This action might not be possible to undo. Are you sure you want to continue?
XI International Conference on Computational Plasticity.
Fundamentals and Applications
COMPLAS 2011
E. O˜ nate and D.R.J. Owen (Eds)
MULTISCALE FAILURE FOR HETEROGENEOUS
MATERIALS: LINK WITH MORPHOLOGICAL MODELING
– COMPLAS XI
E. Roubin, M. Bogdan and J.B. Colliat
LMTCachan
ENSCachan, Paris 6 University, CNRS, UniverSud Paris PRES
61 avenue du Pr´esident Wilson, 94230 Cachan Cedex, France
email: {roubin,colliat}@lmt.enscachan.fr, http://www.lmt.enscachan.fr/
Key words: Heterogeneous materials, FE method, Random ﬁelds, Excursion set, Em
bedded discontinuity
Abstract. A 3D mesoscale model for failure of heterogeneous quasibrittle materials is
presented. At such scale, concrete can be represented as an heterogenous material with
two phases, where aggregates are included within the concrete. The model problem of
heterogeneous materials that is adressed in detail here is based, on the one hand, on
FE models with embedded discontinuities and, on the other hand, on a morphological
representation using Gaussian or Gaussian related random ﬁeld excursion sets.
1 Introduction
In view of the growing complexity of macroscopic models of concrete like materials, the
question of multiscale observation became relevant. It clearly appears that macroscopic
behaviours of such material (cracking, creep. . . ) take their origin at smaller scales (meso
scopic, microscopic. . . ). The framework presented here is to be seen in this context, and
especially in a sequenced way (as opposed to integrated one [1]) where the macroscopic
behaviour comes from a mesoscopic description of the material. At this particular scale,
concrete must be represented as heterogenous materials. Therefore, both mechanical and
geometrical properties have to be represented by the framework.
This communication ﬁrst present a morphological modeling framework for heterogenous
materials. A concrete like material described as a twophase material is considered here,
where inclusions (aggregates) are included within a matrix (cement past and sand). The
idea behind this morphological model is to yield the phases from random ﬁeld excursion
sets. Moreover, adding more phases in order to extend possibilities of representation
is possible by adding excursion sets. If the framework deals with correlated Gaussian
or Gaussian related random ﬁeld (such as the chisquare distribution  χ
2
), an analytic
1
E. Roubin, M. Bogdan and J.B. Colliat
formulae links the random ﬁeld characteristics with geometrical and topological quantities
(volume, surface area, Euler Characteristic...) of the underlying excursion set. This
link has been recently made in [2] giving the possibility of controlling the excursion set
characteristics and applying it to represent material phase with chosen characteristics.
For a realistic modeling of a concrete like material, both in term of geometrical and
topological quantities, due the Gaussian case limitation, an application of the χ
2
random
ﬁeld is made. Both unidimensional KarhunenLo`eve decomposition and turningbands
projectional method are used to simulate three dimensional discrete correlated Gaussian
random ﬁelds.
Eﬀorts of morphological modeling are here made within a multiscale linear frame
work using a FE model with embedded discondinuities [3]. In order to represent these
heterogeneities, those excursions are projected onto the FE mesh, thus deﬁning a set
of discontinuities within the strain ﬁeld interpolation (weak discontinuities [4]). These
kinematics enhancements lead to ”nonadapted” meshes in the sense of independence be
tween heterogeneities morphology and the underlying FE mesh. Application of this linear
implementation is made for a simple hydration process model presented here.
Considering the non linear failure behaviour, weak discontinuities are completed with
a set of strong (displacement ﬁeld) discontinuities within the framework of local enhance
ment [5]. Those discontinuities allow for a simple and accurate representation of the
mesoscale cracks. The macroscopic response of this model is shown for a simple tension
test.
2 Random ﬁeld generation
As the whole morphological framework is based on Gaussian (or Gaussian related)
correlated random ﬁeld, eﬀorts have to be made in the numerical implementation of their
generation. This part explains two methods used to generate realisations of such ﬁelds.
First the KarhunenLo`eve decomposition [6] and then the turning bands projection [7].
Through this paper, we shall call γ(x, w) a Gaussian random ﬁeld over a parameter space
M (which shall always be taken here to be a bounded region of R
N
) which takes values
in R. It is assumed that γ has mean zero, variance σ
2
and is isotropic and stationnary
with a Gaussian covariance function deﬁned as C(x, y) = C(x − y) = E{γ(x)γ(y)} =
σ
2
e
−x−y/Lc
where L
c
is the correlation length.
The orthogonal decomposition of Gaussian correlated random ﬁelds theory stipulates
[8] that mean zero Gaussian ﬁeld with continuous covariance function (such as C) can be
written as follows
γ(x, w) =
∞
n=1
ϕ
n
(x)ξ
n
(w), (1)
where ξ
n
(w) are zero mean, unit variance Gaussian random variables, and ϕ
n
(x) are
functions on M determined by the covariance function C. It is worth noting that eq.(1)
2
E. Roubin, M. Bogdan and J.B. Colliat
allows for stochastic  w  and spatial x  variables separation. Therefore, implementing
this framework comes to put the eﬀort in the determination of the spatial functions ϕ
n
(x).
The KarhunenLo`eve decomposition is based on the previous orthogonal decomposi
tion. It allows us to determine these spatial functions ϕ
n
(x) for simple compact M in
R
N
. Demonstration can be found in [9] that they can be determined by ﬁrst solving the
following eigenvalues problem (known as Fredholm problem):
_
M
C(x, y)ψ(y)dy = λψ(x) (2)
where λ and ψ are respectively the eigenvalues and eigenvector and then by setting
ϕ
n
(x) =
√
λ
n
ψ
n
(x). Theoretically, an inﬁnite sum is needed to deﬁne the exact random
ﬁeld in eq.(1). For the numerical implementation made here, a Finite Element Method is
used to solve a discretized Fredholm problem. Therefore, using a ﬁnite set of eigenvalues
and eigenvectors, the following troncated KarhunenLo`eve decomposition eq.(3) deﬁnes
an approximative realization of the underlying random ﬁeld.
γ(x, w) =
m
n=1
_
λ
n
ξ
n
(w)ψ
n
(x). (3)
The fact that stochastic and spatial variables are still separated is an essential result
for any numerical implementation. Indeed, once the m couples {λ
n
; ψ
n
} of a certain
correlated random ﬁeld are determined, the generation of a realization comes to generate
a set of independent Gaussian variables (which only requires a random number generator).
Moreover, the same couples can be used to produce any other realizations of the same
ﬁeld.
The precision of this method, involving full squared matrix eigenvalues problem, is
quickly limitated by the memory storage when one deals with multidimensional random
ﬁelds of large size. The turning bands projectional method has been developped by Math
eron [7] in order to reduce the amount of numerical ressources. The idea is to generate
several onedimensional realizations of random ﬁelds to produce a multidimensional one.
The algorithm below explains this projectional method with details.
Let M be the discreted multidimensional bounded region where the ﬁnal realization
will be created. Several lines have to be generated (we shall call L their number) with
one arbitrary intersection point 0 and an uniform distribution of directions over the unit
ball (see Fig.1).
Let z(ζ, w
i
), i = 1..L be the L realizations of a onedimensional correlated random ﬁeld
generated over the L lines. For each point N on M, the value of the multidimensional
realization is the average of the onedimensional realization values at the projection of N
on each line i:
γ(N, w) =
1
√
L
L
i=1
z(ζ
N
i
, w
i
) (4)
3
E. Roubin, M. Bogdan and J.B. Colliat
0
z(ζ
Ni
)
line i
u
i
N
M
N
i
ζ
i
x
Ni
Figure 1: Schematic representation of the turning band method (from [10])
In this paper, the application of the method is made for threedimensional random
ﬁelds. The key of this method is the link between the threedimensional covariance func
tion C and the equivalent onedimensional covariance function C
1
we need to generate
the L realizations. Let C(r) be as above (with r = x −y). Following [7] we have
C
1
(r) =
d
dr
(rC(r)) = σ
2
_
1 −
2r
2
L
2
c
_
e
−r
2
/L
2
c
(5)
3 Excursion Set
We call an excursion set the morphology of a subset of a bounded region deﬁned by
thresholding a realization of a random ﬁeld. It allows us to create a set of random shapes.
Let γ be a realization of γ(x, w) : M ⊂ R
N
→ R deﬁne as above and u ∈ R a chosen
threshold. The underlying excursion set A
u
is deﬁned by the points of M where the values
of γ are above u (eq.(6)).
A
u
≡ A
u
(γ, M) {x ∈ M : γ(x) ≥ u} (6)
This principle, applied for M ∈ R is shown on Fig.2.
In our case, random ﬁelds will be yield in a three dimensional space (M ⊂ R
3
) and
therefore deﬁne threedimensional excursion sets. The two excursions represented in Fig.3
are made from the same realization with two diﬀerent threshold values. It is clear that, by
changing this value, a large range of varied morphologies can be generated. This exemple
shows that “low” values of u produce excursions mainly made of handles with high volume
fraction, giving a “sponge” like topology (Fig.3(a)), whereas “high” values of u produce
excursion made of several connected components with a lower volume fraction (Fig.3(b)).
In order to provide a global description of the resulting morphology, the Lipschitz
Killing curvatures, hereafter LKCs, are choosen. In a Ndimensional space N + 1 LKCs
can be deﬁned where each can be thought of measures of the ”jdimensional sizes” of
4
E. Roubin, M. Bogdan and J.B. Colliat
γ
x
u
A
u M
Figure 2: Schematic representation of a onedimensional excursion set A
u
(a) “Low” threshold  sponge
topology
(b) “High” threshold  meatball
topology
Figure 3: Eﬀect of threshold value on tridimensional excursion topology
A
u
. In our threedimensional case, the four LKCs, denoted by L
j
, j = 0..3, provide both
geometrical  L
1
, L
2
, L
3
 and topological  L
0
 descriptions of the morphology A
u
. They
are deﬁned by:
 L
3
(A
u
) is the three dimensional volume of A
u
.
 L
2
(A
u
) is half the surface area of A
u
.
 L
1
(A
u
) is twice the caliper diameter of A
u
.
 L
0
(A
u
) is the Euler characteristic of A
u
, which contrary to the other LKCs is a
topological measure. In threedimension, it can be calculated by:
L
0
(A
u
) = #{connected components in A
u
}−#{“handles” in A
u
}+#{“holes” in A
u
}
For exemple, a ball or a cube are topologicaly identical (Euler characteristic L
0
= 1)
but diﬀer from a hollow ball (L
0
= 2) or a ring torus (L
0
= 0).
5
E. Roubin, M. Bogdan and J.B. Colliat
Following [2], a probabilistic link has been made between excursion set properties and
random ﬁeld thresholding parameters giving an explicit formulae for the expectation of
the LKCs  E{L
i
(A
u
(γ, M))}. It is not the purpose of this paper to give details on these
formulae, however, full proof and details can be found in [9]. The only idea one need
to remember to go through this paper is that this theory gives a new tool helping us to
predict all the geometrical and topological properties of an excursion set from the random
ﬁeld characteristics and the threshold  σ, L
c
, u . These relations have been made explicit
for γ(x, w) as above on a cube M =
3
i=1
[0; T]:
_
¸
¸
¸
¸
_
¸
¸
¸
¸
_
E{L
0
}(A
u
) =
_
√
2
2π
2
T
3
L
3
c
_
u
2
σ
2
−1
_
+
3
√
2
2π
3/2
T
2
L
2
c
u
σ
+
3
√
2
2π
T
Lc
_
e
−u
2
/2L
2
c
+ Ψ
_
u
σ
_
E{L
1
}(A
u
) =
_
√
2
π
3/2
T
3
L
2
c
u
σ
+
3
√
2
4
T
2
Lc
_
e
−u
2
/2L
2
c
+ 3TΨ
_
u
σ
_
E{L
2
}(A
u
) =
√
2
π
T
3
Lc
e
−u
2
/2L
2
c
+ 3T
2
Ψ
_
u
σ
_
E{L
3
}(A
u
) = T
3
Ψ
_
u
σ
_
(7)
Fig.4(a) and Fig.4(b) represent respectively the Euler characteristic and the volume
fraction  directly linked with the fourth LKC by E{L
3
}(A
u
)/T
3
 of excursion sets of
γ(x, w) for u from −20 to 20.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
20 15 10 5 0 5 10 15 20
V
o
l
u
m
e
f
r
a
c
t
i
o
n
[

]
Threshold []
(a)
80
60
40
20
0
20
40
60
20 15 10 5 0 5 10 15 20
E
u
l
e
r
C
a
r
a
c
t
e
r
i
s
t
i
c
[

]
Threshold []
(b)
Figure 4: LKCs of excursion sets of Gaussian random ﬁeld in term of threshold values.
Expected values of LKCs provided by (7), + Numerical values calculated from one
realization of γ(x, w).
The constant decreasing shape of the volume fraction curve in term of u clearly reﬂects
the eﬀect of the threshold level on the “size” of A
u
. Even if more peculiar, the Euler
characteristic curve shape reﬂects also easily the eﬀect of the threshold on excursion sets
topology. For values of u lower than the lowest value of γ, the Euler characteristic is
the one of the full cube (L
0
= 1). By increasing u, several holes appear, counting in
positive for the Euler characteristic (L
0
> 1). Then, the expansion of the holes starts to
form handles which lead to a sponge like topology (L
0
< 0). By increasing u even more,
6
E. Roubin, M. Bogdan and J.B. Colliat
handles disappear forming a “meatball” like topology of connected components (L
0
> 0).
Finally, the Euler characteristic decreases to L
0
= 0 when no more connected components
remain.
From the comparison between theorical values and measures on one realization, we can
point out that the variability of the numerical generation is very low. Therefore, although
eq.(7) gives only expectations of LKCs, for this range of excursion sets we can assume
that V{L
i
(A
u
)} ≪1.
So far, we have seen the eﬀect of the threshold value on excursion sets. But one needs
to remember that, according to eq.(7), both variance and covariance length of γ(x, w)
aﬀect the morphology as well. Understanding the full behaviour of these equations is a
key point for anyone who wants to make excursion set modeling.
4 Application of the modeling framework on concrete like material
The material is represented as an heterogeneous material with two phases. One phase
(aggregates) is represented by an excursion set of a correlated random ﬁeld while its second
phase (concrete) is represented by its complementary. Therefore in this part, the eﬀort
will be put in a “realistic” representation of the aggregates phase. We keep only three
relevant characteristics from the four LKCs: the volume fraction V
v
, the volumic surface
area S and the number of agregates N which are respetively linked with L
3
, L
2
and L
0
.
Thought V
v
and S can be directly estimated, attention must be taken when it comes to
N. Indeed the Euler characteristic does not indicate the number of aggregates for every
topology. In our case, the “meatball” topology has to be targeted and it is only once we
assume that the excursion set is free from holes and handles that N can be estimated by
L
0
. In this speciﬁc kind of topology: N #{connected components} = L
0
.
Once the three characteristics (N, S, V
v
) of the phase are set, the generation of the
underlying excursion set rely on ﬁnding a solution for (u, σ, L
c
) that satisfy the following
system:
_
_
_
E{L
3
}(u, σ) = V
v
T
3
E{L
2
}(u, σ, L
c
) =
1
2
ST
3
E{L
0
}(u, σ, L
c
) = N
(8)
Due to the intrinsic non linearity of eq.(7), depending on the diﬀerent values of (N, S, V
v
)
(especially for “meatball” topology  N ≫ 1) the problem eq.(8) do not always have a
solution. For exemple, we can clearly see on Fig.4 that we can not expect N to be up
per than 40 while keeping a “high” volume fraction (V
v
> 40%). Which in our case of
concrete like material modeling leads to a major issue. So far, the more realistic solution
for “meatball” topology we get with this framework allows us to represent an aggregate
phase with a maximum of 15% volume fraction.
Until now, the framewok has been presented considering Gaussian random ﬁelds. But
estimation of LKCs for excursion set can also be worked out considering Gaussian related
ﬁelds. The application of this paper is made using a chisquare distribution with k degrees
7
E. Roubin, M. Bogdan and J.B. Colliat
of freedom  χ
2
k
. Realizations of such ﬁelds can be seen as sum of k independent squared
realizations of a correlated Gaussian random ﬁeld. Let δ be a realization of such ﬁeld and
γ
i
, i = 1..k be k realizations of the Gaussian ﬁeld γ(x, w) described above. We have :
δ =
k
i=1
γ
2
i
(9)
Although similar to eq.(7), the use of a χ
2
k
distribution add the parameter k to the
system eq.(8). With such ﬁeld, the nearest solution is found for k = 1 and enable us to
double the previous volume fraction V
vmax
≈ 30%. Fig.5 shows a twodimensional slice of
excursions from a Gaussian realization and a χ
2
1
made from the same realization. Fig.5(b),
being the excursion from the squared realization of the excursion Fig.5(a), shows clearly
that, for the same threshold, it is natural to expect the volume fraction to double between
excursions of Gaussian and χ
2
1
random ﬁelds.
(a) Gaussian realization  γ (b) χ
2
1
realization  δ = γ
2
Figure 5: Comparison between Gaussian and χ
2
1
excursion sets for the same threshold value.
The χ
2
1
distribution remains the more suitable solution for meatball topology and high
volume fraction morphology we found.
5 FE model for heterogeneous material  Application to hydration process
modeling
The approach made here relies on a spatial truss, to model pattern of heterogeneities.
The choice of a not adapted meshing process is made here thus, the spatial positions of
nodes are not constrained by the morphology. Therefore, both gemetrical and mechanical
properties have to be handle inside some interface elements. These cut elements are split
into two parts, each having diﬀerent elastic properties by enhancing them with strain
(weak) discontinuities [11]. An elementary enhancements method (EFEM) method for
kinematic enhancement of Finite Element using the HuWashizu variational formulation is
used here. For example, if we consider a twophase material (inclusions within a matrix),
8
E. Roubin, M. Bogdan and J.B. Colliat
three sets of elements are needed: those entirely in the matrix, those entirely in the
inclusions, and those which are split between both (cut elements). To calculate these
elements repartition, a projection of the previous excursion set is made onto the truss.
In order to illustrate this linear framework, a simple hydration process of concrete like
material modeling has been implemented. Considering a simplistic version of the Powers
and Brownyard hydration model [12], with only three phases: unreacted cement, hydration
products (including gel water) and free water, the volume fraction of each one of them
can be calculated according to the following equations:
_
¸
¸
_
¸
¸
_
p =
w/c
w/c+ρw/ρc
V
anh
= (1 −p)(1 −α)
V
h
= 2.12(1 −p)α
V
w
= 1 −V
h
−V
anh
(10)
where p is the initial porosity, α the hydration degree and V
anh
, V
h
, V
w
respectively the
volume fractions of anhydrous cement, hydration products and water.
(a) α = 0.1 (b) α = 0.5 (c) α = 1
Figure 6: Projection of excusion set shapes on FE truss for diﬀerent hydration degrees.
water, hydration products, anydrous cement
As explained previously, thresholding a random ﬁeld with a scalar allows to create a
two phase material. One can easily imagine, that a second threshold, with a diﬀerent
value, will allow to create an additional phase, concentrical to the ﬁrst one. Therefore,
setting two thresholds will allow us to create a three phase material. Thus, for diﬀerent
hydration degrees, each phase’s volume fraction is known and can be linked to the random
ﬁeld’s thresholds u
i
(equation eq.(7)). Eventually, the initial morphology is set up by one
threshold (two phases: water and unhydrated cement), and then, for a growing hydration
degree, two thresholds are calculated and applied to the random ﬁeld, creating a three
phase material (water, unhydrated cement and hydration products).
Within this framework, macroscopic material characteristics like Young modulus can
be estimated over a given hydration degree with simple tension tests. The following
9
E. Roubin, M. Bogdan and J.B. Colliat
characteristics have been chosen E
anh
= 135 000 MPa, E
h
= 25 000 MPa and E
w
=
1 MPa.
0
5000
10000
15000
20000
25000
30000
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Y
o
u
n
g
m
o
d
u
l
u
s
[
M
P
a
]
Hydration degree []
Figure 7: Young modulus of a concrete like material for diﬀerent hydration degrees.
Fig.7 shows that the continuous growing of the macroscopic Young modulus over hy
dration degree is well handled by this FE representation. A slight raising of the slope can
be seen after α = 0.4.
6 FE models with embedded discontinuities
In addition to the geometrical representation of heterogeneities, displacement (strong)
discontinuities are also introduced in the elements, in order to model a nonlinear softening
response based on failure quasibrittle. These discontinuities represent microcracks that
can occurs in both phases as well as at the interfaces (debonding). Details of this FE
numerical implementation can be found in [3].
A other simple tension test is presented here. Material properties are deﬁned according
to Tab.1.
Table 1: Material properties
Matrix Inclusions Interface
E = 10GPa 70GPa −
σ
u
= 3MPa − 3MPa
Gf = 11J/m
2
− 11J/m
2
Two remarks are worthy of attention. The ﬁrst is that the interface is of rigidbrittle
type. The second is that we choosed for inclusions to remains in the linear elastic regime.
10
E. Roubin, M. Bogdan and J.B. Colliat
(a) Displacement ﬁeld and crack pat
tern at last time step
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04
L
o
a
d
[
k
N
]
Displacement [mm]
(b) Load vs imposed displacement
Figure 8: Results for simple tension
The cracking pattern is shown on Fig.8(a) where two zones are splited by a macroscopic
crack (represented by means of the broken elements). Fig.8(b) shows the macroscopic
load vs imposed displacement curve where three steps can be seen. First, a linear part
where no failure occurs. Then, with the apparition of several microscopic cracks, we can
observe a yield behaviour. Finaly, the softening part begin when the localisation of these
microscopic cracks creates a macrosopic one.
7 Concluding remarks
This communication presents a ﬁrst attempt to create a sequential multiscale frame
work where morphology of heterogeneous material is deﬁned by excursion sets of correlated
random ﬁelds. Though, eﬀorts still have to be made in order to generate more realistic
morphologies, advantages have been shown through two examples. We can also add that
this framework is well adapted to other problematics related with concrete like materials
such as the eﬀect of morphological variability on macroscopic behaviour. Indeed, the
use of both KarhunenLo`eve decomposition and nonadapted meshes allows fast compu
tations, limiting the growing amount of numerical ressources needed when dealing with
large sets of morphologies. Futhermore, being able to represent broken elements by means
of a strong discontinuity in the FE method allows calculations of permeability or diﬀusion
in such damaged materials [13].
REFERENCES
[1] Feyel, F. and Chaboche, J.L. Multiscale nonlinear FE analysis of composite struc
ture: damage and ﬁber size eﬀects. Revue europ´eenne des
´
El´ements Finis : NUM
DAM’00 isues (2001) 10:449–472
11
E. Roubin, M. Bogdan and J.B. Colliat
[2] Adler, R.J. Some new random ﬁeld tools for spatial analysis. Stochastic Environmen
tal Research and Risk Assessment (2008) 22:809–822
[3] Benkemoun, N., Hautefeuille, M., Colliat, J.B., Ibrahimbegovic, A. Failure of het
erogeneous materials: 3D mesoscale FR models with embedded discontinuities. Int
J. Num. Meth. in Engng. (2010) 82:1671–1688
[4] Sukumar, N., Chapp, D.L., Mo¨es, N. and Belytshcko, T. Modeling holes and in
clusions by level sets in the extended ﬁnite element method. Computer Methods in
Applied Mechanics and Engineering (2001) 190:6183–6200
[5] Oliver J. Modelling strong discontinuities in solid mechanics via strain softening
constitutive equations. International Journal for Numerical Methods in Engineering
(1996) 39:3575–3623
[6] Loeve, M. Probability Theory. Graduated Texts in Mathematics (1978) Vol. II, 4th
ed..
[7] Matheron, G. The intrinsic random functions and their applications. Advances in
Applied Probability (1973) 5:439–468
[8] Karhunen, K.
¨
Uber lineare Methoden in der Wahrscheinlichkeitsrechnung. Ann.
Acad. Sci. Fennicae. Ser. A. I. Math.Phys (1947) 37:1–79
[9] Adler, R.J. and Taylor, J.E. Random Fields and Geometry (2007) Springer, Boston
[10] Mantoglou, A. and Wilson, J.L The Turning Bands Method for Simulation of Ran
dom Fields using Line Generation with a Spectral Method Water Resources Research
(1982) Vol.II, 2:129–149
[11] Ortiz, M., Leroy, Y. and Needleman, A. A Finite Element method for localized failure
analysis Computer Methods in Applied Mechanics and Enginnering (1987) 61:189–
214
[12] Powers, T.C. and Brownyard, T.L. Studies of the physical properties of hardened
Portland cement paste. J. Am. Concr. Inst., (1947). 43101–132, 249–336, 469–505,
549–602, 669–712, 845–880, 933–992.
[13] Jourdain, X., Colliat, J.B., De Sa, C., Benboudjema, F. and Gatuingt, F. Upscal
ing permeability for fractured concrete : mesomacro numerical approach within a
sequential framework submitted
12
both in term of geometrical and topological quantities. This part explains two methods used to generate realisations of such ﬁelds. Considering the non linear failure behaviour. The macroscopic response of this model is shown for a simple tension test. In order to represent these heterogeneities. 2 Random ﬁeld generation As the whole morphological framework is based on Gaussian (or Gaussian related) correlated random ﬁeld. It is assumed that γ has mean zero. Colliat formulae links the random ﬁeld characteristics with geometrical and topological quantities (volume. those excursions are projected onto the FE mesh. This link has been recently made in [2] giving the possibility of controlling the excursion set characteristics and applying it to represent material phase with chosen characteristics. The orthogonal decomposition of Gaussian correlated random ﬁelds theory stipulates [8] that mean zero Gaussian ﬁeld with continuous covariance function (such as C) can be written as follows γ(x. y) = C( x − y ) = E{γ(x)γ(y)} = σ 2 e− x−y /Lc where Lc is the correlation length. e Through this paper. Both unidimensional KarhunenLo`ve decomposition and turningbands e projectional method are used to simulate three dimensional discrete correlated Gaussian random ﬁelds. due the Gaussian case limitation. (1) where ξn (w) are zero mean. an application of the χ2 random ﬁeld is made. variance σ 2 and is isotropic and stationnary with a Gaussian covariance function deﬁned as C(x. Those discontinuities allow for a simple and accurate representation of the mesoscale cracks. Euler Characteristic. surface area. For a realistic modeling of a concrete like material. It is worth noting that eq. w) = ∞ n=1 ϕn (x)ξn (w). Eﬀorts of morphological modeling are here made within a multiscale linear framework using a FE model with embedded discondinuities [3]. Bogdan and J.. and ϕn (x) are functions on M determined by the covariance function C. Application of this linear implementation is made for a simple hydration process model presented here. weak discontinuities are completed with a set of strong (displacement ﬁeld) discontinuities within the framework of local enhancement [5].E. we shall call γ(x.. w) a Gaussian random ﬁeld over a parameter space M (which shall always be taken here to be a bounded region of RN ) which takes values in R. M. thus deﬁning a set of discontinuities within the strain ﬁeld interpolation (weak discontinuities [4]). These kinematics enhancements lead to ”nonadapted” meshes in the sense of independence between heterogeneities morphology and the underlying FE mesh.(1) 2 . eﬀorts have to be made in the numerical implementation of their generation. unit variance Gaussian random variables. Roubin.) of the underlying excursion set. First the KarhunenLo`ve decomposition [6] and then the turning bands projection [7].B.
wi ). The turning bands projectional method has been developped by Matheron [7] in order to reduce the amount of numerical ressources. wi ) i=1 (4) 3 . using a ﬁnite set of eigenvalues and eigenvectors.(1). m γ(x.E. Colliat allows for stochastic . The algorithm below explains this projectional method with details. Indeed.L be the L realizations of a onedimensional correlated random ﬁeld generated over the L lines. Let z(ζ. Bogdan and J. the generation of a realization comes to generate a set of independent Gaussian variables (which only requires a random number generator).w . M. i = 1. Demonstration can be found in [9] that they can be determined by ﬁrst solving the following eigenvalues problem (known as Fredholm problem): C(x.B. once the m couples {λn . The KarhunenLo`ve decomposition is based on the previous orthogonal decomposie tion.1). Theoretically. For each point N on M. involving full squared matrix eigenvalues problem. a Finite Element Method is used to solve a discretized Fredholm problem. It allows us to determine these spatial functions ϕn (x) for simple compact M in RN . For the numerical implementation made here. the same couples can be used to produce any other realizations of the same ﬁeld. Therefore. The precision of this method. an inﬁnite sum is needed to deﬁne the exact random ﬁeld in eq. Therefore. Moreover. Several lines have to be generated (we shall call L their number) with one arbitrary intersection point 0 and an uniform distribution of directions over the unit ball (see Fig. w) = √ L L z(ζNi . The idea is to generate several onedimensional realizations of random ﬁelds to produce a multidimensional one.and spatial x . y)ψ(y)dy = λψ(x) M (2) where λ √ ψ are respectively the eigenvalues and eigenvector and then by setting and ϕn (x) = λn ψn (x).. (3) The fact that stochastic and spatial variables are still separated is an essential result for any numerical implementation. the value of the multidimensional realization is the average of the onedimensional realization values at the projection of N on each line i: 1 γ(N.(3) deﬁnes e an approximative realization of the underlying random ﬁeld. is quickly limitated by the memory storage when one deals with multidimensional random ﬁelds of large size.variables separation. w) = n=1 λn ξn (w)ψn (x). ψn } of a certain correlated random ﬁeld are determined. the following troncated KarhunenLo`ve decomposition eq. implementing this framework comes to put the eﬀort in the determination of the spatial functions ϕn (x). Roubin. Let M be the discreted multidimensional bounded region where the ﬁnal realization will be created.
applied for M ∈ R is shown on Fig. Let γ be a realization of γ(x.E.(6)). hereafter LKCs. This exemple shows that “low” values of u produce excursions mainly made of handles with high volume fraction. It allows us to create a set of random shapes. The underlying excursion set Au is deﬁned by the points of M where the values of γ are above u (eq. It is clear that. Following [7] we have C1 (r) = 3 Excursion Set 2r 2 d (rC(r)) = σ 2 1 − 2 dr Lc e−r 2 /L2 c (5) We call an excursion set the morphology of a subset of a bounded region deﬁned by thresholding a realization of a random ﬁeld. the LipschitzKilling curvatures. Colliat M N ui xNi Ni ζi line i z(ζNi ) 0 Figure 1: Schematic representation of the turning band method (from [10]) In this paper. M. are choosen. The two excursions represented in Fig. The key of this method is the link between the threedimensional covariance function C and the equivalent onedimensional covariance function C1 we need to generate the L realizations. In order to provide a global description of the resulting morphology. whereas “high” values of u produce excursion made of several connected components with a lower volume fraction (Fig. Roubin. M) {x ∈ M : γ(x) ≥ u} (6) This principle. random ﬁelds will be yield in a three dimensional space (M ⊂ R3 ) and therefore deﬁne threedimensional excursion sets.B. Bogdan and J. a large range of varied morphologies can be generated. Let C(r) be as above (with r = x − y ). by changing this value.3(a)).2. In our case. Au ≡ Au (γ.3 are made from the same realization with two diﬀerent threshold values. w) : M ⊂ RN → R deﬁne as above and u ∈ R a chosen threshold. the application of the method is made for threedimensional random ﬁelds. In a Ndimensional space N + 1 LKCs can be deﬁned where each can be thought of measures of the ”jdimensional sizes” of 4 . giving a “sponge” like topology (Fig.3(b)).
L1 (Au ) is twice the caliper diameter of Au .and topological .L2 (Au ) is half the surface area of Au .L0 (Au ) is the Euler characteristic of Au . They are deﬁned by: .3. Bogdan and J.E.L0 .descriptions of the morphology Au . denoted by Lj . provide both geometrical .sponge topology (b) “High” threshold . In our threedimensional case. the four LKCs. Colliat u γ Au M x Figure 2: Schematic representation of a onedimensional excursion set Au (a) “Low” threshold .. Roubin. L2 . M.B. j = 0. .meatball topology Figure 3: Eﬀect of threshold value on tridimensional excursion topology Au . . a ball or a cube are topologicaly identical (Euler characteristic L0 = 1) but diﬀer from a hollow ball (L0 = 2) or a ring torus (L0 = 0).L1 . In threedimension. . it can be calculated by: L0 (Au ) = #{connected components in Au }−#{“handles” in Au }+#{“holes” in Au } For exemple. 5 . which contrary to the other LKCs is a topological measure. L3 .L3 (Au ) is the three dimensional volume of Au .
a probabilistic link has been made between excursion set properties and random ﬁeld thresholding parameters giving an explicit formulae for the expectation of the LKCs .6 0. u . These relations have been made explicit for γ(x. w). Colliat Following [2]. For values of u lower than the lowest value of γ. Expected values of LKCs provided by (7). w) as above on a cube M = 3 [0. The constant decreasing shape of the volume fraction curve in term of u clearly reﬂects the eﬀect of the threshold level on the “size” of Au . By increasing u even more.2 60 0. the expansion of the holes starts to form handles which lead to a sponge like topology (L0 < 0). Lc . The only idea one need to remember to go through this paper is that this theory gives a new tool helping us to predict all the geometrical and topological properties of an excursion set from the random ﬁeld characteristics and the threshold .directly linked with the fourth LKC by E{L3 }(Au )/T 3 . Then.E {Li (Au (γ. counting in positive for the Euler characteristic (L0 > 1). M.E. It is not the purpose of this paper to give details on these formulae. w) for u from −20 to 20. Even if more peculiar.8 Euler Caracteristic [] 15 10 5 0 Threshold [] 5 10 15 20 0. full proof and details can be found in [9]. M))}.4(b) represent respectively the Euler characteristic and the volume fraction . Bogdan and J. 1 0. the Euler characteristic is the one of the full cube (L0 = 1).4(a) and Fig. Roubin.4 0.B.9 40 0. By increasing u. + Numerical values calculated from one realization of γ(x. the Euler characteristic curve shape reﬂects also easily the eﬀect of the threshold on excursion sets topology. however.1 0 20 80 20 15 10 5 0 Threshold [] 5 10 15 20 20 0 20 40 60 (a) (b) Figure 4: LKCs of excursion sets of Gaussian random ﬁeld in term of threshold values. several holes appear.7 Volume fraction [] 0.5 0. T ]: i=1 √ √ √ 3 2 2 2 3 2 2u 2 T u T u E{L0 }(Au ) = − 1 + 2π3/2 T 2 σ + 32π2 Lc e−u /2Lc + Ψ σ 2π 2 L3 σ2 Lc c √ √ 2 2 3 2 T2 u 2 T3 u e−u /2Lc + 3T Ψ σ E{L1 }(Au ) = 3/2 L2 σ + 4 Lc π (7) c √ 2 T 3 −u2 /2L2 E{L2 }(Au ) = c + 3T 2 Ψ u e π Lc σ u E{L3 }(Au ) = T 3 Ψ σ Fig.of excursion sets of γ(x.σ.3 0. 6 .
(8) do not always have a solution.4 that we can not expect N to be upper than 40 while keeping a “high” volume fraction (Vv > 40%). For exemple. 4 Application of the modeling framework on concrete like material Due to the intrinsic non linearity of eq. L2 and L0 . So far. σ) E{L2 }(u. Therefore. the Euler characteristic decreases to L0 = 0 when no more connected components remain. according to eq. We keep only three relevant characteristics from the four LKCs: the volume fraction Vv .(7). Understanding the full behaviour of these equations is a key point for anyone who wants to make excursion set modeling. Therefore in this part. the more realistic solution for “meatball” topology we get with this framework allows us to represent an aggregate phase with a maximum of 15% volume fraction. So far. the generation of the underlying excursion set rely on ﬁnding a solution for (u. σ. the eﬀort will be put in a “realistic” representation of the aggregates phase. the “meatball” topology has to be targeted and it is only once we assume that the excursion set is free from holes and handles that N can be estimated by L0 . we can point out that the variability of the numerical generation is very low. Lc ) = 1 ST 3 (8) 2 E{L0 }(u. Vv ) of the phase are set. From the comparison between theorical values and measures on one realization. for this range of excursion sets we can assume that V{Li (Au )} ≪ 1. Finally. attention must be taken when it comes to N . both variance and covariance length of γ(x. the framewok has been presented considering Gaussian random ﬁelds.B. But one needs to remember that. In our case. Until now. Roubin. depending on the diﬀerent values of (N . Vv ) (especially for “meatball” topology . σ. the volumic surface area S and the number of agregates N which are respetively linked with L3 . S. Bogdan and J. S. One phase (aggregates) is represented by an excursion set of a correlated random ﬁeld while its second phase (concrete) is represented by its complementary. Lc ) that satisfy the following system: = Vv T 3 E{L3 }(u. Colliat handles disappear forming a “meatball” like topology of connected components (L0 > 0). Which in our case of concrete like material modeling leads to a major issue.N ≫ 1) the problem eq. although eq.(7) gives only expectations of LKCs. M. w) aﬀect the morphology as well. σ.(7). Indeed the Euler characteristic does not indicate the number of aggregates for every topology. Lc ) = N . In this speciﬁc kind of topology: N #{connected components} = L0 . The application of this paper is made using a chisquare distribution with k degrees 7 The material is represented as an heterogeneous material with two phases. But estimation of LKCs for excursion set can also be worked out considering Gaussian related ﬁelds.E. we have seen the eﬀect of the threshold value on excursion sets. Thought Vv and S can be directly estimated. Once the three characteristics (N . we can clearly see on Fig.
k be k realizations of the Gaussian ﬁeld γ(x. For example. Therefore. 1 (a) Gaussian realization .. 5 FE model for heterogeneous material . the use of a χ2 distribution add the parameter k to the k system eq. 8 . the nearest solution is found for k = 1 and enable us to double the previous volume fraction Vvmax ≈ 30%.(7).δ = γ 2 1 Figure 5: Comparison between Gaussian and χ2 excursion sets for the same threshold value. Fig. Realizations of such ﬁelds can be seen as sum of k independent squared k realizations of a correlated Gaussian random ﬁeld. for the same threshold. These cut elements are split into two parts. 1 The χ2 distribution remains the more suitable solution for meatball topology and high 1 volume fraction morphology we found.(8). w) described above. shows clearly that. i = 1.Application to hydration process modeling The approach made here relies on a spatial truss. We have : k δ= i=1 γi2 (9) Although similar to eq. The choice of a not adapted meshing process is made here thus. Colliat of freedom . if we consider a twophase material (inclusions within a matrix).E. the spatial positions of nodes are not constrained by the morphology. Let δ be a realization of such ﬁeld and γi . Bogdan and J. 1 being the excursion from the squared realization of the excursion Fig. to model pattern of heterogeneities. Fig. An elementary enhancements method (EFEM) method for kinematic enhancement of Finite Element using the HuWashizu variational formulation is used here. each having diﬀerent elastic properties by enhancing them with strain (weak) discontinuities [11].5 shows a twodimensional slice of excursions from a Gaussian realization and a χ2 made from the same realization. both gemetrical and mechanical properties have to be handle inside some interface elements.5(b).5(a). Roubin.γ (b) χ2 realization . With such ﬁeld. it is natural to expect the volume fraction to double between excursions of Gaussian and χ2 random ﬁelds.χ2 . M.B.
Considering a simplistic version of the Powers and Brownyard hydration model [12].E. unhydrated cement and hydration products). Thus. Within this framework. thresholding a random ﬁeld with a scalar allows to create a two phase material. Bogdan and J. water. for a growing hydration degree. hydration products (including gel water) and free water. hydration products. that a second threshold. setting two thresholds will allow us to create a three phase material. anydrous cement As explained previously.12(1 − p)α Vw = 1 − Vh − Vanh where p is the initial porosity.1 (b) α = 0. the volume fraction of each one of them can be calculated according to the following equations: w/c p = w/c+ρw /ρc Vanh = (1 − p)(1 − α) (10) Vh = 2. and then. In order to illustrate this linear framework. Roubin. each phase’s volume fraction is known and can be linked to the random ﬁeld’s thresholds ui (equation eq. the initial morphology is set up by one threshold (two phases: water and unhydrated cement). Vh . α the hydration degree and Vanh . a projection of the previous excursion set is made onto the truss. with only three phases: unreacted cement. M. will allow to create an additional phase. One can easily imagine. with a diﬀerent value. those entirely in the inclusions. Eventually. a simple hydration process of concrete like material modeling has been implemented. for diﬀerent hydration degrees. (a) α = 0. Colliat three sets of elements are needed: those entirely in the matrix. macroscopic material characteristics like Young modulus can be estimated over a given hydration degree with simple tension tests. two thresholds are calculated and applied to the random ﬁeld.B. To calculate these elements repartition. The following 9 . concentrical to the ﬁrst one.(7)). and those which are split between both (cut elements).5 (c) α = 1 Figure 6: Projection of excusion set shapes on FE truss for diﬀerent hydration degrees. creating a three phase material (water. hydration products and water. Therefore. Vw respectively the volume fractions of anhydrous cement.
2 0.5 0. A slight raising of the slope can be seen after α = 0. M.E. The ﬁrst is that the interface is of rigidbrittle type. Colliat characteristics have been chosen Eanh = 135 000 MP a.4.7 Hydration degree [] 0. Bogdan and J. A other simple tension test is presented here.3 0. Fig.6 0.1 0.1. Details of this FE numerical implementation can be found in [3]. 6 FE models with embedded discontinuities In addition to the geometrical representation of heterogeneities. Material properties are deﬁned according to Tab. These discontinuities represent microcracks that can occurs in both phases as well as at the interfaces (debonding). displacement (strong) discontinuities are also introduced in the elements.4 0. 10 . in order to model a nonlinear softening response based on failure quasibrittle. 30000 25000 Young modulus [MPa] 20000 15000 10000 5000 0 0. Table 1: Material properties Matrix E = 10GP a σu = 3MP a Gf = 11J/m2 Inclusions 70GP a − − Interface − 3MP a 11J/m2 Two remarks are worthy of attention.9 1 Figure 7: Young modulus of a concrete like material for diﬀerent hydration degrees. The second is that we choosed for inclusions to remains in the linear elastic regime. Eh = 25 000 MP a and Ew = 1 MP a. Roubin.7 shows that the continuous growing of the macroscopic Young modulus over hydration degree is well handled by this FE representation.8 0.B.
8(a) where two zones are splited by a macroscopic crack (represented by means of the broken elements). advantages have been shown through two examples. First.025 Displacement [mm] 0. eﬀorts still have to be made in order to generate more realistic morphologies. and Chaboche.8(b) shows the macroscopic load vs imposed displacement curve where three steps can be seen. Roubin.B. Bogdan and J. Then. F. Indeed. Fig.03 0. Futhermore. Colliat 18000 16000 14000 12000 Load [kN] 10000 8000 6000 4000 2000 0 0 0.015 0. limiting the growing amount of numerical ressources needed when dealing with large sets of morphologies.01 0. Though. being able to represent broken elements by means of a strong discontinuity in the FE method allows calculations of permeability or diﬀusion in such damaged materials [13]. We can also add that this framework is well adapted to other problematics related with concrete like materials such as the eﬀect of morphological variability on macroscopic behaviour. M.005 0. REFERENCES [1] Feyel. a linear part where no failure occurs. Revue europ´enne des El´ments Finis : NUMe DAM’00 isues (2001) 10:449–472 11 . with the apparition of several microscopic cracks. J.L.02 0. the softening part begin when the localisation of these microscopic cracks creates a macrosopic one.04 (a) Displacement ﬁeld and crack pattern at last time step (b) Load vs imposed displacement Figure 8: Results for simple tension The cracking pattern is shown on Fig. 7 Concluding remarks This communication presents a ﬁrst attempt to create a sequential multiscale framework where morphology of heterogeneous material is deﬁned by excursion sets of correlated random ﬁelds.035 0.E. we can observe a yield behaviour. the use of both KarhunenLo`ve decomposition and nonadapted meshes allows fast compue tations. Multiscale nonlinear FE analysis of composite struc´e ture: damage and ﬁber size eﬀects. Finaly.
J.B. and Brownyard. and Wilson. K. Concr.Phys (1947) 37:1–79 [9] Adler. Upscaling permeability for fractured concrete : mesomacro numerical approach within a sequential framework submitted 12 . Acad. I. Probability Theory. Fennicae.L. J. and Belytshcko. Colliat. A Finite Element method for localized failure analysis Computer Methods in Applied Mechanics and Enginnering (1987) 61:189– 214 [12] Powers.C. De Sa. T. (2010) 82:1671–1688 [4] Sukumar. and Needleman.. Benboudjema... Am.. Modelling strong discontinuities in solid mechanics via strain softening constitutive equations. Colliat. Graduated Texts in Mathematics (1978) Vol. Roubin. X. and Gatuingt. II. 469–505. Failure of heterogeneous materials: 3D mesoscale FR models with embedded discontinuities. C. Sci.II.. 4th ed. M. T. N. Bogdan and J. [13] Jourdain. N.. N. A.J. Colliat [2] Adler. T. 249–336. Int J. G. Meth. Ser.L. J. F. Boston [10] Mantoglou. 845–880.. Modeling holes and ine clusions by level sets in the extended ﬁnite element method. J.E. R. A.B. Stochastic Environmental Research and Risk Assessment (2008) 22:809–822 [3] Benkemoun.... M. Hautefeuille.J. Inst. Computer Methods in Applied Mechanics and Engineering (2001) 190:6183–6200 [5] Oliver J. Y. International Journal for Numerical Methods in Engineering (1996) 39:3575–3623 [6] Loeve. F. 669–712. [7] Matheron. The intrinsic random functions and their applications. and Taylor. Ann. (1947).B. Chapp. 933–992. A. Num. Uber lineare Methoden in der Wahrscheinlichkeitsrechnung.. 549–602. Some new random ﬁeld tools for spatial analysis. 43101–132. Ibrahimbegovic. Advances in Applied Probability (1973) 5:439–468 ¨ [8] Karhunen. J. M. Leroy. Random Fields and Geometry (2007) Springer. 2:129–149 [11] Ortiz.L The Turning Bands Method for Simulation of Random Fields using Line Generation with a Spectral Method Water Resources Research (1982) Vol. M. D. R. Studies of the physical properties of hardened Portland cement paste. in Engng. Mo¨s. Math. A.E.