Sie sind auf Seite 1von 10

Metamodeling-Assisted Numerical Homogenization

for Masonry and Cracked Structures


Georgios A. Drosopoulos 1; Konstantinos Giannis 2; Maria E. Stavroulaki 3;
and Georgios E. Stavroulakis 4

Abstract: In this article a numerical approach is proposed for the investigation of composite materials, like masonry. A number of sim-
ulations of a nonlinear, heterogeneous microstructure with gradually increasing loading, are considered. Average stress and stiffness are also
Downloaded from ascelibrary.org by Tufts University on 06/13/18. Copyright ASCE. For personal use only; all rights reserved.

derived. Then, this information is incorporated in databases, and a metamodel is created based on the data and used in an overall macroscopic
analysis. On the micro scale, a classical plasticity model for masonry or a unilateral contact law representing a potential crack are considered.
A satisfactory comparison between the proposed models and direct macroscopic models arises, indicating that the method can be used for an
initial, quick assessment of the failure of heterogeneous materials on the structural scale. Due to its simplicity, it can be applied even by using
commercial software for the creation of the databases. Thus, the creation of complex, sophisticated, nonlinear finite element codes may be
avoided within the framework of this approach. DOI: 10.1061/(ASCE)EM.1943-7889.0001500. © 2018 American Society of Civil
Engineers.

Introduction can be found in Zohdi and Wriggers (2008) and Theodossopoulos


and Sinha (2013).
A multiscale method related to the investigation of the structural Different nonlinearities can also be depicted in a multiscale com-
behavior of composite structures is presented in this work. First, putational homogenization framework (Nguyen et al. 2011). In ad-
independent simulation of a microscopic representative volume dition, both the macroscopic and the microscopic structure are
element (RVE) of the material is performed. Then, average infor- simultaneously solved, and the framework of a nested procedure
mation is obtained and incorporated in a metamodel which repre- (Suquet 1985; Smit et al. 1998; Feyel 1999; Miehe et al. 1999;
sents the homogenized mechanical behavior at the macroscopic Kouznetsova 2002; Miehe and Koch 2002; Massart et al. 2007b;
level. The proposed method can be applied to masonry structures, Drosopoulos et al. 2014b; Massart et al. 2007a; Verhoosel et al.
composite materials, cracked and damaged structures, etc. 2010; Coenen et al. 2012; Drosopoulos et al. 2014a) is incorpo-
Within the framework of numerical homogenization, first a unit rated. A macro-micro-macro computational homogenization for in-
cell is analyzed and its average material properties are estimated. elastic materials has been proposed in Tsalis et al. (2013) and
Then, these average properties are considered for the estimation of Chatzigeorgiou et al. (2016). The main concept related to computa-
the parameters of the macroscopic material law (Luciano and Sacco tional homogenization is the fact that the macroscopic material law
1998; Dascalu et al. 2008). Sacco (2009) proposed transformation is built numerically, during simulation; the microscopic problem
field analysis for the representation of the effective properties of is solved and its average properties are sent to the macro scale.
masonry by using the technique of superposition of the effects By this procedure (which is often called FE2 ), complex microscopic
and the finite element method. In Mercatoris et al. (2009) a closed- patterns can be considered, incorporating every nonlinear law.
form damage model at the mesoscopic level of masonry is used to The idea of deriving the macroscopic material properties of
show that a structural scale localization criterion, based on the heterogeneous materials by using interpolation to average quan-
acoustic tensor adapted to shell kinematics, allows the detection of tities obtained by independent microscopic simulations has been
the structural scale localization. Limit analysis, quadratic program- presented in several articles. In Yvonnet et al. (2009) the effective
ming, and homogenization can also be combined (Milani 2011;
strain-energy potential of nonlinear elastic heterogeneous materials
Milani et al. 2013a, b; Reccia et al. 2014) for the investigation of
is computed in a number of points (nodes) which discretize the
masonry structures. Several aspects of numerical homogenization
macroscopic strain space. Then, an interpolation method is used
1 for the estimation of the strain-energy potential in the whole macro-
Discipline of Civil Engineering, School of Engineering, Univ. of
KwaZulu-Natal, 4041 Durban, South Africa (corresponding author).
scopic strain space, and the effective stress–strain relation and the
Email: DrosopoulosG@ukzn.ac.za tangent stiffness are directly calculated. In Clément et al. (2012), a
2
Institute for Particle Technology, Technical Univ. of Braunschweig, database is used to describe the strain energy density (potential)
Volkmaroder Str. 5, 38104 Braunschweig, Germany. function in both the macroscopic strain space and the geometrical
3
Faculty of Architecture, Applied Mechanics Laboratory, Technical random parameters domain. Finite element analysis on an elemen-
Univ. of Crete, GR-73100 Chania, Crete, Greece. tary cell whose geometry is given by random parameters is used for
4
School of Production Engineering and Management, Institute of the determination of the potential function. A continuous explicit
Computational Mechanics and Optimization, Technical Univ. of Crete, form of the potential, which stands for the evaluation of the macro-
GR-73100 Chania, Crete, Greece.
scopic stress and elastic tangent tensors during the macroscopic
Note. This manuscript was submitted on January 17, 2017; approved
on March 7, 2018; published online on June 13, 2018. Discussion period structural analysis, is estimated by introducing an interpolation
open until November 13, 2018; separate discussions must be submitted for scheme. Artificial neural networks (ANN) have been used in Unger
individual papers. This paper is part of the Journal of Engineering Me- and Könke (2008) in order to link the mesoscopic and the macro-
chanics, © ASCE, ISSN 0733-9399. scopic scale. The mesoscale model is approximated by an ANN,

© ASCE 04018072-1 J. Eng. Mech.

J. Eng. Mech., 2018, 144(8): 04018072


which in turn has been trained from a certain number of experi- The idea which is presented in this article, is related to the usage
ments. Then, the constitutive relation in the macroscopic scale is of databases (instead of RVE simulations) to depict the microscopic
derived by using a homogenization technique. behavior for the chosen structure. These are built, initially, by con-
Within the framework of the present article, some parameters ducting parametric microscopic analyses. For each strain load, an
are incorporated in microscopic finite element models to describe average stress and (tangent) stiffness is assigned in each database,
the loading of each structure. For each loading, average material respectively. These stand for the material properties that will be
properties are calculated. Then, this information is stored in two used by the macroscopic structure in a next step.
databases. For each strain load, an average stress is assigned (first
database) as well as an average stiffness (second database). Accord-
ing to the final step of the method, an interpolation algorithm is Outline of the Method
adopted in MATLAB for the usage of the databases by a homog- The proposed procedure is described by the following steps:
enization (macro) metamodel. This final model works in an FE2 1. The finite element model of the nonlinear RVE is developed.
framework. Two different microscopic constitutive laws are chosen in this
Several microstructures with different material laws can be si- study, one representing a masonry RVE with a proper elasto-
mulated by the proposed scheme. In this article, one elastoplastic
Downloaded from ascelibrary.org by Tufts University on 06/13/18. Copyright ASCE. For personal use only; all rights reserved.

plastic law, and another depicting a two-dimensional plate


masonry structure and one rectangular structure with a crack, which with unilateral contact standing for a potential crack. COMSOL
is simulated by unilateral contact, have been used. Multiphysics version 3.4 and MATLAB have been used, respec-
Finally, among the goals of this work is the proposition of tively, for the analysis of the former and the latter models.
a simple numerical scheme which can be used for the initial assess- 2. A strain load vector, with dimensions 3 × 1 in the plane stress
ment of heterogeneous structures. The proposed approach can conditions used, is applied to the RVE by using either linear
be implemented by commercial software or simple MATLAB displacement or periodic boundary conditions. Three indepen-
codes. Sophisticated localization phenomena are not taken into dent parameters are then adopted for the representation of
account; however, the extension of the method to cover these issues each possible strain vector load case. Variation of these para-
(e.g., Katz et al. 2015) will be the subject of future work. meters leads to a (significant) number of different loading strain
combinations.
3. Following Step 2, average RVE stresses and strains are derived.
Proposed Multiscale Numerical Approach 4. Three test incremental loading strain vectors are used and
Steps 2 and 3 are repeated. Then, by developing the incremental
Multiscale computational homogenization states that two problems stress–strain relation, the instantaneous, average tangent tensor
are formulated (Smit et al. 1998; Kouznetsova 2002): (1) a macro- is obtained from each analysis.
scopic problem, and (2) a microscopic problem. In each Gauss 5. Two databases are derived. Within the first, each strain is
point of a macroscopic model, a microscopic structure (RVE) is assigned an average stress, and in the second, each strain is as-
assigned. Every material law related to the structure is considered signed a (tangent) stiffness. A multiscale homogenization mod-
at this microscopic level. el, which uses these databases, is then developed in MATLAB.
The principle which theoretically supports this method is the Within this metamodel, the required quantity is interpolated
Hill-Mandel energy condition (Hill 1963): from the databases.
6. The results obtained are compared with direct heterogeneous
Z
1 macroscopic models. These are developed within commercial
σ M :δϵM ¼ σ m :δϵm dV m ð1Þ finite element packages and used to verify the proposed
Vm Vm
method.
This energy principle is satisfied when the three most common
types of loading conditions (among others) are applied to the Micro to Macro Relations and the Averaging Procedure
boundaries of the microstructure (RVE): (1) prescribed linear dis-
The macroscopic stress of the final macroscopic metamodel will be
placements, (2) prescribed tractions, and (3) periodic boundary
the average stress estimated in the first database. Moreover, the
conditions. Both linear displacement and periodic boundary con-
macroscopic tangent stiffness matrix will be built based on the ef-
ditions have been selected for this work.
fective tangent tensor, obtained by the second database.
In particular, each RVE is loaded by a strain, the strain of each
For the first phase of the presented concept, a strain loading is
macroscopic Gauss point, and a finite element analysis takes pla-
applied to the boundaries of the RVE, within the framework of lin-
ces. When convergence is reached, the average material properties
ear displacement or periodic boundary conditions. The former con-
are calculated, average stress and (tangent) stiffness. These proper-
ditions are given by Eq. (2):
ties are then used by the macro Gauss point (Fig. 1). According to
this procedure, the constitutive behavior is built during simulation. uj∂V m ¼ ϵM x ð2Þ

where a loading strain ϵM is applied as displacement loading to the


boundaries ∂V m of the RVE; and x denotes the matrix with the
undeformed coordinates of the boundary nodes of the RVE.
Eq. (2) is further elaborated and Eq. (3) is derived:
ux ¼ ϵxx x þ 0.5γ xy y ð3aÞ

uy ¼ ϵyy y þ 0.5γ xy x ð3bÞ


Fig. 1. Schematic representation of the multiscale, concurrent, compu-
where ux and uy = displacement (boundary) loading along the X
tational homogenization.
and Y axes; and ϵxx , ϵyy , and γ xy = terms for the loading strain tensor

© ASCE 04018072-2 J. Eng. Mech.

J. Eng. Mech., 2018, 144(8): 04018072


(plane stress conditions). For linear displacement boundary condi- To obtain the stiffness of the overall macroscopic model, the
tions, the displacement loading is applied at each boundary node of same procedure is repeated. Thus, for every strain load combination
the RVE; therefore, x and y are the (undeformed) coordinates of which is considered in each microscopic analysis, three test, incre-
each boundary node. mental strain vectors are additionally applied to the RVE [Eq. (9a)],
Periodic boundary conditions require periodic displacements, as resulting in three incremental average stress vectors [Eq. (9b)].
well as antiperiodic tractions, in the opposite boundaries of the Then, the instantaneous, effective tangent tensor is estimated by
RVE. In this framework, the displacements of the opposite boun- using the system of equations given in Eq. (9c). This is used for
daries of the RVE are given by the following equations: the creation of the macroscopic tangent stiffness matrix

uT − uB ¼ u4 − u1 ð4aÞ ½δϵM  ¼ ½ δϵM


1 δϵM
2 3 
δϵM ð9aÞ

uL − uR ¼ u1 − u2 ð4bÞ ½δσ M  ¼ ½ δσ M
1 δσ M
2 3 
δσ M ð9bÞ
where the displacements in the top, bottom, left, and right boun-
daries are estimated by using the prescribed displacements of three ½δσ M  ¼ CM ½δϵM  ⇒ CM ¼ ½δσ M ½δϵM −1 ð9cÞ
Downloaded from ascelibrary.org by Tufts University on 06/13/18. Copyright ASCE. For personal use only; all rights reserved.

corner nodes of the RVE, namely 1, 2, and 4, given by Eq. (2).


For the simplest case of a single rectangular (four-node) In plane stress problems, the tangent tensor CM is a 3 × 3 tensor,
element, with dimensions Lx , Ly and origin of the coordinate sys- while δϵM 1 , δϵ2 , δϵ3 , δσ 1 , δσ 2 , and δσ 3 are 3 × 1 tensors; thus,
M M M M M

tem at the bottom-left node, the strain loading of the four corner δϵ and δσ are both 3 × 3 tensors.
M M

nodes (Nodes 1–4 defined anticlockwise) is described by the fol- To use the microstructure databases on the macroscopic model,
lowing equations: the data included in each database (taken from COMSOL) are
saved in spreadsheet files. Each file has 6 columns, representing
u1x ¼ 0; u1y ¼ 0 ð5aÞ each of the (three) strain components and the corresponding (three)
stress components. On each line, one strain load combination and
u2x ¼ ϵxx Lx ; u2y ¼ 0.5γ xy Lx ð5bÞ the corresponding stress components derived from the parametric
(strain) loading are depicted. These files are then used by the
overall multiscale scheme to build the stress and the stiffness
u3x ¼ ϵxx Lx þ 0.5γ xy Ly ; u3y ¼ ϵyy Ly þ 0.5γ xy Lx ð5cÞ corresponding to each Gauss point. In a preliminary step, Eq. (9)
is implemented and the strain–tangent stiffness database is built.
u4x ¼ 0.5γ xy Ly ; u4y ¼ ϵyy Ly ð5dÞ Then, in the final multiscale homogenization analysis, the strain–
stress and strain–tangent stiffness databases are used to provide the
As previously mentioned, three independent parameters are stress and the tangent stiffness within the framework of the non-
used in the COMSOL finite element package for the representation linear finite element analysis.
of each possible strain vector. To capture all the possible strain vec- It should be mentioned that according to the description given
tors that the macroscopic structure may potentially encounter at here, on the microscopic scale only the averaging relations are used
every location in the analysis space, some limiting (maximum for the estimation of the necessary information for the final meta-
and minimum) values for the components of the loading strain model (stress and stiffness). Thus, no manipulation of the final tan-
are defined. Then, a great number of RVE simulations takes place gent stiffness matrix of the RVE is needed for the representation of
within these limits, with every possible combination between strain the average stress–strain law, contrary to classical computational
components. To define these limit values, one may consult the homogenization (Kouznetsova 2002; Drosopoulos et al. 2014b).
stress–strain behavior of the chosen material given in the literature This makes the implementation of the proposed scheme possible
(and find yield strain, ultimate strain, etc.) or run a simple, linear even by using a commercial software package, provided that vol-
analysis problem of the macroscopic structure. ume average quantities can be calculated and a simple finite
After each strain load is applied to the boundaries of the RVE, element metamodel incorporating relation (9c) can be developed.
the average strain and stress are obtained by implementing averag- Regarding the computational effort which is required for the im-
ing relations: plementation of the proposed method, it is mainly significant for
Z Z the creation of the databases. It depends on specific parameters
1 1
<ϵ>V m ¼ ϵm dV m ; < σ>V m ¼ σ m dV m ð6Þ of the microscopic model—mainly the density of the mesh—as
Vm Vm V m Vm well as the number of loading strain vector combinations. There-
fore, it can vary from several minutes to several hours. In order to
The volume average strain of the microstructure is equal to the consider 9,000 strain loading combinations and a mesh with 1,767
(known) load strain vector: nodes (3,534 degrees of freedom), COMSOL 3.4 was used and ap-
<ϵ>V m ¼ ϵM ð7Þ proximately 15 h were needed for a quad-core processor for the
creation of each database. For the solution of each elastoplastic
For the model, developed in MATLAB, of a two-dimensional model, the Newton-Raphson procedure was used. For the second,
plate with the unilateral contact and periodic boundary conditions final phase, where the databases are used on a structural level, the
applied to the RVE, the following relation for the average stress is computational time is relatively small (depending on the mesh den-
adopted (Kouznetsova 2002; Miehe and Koch 2002): sity of the macro structure). Only some minutes were needed for the
simulations presented in this article.
1 It is worth noticing that the alternative of simulating the direct
< σ>V m ¼ f x ¼ σM ð8Þ
Vm p p heterogeneous macroscopic model, taking into account the stone
blocks and mortar of masonry structures, results in large numerical
where fp = external forces in three corner nodes (of the RVE) with models requiring significant computational effort to be solved. This
prescribed displacements; and xp = undeformed coordinates of is attributed to the fact that in order to simulate the small thickness
these nodes. of the mortar, the larger thickness of the stone, and extend for the

© ASCE 04018072-3 J. Eng. Mech.

J. Eng. Mech., 2018, 144(8): 04018072


Fig. 2. Multiscale homogenization scheme.
Downloaded from ascelibrary.org by Tufts University on 06/13/18. Copyright ASCE. For personal use only; all rights reserved.

overall masonry structure, a fairly dense mesh (consisting of thou- Overall Macroscopic Metamodel
sands of elements and nodes) is required. On the contrary, within
The final macroscopic metamodel proposed in this work is shown
the framework of the proposed scheme, the final multiscale model schematically in Fig. 2. A finite element model representing the
consists of a small number of elements. macroscopic structure is developed in MATLAB. Plane stress
analysis, with first order, full integration elements, has been se-
lected within the framework of the Newton-Rapshon procedure.
For each strain, a stress and an instantaneous tangent tensor should
be selected from the databases.
In particular, the nonlinear macroscopic equation is given by
Kðui Þδuiþ1 ¼ −Gðui Þ ð10Þ

where K = tangent stiffness matrix; and G = residual force vector,


given by
Z
Gðui Þ ¼ BT σðui ÞdV − F ð11Þ
V

To build the tangent stiffness matrix K and the internal force


vector given by the integral of Eq. (11), the two databases obtained
from the microstructural analysis are used; the strain to stress data-
base is used to create the internal force vector, and the strain to
stiffness database is used to create the tangent stiffness matrix.
The flow chart shown in Fig. 3 depicts the overall multiscale
scheme.
To obtain these quantities from the databases, an interpolation
technique should be used. According to the method chosen here, a
scattered dataset, defined by locations X and corresponding values
Y, is interpolated using a Delaunay triangulation of X. By doing
this, a surface which has the form Y ¼ fðXÞ is built. Then, the in-
terpolation function f is used for the evaluation of the surface at any
query location qX, using qY ¼ fðqXÞ, where qX lies within the
convex hull of X. In this work, a MATLAB toolbox is used for
the implementation of this interpolation [MATLAB R2010a
Fig. 3. Flow chart of the proposed multiscale scheme.
(MATLAB 2010)]; however other approaches could be considered,
such as Neural Networks (Lefik 2004).

Fig. 4. (a) Geometry of the masonry RVE (in millimeters); and (b) mesh of the RVE.

© ASCE 04018072-4 J. Eng. Mech.

J. Eng. Mech., 2018, 144(8): 04018072


× 10-5
0.1
0.09
0.08
0.07
10
0.06
0.05
0.04
0.03
0.02
0 0.01
(a) (b) 0

Fig. 5. Plastic strain distribution in the RVE, for gradually increased strain loading, from (a) to (b).
Downloaded from ascelibrary.org by Tufts University on 06/13/18. Copyright ASCE. For personal use only; all rights reserved.

Numerical Examples For this model, a total of 1,680 four-node, quadrilateral, isopara-
metric plane stress elements (1,767 nodes) and an out-of-plane thick-
Simulation of Masonry Structures ness of 70 mm were used (Fig. 4). The properties of the materials
were (Lourenço and Pina-Henriques 2006): for bricks Eb ¼
The microscopic masonry structure (RVE) shown in Fig. 4 is made 4865 N=mm2 , nb ¼ 0.09; and for mortar Em ¼ 1180 N=mm2,
of bricks and mortar joints. In this work, only the nonlinear nm ¼ 0.06. The tensile and compressive strength of the mortar, were
behavior of the mortar has been considered. For this, a Drucker- 0.9 and 3.2 N=mm2 , respectively.
Prager material law has been chosen (Drucker 1954; COMSOL Fig. 5 shows failure of the mortar joint in the RVE as the loading
Multiphysics 2007). strain is gradually increased. This behavior, which is obtained after

−1

−2
(a) (b)

-1

(c)

Fig. 6. (a) Masonry wall (in meters); (b) maximum principle stresses from the proposed approach (MPa); and (c) maximum principle stresses from the
direct macroscopic simulation (MPa).

© ASCE 04018072-5 J. Eng. Mech.

J. Eng. Mech., 2018, 144(8): 04018072


The element type and material properties within the DNS model
were the same as those taken into account in the RVE.
A first example, a rectangular masonry wall, was examined. The
dimensions, displacement loading, and boundary conditions for
this structure are shown in Fig. 6(a). For this DNS analysis, a total
of 40,768 four-node, quadrilateral, isoparametric plane stress ele-
ments (41,173 nodes) were used. A significantly smaller number of
elements were used for the macroscopic model of the proposed
multiscale scheme; thus, 400 elements (441 nodes) were considered
for this model.
In Figs. 6(b and c) the maximum principle stress distribution
from the proposed homogenization scheme [Fig. 6(b)] and the
DNS model [Fig. 6(c)] is given. According to these figures, a sim-
ilar distribution for the maximum principle stress arises from the
Fig. 7. Force-displacement diagrams from both the multiscale and
Downloaded from ascelibrary.org by Tufts University on 06/13/18. Copyright ASCE. For personal use only; all rights reserved.

two approaches.
DNS approaches. A comparison between the force-displacement diagrams ob-
tained from both the DNS and the multiscale models is given in
Fig. 7. Based on this figure, it can be stated that the recommended
RVE analysis, is projected on the macroscopic scale by use of the approach can be used as a tool for a first evaluation of a composite
databases, within the final macroscopic metamodel. structure.
A direct heterogeneous model was created (within commercial A second example of a masonry structure with two openings
finite element software) and direct numerical simulation (DNS) was then examined. The dimensions, displacement loading, and
was used for the verification of the method presented in this work. boundary conditions for this structure are shown in Fig. 8(a).

0.5

–0.5

(a) (b)

1.5
1.0
0.5
0.0
–0.5
–1.0
–1.5

(c)

Fig. 8. (a) Masonry wall with two openings (in meters); (b) maximum principle stresses from the proposed approach (MPa); and (c) maximum
principle stresses from the direct macroscopic simulation (MPa).

© ASCE 04018072-6 J. Eng. Mech.

J. Eng. Mech., 2018, 144(8): 04018072


A dense mesh of 103,805 four-node, quadrilateral, isoparametric Simulation of Material with Uniform Distributed
plane stress elements (104,825 nodes) was used for the DNS analy- Microcracks
sis. The macroscopic multiscale mesh consists of 539 elements
In this section, the proposed multiscale computational homogeni-
(612 nodes).
zation scheme was applied to a different material, a microstructure
In Figs. 8(b and c) the maximum principle stress distribution
with a preexisting, open crack [Figs. 10(b and c)]. Between the
from the proposed homogenization scheme [Fig. 8(b)] and the
faces of the crack, a unilateral contact condition without friction
DNS model [Fig. 8(c)] is given. Although some differences
was introduced. This law induced the only nonlinearity of the
arose, mainly in peak negative stresses, the results are qualita-
model.
tively comparable in depicting values with similar orders of
When only tensile loading is applied to the macroscopic struc-
magnitude.
Finally, the stress contour received from the two models is given ture [Fig. 10(a)], the microcrack tends to open and the analysis is
in Fig. 9. According to this figure, a sufficient comparison arises linear, since for a nonpropagating crack no further change in the
for the stress distribution and the limit stress values. Some differ- stiffness of the structure takes place. On the contrary, for a com-
ences, which are observed in the Sxx values, can be attributed to pressive load case, the analysis becomes nonlinear due to the differ-
some peaks that appear in the corners of the windows in the direct ent stiffnesses of the system, before and after the appearance of
Downloaded from ascelibrary.org by Tufts University on 06/13/18. Copyright ASCE. For personal use only; all rights reserved.

heterogeneous model. These peak values of stresses, which are contact in the RVE. Thus, when contact occurs, the stiffness of
depicted by the very dense mesh of the direct heterogeneous model, the structure is increased.
cannot be depicted by the homogenization scheme. However, The material properties used in this problem were taken from
the differences only appear in very small areas of the structure the literature and can be considered similar to a ceramic material:
(e.g., near the corners of the windows). For the same reason, some E ¼ 328,389 N=mm2 , n ¼ 0.229. Figs. 10(b and c) show the di-
positive values for the Sxy stresses, obtained from the direct hetero- mensions of the microscopic structure (RVE) and the mesh,
geneous model [Fig. 9(c)], are not captured by the multiscale model consisting of 412 four-node, isoparametric plane stress elements
[Fig. 9(f)]. The contour of Sxy stresses shown in Fig. 9(c) indicates (462 nodes). Periodic boundary conditions were used for the load-
that these positive values represent only small areas of the wall ing of the microstructure. For the parametric microscopic simula-
(the authors can recognize four small areas, two near the top edge tions, a MATLAB finite element code was developed.
and two below the bottom window, with a value approximately Fig. 10(a) shows the dimensions, the (tensile) loading, and the
equal to 0.09 MPa). boundary conditions of the macrostructure. For the multiscale

Fig. 9. Stresses Sxx , Syy , and Sxy (MPa) obtained (a–c) from the direct macroscopic model; and (d–f) from the proposed homogenization approach
(last time step).

© ASCE 04018072-7 J. Eng. Mech.

J. Eng. Mech., 2018, 144(8): 04018072


(a) (b) (c)

Fig. 10. (a) Macroscopic structure (in millimeters); (b) RVE (in millimeters) with an initial opening (preexisting crack); and (c) mesh of the RVE.
Downloaded from ascelibrary.org by Tufts University on 06/13/18. Copyright ASCE. For personal use only; all rights reserved.

−3
x 10
9

0
(a) (b)

Fig. 11. Vertical displacements (in meters) obtained from (a) the proposed multiscale approach; and (b) the direct macroscopic simulation with tensile
macroscopic load.

−4
x 10
0

−2

−4

−6

−8

−10

(a) (b)

Fig. 12. (a) Nonlinear force-displacement diagram from the overall multiscale scheme and the direct macroscopic simulation; and (b) vertical
displacements (in meters) of an RVE at the end of analysis.

analysis, 100 first order, plane stress elements (121 nodes) were According to the results shown in Fig. 11, a quite similar vertical
selected. Identical characteristics were adopted in the direct displacement distribution was generated by the two numerical
numerical simulation of the macroscopic heterogeneous structure. schemes.
The mesh of this model consisted of 41,200 elements (42,411 When a compressive load is applied to the macroscopic struc-
nodes). ture, either no contact or contact conditions appear in the RVE.

© ASCE 04018072-8 J. Eng. Mech.

J. Eng. Mech., 2018, 144(8): 04018072


Downloaded from ascelibrary.org by Tufts University on 06/13/18. Copyright ASCE. For personal use only; all rights reserved.

Fig. 13. Vertical displacements (in meters) obtained from (a) the proposed multiscale approach; and (b) the direct macroscopic simulation with
compressive load.

When the gap closes and contact occurs, a higher value in the y structures (masonry) or structures with cracks by using commercial
direction of the effective tangent tensor is expected. In this case, software and/or simplified MATLAB codes in the form of interpo-
the effective tangent tensor is similar to the initial elasticity tensor lation from a database of results coming from suitable models, in-
of the structure. stead of sophisticated models, which need specialized knowledge
An augmented Lagrangian formulation has been used to simu- in computational mechanics (to be formulated) and advanced pro-
late contact conditions in the microscopic analysis (Drosopoulos gramming capabilities (to be implemented).
et al. 2014b). In addition, although the authors did simulate a problem close to
According to the diagrams shown in Fig. 12(a), the expected real cases—a two-story masonry wall with windows—it should be
nonlinearity due to the difference in the stiffness before and stated that this approach can be used in even bigger structures, pro-
after contact occurrence can be successfully depicted. At the end vided that one uses a commercial finite element analysis package
of the analysis, the gap is closed and a contact state on the micro- for conducting parametric analyses for building the databases, and
scopic level is received [Fig. 12(b)]. The depicted change in the a finite element package or a simple MATLAB code for using the
contact/no-contact state shown in this work also characterizes databases. There are commercial finite element packages which can
more complicated microscopic structures with potential periodic easily incorporate the parameters related to materials, loading, etc.
microcracks—for instance, composite materials [matrix-fiber and (e.g., ANSYS or COMSOL). Modern approaches, like parallel sol-
its interface, simulated within a classical FE2 approach without ution by several computer cores, usage of clusters, etc., could be
databases in Drosopoulos et al. (2014b)]. used to reduce computational effort, which is mainly required for
It is worth noticing that a step-by-step increasing of the loading the creation of the databases.
allows for gradual activation of contact along the crack faces and Complex phenomena, such as localization, cannot yet be inves-
clear observation of the nonlinear effects shown in Fig. 12(a). tigated with this approach. Consequently, it can be considered as an
Finally, Figs. 12(a) and 13 show the comparison between the initial tool for the assessment of the mechanical behavior of hetero-
multiscale homogenization scheme and the direct heterogeneous geneous structures.
model for a compressive load case. The proposed multiscale algorithm has been used on masonry
structures as well as on materials with a microscopic open crack.
In the former model, degradation of strength due to failure in the
Conclusions mortar is depicted. In the latter, alterations in the stiffness of the
macroscopic system, due to changes in the contact conditions of
In this work a multiscale approach is proposed for the investigation the open microcrack, are given.
of the behavior of heterogeneous structures. According to the main Possible future work could incorporate the study of localization,
concept, two databases, which include information for the average thus softening traction-separation laws describing the (unstable)
stress–stiffness of a macroscopic structure, replace each RVE sim- behavior of materials. Moreover, a cyclic (rather than monotonic)
ulation that normally takes place within the framework of an FE2 loading could be considered to further extend the applicability of
analysis. To build the databases some independent parametric mi- the proposed scheme. Thais would involve the consideration of
croscopic analyses are initially considered. An overall metamodel, both loading and unloading during the creation of the databases.
which uses the databases to build the constitutive behavior, is then
developed. Comparison with direct numerical heterogeneous mod-
els has been used to verify the proposed approach.
References
The proposed scheme can be implemented even by commercial
software, since only the averaging relations in the microscopic Chatzigeorgiou, G., N. Charalambakis, Y. Chemiski, and F. Meraghni.
scale are necessary for the macroscopic model (stress and stiffness). 2016. “Periodic homogenization for fully coupled thermomechanical
No manipulation of the final tangent stiffness matrix of the RVE is modeling of dissipative generalized standard materials.” Int. J. Plast.
needed, contrary to classical computational homogenization 81: 18–39. https://doi.org/10.1016/j.ijplas.2016.01.013.
(Kouznetsova 2002). Therefore, this method offers an alternative Clément, A., C. Soize, and J. Yvonnet. 2012. “Computational nonlinear
way of estimating the structural evaluation of heterogeneous stochastic homogenization using a nonconcurrent multiscale approach

© ASCE 04018072-9 J. Eng. Mech.

J. Eng. Mech., 2018, 144(8): 04018072


for hyperelastic heterogeneous microstructures analysis.” Int. J. Numer. Miehe, C., and A. Koch. 2002. “Computational micro-to-macro transitions
Methods Eng. 91 (8): 799–824. https://doi.org/10.1002/nme.4293. of discretized microstructures undergoing small strains.” Archit. Appl.
Coenen, E. W. C., V. G. Kouznetsova, E. Bosco, and M. G. D. Geers. 2012. Mech. 72 (4–5): 300–317. https://doi.org/10.1007/s00419-002-0212-2.
“A multi-scale approach to bridge microscale damage and macroscale Miehe, C., J. Schröder, and J. Schotte. 1999. “Computational homogeni-
failure: A nested computational homogenization-localization frame- zation analysis in finite plasticity simulation of texture development
work.” Int. J. Fract. 178 (1–2): 157–178. https://doi.org/10.1007 in polycrystalline materials.” Comput. Methods Appl. Mech. Eng.
/s10704-012-9765-4. 171 (3–4): 387–418. https://doi.org/10.1016/S0045-7825(98)00218-7.
COMSOL Multiphysics. 2007. User’s guide. Stockholm Sweden: COM- Milani, G. 2011. “Simple lower bound limit analysis homogenization
SOL AB. model for in- and out-of-plane loaded masonry walls.” Constr. Build.
Dascalu, C., G. Bilbie, and E. K. Agiasofitou. 2008. “Damage and size Mater. 25 (12): 4426–4443. https://doi.org/10.1016/j.conbuildmat.2011
effects in elastic solids: A homogenization approach.” Int. J. Solids .01.012.
Struct. 45 (2): 409–430. https://doi.org/10.1016/j.ijsolstr.2007.08.025. Milani, G., Y. W. Esquivel, P. B. Lourenço, B. Riveiro, and D. V. Oliveira.
Drosopoulos, G. A., M. E. Stavroulaki, K. Giannis, L. Plymakis, G. E. 2013a. “Characterization of the response of quasi-periodic masonry:
Stavroulakis, and P. Wriggers. 2014a. “Nonlinear homogenization in Geometrical investigation, homogenization and application to the
masonry structures.” In Proc., WCCM XI-ECCM V-ECFD VI., edited Guimarães castle Portugal.” Eng. Struct. 56: 621–641. https://doi.org/10
by E. Oñate, X. Oliver, and A. E. Huerta, 6795–6806. Barcelona, Spain: .1016/j.engstruct.2013.05.040.
Downloaded from ascelibrary.org by Tufts University on 06/13/18. Copyright ASCE. For personal use only; all rights reserved.

International Center for Numerical Methods in Engineering. Milani, G., M. Pizzolato, and A. Tralli. 2013b. “Simple numerical model
Drosopoulos, G. A., P. Wriggers, and G. E. Stavroulakis. 2014b. “A multi- with second order effects for out-of-plane loaded masonry walls.” Eng.
scale computational method including contact for the analysis of dam- Struct. 48: 98–120. https://doi.org/10.1016/j.engstruct.2012.08.029.
age in composite materials.” Comput. Mater. Sci. 95: 522–535. https:// Nguyen, V. P., M. Stroeven, and L. J. Sluys. 2011. “Multiscale continuous
doi.org/10.1016/j.commatsci.2014.08.004. and discontinuous modeling of heterogeneous materials: A review on
Drucker, D. C. 1954. “Coulomb friction, plasticity and limit loads.” J. Appl. recent developments.” J. Multiscale Model 3 (4): 1–42. https://doi.org
Mech. 21: 71–74. /10.1142/S1756973711000509.
Feyel, F. 1999. “Multiscale FE2 elastoviscoplastic analysis of composite Reccia, E., G. Milani, A. Cecchi, and A. Tralli. 2014. “Full 3D homogeni-
structures.” Comput. Mater. Sci. 16 (1–4): 344–354. https://doi.org/10 zation approach to investigate the behavior of masonry arch bridges:
.1016/S0927-0256(99)00077-4. The Venice trans-lagoon railway bridge.” Constr. Build. Mater.
Hill, R. 1963. “Elastic properties of reinforced solids: Some theoretical 66: 567–586. https://doi.org/10.1016/j.conbuildmat.2014.05.096.
principles.” J. Mech. Phys. Solids 11 (5): 357–372. https://doi.org/10
Sacco, E. 2009. “A nonlinear homogenization procedure for periodic ma-
.1016/0022-5096(63)90036-X.
sonry.” Eur. J. Mech. A/Solids 28 (2): 209–222. https://doi.org/10.1016
Katz, A., C. Trinh, J. Wright, W. Tu, and M.-J. Pindera. 2015. “Plastic-
/j.euromechsol.2008.06.005.
strain localization in periodic materials with wavy brick-and-mortar ar-
Smit, R., W. Brekelmans, and H. Meijer. 1998. “Prediction of the mechani-
chitectures and its effect on the homogenized response.” Compos.
cal behavior of non-linear heterogeneous systems by multi-level finite
Part B Eng. 68: 270–280. https://doi.org/10.1016/j.compositesb.2014
element modeling.” Comput. Methods Appl. Mech. Eng. 155 (1–2):
.08.037.
181–192. https://doi.org/10.1016/S0045-7825(97)00139-4.
Kouznetsova, V. 2002. “Computational homogenization for the multi-
scale analysis of multi-phase materials.” Ph.D. thesis, Technical Univ. Suquet, P. M. 1985. Local and global aspects in the mathematical theory
Eindhoven. of plasticity. Plasticity today: Modelling, methods and applications.
Lefik, M. 2004. “Hybrid, finite element-artificial neural network model for London: Elsevier.
composite materials.” J. Theor. Appl. Mech. 42 (3): 539–563. Theodossopoulos, D., and B. Sinha. 2013. “A review of analytical methods
Lourenço, P. B., and J. Pina-Henriques. 2006. “Validation of analytical and in the current design processes and assessment of performance of ma-
continuum numerical methods for estimating the compressive strength sonry structures.” Constr. Build. Mater. 41: 990–1001. https://doi.org
of masonry.” Comput. Struct. 84 (29–30): 1977–1989. https://doi.org/10 /10.1016/j.conbuildmat.2012.07.095.
.1016/j.compstruc.2006.08.009. Tsalis, D., T. Baxevanis, G. Chatzigeorgiou, and N. Charalambakis. 2013.
Luciano, R., and E. Sacco. 1998. “A damage model for masonry struc- “Homogenization of elastoplastic composites with generalized perio-
tures.” Eur. J. Mech. A/Solids 17 (2): 285–303. https://doi.org/10 dicity in the microstructure.” Int. J. Plast. 51: 161–187. https://doi
.1016/S0997-7538(98)80087-9. .org/10.1016/j.ijplas.2013.05.006.
Massart, T. J., R. H. J. Peerlings, and M. G. D. Geers. 2007a. “An enhanced Unger, J. F., and C. Könke. 2008. “Coupling of scales in a multiscale sim-
multi-scale approach for masonry wall computations with localization ulation using neural networks.” Comput. Struct. 86 (21–22):
of damage.” Int. J. Numer Methods Eng. 69 (5): 1022–1059. https://doi 1994–2003. https://doi.org/10.1016/j.compstruc.2008.05.004.
.org/10.1002/nme.1799. Verhoosel, C. V., J. J. C. Remmers, M. A. Gutiérrez, and R. de Borst. 2010.
Massart, T. J., R. H. J. Peerlings, and M. G. D. Geers. 2007b. “Structural “Computational homogenization for adhesive and cohesive failure
damage analysis of masonry walls using computational homogeniza- in quasi-brittle solids.” Int. J. Numer. Methods Eng. 83 (8–9):
tion.” Int. J. Damage Mech. 16 (2): 199–226. https://doi.org/10.1177 1155–1179. https://doi.org/10.1002/nme.2854.
/1056789506064943. Yvonnet, J., D. Gonzalez, and Q.-C. He. 2009. “Numerically explicit po-
MATLAB R2010a. 2010. User’s guide. Natick, Massachusetts: MathWorks. tentials for the homogenization of nonlinear elastic heterogeneous ma-
Mercatoris, B. C. N., P. Bouillard, and T. J. Massart. 2009. “Multi-scale terials.” Comput. Methods Appl. Mech. Eng. 198 (33–36): 2723–2737.
detection of failure in planar masonry thin shells using computational https://doi.org/10.1016/j.cma.2009.03.017.
homogenisation.” Eng. Fract. Mech. 76 (4): 479–499. https://doi.org/10 Zohdi, T. I., and P. Wriggers. 2008. An introduction to computational mi-
.1016/j.engfracmech.2008.10.003. cromechanics. Dordrecht, Netherlands: Springer.

© ASCE 04018072-10 J. Eng. Mech.

J. Eng. Mech., 2018, 144(8): 04018072

Das könnte Ihnen auch gefallen