical results both for (EPJ) and (EPK) material. A gradient based
least-squares algorithm for parameter identification is applied to the pointwise correction
model. The gradient evaluation with respect to the parameters, which is very large in scale
for (EPJ) material, is realised with the modern algorithmic/automatic forward differentiation
technique. It makes the parameter identification feasible for practice. The results of the
correction model agree very fine both with numerical stress-strain analyses (transient elastoplastic finite element method) and physical strain measurements.
With our correction model, we lie somewhere between lcf and hcf: It is fast enough for hcf,
cf. chapter 7 and remark 8.3.2, the estimates will assure a small error for lcf, if appropriate
elastic parameters have been determined, cf. theorem 3.1.3.
12
introduction
Part II
THEORY
13
Chapter 1
behaviour are the so-called stop and play operators. Here the works of Brokate/Krejc
[9], ..., [14], [75], ..., [77], Krasnoselskii/Pokrovskii [74] or Visintin [108] have been pathbreaking. We will need those operators in this whole part II, e. g. to express the difference of
the solutions of the elastic and elastoplastic boundary value problems in chapters 2 and 4.
This first chapter has preparing character. In section 1.1, we give a definition of stop and
play that allows to take different scalar products into account. Krejcs theorem can be
straightforwardly adapted, see theorem 1.1.1 and appendix A for a detailed proof.
Section 1.3 gives some Lipschitz estimates for the differences of stop/play operators that
correspond to different scalar products. We need the latter in order to estimate the model
differences in chapters 2 and 4 and in order to give estimates for our correction model in
chapter 3. The main theorem of this chapter is theorem 1.3.1. Further in section 1.4, we
prove some useful transformation formula for stop and play.
The results of this chapter are formulated for both the AC = W 1,1 and the CBV class of
input/output functions.
1.1
Definitions, Wellposedness
For the scalar product h, i on a separable Hilbert space X, a convex and closed characteristic
C X such that 0 C, an initial memory s0 C and a symmetric, strongly positive operator
A L(X, X), there exists a unique decomposition of the unity
I = PC,A,h,i + SC,A,h,i
(1.1)
(1.2)
into operators
named
play p(t) = PC,A,h,i (f, s0 )(t)
and
for t [0, T ],
16
the stop and the play operator w.r.t. different scalar products 1
whose outputs are defined by the uniquely determined solution of the (differential) evolution
variational inequality
Z T
A dp( ), (s )( ) 0
for all C([0, T ], C)
0
s(t) + p(t) = f (t)
e. in [0, T ] .
(1.3)
s(t)
C
e.
in
[0,
T
]
s(0) = s0
for CBV resp.
1
A p(t),
s(t)
s(t) + p(t)
s(t)
s(0)
0
f (t)
C
s0
for all C
a. e. in [0, T ]
e. in [0, T ]
.
e. in [0, T ]
(1.4)
for W 1,q . Here the integral in (1.3) is to be understood in the Riemann-Stieltjes sense with
Hilbert-space valued integrands. The spaces CBV and W 1,q are defined by
W 1,q ([0, T ], X) = f AC([0, T ], X) : f Lq ([0, T ], X) ,
(1 q )
CBV ([0, T ], X) = C([0, T ], X) BV ([0, T ], X),
(1.5)
for f W 1, ([0, T ], X)
(1.6)
kf kW 1,q, = kf (0)k +
for 1 q < ,
for q = . Here
Z
kf( )kq d
kf k = max kf ( )k
[0,T ]
sup
N
X
kf (i ) f (i1 )k =
T
0
kf( )kd
is the total variation of f CBV ([0, T ], X). In the latter, the left hand identity is the general
definition in CBV , the right hand equality holds on the subspace AC = W 1,1 . Note that the
inclusions
W 1,q ([0, T ], X) W 1,1 ([0, T ], X) = AC([0, T ], X) CBV ([0, T ], X) C([0, T ], X)
for 1 q hold.
Krejcs theorem 4.1 [76] can be easily adapted for scalar products, which are possibly different
than the original one. A detailed proof of the next theorem is given in appendix A.
1 the stop and the play operator w.r.t. different scalar products
17
1.1.1 Theorem (Stop and play operator, CBV and W 1,q ) Let (X, h, i) be a separable Hilbert space, C X convex and closed, containing the origin, and A L(X, X) symmetric
and strongly positive. Then the following assertions hold.
(a) For each f CBV ([0, T ], X), s0 C, there exist uniquely determined functions p and
s in CBV ([0, T ], X), such that (1.3) holds.
(b) For each f W 1,q ([0, T ], X), s0 C, there exist uniquely determined functions p and s
in W 1,q ([0, T ], X), such that (1.4) holds.
Here 1 q in (1.4) .
Proof: According to section B.2 in the appendix, h, iA = hA1 , i defines an equivalent
scalar product on X. Thus the topologies on X X , induced by k k and k kA , are equal.
Thus (X, k kA ) is as well separable, and C is as well closed with respect to k kA . Clearly,
the origin remains in C, and the convexity, which is a purely algebraic property, is as well not
affected.
[75] or theorem 4.1 in Krejc
[76] to (X, k kA ) to receive
Now apply theorem 3.1 in Krejc
(a). Apply proposition 3.9 in [75] or proposition 4.1 in [76] to (X, k kA ) to receive (b).
1.1.2 Remark. (a) In fact, (1.3) enjoys a higher generality than (1.4). Clearly, (1.4) implies (1.3)
[75].
because of lemma 1.25 in Krejc
(b) Note that the proof above does not need Int(C) 6= , cf. remark 1.2.1 as well. (For higher
regular sets C, even inputs of class C([0, T ], X) may be admitted, cf. section 4.2 in [76].)
(c) If 1 q < , stop and play are continuous w.r.t. the norm k kW 1,q, , cf. theorem 4.2 in
[76], and thus w.r.t. k kW 1,q due to lemma 1.1.3 below. For q = this is not true
Krejc
any more, see the counterexample in remark 4.3 in [76].
We introduce
q =
q/(q 1) if q > 1
,
1
if q = 1
thus
q
= q 1,
q
(1.7)
the (quasi) H
older conjugate for a given real number 1 q < .
1.1.3 Lemma. Let (X, h, i) be a separable Hilbert space. On the space W 1,q ([0, T ], X), the standard Sobolev norm
kf kW 1,q =
Z
kf ( )k d
1/q
Z
T
0
kf( )kq d
1/q
(1.8)
and the non-standard Sobolev norm (1.5) are equivalent for each 1 q < . Additionally,
we have
kf (t)k kf k kf kW 1,1,
(1.9)
Proof: (1) Norm k kW 1,q, is stronger than norm k kW 1,q . First the case q = 1. According
18
the stop and the play operator w.r.t. different scalar products 1
kfkdt
f ( )d
dt +
=
f (0) +
0
0
0
Z T
Z T
Z T
kfkdt
kf kd dt +
kf (0)k +
0
T kf (0)k + (T + 1)
(T + 1)kf kW 1,1, .
kfkdt
Now the case 1 < q < . Let q be defined by (1.7). Then with Holders inequality, there
follows
Z t
q Z t
q (HI)
Z t
q/q
( )kd
( )d
kf( )kq d
(1.10)
t
1
k
f
f
and
Z t
q
Z
(1.10)
f d
dt
0
kfkq d
Z
q1
dt
Tq
=
q
kfkq d.
(1.11)
kf (t)k dt
+
kf (t)k dt
=
kf kW 1,q
0
(F T )
Z
(M I)
(1.11)
Z
q 1/q Z T
1/q
Z t
q
f (0) +
kf k dt
+
f ( )d
dt
0
0
q 1/q Z T
1/q
1/q Z T
Z t
q
q
kfk dt
+
fd
+
kf (0)k dt
dt
Z
T
+
1
1/q
T 1/q kf (0)k +
const(T, q) kf kW 1,q, .
kf( )kq d
1/q
(2) Norm k kW 1,q is stronger than norm k kW 1,q, . Due to the continuous embedding (see
Evans [41], chapter 5.9, theorem 2, or Zeidler [110], proposition 23.23)
W 1,q ([0, T ], X) C([0, T ], X)
(for each 1 q )
such that
kf k const(T )kf kW 1,q ,
where the constant can be chosen in dependence only on T , we easily find the contrary
inequality
kf kW 1,q, (1 + T )const(T )kf kW 1,q .
(3) Additionally, let s argmax kf (t)k, t [0, T ] , then there follows
Z s
Z s
kf k = kf (s)k =
f (0) +
f ( )d
kf (0)k +
kf( )kd kf kW 1,1, .
0
1 the stop and the play operator w.r.t. different scalar products
19
1.1.4 Lemma. Let X be a vector space with two norms k k1 and k k2 , such that (X, k k1 ) and
(X, k k2 ) are Banach spaces. Then k k1 ck k2 automatically implies k k2 Ck k1 .
Proof: See Heuser [55], exercise 39.6. Use the open mapping theorem.
1.2
From the defining variational inequality (1.4), by constructing difference quotients and passing
[75], one sees that an A-orthogonal
to the limit to zero, cf. proof of theorem 3.9 in Krejc
decomposition of the input derivative
p(t),
s(t)
=0
a. e. in [0, T ],
A
2
2
kf(t)k2A = ks(t)k
A + kp(t)k
A
from which
kp(t)k
A kf (t)kA ,
a. e. in [0, T ],
ks(t)k
A kf (t)kA
a. e. in [0, T ].
(1.12)
q
1/q
,
(1.13)
kf( )kA d ks(0)kA + t
kfkA d
ks(t)kA ks(0)kA +
0
(F T )
kp(t)kA kp(0)kA +
(HI)
Z
kfkqA d
1/q
(1.14)
everywhere in [0, T ], if f W 1,q ([0, T ], X). Here q is defined by (1.7). Further, we have for
two decompositions
pi = PA (fi , si,0 ),
si = SA (fi , si,0 ),
f i = pi + s i
1
and = 2 = 1 2 the standard Lipschitz estimates
for stop
ks(t)kA ks0 kA +
for play
kp(t)kA
(F T )
ks0 kA + kf (t)kA +
(i {1, 2})
kf( )kA d,
0Z
(1.15)
kf( )kA d
t
kf( )kA d
ks0 kA + kf (0)kA + 2
0
Z t
kf( )kA d ,
kp(0)kA + 2 kf (0)kA +
0
(1.16)
20
the stop and the play operator w.r.t. different scalar products 1
with respect to both norms. The constants w. r. t. k kW 1,q, are equal to 1 resp. 2. For the
ones w. r. t. k kW 1,q cf. the proof of the lemma 1.1.3.
1.2.1 Remark. (a) The proofs of the estimates (1.15) and (1.16) do not need the interior of C
being non-empty.
[14], propositions A.1 and A.3, can be
The assumption 0 Int(C) in Brokate/Krejc
[75], theorem 1.9, proposition 3.9 and remark
weakened to 0 C. For this, see Krejc
3.10.
With the aid of the Gronwall-kind lemma B.1.1, one can get rid of the factor 2 before
[75].
the integral in (3.24) of remark 3.10 in Krejc
(b) Note that the smallness result (1.13) for stop follows as well from the Lipschitz result
(1.15) for stop by taking f1 f , f2 0, s1 (0) = s(0), s2 (0) = 0 as inputs.
(c) The detailed proof of theorem 1.1.1 in appendix A for CBV functions shows the following
Sub-Pythagoras relation for the piecewise linear interpolants N s, N p, N f , which are
used therein in order to construct a solution, i. e.
(A.6)
ksn kA kfn kA .
A (N s) A (N f ) A (f )
for each N N. Taking the lim inf N on both sides, with proposition V1.18 in Krejc
[75], we conclude
A (p) A (f ),
A (s) A (f ),
(1.17)
in generalisation of (1.12).
1.3
We now generalise Brokates and Krejcs results, in the sense that we allow different input
functions f1 , f2 , different initial memories s0,1 , s0,2 and different scalar products hA1
1 , i,
1
hA2 , i, which are induced by different symmetric and strongly positive operators A1 , A2 .
It turns out, that the output of stop and play depends as well in a Lipschitz manner on these
scalar products. We first consider the class of W 1,1 = AC functions. The main tool of proof
is the Gronwall-kind lemma B.1.1.
1 the stop and the play operator w.r.t. different scalar products
21
1.3.1 Theorem (Stop operator, W 1,1 ) Let (X, h, i) be a separable Hilbert space, A1 , A2
L(X, X) symmetric and strongly positive, inducing equivalent scalar products
x, y A1 = A1
x, y A2 = A1
(1.18)
1 x, y ,
2 x, y
and equivalent norms
(1.19)
Let further a convex, closed subset C X, such that 0 C, and two inputs
f1 , f2 W 1,1 [0, T ], X ,
s0,1 , s0,2 C,
be given. For the outputs
p1 = PA1 (f1 , s0,1 ),
Z t
CA1
1
1
kf ( )k + 2 kA p2 ( )k d,
cA1
cA1
0
Z t
CA2
CA2
1
1
ks0 k +
kf ( )k + 2 kA p1 ( )k d.
cA2
cA2
cA2
0
CA1
ks0 k +
cA1
1
= 2 = 1 2 ,
especially
pi (t), si (t) A 0
si (t) C
si (0) = si,0
(1.20)
(1.21)
1
A1 = A1
1 A2 .
(1.22)
Letting s1 (t) and s2 (t) chiastically into the defining variational inequalities (1.22), we see
(
p 1 , s1 s2 0
p 1 , s1 s2 A1 = A1
1
,
p 2 , s2 s1 A2 = A1
0
2 p 2 , s2 s1
thus by addition
1
A1
1 p 1 A2 p 2 , s
0.
(1.23)
1
1
1
Adding a zero term 0 = A1
1 p 2 + A1 p 2 resp. 0 = A2 p 1 + A2 p 1 in the left side and
using the second relation in (1.22), we get after rearranging both
and
1d
1
ksk2A1 = A1
s A1
1 s,
1 f + A p 2 , s
2 dt
1d
1
ksk2A2 = A1
s A1
2 s,
2 f + A p 1 , s .
2 dt
(1.24)
(1.25)
22
the stop and the play operator w.r.t. different scalar products 1
(F T )
(CS)
(1.19)
Z t
f, s
A1
+ A1 p 2 , s d
Z t
1
Z t
0
1
1
kA p 2 k d.
CA1 kf k +
cA1
(1.26)
Z t
0
1
1
CA1 kf k +
kA p 2 k d,
cA1
which finally gives (1.20). The same procedure for p1 instead of p2 , h, iA2 instead of h, iA1 ,
h, iA1 instead of h, iA2 together with (1.25), (1.19) and (1.18) yields (1.21).
We directly have the following consequences.
1.3.2 Corollary (Special cases, W 1,1 ) Let the assumptions of theorem 1.3.1 hold. Then we have
especially the following estimates.
(a) Same inputs, different scalar products. If f1 = f2 , then there holds
s(t) = p(t)
(1.27)
and
kp(t)k = ks(t)k
1
CA1
ks0 k + 2
cA1
cA1
kp(t)k = ks(t)k
CA2
1
ks0 k + 2
cA2
cA2
Z
Z
t
0
t
0
kA1 p2 ( )kd,
(1.28)
kA1 p1 ( )kd.
(1.29)
ks(t)k
kf ( )kd .
(1.30)
ks0 k +
cA
0
(c) Original scalar product each, different inputs. If A1 = A2 = A = I, which allows
cA1 = cA2 = 1 = CA1 = CA2 , then
Z t
kf( )kd,
ks(t)k ks0 k +
0
1 the stop and the play operator w.r.t. different scalar products
23
1.3.3 Corollary (W.r.t input, W 1,1 ) Let the assumptions of theorem 1.3.1 hold. Then in any
case, we have the estimates
Z t
Z
CA2 kA1 k t
CA1
kf ( )kd +
ks0 k+
kf2 ( )kd, (1.31)
ks(t)k
cA1
c2A1 cA2
0
0
Z t
Z
CA1 kA1 k t
CA2
kf( )kd +
ks0 k+
kf1 ( )kd
(1.32)
ks(t)k
cA2
cA1 c2A2
0
0
with respect to the inputs everywhere in the interval [0, T ].
Proof: Just note
kA1 p 2 k
(1.19)
kA1 k kp2 k
1
cA2
(1.12)
(1.19) C
A2
kA1 k kf2 kA2
kA1 k kf2 k
cA2
cA2
1.3.4 Corollary (Play operator, W 1,1 ) Let all assumptions of theorem 1.3.1 hold. Then we have
in any case
Z t
Z
CA2 kA1 k t
CA1
kfkd +
ks0 k +
kp(t)k kf (t)k +
kf2 kd, (1.33)
cA1
c2A1 cA2
0
0
Z
Z t
CA2
CA1 kA1 k t
kf1 kd (1.34)
kp(t)k kf (t)k +
kf kd +
ks0 k +
cA2
cA1 c2A2
0
0
everywhere in the interval [0, T ].
Proof: Note s(t) = f (t) p(t), implying
ks(t)k kp(t)k kf (t)k kp(t)k kf (t)k.
1.3.5 Theorem (Play operator, W 1,1 ) Let the assumptions and notations of theorem 1.3.1 hold.
Then we have
Z t
2
CA
2
2
1
1
2
1
kp(t)k
(A p),
f A p2 , p d,
kp(0)k + 2
c2A1
cA1 0
Z t
2
CA
2
1
1
2
2
2
(A p),
f A p1 , p d.
kp(0)k + 2
kp(t)k
c2A2
cA2 0
24
the stop and the play operator w.r.t. different scalar products 1
Here especially
1
A1 = (A1 ) = A1
1 A2 ,
1
(A1 p)
= A1
1 p 1 A2 p 2 .
1
A1 p1 A1
A1
2 p 2 , p
1 p 1 A2 p 2 , f ,
thus by adding appropriate zeros
1d
kpk2A1 = p,
p A1
2 dt
1d
kpk2A2 = p,
p A2
2 dt
(A1 p),
f A1 p2 , p ,
(A1 p),
f A1 p1 , p .
(1.36)
(1.37)
1.3.6 Corollary (Special cases, W 1,1 ) Let all assumptions of theorem 1.3.5 hold. Then we have
especially the following estimates.
(a) Same inputs, different scalar products. If f1 = f2 , then there holds
ks(t)k = kp(t)k
1
CA1
kp(0)k + 2
cA1
cA1
ks(t)k = kp(t)k
1
CA2
kp(0)k + 2
cA2
cA2
kA1 p2 k d,
kA1 p1 k d,
kp(t)k
2
CA
2
kp(0)k2 + 2 kf k,t
2
cA
cA
kA1 p(
)k d.
t
0
kp(
)k d,
(CS),(1.19)
1
1
kp(0)k2A1 +
2
cA1
kA1 p 2 k kpkA1 d.
1 the stop and the play operator w.r.t. different scalar products
25
kp(0)kA1 +
1
cA1
kA1 p 2 k d
and, in view of (1.19), the assertion. (b), (c) Use the Cauchy-Schwarz inequality.
The transfer of theorem 1.3.1 and corollary 1.3.3 for CBV functions is straightforward. But
lemma B.1.1 is not applicable. But we may argue in a different way. (Then we have to accept
a factor of two.)
1.3.7 Theorem (Stop operator,
CBV ) Let the assumptions of theorem 1.3.1 hold with inputs
f1 , f2 CBV [0, T ], X . Then there follows
ks(t)k
ks(t)k
CA1
ks0 k + 2(f ) +
cA1
CA2
ks0 k + 2(f ) +
cA2
c2A1
2
c2A2
(A1 p2 ),
(1.38)
(A1 p1 ).
(1.39)
Proof: Similar as in the proof of 1.3.1, we find, cf. (1.24) and remark 1.1.2,
Z
d(A1
1 s), s
d(A1
1 f ), s
d(A1 p2 ), s .
[75] and the fundamental estimate for RiemannThus with the aid of exercise 1.24 in Krejc
Stieltjes integrals
1
ks(t)k2A1
2
1
ks0 k2A1 + kskA1 , A1 (f ) + ksk (A1 p2 )
2
1
1
1
kskA1 ,
ks0 kA1 + CA1 (f ) +
(A p2 ) .
2
cA1
Here the maximum-norm is taken over the whole interval [0, T ]. Therefore, taking the maximum at the left hand side, we arrive at
1
1
1
ksk,A1 ks0 kA1 + CA1 (f ) +
(A1 p2 )
2
2
cA1
after cancelling. This yields (1.38) with the aid of (1.19).
1.3.8 Corollary (W.r.t input, CBV ) Let the assumptions of theorem 1.3.7 hold. Then in any
case, we have with respect to the inputs
2CA2 kA1 k
CA1
ks0 k + 2(f ) +
(f2 ),
cA1
c2A1 cA2
2CA1 kA1 k
CA2
(f1 )
ks0 k + 2(f ) +
ks(t)k
cA2
cA1 c2A2
ks(t)k
(1.40)
(1.41)
26
the stop and the play operator w.r.t. different scalar products 1
Proof: Use the following rules for the total variation, cf. (1.19), namely
(A1 ) kA1 k (),
(i = 1, 2)
1.4
In this section, we present just one lemma, which gives some transformation formula for stop
[76], section 2.3, for the
and play. Its part (b) is a generalisation of lemma 2.1 in Krejc
infinitely dimensional Hilbert space case. The proof is adapted straightforwardly.
1.4.1 Lemma (Transformation rules for stop and play) Let (X, h, i) be any real separable
Hilbert space, C X convex, closed, containing the origin. We identify X X. Then the
following transformation formula for stop and play hold.
(1) For each symmetric and strongly positive A L(X, X ), there holds
PA (p + s, s0 ) = AP(A1 p + s, s0 ),
SA (p + s, s0 ) = AS(A
SA (p + s, s0 ) =
S(A
(1.42)
p + s, s0 ) + (I A)s,
(1.43)
p + s, s0 )
(1.44)
s = SA (p + s, s0 ).
(1.45)
(2) For each symmetric and strongly positive A1 , A2 L(X, X ), there holds
SA1 +A2 f + A2 A1
1 PA1 (f, s0 ), s0
PA1 +A2 f +
A2 A1
1 PA1 (f,
s0 ), s0
= SA1 (f, s0 ),
= (I
+ A2 A1
1 )PA1 (f,
(1.46)
s0 )
(1.47)
s = SA1 (f, s0 ),
f = p + s.
Here P and S without subscript always denote the play and the stop operator w. r. t. the
original scalar product h, i (i. e. for A = I).
Proof: (1) We write (1.45) as
p(t),
s(t) A
p(t) + s(t)
s(t)
s(0)
0
p(t) + s(t)
C
s0
(1.48)
1 the stop and the play operator w.r.t. different scalar products
or equivalently by virtue of
1
A p(t),
s(t)
A1 p(t) + s(t)
s(t)
s(0)
27
(B.11) as
0
A1 p(t) + s(t)
C
s0
(1.49)
In (1.48) and (1.49) the second identity is trivial. Now (1.49) with (1.4) yields
A1 p = P(A1 p + s, s0 ),
s = S(A1 p + s, s0 ),
which gives (1.42) and (1.44). Rule (1.43) is an easy consequence of (1.44), or can alternatively be derived by (1.42) and (1.1).
(2) We set A1 + A2 is again symmetric and strongly positive
p+ = PA1 +A2 (f+ , s0 ),
f+ = f + A2 A1 1 p
and compute
1
1
f+ = f + A2 A1
1 PA1 (f, s0 ) = (I + A2 A1 )f A2 A1 s
1
= (A1 + A2 )A1
1 f A2 A1 s.
1
p,
s A1
=
A1 p,
s 0 for all C,
p+ , s+ A1 +A2 = (A1 + A2 )1 p + , s+ 0 for all C.
s+ s 0.
(A1 + A2 )1 p+ A1
1 p,
(1.50)
We compute
1
1
A1
= (A1 + A2 )1 (I + A2 A1 1 )
1 = (A1 + A2 ) (A1 + A2 )A1
= (A1 + A2 )1 + (A1 + A2 )1 A2 A1
1 ,
(A1 + A2 )1 p + A1
(A1 + A2 )A1
1 p = (A1 + A2 )
1 f A2 A1 s s + A1 f s
1
1
1
= [A1
1 (A1 + A2 ) A2 A1 ]s (A1 + A2 ) s +
= (A1 + A2 )1 (s s + ).
1 d
ks s+ k2A1 +A2 = s s + , s s+ A1 +A2 = (A1 + A2 )1 (s s + ), s s+ 0.
2 dt
Integration with the initial value (s s+ )(0) = 0 yields s s+ everywhere in [0, T ], giving
(1.46). Rule (1.47) follows again easily with the aid of (1.1).
28
the stop and the play operator w.r.t. different scalar products 1
Chapter 2
R33
R33
h
d
(2.1)
30
1
dev : R33 R33
d , A 7 A 3 tr(A)I, (deviatoric projection)
1
(hydrostatic projection)
hyd : R33 R33
h , A 7 3 tr(A)I,
we define the following deviatoric subspaces of R33
s ,
33
R33
= A R33
: a33 = a11 a22 R5 ,
s
sd = dev Rs
33
33
R33
: a13 = a23 = 0, a33 = a11 a22 R3 ,
spd = dev Rsp = A Rs
(2.2)
(2.3)
(2.4)
2.1
The Newtonian balance equations, the partial differential coupling in the domain , are written
in short-hand operator notation
e
(t) = D eu(t),
D e(t) = F (t)
e. in [0, T ].
(2.5)
(t) = C e(t)
e. in [0, T ].
(2.6)
For each fixed t [0, T ], equations (2.5) and (2.6) build up a (static) linear elliptic problem. It
is in fact a reformulation of (8) and (9). For detailed discussion of model (E), we recommend
Ciarlet, Lions and Duvaut [24, 25, 38].
2.2
We consider the simplified elastoplastic boundary value problem (5), ..., (7) instead of the full
general formulation (2), ..., (4). The reader may find many additional details in Zeidler [113],
section 66, or Han/Reddy [49, 50]. In the latter reference, our formulation corresponds to
the so-called dual variational problem. Here as well, the equivalence to the primal variational
problem is proven, cf. theorem 8.3 therein.
The weak Newtonian balance equations are again
(t) = Du(t),
D (t) = F (t)
e. in [0, T ].
(2.7)
e. in [0, T ],
(2.8)
31
pl (t) C (t)
e. in [0, T ]
(2.9)
a. e. in [0, T ]
(2.10)
(2.11)
2.2.1 Remark. (a) We apply the same given exterior force F to both bodies. We compare the
difference of (EPK) and (E) solutions in section 2.7. The difference of (EPK) and a globally
corrected version of (E) is considered in chapter 3.
(b) Compared to the full problem (2), ..., (4), the simplifications mentioned are the following.
Quasi-staticity: No dynamic or viscous effects are considered. Our body has density = 0. The model is rate-independent, as a reparametrisation = (t) with a
monotonically increasing, absolutely continuous map : [0, T ] [0, T ] shows.
Linearised geometry: We neglect all nonlinear terms in the displacement gradients
u resp. eu. The spatial differential operator D (and consequently its adjoint D ) are
assumed to be linear.
(c) The normality rule (2.10) is equivalent to the maximum dissipation principle, see section
4.2 in Han/Reddy [49].
2.3
In this section we specify the underlying operator/function spaces, the linear operators B, C,
D (and its adjoint D ), the convex set C (and its indicator function C) for the two boundary
value problems.
2.3.1
The spaces
u(x) v(x) + u(x) : v(x) dV(x),
u, v U =
Z
(x) : (x) dV(x),
, =
kuk2U
kk2
u, u U ,
= , ,
=
(U, h, iU ) and (, h, i ) are separable Hilbert spaces, see Zeidler [113], section 55.3. Here
and : are the Euklidian/Frobenius scalar products from (C.1) and (C.2). We have
U
U
=
=
=
=
space
space
space
space
of
of
of
of
displacements:
exterior forces:
strains:
stresses:
u, v, eu, . . .
F
.
, , el , pl , e, eel , epl , . . .
, , , , e, e, e, . . .
(2.12)
32
F (u) = F, u ,
() = ,
for F U , u U, , .
f (t) v dV +
F (t), v =
g(t) v dA
vU
2.3.2
(1 q ).
(2.13)
(2.14)
The operators
D L(U, ),
= adjoint of D
1
u + ut ,
2
D , v = , Dv =
: Dv dV
(2.15)
(for u, v U, ).
kDvk D ()kvkU
v U
R33
with the positive constant D () > 0. If C 1 (,
s ), then
Z
Z
(IP )
n v dA
D , v = DIV v dV +
1
(2.16)
v U.
With the choice (2.13), (2.14), with the divergence theorem and with the aid of the calculus
rule
div(At v) = DIV A v + A : v,
for A C 1 (, Rnn ), v C 1 (, Rn )
relation (2.7) gives the strong form of the Newton balance equations
DIV (t) + f (t) = 0
in
.
(t)n = g(t) on 1
Explicitly written, relation D (t) = F (t) from (2.7) with F (t) as in (2.13) becomes
Z
Z
(t)n g(t) v dA
v U,
f (t) + DIV (t) v dV =
which is the principle of virtual displacements. The same, of course, holds analogously for
(2.5) and e of the elastic model (E).
33
kf (t)kL2 () + C()kg(t)kL2 (1 ) kvkW 1,2 () ,
(HI)
F (t), v
Z
where C() > 0 not depending on time t is the constant of the continuous trace-embedding,
cf. Evans [41], section 5.5, theorem 1,
W 1,2 () L2 (),
The same holds for
F (t), v
kf(t)kL2 () + C()kg(t)k
L2 (1 ) kvkW 1,2 () ,
L2 (1 )
are Lq -integrable due to the assumptions (2.14) on the functions f and g, so we have in fact
F W 1,q ([0, T ], U ).
The linear kinematic hardening operator and Hookes linear operator
B, C L ,
(2.17)
B = B,
kBk kBk kk ,
B, = , B ,
B, B kk2
(2.18)
and
C = C ,
kCk kCk kk ,
C, = , C ,
C, C kk2
(2.19)
B(x)(x) : (x) dV(x)
C(x)(x) : (x) dV(x),
B, =
C, =
33
C(x), B(x) L R33
s , Rs
for a. e. x
(2.20)
34
B (x) B > 0.
Then the strong positivity property in (2.18), (2.19) is globally satisfied. Hookes (pointwise)
operator C(x) is usually of the form (10). The (pointwise) linear kinematic hardening operator
B(x) is chosen in a similar fashion.
Looking forward, the right hand condition in (2.20) for B(x) will be exactly the pointwise
condition (5.1) later. Almost every point x satisfies the pointwise variational inequality
(5.4) with the scalar product induced by B(x).
For later use, we introduce the following linear, continuous operators.
(1) The solution operator for the elastic stress,
S = CD(D CD)1 L(U , ).
(2.21)
P = I Q L( , ).
(2.22)
(2.23)
im Q = ker P = im CD,
= ker D im CD,
(2.24)
2.3.3
The elastic domain C is allowed to be any convex, k k -closed subset of , which contains
the origin.
R, x , that are convex, lower
2.3.3 Lemma. For any collection of functions x : R33
s
semicontinuous and satisfy x (0) = 0 for almost every x and a measurable radius
function : R s. t. (x) 0 a. e. in , we define
C = C, = : x ((x)) (x) for almost every x in .
Then C is closed, convex and contains the origin.
for each 1 , 2 C and [0, 1]. This implies the convexity of C. Third, we show the
k k -closedness. To this end, let
n
C n
in .
(2.25)
35
1
33
As is complete, . By Holder, = L2 (, R33
s )-convergence implies L (, Rs )convergence. Consequently, there exists a subsequence, denoted again by n , such that
n
n (x) (x)
Therefore,
in k kR33
s
(x) lim inf x n (x) x (x)
n
a. e. in .
a. e. in
(2.26)
C() = : C() C() + , for all ,
(2.27)
see chapter 8 in Deimling [30], is equal to the normal cone at C in the point , cf. Brokate
[9], Han/Reddy [49, 50], Urruty [58] or Rockafellar [100]. We distinguish three cases:
(1) We have
C() = {0},
if lies in the interior of C (for which we have purely elastic behaviour, the plastic
strain does not change: pl = 0 according to (2.10)).
(2) We have
C() = : , 0 for all C ,
36
Note that for the Mises set (2.26), we have Int(C) = with respect to k k . Thus the
first case cannot occur. In fact, for this choice of C, every C is a boundary point, since
= Int(C) C. We cannot conclude () = {0} (and the body behaving globally
C=C
elastic) from Int(C), as the latter set is empty. But we have the following lemma for
global elastic behaviour.
2.3.5 Lemma (Mises set, elastic behaviour) Consider the Mises set (2.26), where we assume
L2 (, R). Let and L2 (, R) be given such that
k dev (x)k (x) < (x)
for a. e. x .
(x) (x)
if (x) 6= 0
(x) +
for x .
(2.28)
(x) =
k dev k k (x)k
(x)
if (x) = 0
if (x) 6= 0, and k dev (x)k (x) < (x) otherwise. From the subdifferential definition
(2.27), where () = () = 0, we infer
(2.28)
0 , =
1
k dev k
Z
2.4
dV (x),
We start with the much simpler model (E). In its quasistatic formulation, the problem has not
been considered so far, probably because it is trivial and has therefore not been of interest.
For each fixed t [0, T ] it is a static, elliptic problem without any hysteretic evolution over
time [0, T ]. As a consequence,
it is fully reversible, there is no energy dissipation, and consequently there is no damage
accumulation. (This is as well true for the non-linear elastic case, cf. Bonet/Wood
[6].)
it admits arbitrary input forces of class M AP ([0, T ], U ), yielding outputs of class
M AP ([0, T ], ).
37
We solve the problem in compact operator form. First, we premultiply (2.5) for fixed t [0, T ]
with Hookes operator C, yielding
e
(t)
(2.6)
C e(t)
(2.5)
CD eu(t),
(2.5)
(2.29)
(2.16)
(2.19)
D CDv, v = CDv, Dv C kDvk2 C 2D kvk2U
v U,
(2.30)
thus invertible. So we can explicitly and uniquely solve (2.29) for eu. We obtain for each
t [0, T ]
(1) the elastic displacement
e
(2.31)
(2.32)
(2.33)
Have a look at the following diagram for an overview. Do not mistake the (linear) operators
S, P for the (nonlinear) operators stop S and play P.
-U
S
D CD
(D CD)1
?
C 1
C
- U
6
D
(2.34)
-
6
P, Q
The static solution operators for the linear elastic model are
(D CD)1 L(U , U ),
S L(U , ).
(2.35)
We have the following wellposedness result for the quasistatic elastic model.
2.4.1 Theorem (Wellposedness model E in W 1,q ) For each given exterior force
(a) F M AP [0, T ], U ,
(b) F W 1,q [0, T ], U , 1 q ,
there exists a uniquely determined triple
eu M AP [0, T ], U ,
e M AP [0, T ], ,
(a)
eu W 1,q [0, T ], U ,
e W 1,q [0, T ], ,
(b)
(2.36)
M AP [0, T ], ,
e W 1,q [0, T ], ,
given by (2.31), ..., (2.33), satisfying the elastic relations (2.5), (2.6).
38
Proof: (a) In fact (2.31), ..., (2.33) hold for each given F : [0, T ] U , as for each fixed t
we have an elliptic, uniquely solvable, elastic problem. (There are no time derivatives in the
model relations.)
(b) If now, in addition, F W 1,q [0, T ], U , we identify the operators in (2.35) with their
hatted counterparts from section B.2, we have
keukW 1,q ([0,T ],U )
k(D CD)1 k kF kW 1,q ([0,T ],U ) < ,
kekW 1,q ([0,T ],) kD(D CD)1 k kF kW 1,q ([0,T ],U ) < ,
kekW 1,q ([0,T ], ) kCD(D CD)1 k kF kW 1,q ([0,T ],U ) < ,
each of the right hand sides being finite by assumption.
The basis for high-cycle fatigue analysis (with or without subsequent correction) is the following space-time-decoupling lemma. This corresponds to the part on top-left in figure A on
page 3.
2.4.2 Lemma (Space-time-decoupling) Assume, the exterior force F M AP ([0, T ], U ) is
given by a linear superposition of finitely many static load cases Fi U , i. e.
X
F (t) =
Li (t) Fi
e. in [0, T ]
(2.37)
i
with arbitrary scalar load functions Li M AP ([0, T ], R). Further for each i, let eui U ,
e
i , i denote the solution triple of the static, elastic problems, that is corresponding
to Fi . Then the quasistatic solution triple
M AP ([0, T ], ),
M AP ([0, T ], ),
corresponding to F is given by
X
e
(t) =
Li (t) e i ,
(t) =
everywhere in [0, T ].
u M AP ([0, T ], U )
Li (t) ei ,
Li (t) eui
(2.38)
i = CD(D CD)1 Fi
(2.39)
u(t) =
ui = (D CD)1 Fi ,
i = D(D CD)1 Fi ,
(2.31)
(2.32)
(2.33)
u(t)
(t)
(t)
X
Li (t)(D CD)1 Fi =
Li (t)eui ,
i
i
X
(2.37) X
D(D CD)1 F (t) =
Li (t)D(D CD)1 Fi =
Li (t)ei ,
i
i
X
(2.37) X
CD(D CD)1 F (t) =
Li (t)CD(D CD)1 Fi =
Li (t)e i ,
(2.37)
(D CD)1 F (t) =
2.4.3 Corollary. If there is a 1 q , such that for each i the function Li from lemma 2.4.2 is
in W 1,q ([0, T ], R), then e W 1,q ([0, T ], ), e W 1,q ([0, T ], ) and eu W 1,q ([0, T ], U ).
Proof: Clear.
2.5
39
kRk kRk kk ,
R, = , R ,
R, R kk2 .
(2.40)
Consequently the inverse R1 exists and is strongly positive and symmetric. The same holds
for B, cf. (2.18), and its existing inverse B 1 . We define, cf. (B.11), equivalent scalar products
on by
, B = B 1 , ,
, R = R1 , .
(2.41)
We denote the norm equivalence constants by
cB k k k kB CB k k,
cR k k k kR CR k k.
(0) = 0 C,
(S F )(t)
R C (t) ,
(2.42)
(2.43)
where
R C() = RC()
= R : C() C() + , for all C
= : C() C() + , R for all C
+ (B + C)pl = CDu.
(2.44)
Second, left-multiplication with S (at the left hand identity) and P (at the right hand identity)
with (2.22), (2.21) yields
Q( + Bpl ) = SF,
P + P (B + C)pl = 0,
since ker P = im(CD), see (2.24). Third, addition gives with (2.23),
+ Rpl = + (P C + B)pl = (Q + P ) + (Q + P )Bpl + P Cpl = SF,
yielding
pl = R1 (SF ).
(2.45)
Fourth, left-multiplication of the right hand identity in (2.44) with D and (D CD)1 gives
u = (D CD)1 D + (B + C)pl .
(2.46)
40
Finally, (2.45) together with (2.10), (2.11) gives (2.43), as desired. Thus every solution of
(2.7) ..., (2.11) necessarily yields a solution of (2.43). And vice versa, each solution of (2.43)
allows the construction of pl resp. u via (2.45) resp. (2.46), from which it is then possible to
construct all the remaining unknowns in terms of
(t)
and
(t) = SF (t),
(2.47)
namely
(2.48)
(2.49)
(2.50)
el (t) = C 1 BR1 (t) + (t) ,
(t) = C 1 B + I R1 (t) + C 1 (t),
u(t) = (D CD)1 D (B + C)R1 (t) + (t) .
(2.51)
(2.52)
(2.53)
The essential observation now is that we are able to express resp. as a play resp. a stop
with the input SF , which is nothing but the elastic stress e.
(R) W. r. t. the input e = SF and
(2.43) and (2.47) equivalently as
(t),
(t)
(t) + (t) =
(t)
(0) =
(2.54)
= SR (SF, 0 ).
(2.55)
, , , W 1,q [0, T ], ,
given by (2.43), (2.48), ..., (2.53), satisfying the elastoplastic relations (2.7), ..., (2.11).
41
Proof: From the preliminaries, we have: System (2.7), ..., (2.11) is uniquely solvable, if and
only if (2.54) is uniquely solvable. By theorem 1.1.1, the solution (, ) of the latter exists and
is unique, given by (2.55), since SF = e W 1,q ([0, T ], ), according to theorem 2.4.1. The
remaining unknowns (exist) and are uniquely given by (2.47), (2.48), ..., (2.53). As according
to theorem 1.1.1, the operators P, S map W 1,q into W 1,q , we have and in W 1,q ([0, T ], ).
There holds
kpl kW 1,q ()
kkW 1,q ( )
kkW 1,q ( )
kel kW 1,q ()
kkW 1,q ()
kukW 1,q (U )
(2.48)
(2.49)
(2.50)
(2.51)
(2.52)
(2.53)
kC 1 k kBR1 k kkW 1,q ( ) + kkW 1,q ( ) ,
k(D CD)1 D k kB + C)R1 k kkW 1,q ( ) + kkW 1,q ( ) ,
according to section B.2 in the appendix, the right hand sides each being finite.
2.5.2 Remark. (a) In the literature, sometimes the terminology plastic stress for and elastic
stress for is used, e. g. [9, 14, 70, 75, 82].
(b) Note that B, C, D and R are purely spatial operators, and the transformation above is of
purely spatial kind.
In the same way as before, the pair resp. can be expressed as a play resp. a stop operator
with respect to the elastoplastic stress .
input and the scalar product h, iB from (2.41): Rewriting
equivalently as
for all C a. e. in [0, T ]
(t),
(t) B 0
(t) + (t) = (t)
e. in [0, T ]
,
(t) C
e. in [0, T ]
(0) = 0
(B) W. r. t. the
(2.8), ..., (2.11)
= PB (, 0 ),
= SB (, 0 ).
(2.56)
This is the infinitely dimensional, globally acting counterpart of (5.5) later and the exposi [76], section 2.3. As a consequence, the construction of global stress-strain
tion in Krejc
operators, cf. theorem 2.5.6, is possible.
2.5.3 Remark (Virginity) We have from (2.48) the following equivalence
pl (0) = 0
(2.57)
42
2.5.4 Remark (Monotonicity) Due to Rockafellars theorem, see Zeidler [112], theorem
47.F, Deimling [30], theorem 23.5, or Urruty [58], Rockafellar [100], R C is maximal
monotone. Monotonicity for such a multivalued map means
1 2 , 1 2 R 0
for all 1 R C(1 ), 2 R C(2 ).
(2.58)
Thus for given exterior forces F1 , F2 U , the corresponding solutions 1 , 2 the choice
1 = 1 = S F1 1 , 2 = 2 = S F 2 2 and = |12 , we conclude
thus
,
R
1
k(t)k2R
2
+ SF ,
,
,
=
S
F
R
R
R
(2.58)
(F T )(CS)
1
k(0)k2R +
2
(2.59)
With the same trick one can prove the standard Lipschitz estimate (1.15) for S. In fact,
both derivations are equivalent, but formulated in different languages. And it is in fact this
observation of the author to make him reveal = PR (SF, 0 ) and = SR (SF, 0 ), cf. (2.55).
The Lipschitz result (1.15) is in fact nothing else but (2.59).
2.5.5 Corollary. For the elastoplastic stress , corresponding to a given F W 1,q ([0, T ], U ) and
0 C, there holds
(2.60)
SB+C + CB 1 PB (, 0 ), 0 = SB (, 0 ),
1
1
(2.61)
PB+C + CB PB (, 0 ), 0 = (I + CB )PB (, 0 ).
Proof: Apply lemma 1.4.1 (1.46), (1.47) with A1 = B, A2 = C, s0 = 0 and (2.56).
(, 0 ) 7 C 1 + B 1 PB (, 0 )
and
G : W 1,q [0, T ], C W 1,q [0, T ], ,
(, 0 ) 7 C(B + C)1 B + SB+C (C, 0 ) .
We then have the following theorem, which is the infinitely dimensional global counterpart of
[76].
the stress/strain representation in section 2.3 in Krejc
2.5.6 Theorem (Stress-strain operators) Let F W 1,q ([0, T ], U ), 1 q , and 0 C be
given with the corresponding elastoplastic solutions
W 1,q ([0, T ], ),
W 1,q ([0, T ], ).
= F(, 0 ),
F(, 0 ) = G 1 (, 0 ),
i. e. F and G are turning stress into strain and vice versa. They are continuous for each
1 q < .
43
Proof: Clearly,
(2.9)
(2.8)
(2.56)
= el + pl = C 1 + B 1 = C 1 + B 1 PB (, 0 ) = F(, 0 ).
We have
+ CB 1 PB (, 0 )
(2.56),(2.8)
(2.8)
(2.9)
(2.9)
SB+C (C, 0 ) = SB (, 0 ) = =
thus
(2.8),(2.9)
B( C 1 ) = (B + C)C 1 B,
= C(B + C)1 B + SB+C (C, 0 ) = G(, 0 ).
2.6
We are now able to give a new proof of the stability results and error estimates of problem
(EPK) compared to the one in Han/Reddy [49], section 7.3. New is to be understood in the
following sense.
(1) We use the language of stops and plays, which is very convenient and concise. We
consider W 1,q for arbitrary 1 q , especially W 1,1 = AC.
(2) In contrast to Han/Reddy [49], we consider directly, what they call the dual problem
(in the context of convex analysis). This equivalent formulation (cf. Han/Reddy
[49], section 4.2) is by far the most widespread in both mathematical and engineering
literature.
(3) We have the case of mixed Dirichlet and Neumann boundary conditions. In Han/Reddy [49] only homogeneous Dirichlet conditions u = 0 on whole are studied, and
there is no indication, how to handle the mixed problem, which is essential for practice.
The robustness of model (EPK) is a consequence of the generalised Lipschitz estimates of
chapter 1. In the same way, as well linear kinematic hardening plus isotropic hardening, cf.
section 4.1, and the spaces CBV , cf. section 4.2, can be treated.
For the proof of the following theorem, stating the input stability of model (EPK), we just
need the standard estimates (1.15) and (1.16). They will turn out to be special cases of
corollary 2.6.4. Nevertheless, we mention it separately.
(0)
(1)
44
(1)
(2)
(0)
(1)
(2)
k(t)k
k(t)k
k(t)k
(2)
C k0 k
+ C F (t),
Here
= |12 = 1 2 ,
F (t) =
t
0
kSF ( )k d.
All estimates are valid everywhere in [0, T ]. They all remain valid, if is omitted.
Proof: We use (2.47). The explicit formulas (2.48), ..., (2.53) and the norm equivalence (2.42)
give us with the definition of the linear operator norm
k(t)k
k(t)k
and
kpl (t)k
k(t)k
k(t)k
kel (t)k
k(t)k
ku(t)kU
()
()
c (R)k(t)kR ,
()
c (R)k(t)kR ,
(2.50)
(2.51)
(2.48)
cpl (R)k(t)kR ,
c()
(R, B)k(t)kR ,
(2.49)
c (R)k(t)kR ,
()
()
c()
(R, B, C)k(t)kR
(2.63)
()
c()
(R, B)k(t)kR +
(2.53)
(2.62)
()
(2.52)
where c = 1/cR ,
where c()
= 1/cR ,
+
()
(2.64)
()
c (R, C)k(t)kR ,
c()
u (R, B, C, D, D )k(t)kR + cu (R, C, D, D )k(t)kR
for each t [0, T ]. Now apply estimate (1.15) for k(t)kR and (1.16) for k(t)kR , the triangle
inequality and the norm equivalence (2.42). The relations without are obtained by taking
F1 F , F2 0, 1 (0) = (0), 2 (0) = 0.
As a consequence of the Pythagoras relation (1.12) for the time derivatives of stop and play,
we have insight into the following fact.
45
2.6.2 Theorem (Robustness model EPK, II) There exist non-negative constants, named C ,
such that for each exterior force F W 1,q ([0, T ], U ) and its corresponding solution octuple
there holds
kpl k Cpl kS F k,
kel k Cel kS F k,
kk
C kS F k
(2.65)
C kS F k,
kk
kk
C kS F k,
kk
C kS F k (2.66)
(2.67)
(for the displacement). Here all the right hand side norms are k k = k k . All estimates
are valid a. e. in [0, T ].
Proof: We have
R kS F kR CR kS F k ,
kk
kk
R kS F kR CR kS F k
a. e. in [0, T ]
from (2.55), (2.33) and (1.12), where the constant CR is from (2.42). Now apply the explicit
expressions (2.48), ..., (2.53) and the triangle inequality in order to obtain dotted equivalents
of the relations (2.62), ..., (2.64).
It is now interesting to realise, that the generalised Lipschitz estimates of chapter 1 yield local
parameter stability for the model (EPK).
2.6.3 Theorem (Robustness model EPK, III) We consider two sets of symmetric and strongly positive data Bi , Ci , inducing four pairs of operators
(2.21)
Si = Ci D(D Ci D)1 ,
(2.22)
Qi = S i D ,
(2.22)
Pi = I Qi ,
(2.23)
Ri = Pi Ci + Bi , (2.68)
two inputs Fi W 1,q ([0, T ], U ) and 0,i C. Always i {1, 2}. With the abbreviations
B = min B1 , B2 ,
and
b = max kB1 k, kB2 k ,
C = min C1 , C2 ,
c = max kC1 k, kC2 k ,
d = max kDk, kD k
k0 = (2 C )1 ,
D from (2.30),
r = b + c 1 + d2 ck0 ,
k3 = dck0 ,
k2 = d2 ck0 d2 ck0 + 2 + 1,
k1 = dk0 d2 ck0 + 1 ,
kF kd
k0 k + k3
k(t)k
B
0
Z t
rk
r 3 k3
1
1 k + kF2 k d
kBk
+
k
kCk
+
+
k
F
kCk
2
2B
4B
0
(2.69)
46
and
k1
k(t)k k3 kF (t)k + kCk kF1 (t)k + kF2 (t)k
2 Z t
r
+
kF kd
k0 k + k3
B
0
Z t
rk
r 3 k3
1
kBk + k2 kCk +
kF1 k + kF2 k d
+
kCk
4
2B
B
0
(2.70)
everywhere in [0, T ].
R = (P C) + B.
(2.71)
P = Q
(2.72)
Further
(P C) = C (QC),
Q = S D ,
(S F ) = S F1 + S2 F = S F2 + S1 F = S
F +
S F .
(2.73)
(2.74)
We decompose
S = C1 (C 1 S) + C C21 S2 = C2 (C 1 S) C C11 S1 .
From (2.68) we infer Ci1 Si = D(D Ci D)1 and Ci1 Qi = Ci1 Si D . Thus
(C 1 S) = D (D C1 D)1 (D C2 D)1
= D(D C1 D)1 D C2 D D C1 D (D C2 D)1
= D(D C1 D)1 D C D(D C2 D)1
(2.75)
(2.76)
(2.68)
k(D Ci D)1 k k0 ,
and
(2.19)
kCi1 k
1
,
C
kSi k dck0 = k3 ,
(2.18)
kBi1 k
1
,
B
(2.71)
kRi k r
(2.77)
(2.68)
kQi k d2 ck0 ,
(2.78)
where we have applied (B.13) for k(D Ci D)1 k, kCi1 k and kBi1 k. There holds by the submultiplicativity of the operator norm
kSk
kQk
k(QC)k
(2.75)
(2.72)
(2.73)
(2.76)
kCk kQk +
kQk kCk d2 ck0 d2 ck0 + 2 kCk,
kR1 k
1
kBk
+
k
kCk
.
2
2B
(2.79)
47
(2.80)
(2.81)
Step II. Application of the estimates (1.31), (1.32) with appropriate symmetrisation yields
Z t
1 CR1
CR2
k(t)k
k(S F )kd
+
k0 k +
2 cR1 cR2
0
Z t
(2.82)
CR1
1
CR2
1
kS1 F1 k + kS2 F2 k d.
+
+
kR k
2cR1 cR2 cR2
cR1
0
Ri , Pi Ci , + Bi , Bi kk2 ,
(2.83)
Ri Bi
(2.84)
in (2.40). Having a look at section B.2 in the appendix, we see that the choices
CRi
(B.14)
1
Ri
(2.84)
1
,
Bi
(B.14)
cRi
kRi k
Ri
(2.84)
r
Bi
(2.85)
,
+
2.
2 cR1
cR2
B
2cR1 cR2 cR2
cR1
B
(2.86)
Inserting this and (2.79), (2.80) into (2.82) finally gives the assertion (2.69) after rearranging.
Step III. The assertion for follows from (1.33), (1.34), giving
Z t
1 CR1
CR2
2.6.4 Corollary (Robustness model EPK, IV) For each pair of symmetric, strongly positive
(6)
(1)
data Bi , Ci , i {1, 2}, there exist non-negative constants C , ..., C , depending on the
norms kBi k, kCi k, the coercivity constants Bi , Ci , Korns constant D and kDk, kD k, such
that for each pair Fi and 0,i , i {1, 2}, the corresponding elastoplastic solution differences
k(t)k,
k(t)k,
k(t)k,
k(t)k,
kel (t)k,
C k0 k ,
and
(4)
C
Here =
|12
(2)
(3)
C kBk + C kCk
kF kU d,
= 1 2 as usual.
kpl (t)k,
k(t)k,
ku(t)k
kF1 kU + kF2 kU d
(6)
(5)
C kBk + C kCk kF1 (t)kU + kF2 (t)kU .
48
2.7
In this section, we compare the solutions of (E) to the solutions of (EPK). The next theorem
2.7.1 is the main theorem of this chapter, as it gives the structural result, that the difference
of both models are play operators.
But let us first introduce the linear operators
H,R =
P CR1 ,
1
H,R =
(C P C I)R1 ,
Hu,R = (D CD)1 D CR1 ,
H,B =
P CB 1 ,
1
H,B =
(C P C I)B 1 ,
Hu,B = (D CD)1 D CB 1 .
(2.87)
There holds
H, L( , ),
H, L( , ),
Hu, L( , U ).
We let further
kH,R k ,R = sup
6=0
kH,R ()k
,
kkR
kH,R ()k
,
kkB
...
(2.88)
(2.89)
(2.90)
Proof: Note
(2.23)
(2.48)
(2.49)
pl = R1 = B 1 .
R B = P C,
(2.91)
Thus first,
e
(2.33)
(2.47)
(2.91)
(2.91)
49
second,
e
(2.6),(2.8),(2.9)
C 1e (C 1 + pl ) = C 1 P Cpl pl = (C 1 P C I)pl ,
third,
e
uu
(2.31),(2.53)
(2.8),(2.9)
From (2.55), (2.56) and (2.91), the assertion follows with definition (2.87).
2.7.2 Remark. (a) Theorem 2.7.1 states that it is possible to obtain the elastic stress e from the
elastoplastic one , and vice versa, i. e.
e
= e H,R PR (e, 0 ).
= H,B PB (, 0 ),
For the stresses, the situation is even better, cf. theorem 2.8.1 further below.
(b) For u and eu note that there holds
u = (D CD)1 D ( + Cpl ) = (D CD)1 D (Cel + Cpl ) = S ,
Cf. proof of theorem 2.7.1 and diagram (2.34). From this we have e el ker S .
We are now able to apply the estimates (1.14), which were resulting from Pythagoras relation
for the time derivatives of stop and play, in order to obtain estimates for the difference of the
model solutions. This is done in the next theorem.
2.7.3 Theorem (Difference of models E and EPK, II) Let the assumptions of theorem 2.7.1
hold. The following estimates hold everywhere in the interval [0, T ].
(R) In terms of SF = e and = SF , there holds
for the stresses:
k(e )(t)k
for the strains:
k(e )(t)k
for the displacements: k(eu u)(t)kU
where
(t) = k(0)k +
CR kH,R k ,R (t),
CR kH,R k,R (t),
CR kHu,R kU,R (t),
kS F ( )k d.
(2.92)
Note that e = SF is up to a linear operator equal to the input, cf. (2.33), and that
estimates w. r. t. F are easily obtained.
(B) In terms of and = , there holds
for the stresses:
k(e )(t)k
for the strains:
k(e )(t)k
for the displacements: k(eu u)(t)kU
where
(t) = k(0)k
CB kH,B k ,B (t),
CB kH,B k,B (t),
CB kHu,B kU,B (t),
k(
)k d.
(2.93)
50
2.7.4 Theorem (Difference of models E and EPK, III) Let the premises of theorem 2.7.1
hold. The following estimates hold almost everywhere in the interval [0, T ].
(R) In terms of e = S F , there holds
for the stresses:
k(e )(t)k
e
for the strains:
k( )(t)k
e
for the displacements: k( u u)(t)k
CR kH,R k ,R kS F (t)k ,
CR kH,R k,R kS F (t)k ,
CR kHu,R kU,R kS F (t)k .
(B) In terms of ,
there holds
for the stresses:
k(e )(t)k
e
for the strains:
k( )(t)k
CB kH,B k ,B k(t)k
,
CB kH,B k,B k(t)k
,
CB kHu,B kU,B k(t)k
2.8
In this section, we especially have a look at the stress difference. The following theorem is
very remarkable for practice, if no stress correction is applied. This corresponds to the very
left branch in figure A on page 3.
2.8.1 Theorem (Stress difference) Let 1 q . For e and from any elastoplastic model,
for which D = F with F W 1,q ([0, T ], U ) holds, we have
e
= Q,
= Q,
e = P ,
e = P ,
(2.94)
and
ke(t)k CQ k(t)k ,
ke(t)k
CQ k(t)k
,
k( e)(t)k CP k(t)k
k( e)(t)k
CP k(t)k
e. in [0, T ],
a. e. in [0, T ].
The constants CQ , CP depend only on D, D and C, but not on the special plastic constitutive
operators.
51
(2.33)
(2.7)
(2.22)
= SF = SD = Q.
2.8.2 Remark (Importance) (a) In theorem 2.8.1, the special plastic constitutive relations do
not play any role. It holds for any elastoplastic constitutive model with linearised geometry
and quasi-staticity. Theorem 2.8.1 now gives the guarantee that e will not be too bad, which
is of importance in high cycle fatigue analysis.
(b) The statement of theorem 2.8.1 does not carry over for e and or eu and u. Instead, we
have just
e
e
= D(D CD)1 D ,
u = (D CD)1 D .
Especially for constitutive models (as Jiangs [64] or Chaboches [20, 21, 79]), which are
comprising the ratchetting effect, may become arbitrarily large, even if remains bounded,
[12].
see as well Brokate and Krejc
(c) As the operator Q is singular, it is impossible to solve (2.94) for . See (2.24) for the
kernel of Q.
52
Chapter 3
3.1
Global correction
Apart from the linear kinematic hardening tensor B, which is the elastoplastic material parameter, and the elastoplastic initial memory 0 , let another symmetric, strongly positive
operator eB L(, ), which is the elastic parameter,
e
e
B = eB ,
keBk keBk kk ,
B, = , eB ,
B, eB kk2 ,
and an elastic initial memory e 0 C be given. For eB, we have the norm equivalence, cf.
section B.2,
54
eB
e(0) = e
0
i. e. with theorem 1.1.1,
e
= PeB (SF, e 0 ),
= SeB (SF, e 0 ).
(3.2)
This is the infinitely dimensional global counterpart of the pointwise elastic decomposition
(5.3) later. The tilde denotes the corrected quantities, which are introduced in the sequel.
Based on decomposition (3.2), we define the corrected stresses by
= e,
= B eB 1e,
=
+ ,
(3.3)
el = C 1
,
= el + pl
(3.4)
(3.5)
This is again the infinitely dimensional, globally acting counterpart of the pointwise correction model in chapter 5.
We now arrive at the following difference (or error) expressions compared to the elastoplastic
model,
=
= B eB1 PeB (SF, e 0 ) BR1 PR (SF, 0 ),
= =
SeB (SF, e 0 )
SR (SF, 0 ).
(3.6)
(3.7)
Further we arrive at
= + ,
u = (D CD)1 D C
(3.8)
and
pl = B 1 ,
el = C 1 ,
= el + pl .
(3.9)
Here = is the difference of the corrected elastic and the elastoplastic solution. Setting
= B, we find that
eB
=: e,
=: e,
= e,
pl =: epl ,
Setting eB = R, we find
= ,
= ,
el = e =: eel ,
= ,
= ,
pl = pl ,
= e + epl ,
u
= S . (3.10)
= ,
el = el ,
= ,
u
= u,
(3.11)
this means the exact elastoplastic solution is recovered. And it is clear that the best choice
for the elastic parameter will be eB = R.
55
3.1.1 Remark. Alternative equivalent expressions for the differences (3.6) and (3.7) are
=
=
=
BP(eB1e + e, e 0 ) BP(R1 + , 0 ),
S(eB 1e + e, e 0 )
S(R1 + , 0 ),
where the transformation formula of lemma 1.4.1, which is especially valid for A B, R, eB ,
have been applied.
3.1.2 Example. For eB we may e. g. choose, cf. (2.20),
e
B, =
e
B(x)(x) : (x) dV(x)
(, )
with the assumptions on eB(x) that are analogous to example 2.3.2. Looking forward, almost
every point x will satisfy the pointwise variational inequality (5.5) with the scalar product
induced by eB(x), or equivalently the differential equations (5.18), (5.19).
We are now going to apply the generalised Lipschitz estimates of chapter 1, in order to receive
estimates for the errors . We may apply them this is the essential thing with the same
input, namely SF , cf. corollary 1.3.2. We set for abbreviation
F (t) =
t
0
kS F ( )k d.
(3.12)
3.1.3 Theorem (Error correction) For each exterior force F W 1,q ([0, T ], U ) and initial memories 0 , e 0 C we have the following estimates.
(1) Estimates for the stresses. For , there holds
e
k(t)k
k(t)k
(3.13)
k(t)k
k(t)k
qR k0 k + qeB c2
R h F (t),
qeB k0 k + qR c2
eB h F (t).
(3.14)
= e,
e
r1b + 1 k(t)k + qeB hB F (t),
ebb + 1 k(t)k + q h (t).
R B F
1
k(e )(t)k = k(t)k ,
(3.15)
(3.16)
56
(3.17)
(3.18)
e
r1 + c1 (r1b + 1) k(t)k + qeB h + c1 hB F (t),
eb + c (ebb + 1) k(t)k + q
1
1
R h + c1 hB F (t).
1
(3.19)
(3.20)
Here = denotes the difference between the correction of (E) and the elastoplastic
solution (EPK). We have used the abbreviations
e
c1 =
C 1
, ,
b1 =
eB 1
, ,
r1 =
R1
, , (3.21)
CR
CeB
e b
,
qeB =
,
r1b =
BR1
, ,
b1 =
B eB 1
,
cR
ceB
and the small operator differences
h =
eB1 R1
,
hB =
B(eB 1 R1 )
.
qR =
(3.22)
Proof: Step I. For the estimates of , we just have to apply corollaries 1.3.2 (a) and 1.3.3
with
(1) the inputs
(2) the stops
f1 = f2 = SF = e ,
s1 = e,
s2 = ,
p1 = e,
p2 = ,
s0,2 = 0 ,
(3.23)
, A1 = , eB ,
57
, A2 = , R ,
and the estimates (3.1) and (2.42) for (1.19). We find with a combination of (1.28) and (1.31)
easily (3.14).
Step II. Assertion (3.16) follows directly from (1.27).
Step III. There holds
(3.24)
= Bpl = B eB1e R1 .
k(t)k
kB B
e 1
(t)k + kB( B
)(t)k .
(3.25)
(3.26)
The second summand in (3.25) resp. (3.26) is estimated with (1.14) and (3.1) resp. (2.42) in
the usual way.
Step IV. For pl = eB 1e R1 , the proof is like for , but not that the operator B is
lacking, cf. (3.24).
Step V. Estimates for the remaining corrected quantities can straightforwardly by derived
via (3.8), (3.9) and the estimates for , and pl .
Alternatively from (3.24) we have
= B eB 1 (SF e) R1 (SF ) ,
The latter could serve as well as a basis for error estimates with the aid of (1.13).
3.1.4 Corollary. If the initial values are chosen in a consistent manner, i. e. 0 = e 0 , and the
material is virgin, cf. remark 2.5.3, there exist non-negative constants c dependencies in
brackets such that the following estimates hold.
(1) Estimates for the stresses. There holds
k(t)k
k(t)k
k(t)k
k(e )(t)k
c (eB, R, B)h + k (eB, R)hB F (t),
c (eB, R, B)h + k (eB, R)hB F (t),
cel (eB, R, B, C)h + kel (eB, R, C)hB F (t),
c (eB, R, B, C)h + k (eB, R, C)hB F (t).
58
cu (eB, R, B, C, D)h + ku (eB, R, C, D)hB F (t).
Here again = and F (t) from (3.12). All estimates are valid everywhere in [0, T ].
Proof: We have
= (0) = 0
0 = e 0
and
(2.57)
(3.16)
= (0) = e(0) = 0 .
pl (0) = 0
k(t)k
1
2
h F (t) =: c (eB, R)h F (t).
qR c2
eB + qeB cR
2
Inserting this into the other error expressions of theorem 3.1.3, rearranging and taking the
average of each pair finally gives the desired results.
3.1.5 Remark. (a) Note that we have
e 1
R1 = eB 1 (I eBR1 ) = eB 1 (R eB)R1 ,
(3.27)
(3.28)
and
(b) Note, that have at the look at the proof of theorem 1.3.1 there hold the inequalities
or equivalently,
B 1 ,
0
e 1e
(eB 1 R1 )S F , .
B R1 ,
(c) In order to get better error estimates, one could consider ansatzes, which are decomposing
the correction error in the form
e
k correB,e 0 (e)k k correB,
e () correB,e 0 ( )k
e ()k + k correB,
0
(with possibly different eB and e0 ) in the hope that the first summand is very close to zero.
This approach seems not to be successful, since in the second summand we have to consider
the difference between
and
59
which inherits again the problems, which will occur in chapter 5. To illustrate this, we have
a look at the stop difference at the right hand side above. We have
Z t
Z t
CeB
CeB
e 1
ke kd
+
ke 0 k +
k
kd
k
B
k
kSeB (e, e 0 ) SeB(, e0 )k
ceB
ceBc2eB
0
0
according to (1.31). The choice eB = eB, e0 = e 0 makes the first and last summand vanish.
In contrast to chapter 5, we are now able to apply theorem 2.7.4, i. e. ke k
kS F k, but
this does not give any smallness indication at all.
We have a counterpart to theorems 2.6.2 and 2.7.4. In fact as the choice eB = R shows
it generalises theorem 2.7.4. The fact that SF = e is common input into (2.55), cf. (3.2), is
essential.
3.1.6 Theorem (Derivatives correction) There exist non-negative constants C and k , such
that for each exterior force F W 1,q ([0, T ], U ) there holds
Ce kS F k
kek
(3.29)
kpl k kpl kS F k ,
kel k kel kS F k ,
kk
k kS F k
(3.30)
C kS F k ,
C kS F k ,
C kS F k ,
kk
k kS F k ,
k kS F k ,
kk
kk
k kS F k
(3.31)
Cu kS F k ,
kuk
ku kS F k
(3.32)
kek
Ce kS F k ,
(for the elastic quantities),
kdt pl k Cpl kS F k ,
kdt el k Cel kS F k ,
kdt k
C kS F k ,
(for the corrected strains),
kdt
k
kdt k
kdt
k
(for the corrected stresses),
kdt u
kU
(for the corrected displacement). Here = . All estimates are valid a. e. in [0, T ].
Proof: We have
eB kS F keB CeB kS F k ,
kek
kek
eB kS F keB CeB kS F k ,
a. e. in [0, T ]
from (3.2), (2.33) and (1.12), where CeB from (3.1) in order to receive (3.29). Now apply
relations (3.3), ..., (3.5) and the triangle inequality to obtain all the left hand side estimates
in (3.30), ..., (3.32).
The right hand side estimates follow from the left hand ones with (2.65), ..., (2.67) and the
abused triangle inequality k k k k + k k.
We now have a stability result for small disturbances of the elastic parameter, which is similar
to theorem 2.6.3.
60
3.1.7 Theorem (Robustness correction) Let two symmetric and strongly positive elastic data
eB L(, ), and two inputs F W 1,q ([0, T ], U ) and e
i
i
0,i C be given. Always i {1, 2}.
With the abbreviations
e
eB = min{eB1 , eB2 },
d = kDk,
s=
dc
,
2D C
there hold in terms of the input disturbances kF k, ke 0 k and elastic parameter disturbances keBk
e3
Z t
Z t
eb
b s e
e
e
(3.33)
kF1 k + kF2 k d
kF kd + 4 k Bk
k 0 k + s
k (t)k
eB
eB
0
0
and
ke(t)k
skF (t)k +
eb
eB
k 0 k + s
kF kd
eb3 s
4eB
k Bk
t
0
kF1 k + kF2 k d
(3.34)
everywhere in [0, T ].
Proof: Use (1.31), ..., (1.34) and
e 1
e
e 2
e
keB1 k keB1
1 k k B 2 k k Bk b k Bk,
CeBi =
1
,
eBi
keB i k
ceBi =
.
eBi
3.1.8 Corollary (Robustness correction II) For each given symmetric and strongly positive
(4)
(1)
pair eBi L(, ) there exist non-negative constants C , ..., C , depending on the norms
keB i k, the coercivity constants eBi , Korns constant D and d = max{kDk, kD k}, such that
for each pair Fi and e 0,i the difference k()(t)k can be written as sums of
(1)
C k0 k ,
and
(2)
t
0
kF kU d,
(3)
C keBk
t
0
(4)
C keBk kF1 (t)kU + kF2 (t)kU .
kF 1 kU + kF2 kU d
Here represents one quantity of the corrected elastic septuple, i. e. one of
, ,
, , el , pl
or u
.
Proof: Clear.
3.2
Localisation of correction
In this section, we inspect the widespread case, that the overall body behaviour is elastic,
whereas the plasticity regions, the active regions according to Han/Reddy [49, 50], are small
in volume. We cannot give any sufficient conditions for this behaviour, which is usually called
the effect of elastic support in engineering literature, see e. g. Glinka et al. [18, 44, 83, 104],
ttgen et al. [72, 73]. Thus we cannot help
Hertel [53], Hoffmann/Seeger [61] or Ko
assuming it, cf. (3.38) and (3.40).
61
Numerical examples and examples, where analytical solutions for e = SF are available (see
monograph Neuber [90]), suggest that such regions as in (3.37) exists. But a rigorous proof
of these facts is not known to the author.
f (x,6t)6
ep
Y
H
H
?
?
'$
0
?
u=0
corr
-
&
Y
H
H
1,g
1
g(x, t)
?
(3.35)
Having a closer look at the proof of theorem 1.3.1, we see that, if we apply it to the same
quantities as in the proof of theorem 3.1.3,
Z t
(1.26) Ce
1
B
k(t)k
k0 k + 2
k(, )k d
(3.36)
ceB
ceB 0
holds, with a function
W 1,q ([0, T ], ), [0, T ] t 7 (t, ) ,
which is given by
(2.48)
(3.37)
a. e. in [0, T ] e .
(3.38)
(3.39)
(3.40)
(2) The operator eB is somehow sufficiently close to R locally in the correction domain
corr , i. e.
k=
sup
k(eB 1 R I)kL2 (corr ,R33
(3.41)
) 1.
s
kk 2 corr 33 =1
L (
,Rs
)
(This should be achieved by appropriate parameter identification technique for eB, see
chapter 8.)
62
Relations (3.38), (3.40) provide a definition for the effect of elastic support.
3.2.1 Remark (Pointwise correction model) The good thing is, that the numerical computations in chapters 7 and 8 show, that our correction model works as well in the degenerate
case, where we simply correct at one single isolated boundary point of , which means
Vol(corr ) = 0. A rigorous proof for the transfer to this (degenerate) pointwise case is missing so far.
We decompose the L2 -norm of the right hand error integrand in (3.36) into three parts,
k(t, )k =
Z
1/2
=:
k(t, x)k dV(x)
2
1/2
I02 (t) + I12 (t) + I22 (t)
.
Here we have
I0 (t) =
Z
1/2
(3.38)
= 0
k(t, x)k2 dV(x)
a. e. in [0, T ],
Z
1/2
kkpl kL2 (corr ,R33
k(t, x)k dV(x)
)
s
2
corr
a. e. in [0, T ]
Z
e 1
ep \corr
k( B
1/2
R I) k dV(x)
pl 2
a. e. in [0, T ].
Due to (3.39) and the typical situation that pl is small in ep \ corr , this contribution to the
total error should be as well small, even if keB 1 B 1 k 1 and kB 1 R1 k 6 1. Thus,
under the assumptions of elastic support, the total error will be small.
3.2.2 Remark. Looking one section forward, where we construct a homotopy between models (E)
and (EPK), the choice
e
B = B + P C,
e
B, =
(B + P C) (x) : (x)dV(x)
(x) 1 a. e. in \ corr ,
0 (x) 1 arbitrary a. e. in e ,
3.3
63
It is now a very interesting and very beautiful fact, that we may construct an explicit homotopy between the solutions of the elastic and elastoplastic models in an easy fashion, namely
by a continuous connection of operator B to operator R, i. e. by a continuous connection of
the corresponding scalar products k kB and k kR .
We hope that in the future numerical homotopy (or continuation) methods, cf. Allgower/Georg [1], Deuflhard [31], will provide good preconditioners for transient nonlinear
elastoplastic finite element solvers. We construct the homotopy as follows.
Let the two initial memories
e
0 , 0 C
(3.42)
be given. 0 is the initial memory for the elastoplastic model, e 0 is somewhat arbitrary. For
a homotopy parameter [0, 1], let us take the choice
e
B = R + (1 )B = B + P C
(3.43)
0, = 0 + (1 )e 0
(3.44)
in (3.2). That is
= PeB (SF, e 0, ),
= SeB (SF, e 0, ),
+ e = SF.
(3.45)
= B eB 1
,
= e ,
=
+ ,
(3.46)
the -strains by
e 1 e
pl
= B ,
pl
= el
+
1
,
el
=C
(3.47)
(3.48)
All these definitions are according to (3.3), ..., (3.5), expect the one for pl
, which additionally
has a prefactor . Then we receive
e
B B = P C,
R eB = (1 )P C,
R B = P C.
Further
for = 0, i. e. for eB 0 = B and e 0,0 = e 0 , the elastic solution
0 = e,
0 = e,
u
0 = eu
1 = ,
1 = ,
1 = ,
1 = ,
pl
pl
1 = ,
el
el
1 = ,
u
1 = u
64
is recovered. The choice = 0 gives rise to define the other e -quantities, which are originally
not present in model (E), namely
:= 0 ,
:=
0 ,
e el
:= 0 = el
0,
:= pl
0 = 0,
e pl
so that the elastic solution is now as well a septuple. The essential fact is that we are able
to connect at least one of e, e, eu (namely e) with its elastoplastic counterpart (namely ).
Cf. diagram (3.49) as well.
BB1
= e
= e
6 = 0
B eB1
= e
(3.49)
e
= e
BR1
= e
? = 1
3.3.1 Remark. Due to the convexity of C it is assured that e 0, is in C again. eB is clearly linear,
continuous and symmetric, since the endpoints R and B are so. We estimate below,
(3.43)
(2.18),(2.40)
e
B , = R, + (1 ) B,
(3.50)
R,B = min B , R .
eB = R + (1 )B R,B ,
This means that eB is in fact strongly positive, so that the inverse eB1
exists. We use the
results of section B.2, giving
(B.13)
keB1
k
R + (1 )B
We further have
k keB
where
CeB = p
(B.14)
1
min{B , R }1 .
CeB k k CRB k k,
1
,
R + (1 )B
CRB = p
1
.
min{R , B }
(3.51)
(3.52)
All rightmost estimates are independent of the parameter . Now the other way round. We
estimate above,
(3.43)
keB k kRk + (1 )kBk max kRk, kBk ,
thus
keB 1
k
1
kRk + (1 )kBk
max{kRk, kBk}1
65
and
(B.14)
cRB k k ceB k k
where
ceB
R + (1 )B
,
=
kRk + (1 )kBk
k keB ,
cRB =
(3.53)
min{R , B }
.
max{kRk, kBk}
=
,
=
,
el
el
.
=
(3.54)
For = 1, we have all the tilded corrected quantities identical to the hatted corrected
quantities, cf. (3.11).
We now prove that our construction depends continuously on the homotopy parameter
and the exterior force F . The main ingredients of proof are again the generalised Lipschitz
estimates of chapter 1.
3.3.3 Theorem (Homotopy) (a) The map
H : [0, 1] W 1,q [0, T ], U C [0, T ], ( )3 3 U
(, F ) 7
, ,
, el , pl , , u
(F )
is a homotopy
H(0, F ) = e, e, e, eel , epl , e, eu (F )
H(1, F ) = , , , el , pl , , u (F ).
H(, ) : W 1,q [0, T ], U W 1,q [0, T ], ( )3 3 U
(, F ) 7 (F ).
(2) From (3.52), (3.53) in remark 3.3.1, we have the following norm equivalence, which is
independent of , namely
cRB k k ceB k k k keB CeB k k CRB k k.
(3.55)
66
(3) Let a sequence (n ) [0, 1] be given such that n [0, 1]. Then there holds
trivially
n
in L , .
B n eB
e 1
e 1 e
e 1
e
keB 1
n B k = k B n ( B B n ) B k
(3.51)
e
e 1
e
keB1
n k k B k k B n B k
i. e.
e 1
Bn
Thus
e 1
Bn
and
eB 1
e 1
Bn
Clearly,
eB 1
keB n eB k n
0,
min2 {B , R }
in L , .
eB 1
(3.56)
in L W 1,q ([0, T ], ), W 1,q ([0, T ], )
in L C([0, T ], ), C([0, T ], ) .
n
0,n = n 0 + (1 n )e 0 0 + (1 )e 0 = e 0, .
(3.57)
(3.58)
n in [0, 1],
Fn F in W 1,q [0, T ], U
e
e
(F ) = B eB 1
n (Fn )
B eB1
PeB (SF, 0, ),
n PeBn (SFn , 0,n ) =
n
SeBn (SFn , e 0,,n ) = n (Fn ) (F ) = SeB (SF, e 0, )
in C [0, T ], . To this end we show amongst others
n
(F ) = PeB (SF, e 0, )
(3.59)
(3.60)
(3.61)
in C([0, T ], ). First, clearly due to the continuity of S note the identification in section
B.2 of the appendix ,
n
n = SFn SF = e
in W 1,q ([0, T ], ).
(3.62)
in C([0, T ], ).
(3.63)
n = SFn SF = e
We apply the generalised stop estimate theorem 1.3.1 with (3.55) in (1.19) in order to receive
SeB (SFn , e 0, )(t) SeB (SF, e 0, )(t)
=
kn (Fn )(t) (F )(t)k
n
n
Z t
(1.20) C
RB
e
e
kS(Fn F )kd
k 0,n 0, k +
cRB
0
Z t
1 e 1 e 1
+ 2 k Bn B k
kdt PeB (SF, 0 )kd.
cRB
0
67
B
k
kS(Fn F )kd + keB 1
kn (Fn ) (F )k c ke 0,n e 0, k +
n
0
This gives (3.60). We next show (3.61) in the same way. Applying the generalised play
estimate from theorem 1.3.4, to obtain
PeB (SFn , e 0, )(t) PeB (SF, e 0, )(t)
=
ken (Fn )(t) e (F )(t)k
n
n
(1.33)
kS(F Fn )(t)k
Z t
CRB e
e
kS(Fn F )kd
k 0,n 0, k +
+
cRB
0
Z t
1
e 1
+ 2 keB 1
kdt PeB (SF, 0 )kd.
B
k
n
cRB
0
Thus, taking the supremum at the right hand side, letting t = T , with a constant c, independent of n, because of (3.56), (3.62), (3.63) and (3.58),
e
e
k n (Fn ) (F )k c ke 0,n e 0, k + kS(Fn F )k
Z T
n
e 1
e 1
kS(Fn F )kd + k Bn B k 0.
+
0
e 1 e
e
B eB1
n n (Fn ) B B (F )
in C([0, T ], ).
This is (3.59).
(b) This assertion is clear because of the continuity of the stop and play W 1,q ([0, T ], )C
W 1,q ([0, T ], ) for 1 q < , see remark 1.1.2.
3.3.4 Remark. (a) The assertion (b) of theorem 3.3.3 is not true any more for the case that q = ,
cf. remark 1.1.2.
(b) In the same way as in theorem 3.3.3, it can be seen that
H : [0, 1] W 1,q [0, T ], U C [0, T ], ( )3 3 U
(F )
(, F ) 7
, ,
, el , pl , , u
68
k (t)k
For u, there holds
k u(t)kU
k u(t)kU
e
r1 + c1 (r1b + 1) k (t)k + qeB h() + c1 hB F (t),
eb1 + c1 (ebb1 + 1) k (t)k + qR h() + c1 hB F (t).
e
d r1 + c1 (r1b + 1) k (t)k + qeB d h() + c1 hB F (t),
d eb1 + c1 (ebb1 + 1) k (t)k + qR d h() + c1 hB F (t).
All estimates are valid everywhere in [0, T ]. All abbreviations are as in theorem 3.1.3 with
eB for eB. Further
1
h() =
eB1
R
.
(3.64)
,
For = 1, all rightmost sides are vanishing.
Proof: It is similar to the proof of theorem 3.1.3. The estimates for , , and el
are completely unchanged because of (3.54). But in the present situation, we have
e
1
pl = eB1
R
(3.65)
k (t)k
1 e
kR1 (t)k + k(eB 1
R ) (t)k ,
keB1
(t)k
k(eB 1
)(t)k .
(3.66)
(3.67)
The second summand in (3.66) resp. (3.67) is estimated with (1.14) and (3.1) with eB resp.
(2.42). The rest is again straightforward.
3.3.6 Corollary (Error homotopy II) Assume e 0 = 0 and pl (0) = 0. There exist nonnegative constants dependencies in brackets such that the following estimates hold.
(1) Estimates for the stresses. There holds
k (t)k
k (t)k
k (t)k
k(e )(t)k
cpl (R, B)h + kpl (R, B)h() F (t),
c (R, B, C)h + k (R, B, C)h() F (t).
69
cu (R, B, C, D)h + ku (R, B, C, D)h() F (t).
All estimates are valid everywhere in [0, T ]. Here = and F (t) from (3.12). The
function
kR1 k, kP Ck ,
[0, 1] 7 h = (1 )
min{B , R }
h0 =
kR1 k, kP Ck ,
,
min{B , R }
h(0) = kR1 k, .
(3.68)
Proof: Take note of estimates (3.51) and (3.55). Thus, by taking infimum and supremum,
eb1 eb1 ,
b1
1
,
min{R , B }
e b
b1
kBk ,
,
min{R , B }
qeB
CRB
cRB
in theorem 3.3.5 and (3.21), (3.22) is possible. (3.27) and (3.28) give h h and hB kBk h .
For h() such an estimate is not possible.
We now have an alternative proof of theorem 2.7.3 for the differences of models (E) and
(EPK) with the aid of our homotopy.
3.3.7 Corollary (Difference of models E and EPK, IV) Let the premises of theorem 2.7.1
hold. Assume e 0 = 0 and pl (0) = 0. There exist non-negative constants dependencies in
brackets such that
k(e )(t)k
e
k(eu u)(t)kU
Here again F (t) from (3.12). All estimates are valid everywhere in [0, T ].
Proof: Let = 0 in theorem 3.3.5 and corollary 3.3.6, giving h0 and h(0) as in (3.68).
3.3.8 Remark (Preconditioners) Our homotopy will hopefully give rise to the construction of
preconditioners for the elastoplastic boundary value problem. For details, the reader is referred to Allgower/Georg [1] or section 4.4.2 in Deuflhard [31].
The results of this section can be adapted for linear kinematic plus isotropic hardening material (EPKI), cf. section 4.1, and for the class of CBV -functions, cf. section 4.2.
70
Chapter 4
Two generalisations
During the becoming of this thesis, some facts of the preceding chapters 2 and 3 turned out
to be straightforwardly generalisable.
(1) First, in an appropriate model formulation, linear kinematic plus isotropic hardening
material (EPKI) is considered in section 4.1. The main result is again, that the difference
of elastic and elastoplastic model solutions can be written up to some linear operators
as (a projection of) a play operator, cf. theorem 4.1.7. It makes another contribution to
the central question on top of figure A. In the same way as for (EPK) material, difference
estimates, robustness results and a homotopy are received for (EPKI) material .
(2) Second, as the spaces W 1,1 = AC and CBV = C BV do not differ too much, the
wellposedness and stability results can be adapted for the more general CBV class, cf.
theorems 4.2.4 and 4.2.3. The Sub-Pythagoras relation (1.17) for CBV simply takes
the place of the Pythagoras relation (1.12) for AC. This is done in section 4.2.
Throughout this chapter, we let again the assumptions on from part I hold. We pick up
all notations of chapter 2.
4.1
We now consider linear kinematic plus isotropic hardening material. We call it the (EPKI)
model. The governing relations are given by the Newtonian balance equations
(t) = Du(t),
D (t) = F (t)
e. in [0, T ],
(4.1)
= Bpl (t),
e. in [0, T ]
(4.2)
and the additive decomposition of the total strain and total stress
(t) = el (t) + pl (t),
e. in [0, T ].
(4.3)
72
two generalisations 4
These are identical to (2.7), ..., (2.9), The difference is, that we replace (2.10) and (2.11) by
the extended normality rule
pl (t), e(t)
C (t), s(t)
a. e. in [0, T ]
(4.4)
and the extended initial memory
(, s)(0) = (0 , s0 ) C.
(4.5)
e. in [0, T ]
(4.6)
between the new introduced scalar strain e and scalar stress s. According to section 2.4 in
[76], combined linear kinematic and isotropic hardening is modelled.
Krejc
(
(
(
(((
(((((
(
(((
(
(
(
((((
Figure B. The uniaxial equivalent cases. Left hand side: Linear kinematic hardening (EPK). Right hand
side: Linear kinematic plus isotropic hardening (EPKI).
The spaces.
We set
= L2 , R33
L2 , R .
s
(1 , e1 ), (2 , e2 ) = 1 , 2 L2 (,R33 ) + e1 , e2 L2 (,R) ,
k k2 = , ,
which are appropriate spaces for the extended stress and strain. There clearly holds
k k k k ,
k kL2 (,R) k k .
(4.8)
4 two generalisations
73
The operators. The operators B, C, R and D are the same as in (2.17), (2.23) and (2.15).
We assume that the new introduced operator
K L L2 (, R), L2 (, R) ,
defined by
Ke, f =
K(x)e(x)f (x)dV(x),
e, f L2 (, R) ,
K(x) K > 0,
(4.9)
kKekL2 (,R) kKk kekL2 (,R) ,
Ke, f = e, Kf ,
Ke, e K kek2L2 (,R)
BK = (B, K),
CK = (C, K),
RK = (R, K),
(4.10)
. It is straightforward to see, that BK , CK and RK are symmetric
which are all in L ,
and strongly positive, i. e.
BK , BK k
k2 ,
CK , CK k
k2 ,
RK , RK k
k2 .
The choice
for .
BK min{B , K },
CK min{C , K },
RK min{R , K }
(4.11)
for the coercivity constants is possible. The inverse operators exist and are given by
1
= (B 1 , K1 ),
BK
1
= (C 1 , K1 ),
CK
1
1
R1
K = (R , K ),
,
. The equivalence constants are denoted by
lying in L
cBK k k k kBK CBK k k,
(4.12)
(4.13)
(4.14)
x : R R,
(x )
74
two generalisations 4
x is concave, upper semicontinuous, 0 (x) x ().
Define
=C
, = (, s)
: x (x) x s(x) for almost every x in .
C
,
n (x) (x) in k kR33
s
sn (x) s(x) in | |
a. e. in .
x (x) lim inf x n (x) lim sup x sn (x) x s(x)
n
a. e. in .
due to (4.15) and the semicontinuity properties of almost every x , x . Therefore, finally
(, s) C.
4.1.2 Lemma (Expanding Mises set) The expanding Mises set, given by
= (, s)
: k dev (x)k (s(x)) for almost every x in
C
(4.16)
Lemma 2.3.5 can be adapted straightforwardly for the expanding Mises set (4.16).
4.1.3 Lemma (Expanding Mises set, elastic behaviour) Consider the expanding Mises set
and L2 (, R) be given such that
(4.16). Let (, s)
0 < 0 (x),
k dev (x)k (x) < s(x)
Then we have C , s = {(0, 0)}.
for a. e. x .
4 two generalisations
75
(S F )(t),
s(t)
RK C (t), s(t) ,
(, s)(0) = (0 , s0 ) C.
(4.17)
(,
s),
(, s) R
(, s) + (, s)
(, s)
(, s)(0)
0
(SF, 0)
a. e. in [0, T ]
for all C
e. in [0, T ]
,
e. in [0, T ]
(4.18)
(0 , s0 )
(, s) = PRK (SF, 0), (0 , s0 ) ,
(, s) = SRK (SF, 0), (0 , s0 ) .
(4.19)
We have the following wellposedness result, which is essentially based on theorem 1.1.1.
4.1.4 Theorem (Wellposedness model EPKI in W 1,q ) For each given exterior force and initial memory
F W 1,q [0, T ], U , 1 q ,
and
(0 , g0 ) C,
there exists a uniquely determined decuple
u W 1,q [0, T ], U ,
, el W 1,q [0, T ], ,
,
(pl , e) W 1,q [0, T ],
, , W 1,q [0, T ], ,
,
(, s) W 1,q [0, T ],
given by (2.43), (2.48), ..., (2.53), satisfying the elastoplastic relations (4.1), ..., (4.6).
Proof: Step I. The transformation of the (EPKI) model to (4.18) is done exactly in the same
way as in section 2.5. We have from (4.4)
= R1 (S F ),
K1 s = (pl , e)
C (, s),
R1
K (S F , s)
(i) The validity of the elastoplastic relations (4.1), ..., (4.6) implies (4.17), yielding uniqueness.
(ii) The validity of (4.17) allows the construction of pl and u via (2.45) and (2.46), yielding
existence.
76
two generalisations 4
Therefore, each solution of (4.1), ..., (4.6) induces a solution of (4.17) and vice versa. The
validity of (2.48), ..., (2.53) is then verified straightforwardly.
), according to theorem 2.4.1.
Step II. The elastic stress (e, 0) = (SF, 0) is in W 1,q ([0, T ],
Now theorem 1.1.1, gives us the unique solvability of (4.17), the solution given by stop and
). The rest is analoplay in (4.19). Due to (1.2), we have (, s) and (, s) in W 1,q ([0, T ],
gous to the proof of 2.5.1.
As for linear kinematic hardening, (, s) = (, 0) (, s) resp. (, s) can be expressed as
a play resp. a stop operator.
(BK ) W. r. t. the input and the
(4.1), ..., (4.6) equivalently as
(
,
s),
(,
s)
BK
(, s) + (, s) =
(, s)
(, s)(0) =
C
(0 , s0 )
a. e. in [0, T ]
for all C
e. in [0, T ]
,
e. in [0, T ]
(4.20)
(, s) = PBK (, 0), (0 , s0 ) ,
(, s) = SBK (, 0), (0 , s0 ) .
(4.21)
This is the infinitely dimensional, globally acting counterpart to the exposition in section 2.4
[76].
in Krejc
The scalar stress s resp. scalar strain e can be interpreted as scalar accumulated stress resp.
strain.
4.1.5 Lemma (Accumulation) If the functions x : R R in lemma 4.1.1 are non-decreasing
for a. e. x , then for the scalar stress s and strain e, corresponding to a given F
there holds
W 1,q ([0, T ], U ) and initial memory (0 , s0 ) C,
e(x)
0 s(x)
for a. e. x .
(4.22)
then (, s) C
for each s L2 (, R) , which satisfies s s a. e. in .
Proof: If (, s) C,
Inserting (, s) into (4.18) or (4.20), we find
hK1 s,
s si 0
for each such s, especially for the choice s s + 1. Thus (4.22) holds due to (4.9) and (4.6).
As for linear kinematic hardening, it is possible to express all remaining unknowns in terms
of
(, s)(t)
and
(, s)(t) = (SF, 0) (, s) (t),
(4.23)
namely
(4.24)
4 two generalisations
77
(4.25)
(, 0)(t) = BK R1
K (, s)(t) + (, s)(t),
(4.26)
(4.27)
1
1
BK + I R1
(, e)(t) = CK
K (, s)(t) + CK (, s)(t),
(4.28)
u(t) = (D CD)1 D (BK + CK )R1
K (, s)(t) + (, s)(t) .
(4.29)
Formulas (4.23), (4.24), ..., (4.29) are the counterparts of (2.47), (2.48), ..., (2.53). It is convenient to abbreviate the extended stresses and strains by
= (, 0),
= (, e),
= (, s),
pl = (pl , e),
= (, s),
el = (el , 0).
= (, s),
(4.30)
,
, ,
With the linear projection apply Riesz-Frechet identification
,
:
,
:
(, e) 7 ,
(, s) 7 ,
there holds
kk, 1,
k( , 0)k = k k = k k = k( , 0)k ,
k k , 1
(4.31)
and
= ,
= ,
= ,
= ,
= ,
el = el ,
pl = pl .
W 1,q ([0, T ], ),
).
W 1,q ([0, T ],
= G(
, 0 ),
= F(
F(, 0 ) = G 1 (, 0 ),
78
two generalisations 4
W 1,q [0, T ],
C
,
F : W 1,q [0, T ],
W 1,q [0, T ],
C
,
G : W 1,q [0, T ],
1
1
PBK (, 0), (0 , s0 ) ,
(, 0) + BK
F (, 0), (0 , s0 ) =
CK
G (, e), (0 , s0 ) = CK (BK + CK )1 BK (, e) + SBK +CK CK (, e), (0 , s0 ) ,
are turning stress into strain and vice versa. They are continuous for 1 q < .
Proof: Similarly as in corollary 2.5.5, there holds
1
PBK (
, 0 ),
+ CK B K
SBK +CK
, 0 ),
+ CK B 1 PB (
PB +C
K
, 0 ),
0 = SBK (
1
0 = (I + CK BK
)PBK (
, 0 ).
(4.32)
(4.33)
Namely, apply lemma 1.4.1 (1.46), (1.47) with A1 = BK , A2 = CK , s0 = 0 and (4.21). The
rest is done as in the proof of theorem 2.5.6 with (4.30).
Now, the next theorem is the main one of this section.
4.1.7 Theorem (Difference of models E and EPKI, I) Let F W 1,q ([0, T ], U ), 1 q ,
be given. Up to some linear operators, the difference of the solutions of
and (0 , s0 ) C
model (E), i. e. e, e and eu, which are corresponding to F , and model (EPKI), i. e. ,
and u, which are corresponding to F and (0 , s0 ), are projections of play operators. More
precisely, there holds
= H,R PRK (SF, 0), (0 , s0 ) = H,B PBK (, 0), (0 , s0 ) ,
e
= H,R PRK (SF, 0), (0 , s0 ) = H,B PBK (, 0), (0 , s0 ) ,
e
u u = Hu,R PRK (SF, 0), (0 , s0 ) = Hu,B PBK (, 0), (0 , s0 ) ,
4 two generalisations
79
Here CRK , CBK are from (4.13), (t), (t) from (2.92), (2.93) and the linear operators H,
from (2.87).
Proof: Analogous to theorem 2.7.3. Use further (4.31).
4.1.9 Theorem (Difference of models E and EPKI, III) Let the premises of theorem 4.1.7
hold. For the difference of the solutions of models (E) and (EPKI), the following estimates
hold almost everywhere in [0, T ].
(RK ) In terms of e = S F , there holds
for the stresses:
k(e )(t)k
e
for the strains:
k( )(t)k
e
for the displacements: k( u u)(t)k
(BK ) In terms of ,
there holds
for the stresses:
k(e )(t)k
e
for the strains:
k( )(t)k
,
CBK kH,B k,BK k(t)k
,
CBK kHu,B kU,BK k(t)k
4.1.10 Theorem (Robustness model EPKI) We consider two sets of data Bi , Ci , Ki (symmetric and strongly positive), inducing the operators in (2.68) and
BK,i = (Bi , Ki ),
CK,i = (Ci , Ki ),
RK,i = (Ri , Ki ),
(4.34)
Always i = 1, 2. With the abbreviaand two inputs Fi W 1,q ([0, T ], U ) and (0,i , s0,i ) C.
tions as in theorem 2.6.3 and
K = min K1 , K2 ,
k = max kK1 k, kK2 k ,
the following estimates hold in terms of the input kF kU , k(0 , s0 )k and parameter
disturbances kBk, kCk, kKk, namely
Z t
kKk2 1/2
(kBk + k2 kCk)2
+
+ k6 kCk
+ k5
4B
4C
Z t
kF1 kU + kF2 kU d
0
80
two generalisations 4
and
k(, s)(t)k k3 kF (t)k +
Here
r2 + k2
,
k4 =
min{B , K }
k1
kCk kF1 (t)k + kF2 (t)k + k(, s)(t)k .
2
p
k3 (r 2 + k2 )3
k5 =
,
min2 {B , K }
k6 =
rk1
.
2B
Here kBk, kCk, kKk are w. r. t. their respective L2 -norms. All estimates are valid
everywhere in [0, T ].
Proof: Analogously to theorem 2.6.3 and corollary 2.6.4. Note, that for the norm equivalence
constants, we have
(B.14)
C(Ri ,Ki )
and
(B.14)
1
c(Ri ,Ki )
(4.11)
1
(Ri ,Ki )
(2.84)
1
min{R , K }
1
min{B , K }
(2.84)
r2 + k2
r2 + k2
k(Ri , Ki )k (4.14)
.
p
p
(Ri ,Ki )
min{B , K }
min{R , K }
(2.79)
1
1
(kBk + k2 kCk)2 + 4 kKk2
4
B
K
with the usual trick K1 = K11 K K21 . Note further estimate (4.31).
4.1.11 Corollary (Robustness model EPKI, II) For each pairs Bi , Ci , Ki , there exist non-ne(j)
gative constants c , depending on the norms kBi k, kCi k, kKi k, the coercivity constants Bi ,
Ci , Ki , and , d, such that for each pair Fi and (0,i , s0,i ) the difference k()(t)k can be
written as sums of
Z
t
(4)
(3)
(2)
(1)
kF1 kU + kF2 kU d
c kBk + c kCk + c kKk
c k(0 , s0 )k ,
0
and
(5)
c
t
0
kF kU d,
(7)
(6)
c kBk + c kCk kF1 (t)kU + kF2 (t)kU .
kel k cel kS F k,
kk
c kS F k
(4.35)
4 two generalisations
81
kk
c kS F k,
(for the stresses),
kk
c kS F k,
kk
c kS F k (4.36)
kuk
U cu kS F k
(4.37)
kek
L2 (,R) ce kS F k
(4.38)
(for the scalar stress and strain). All the right hand side norms are kk = kk . All estimates
are valid a. e. in [0, T ].
Proof: Like the proof of theorem 2.6.2. Use (4.31) and (4.8).
As in section 3.3, we are able to connect the solution of the (EPKI) model continuously to
the solution of model (E). For [0, 1], we define the elastic parameter
e
we set
For a given elastic initial memory (e 0 , es0 ) C,
(e 0, , es0, ) = (0 , s0 ) + (1 )(e 0 , es0 ).
(4.39)
(e , es ) = PeBK , (SF, 0), (e 0, , es0, ) ,
= e ,
= B eB 1
,
=
+ ,
s = es ,
(4.40)
e = K
s ,
(4.41)
the -strains by
e 1 e
pl
= B ,
pl
= el
+
,
,
el
=C
the -displacements by
u
= (D CD)1 D C .
(4.42)
We receive
for = 0, i. e. for eB K,0 = BK and (e 0,0 , es0,0 ) = (e 0 , es0 ), the elastic solution
0 = e,
0 = e,
u
0 = eu.
1 = , 1 = ,
1 = ,
pl
el
s1 = s, 1 = , pl
el
1 = e,
1 = , e
1 = ,
u
1 = u.
:=
0 ,
:= 0 ,
e el
:= 0 = el
0,
:= pl
0 = 0,
e pl
s := s0 = 0,
e := e0 = 0.
82
two generalisations 4
is a homotopy
H(0, F ) = e, e, e, es, eel , epl , e, ee, eu (F )
H(1, F ) = , , , s, el , pl , , e, u (F ).
4.2
Both the (EPK) and (EPKI) problems may as well be posed and uniquely solved in the space
CBV = C BV , because of three reasons.
The main ingredient of proof for the wellposedness results, theorems 2.5.1 for (EPK)
resp. 4.1.4 for (EPKI), is the wellposedness theorem 1.1.1 for stop and play.
The transformation of the governing model relations (2.7), ..., (2.11) to (2.55) resp.
(4.1), ..., (4.6) to (4.19) is of purely spatial nature.
The Sub-Pythagoras relation (1.17) for the CBV class replaces Pythagoras relation
(1.12) for the W 1,q class.
So it is an interesting theoretical aspect to admit inputs of the larger class CBV , yielding
outputs of CBV .
But first, we concern ourselves with the elastic boundary value problem (E), which is trivial.
4.2.1 Theorem (Wellposedness model E in CBV ) For each given exterior force
F CBV [0, T ], U
there exists a uniquely determined triple
e
e
u CBV [0, T ], U ,
CBV [0, T ], ,
(4.43)
CBV [0, T ], ,
given by (2.31), ..., (2.33), satisfying the elastic relations (2.5), (2.6).
Proof: In the same way as the proof of theorem 2.5.1 for W 1,q .
4 two generalisations
83
4.2.2 Corollary. If in lemma 2.4.2, the scalar load functions Li are of class CBV ([0, T ], R) for
each i, then e CBV ([0, T ], ), e CBV ([0, T ], ) and eu CBV ([0, T ], U ).
Proof: Clear.
Now the elastoplastic boundary value problem. We replace the normality rules (2.10) resp.
[75].
(4.4) by the following weakened forms, cf. lemma 1.25 in Krejc
For linear kinematic hardening (EPK), the normality rule (2.10) is replaced by the
weakened normality rule
Z
(t) C,
T
0
( )( ), dpl ( ) 0
(4.44)
For linear kinematic hardening plus isotropic hardening (EPKI), the extended
normality rule (4.4) is replaced by the weakened extended normality rule
(t), s(t) C,
E
(, s) (, ) ( ), d(pl , e)( ) 0
(4.45)
The integral is the Riemann-Stieltjes integral for functions with values in a separable Hilbert
space. We easily arrive at the following results, where now part (a) of Krejcs theorem 1.1.1
for CBV functions yields wellposedness.
4.2.3 Theorem (Wellposedness model EPK in CBV ) For each given initial memory and exterior force
F CBV [0, T ], U
and
0 C,
there exists a uniquely determined octuple
u CBV [0, T ], U ,
, el , pl CBV [0, T ], ,
, , , CBV [0, T ], ,
satisfying the elastoplastic relations (2.7), ..., (2.9), (2.11) with the weakened form of the
normality rule (4.44).
Proof: In the same way as in the proof of theorem 2.5.1, rule (4.44) of the problem (EPK)
can be transformed to
Z T
(t) C,
( )( ), R1 d( ) 0
0
for all t [0, T ] and C([0, T ], C), which is the weakened form of (2.43). The rest is done
as in the cited proof.
4.2.4 Theorem (Wellposedness model EPKI in CBV ) For each given exterior force and initial memory
F CBV [0, T ], U ,
and
(0 , g0 ) C,
84
two generalisations 4
satisfying the elastoplastic relations (4.1), ..., (4.3), (4.5), ..., (4.6), with the weakened form of
the normality rule (4.45).
Proof: In the same way as in the proof of theorem 4.1.4, rule (4.45) of the (EPKI) problem
can be transformed to
Z TD
E
d(,
s)(
)
0
(, s) (, ) ( ), R1
(, s)(t) C,
K
0
Theorems 5.3.1, 5.4.1 for the isolated stress-strain laws and theorems/corollaries 2.5.6 (resp.
4.1.6), 2.6.3 (resp. 4.1.10), 2.6.4 (resp. 4.1.11), 2.7.1 (resp. 4.1.7), 2.7.3 (resp. 4.1.8), 3.1.3
and 3.1.7 for (EPK) (resp. (EPKI)) can be easily adapted for the CBV input and output
class.
Chapter 5
5.1
86
33 , such that
C, B and eB in L R33
s , Rs
C = C ,
B = B ,
B = eB ,
C : C kk2 ,
B : eB kk2 ,
(5.1)
be given. B represents the elastoplastic linear kinematic hardening parameter, eB a so-called
elastic linear kinematic hardening parameter and C is Hookes operator with components
given in (10). C , B , eB are positive coercivity constants. We assume that they leave the
deviatoric parts invariant, i. e.
dev(eB) = eB dev(),
B : B kk2 ,
dev(B) = B dev()
(5.2)
: (e d (t) ) 0
B
e (0)d
e.
in [0, T ]
e (0) = e
d
0,d
: (d (t) ) 0
B
(t)
B
(0)
e. in [0, T ]
d
d
d (0) = 0,d
Here
e
resp.
(0)d = B
(0) R33 ,
B
sd
87
and
(0)d RI,
C=B
e (0)d , B
(0)d (RI)
B
are convex cylinders, and we have from remark 4.5 in [76] that the hydrostatic and the
deviatoric components decouple. This means
e
= PeC,eB (e, e 0 ),
= SeC,eB (e, e 0 ),
= PC,B (, 0 ),
= SC,B (, 0 ), (5.5)
imply e = e + e, = + ,
=
+ (in total), es = e + e d , s = + d , s =
+ d (for
e
e
e
d = es e,
(5.6)
pl = eB 1e,
d = e e d ,
= B eB 1e,
d = s ,
(5.7)
cf. (2.8), (2.9). Now according to remark 4.4 in [76], an equivalent formulation via the
following differential equations (with discontinuous right hand side) is possible.
elastic stress space
e
n
6
76
e
d
e
s
3A
e
Ae
A
A
A
r
Se (e)
0
6
r
X
d
XX
XXX
0
XXX
z
@
@
@
S (
)
(5.8)
d = e e d ,
(5.9)
(5.10)
88
(x R),
and en = en(e d ) = (e d ), n = n(d ) = (d ) are the outward unit normals at the elastic
resp. elastoplastic yield surfaces, cf. (C.3). The situation is depicted in diagram (6.5), where
e
e
V = R33
s . Note that there holds n = (d ) = ( d ).
The closed-form differential equations (5.9), (5.10) are very cumbersome for use. Therefore,
we give an equivalent detailed presentation, which allows the transfer to Jiangs model (and
any other constitutive model of the collection on page 5) later in the practical part of this
thesis. This procedure is standard, cf. Lemaitre/Chaboche [79].
The corrected quantities. In the case of elastic behaviour es(t) Be (e), where the elastic
stress deviator is inside the interior of the elastic domain, we have
dt = 0,
dt pl = 0,
dt ei = 0,
dt
i = 0,
(5.11)
(5.12)
In the case of neutral loading es n 0, where the elastic load is not directed into the outward
direction, (5.11) holds again. In the case of active plastic yielding, we have
d
= B en,
d e = eB en,
H = en : eB en
(5.13)
es
n
,
eH
dt pl = dt en,
dt e = dt d e,
dt
= dt d
.
(5.14)
s e) = B pl +
s eB pl ,
(5.15)
which we call the simultaneity relation, which will turn out to be one of the key features for
numerical implementation, cf. section 7.2. For illustration see (5.8) and figure C on page 102.
Note that, if e = , the simultaneity relation is just d = e d , in accordance to (3.3).
Now, in order to get the corrected stress
from its deviator s, we use a trick, which is possible,
if is a free boundary point, i. e. 1 \ 1,g , cf. the introduction. Then the elastic stress
e is plane, i. e.
e
e
AC([0, T ], R33
s AC([0, T ], R33
sp ),
spd ).
We call this situation the free boundary case. Clearly, is this case, as well the elastoplastic
stress is plane, i. e.
AC([0, T ], R33
sp ),
s AC([0, T ], R33
spd ).
89
sp ), Now, since, s
33
AC([0, T ], Rspd ), we simply may pull back via
= D 1 s.
(5.16)
where D from (5.20) is the restriction of the deviatoric projection, explained in remark 5.1.2.
Now finally,
el = C 1
,
= pl + el ,
(5.17)
in accordance to (3.4). Equations (5.11), ..., (5.17) now constitute the pointwise correction
model for linear kinematic hardening material.
Summarising, we have written the pointwise correction model, equations (5.11), ..., (5.17), as
some differential equations with discontinuous right hand side,
e,
e,
,
(5.18)
= Jes,e, t, ,
dt pl , ,
followed by some explicit algebraic equations,
(
, s, , el )(t) = Les,e, (t, pl , e,
).
(5.19)
R33
sp A = (aij )i,j=1,2,3 7 A = (a11 , a22 , a12 ) R
resp.
t
3
R33
spd A = (aij )i,j=1,2,3 7 A = (a11 , a22 , a12 ) R .
In the latter, we simply forget the component a33 = a11 a22 . Now notice that the
restriction
33
: R33
(5.20)
D = R33 | dev |R33
sp Rspd
sp
spd
2/3 1/3 0
= 1/3 2/3 0 ,
D
0
0
1
| dev |R33
R33
s
sd
is obviously commutative.
1
D
2 1 0
= 1 2 0 ,
0 0 1
R33
sp
- R33
spd
?
- R3
?
R3
(5.21)
90
For the elastoplastic quantities, i. e. the elastoplastic BVP solution, there holds analogously in the case of elastic behaviour s(t) B ()
dt pl = 0,
If s(t) B (),
dt = 0,
n = n(t, ) = (s(t) ),
dt = 0.
(5.22)
s n = s(t)
: n.
(5.23)
In the case of neutral loading s n 0, (5.22) holds again, otherwise in the case of active plastic
loading s n > 0
d = B n,
H =n:Bn
(5.24)
and again after reparametrisation w. r. t. time
dt =
s : n
,
H
dt pl = dt n,
dt = dt d ,
(5.25)
System (5.22), ..., (5.25) can be seen as a subsystem of system (5.11), ..., (5.14) with e s and
s cancelled.
5.1.3 Remark. (a) Note that if we choose especially eB = B and = e, we receive s = es,
= e,
i. e. the pointwise correction model reduces to the identity in accordance to (3.10).
(b) In general, the normals en and n will not be equal. But all our numerical experiences in
practice show, that en and n are very close to each other.
5.1.4 Remark (Pointwise correction model for Jiang material) We take a short look forward. Compared to the pointwise correction model for (EPJ) material later in chapter 6, the
situation here is simpler.
The radii e resp. of the elastic resp. yield surfaces are constant.
The total backstresses e and
are in linear relation to the plastic strain pl . There is
no additive decompositions into partial backstresses.
There are no memory surfaces present.
In fact, the function Jes,e, does not depend on and Les,e, does not depend on e,
.
(a) Equations (5.11), ..., (5.17) correspond to (6.8), ..., (6.24), (6.29), (6.30), equations (5.18),
(5.19) correspond to (6.2), (6.6), diagram (5.8) corresponds to diagram (6.5).
(b) In contrast to Jiang material in chapter 6, where softening eH < 0 resp. H < 0 is possible,
due to the positively definiteness of eB resp. B, we have
(5.1)
H = en : eB en eB > 0,
(5.1)
H = en : B en B > 0,
and therefore the introduced internal scalar variables and are total corrected resp.
elastoplastic accumulated plastic strain, i. e.
= (0)
+
(t)
t
0
pl
kd ( )kd,
(t) = (0) +
kd pl ( )kd.
91
(c) If we already take a look forward to the flux computations for Jiangs model in section
6.3.2, the linear system of equations (6.90) is in the present context given by
and
e,
x = (t, y) = (t, pl , ,
),
I (x) = 1, 0, 0, 0, 0 ,
I + (x) = 1, , , d e,
e (x) = kes(t) ek e,
es
en
: en e e e
n : B n = es : en.
: eB en
The linearisation of the pointwise correction model in the case of active yielding, cf. section
7.3, reads
s
pl e 1 e
e 1 e
e
e
e
=
B
B
= B n en.
n
n
+
I
n
,
e
es
es
(d) Clearly, the algorithms I, III and III of chapter 7 are applicable in the same way as they
will be for the correction model for Jiang material.
The estimates of the next three sections 5.2, ..., 5.4 were the first approaches for error estimates
for our correction model. They are of course true, but they do not at all explain, why the
output of the correction model k
s(t)s(t)k should be better than the elastic error kes(t)s(t)k.
Therefore, as they lead us to be a dead-end street, they might be skipped by the reader in
the first instance.
Just have a look at theorem 5.4.1 for example. The best choice in order to make the right
e
hand
R t e side in (5.41)
R t resp. (5.42) as small as possible is B = B. Then the bad integrals
resp. 0 kskdt
k s skd
.
k(d d )(0)k +
k(d d )(t)k
cB
0
92
5.2
A first estimate
We consider in this section the pointwise correction model, where we assume the operators B
and eB to be just positive multiples of the identity operator, i. e.
B = b I,
B = eb I,
( eH = eb)
(5.26)
with scalars b, eb > 0. The deviatoric invariance (5.2) and the positive definiteness (5.1) are
then clearly valid. Systems (5.3) resp. (5.4) then simplify to
e
e (0)d a. e. in [0, T ]
(t)
(t)
: (d (t) ) 0
...
(0)d a. e. in [0, T ]
for all B
(5.28)
since we are able to cancel the positive scalars eb resp. b. We allow the yield radii e and to
be different. We introduce the abbreviations
d = e e d =: e d ,
b
eb
=: e
(= b pl )
for the ratios b/eb and /e. First, we decompose the correction model error into three parts
k(
s s)(t)k = k(
(t) + d (t)) ((t) + d (t))k
k es(t) s(t)k + k e(t) (t)k + k e(t) (t)k.
(5.29)
Let us assume
1
i. e.
b e eb
(5.27),(5.28)
d : d + 2
I1 (t) =
I2 (t) =
I3 (t) =
d e : e d = I1 (t) + I2 (t) + I3 (t)
d ( ) e d ( ) : (
) e(
) d,
e d ( ) ( ) : (
) d,
d ( ) e( ) : e(
) d.
This allows us to treat the second and the third summand in (5.29) similarly. The first
summand in (5.29) remains unchanged. We consider the second summand in (5.29). Let
= and set
d
e d
if ke d k
e
,
=
.
=
e
( d ) if ke d k >
93
Since
kd k and ke d k e
kek e and kk ,
we may use them as test functions into the variational inequalities (5.27) and (5.28). We have
I3 0 by construction of e, and
Z t
Z t
(s es) : ( e)
( e) : ( e)
I1 (t) =
d
d
0
0
Z t
1
e
( e)(t)
2
( e)(0)
2 .
kd ( e)( )kd
k s sk,t
2
0
The integrand of I2 is zero, if ke d k / , otherwise the estimate
I2 (t) k d (t)k e
k(t)k
e ( ) k(t)k
k d (t)k
e(0)
(5.30)
d
,
= e d ,
Putting those three estimates together, we have an upper bound for the total difference (5.29)
of the approximation above.
[75], proposition 3.5, is
5.2.1 Remark. For 1, which enforces 1, estimate in Krejc
revealed as a special case.
5.3
A second estimate
We consider the same situation as in section 5.2. Now we give an error estimate for the
pointwise correction model for the case that we have
= =
b
eb
and
= 1.
(5.31)
94
Thus the elastic and the corrected yield surface are assumed to have the same radii. We
introduce the following abbreviations
1/2
F(, e d , d )(t) = ke d (t) d (t)k2 + k(e d d )(t)k2
,
1/2
2
F (, ) = + 6 + 1
.
There holds
(CS)
F (, ) : [0, 1] R, 7 F (, ),
F (0, ) = ,
F (1, ) = 8
is concave and strictly increasing. We first distinguish two cases.
Case A. If at time t both stop operators d and e d are in touch with the yield surface, i. e.
ke d (t)k =
and
kd (t)k =
(5.32)
we have
dt
= (es : en)en =
(es : e d ) e
d,
2
dt = (s : n)n =
s : d
d .
2
+
(
s
(
s
)
(
s
s)
d d
d d
d
d
2
1
1
(es s)
: d d
+ 2
(es : e d )e d (es : d )d
.
(5.33)
(5.32)
2 2 (es : e d )2 + (es : d )2 2(es : e d )(es : d )
(es : e d )(es : d ) ke d k2 + kd k2 2(e d : d )
2
k(es : e d )e d (es : d )d k2 2 es : (e d d )
=
(CS)
2
2 es : (e d d ) + (es : e d )(es : d )ke d d k2
2 kesk
2 ke d d k2 + 2 kesk
2 ke d d k2
2 kesk
2 F 2 (e d , d , ),
=
thus we conclude
kdt (
)k
(CS),(5.33)
kes sk
+
1 e
k sk
F(e d , d , ).
95
Case B. In the case where at least one of the two stop operators d or e d is in the elastic
domain, i. e.
ke d k <
or
kd k < ,
(5.34)
we find in view of
kdt
k =
1
e
2
s : d dt k d k ,
2
1
1
2
kk
=
s : d dt kd k
the identity
1
1 d e 2
kdt
k kk
+
k d k kd k2 =
(es : e d ) (s : d ) .
2 dt
(5.35)
and
kk
> 0 ke d k =
sign kdt
k kk
= sign ke d k2 kd k2 .
(5.36)
(5.37)
1
1 d e 2
e e
2
k
k
s
(
s
(5.38)
k
k
+
k d
(
kdt
d
d .
d
2 dt
On the set t : ke d (t)k = kd (t)k this is obvious from (5.35), as the second summand on
the left hand side in
equal to zero. Then there
even holds equality in (5.38).
On the complement t : ke d (t)k =
6 kd (t)k it follows from (5.35) by multiplication with
the sign in (5.37).
We have
kdt (
)k = kdt
k kk
at the left hand side of (5.38) by virtue of (5.34) and (5.36). Thus there holds, by smuggling
a zero,
1 d e 2
dt (
)
+
k d k kd k2
2 dt
(5.38)
1 e e
s : d s : d
1
1 e
: d + es : (e d d )
( s s)
1
ke d d k
kes sk
+ kesk
1
F(, e d , d ).
kes sk
+ kesk
(5.39)
96
which is valid a. e. in the interval [0, T ]. We note, that with the binomial identity there holds
e
k d (0)k2 kd (0)k2
(5.31)
ke d (0)k + kd (0)k ke d (0)k kd (0)k
( + 1) ke d (0)k kd (0)k
2 ke d (0)k kd (0)k.
(5.40)
5.3.1 Theorem. For the pointwise correction model for linear kinematic hardening material, there
holds for each given es, s AC([0, T ], R33
sd )
Z
Z t
0
1 e 2
2
k d k kd k +
k(es s)(
)kd
2
0
t
Z
1 t e
e
+
k skF(,
d , d )d
0
1
e d (0)
d (0)
kd (t)k2 kd (t)k2
2
Z t
Z
F (, ) t e
e
k( s s)(
)kd +
k s(
)kd
+
0
0
kd (
)( )kd
1
kd (
)( )kd =
b
pl
pl
kd (
)( )kd
Here again for consistent initial memories e d (0) = d (0), the first summand at the right hand
side vanishes.
[14], which is
5.3.2 Remark. If = 1, we rediscover theorem A.5 in Brokate/Krejc
Z
t
0
kd (
)( )kd
(e d d )(0)
+
k(es s)(
)kd
Z t
2
kes(
)k k(e d d )( )kd.
+
0
2 k d d k 8 = F (1, ).
5.4
97
A third estimate
Now we pick up the general situation (5.3), (5.4) under the assumption = e. Considering
the simultaneity formula, we find a. e. in the interval [0, T ]
kdt (es s)(t)k
(5.15)
(5.14)
(5.14)
k(eB B)en(t)k
dt (t)
(CS),(5.1) keB Bk
: es(t)
e
k
B
Bk
kes(t)k.
en(t) : eB en(t)
ceB
en(t)
Note here that dt 0 and that we have put en(t) = 0 in the elastic regime. By integration
over the interval [0, T ], we find with the triangle inequality
k(s s)(t)k k(s es)(t)k
k(s s)(t)k
k(s s)(0)k +
(F T )
kd (es s)(t)kd
0
Z
keB Bk t e
k s(
)kd
k(s s)(0)k +
ceB
0
t
0
kes(
)kd.
k( d d )(t)k
ceB
0
Z
CB keB 1 B 1 k t
kskd
+
c2eB cB
0
(5.41)
and
k(e
d )(t)k
Z t
CB
e
e
k s skd
k( d d )(0)k +
cB
0
Z
CeB keB 1 B 1 k t e
+
k skd.
ceB c2B
0
(5.42)
98
f2 = s,
(5.43)
s1 = e d ,
s2 = d ,
(5.44)
p1 = e,
p2 = ,
(5.45)
s0,2 = d,0 ,
(5.46)
, A1 = eB 1 : ,
, A2 = B 1 : .
(5.47)
Part III
PRACTICE
99
Chapter 6
102
notice that in contrast to the original works Jiang/Sehitoglu [63], ..., [67] we use
, R, cR , a , b , , I
instead of
p, RM , cM , ak , bk , 0 , M,
(in this
order) and we substitute the Mises factor 2 into the yield surface radius, i. e.
= 2 k. Thus the overall notation becomes much more consistent. The same nomenclature
will be applied for the elastic e - and the corrected -quantities in the correction model.
The reader should pay careful attention: There exist two constitutive material laws in literature, their inventors named Jiang. We mean Y. Jiang [63], ..., [67], not W. Jiang [68, 69]!
6.1
As in section 5.1 for (EPK) material, we assume now for (EPJ) material that symmetric elastic
(which corresponds to the general situation) or
stresses e AC([0, T ], V ), where V = R33
s
33
more specially V = Rsp (which corresponds to the free boundary case), are given. Their
deviatoric parts are denoted by
e
e
s(t) = dev e(t) ,
s(t)
= dt es(t) = dev e(t)
.
(6.1)
They are external input/force into the (EPJ) correction model, which can be written as a
system of first order discontinuous ordinary differential equations, namely
e1 , . . . , eI , eR,
e1 , . . . , eI , eR,
= Jes t, ,
dt pl , ,
1 , . . . ,
I , R
(6.2)
.
1 , . . . ,
I , R
We obtain them by doubling the Jiang constitutive stress controlled equations. The unfamiliar
reader finds the latter again in section 6.2.1. That way, Jes will be explicitly time dependent,
the so-called elastic stress control. (As e(t) = C e(t), it can as well be considered elastic
strain controlled. Note, that, as dev is linear, dt and dev commute.)
Figure C. It depicts the movement of the elastic e d /e (red) and the corrected d /
(orange) compared
to the elastoplastic d / (green) on the unit yield surface. The red and orange paths are identical due to
simultaneity relation (6.3). The example is taken out of section 8.3, named the butterfly path [d] therein.
103
The movements of the elastic e -variables are thought in an elastic stress space, those of
the corrected -variables in a corrected stress space. The correction idea is based on a
simultaneity assumption. This means, the corrected stress deviator difference d = s
is defined
simultaneously (i. e. parallelly and rescaled)
to the elastic stress deviator difference e d = es e at the corresponding point in the corrected
stress space, cf. (6.3). The movements in the corrected stress space, in order to define s, follow
the movements in the elastic stress space, controlled by es. The rescaled difference tensors
es e
e
s
d
=
= e
= ed
(6.3)
are equal at each point in time. Through the rescaling by the ratio of the yield surfaces
radii e and , it is guaranteed that es(t) and s(t) get into contact/lose contact with their
respective yield surfaces at exactly the same point in time. As a consequence, the normals to
the corresponding yield surfaces
(6.3)
n
= (d ) = (e d ) = en
(6.4)
are always equal and active plastic yielding starts/stops at exactly the same time. For illustration see diagram (6.5) and figure C.
elastic stress space
corrected stress space
e
SR
e
Vd
SR
n6
76
e 6
R
n
s
:
e e
6
s
d
e
R
r
X
d
X
0@ XXXX
3A
e
X
@
XX
z
A
R
@
@
6
- Ae
@
@
i@
R
A
e
H
R e
@
*
e H
A
S ()
i
r
j
H
e
Se ( )
0
(6.5)
, s, , el , eel , epl , e, e d , e,
Les t, pl , e1 , . . . , eI , eR,
=
,
, d , (t)
1 , . . . ,
I , R
(6.6)
, s, , el , eel , epl , e, e d , e,
Kes t, pl , e1 , . . . , eI , eR,
=
1 , . . . ,
I , R,
h ,
, d , (t)
(6.7)
104
for each t [0, T ] being a system of simple explicit algebraic equations. The free boundary
case is the essential one for practice, since cracks usually emerge at the free boundary of the
body, see Lemaitre/Chaboche [79]. The general case is mentioned just for the sake of
completeness.
For the elastic and the corrected stress space, the stress-like internal Jiang state variables
are listed in the following table.
e
Se (e)
e
R
SeR
S()
R
SR
6.1.1
Function Jes defines the ordinary differential equation part of the correction model. Its domain
of definition is
D = [0, T ] R VdI [0, ) VdI [0, ),
where
Vd = dev V from (2.3), (2.4), i. e. V = R33
or V = R33
s
sp .
e1 , . . . , eI , eR,
D we set
For (t, ,
1 , . . . ,
I , R)
e
e(e
e
1 , . . . , I ) =
I
X
i ,
i=1
=
e
d
e
(
1 , . . . ,
I ) =
e
d (t,
e
1 , . . . , I )
ef (eR)
=
=
I
X
and
Se (e) = Be (e),
(6.8)
i,
i=1
es(t) e,
e e
1 + ea e b R
SR = BR (0).
(6.9)
dt = 0,
dt ei = 0,
dt eR = 0,
dt
i = 0,
= 0,
dt R
(6.10)
105
s n = es n (t, e1 , . . . , eI ) = es(t)
: en.
(6.11)
If es n 0 (neutral loading), we let (6.10) hold again, else (active plastic loading) we set for
i = 1, . . . , I
e e
e
(6.12)
i = eQi 2 en : (ei ) 1 + ea e b R ,
K
X
(k)
e (k) ebi
1
+
if 0
ai e
e
e
k=1
(6.13)
ci = ci
K
X
(k)
e
ai
if < 0
1+
k=1
and
dei =
de =
e e
ci r i
I
X
kei k
er
i
ei +1
(ei ) ,
dej ,
(6.14)
(6.15)
j=1
e
e
e
cR 1 k k/ R
+
(e) : de
deR =
kdek
(6.16)
in the elastic stress space. And analogously with the same yield surface normal en according
to the simultaneity (6.4) we define
(6.17)
i ) 1 + a eb R ,
i = Qi 2 en : (
K
X
(k)
(k)
1+
ai ebi if 0
k=1
ci = c
(6.18)
i
K
(k)
1+
ai
if < 0
k=1
and
d
i = ci ri
d
=
I
X
k
i k
ri
i +1
(
i ) ,
d
j ,
(6.19)
(6.20)
j=1
k/R
cR 1 k
+
=
(
) : d
dR
kd
k
if
if
if
> 0,
R
BR (0)
> 0,
R
BR (0)c
=0
R
(6.21)
106
H = en : de + de.
dt R
dt
i = dt d
i.
(6.22)
(6.23)
(6.24)
Putting all those pieces together, the dependence structure of the tensor and scalar valued
function
Jes = (Jpl , J, Jei , JeR , J i , JR )
is given by
dt pl
dt
dt ei
dt eR
= Jpl (t, e1 ,
=
J(t, e1 ,
= Jei (t, e1 ,
= JeR (t, e1 ,
and
dt i =
dt R =
...,
...,
...,
...,
e ,
I
e ,
I
e ,
I
e ,
I
),
),
eR, ),
eR, ),
eR,
eR,
(6.25)
J i (t, e1 , . . . , eI , eR, ,
1, . . . ,
I , R),
JR (t, e1 , . . . , eI , eR, ,
1, . . . ,
I , R).
(6.26)
Of course, we have dt ei , dt
i , dt pl in Vd , thus ei ,
i , pl remain in Vd due to lemma 6.3.6,
since Vd is a linear space. We remark that (6.25) for pl is just an integration, not a genuine
differential equation, since the right hand side does not depend on pl . Equations (6.14) and
(6.19) for the partial backstresses seem to be discontinuous, but in fact are not, as expansion
shows.
The initial values/conditions for Jes . We assume that the initial values pl (0), (0),
k i (0)k < r i ,
k
i (0)k < ri ,
(6.27)
R(0),
(0)
(0)
R(0),
j
j
e (0), eR(0),
i
j=1
for i = 1, . . . , I and
pl (0) Vd ,
R,
(0)
j=1
I
X
e e
e
e
j (0)
e 1 + ea e b R(0) .
s(0)
(6.28)
j=1
i (0) =
i (0) = i (0) for all i = 1, . . . , I,
R(0) = R(0)
= R(0),
where i (0), R(0) are the initial values for the transient elastoplastic boundary value problem
e(0) (e(0)).
(5), ..., (7) and kei (0)k 1, eR(0) 1 and es(0) B
6.1.1 Remark. en, eH, dt pl and are common for both stress spaces, and already uniquely defined
by the movements in the elastic stress space.
If we leave aside the corrected stress space, i. e. the additional equations (6.26), the elastic
stress space alone, i. e. system (6.25) with input e, epl , yields the corrected plastic strain pl
as output, according to the ideas of [72, 73, 53].
6.1.2
107
The functions Les and Kes define the algebraic part of the correction model. The variables e,
e,
, , s,
, el , , e d , d , eel and epl depend explicitly on the independent variables pl ,
e1 , . . . , eI , eR,
,
1 , . . . ,
I , R.
Definition of function Les .
domain of definition is
, e,
, e d = as in (6.8),
S(
) = B(
)
and
d = e e d ,
= 1 + a eb R ,
= f (R)
s =
+
= D 1 s,
el = C 1
,
= el + pl .
s(t) e .
(6.29)
(6.30)
= C 1e,
e pl
= eel .
(6.31)
6.1.3
The parameters
Systems (6.2), (6.6) resp. (6.2), (6.7) depend on a lot of parameters. There are some elastic
parameters epl for the elastic stress space, and the elastoplastic parameters pl for the corrected
108
stress space, which are chosen equal to the ones for the transient elastoplastic boundary value
problem. For later use, we combine the parameters in two vectors
p = (p1 , . . . , pL )
= cR , , a , b , a , b , c
1 , . . . , cI , r1 , . . . , rI , Q1 , . . . , QI ,
(1)
(1)
(K)
(K) (1)
(K)
(1)
(K)
a1 , . . . , aI , . . . , a1 , . . . , aI , b1 , . . . , bI , . . . , b1 , . . . , bI ,
ep = (ep , . . . , ep )
1
L
e e
e
e
e
= ecR , e , ea , eb , ea , eb , ec
1 , . . . , cI , r 1 , . . . , r I , Q1 , . . . , QI ,
ea(1) , . . . , ea(1) , . . . , ea(K) , . . . , ea(K) , eb(1) , . . . , eb(1) , . . . , eb(K) , . . . , eb(K) .
1
1
1
1
I
I
I
I
(6.32)
Each elastic parameter has exactly one elastoplastic counterpart. They have to obey the
following restrictions
e (k) e
ai , a
b 0 eQi ;
R;
(k)
e
b < 0 < ebi , ecR , eri , ec
i ,
(6.33)
1 < ea < 0;
and
(k)
ai , a R;
(k)
(k)
b 0 Q i ;
(k)
(k)
I i := inf
1 < a < 0;
(k)
b < 0 < bi , cR , ri , c
i , .
(6.34)
K
X
k=1
(k)
e (k) ebi
ai e
> 1,
Ii := inf
K
X
(k)
(k)
ai ebi
k=1
> 1.
(6.35)
Y = e (1 + ea ) = (1 + a ) = Y ,
(6.37)
= limeR0 ef (eR) must hold for virgin material until the first active plastic
as limR0
f (R)
6.1.4
109
It makes sense, but is not necessary from the simultaneity idea, that in both stress spaces
the same constitutive model is used, here in our case Jiangs model with I backstresses. So,
each constitutive parameters pl , determining the movements in the corrected stress space, has
exactly one elastic counterpart epl in order to determine the movements in the elastic stress
space. Of course, the number I of backstresses has no influence on the whole simultaneity
idea as well.
But the movements of the ei and
i (and thus those of the total backstresses e and
) and
(and thus those of the yield surface
the development of the memory surface radii eR and R
radii e and due to lemma 6.3.1) differ in general, unless epl are pl equal. Note that, if all
ep = p , then
= e, i. e. the pointwise correction model is nothing but the identity. The
l
l
same we had for (EPK) material in chapters 3 and 5.1, see (3.10) and remark 5.1.3. For
illustration see (6.5), where I = 5 or (6.40).
ttgen,
Defining Jes , Les and Kes , we clearly pick up the pseudo stress approach from Ko
Barkey, Socie [72, 73] and Hertel [53]. The former developed this kind of scheme for the
z [88, 89], cf. as well Brokate, Dressler and Krejc
[16, 13]. In
constitutive model of Mro
contrast to [53, 72, 73], we prefer the terminology elastic instead of pseudo. But compared
to our formulation, there are some differences.
First, the integration of the differential equations for the corrected stress in the corrected
stress space, namely
H
= en : d
H
,
(6.38)
dt s = e dt es,
+ d
H
which was proposed in [72, 73], cannot be realised in a consistent manner with the fact that
Jiangs model comprises expansion/contraction of the yield surface. During integration of
(6.38), we have two yield surfaces, whose radii and e will differ in general. Thus we have
two yield conditions to be satisfied at the same time, which cannot be realized in a consistent
manner consistent with the simultaneity idea. This explains some numerical problems in [53].
stress
6
pl
e pl
strain
el
e el
(6.39)
In contrast to our case, a purely kinematic version of Mrozs model with identical yield radii
e =
is referred to in [72, 73]. Just compare our equivalent uniaxial case in (6.39) with the
corresponding in [72].
Second, an additional advantage of our formulation is that we get rid of the differential equation (6.38), so the number of equations to be integrated is reduced by three (free boundary
110
(t)
s(t)
pl
pl
Jes
elastic stress space
Jes
Jes
h (t)
Jes
Kes /Les
Kes
Kes /Les
pl
(6.40)
(t)
(t)
s
s
@
R
@
(t) = pl (t) + el (t)
Ke /Le
Kes /Les
D1
el (t) = C 1
(t)
Each path therein should be understood as has influence on. The next input/output-diagram
is commutative. Here a double arrow indicates the correction model and a single arrow
indicates the (stress/strain controlled) constitutive model.
e XXX
e
:
pl
ep
9
l
6 pl |pl
pl + e
XX
XX
z
z
X
= pl + el
X
y
:
X
yXX
pl
XX pl |epl ?
p
l
9
XX
d eR = d deR
and vice versa. The same applies to all equations in (6.23), (6.24), and trivially to (6.10).
So the pointwise correction model is rate-independent as well as the constitutive model is.
Originally in Jiang/Sehitoglu [64], the equations for plastic loading are parametrised as an
system of differential algebraic equations over the variable , but if we have > 0, cf. lemma
6.3.4, our formulation as an system over the fictive time t is possible and independent up to
time scaling. It functions as well, if > 0 is not satisfied.
6.2
111
The internal state variables in (1) for Jiangs constitutive model are given by pl , , 1 , ..., I
(and dependently = 1 + . . . + I ) and R (or equivalently = f (R) with a scalar function
f , which is one-to-one and onto, cf. lemma 6.3.1). This table gives an overview about the
internal stress-like state variables.
i
S ()
R
SR
The strain-like ones are are the plastic strain pl and the accumulated plastic strain .
Thus the constitutive Jiang hysteresis operator in (1) may be written as
(x, t) =
t
0
C ep (x, ), (x,
), pl (x, ), (x, ), i (x, ), R(x, ) (x,
) d.
(6.41)
For an explicit representation and derivation of the local elastoplasticity tensor C ep we refer
to (6.62).
For the (EPJ) constitutive model, we consider the general (non-plane) space V = R33
s
and its deviatoric projection Vd = R33
sd . We formulate the models for given s = dev()
AC([0, T ], Vd ) (explicit stress controlled case) resp. AC([0, T ], V ) (explicit strain controlled case).
6.2.1
i,
Since the right hand side functions Jpl , J, Jei and JeR in (6.25) do not depend on
i, R
you rediscover the original stress controlled Jiang constitutive model as a subsystem of system
(6.2). Just drop the additional equations (6.26), which build up the corrected stress space,
and e and the at each variable and each parameter. Then we receive the stress controlled
model from Jiang/Sehitoglu [63], ..., [67] in the form
1 , . . . , I , R = Js t, , 1 , . . . , I , R
pl , ,
on
(6.42)
I
X
i=1
i ,
d = s(t) ,
= f (R) = 1 + a eb R
(6.43)
SR = BR (0).
(6.44)
and
S () = B (),
112
pl = 0,
R = 0.
i = 0,
(6.45)
s n = s(t)
: n.
(6.46)
If s n 0 (neutral loading), let (5.11) hold again, else (plastic loading) set for i = 1, . . . , I
i = Qi 2 n : (i ) 1 + a eb R ,
K
X
(k)
(k)
1+
ai ebi if 0
k=1
ci = c
i
K
X
(k)
1+
ai
if < 0
(6.47)
(6.48)
k=1
and
d i = ci ri n
I
X
d =
ki k
ri
i +1
(i ) ,
(6.49)
d j ,
(6.50)
j=1
Further
cR 1 kk/R
+
d R =
() : d
kd k
if R > 0, BR (0)
if R > 0, BR (0)c .
if R = 0
d = f (R) d R = a b eb R d R,
H = n : d + d .
(6.51)
(6.52)
pl = n,
R = d R,
i = d i .
(6.53)
Here there exists just one stress space, in particular SR in (6.9) does not exist at all. The
conditions for the initial values i (0), R(0), pl (0), (0) are
ki (0)k < ri ,
and
I
X
j (0)
R(0),
j=1
pl (0) Vd ,
I
X
j (0)
1 + a eb R(0)
s(0)
j=1
(0) R
(6.54)
(6.55)
6.2.2
113
The strain controlled counterpart of (6.42) is never outlined at all in the works Jiang/Sehitoglu [63], ..., [67]. It can be written in the form
1 , . . . , I , R = J t, , , 1 , . . . , I , R
,
,
on
(6.56)
tr
s tr
n = s : n,
s tr = dev( tr ),
n = (s ),
s = dev(), (6.57)
cf. Simo [103], ch. 1, section 3. In the case of plastic loading, i. e. s S (), s tr
n > 0, the
inversion of the projection (6.53) and the additive decomposition of into el and pl can be
formulated as follows. We have
1
: n = : n,
I : n = 0,
(6.58)
s : n = tr()I
3
due to the orthogonality relation (2.1) and the fact that n is deviatoric. So after change to
matrix-vector notation
C 1 = pl
Consequently,
(6.53),(6.58)
: n
1 t
n=
n
n
.
H
H
(6.59)
1
1
1
=: C ep ,
n
C + n
H
(6.60)
(u, v, w RN )
: R33
R6 ,
s
and
C 1
1/E /E /E
0
0
0
0
0
0
/E 1/E /E
/E /E 1/E
0
0
0
0
0
0
(1 + )/E
0
0
0
0
0
0
(1 + )/E
0
0
0
(1 + )/E
0
0
0
(6.61)
and
n 0, there simply hold
(6.45).
114
6.2.1 Lemma. For each H 6= 2, where from (10), the inverse in (6.60) exists and is explicitly
given by
2
ep
n
n
.
(6.62)
C = C I
2 + H
In particular, since > 0, the inverse always exists, if H > 0.
Proof: To verify this, we need three facts.
= 2
(i) As n R33
n.
d , one can show C n
(ii) As n has unit length, we have (
nn
)2 = n
(
nt n
) n
t = n
knk2 n
t = n
n
t = n
n
.
(iii) C is in R66
s .
So
1
I + (
=
nn
) C C 1
H
1 (i)
(iii)
1
2
I+
=
n
Cn
C
= I+
n
n
C 1 ,
H
H
1
n
C 1 + n
H
thus
2
1
1
2
1
n
n
C C I
n
n
= I + 2
n
n
I+
H
2 + H
H
2 + H
(ii)
(2)2
(
nn
)2 = I,
H(2 + H)
replace the first two equations in (6.53). The conditions for the initial values (0), i (0),
R(0), (0) are
I
X
(0) R,
(6.63)
j (0)
R(0),
ki (0)k < ri ,
j=1
and
I
X
j (0)
1 + a eb R(0) .
s(0)
(6.64)
j=1
Of course, a sufficient condition for H > 2 such that the model makes sense, cf. lemma
6.2.1 is by virtue of the estimate (6.72) for H given by
a b cR 2.
(6.65)
But it is important to realise that the strain controlled model is defined as well for negative
H, i. e. where the uniaxial equivalent stress-strain curve has a critical point, see (6.66).
115
H H > 0
H=0
@
@
@
@
H 2 @@ H < 0
(6.66)
Note that, whereas in the (elastic) stress controlled cases (6.2) resp. (6.42) the equation for
pl is just an integration, the right hand side of (6.56) does depend on , as the yield condition
depends on dev(). For the explicit strain controlled Jiangs model, we find from (6.52) that
H = H(t, , 1 , . . . , I , R).
6.2.2 Remark. (a) Alternatively for the boundary value problem and corresponding finite element
computations, where not an explicit strain path t 7 (t) is given, the strain controlled model
(6.56) can be written
1 , . . . , I , R = J , ,
pl , ,
pl , , 1 , . . . , I , R
so that we have H and C ep as functions of , ,
pl , , 1 , . . . , I , R.
(k)
6.3
The explicit time-dependent (elastic) stress- resp. strain controlled functions Jes , Js , J in
(6.2), (6.42), (6.56) are discontinuous. So unfortunately, the standard theory of systems of
ordinary differential equations, where usually at least local partial Lipschitz continuity in the
space variable is required, does not apply.
For other elastoplastic constitutive laws e. g. the models of Melan and Prager [86, 95,
96] (linear kinematic hardening), Armstrong and Frederick [2] or Chaboche [20, 21]
(nonlinear kinematic hardening) results of wellposedness are already available. The reader
[9, 10, 11, 12, 14, 17]. Mathematically, it is
is referred to the works of Brokate and Krejc
possible to transform those models to operator equations, containing the stop and the play
operator, cf. chapter 1. Such a transformation of Jiangs differential equations is not easily
possible. Jiangs model additionally comprises nonlinear isotropic hardening/softening. The
yield surface the convex set in Brokate and Krejcs theory may contract and expand,
depending of the position of the total backstress and the memory surface.
So in this section 6.3.2, we can just give some hints of local existence and uniqueness by
considering the fluxes if the solution reaches a surface of discontinuity, or an intersection of
several such surfaces.
Further in this section 6.3, we discuss some sub-parts of Jiangs model, which are helpful to
understand its nonlinear behaviour. We formulate the ideas for the Jiang correction model,
always having in mind that Jiangs constitutive model is a subsystem, cf. (6.23) and (6.24).
Some of these reflections can be carried over to the Doring model, cf. appendix B.3.
116
6.3.1
This section gives quite a couple of lemmas that help to understand how Jiangs model and
the pointwise correction model behave. We hope that these insights will provide useful tools
for future research, especially for a final proof of the wellposedness of Jiangs constitutive
model both in isolation and in coupling with Newtons balance equations.
Figure D. Left hand side: The graphs of ef and f . Right hand side: The correction region {(eR, R)
e e
e
e
6.3.1 Lemma (Coupling of radii) The yield and memory surface radii are coupled in a one-toone manner. The mappings
ef
[0, ) eR 7 e [e Y , e ) in (6.8),
7 [Y , ) in (6.29),
[0, ) R
eR0
e Y := e (1 + ea ),
and
e
eR
e ,
R0
Y := (1 + a )
(6.67)
(6.68)
(6.69)
Especially,
0 < e Y e < e < ,
The inverse mappings are given by
e e
e 1
/ 1
1
e f
,
7 e ln
ea
b
1
ln
7
b
f 1
/ 1
.
a
(6.70)
The same holds of course for = f (R) of the Jiang constitutive model without e s and s.
e and are the initial yield stresses for virgin material in both stress spaces.
Y
Y
117
(R) = a b eb R ,
dR
def e
e e
( R) = e ea eb e b R
deR
(0) =
1 (0) + . . . +
I (0) Vd and positive initial radii R(0),
e
(0) =
(0)
where
(0) = ef 1 eR(0) ,
Set
and
(0) = (0),
(0) = f 1 R(0)
f, ef from (6.70).
(0)
= B
e(0) e(0) = B
(0) .
B
s[0, ] B,
[0, ] e [0, ] ,
d [0, ] e d [0, ] ,
[0, ] e[0, ] e(0).
Thus on the interval [0, ], the correction model reduces to the identity.
Proof: Clear with simultaneity from (6.1), (6.10), (6.29) and (6.30).
The next lemma is concerned with the moving directions of the partial (and the total) backstresses.
6.3.3 Lemma (Backstress movements) For the pointwise correction model for Jiang material,
the following assertions hold in the elastic stress space.
(a) If kei k er i for any 1 i I, then dei : en 0.
(b) If kei k < er i for any 1 i I, then dei : en > 0.
(c) If kei k er i for all i = 1, . . . , I, then de : en 0.
118
(d) If kei k eri for all i = 1, . . . , I and at least one of the inequalities is strict, then
de : en > 0.
(e) If kei k = er i for any 1 i I, then dei : ei 0.
The analogous assertions hold for d
and en in the corrected stress space and for d and n
in the elastoplastic stress space, when considering the constitutive Jiang model.
Proof: We simply apply Cauchy-Schwarz inequality. Due to (6.13), (6.33) and (6.35), there
holds ei , eci 0, er i > 0. We have
!
e ei +1
k
k
(6.14)
i
e
n : (ei ) ,
dei : en = eci er i 1
er
i
from which
e e
ci r i
kei k
er
i
e ei +1 !
ei +1 ! (CS)
(CS)
k i k
e
e
e e
d i : n ci r i 1 +
.
er
i
The left hand estimate especially yields (a) and (b). After summing up over i = 1, . . . , I, we
find
e ei +1 !
e ei +1 !
I
I
X
X
k i k
k
k
i
e e
e e
ci ri 1 +
de : en
,
ci r i 1
er
er
i
i
i=1
i=1
the left hand estimate giving (c) and (d). Similarly from
e
(6.14) e
dei
: i =
kei k
er
i
ei
e e
ri n
ci
: i
kei k
er
i
ei
k i k
there follows
eci kei k
ri +
kei k
(CS)
(CS)
ri
kei k
er
i
ei
kei k ,
the right hand estimate being positive, if kei k < er i , and equal to zero, if kei k = er i . This
especially yields (e).
During elastic/neutral loading, dt = kdt pl k =
We now have a look at the internal variable .
0 according to (6.10). During plastic loading, we conclude from the last equation in (6.23)
the next lemma will give a sufficient condition for eH > eH min > 0 that 0 < dt = kdt pl k
(dt )max . Therefore altogether dt = kdt pl k 0 and
(0)
=
(t)
t
0
) d =
dt (
kdt pl ( )k d
R is
is the total (corrected) accumulated plastic strain. The function [0, T ] t 7 (t)
then constant during elastic/neutral loading and strictly increasing during plastic loading, in
0.
particular non-negative, if (0)
119
6.3.4 Lemma. Assume for the Jiang correction model, that the elastic stress deviator es is of class
1 . Then if
Cpw
kei k eri for i = 1, . . . , I
and
kek eR,
(6.71)
the following assertions hold for the movements in the elastic stress space.
(a) There holds
e
H e ea eb ecR .
(6.72)
max < , is
(b) A sufficient condition for 0 < eH min eH, or equivalently 0 < dt (dt )
given by
I
X
e
cj er j kei k e ea eb ecR +
(6.73)
j=1
(6.14),(6.15),(6.22)
I
X
e e
cj rj
j=1
(CS),(6.74)
I
X
j=1
e e
cj rj
kej k
er
j
kei k
1 e
ri
(6.71) e
cR .
ej +1
n : (ej )
(6.74)
+ e ea eb e b R deR
e ea eb ecR ,
from which (6.72) and (6.73) follow. Note here that ei 0 > ea , eb . As es is bounded over
any interval [tn1 , tn ], where es C 1 and es C 0 , it is in fact globally bounded on [0, T ], thus
max < .
dt (dt )
The norm kei k, in the worst case each, may asymptotically reach eri and the first summand
in (6.73) may become too small. This usually occurs, if the inputs es(t) are too large. Anyway
the model (6.2) is just able to get along with inputs that are small enough, cf. remark 6.3.10.
So far, we have not been able to prove a sufficient a-priori-estimate depending on the Sobolev did in [9, 12, 14] for other constitutive models
norm (1.5) of the input like Brokate/Krejc
of elastoplasticity.
A way to find a remedy in practice may be to use more backstresses. Unfortunately, a larger
I yields more differential equations and more parameters.
The next lemma proves that the ei resp.
i remain in their limiting radii er i resp. ri , as
already stated in Jiang/Sehitoglu [65] in the case of proportional monotonous loading.
But was not proven, least of all for general loading input.
120
6.3.5 Lemma (Limiting radii) Consider the differential equations (6.12), ..., (6.14) for the partial
backstresses, and let a function
e
1
n Cpw
[0 , T ], Vd B1 (0)
(0 0 < T < )
(6.75)
be given as driving input for them. Then here holds:
1 [ , ], V .
(a) Each ei is in Cpw
0 T
d
interval [0 , T ].
(c) Necessarily for the total backstress, there holds in the whole interval [0 , T ]
I
e
X
e
()
ri .
(6.76)
i=1
k=1
where eI i denotes the infimum in (6.35). Let us first consider the interval [0 , 1 ], where en is
C 1 . Here we have the polar decomposition
e
i = kei k (ei ).
(6.14)
dei = eci eri
kei k
er
i
ei +1
(ei ) .
(6.78)
1 d
k(ei )k2 = (ei ) : d(ei ),
2 d
C i ri 1
dk i k = ci ri n : ( i )
er
er
i
i
As eb 0 according to (6.33), we have
1 ei + 1
(CS),(6.12)
1 + 3 eQi 1 + |ea | =: eKi .
With the substitution e i = kei k/er i , we obtain the scalar radial differential inequality
e
de i eC i 1 e iKi .
121
(i) In the case eKi > 1, i. e. eQi > 0, separating the variables leads us (geometric series
expansion, 0 e 0i < 1 according to our premise) to
e
e
e i
X
Ki +1 i
fi ()e 0 :=
eK + 1 0
i
e
i
=0
i
Z e i X
Z e i
Z
d
eK
e
i
=
d =
C i d ,
eK
i
e 0
e 0 1
0
i
i
=0
where the function fi : [0, 1) [0, ) at the left hand side is strictly increasing, thus
bijective, s. t. fi (0) = 0 and lim1 fi () = .
(ii) In the case eKi = 1, i. e. eQi = 0, the same technique leads us to the function
fi : [0, 1) [0, ), 7 ln(1 ),
which has the same properties.
In any case, as 1 , we find fi (e 1i ) fi1 fi (e 0i ) + eC i (1 0 ) [0, 1). The rest of (a)
and (b) is induction.
(c) Triangle inequality on (6.8). The rest: Discard all the s and e s, except for the one at
the normal en.
The following lemma comprises an interesting property of the correction model as a consequence of the simultaneity.
6.3.6 Lemma (Subspace lemma) Let V be any linear subspace of R33
s . For the Jiang correction
model, the following assertions hold.
(a) At any point t [0, T ] there holds
e
s, e1 , . . . , eI ,
1 , . . . ,
I V e 1 , . . . , e I , dt
1 , . . . , dt
I , dt pl V,
then
1 (0), . . . , eI (0),
1 (0), . . . ,
I (0) V
and
e
s(t) : t [0, T ] V
e
i (t) : t [0, T ] ,
i (t) : t [0, t]
V
for i = 1, . . . , I,
s(t) : t [0, T ]
V,
pl
(t) : t [0, T ]
pl (0) + V.
Analogous assertions holds for the stress controlled constitutive Jiang model.
Proof: (a) We have directly by definition of Jes and Les
e
s, e1 , . . . , eI ,
1 , . . . ,
I V
(6.8)
(6.11),(6.29)
(6.14),(6.19)
de1 , . . . , deI , d
1 , . . . , d
I V
(6.23),(6.24)
,
, e d V
n, d , s V
e
=
1 , . . . , e I , dt
1 , . . . , dt
I , dt pl V.
(b) Clearly es(t) : t V implies es(t)
: t V , with (a) proving the proposition.
122
6.3.2
Flux computations
As already announced, we now start computing the fluxes at the discontinuity surfaces. It is
a pity, but the theory of Filippov [42] is not applicable, as the position of the discontinuities
depends on the actual solution. But the fluxes at the touching point make local wellposedness,
let us say, probable.
6.3.7 Remark (Embeddings) Let the dot denote the standard scalar product on the spaces R4
resp. R6 . The linear embedding
4
A 7 (a11 , a22 , a11 a22 , 2a12 )t
: R33
(6.79)
spd , : R , ,
is an isometry, since
= A B
A : B = At M B
where the matrix
for A, B R33
spd ,
2 1 0
M = 1 2 0
0 0 2
(6.80)
(6.81)
is symmetric and positively definite, its eigenvalues being equal to 1, 2 and 3. Similarly, the
map
6
(6.82)
A 7 (a11 , a22 , a11 a22 , 2a12 , 2a23 , 2a13 )t
: R33
sd , : R , ,
We consider first the correction model, knowing that the stress controlled Jiang model is just
a submodel.
For our theoretical purposes, we rewrite (6.2) equivalently in the redundant form
pl , ,
ec
c
c
y = Jes(t, y),
y = c
1 , . . . , ec
I , eR,
1, . . . ,
I , R
(6.83)
with the embedding from (6.79) in the free boundary case V = R33
sp resp. from (6.82) in
.
Now
(6.83)
is
equivalent
to
the general case V = R33
s
x = Ies(x) = 1, Jes(x) ,
x = (t, y).
(6.84)
Through the introduction of time t as an artificial space variable x0 , with x 0 = 1 and initial
condition x0 (0) = 0 thus x0 (t) = t, the system (6.84) is autonomous and the time discontinuity is pushed into space.
System (6.83) consists of
4 (2I + 1) + 3 equations, of which 3 (2I + 1) + 3 are independent (in the free boundary
case V = R33
sp ).
6 (2I + 1) + 3 equations, of which 5 (2I + 1) + 3 are independent (in the general case
V = R33
s ).
123
System (6.84) additionally includes the dummy equation for x0 = t. As the image of is a
linear subspace, it is clear that the solution will lie here, if the initial values do so, cf. lemma
6.3.6 as well. From now on, we will drop the hat for this section.
We consider the autonomous system (6.84) in the form of a differential inclusion
n
o
k
k
x Ies (x) = conv v : (xk )kN D s. t. xk x and Ies (xk ) v ,
(6.85)
Of course, if Ies is continuous in x, we have Ies (x) = Ies (x) . Se set of discontinuity points
D with the
of Ies has measure zero. More precisely speaking, it is given by SeD (e) SeDR SR
smooth submanifolds (cylinders)
D
SeD (e) = x : e (x) = 0 ,
SeDR = x : eR (x) = 0 ,
SR
= x : R (x) = 0 , (6.86)
where
I
I
I
X
X
X
e
e
e
e
e
e
(6.87)
j R.
e (x) = s(t)
j R, R (x) =
j f ( R), eR (x) =
j=1
j=1
j=1
= x : R < 0 ,
DR
+
= x : R > 0 ,
DR
(6.88)
(6.89)
where Ies is single-valued and Lipschitz and the solution x therefore will be smooth. All of
the domains (6.88) and (6.89) and all their mutual intersections
+
De+
eR = De DeR , . . . ,
+
+
De++
eRR = De DeR DR , . . .
are locally connected, as a consequence they satisfy the so-called -condition in [42], 6, which
means
(De ) t = const = (De t = const ) for almost all t, . . .
where the boundary at the right hand side is to be taken in the relative topology of the
hyperplane {t = const}.
For the computation of fluxes, we use the brackets h, i for the Euklidian scalar product on
Rr(2I+1)+4 and
= x = t , y = t , pl
I , R
1 , . . . ,
, , e1 , . . . , eI , eR ,
124
denotes an arbitrary scalar, an arbitrary element of Rr , 0 the zero in Rr . For any manifold
S and x S, Tx S denotes the tangential space of S at x. We let and | | denote the standard
Euklidian scalar product/norm on Rr .
In the sequel, we assume that es n = en es > 0, where the discontinuities at the surfaces
D occur. In the case es 0, due to (6.10), the whole movements, i. e. the
SeD (e), SeDR and SR
n
e1 , ..., eI , eR,
remain constant.
functions pl , ,
1 , ...,
I and R,
(i) Active loading of SeD (e). We consider the case, where the solution reaches a point x
S DeR , where S = SeD (e) = De = De+ . Here we have Ies = conv I , I + with functions
I (x) =
lim
De x x
Ies (x ) = (1, 0, 0, 0, . . . , 0, 0, 0, . . . , 0, 0)
and
I + (x) =
=
lim
De+ x x
Ies (x )
e
1, , , c1
e
|e1 | 1 e
...,
1 dt ,
r1 n e
r1
!
e eI
e|
|
|
|
I
e
e
, . . . , , ,
ecR 1
I dt ,
cI erI en e
dt ,
eR
rI
e and D
e+ respectively. To have a unique element I 0 = I +
which are smooth up to D
+
+
I Ies Tx S, it is necessary and sufficient that the system of linear equations
1
1
=
(6.90)
+
e , I
e , I
0
has one and only one solution in [0, 1]2 . Vector calculus leads us to
I
X
e
= es n (t, e1 , . . . , eI ),
j = en(t, e1 , . . . , eI ) es(t)
t e = t es(t)
j=1
I
X
e
(i = 1, . . . , I)
j = en(t, e1 , . . . , eI )
ei e = ei es(t)
j=1
and
pl e = 0,
e = 0,
eR e = e ea eb e b R ,
The projections of the function value limits on the surface normal are
e , I = es n > 0
(6.91)
125
and
e , I +
ej
eb eR
|e|
e e
e
cR 1 e
j dt + a b e
dt
n cj rj n
=
s n
R
j=1
)
(
e ej
I
X
|
|
e
e
(6.16)
j
e
e
e
j + e ea eb e b R deR dt
=
s n en
cj erj en e
rj
j=1
n
o
(6.14),(6.15),(6.22) e
=
s n en d e + d e dt
(6.22),(6.15)
(6.23)
I
X
|ej |
er
j
s n eH dt
0.
Hence = 0, + = 1 and Ies (x) Tx S = I + (x) We see that the solution curves reach S
and a sliding motion occurs as long as es n > 0. The solution x cannot enter De+ . As I + and
S are smooth, x should be smooth during the sliding motion along surface S.
The case, where x additionally meets SeDR , is considered in (ii). The cases, where the solution
D D , cannot occur.
x meets a point on SeDR De or SR
e
(ii) Active loading of SeD (e) and SeDR . Here the surface of discontinuity is given by S =
SeD (e) SeDR , where the normals
e = same as in (6.91),
eR =
0, 0, 0, (e), . . . , (e), 1, 0, . . . , 0, 0
are linearly independent (obviously due to the separation into the elastic and corrected stress
space). So SeD (e) and SeDR locally divide D at x S into four quadrants, so we have four
limiting values of Ies
+
(x) =
I
i. e.
lim
x x
De+
eR
Ies (x ),
+
(x) =
I+
lim
De++
x x
eR
Ies (x ),
...
+ +
, I , I+ , I+
Ies = conv I
(x)
I+
1, , , c1
e
|e1 | 1 e
...,
1 dt ,
r1 n e
r1
!
e eI
|
|
I
e
e
0, , . . . , ,
I dt ,
cI er I en e
rI
126
and
+
(x)
I+
1, , , c1
e
|e1 | 1 e
...,
1 dt ,
r1 n e
r1
!
e eI
+
|
|
I
e
e
(e) d e dt ,
, . . . , ,
I dt ,
cI erI en e
rI
e
are Lipschitz in D
eR , etc. There exists one unique element
+ +
+ +
I 0 =
I + I + + I+ + + I+ Ies Tx S,
1
1
1
1
1
+
e , I
e , I
e , I
e , I
=
0
+
+
0
eR , I
eR , I
eR , I+
eR , I+
+
+
(6.92)
has one unique solution in [0, 1]4 . Similarly as in (i), one shows
+
e , I+
= e , I+
= 0,
e , I
= e , I
= s n > 0.
Obviously
Straightforwardly, we calculate
eR , I+
I
X
j=1
(6.14),(6.15)
Further,
+
eR , I
= eR , I
= 0.
(e) ecj erj
|ej |
er
j
ej +1
(ej ) dt
dt (e) de.
+
+
eR , I+
= dt (e) de e/eR de
= 0,
since |e| = eR, (e) = e/eR on SeDR , if the loading is active. This yields the unique solution
e
e
e
e
e e e
+
+ = 1, = = + = 0 of (6.92), since dt = s n / H, s n > 0 and H = n : d +d > 0.
e
e
e
e
The latter because n : d > 0 according to lemma 6.3.5 and d , d R 0 according to
+
(x) . Here as well a sliding motion
lemma 6.3.1 and (6.16). Consequently Ies (x) Tx S = I+
+
is Lipschitz, x
occurs, the solution can neither enter De+ nor De+R . As S is smooth and I+
will be smooth during the sliding motion along S.
D . Here we have S = S D S D . Just erase appropriate e s in
(iii) Active loading of SeD and SR
e
R
paragraph (ii) above.
D . Just apply the same techniques as above to S =
(iv) Active loading of SeD , SeDR and SR
D
D
D
Se SeR SR . Now D is locally divided into octants, the solution x can neither enter De+
+
. No additional ideas/tricks are needed to verify this.
nor De+R nor DR
127
6.3.8 Remark. Note that the first equation in (6.90) and (6.92) reflects the fact that I 0 is a
convex combination, whereas the last (two) equation(s) in (6.90) resp. (6.92) express the
tangentiality to S.
The results for the original stress controlled constitutive Jiang model can be easily derived, as
it is a submodel. Drop each e andand the additional equations (6.26). So in (6.87) consider
D instead of S D (e) resp. S D , but there is no third
the discontinuity surfaces SD () resp. SR
e
eR
discontinuity surface.
For the strain controlled constitutive Jiang model (6.56) the computations are similarly. We
consider the differential inclusion y J(t, y), generalising
y = J(t, y),
y = (
, ,
1 , . . . ,
I , R)
with
D = Rr R RrI (0, ) Rr(I+1)+2 ,
r = 6.
We drop again the hat . The main difference compared to the (elastic) stress controlled
case lies in the fact, that all discontinuities are in the space variable y, not in time t. The
discontinuity surfaces are given by
where
SeD (e) = y D : (y) = 0 ,
D
SR
= y D : R (y) = 0 ,
I
X
j 1 + a eb R ,
(y) = dev()
j=1
I
X
j R.
R (y) =
j=1
With
= y = ( , , 1 , . . . , I , R )
the respective normals are
=
R =
Note here that
n, 0, n, . . . , n, a b eb R ,
0, 0, (), . . . , (), 1 .
= s .
J ,
= C (t)
: n = tr : n = ntr > 0,
+
J , = : n H = 0,
...
128
Nevertheless, we cannot help making the following reasonable assumptions concerning Jiangs
constitutive model, both in isolation and in coupling with the Newtonian balance equations.
6.3.9 Assumption (Jiangs model) For Jiangs model, we assume the following.
(a) Elastic stress controlled Jiang correction model. Let V = R33
or V = R33
s
sp .
e
pl
For a given set of parameters pl , pl , satisfying (6.33), ..., (6.35), and initial values (0),
e
an appropriate subset AC([0, T ], Vd ), depending on those parameters and initial values,
such that for all elastic stress deviator inputs
e
s e
AC([0, T ], Vd )
there exists one uniquely determined solution
e1 , . . . , eI , eR,
AC [0, T ], Vd R (V I [0, ))2 ,
(
pl , ,
1 , . . . ,
I , R)
d
of (6.2) under the initial values/conditions (6.27), (6.28). The cited conditions remain valid
for each t [0, T ].
(b) Stress controlled Jiang constitutive model. Let V = R33
s . For a given set of
parameters pl , satisfying (6.34), (6.35), and initial values pl (0), (0), 1 (0), . . . , I (0), R(0),
satisfying (6.54), (6.55), there exists an appropriate subset AC([0, T ], Vd ), depending on
those parameters and initial values, such that for all stress deviator inputs
s
AC([0, T ], Vd )
there exists one uniquely determined solution
(pl , , 1 , . . . , I , R) AC [0, T ], Vd R VdI [0, )
of (6.42) under the initial values/conditions (6.54), (6.55). The cited conditions remain valid
for each t [0, T ].
(c) Strain controlled Jiang constitutive model. Let V = R33
s . For a given set of
pl
parameters pl , satisfying (6.34), (6.35), (6.65) and initial values (0), (0), 1 (0), . . . , I (0),
R(0), satisfying (6.63), (6.64), and all strain inputs
AC([0, T ], V )
there exists one uniquely determined solution
(, , 1 , . . . , I , R) AC [0, T ], V R VdI [0, )
129
of (6.56) under the initial values/conditions (6.63), (6.64). The cited conditions remain valid
for each t [0, T ].
(d) Boundary value problem with Jiangs constitutive model. We assume that the
transient elastoplastic boundary value problem (5), ..., (7) with Jiangs constitutive operator
and given parameters pl , satisfying (6.34), (6.35), has got a unique solution, such that for a. e.
point x there holds (1), (6.62) where H from (6.52) and n from (6.46).
pl , e1 , ..., eI eR,
are of class AC, as well
6.3.10 Remark. (a) If the functions ,
1 , ...,
I and R
el
e
e
e
e
el
e
pl
the functions
, , , , d , ,
, d , , and will be. This is clear by the definition
of Les resp. Kes , if
h AC.
(b) If assumption 6.3.9 is satisfied, epl = pl for all l = 1, . . . , L and the initial values are
= eR, = e.
consistently chosen, then
= e,
i = ei , d = e d , R
(c) As the triangle inequality shows, there necessarily holds at any time, since we have es =
e + e ,
d
e
(t)
e
The analogous estimates
(6.69)
e
(t)
e(t)
I
(6.69),(6.76) X
e
e
s(t)
e(t)
+ e(t)
r i + e
i=1
I
X
ri + ,
(t)
s(t)
i=1
I
X
(t)
s(t)
ri + ,
i=1
hold for the corrected stress and the elastoplastic stress deviator. From this, we see that we
must necessarily have at least
I
X
e
e
e
e
e
r i + in [0, T ]
s AC([0, T ], Vd ) : k s(t)k
i=1
resp.
I
X
ri + in [0, T ] ,
s AC([0, T ], Vd ) : ks(t)k
i=1
6.3.3
A special case
If for any i {1, . . . , I} the ratchetting parameter eQi is equal to zero, i. e. the ratchetting
exponent e + 1 is constantly equal to one, the differential equation for the partial backstress
e simplifies to
i
(6.14)
(6.23)
e
i = eci eri en ei dt = eci eri dt pl dt ei ,
(6.93)
the right hand side being a generalisation of the backstress evolution equations of Arm [14], since
strong/Frederick [2] resp. Chaboche [20, 21], see as well Brokate/Krejc
ec = ec ()
is non-constant. If we replace
i
i
es e
,
dt pl = dt e
(6.94)
130
e
s(t) e(t) e
e
ri
dt (t)
i = eci (t)
i ,
e(t)
(6.95)
a system of ordinary differential equations of first order. If the functions e, es, e, are of
class AC and satisfy
e
s e
e e. in [0, T ],
dt 0 a. e. in [0, T ],
(6.96)
we can solve (6.95) by the variation of constants method, yielding the explicit expression
Z te
1
s( ) e( ) e
e
e
e
i (t) = e
i (0) + ri
Wi ( ) d
(6.97)
e( )
Wi (t)
0
for the ith partial backstress. Here
Z
Z t
e
e
ci ( ) dt ( ) d = exp
Wi (t) = exp
0
and
exp
(6.13)
ec
i
ci (p) dp = e
QK
k=1
QK
k=1
ee
(k)
eb
0
i
ee
(t)
(0)
ci p dp
e (k) e (k)
ec
i ai / bi
(6.98)
With the aid of (6.97), we are now able to give an alternative proof of lemma 6.3.5.
6.3.11 Lemma. There holds kei (t)k er i everywhere in the interval [0, T ], if the initial value ei (0)
satisfies kei (0)k er i .
Proof: eWi in (6.98) is obviously a nondecreasing function in t, satisfying eWi (0) = 1. Thus
we may estimate
e
e
Wi (t)
ei (t)
Z t
e
(6.97)
s e
e
e
ri
e
Wi ( ) d + 1
0
Z t
(6.96)
e
e
Wi ( ) d + 1
ri
=
0
e e
r i Wi (t),
An analogous equation to (6.95) and solution formula (6.97) holds of course in the corrected
stress space. (6.94) may as well be written as
s
dt pl = dt
because of simultaneity (6.3), (6.5) and the fact that is common for both the elastic and
the corrected stress space.
6.3.4
131
Phases of relaxation
In this very short section, we have a closer look at the effect of relaxation of the memory
surface, i. e. where the input is such that 0 < emin kek < eR over a long -interval.
e
Separating the variables in the differential equation for R leads us to
Z
Z
Z
deR
deR
(6.16)
e
cR d =
e
e
kek/eR 1
min / R 1
This gives the supersolution
cR ( 0 ) emin log(eR emin ) eR,
for .
This behaviour traps e between emin and eR with increasing plastic flow, as long no active
loading of the memory surface for kek = eR takes place. Just have a look at figure J on page
150, where relaxation of the memory surface is illustrated.
in the corrected stress space and , R, in the elastoThe same holds of course for
, R,
plastic stress space of the Jiang constitutive model.
6.4
We consider here in section 6.4 the plane free boundary case V = R33
sp and 1 \ 1,g .
As explained in the introduction this case is the most relevant in practical fatigue analysis.
Direction 1 and 2 are assumed to be in T , i. e. tangential to the boundary, direction 3 in
(T ) , i. e. collinear with the outer normal at .
By considering circumferential sharply notched axles, Seeger et al. [59, 60, 61] found that
that the ratios e11 /e22 of the elastic tangential strains and 11 /22 of the elastoplastic tangential strains almost coincide, i. e.
e (t)
11 (t)
11
,
e
22 (t)
22 (t)
(6.99)
if straining disability is present. (Here the dependence on has been dropped again.) This is
a purely empirical result, but a very fruitful achievement of lifelong experience. So it suggests
itself to impose the constraint
e (t)
11 (t)
11
,
(6.100)
=e
22 (t)
22 (t)
on the corrected tangential strain components 11 and 22 in such a case. Here the right hand
side is known from the elastic boundary value problem.
In a more general setting we are motivated to modify the correction model by the introduction
of one of the following algebraic constraints (which may be of any empirical nature). Let us
assume that external scalar functions t 7 f1 (t), t 7 f2 (t) of class AC are given.
132
(6.102)
(S2) Or we impose
(2)
p 22 = f1 (t) 11 f2 (t) 22 = 0, (6.103)
Se11 ,e12 t, pl , e1 , . . . , eI , eR,
1 , . . . ,
I , R,
(6.104)
In the case (S1) t 7 e 22 (t) and t 7 e 12 (t), in the case (S2) t 7 e 11 (t) and t 7 e 12 (t)
are the given elastic stress components of class AC. We will call the functions p 11 (which
replaces e 11 in the case (S1)) resp. p 22 (which replaces e 22 in the case (S2)) the pseudo
elastic stress 11- resp. 22-component. We dont know any physical meaning so far.
6.4.1 Remark (Seegers constraint) (a) Seegers empirical constraint (6.100) for the corrected
strains, i. e. 11 e22 = 22 e11 , is revealed by the choice
f1 (t) = e22 (t),
(6.105)
where
e
11 (t) =
1 e
11 (t) e 22 (t) ,
E
22 (t) =
1 e
22 (t) e 11 (t)
E
(6.106)
from Hookes law e(t) = C 1e(t) for plane stress. In section 8.3, we will see that the
introduction of Seegers constraint will lead to significant improvements of the correction
model results.
(b) In the case of a linear superposition of combined tension force and torsion moment, we
are able to lift the zeros of e22 , so that
k=
e (t)
11
e (t)
22
e 22 (t)
= const
e (t) e (t)
22
11
11 (t)
(6.107)
is in fact independent of t. This will be done in detail in section 8.3. Therefore the constant
choice
f1 (t) = 1,
f2 (t) = k,
(6.108)
is possible instead in (6.105), so that due to lemma 6.4.2 there are no troubles about nonuniqueness, if k 6= .
We formulate a modified system of equations, compared to section 6.1. In addition to the
ei , eR,
we are looking for the new unknown input function
preexisting unknowns pl , ,
i , R,
p
p
11 = f2 (t)
22 is satisfied.
11 resp. 22 , such that the constraint f1 (t)
133
p
e
e (t) 0
11
12 (t) 0
11
12
e
e
= e 12 (t) e 22 (t) 0 .
= e 12 (t) e 22 (t) 0 ,
0
0
0
0
0
0
(6.109)
11 (t) e 12 (t) 0
11 (t) e 12 (t) 0
e
e
0 ,
0 .
= e 12 (t) p 22
= e 12 (t) p 22
0
0
0
0
0
0
(6.110)
Then
e
s = D e,
s = D e
(6.111)
e
e(t, p
is again the (pseudo) elastic stress deviator. We now have the dependence =
11 ) in
e
e
p
e
e
the case (S1) and = (t, 22 ) in the case (S2), whereas = (t) in (6.2), (6.6) before.
(1)
V = R33
sp .
We set
e
, e,
, e d = as in (6.8),
d , , s = as in (6.29),
, el , = as in (6.30)
where e from (6.109) and (6.111). This defines (6.102) and finally (6.101).
(2)
Definition of function Se11 ,e12 . Completely symmetrically with e from (6.110) and
(6.111), defining (6.104) and (6.103).
All other variables are defined formally completely analogously as in Jes in (6.2) and in Les
in (6.6), cf. section 6.1. (We do not repeat all those definitions for the sake of brevity.) We
may now write everything together into a system of differential algebraic equations
ei , eR,
e e i , R,
= Je(1) e
p 11 , p 11
i , R
dt pl , ,
22 , 12 t, , i , R,
(1)
p 11
(6.112)
, , eel , epl , s, e, e d , e, ,
d , = Le22 ,e12 t, pl , ei , eR,
i , R,
(1)
pl
e
e
p
11
0 = Se22 ,e12 t, , i , R,
i , R,
in the case (S1), resp.
ei , eR,
e e i , R,
= Je(2) e
p 22 , p 22
i , R
dt pl , ,
11 , 12 t, , i , R,
(2)
p 22
, , eel , epl , s, e, e d , e, ,
d , = Le11 ,e12 t, pl , ei , eR,
i , R,
(2)
p 22
0 = Se11 ,e12 t, pl , ei , eR,
i , R,
(1)
(2)
(6.113)
in the case (S2). The functions Je22 ,e12 resp. Je11 ,e12 have the same discontinuities as
Jes , see section 6.3. It is not possible any more to write the differential equations in explicit
(1)
(1)
(2)
(2)
form like in (6.2). Je22 ,e12 plus Le22 ,e12 resp. Je11 ,e12 plus Le11 ,e12 contain exactly as
many equations as Jes plus Les do, but there is one additional unknown more, this lack is
compensated by the scalar algebraic constraint (6.101) resp. (6.103).
134
and pl , the
6.4.2 Lemma. For each admissible state of variables t, e1 , ..., eI ,
1 , ...,
I , eR, R
following assertions hold.
(S1) Constraint (6.101), (6.102) has got a unique solution for p 11 , if the prescribed functions
satisfy
f1 (t) 6= f2 (t).
(6.114)
(S2) Constraint (6.103), (6.104) has got a unique solution for p 22 , if the prescribed functions
satisfy
f2 (t) 6= f1 (t).
(6.115)
By admissible, we mean that conditions (6.27),(6.28) are satisfied at time t instead of time
zero.
Proof: Analogously to (6.106) from Hookes law, we have for the elastic part of the corrected
strain
(6.30) 1
(6.30) 1
el
(6.116)
11
22 ,
22
11 .
el
22 =
11 =
E
E
Equations (6.101), (6.103) are therefore equivalent to
pl
pl
0 = E f1 (t)
11 + f2 (t)
22 +
11 f1 (t) f2 (t) +
22 f2 (t) f1 (t) .
(6.117)
Pulling back the yield surface centres via
A = D 1
and
A = D 1e
(6.118)
11 = A11 + e p 11 eA11 ,
e
11 = A11 + e 11 eA11 ,
22 = A22 + e e 22 eA22
in case of (S1),
22 = A22 + e p 22 eA22
in case of (S2).
P
e = ef (eR), e = P ej ,
Inserting this into (6.117), bearing = f (R),
= j
j in mind,
j
shows p 11 resp. p 22 explicitly as function of the state variables.
It is not clear, which differential index, cf. Hairer/Wanner [52] section VI.5, systems
(6.112) resp. (6.113) have, as we do not know how to transform them into an underlying
ODE system. The time derivatives p 11 resp p 22 at the right hand side cause this trouble.
And it is not known to the author, if there already exist a definition of an index (differential,
perturbation, ...) for DAE systems with discontinuous right hand sides.
6.4.3 Remark (Multiaxial Neuber correction) We have applied the modification of this section 6.4 to the multiaxial Neuber approaches of Glinka et al. [18, 44, 83, 104] and Chu [23],
where the (due to tensor symmetries) six componentwise algebraic constraints
e
ij eij =
ij ij
i, j {1, 2, 3}
(6.119)
are imposed. Here as well, this leads to a significant improvement, qualitatively comparable
to the results in section 8.3, if we impose the Seeger constraint (6.100) instead of (6.119) for
135
(i, j) = (2, 2). By a dense output method, cf. Enright/Jackson [39, 40], in order to detect
the zeros of the strain ij and to define memories for turning point, we were able to develop
an algorithm for cyclic loading paths. But in any computed example, our approach yields
better results, especially they are much better for ratchetting, cf. paths [e] and [f] in section
8.3, where the multiaxial Neuber approach fails.
136
Chapter 7
It is faster than existing ones, cf. [53, 72, 73], since simultaneous integration in the elastic
and corrected stress space is performed. Numerical tests are presented in section 7.2. An
easy predictor-corrector algorithm is designed for the modified correction model, where Taylor extrapolation reduces the number of corrector steps for smooth loading paths.
For the numerical solution of the elastoplastic BVP, we have tested our variable step size
algorithm II of this chapter against existing well established Abaqus material subroutines
(for linear kinematic hardening, Jiang is of course non-standard). Our method turned out to
be more stable and to yield less convergence problems, when solving the full boundary value
problem.
Throughout this chapter, except for algorithm II, we are exclusively concerned with the
free boundary case, i. e. we consider the correction model for (EPJ) material given by the
differential equations (6.2) and algebraic equations (6.6). We write it in the non-redundant
form
y = Jebs (t, y)
,
(7.1)
z = Lebs (t, y)
where
pl , ,
ec
c
c
c
1 , . . . , ec
I , eR,
1 , . . . ,
I ,
b
d , e,
,
pl , eb, ec
el , ef
z =
, b
s, e
, f
el , ef
y =
R3(2I+1)+3 ,
R
c
d , R63+44+2 = R36 .
The wide tilde resp. the hat stand for the linear embedding
4
e: C 1 R33
sp R ,
(7.2)
(7.3)
138
33 , is
The scalar product h, iM on R3 R33
spd = Vd , which is considered as subspace of R
given by
x, y M = xt M y,
kxk2M = x, x M
(x, y R3 ).
The symmetric, positively definite matrix M is the same as in (6.81). From now on in this
chapter we let drop the b and e , as it is clear what is meant.
7.1
The usual case in application practice is that only sampled data or signals of the elastic
stress tensor are available. If the sample times tn are small enough, it is justified to
interpolate linearly. Therefore, we consider a discrete finite sequence of given
elastic stress tensors e 0 , . . . , e N
at times
t0 = 0 < . . . < tN = T, N N
(7.4)
(t) = e n1 +
t tn1 e
n,
tn
(t)
e n
tn
(7.5)
where
e n = e n e n1 ,
tn = tn tn1
(7.6)
tn = n
(n = 0, . . . , N )
tn = tn tn1 = 1
(n = 1, . . . , N ).
s(t)
= esn .
(7.7)
The basic algorithm I presented below consists of an outer and and inner loop. In the outer
loop, the discontinuity at the yield surface is detected, the inner loop consists of a standard
integration scheme for ordinary differential equations.
7.1.1
The idea is simply to detect the discontinuity at the elastic yield surface Se (e). To this end,
we reparametrise the time in the the form
t [n 1, n]
tn = t n + 1
tn [0, 1]
for n = 1, . . . , N and calculate the future intersection point esn of the straight arc
An = esn1 + tn esn : tn [0, ) ,
139
which is obtained by prolongation of esn , with the current elastic yield surface
e
e
Sen1
n1 n1 = Ben1 n1 ,
whose position remains constant on the interval [0, tn ]. With e d,n1 = esn1 en1 the
intersection point is given by
e
sn
e
= esn1 + tn esn An Sen1
( n1 ),
where
tn =
2
1/2
(7.8)
e
if esn 6= 0. Then An Sen1
( n1 ) has exactly one element. There holds
e
Sen1
( n1 )
tn =
otherwise
(7.9)
e(t) (e(t)) = B
e(0) (e(0))
s(t) B
[tn , 1]
n =
tn tn
1 tn
n [0, 1].
This front tracking method will lead to a significant reduction of steps that are rejected by
the solver, see chapter II.6 in Hairer et al. [51] and the numerical tests in section 7.2.
The discontinuities at SeR and SR are left to the local error control of the standard integrator. Using a constant step size would yield too large errors when crossing these discontinuities.
Algorithm I.
It turns the
e
Input
into the
Output
es ,
n
e ,
n
sn ,
n ,
eel ,
n
el
n,
epl ,
n
pl
n,
ei ,
n
in ,
n,
n,
eR
n,
n,
R
e ,
n
n ,
d,n ,
d,n , n .
Here
n = 0, . . . , N,
i = 1, . . . , I,
l = 1, . . . , L,
L = (2K + 3)I + 6.
(7.10)
Lower indices n are outer loop indices, whereas the upper indices i enumerate the backstresses. It reads as follows.
140
e i e
0 as in (6.27), (6.28).
A Initialisation e 0 , pl
i0 , R
0 , 0 , 0 , R0 ,
e 0 as in (7.7).
B Set es0 := D
C Set e0 , e d,0 , e0 ,
0 , d,0 , 0 , s0 as in (6.8), (6.29).
1 s0 , el := C 1
1 pl as in (5.21), (6.30).
D Set
0 := D
0 , 0 := el
0
0 +D
0
1e 0 , epl := 0 eel as in (6.31).
E Set eel
0
0 := C
0
F For n := 1, . . . , N
G
If 0 tn < 1
Set esn := esn1 + tn esn , esn := esn esn .
e i e
n :=
Set pl
in , R
n , n , n , Rn ,
Dormand/Prince-solution of y = Jes (, y)
over n [0, 1] at n = 1
e i
e
n1
in1 , R
with initial values pl
n1 , n1 , n1 , Rn1 ,
Else
e i e
n := pl , n1 , ein1 , eRn1 ,
n1
Set pl
in , R
in1 , R
n , n , n , Rn ,
n1
acc. to (6.10).
N
Set en , e d,n , en ,
n , d,n , n , sn as in (6.8), (6.29).
1 pl
1 sn , el := C 1
n , n := el
Set
n := D
n as in (5.21), (6.30).
n +D
n
Here
C 1
0
1/E /E
/E
0
1/E
,
=
/E /E
0
0
0 (1 + )/E
1
D
1
0 0
0
1 0
=
1 1 0 .
0
0 1
7.1.2
141
For the integration of the constitutive equations in the plastic regime, i. e. during active plastic
loading of the elastic yield surface in line K, we use the Dormand/Prince method with
variable step size control. It is an embedded Runge-Kutta scheme with Butcher coefficients,
which are given by
1
1
5
ai
3
10
4
5
8
bij
=
9
ej
1
5
3
40
44
45
19372
6561
9017
3168
35
384
71
57600
9
40
56
15
25360
2187
355
33
0
0
32
9
64448
6561
46732
5247
500
1113
71
16695
212
729
49
176
125
192
71
1920
5103
18656
2187
6784
17253
339200
11
84
22
525
1
40
Let > 0 denote a prescribed relative error tolerance for the integrator. Initially for :=
0 := 0, j := 1, we set
(
1 )
|y 0i |
1/5
:= min 1,
max
,
i=1,...,6I+6 max |y0i |, 106 /
J := y0 |0| . . . |0 R(6I+6)7 .
Cf. [102] or section II.4 in Hairer et al. [51]. If the -step
(i = 1, . . . , 6),
Ji+1 := Jes + ai , y + J (bi )t ,
y := y + J (b6 )t,
:= + a6
max
k=1,...,6I+6
)
P7
ke
J
i
i=1 i
max |y i |, |
y i |, 106 /
:= ,
y := y,
j := ,
j := ,
j := j + 1.
The step size control is performed in the following way. If the step with is successful, we
reset
n
o
{1 } [4/5, 5],
:= min 1 , min 4/5 (/)1/5 , 5
if there havent been any failures before, otherwise
:= min 1 , .
:= /2
142
after each following failure. That way, if does not shrink to zero, a partition of the unit
interval
0 = 0 < . . . < J = 1,
j = j j1
(j = 1, . . . , J),
(7.11)
is generated automatically, which is non equidistant in general.
7.1.1 Remark (Total amounts for algorithm I) Altogether we have for the Jiang correction
model in the free boundary case
total number of differential equations = 3(2I + 1) + 3 = 6I + 6
(7.12)
and
2L = number of parameters pl , epl
number of remaining parameters
number of variables , outer loop
number of variables , inner loop
total number of vars and params
=
=
=
2 (2K + 3)I + 6
2
2(26 + 3I) + 1 (N + 1) + 14N + 33 .
2(6I + 7) + 2(6I + 6) + 61
(67 + 6I)N + (4K + 36)I + 187
(7.13)
Here we
(1) do not forget intermediate help variables (like tn , ...) of the outer loop.
(2) do forget the intermediate variables of the inner loop. (Independent of n, we use them
only once.)
( They are E and .
7.1.3
Parameters
excluded.)
Algorithm I is as well applicable to the stress controlled (6.42) and the strain controlled Jiang
constitutive model (6.56) with the isometric isomorphism
b = (a11 , a22 , a11 a22 , a12 , a13 , a23 )t
A 7 A
5
b: R33
sd R ,
and
x, y
= xt M y,
kxk2M = x, x M
2 1 0
M = 1 2 0 ,
0 0 2I
(x, y R5 )
1 0 0
I = 0 1 0 .
0 0 1
(7.14)
(7.15)
The stress controlled model. As it is a subsystem of the elastic stress controlled correction
model, not many further details have to be explained. We just drop the addition equations
(6.26), the e -s and -s.
The strain controlled model. It is necessary at integration points in finite element computations, cf. as well remark 8.3.3 further below. We modify algorithm I slightly according to
143
Simo [103] and section 6.2.2. The following algorithm II integrates along the linear strain
path
(t) = 0 + t,
t [0, 1],
(0) = 0 ,
(1) = 1 ,
= 1 0 ,
(t)
= ,
where we assume parametrisation (rate-independence!) over the unit interval [0, 1]. We as33
sume further that 0 , R33
s , or equivalently 0 , 1 Rs , are given.
Algorithm II.
It turns the
Input
0 ,
(the strain state and derivative)
E, , epl , pl
(the parameters)
i
0 , 0 , R0 , 0 (the actual state of internal variables)
into the
Output
pl
i
1 , s1 , 1 , el
1 , 1 , 1 , 1 , R1 , 1 , 1 , 1,d , 1 .
Here again, i and l may vary as in (7.10). Upper indices i enumerate the backstresses, whereas
indices 0 resp. 1 denotes the values at t = 0 resp. t = 1. It reads as follows.
A Set 0 , 0 , i0 , R0 as in (6.63), (6.64).
B Set 0 , 0 , as in (6.8), (6.29), (6.31), s0 := dev 0 , 0 := 0 0 , 0,d := dev 0 .
1 0 , pl := 0 el as in (5.21), (6.30), C 1 from (6.61).
C Set el
0
0 := C
0
D Set tr := C,
str := dev tr acc. to (6.57).
E Set t as in (7.16).
F If 0 t < 1
G
Set := 0 + t , := 0 + t tr , := (1 t ).
Set 1 , 1 , i1 , R1 :=
Dormand/Prince-solution of y = J (, y) at = 1
over [0, 1]
144
2
1/2
d,0 , str M + d,0 , str M kesn k2M kd,0 k2M e2n1
kstr k2M
(7.16)
if str 6= 0. Otherwise t = . The matrix M and the induced scalar product h, iM are
from (7.14) and (7.15).
7.2
Numerical tests
In these section we consider some numerical examples for our algorithm I. Appropriate identification of the parameters pl , epl is studied and presented in chapter 8 in comparison with
transient elastoplastic boundary value problem solutions and measurements.
In these tests, we consider the elastic stress given in the form of a linear superposition of two
load cases in the form
e
in which
e
11 = (1, 0, 0)t ,
22 = (0, 1, 0)t ,
12 = (0, 0, 1)t
(11)
= 3.0e1
= 1.604e5 ,
= 1.203e5
rj
9.500e1
7.000e1
1.000e2
1.700e2
8.050e2
rj
9.000e1
5.000e1
4.000e1
4.000e1
7.000e1
e
cj
3
1.099e
3.530e2
1.790e2
7.200e1
3.710e1
c
j
1.164e3
3.580e2
1.860e2
7.100e1
3.740e1
e (1)
aj
2
4.300e
6.800e2
7.100e2
8.000e2
7.000e2
a1j
2.360e1
4.000e1
2.000e1
3.000e1
2.500e1
e (1)
bj
1
7.740e
4.440e1
3.460e1
3.000e1
2.700e1
b1j
1.880e1
1.880e1
3.000e1
2.500e1
2.000e1
e (2)
aj
2
4.800e
1.090e1
9.400e2
1.320e1
7.600e2
a2j
2.950e1
3.860e1
3.000e1
3.500e1
6.000e1
e (2)
bj
3
7.000e
1.500e2
2.000e2
2.000e2
3.500e2
b2j
1.710e+0
2.990e+0
2.000e2
2.000e2
3.000e2
Qj
1.390e+0
1.390e+0
1.390e+0
1.390e+0
1.390e+0
Qj
1.000e2
1.000e2
1.000e2
1.000e2
1.000e2
e
a
e
b
e
a
e
b
e
cR
a
b
a
b
cR
(7.17)
2.015e+2
2.880e1
6.160e3
2.850e1
6.450e3
1.000e+2
2.107e+2
3.200e1
7.600e3
5.000e+4
2.600e2
1.000e+2
K = 2,
L = 41.
(7.18)
So altogether, the number of differential equations is 36, the total number of parameters is
84. Condition (6.73) is never violated during the integration process. The critical value at
its right hand side is e ea eb ecR 35.75. For the strain controlled model (in finite element
computations), we find a b cR 51.25, cf. (6.65).
7.2.1
145
A typical application field is the subjection of the body to cyclic loading. Therefore in this
first test, we consider the periodic analytical input
L11 (t) = 621.63 sin(2t),
L22 (t) = 0,
(7.19)
over the time interval [0, T ] with T = 80, = 2/8, these are 10 cycles. Note that (7.19)
is obtained from path [d] in section 8.3 by a rotation of the coordinate system. (So we have
here in fact just a 2D problem.) All other paths in the section mentioned have been checked
in the same way with similar results concerning accuracy and speed.
We compare the solution, obtained by the Dormand/Prince algorithm with smooth input and
variable step size control, with our algorithm I with sampled data
tn = 4k ,
N = 80 4k sample intervals,
k {0, . . . , 5}.
(7.20)
Of course, we falsify the loading path, the error in the input will be Taylor expansion of
order O(tn ). Figure E shows the results qualitatively.
Figure E. The results qualitatively. Left hand side: Plane stress phase plot. Right hand side: The stresses
e
and
over time.
Figure F shows three triples of computations. First, we check the pure Dormand/Prince
algorithm with smooth input (7.19), where no discontinuity detection at Se (e) is applied.
Second, the pure Dormand/Prince algorithm has to handle the same input piecewise linearly
interpolated according to (7.20). Third, the same sampled input is fed into algorithm I.
We point out the following observations.
The pure Dormand/Prince-method for both piecewise linear and smooth input struggles very hard to manage the discontinuity at Se (e). The rate of rejected -steps is
huge.
146
Further improvement may be achieved by the use of dense output methods in Enright et al.
[39, 40]. (The relative computation times of the piecewise linear computation for k {4, 5}
lie much beyond the scope of the depicted region.)
Figure F. Left hand side: Relative computation times. Right hand side: Percentage of rejected steps during
the inner loop in line K, algorithm I. plw means piecewise linear.
Figure G shows the accuracy of algorithm I (with sampled input data and low relative accuracy , fast!) compared to the pure Dormand/Prince integrator (with smooth input and high
relative accuracy , slow!). For the use of , see section 7.1.2. In the relative error, there are
some isolated peaks appearing at whose points, where
has a zero, cf. as well figure E. But
apart from these, the relative error is in the region of approximatively 104 .
Closing this section, we want to compare our method to the existing approaches in [53, 72, 73].
We are integrating the equations in elastic and corrected stress space simultaneously, with
only one Mises yield condition es Se (e), es n > 0, and not in series
e(t)
parameters epl
- pl (t)
parameters pl
-
(t)
(7.21)
as it is done in [53, 72, 73]. As we dont have a stress controlled integration followed by a
completely separate (total) strain controlled integration, our procedure yields more accuracy
and speed. This has several reasons.
147
First, we have to handle the problems with discontinuities only once, as we have the guarantee
that plastic yielding starts/stops exactly at the same point in time.
Second, it is not clear how to interpolate the output pl of the stress controlled step in
(7.21) before feeding into the strain controlled constitutive model. In Hertel [53], linear
interpolation is assumed.
Figure G. Accuracy of exact solution with smooth input and = 109 compared with solution with sampled
input = 103 , N = 80 42 . Left hand side: Absolute error
ij , Right hand side: Relative error
ij /
ij .
Third, for the plastic strain controlled step in (7.21), a nonlinear system of equations of the
form
Jn (
pl
pl
pl
n ) =
(n = 1, . . . , N )
(7.22)
n =
n
n1
has to be solved in each substep, i. e. for each corrected plastic strain increment
pl
n . The
evaluation of Jn involves a strain controlled integration of the form (6.56) for the current
target
pl
n , i. e.
Z 1
Z 1
pl
dt
Jn (
dt (t) dt =
J (t,
, ,
1 , . . . ,
I , R)
n ) =
0
where
(t) = n + t
n ,
n = 0 +
n1
X
j .
j=0
A numerical solution of system (7.22) has been computed in Hertel [53] with the iterative
scheme
(
(0)
n = en ,
(i = 0)
(7.23)
(i)
(i1)
(i1)
pl
n =
n
Jn (
n )
n ,
(i = 1, 2, . . .)
and strain controlled forward integrations, cf. section 7.4.2.2, scheme (I), in Bronstein et
al. [19].
148
Now on figure H you see, which low convergence speed (7.23) exhibits in a practical example.
It is not clear whether a solution of (7.22) actually exists nor whether it is unique. However,
the numerical convergence of (7.23) indicates that it is stable.
The costs for a single function evaluation of Jes are just only double as much as the one of
J and that the average -step size during the increment es is just slightly smaller than
during the increments n .
In this example for algorithm (7.23), we have a ratio of about 21.62 = 250.1/11.57 in the
number of function evaluations and 21.58 = 18.13/0.84 in the number of -steps, averaged
(k)
over all outer increments tn . The criterion for iteration break is chosen as kJn (
n )
7
pl
n k 1.0e .
Figure H. Comparison of our algorithm I with algorithm (7.23), where N = 80 45 and = 103 . Left hand
side: Number of iterations. Right hand side: Total number of function evaluations.
sn = esn ,
(i)
(i1)
(i1)
sn =
sn
k Jsn (
sn )
pl
n ,
(i = 0)
(i = 1, 2, . . .)
(7.24)
may be applied. Here k > 0 is a preconditioning factor, chosen small enough to achieve
stability, and as large as possible to accelerate convergence. Experiments show that iteration
(7.24) needs much more steps, even if k is chosen largest possible.
pl /b
Note that Newtons method instead of (7.24) is not applicable, since the Jacobian c
s
is very close to be singular, because each pointwise Jacobian
dt c
pl
1
en
,
(7.25)
= e en
b
H
dt s
is dyadic and therefore singular. (Here we use the matrix-vector notation from section 6.2.2.)
The derivation of (7.25) is similar to the one in section 6.2.2, cf. (6.11), (6.23) and (6.59).
Note that the output of the stress controlled step (7.21) consists of pl and eel , whose sum is
in general not equal to after the first plastic yielding, cf. (6.39) and the remarks in section
ttgen et al. [72].
4 of Ko
7.2.2
149
We now have a look at another numerical test, where we correct some unsmoothed white
noise. In engineering applications (measurements), such noise is sometimes unavoidable,
when a canonical smoothing without smearing out extrema is not clear. It is important to
verify that our algorithm is able to manage these difficulties.
Figure I. Left hand side: Relative computation times. Right hand side: Average number of -steps on the
interval [0, 1]. Both as functions of relative error tolerance .
For
T = N = 1000,
tn = n,
tn = 1
150
cf. section
(iii) phases of expansion and relaxation of the memory surface radii eR and R,
6.3.4 and the right hand side.
e (e), B
(), B
eR (0), B
R (0). None of
(iv) the forbidden domains outside the closed balls B
them is intruded.
Figure J. Distances and radii. Left hand side: Development of the yield surface radii e and . Right hand
It is again worth mentioning that discontinuity detection at the memory surfaces SeR and SR
is indeed possible, cf. [39, 40], but not really necessary here.
7.3
The algorithm III of this section for the modified correction model is a predictor (P) - corrector
(C) method. For the sake of briefness, we present here the (S2) version of section 6.4, as we
need it later in chapter 8.
The task is to find the input function p 22 such that correction model outputs 11 , 22 satisfy
constraint (6.103) for given functions f1 , f2 . As before, we consider the piecewise linear
problem, i. e. we consider the elastic stress components e 11 , e 12 and the pseudo stress p 22
to be in P W L([0, T ], R). As usual, we let
e
ij,n = e ij (tn ),
ij,n = p ij (tn ),
ii,n = ii (tn ),
fi,n = fi (tn ),
...
We present here two choices for the initial guess, both are easily cheap to
The Elastic predictor. The choice, which one clearly first thinks of, is the elastic
predictor
p
22,n = e 22,n ,
(7.26)
since p 22,n should not be too far away from e 22,n .
151
As already pointed out in Han/Reddy [49] or Comi/Maier [29], the elastic predictor in
the context of finite element computations is not always the best choice. Our experiments
confirm this here as well, see figure L. Hence in the following, we give a heuristic alternative,
which will be quite successful.
The Taylor predictor. We guess a good initial value for the searched input p 22 ,
assuming that it is in some sense smooth enough. We are looking for a better
choice by considering the equidistant extrapolation problem, i. e. to extrapolate given
equidistant data points
p
p
22 t(K1)t , . . . , p 22 tt , p 22 t
22 t+t
in order to foreguess the solution at the next time step. By virtue of lemma B.1.2 we
set for the initial guess
K1
X
K
p
k
p
22,n =
(1)
22,nk1 for n K
(7.27)
k+1
k=0
22,n1 , . . . , p 22,nK
Figure K. Left hand side: Number of Newton corrections for each of the smooth paths of section 8.3 for
K {1, . . . , 4}. K = 1 corresponds to the elastic predictor. Right hand side: Constraint deficiency
Sn /k = 22,n 11,n /k for the choice (6.108) and 1/k {7, . . . , 5} as function of unknown p 22,n . The
orange plot corresponds to the example in section 8.3.
152
Corrector (C). We repeat correcting with Newtons method until the algebraic constraint
(6.103) is satisfied. At time tn , we have to solve the nonlinear equation
Sn (x) = f1,n 11,n (x) f2,n 22,n (x) = 0
(7.28)
(7.29)
According to the proof of lemma 6.4.2, equation (7.28) can equivalently be expressed with
the aid of the internal state variables, namely
pl
pl
e (x) E f
(x)
+
f
(x)
+
11,n (x)(f1,n f2,n )
1,n
2,n
11,n
22,n
n
e
A22,n (x) ,
x A22,n (x) =
n (x)
f1,n f2,n
where
n (x) e
11,n
e (x)
n
n1 ( eAn1 ),
Rn1 ( en1 ),
eA11,n (x) ,
A, A as in (6.118).
n1 ( An1 ),
n1 ( n1 ),
R
pl
n1
f2,n ,
11,n
n ( eAn ),
Rn ( en ),
n ( An ),
n ( n ),
R
pl
n
depend solely on x.
We denote
= e = e = e11 , p22 , e12
(7.30)
the Jacobian with respect to the modified plane elastic stress tensor, written as a row-vector,
cf. (6.110). The Jacobian matrix
is evaluated further below in the next paragraph.
Algorithm III.
It turns the
Input
into
Output
22,n ,
n ,
es ,
n
sn ,
e ,
n
n ,
eel ,
n
el
n,
epl ,
n
pl
n ,
ei ,
n
in ,
n,
n,
eR
n,
n,
R
e ,
n
n ,
d,n ,
d,n , n
Lower indices n are outer loop indices, whereas the upper indices i enumerate the backstresses. n, l and i vary again as in (7.10). For the sake of simplicity, we formulate it
assuming that the initial constraint for t = 0 is already satisfied for the elastic stress, i. e.
p
22,0 = e 22,0 .
It reads as follows.
(7.31)
153
22,n =
as in (7.26)
as in (7.27)
if n < K
,
if n K
p 22,n = p
22,n p 22,n1 .
While true
e i e
n :=
Set pl
in , R
n , n , n , Rn ,
Solution of forward elastic stress controlled integration sweep as in
algorithm I, lines D - K, over [0, 1] at = 1
e i
e
n1
with initial states pl
in1 , R
n1 , n1 , n1 , Rn1 ,
Compute Jacobian (
n ) as in the next paragraph simultaneously.
2 + f2 )
If |Sn | < Tolerance(f1,n
2,n
K
L
Break While.
Else
Tests show that there seems to exist exactly one solution of (7.28) in each incremental outer
step, cf. figure K, that Newtons method converges for arbitrary real initial guess. It does
not matter which nonlinear equation solver is used (here: Newton) in order to find it.
We have tested algorithm III for the six cyclic, smooth input paths from section 8.3. Using
the elastic predictor (7.26), about four Newton iterations are needed for correction.
But the Taylor predictor (7.27) turns out to yield much better initial values. Here it is possible
to have just about two Newton corrections, depending of the choice of number of backward
points K, cf. as well figure K. The optimal choice of K depends on the degree of smoothness
of p 22 and is not known a priori. The numerical experiments here show what is the optimal
K. (But clearly, p 22 is not of class C K for any K.)
154
7.3.1 Remark (Total amounts for algorithm III) Compared to (7.12), we have altogether for
the free boundary case
total number of differential equations:
total number of algebraic equations:
3(2I + 1) + 2 = 6I + 5
1
= e 0 + te,
e,
s=D
= e,
e,
s = D
(7.32)
e = I,
es = tD,
es = D.
x, y M = y t M x + xt M y,
kxkM =
xt M x
,
kxkM
(x) = x + x
for R3 -valued functions x and y and scalar valued functions , we arrive after some computation at
t
(6.11)
e
e 2
e
e
e
e
en = e
k
s
k
I
(
s
(
s
)M
D
M
k s ek3M
e e
(6.12)
ei = eQi 1 + ea e b R (ei )t M en
(6.13)
eci = 0,
de
I
X
j=1
dej
(6.16)
de
(6.22)
e e
0
e
t
( ) M de
if e BeR (0) or (e), de M < 0
otherwise
f ( R) deR
(6.22)
(6.23)
(de)t M en + ent M de + de
1 e t e t
e
e
n M D + s M n dt H
eH
155
(for tangent modulus and the accumulated strain). After reparametrisation from to t, we
receive
dt ei
(6.23)
dt dei + dei dt ,
dt eR
(6.23)
dt deR + deR dt .
Similar formulas hold of course again for the remaining quantities of the corrected stress
space. The simultaneity implies that the time derivative of the corrected stress is given by
1
e e
e
(6.29)(6.30)
1 e
1
1e .
D
(7.33)
dt
=
D dt
+ e D dt + e dt e dt
Pulling back the yield surfaces centres via (6.118), and denoting their radiis ratio by r = /e
we get
1
1 dt e,
1 dt
(7.34)
dt eA = D
dt A = D
,
r = e dt rdt e ,
(7.33)(7.34)
(6.23)
(6.30)
dt pl = dt en + en dt ,
dt = dt el + dt pl .
R1
Differentiation under the integral sign yields (
) = 0 dt ( ) d , which especially gives
us the desired partial derivatives 11 /p 22 and 22 /p 22 .
156
Chapter 8
e2 - e1
forces
f (, t)6= 0?
@
I
@
1 e1
e2
0
B
S
(encastre)
NB
e3
k
'
6
?S
e3
(k , t), (k , t)
(k , t), e(k , t)
tr
(k , t), tr(k , t)
moments
@
I
@
1,g (rigid)
(8.1)
@
I
@
1 \ 1,g
g(, t) =
Li (t)gi ()
Loads, here normal force and torsional moment of scalar magnitude Li (t), are applied at
the point , along the global 1 -direction. All stresses and strains at k , appearing in the
sequel, are measured in the local k -coordinate-system S, whose 3-direction is collinear with
157
158
(b)
(c)
geometry
elements H
HH
j
?
?
?
elastoplastic BVP
E,
specimen
elastic
BVP
Y
H
H
B
B
plasticity
model, pl A B
?
?
?
B
A Be
e
tr( , t), tr( , t)
(k , t), (k , t)
k
k
A B (k , t), (k , t)
P
i
PP
A B
PP
AKA
PP
PP
ep(k)
A
PP
AA error
H
P
P
error
l
HHAU NB
PP
?
P
A
P
P
j
H
AU
P
q
P
multiaxial
(k , t), (k , t)
correction
physics
mathematics
Let us denote the values of the tensors at the time points 0 = t0 < . . . < tN = T of
consideration by (7.4) as in chapter 7, cf. algorithm I and III, and set
e (k)
(k)
(k)
= e 0 , . . . , e N ,
(k)
(k)
(k) = 0 , . . . , N ,
(k)
(k)
(k) =
0 , . . . ,
N ,
...
159
additionally depends on Youngs modulus E and Poisson ratio , but they are considered
constant and remain unchanged. We present in this section two results (tg means target).
(a) We target at some notch strain measurements tg k (t) = tr(k , t),
Here we have just the 11- and 12-components at our disposal.
tg (t)
k
= tr(k , t).
(b) We target at the nonlinear finite element results with parameters pl , i. e. tg k (t) =
(k , t), tgk (t) = (k , t), which will be of course of the main interest from the mathematical point of view. We investigate both (EPJ) and (EPK) material.
(k)
(ii) We vary both the epl and the pl . A higher dimensional parameter space has better
chances to reflect reality (a). The additional costs for the gradient evaluation are low,
as we will see below in figure L on page 165.
We justify (ii) in connection with (a), since we know that Jiangs constitutive law for S460N
is hard to parametrise. As the elastoplastic BVP solution suggests, see page 173, the material
parameters p1 , . . . , pL are not very certain. We refer to the following former works.
The material. For a set of Jiang material parameters pl of S460N with I = 5 backstresses, we refer to Hoffmeyer et al. [62].
The axle. The geometry of the axle (8.1) is described in detail in Hertel [53], section
4.1, 4.2, so that we have been able to build up a finite element model of it. In addition,
notch strain measurements have been performed.
The total parameter identification procedure consists of two steps.
(k)
160
8.1
Parameter identification
Let tg = (tg 0 , . . . , tg N ) and tg = (tg0 , . . . , tgN ) denote the targets, which we want to
approximate optimally. The actuals corresponding to the actual parameters set (p, ep) are
given by the evaluation of the correction model with the constant input e, which is not
depending on (p, ep),
ep), (p, ep), el (p, ep), pl (p, ep), . . . .
(p, ep),
(p, ep), (p,
(8.3)
e, p, ep =
For the objective function F , giving the error between actuals and targets, and which has to
be minimized, we choose the weighted least squares residue
X
X
F (p, ep) =
Fn(ij) (p, ep),
(8.4)
n=0,...,N i,j=1,2,3
(ij)
2 w(ij)
w (ij) e
(ij) 2
tg (ij)
+
n (p, p) n
(8.5)
(ij)
(8.6)
161
(ij)
with respect to the parameters epl (case (i)) resp. (pl , epl ) (case (ii)) of each summand Fn
in (8.4) is computed via
(ij)
Fn (p, ep)
(CR)
(ij)
(ij)
(ij)
n (p, ep)
n (p, ep) tg n
(ij)
(ij)
(ij)
(ij)
n (p, ep)
n (p, ep) tgn
+ w
(ij)
+ ... .
Then obviously
F =
n=0,...,N i,j=1,2,3
Fn(ij) .
each time, in (8.3) is evaluated, we have in principal three possibilities, which are more or
less elemental.
ED External Differentiation or finite differences, based on the definition of the partial
derivatives,
e
( , p, ep) =
pl
e
( , p, ep) =
epl
1 e
( , p + pl el , ep) (e, p, ep) + O p2l ,
pl
1 e
e
e
e
e
e 2
(
,
p,
p
+
p
e
)
(
,
p,
p)
+
O
p
.
l
l
l
epl
(8.7)
(8.8)
el is the lth unit vector of length L and |epl |, |pl | 1, l = 1, . . . , L. The drawbacks
of ED have been pointed out in Hairer et al. [51], section II.6, and prove true for our
application, see section 8.2.
ID Internal Differentiation. For numerical analysis of parameter sensitivities for differential equations, it is recommended in [51], section II.6, to integrate the linearised
differential equations for the parameter derivatives, i. e.
d y
dt pl
d y
dt epl
=
=
y
Jes
Jes
+
t, y(t), p, ep
t, y(t), p, ep ,
y
pl
pl
y
Jes
Jes
t, y(t), p, ep e + e t, y(t), p, ep ,
y
pl
pl
y
(0) = 0,
pl
y
(0) = 0,
epl
(l = 1, . . . , L). Here y from (7.1), (7.2) additionally depends on p and ep. This is as well
very expensive, since the total number of differential equations will be very large and
the evaluation of the Jacobians Jes /y, Jes /pl , Jes / epl not really cheaper.
AD Algorithmic Differentiation, or Automatic Differentiation, is based on the multidimensional chain rule.
We clearly prefer the AD method. This modern technique and the reasons for its usage are
explained in the next section 8.2.
162
8.2
Algorithmic differentiation
:=
:=
:=
:=
:=
:=
e
i
ep
iN
(i = 0, . . . , N )
in
(i = N + 1, . . . , N + L)
in
piN L
(i = N + L + 1, . . . , N + 2L) in
,
(i = N + 2L + 1, . . . , M )
i (xj )ji
(xj0 , . . . , xjN ) (0 j0 , . . . , jN M )
(xi0 , . . . , xiN ) (0 i0 , . . . , iN M )
(8.9)
:=
:=
:=
(CR)
:=
:=
:=
:=
:=
:=
(CR)
:=
:=
:=
epl e i = 0
epl epiN = l,(iN )
epl piN L = 0
X
xj i (xj )ji epl xj
ji
epl xj0 , . . . , epl xjN
epl xi0 , . . . , epl xiN
pl e i = 0
pl epiN = 0
pl piN L = l,(iN L)
X
xj i (xj )ji pl xj
ji
pl xj0 , . . . , pl xjN
pl xi0 , . . . , pl xiN
(i = . . .)
(i = . . .)
(i = . . .)
(i = . . .)
(8.10)
(0 . . .)
(0 . . .)
(i = . . .)
(i = . . .)
(i = . . .)
(i = . . .)
(0 . . .)
(0 . . .)
(8.11)
163
(each code line for l = 1, . . . , L). The original code lines remain without exception. The
symbol denotes Kroneckers delta
xy =
1 if x = y
0 if x 6= y
The first equations in (8.10) resp. (8.11) express the fact, that the input e is independent
of all pl resp. epl . The second and third equations in (8.10) and (8.11) express the fact, that
each parameter is independent of all others (surely, except for itself, where the derivative is
trivially equal to unity).
8.2.1 Remark (Choice of AD mode) The number of output variables 2N (usually several thousands) is much larger than the number of parameters 2L in case (i) resp. 2L in case (ii). Thats
why the AD forward mode is to be preferred to the backward mode for the sakes of speed
and memory. For details/theory and further explanation, the reader is referred to Griewank
[46], Rall [97] or Rich/Hill [99].
8.2.2 Examples. (a) Consider the general exponential function x3 = 3 (x1 , x2 ) = xx1 2 , where
x1 > 0, x2 R. Here we have
x1
x2
x3 (CR) x2 x1 x2 x2
=
+
= x2 xx1 2 1
+ ln(x1 ) xx1 2
x1
x1
pl
x1
pl
x2
pl
pl
pl
(8.12)
x2
x3 (CR) x1
=
=
x2 + x1
(x1 x2 ) +
(x1 x2 )
pl
pl x1
x2
pl
pl
pl
(8.13)
164
8.2.3 Remark (AD in MATLAB) An AD tool called Admat, developed by Verma et al., for
the language Matlab is free available under
http://www.cs.cornell.edu/home/verma/AD
(8.14)
but still not free of bugs and sophisticated errors. At the moment, a new site is arising with
many links to AD tools for other languages
http://www.autodiff.org/
Performance of Matlab-code: If the number of parameters is very large, as in our case,
you should realise the code for the derivatives not with for-loops, as it is (easily) formulated
in (8.10), ..., (8.13), Griewank [46] and realised in the tool (8.14), but use the typically
Matlab-kind operations ., ./, repmat, vertcat, etc.
We give two examples. The lines (8.12) resp. (8.13) from the example 8.2.2 are implemented
much more effectively in the form
pl x3 = x2 x1(x2 1) . pl x1 .+ ln(x1 ) x1x2 . pl x2
resp.
Here
pl x3 = pl x1 . x2 .+ x1 . pl x2 .
xi = xi .V,
v . w = (v1 w1 , . . . , vL wL ).
Analogously for the elastic parameters epl . In our case, we have designed our own AD code
optimised especially for a large number of parameters, such that we have gained a speed factor
of about 80 compared to the tool (8.14).
Why should AD be preferred to ED or ID? As code includes the integration of
discontinuous, highly nonlinear differential equations, it is very complex. Simple computation
of the derivatives by hand or adjoining the differential equations is very elaborate and prone
to errors. Thats why we do not apply ID.
In the following, we discuss the three aspects of computational time, accuracy and memory
requirements, why we do not apply ED.
(1) Computational time. It is clear that the computation times for ED process almost linearly
in the total numbers of parameters: For the forward/backward differences (8.7), (8.8), you
need 2L evaluations of at the disturbed operating points (e, p+pl el , ep) resp. (e, p, ep+
epl el ) plus 1 evaluation of at the operating point (e, p, ep).
The averaged computation times for the six loading paths [a], ..., [f] in section 8.3 are depicted
in figure L on page 165, where our own designed AD tool, see remark 8.2.3, is compared to
ED.
Performing AD, we have gained a speed factor of more than 5.8 in the case (ii) of 2L = 82
parameters (pl , epl ) and a factor of more than 3.3 in the case (i) of L = 41 parameters epl ,
compared to ED.
165
Further, the additional costs for the derivatives with respect to the pl are less than 20%, and
not L/(L + 1) 98% as it is when applying ED.
(2) Accuracy. When applying AD, the only occurring errors are roundoff errors (due to the
limited relative machine accuracy 2.22e16 ), there are no truncation errors (due to the
neglecting of higher order terms in (8.7), (8.8)). And there is no dependence on the ambiguous
choice of the perturbations epl and pl .
Figure L. Left hand side: Averaged computation times vs. number of parameters. Right hand side: Maximum
memory requirements as function of I and N .
(3) Memory requirements. For each variable, that has been used in code for , you additionally
need 2L variables in the extended code for (f, f ). The original code has
V =
ri
M Y
X
ni,j
i=0 j=1
variables of type single/double, so the extended code has (2L + 1)V = (2L + 1)
ones. Here in our case for algorithm I we have
PM Qri
i=0
j=1 ni,j
(7.13)
The right hand upper bound is for fixed K linearly increasing in the number of time
points N and quadratically increasing in the number of backstresses I. The special choice
I = 5, K = 2 in (7.18) gives
(2L + 1)V 8051N + 33781.
166
It follows, that for N = 10 000 at most 645Mbytes of memory are needed, where 1 variable
= 1 double = 8 Bytes. This is practicable with contemporary technique. We should mention
that the estimate above contains the sensitivities of the whole output time signals of the whole
Jiang output variables (which are of course not all needed). The memory requirements for
algorithm III are estimated similarly.
H
Y
H
@
I
@
Figure M. The two load cases in the 2D model: Tension force (above), Torsion moment (below). Colored is
the Mises stress. lies in the red high-stressed notch region, cf. (8.1).
8.3
Results
We consider the rotational symmetric axle as in the beginning of chapter 8 and subject it to
simultaneous tension and torsion. The elastic stress is given by linear superposition
e
(8.15)
2.375e0
0
0
linear elastic BVP solution to static unit
e
1
0
0
5.299e
1 () =
= tension force along the rotational axis
(positively directed, i. e. in 1 -direction)
0
0
0
and
0
1.559e0 0
linear elastic BVP solution to static unit
e
0
0 = torsional moment along the rotational axis
2 () = 1.559e0
(positively directed, i. e. in 1 -direction)
0
0
0
cf. lemma 2.4.2. The elastic strain ratio constant k in (6.107) for the Seeger modified correction model is
k
(6.107)
e
e
L1 (t)e 1,11 () L1 (t)e 1,22 ()
1,11 () 1,22 ()
=
e
e
L1 (t)e 1,22 () L1 (t)e 1,11 ()
1,22 () 1,11 ()
167
since
e
2,11 () = e 2,22 () = 0.
It depends just on the tension load case e 1 (), but neither on the torsion load case e 2 ()
nor the applied load functions t 7 Li (t). The zeros of function t 7 L1 (t) can be lifted. Here
k = 1.214e1 .
We consider the following six periodic loading paths below, numbered by [a], ..., [f]. Paths
of this kind are widespread in engineering in order to check out the material model properties,
ttgen et al. [72, 73], Glinka et al. [18, 83, 44, 104].
cf. Hertel [53], Ko
> 0 denotes a fixed time period, = 2/ the corresponding circular frequency. (Clearly,
because of rate-independence, has in fact the meaning of a fictive time period.)
[a] The proportional path (class C ),
L1 (t) = A1 sin(t),
L2 (t) = A2 sin(t).
L1 (t)
=
A1
2t/
1
cos 2t
1
cos 2t
1
t /2
/2 t
t , 0 4t/
t , 1 4t/
t , 2 4t/
t , 3 4t/
1
,
2
3
4
Here t = t mod .
2t/
1
L2 (t)
1
=
cos 2t
A2
cos 2t
t/
t
,
cos t t
L2 (t) = A2
...
...
...
.
...
...
...
0
t
.
sin t t
L2 (t) = A2 sin(t/2).
[e] Periodic torsion with constant tension, non-zero elastic mean stress (class C for t ),
L1 (t) = A1
t/
1
t
,
t
L2 (t) = A2
0
t
.
sin t t
[f] Out of phase circular path with non-zero elastic mean stress (class C for t ),
L1 (t) = A1
0
t
,
1 cos t t
L2 (t) = A2
t/
t
.
1 + sin t t
168
Loads [e], [f] will produce the ratchetting effect when considering Jiang material, since
the elastic mean stress and thus the elastoplastic mean stress is not equal to zero. Paths
[a], ..., [d] have a zero elastic mean stress. Paths [b], ..., [f] are usually called non-proportional in the community of material scientists.
The targets and weights. Three identification examples are presented here, combining (a)
with cases (i), (ii) and (b) with case (i).
(a) Measurements have been carried out, cf. section 7 in Hertel [53], for the normal resp.
shear strain tr11 (, t) resp. tr12 (, t) at the notch root, but not for the stresses. Here
we set
(t) : t [0, T ] V,
i. e. the trace of
(blue) trapped exactly in the same V . It follows from lemma 6.3.2, that
from t = 0 until the first active yielding at
e( ) (0)
t = t = sup [0, T ] : dev es( ) B
(orange).
It would be very interesting to have the measurements tr at additional disposal in order to
plot.
(front view)
(side view)
169
(top view)
Correction
Correction
In the proportional case path [a], where L2 (t) = L1 (t) with a constant > 0, we even have
e
dim V = 1,
(t) = (1 + c)L1 (t) 1 + 2 : t [0, T ] LH 1 + 2 = V,
and consequently
(t) : t [0, T ] V,
the trace of
constrained in the one-dimensional linear subspace V .
170
8.3.2 Remark (Feasibility) It is important to note, that our correction model makes high-cycle
fatigue analysis feasible. The rough computing times for (EPJ) material are
10 cycles high end finite element solver Abaqus, approximately 40 minutes.
10 cycles slow interpreter Matlab, approximately 5 seconds (algorithm I), 12 seconds
(algorithm III).
Hereby, the same step-size controlled integrator for the inner loop, see section 7.1.2, algorithms
I, II and III in chapter 7 and remark 8.3.3, was used. The mesh had 4877 nodes and 1586
elements.
8.3.3 Remark (Finite element computations) Both the solution of the elastic BVP (8), (9)
and of the elastoplastic BVP (5), ..., (7) have been performed with the finite element solver
Abaqus. In both cases, eight node, biquadratic, axial symmetric solid elements with reduced
integration have been chosen, cf. Kardestuncer [71], part 3, section 1.4a. The number of
degrees of freedom per node is three. They are the displacements in 1 and 3 direction and
the twist angle around the 1 direction.
Some remarks for the elastoplastic case. Jiangs model (EPJ) and the linear kinematic hardening model (EPK) have been implemented via the user subroutine interface Umat for user
defined constitutive material. The overall algorithm to iterate forces and displacements into
equilibrium is realised with Newton-Raphson algorithm (see Zienkiewicz/Taylor [115],
section 2.2.2), where Umat computes the stress increments at the elements integration
points from the strain increments = (u + ut )/2 through integration of the constitutive differential equations of (EPJ) or (EPK) by Dormand/Prince with discontinuity
detection. See algorithm II in section 7.1. This is an improvement of the subincrementation method of [115], section 3.4.1, as we have a variable step size controlling the local
time discretisation error. It leads to less convergence problems of the Newton-Raphson solver
compared to the use of the standard (EPK) implementation in Abaqus.
In order to assembly the total stiffness matrix, there is the necessity for computation of the
local stiffness matrix /. It is computed from
Z 1
Z 1
ep
ep
ep
C (t) dt ,
(8.16)
C (t)(t)
dt =
= C ,
=
0
C ep (t) d.
=
0
The integration of C ep can be realised parallelly during integration of Jiangs strain controlled
model, cf. algorithm II, via the simple quadrature scheme
Z 1
J
X
ep
j C ep ( j )
C (t) dt
t C + 1t
0
j=1
(6.62)
J
X
j I
t C + 1 t C
j=1
C I 2 1 t
J
X
j=1
2
j
j
n
(
)
(
)
2 + H( j )
!
j
n
( j ) n
( j )
2 + H( j )
171
(notation of section 6.2.2) and needs therefore only little additional cost compared to the
evaluation of /, via finite differences or algorithmic differentiation. Here the j ,
j = 0, . . . , J, are from the automatically generated partition (7.11) of [0, 1].
The following table gives an overview about the results, which are depicted on the following
pages 173-194. Correction means algorithm I, Correction means algorithm III.
(A)
(A)(i)
(A)(ii)
(B)(i)
(B)(i)
target
start of iteration
measurements (tr)
measurements (tr)
transient
elastoplastic BVP
solution ()
transient
elastoplastic BVP
solution ()
constitutive law
plotted
, e,
tr,
, e
tr,
, e
,
, e
(, ), (
,
), (e, e)
, , e
,
, e
(, ), (
,
), (e, e)
, , e
tr,
(EPJ)
(EPJ)
(EPK)
pages
173
174
175
176, ..., 178
179, ..., 181
182, ..., 184
185, ..., 187
188, ..., 190
191, ..., 193
172
badly identified parameters, as already mentioned above. You can see, that the computed
strains are still slightly overestimating the measured strains, which may usually be desired, if
a subsequent fatigue analysis is performed.
(b) The error curves for the Seeger modified correction model (algorithm III) lie much below
the ones for the pure correction model (algorithm I), cf. remark 8.3.1. It is seen in the
plots, that the maxima/minima (the turning points) in the stress phase diagrams are hit
very exactly. The results for the corresponding strains are a little bit worse compared to the
stresses, but they are still and this is decisive for practice better than the elastic solution.
They lie somewhere between.
Figure O. Standardised least-squares residues (F (p() , ep() )/F (p(0) , ep(0) ))1/2 as functions of the iteration step
number are depicted in these graphs, both for Jiang material, (a) at left hand side, (b) at right hand side.
Why is path [c] in case (a) so bad? Due to Hertel [53], it is Jiangs model that exhibits bad
properties in out-of-phase loadings. The steel S460N shows out-of-phase hardening, i. e. the
measured strains in the cyclic stabilised regime are smaller than the computed ones. With
ring
improved constitutive models, you may get rid of this drawback, cf. Tanaka [106], Do
nlich [57].
[33, 34] and Herz [54]. For path [b], cf. as well the numerical results in Hu
8.3.4 Remark. Within the framework of this thesis, AD-sensitivity analysis of the whole left hand
high cycle chain on page 3 has been performed and is available from the author, starting
from the elastic BVP output until the damage. It includes
(1) Rainflow counting (uniaxial, multiaxial, critical plane approach), see [4, 35, 36, 37, 15,
26, 101].
(2) Uniaxial corrections (Neuber [90, 91], ESED [45]).
(3) Damage parameters: Manson, Coffin and Morrow [84, 28, 87], Smith, Watson and
Topper [105], Goodman, Bergmann, and many more.
Further, within the framework of a project with the technical university of Darmstadt, our
AD based parameter identification technique has successfully been tested for the DoringVormwald constitutive model [33, 34, 54], cf. section B.3 as well.
Correction
173
Material: (EPJ)
Measurements
174
Correction
Material: (EPJ)
Measurements
Correction
175
Material: (EPJ)
Measurements
176
Correction
Material: (EPJ)
Correction
Correction
177
Material: (EPJ)
Correction
178
Correction
Material: (EPJ)
Correction
Correction
179
Material: (EPJ)
Correction
180
Correction
Material: (EPJ)
Correction
Correction
181
Material: (EPJ)
Correction
182
Correction
Material: (EPJ)
Correction
Correction
183
Material: (EPJ)
Correction
184
Correction
Material: (EPJ)
Correction
Correction
185
Material: (EPK)
Correction
186
Correction
Material: (EPK)
Correction
Correction
187
Material: (EPK)
Correction
188
Correction
Material: (EPK)
Correction
Correction
189
Material: (EPK)
Correction
190
Correction
Material: (EPK)
Correction
Correction
191
Material: (EPK)
Correction
192
Correction
Material: (EPK)
Correction
Correction
193
Material: (EPK)
Correction
194
Correction
Material: (EPK)
Correction
195
Lists of identified parameters. The constants E, (and , ) for all computations are
chosen as in (7.17).
(A) (ii) Target: Measurements (strain tr), Constitutive model: (EPJ)
i
1
2
3
4
5
i
1
2
3
4
5
ri
9.701e1
7.055e1
1.005e2
1.701e2
8.051e2
ri
8.991e1
5.263e1
5.091e1
3.830e1
6.890e1
e
ci
3
1.102e
3.509e2
1.828e2
7.211e1
3.717e1
c
i
1.153e3
3.601e2
1.884e2
7.065e1
3.665e1
e (1)
ai
1
2.205e
7.980e1
7.390e1
1.051e1
9.065e2
a1i
1
4.378e
4.524e1
1.365e1
5.880e1
3.137e1
e (1)
bi
1
7.171e
4.285e1
4.092e1
3.075e1
2.772e1
b1i
1.847e1
1.877e1
2.973e1
2.383e1
1.948e1
e (2)
ai
2
e (2)
bi
6
9.471e
3.640e1
7.357e1
4.166e2
6.392e3
a2i
1
1.299e
5.981e1
5.342e1
4.902e1
7.142e1
7.554e
2.808e1
7.008e2
2.567e1
2.262e1
b2i
+0
1.998e
2.717e+0
3.851e1
3.702e1
4.443e1
Qi
1.307e+0
1.567e+0
1.417e+0
1.036e+0
1.032e+0
Qi
2.487e2
3.699e2
7.742e3
4.143e3
7.428e3
e
a
e
b
e
a
e
b
e
cR
a
b
a
b
cR
2.037e+2
1.005e1
4.558e2
2.871e+1
2.584e3
9.583e+1
2.150e+2
1.122e1
2.737e2
4.999e+4
4.864e2
9.924e+1
(B) (ii) Target: Transient elastoplastic BVP solution (stress ), Constitutive model: (EPJ)
i
1
2
3
4
5
ri
8.389e1
6.987e1
1.065e2
1.719e2
8.135e2
e (1)
ai
2
e
ci
3
1.095e 2.766e
3.552e2 1.870e1
1.796e2
1.651e2
1
7.304e
1.304e1
1
3.771e 6.982e2
e (1)
bi
1
e (2)
ai
2
e (2)
bi
6
7.050e 6.909e
4.751e1
5.267e1
1
3.540e
1.925e1
1
2.526e
5.731e2
1
2.822e
1.148e1
1.581e
2.225e1
1.128e2
3.338e1
6.848e1
Qi
1.491e+0
7.068e+0
6.199e+1
1.360e+0
2.512e+0
1.952e+2
e
a 3.022e1
e
b 7.021e3
e
a
2.707e+1
e
b 1.195e+1
e
1.100e+2
cR
For the pl see page 144. The epl on the same page have been the initial parameters for each
identification procedure. We have I, K and L as in (7.18).
(B) (ii) Target: Transient elastoplastic BVP solution (stress ), Constitutive model: (EPK)
e
b11
1.191e5
b11
1.111e5
b22
1.183e4
b22
1.111e5
b33
0.000e0
b33
0.000e0
b12
1.297e5
b12
1.111e5
b13
0.000e0
b13
0.000e0
b23
0.000e0
b23
0.000e0
1.433e2
1.433e2
Here eB = (ebij )i,j=1,2,3 and B = (bij )i,j=1,2,3 . The initial parameters for the iteration have
been the material parameters itself. The identification could as well be done by hand here.
196
Part IV
APPENDIX
197
Appendix A
Proof of Krej
cs theorem
We give here the detailed proof of theorem 1.1.1 for the sake of completeness and because of
[75] for the proof of part (a) and proposition 4.1
its beauty. We adapt theorem 3.1 in Krejc
[76] for the proof of part (b) by the use of equivalence (B.12). The main ingredient
in Krejc
is the Hahn-Banach convex separation theorem.
(a) Proof for the class of CBV -functions. Step I. Let a function f CBV ([0, T ], X)
and a point s0 C be given. For N N, we introduce the equidistant partition
t0 = 0 < t1 = tN < t2 = 2tN < . . . < tN = N tN = T,
tN = T /N
p0 = f 0 s 0 ,
pn = f n s n ,
s0 = s0 ,
sn = A (fn pn1 ),
(n = 0)
(n = 1, . . . , N )
(A.1)
A : X X,
A +
A = id
are defined by
A (x) C,
k
A (x)kA = distA (x, C) = min kx ckA ,
cC
A (x) +
A (x) = x
(A.2)
sn = sn sn1 ,
pn = pn pn1
(n = 1, . . . , N ).
pN
,
tN
N s)(t) =
(
sN
,
tN
1
A
A (x), A (x) c 0
199
N f )(t) =
(
for all c C.
fN
.
tN
(A.3)
(A.4)
s theorem A
proof of krejc
200
(Hahn-Banach convex separation theorem, cf. lemma 2.2 (i) in [75].) There holds by construction
(A.2)
(A.1)
(A.1)
(A.4)
A1 (pn pn1 ), sn c = A1
A (fn pn1 ), A (fn pn1 ) c 0
(A.5)
Therefore,
(A.1)
0 A1 pn , sn A1 pn , fn .
kpn k2A
(A.6),(A.1)
and
A1 pn , fn
(B.12)
cA kpn k kpn kA
(A.6)
(CS)
kpn kA kfn kA
(B.12)
CA kfn k
N
X
n=1
N
CA
CA X
kfn k
(f ).
kpn k
cA
cA
(A.7)
n=1
1
t
t
(A.3)
n1
1
N p)(t), (N s)(t) c
=
A pn , sn1 +
tN A (
sn c
tN
=
A pn , sn c A1 pn , sn
t tn1
1
+
A pn , sn
tN
(A.5),(A.6)
A pn , sn
(A.6)
A1 pn , fn
(CS),(B.12)
2
CA
kpn k f (tN )
sup
|1 2 |tN
kf (1 ) f (2 )k.
Z tD
E (A.7) C 3
N
A1 d(N p)( ), (N s )( ) A (f )f (tN ) 0.
cA
0
(A.8)
s theorem
A proof of krejc
201
Np
M p)( ), (N f M f )( ) d
A1 (
Z t
1
Np
M p)( ), (N s M s)( ) d
A (
(A.8)
2
CA
kN f M f k N p M p
3
CA
(f ) f (tN ) + f (tM )
cA
CA
2
f (tN ) + f (tM )
CA (f ) 2 kN f M f k +
cA
(A.7)
by the fundamental triangle estimate for Riemann-Stieltjes integrals, cf. section V.1 in
N
N
[75]. Since N f f in C([0, T ], X) and f (tN ) 0, the sequence N p is Cauchy in
N
the first condition in (1.3). There clearly holds N s(t) s(t) in X for each t [0, T ], giving
especially the fourth condition in (1.3). As C is closed, there holds s C([0, T ], C), the third
condition in (1.3).
Because of proposition 1.18, chapter V, in [75], we see that p is in CBV ([0, T ], X) again,
satisfying
(A.7)
CA
(f ) < .
cA
1+
CA
(f ) < .
cA
Altogether, the existence of a solution of (1.3) in the class of CBV functions is proven.
Step IV. The uniqueness follows in the same way as in the proof of the Lipschitz estimates
in theorem 1.3.7.
(b) Proof for the class of W 1,q -functions. Step I. Let f even satisfy f AC([0, T ], X).
It follows by lemma 4.1 in [76] that the integral condition in (1.3) implies
Z
t2
t1
A1 dp( ), (s )( ) 0
C([t1 , t2 ], C)
(A.9)
for any choice 0 t1 < t2 T . Therefore, as p and s are of class CBV according to (a), by
s theorem A
proof of krejc
202
taking the test function = s(t1 ), we find
1
kp(t2 ) p(t1 )k2A
2
(A.9)
(IP )
t1
t2
t1
t2
t2
p( ), f ( ) f (t1 )
t1
t2
t1
D
E
A1 dp( ), f ( ) f (t1 )
1
A p(t2 ), f( ) d
t2
t1
t2
t1
A1 p( ), f( ) d
A1 p( ), f( ) d
A1 p(t2 ) p( ) , f( ) d
t2
t1
kf( )kA d,
where integration by parts for Riemann-Stieltjes integrals has been applied, cf. section 9.3 in
[76]. Therefore,
(B.12)
cA
kp(t2 ) p(t1 )k
2
Z t2
1
kf( )kA d
kp(t2 ) p(t1 )kA
2
t1
Z
Z T
(B.12)
kf ( )kA d CA (t2 t1 )
(t2 t1 )
0
T
0
kf( )kd,
where 0 t1 < t2 T maybe arbitrarily chosen. Now let > 0. We set, where we may
w.l.o.g. assume f 6 const,
1
Z T
cA
.
(A.10)
kf ( )kd
=
3CA
0
We let further 0 a1 < b1P< . . . < aN < bN T be any partition of [0, T ] such that
P
N
N
n=1 (bn an ) < . Then
n=1 kp(bn ) p(an )k < by construction (A.10). This means
p AC([0, T ], X). Therefore, s = f p AC([0, T ], X) as well. Thus (A.9) implies
Z
t2
t1
A1 p(
), (s )( ) d 0
C([t1 , t2 ], C)
for any choice 0 t1 < t2 T . This is equivalent to the pointwise inequality hA1 p(t),
s(t)
ci 0 for all c C and a. e. t [0, T ] in (1.4).
Step II. Now, if we assume that f even satisfies f W 1,q ([0, T ], X) for any 1 q ,
relation (1.12) yields
kp(t)k
CA
kf (t)k,
cA
ks(t)k
CA
kf (t)k
cA
a. e. in [0, T ].
Appendix B
Miscellaneous
B.1
The two lemmas presented here are of general nature and therefore extracted here into the
appendix. The following lemma is similar to a one in Br
ezis [8].
B.1.1 Lemma (Gronwall-kind estimate) Let two real valued-functions f L1 ([0, T ], R) and
g C([0, T ], R) be given, such that there holds
f (t) 0 for a. e. t [0, T ],
g(0) 0.
1 2
1
g (t) g2 (0) +
2
2
f ( )g( )d
0
(B.1)
Thus
f ( )d
2
1
h (t) := g(0) + +
2
1
1
h (t) g2 (0) + 2 +
2
2
Z
Z
(B.2)
for t [0, T ].
f ( )g( )d
0
(B.1)
(B.3)
f ( )g( )d
1 2
1
g (t) + 2
2
2
1 2
h (t)
2
e. in [0, T ].
(B.4)
a. e. in [0, T ].
(B.5)
and
(B.4)
p
d
h (t) = f ( )g( ) f ( ) 2h ( )
dt
203
204
miscellaneous B
Thus
and
(B.4)
|g(t)|
h (t)
2h (t)
(F T ) 1
h (0)
2
2h (0) +
a. e. in [0, T ].
f ( )d
(B.3)
f ( )d = |g(0)| + +
f ( )d,
0
which is valid for each > 0. Letting 0 gives the desired result (B.2).
The following lemma is standard, since it is a direct consequence of Taylors expansion theorem.
B.1.2 Lemma (Extrapolation lemma) Let the function x be of class C K (U, CN ) in an open
neighbourhood U R of the compact interval [t (K 1)t, t + t] for an appropriate
K N. Then the discrete differential operator with K backward steps of size t, defined by
K
X
K
k K
t+t x =
(1)
xt+(1k)t ,
(B.6)
k
k=0
satisfies
K t+t x = O tK
for t 0.
(B.7)
(B.8)
k=0
for the binomial coefficients, whose validity can be verified inductively, we straightforwardly
arrive at
K xt+t xt+t
K
X
(B.6)
k K
=
(1)
xt(k1)t
k
k=1
!
K1
K1
X
X x() |t
(T E)
K
k+1
K
(1)
=
(kt) + O t
k+1
!
=0
k=0
!
K1
K1
X x() |t
X
K
=
+ O tK
(t)
(1)k++1 k
k+1
!
=0
(B.8),(T E)
t+t
k=0
+ O tK
where we have used Taylor expansion theorem twice. This yields (B.7). Uniqueness follows
from the identity theorem for polynomials.
B miscellaneous
B.2
205
In this section, we resume some basic facts and estimates for linear operators and function
spaces. Let 1 q and T > 0 be real numbers.
For two separable Hilbert spaces X, Y
and a linear, continuous operator A L X, Y , let its corresponding hatted operator A be
defined by
A : M AP ([0, T ], X) M AP ([0, T ], Y ),
(A)(t)
= A((t)).
such that
A L (W 1,q ([0, T ], X), k kW 1,q ), (W 1,q ([0, T ], Y ), k kW 1,q )
d
(A)(t) = A (t)
dt
Further,
a. e. in [0, T ].
A L (C([0, T ], X), k k ), (C([0, T ], Y ), k k ) ,
A L (CBV ([0, T ], X), k kCBV ), (CBV ([0, T ], Y ), k kCBV ) .
Z
kA( )k d
kAk
1/q
k( )k d
Z
1/q
kA(
)k d
Z
+ kAkq
T
0
1/q
q
k(
)k d
1/q
for the standard Sobolev norm in the case 1 q < (for q = similarly), and (again
similarly)
W 1,q, kAk kkW 1,q, ,
kAk
kAk kk ,
kAk
for the other (semi-)norms. The operator A is continuous, its norm bounded by the original
operator norm
W 1,q kAk,
kAk
W 1,q, kAk,
kAk
kAk,
kAk
CBV kAk.
kAk
Similar inequalities hold, if A has a continuous inverse A1 L(Y, X), cf. theorem 10.5 in
Heuser [55] or theorem 26.A in Zeidler [111]. Then from
kAk A kk,
there follows
W 1,q A kkW 1,q ,
kAk
A kk ,
kAk
(B.9)
206
miscellaneous B
A 7 A,
: L(X, Y ) L W 1,q ([0, T ], X), W 1,q ([0, T ], Y ) , A 7 A
and : L(X, Y ) L CBV ([0, T ], X), CBV ([0, T ], Y ) , A 7 A
are all linear and continuous with norms kk, which is bounded above by unity. Operator
convergence
n
An A
in L(X, Y )
implies
n
An A
n
An A
n
An A
in L C([0, T ], X), C([0, T ], Y ) ,
in L W 1,q ([0, T ], X), W 1,q ([0, T ], Y ) ,
in L CBV ([0, T ], X), CBV ([0, T ], Y ) .
(b) We consider the special case that Y = X . If the operator A L(X, X ) is symmetric
and strongly positive, i. e.
A = A ,
A, A kk2 ,
(B.10)
definition
A1 , ,
kk2A =
(B.11)
yield a scalar product and a norm on X X . This new norm, induced by A, is equivalent
to the original one,
cA k k k kA CA k k.
(B.12)
It should be mentioned that the inverse operator A1 L(X , X) exists, because the estimate
A kk2
(B.10)
(CS)
A, kAk kk
implies estimate (B.9), which in turn yields the invertibility. As well, A1 is symmetric and
strongly positive. There holds
kA1 k = sup
6=0
1
kk (B.9)
kk
kA1 k
= sup
=
.
sup
kk
kAk
kk
A
6=0
6=0 A
(B.13)
kk2A
1
= A
,
sup A1 , = kA1 k
2
kk
kk kk
k k=1
and below
kk2A
hA1 , i
h, Ai
A kk2
A
=
=
=
.
kk2
h, i
hA, Ai
kAk2 kk2
kAk2
By taking the infimum over all such CA resp. the supremum over all such cA in (B.12), one
sees that the choices
kAk
1
,
A
cA
CA kA1 k1/2
(B.13)
(B.14)
B miscellaneous
207
for the equivalence constants are possible. Similarly as above in part (a), we are able to
estimate
Z T
Z T
2
2
k(t)k
and
A,
A(t), (t) dt +
T
0
A(t),
(t)
dt
1,2,
A kk2W 1,2, ([0,T ],X)
A,
W
([0,T ],X)
A,
= A(0), (0) +
T
0
A(t),
(t)
dt.
as it is
Summarising, it is possible to identify operator A with its hatted counterpart A,
sometimes done in the theoretical part of this thesis.
B.3
Remarks on the D
oring-Vormwald constitutive model
ring,
In the meantime, an improved version of Jiangs model has been developed by Do
Vormwald et al. [33, 34] and Tanaka [106], which takes out-of-phase phenomena better
into account. We have a short look at it in this section, since some theoretical pieces of section
6.3.1 are still valid here.
We consider Dorings model in its simplified version without Tanaka tensor, cf. Tanaka
[106]. We discuss it in the form given in Herz [54]. For the sake of shortness we have just a
look at its explicit stress controlled version. The following differential equations
d = b T ,
(B.15)
i
d ri = b rT ri ,
(B.16)
!
i +1
i
ki k
(i ) +
d ri ,
(B.17)
d i = ci ri n
ri
ri
d =
I
X
d j
(B.18)
j=1
replace (6.47), ..., (6.50). Here again always, if not otherwise stated, i = 1, . . . , I. The socalled target functions T , rTi : [0, ) R are given by
!
!
I
I
i
X
X
aj
a
j
i
1+
,
rTi () = r
T () = 1 +
(1 + bj )2
(1 + bj )2
j=1
j=1
i = i0 + (i0 + 1)(ci 1) 1 |n : (i )| .
(B.19)
208
miscellaneous B
exactly in the
The reparametrisation w. r. t. time t is again done by multiplication with ,
same way as in (6.22) and (6.23). The essential differences compared to Jiangs model are
that
(1) the partial backstress limiting radii ri themselves satisfy differential equations, namely
and (B.16), and are not constant.
(2) the yield surface radius is not coupled to the memory surface radius R any more.
The governing equation (B.15) is the same as those for the ri . (The memory surface is
realised in the plastic strain space, which we do not consider here.)
(3) the ci are constant parameters and not functions of any more.
Compared to Jiangs model (6.42), the Jiang-like subsystem of Dorings model, which we
examine here, can be written as
1 , . . . , I , r1 , . . . , rI , ,
pl , ,
R = Js t, , 1 , . . . , I , r1 , . . . , rI , , R .
For the parameters, which are appearing in the model, we assume that they satisfy the
conditions
i
Qi , ci 0,
ci > 1,
a > 1,
b, bi , r
, > 0
(B.20)
and (B.21) for the aij and aj . The following lemma is the analogon to lemma 6.3.5.
B.3.1 Lemma (Limiting radii) Consider the differential equations (B.16), ..., (B.18) for the partial backstresses, and let a function
1
n Cpw
[0 , T ], Vd B1 (0) ,
(0 0 < T < )
be given as driving input for them. Then here holds:
1 [ , ], V .
(a) Each i is in Cpw
0 T
d
(b) If the initial value satisfies ki (0 )k < ri (0 ), then there holds ki ()k < ri () in the
whole interval [0 , T ].
P
(c) Necessarily, there holds
()
Ii=1 ri () in the whole interval [0 , T ].
Proof: Obviously,
d ki k = d i : (i ),
i : (i ) = ki k.
We have
d
ki k
ri
!
ki k i +1
1
ki k
(B.17)
d ri
.
=
d ki k
= ci n : (i )
ri
ri
ri
(B.19),(CS)
1 + (1 + a )Qi + 2(ci 1) (1 + a )Qi + 1 =: Ki ,
i () =
ki ()k
ri ()
B miscellaneous
209
And now the same technique as in the proof lemma 6.3.5 can be applied to construct a supersolution.
It is essential for the radii ri and to remain positive for all times. The parameter condition
ring et al. [33, 34, 54] is by far not sufficient.
for the aij resp. the aj , given in Do
B.3.2 Lemma. Sufficient conditions on the model parameters for the radii to remain positive, i. e.
i
ri rmin
> 0 resp. min > 0, are given by
I
X
aij >
j=1
1
1,
i
r
I
X
resp.
aj >
j=1
rTi ()
rTi (0)
i
r
1+
I
X
j=1
aij
1
1.
(B.21)
> 1.
(B.21)
(B.22)
i
i
Set rmin
= Ji 1. Then rmin
> 0 and we find
d ri
(B.16)
= b rTi () ri
(B.22)
b Ji ri .
i
ri Ji eb Ji 1 rmin
For the yield surface radius , the proof is completely the same.
In the next lemma, we prove that the terminology target radii for the functions T and rTi
is in fact justified.
B.3.3 Lemma (Target radii) For each i = 1, . . . , I, we have got the following asymptotic behaviour
i
i
ri () rTi ()
0,
rTi () r
,
r i () r
.
(B.23)
for the limiting radii. The same asymptotics
() T ()
0,
T () ,
()
(B.24)
Proof: We just prove (B.23). Asymptotics (B.24) are then proven in exactly the same way.
As trivially
MTi := sup rTi < ,
[0 ,)
210
miscellaneous B
|ri ()| |ri (0 )|eb(0 ) + MTi 1 eb(0 ) |ri (0 )| + MTi =: M i < .
The variation of constants method for equation (B.16) yields the explicit solution
Z
b( )
eb(p) rTi (p) dp,
ri () = ri ( )e
+b
(B.25)
which holds for each , where 0 is fixed. Thus, we are able to give an upper
estimate of the error between the ith radius and its target radius function
Z
i
h
(B.25)(IP
)
ri () rTi ()
eb(p) dp rTi (p) dp
rTi ()
=
ri ( )eb( ) + rTi (p)eb(p)
Z
eb(p) dp rTi (p)dp,
ri ( ) rTi ( )eb( ) +
j=1
i
2r
I
X
j=1
|aij |bj
I
X
|aij |
dp
i
=
r
=: i ( ).
(1 + bj p)3
(1 + bj )2
j=1
Let > 0 be arbitrary. We now choose 0 large enough such that i ( ) < /2, and
set = ln /(2b(M i + MTi )). Thus we have |ri () rTi ()| < for all max{ , },
proving the first limit in (B.23). The second limit in (B.23) is trivial, from where the third
limit in (B.23) easily follows.
We would like to have an analogon to lemma 6.3.3, i. e. that the movement of i is directed
towards n. Unfortunately, this cannot be guaranteed in general. Multiplication of (B.17)
with n yields
!
i : n
ki k i +1
(i ) : n +
d ri
(B.26)
d i : n = ci ri 1
ri
ri
We discuss this, assuming that the parameters satisfy (B.21) and (B.20). As well, the premises
of lemma B.3.1 are assumed to hold.
First, if i : n 0, the first summand in (B.26) is non-negative due to the Cauchy-Schwarz
inequality, but can be arbitrarily close to zero, so that we have d i : n 0 only if d ri 0,
i. e. if ri rTi due to (B.16). Second, if i : n 0, we have d i : n 0, if
b
r i ri ,
b + ci T
(B.27)
as this implies
(CS)
0 d ri
i : n
1+
ri
(B.27)(B.16)
i : n
d ri d i : n
ci ri +
ri
in the case d ri 0. (In the opposite case the desired inequality holds trivially.) Due to
lemma B.3.3 there exists 0 such that (B.27) is satisfied for all . This implies the
intuitive movement.
B miscellaneous
B.4
211
Of course, not all questions could be answered by this thesis. As usual, new questions like
the following arose.
Application of AD techniques for parameter identification for nonlinear elastoplastic
BVP computations.
Further reduction of rejected steps through detecting the discontinuities at SeR and
SR , cf. section 7.2, with dense output methods for Runge-Kutta algorithms from Enright/Jackson [39, 40].
The proof of assumption 6.3.9. Derivation of error estimates for correction model for
Jiang material.
Proper limit corr {}, cf. section 3.2. Improvement of local error estimates for the
correction model.
The dynamic, viscoelastic, viscoplastic, nonlinear elastic case. The boundary value
problem with nonlinear geometry.
We hope that these questions will provide new open mathematical challenges and tasks for
the future.
212
miscellaneous B
Appendix C
Notations
Notations I. Throughout the thesis, we make frequently use of the following practical shortcuts.
(CR)
(CS)
(FT)
(HI)
(MI)
(IP)
(TE)
Chain rule
Cauchy-Schwarz inequality
Fundamental theorem
Holders inequality
Minkowskis inequality
Integration by parts
Taylor expansion theorem
a. e.
e.
LH
n
almost everywhere
everywhere
linear hull
limit n
transpose
complement
orthogonal complement
L(X, Y )
= L(X, R)
=
=
Second, let
C k ([0, T ], )
C ([0, T ], )
k ([0, T ], )
Cpw
=
=
=
where the respective image depends on the context. Third, let for a separable Hilbert space
X, the following notations hold for X-valued functions
M AP ([0, T ],
Lq ([0, T ],
BV ([0, T ],
CBV ([0, T ],
C([0, T ],
AC([0, T ],
W 1,q ([0, T ],
X)
X)
X)
X)
X)
X)
X)
=
=
=
=
=
=
=
P W L([0, T ], X)
214
notations C
Note that AC = W 1,1 holds, i. e. its elements are differentiable a. e. with L1 -integrable weak
derivative, and for which the fundamental theorem holds. Fourth, let
C k (,
C k (,
q
L (,
Lq (1 ,
W k,q (,
k,q
W0 (, 0 ,
Rn )
Rn )
Rn )
Rn )
Rn )
Rn )
=
=
=
=
=
=
space
space
space
space
space
space
of
of
of
of
of
of
Notations III. (a) We let and | | denote the standard Euklidian scalar product/norm on
Rn , i. e.
n
n
X
X
2
a2i
(a, b Rn ).
(C.1)
ai bi ,
kak =
ab =
i=1
i=1
We let further denote the standard dyadic product on Rn , its elements considered as column
vectors, i. e.
a b = a bt ,
(a b)ij = ai bj
(a, b Rn ).
2
(b) We let : and k k denote the standard Frobenius scalar product/norm on Rnn Rn ,
i. e.
n
n
X
X
a2ij
(A, B Rnn ). (C.2)
aij bij = tr(At B) = tr(B t A),
kAk2 =
A:B=
i,j=1
i,j=1
w C 1 (, Rn ),
v C 1 (, Rm ),
we define
div w C 0 (, R),
DIV A C 0 (, Rm ),
v C 0 (, Rmn )
by
div w(x) =
n
X
j=1
xj wj (x),
n
X
xj Aij (x),
DIV A(x) i =
j=1
v(x) ij = xj vi (x),
v 7
v/kvk
0
if v 6= 0
if v = 0
(C.3)
Bibliography
[1] Allgower E. L., Georg K.: Introduction to numerical continuation methods. Classics in applied mathematics. SIAM Philadelphia, 2003.
[2] Armstrong P. J., Frederick C. O.: A mathematical representation of the multiaxial Bauschinger effect. CEGB, report RD/B/N 731, 1966.
[3] Barkey M. E., Socie D. F., Hsia K. J.: A yield surface approach to the estimation of
notch strains for proportional and nonproportional cyclic loading. Journal of engineering
materials and technology, Vol. 116, No. 2, pp. 173-180, 1994.
tzle H., Kru
ger W., Maier B., Petersen J.: Mul[4] Beste A., Dressler K., Ko
tiaxial Rainflow: A consequent continuation of professor Tatsuo Endos work. The Rainflow Method in Fatigue (ed. by Y. Murakami), Oxford, 1992.
[5] Besseling J. F.: A theory of elastic, plastic and creep deformations of an initially
material showing anisotropic strain hardening, creep recovery and secondary creep.
Transactions of the ASME, Journal of applied mechanics, Vol. 25, pp. 529-563, 1959.
[6] Bonet J., Wood R. D.: Nonlinear continuum mechanics for finite element analysis.
Cambridge University Press, New York, 1997.
[7] Bower A. F.: Cyclic hardening properties of hard-drawn copper and rail steel. Journal
of the mechanics and physics of solids, Vol. 37, No. 4, pp. 455-470, 1989.
[8] Br
ezis H.: Operateurs maximaux et semi-groupes de contractions dans les espaces de
Hilbert. North-Holland Publishing, Amsterdam, 1973.
[9] Brokate M.: Elastoplastic constitutive laws of nonlinear kinematic hardening type.
Pitman research notes in mathematics series, Vol. 377, pp. 238-272, 1998.
[10] Brokate M.: Hysteresis operators. Phase transitions and hysteresis. Lecture notes in
mathematics, Vol. 1584, Springer, Berlin, pp. 1-38, 1994.
P.: Maximum norm wellposedness of nonlinear kinematic hard[11] Brokate M., Krejc
ening models. Continuum mechanics and thermodynamics, Vol. 9, pp. 365-380, 1997.
P.: On the wellposedness of the Chaboche model. International
[12] Brokate M., Krejc
series of numerical mathematics, Vol. 126, pp. 67-79, 1998.
P., Rachinskii D.: Some analytical properties of the multidi[13] Brokate M., Krejc
mensional continuous Mroz model of plasticity. Control and cybernetics, Polish academy
of sciences, Vol. 27, No. 2, pp. 199-215, 1998.
215
216
references
references
217
[29] Comi C., Maier G.: On the convergence of a backward difference iterative procedure
in elastoplasticity with nonlinear kinematic and isotropic hardening. Computational
plasticity: Models, software and applications, Pineridge Press, Swansea, pp. 323-334,
1989.
[30] Deimling K.: Nonlinear functional analysis. Springer, Berlin, 1985.
[31] Deuflhard P., Hohmann A.: Numerische Mathematik. Eine algorithmisch orientierte Einf
uhrung. DeGruyter, Berlin, 1993.
[32] Dormand J. R., Prince P. J.: A family of embedded Runge-Kutta formulae. Journal
of computational and applied mathematics, Vol. 6, No. 1, pp. 19-26, 1980.
ring R., Hoffmeyer J., Seeger T., Vormwald M.: A plasticity model for cal[33] Do
culating stress-strain sequences under multiaxial nonproportional cyclic loading. Computational materials science, Vol. 28, pp. 587-596, 2003.
ring R.: Zum Deformations- und Schadigungsverhalten metallischer Werkstoffe
[34] Do
unter mehrachsig nichtproportionalen zyklischen Beanspruchungen. Dissertation, TU
Darmstadt, 2006.
[35] Dressler K., Hack M.: Fatigue lifetime estimation based on rainflow counted data
using the local strain approach. European Journal of Mechanics, A/Solids, Vol. 15, No.
6, pp. 955-968, 1996.
ger W., Beste A.: Rainflow - das Werkzeug f
[36] Dressler K., Kru
ur den Lebensdauernachweis von Fahrzeugen, Bauteillebensdauer - Rechnung und Versuch. 19. Vortragsveranstaltung des DVM-Arbeitskreises Betriebsfestigkeit, M
unchen, 1993.
ger, W.: The multiaxial rainflow method. Low
[37] Dressler, K., Carmine, R., Kru
cycle fatigue and elasto-plastic behaviour of materials (ed. by K. T. Rie). Elsevier
Science Publishing, London, 1992.
[38] Duvaut G., Lions J. L.: Inequalities in mechanics and physics. Die Grundlehren der
mathematischen Wissenschaften, Vol. 219, Springer, 1976.
[39] Enright W. H., Jackson K. R.: Interpolants for Runge-Kutta formulas. ACM transactions on mathematical software, Vol. 12, No. 3, pp. 193-218, 1986.
[40] Enright W. H., Jackson K. R.: Effective solution of discontinuous IVPs using a
Runge-Kutta formula pair with interpolants. Applied mathematics and computation,
Vol. 27, pp. 313-335, 1988.
[41] Evans L. C.: Partial differential equations. Graduate studies in mathematics, Vol. 19,
American mathematical society, 1998.
[42] Filippov A. F.: Differential equations with discontinuous righthand sides. Kluwer
Academic Publishers, 1988.
[43] Garud Y. S.: A new approach to the evaluation of fatigue under multiaxial loading.
Transactions of the ASME, Vol. 103, pp. 118-125, 1981.
218
references
[44] Glinka G., Bucynski A., Ruggeri A.: Elastic-plastic stress-strain analysis of
notches under non-proportional loading paths. Archives of mechanics, Vol. 52, No. 4-5,
pp. 589-607, 2000.
[45] Glinka G., Molski K.: A method of elastic-plastic stress and strain calculation at a
notch root. Materials science and engineering, Vol. 50, pp. 93-100, 1981.
[46] Griewank A.: Evaluating derivatives. Principles and techniques of automatic differentiation. Society for industrial and applied mathematics. Frontiers in applied mathematics, Vol. 19, 2000.
ger K.: Initial-value problems for elastoplastic and elasto-viscoplastic systems. In:
[47] Gro
Fucik S. and Kufner A. [Eds.], Nonlinear analysis. Function spaces and applications,
Teubner, Leipzig, pp. 95-127, 1979.
[48] Hack M.: Schadigungsbasierte Hysteresefilter. Dissertation, TU Kaiserslautern, D386,
Shaker Aachen, 1998.
[49] Han W., Reddy R.B.: Plasticity. Mathematical theory and numerical analysis. Interdisciplinary applied mathematics 9, Springer, New-York, 1999.
[50] Han W., Reddy R.B.: Computational plasticity: The variational basis and numerical
analysis. Computational mechanics advances, Vol. 2, pp. 283-400, 1995
[51] Hairer E., Noerset S.P., Wanner G.: Solving ordinary differential equations I:
nonstiff problems. Springer, Berlin, Heidelberg, 1993.
[52] Hairer E., Wanner G.: Solving ordinary differential equations II: stiff and
differential-algebraic problems. Springer, Berlin, Heidelberg, 1991.
[53] Hertel O.: Weiterentwicklung und Verifizierung eines Naherungsverfahrens zur
Berechnung von Kerbbeanspruchungen bei mehrachsiger nichtproportionaler Schwingbelastung. Masters thesis, Bauhaus-Universitat Weimar, 2003.
[54] Herz E., Thumser R., Bergmann J. W., Vormwald M.: Endurance limit of autofrettaged Diesel-engine injection tubes with defects. Engineering fracture mechanics,
Vol. 73, pp. 3-21, 2006.
[55] Heuser H.: Funktionalanalysis. Theorie und Anwendung. Teubner Stuttgart, 1992.
[56] Hill R.: The mathematical theory of plasticity. Oxford classical texts in physical
sciences. Claredon press, Oxford, 2004.
nlich R.: On simultaneous torsion and tension of a circular cylindrical bar consist[57] Hu
ing of an elastoplastic material with linear hardening. Journal of applied mathematics
and mechanics, Vol. 58, pp. 509-516, 1979.
[58] Hiriart-Urruty J. B., Lemar
echal C.: Fundamentals of convex analysis. Springer,
Berlin, 2004.
[59] Hoffmann M., Seeger T.: A generalized method for estimating multiaxial elasticplastic notch stresses and strains. Part 1: Theory. Transactions of the ASME, Vol. 107,
pp. 250-254, 1985.
references
219
[60] Hoffmann M., Seeger T.: A generalized method for estimating multiaxial elasticplastic notch stresses and strains. Part 2: Application and general discussion. Transactions of the ASME, Vol. 107, pp. 255-260, 1985.
[61] Hoffmann M., Seeger T.: Mehrachsige Kerbbeanspruchungen bei proportionaler
Belastung. Konstruktion, Vol. 38, H.2, pp. 63-70, 1986.
ring R., Schliebner R., Vormwald M., Seeger T.: Lebens[62] Hoffmeyer J., Do
dauervorhersage f
ur mehrachsig nichtproportional schwingbeanspruchte Werkstoffe mit
Hilfe des Kurzrissfortschrittkonzeptes. FD-2/2000, FG Werkstoffmechanik, TU Darmstadt, 2000.
[63] Jiang Y.: Cyclic plasticity with an emphasis on ratchetting. PhD thesis, University of
Illinois, Urbana-Champaign, 1993.
[64] Jiang Y., Sehitoglu H.: Modeling of cyclic ratchetting plasticity, part I: Development of constitutive relations. Transactions of the ASME, Vol. 63, pp. 720-725, 1996.
[65] Jiang Y., Sehitoglu H.: Modeling of cyclic ratchetting plasticity, part II: Comparison of model simulations with experiments. Transactions of the ASME, Vol. 63, pp.
726-733, 1996.
[66] Jiang Y., Sehitoglu H.: Cyclic ratchetting of 1070 steel under multiaxial stress
states. International journal of plasticity, Vol. 10, No. 5, pp. 579-608, 1994.
[67] Jiang Y., Sehitoglu H.: Multiaxial cyclic ratchetting under multiple step loading.
International journal of plasticity, Vol. 10, No. 8, pp. 849-870, 1994.
[68] Jiang W.: A new constitutive model in the theory of plasticity. Part I: Theory. Journal
of pressure vessel technology, Vol. 117, pp. 365-370, 1995.
[69] Jiang W.: A new constitutive model in the theory of plasticity. Part II: Examples.
Journal of pressure vessel technology, Vol. 117, pp. 371-377, 1995.
[70] Kaliszky S.: Plastizitatstheorie. VDI-Verlag, D
usseldorf, 1984.
[71] Kardestuncer H.: Finite element handbook. McGraw-Hill, New York, 1987.
ttgen V. B., Barkey M. E., Socie D. F.: Pseudo stress and pseudo strain based
[72] Ko
approaches to multiaxial notch analysis. Fatigue and fracture of engineering materials
and structures, Vol. 18, No. 9, pp. 981-1006, 1995.
ttgen V. B., Barkey M. E., Socie D. F.: Structural stress-strain analysis for
[73] Ko
non-proportional loading suitable for FEM-postprocessing. International proceedings in
fatigue design, Helsinki, 1995.
[74] Krasnoselskii M. A., Pokrovskii A. V.: Systems with hysteresis. Nauka, Moscow,
1983.
P.: Hysteresis, convexity and dissipation in hyperbolic equations. Gakuto In[75] Krejc
ternational series, Mathematical sciences and applications, Vol. 8, Gakkotsho, Tokyo,
1996.
220
references
references
221
[93] Ohno N., Wang J. D.: Kinematic hardening rules with critical state of dynamical
recovery. Part II. International journal of plasticity, Vol. 9, No. 3, pp. 391-403, 1993.
lya G.: Liegt die Stelle der grossten Beanspruchung an der Oberflache? Journal of
[94] Po
applied mathematics and mechanics, Vol. 10, pp. 353-360, 1930.
[95] Prager W.: Recent developments in the mathematical theory of plasticity. Journal of
applied physics, Vol. 20, pp. 235-241, 1949.
[96] Prager W.: The theory of plasticity: A survey of recent achievements. James Clayton
lecture. Proceedings of the institution of mechanical engineers, Vol. 169, pp. 41-57, 1955.
[97] Rall L. B.: Automatic differentiation. Techniques and applications. Lecture notes in
computer sciences. Springer, Berlin, 1981.
[98] Reismann H., Pawlik P. S.: Elasticity. Theory and applications. Wiley, New York,
1980.
[99] Rich L. C., Hill D. R.: Automatic differentiation in Matlab. Applied numerical
mathematics, Vol. 9, pp. 33-43, 1992.
[100] Rockafellar R. T.: Convex analysis. Princeton university press, Princeton, New
Jersey, 1970.
[101] Rychlik I.: A new definition of the rainflow cycle counting method. International
journal of fatigue, Vol. 9, No. 2, pp. 119-121, 1987.
[102] Shampine L. F., Reichelt M. W.: The Matlab ODE suite. SIAM journal on scientific
computing, Vol. 18, No. 1, 1997.
[103] Simo J. C.: Topics on the numerical analysis and simulation of plasticity. Handbook
of numerical analysis, Vol. VI, Numerical methods for solids (pt. 3), Edited by Ciarlet
and Lions, Elsevier Science B.V., North Holland, pp. 183-499, 1998.
[104] Singh M. N. K., Glinka G., Dubey R. N.: Elastic plastic stress-strain calculation in
notched bodies subjected to non-proportional loading. International journal of fracture,
Vol. 76, pp. 39-60, 1996.
[105] Smith K.N., Watson P., Topper T.H.: A stress-strain function for the fatigue of
metals. Journal of materials, Vol. 5, No. 4, 767-778, 1970.
[106] Tanaka E.: A nonproportionality parameter and a cyclic viscoplastic constitutive
model taking into account amplitude dependencies and memory effects in isotropic
hardening. European journal of mechanics. A/Solids, Vol. 19, pp. 173-188, 1994.
[107] Temam R.: Mathematical problems in plasticity. Gauthier-Villars, Paris, 1985.
[108] Visintin A.: Differential models of hysteresis. Applied Mathematical Sciences, Vol.
111, Spinger, Berlin, 1994.
[109] Wheeler L. T.: Maximum principles in classical elasticity. In: Mathematical problems
in elasticity (ed. by Russo R.), World scientific publishing, Singapore, pp. 157-185, 1996.
222
references
[110] Zeidler E.: Nonlinear functional analysis and its applications II/A. Linear monotone
operators. Springer, New York, 1990.
[111] Zeidler E.: Nonlinear functional analysis and its applications II/B. Nonlinear monotone operators. Springer, New York, 1990.
[112] Zeidler E.: Nonlinear functional analysis and its applications III. Variational methods
and optimization. Springer, New York, 1985.
[113] Zeidler E.: Nonlinear functional analysis and its applications IV. Applications to
mathematical physics. Spinger Verlag, New York, 1988.
[114] Ziegler H.: A modification of Pragers hardening rule. Quarterly of applied mathematics, Vol. 17, pp. 55-65, 1959.
[115] Zienkiewicz O. C., Taylor R. L.: The finite element method. Volume 2: Solid
mechanics. Butterworth-Heinemann, Fifth edition, Burlington, 2003.
Thank you!
Thanks to the Deutsche Forschungsgemeinschaft, the Graduiertenkolleg Mathematik und Praxis, TU Kaiserslautern, for the financial support, the Fraunhofer
Gesellschaft for the technical and partial financial support, professors Rene Pinnau, TU Kaiserslautern, and Martin Brokate, TU M
unchen, for scientific support.
Thanks to Olaf Hertel and Michael Vormwald, TU Darmstadt, for their kind
provision of experimental data.
Thanks to our whole MDF group, in particular to Albert, Fredrik, Ilker, Sabrina,
Sascha, Thomas and Thorsten for many cigarette/coffee/curry-sausage breaks and
donerler/kebablar, especially to Anja for postprocessing orthographic correction
and exceedingly to Gerd, Joachim, Michael and Klaus for this interesting mathematical subject and many fruitful discussions.
Further thanks to Andre, Andrea, Baschdi, Bolipoto, kabhi Carsten, Charly,
Ernst, Heinz, the Horts, Hwajeong, J
urgen, Judith, J
udith, K
uwi, the Langs,
Marco, Martin, the Michaels (D., G. and R.), Moni, nige Lijun, nige Lily, Peter,
Ratna, Ruthard, kabhi Scholli, Torsten and Waltrout.
Curriculum Vitae
Me
Name
Holger Lang
Parents
Birth
Education
1983 - 1987
1987 - 1996
1996 - 1997
1997 - 2002
2000 - 2002
Member of the Group for Differential Geometry and Topology at the UdS
2002 - 2004
2002
2003
2004 - now
2004 - now
2004 - now
Member of the department Mathematical Methods in Dynamics and Durability at the Fraunhofer ITWM Kaiserslautern
Teaching
1999 - 2004
Courses in Mathematics for coming teachers, engineers and computer scientists at the UdS
2002 - 2003
2003 - 2004
Teacher for Mathematics and Physics at the Stadtgartengymnasium Saarlouis, grammar school, Saarland
Publications Overview
Holger Lang
(A) Lang, H.: Die Flache von Costa, Hoffman und Meeks: Eine vollstandige, in den
R3 eingebettete Minimalflache von Geschlecht Eins, mit drei Enden und endlicher
Gaussscher Totalkr
ummung 12. Diploma thesis, Universitat des Saarlandes, 2002.
(B) Lang H., Dressler K., Pinnau R. : A multiaxial stress-strain correction scheme.
Report of the Technomathematics group, TU Kaiserslautern, No. 263, 2005.
(C) Lang H., Dressler K., Pinnau R.: Parameter optimization for a multiaxial stressstrain correction scheme. Report of the Technomathematics group, TU Kaiserslautern,
No. 264, 2005.
(D) Lang H.: Notes at the embeddedness of the minimal surface of Costa, Hoffman and
Meeks. Report on differential geometry and topology, TU Kaiserslautern/Universitat
des Saarlandes, 2005.
(E) Lang H., Dressler K., Pinnau R.: Lipschitz estimates for the stop and the play
operator. Report of the Technomathematics group, TU Kaiserslautern, No. 266, 2006.
(F) Lang H., Dressler K., Pinnau R., Speckert M.: Error estimates for quasistatic
global elastic correction and linear kinematic hardening material. Report of the Technomathematics group, TU Kaiserslautern, No. 267, 2006.
(G) Lang H., Dressler K., Pinnau R., Bitsch G.: A homotopy between the solutions
of the elastic and the elastoplastic boundary value problem. Report of the Technomathematics group, TU Kaiserslautern, No. 268, 2006.
(H) Lang H., Bitsch G., Dressler K., Speckert M.: Comparison of the solutions of
the elastic and elastoplastic boundary value problems. Report of the Fraunhofer ITWM,
No. 99, 2006.
Publications (B), (C), (E), (F), (G) and (H) are related to this work.
Sworn Declaration
Hereby, I declare on oath that
I produced this work on my own, only using the devices and the literature
that are cited.
neither this work nor a similar work has been submitted as an examination
thesis for any state or academic exam.
Michelbach, 3rd of May 2007
Holger Lang