Computer Methods and Recent Advances in Geomechanics Oka, Murakami, Uzuoka & Kimoto (Eds.

)
2015 Taylor & Francis Group, London, ISBN 978-1-138-00148-0

Numerical modeling of discrete spatial heterogeneity in seismic

risk analysis: Application to treated ground soil foundation
S. Montoya-Noguera, F. Lopez-Caballero & A. Modaressi-Farahmand-Razavi
Laboratoire MSSMat CNRS UMR, Ecole Centrale Paris, France

ABSTRACT: A discrete autoregressive model is used to analyze the behavior of heterogeneous soil in a
probabilistic framework. Two-phase mixtures are common in geotechnical engineering either to represent inherent spatial variability or the effect of a soil improvement technique. In this study, the spatial distribution of
two types of soils (a liquefiable and a treated sand) is added on a two-dimensional finite-element soil-structure
interaction model. The soil behavior is represented by an elasto-plastic multi-mechanism model (Aubry et al.
1982). To account for the natural hazard, real input motions from the PEER database were used. The co-seismic
settlements of the structure are estimated for different mixture fractions and for different spatial distributions. The results show that the traditional homogenization techniques may lead to unsafe design while the
Generalized-Effective Medium (GEM) equation accounts for interaction behavior between loose and dense
deposits.

INTRODUCTION

In geotechnical earthquake engineering, there are

many sources of variability and uncertainty. Natural
processes and the geological formations present inherent spatial variability; in addition, limited observations
and measurement errors add an important amount of
uncertainty. Commonly, variabilities and uncertainties in the soil profile are addressed with probabilistic
analyzes. The soil properties change randomly following a specified probabilistic distribution; though,
spatial variability caused by different geological mechanisms is more likely to be a distribution of different
soil deposits with horizontal and vertical correlations
than a random continuous change.
Liquefaction is described as the loss of strength
and stiffness caused by the increase of excess pore
water pressure (
pw ). In heterogeneous soils it has
been observed that the
pw is generated first in
loose deposits and when water migrates to neighboring dense sand, it is softened and eventually liquefied.
This interaction affects also the liquefaction-induced
settlements. Extensive research has been performed
on the calculation of these settlements but rarely
they include the variability of soil properties and modeling of different soil deposits (Koutsourelakis et al.
2002, among others).
Furthermore, soil-mixing and densification techniques to strengthen soft soils, strengthen foundations and mitigate liquefaction are becoming widely
used even though there remains some uncertainties
given the spatial variability introduced in the design
(Kasama et al. 2012). One of the most influential
factors on the treated ground is the effectiveness of the

mixing process (i.e. spatial fraction of the treated soil)

which adds important variations with different spatial
configurations on the vertical as well as in the horizontal direction. In this analysis, the treated ground
soil foundation is a two-phase mixture composed of
two materials: a medium-to-dense sand (MDS) and the
remaining loose-to-medium sand (LMS). The treated
area is below the structure with more than thrice its
width and up to 5 m depth. To assess the effectiveness of the mixing process, the co-seismic settlements
of the structure are estimated for spatial fractions
ranging from untreated to fully-treated and with different spatial distributions for each spatial fraction.
The results are compared with effective values from
homogenization techniques.

NUMERICAL MODEL

A typical layered soil/rock model is considered. The

soil profile is composed of 20 m of loose-to-medium
sand and it is divided in four layers. The average
weighted shear wave velocity (Vs,20 ) of the soil profile
is 154 m/s. The shear modulus increases with depth.
The fundamental elastic period of the soil profile, calculated with a low-strain frequency analysis, is equal
to 0.38 s. An elasto-plastic multi-mechanism model is
used to represent the soil behavior on the top 20 m.
Under the soil deposit, a bedrock representing a halfspace medium is modeled with an isotropic linear
elastic behavior and a shear wave velocity (Vs ) equal
to 550 m/s. The ground water table is 1 m below the
surface.

1277

Concerning the treated ground soil an area of

20 m width and 4 m depth is placed below the water
table. The heterogeneity is added with a discrete
autoregressive-model composed of two materials: a
treated medium-to-dense sand (MDS) and the remaining loose-to-medium sand (LMS). An illustration of
the model is shown in figure 1.

2.1

Finite element model

A 2D dynamic approach derived from the upw version of the Biots generalized consolidation theory
(Zienkiewicz & Taylor 1991) was adopted for the
soil. The model uses quadrilateral isoparametric elements with eight nodes for both solid displacements
and fluid pressures. The thickness of the elements is
0.5 m. For a maximum frequency of 20 Hz, a minimum
of 12 points per wavelength is allowed. An implicit
Newmark numerical integration scheme is used in the
dynamic analysis ( = 0.611 and = 0.301) for the
discretization in time (Kuhl & Crisfield 1999). To
take into account the interaction effects between the
structure and the soil, a modified width plane-strain
condition (Saez et al. 2013) was assumed in the finite
element model. In this case a width of 4 m is used.
2.1.1 Boundary conditions
As the signal propagation is 1D and as the response of
an infinite semi-space is modeled, equivalent boundaries have been imposed on the lateral nodes (i.e. the
normal stress on these boundaries remains constant
and the displacements of nodes at the same depth in
two opposite sides are the same in all directions). For
the bedrocks boundary condition, paraxial elements
simulating deformable unbounded elastic bedrock
have been used (Modaressi & Benzenati 1994). The
vertically incident shear waves, defined at the outcropping bedrock, are introduced into the base of the model
after deconvolution. Thus, the obtained movement at

2.2 Input earthquake motions

The earthquake signals used are near-to-source strong
motions recorded in dense soil conditions. These signals are supposed to have minimal noise and are
appropriate for an outcropping bedrock condition. The
earthquake motions and some intensity measures are
shown in table 1. Figure 2 shows the acceleration time
histories and the response spectra of the signals with
a structural damping () of 5%.

2.3 Structural model

The reinforced concrete building is a large-scale, onespan, two-story frame model proposed by Vechio &
Emara (1992). The total structure height is 4.2 m and
the width is 4.0 m. The mass of the building is equal to
40 T and is assumed to be uniformly distributed along
beam elements while the columns are supposed massless. The fundamental period (Tstr ) is equal to 0.24s.
In order to simulate the structure, plastic hinge beamcolumn elements are used which take into account
axial force (P) and bending moment (M ) interaction
(Prakash et al. 1993). The foundation is modeled as a
rigid block of 0.1 6 4 m.
Figure 3 shows the transfer function of the free-field
(surface/bedrock) and of the structure (top/base) and
the predominant frequencies of the input motions for
the homogeneous untreated case. The relative position of the fundamental frequency of the soil with
respect to the fixed-base frequency of the structure
and the predominant frequency of the input motion is
very important for the inelastic dynamic-soil-structure
interaction (DSSI) effects (Saez et al. 2013).

Figure 2. Input earthquake motions: a) acceleration time

histories and b) Response spectra of acceleration.

Figure 1. Illustration of the numerical model.

Table 1.

the bedrock is composed of both the incident and the

reflected waves.

Ground motion database.

Event

Year

ID*

Mw

R [km]

Vs [m/s]

PHA [g]

IA [m/s]

D595 [s]

PGV [cm/s]

1
2
3

Northridge_01
Kocaeli, Turkey
Friuli, Italy_02

1994
1999
1976

1050
1165
133

6.69
7.51
5.91

4.9
3.6
14.4

2016
811
660

0.43
0.21
0.23

1.79
0.80
0.22

9.84
13.3
2.83

51.23
34.64
112.5

*ID of NGA database (http://peer.berkeley.edu/peer_ground_motion_database)

Joyner-Boore source-to-site distance.

1278

Figure 3. Transfer function of the free-field and the structure and the predominant frequencies of the input motions.

2.4

Spatial discretization

The spatial discretization is based on the homogeneous

autoregressive model derived by Bartlett & Besag
(1969). The heterogeneous zone is a binary mixture
with fractions M 1 and M 2 = 1 M 1 . The model is
discretized into a rectangular grid, where each element
is assigned one of the possible values: 0 (Material 1)
or 1 (Material 2). The expectation of a binary random
variable xij is:

where is the E[xij ] under the condition of homogeneity and 1 and 2 are the auto-regressive coefficients
that give the one step correlation of the process in
each direction. For each element, the generated probability is compared to a random number (uij ) between 0
and 1 making use of the Monte Carlo simulation (i.e.
binarization process). The random number is determined by the uniform distribution function and each
element is independent. If uij > , then xij = 0; otherwise, xij = 1. The procedure to generate the field of
size m n at each simulation is the following:
1. Assign the value for the corner element (x1,1 ):
Binarization process for E[xij ] = .
2. Assign the values for the lower horizontal boundary
xi=2,...,m,1 with 1 = 2 (i.e. 1D model).
3. Repeat step 2 but for the left vertical boundary
x1,j=2,...,n .
4. Fill up the interior of the rectangle xi=2,...,m,j=2,...,n
with equation 1.

Figure 4. Simulated laboratory tests: Volumetric strain (v )

as a function of the deviatoric strain (1 ) and b) Liquefaction
curves for the LMS and MDS.

the deviatoric mechanisms and only volumetric strains

for the isotropic one). To take into account the cyclic
behavior it uses a kinematical hardening which relies
on the state variables at the last load reversal. The
soil behavior is decomposed into pseudo-elastic, hysteretic and mobilized domains. Refer to Aubry et al.
(1982) and Hujeux (1985) among others for a complete
understanding of the ECP model.
3.1 Laboratory tests simulations
The soils used in the model are two sands with different
relative densities: LMS, loose-to-medium and MDS,
medium-to-dense. The soil parameters were determined with the procedure defined by Lopez-Caballero
et al. (2007) and can be found in Saez et al. (2013) and
DAguiar et al. (2011), respectively.
A triaxial drained test was modeled for different
confining pressures. The volumetric strains (v ) as a
function of the deviatoric strains (1 ) for both soils
are presented in figure 4a. For the LMS, the volumetric strains are below zero, which means the soil
deposit is principally contracting. On the contrary, the
MDS presents a dilatant behavior for the same initial
stress.
In addition, an undrained stress controlled cyclic
shear test was modeled. The cyclic stress ratio (SR =


/v0
) as a function of the number of loading cycles
to produce liquefaction (N ) is shown in figure 4b for
both soils.As qualitative comparison, the modelled test
results are compared with the curves given by Byrne
et al. (2004) for Nevada sand at different densities (i.e.
Dr = 40 and 60%). It is noted that the obtained curves
are closer to the reference for a Dr = 40% corresponding to the LMS; while, the MDS curves are closer to
those of a Dr = 60%.
4

LIQUEFACTION ANALYSIS

SOIL CONSTITUTIVE MODEL

The soil behavior is simulated with the ECP elastoplastic multi-mechanism model. It can take into
account a large range of deformations. The model is
written in terms of effective stresses, uses a Coulomb
type failure criterion and follows the critical state
concept. The evolution of hardening is based on the
plastic strain (deviatoric and volumetric strains for

In this section, the response of the model without treatment is analyzed. For the sake of brevity, figure 5a


shows the pore pressure ratio (ru =
pw /v0
) at the
end of shaking (i.e. co-seismic analysis) for only EQ2.
When ru is equal to unity, the soil has loosen all its
strength and its totally liquefied. The liquefied area
below the structure is found above 5 m depth and is
thrice as large; thus this would be the treatment region.

1279

Figure 5. Liquefaction ratio ru at the end of shaking for

EQ2: a) before and b) after soil treatment.

LIQUEFACTION MITIGATION ANALYSIS

Liquefaction mitigation methods like vibrocompaction or preloading use vibratory or potential

energy to densify loose cohesionless soils at depth.
Deep mixing techniques like jet-grouting, biogrouting or wet-soil mixing use stiffer materials to
form foundation elements to stabilize soft soils,
strengthen foundation-soils and mitigate liquefaction.
These techniques are widely used even though there
remains some uncertainties given the spatial variability introduced in the design (Kasama et al. 2012).
Sometimes, given the poor soil conditions and the
high seismic demand, soil mixing is used to replace
an entire portion of the deposit. This case is analyzed and the liquefaction ratio at the end of shaking
is shown in figure 5b for the EQ2. It is interesting to note that while ru decreases directly under
the structure, it increases for higher depths and on
free-field. These results are in accordance with those
of Coelho et al. (2004) and Lopez-Caballero &
Modaressi-Farahmand-Razavi (2013). Furthermore,
on the boundaries between both soils, a pore-pressure
migration can be seen from loose untreated sand.
An important interaction is therefore present in the
mixture.
5.1

Heterogeneous model

One of the most influential factors on the treated

ground is the effectiveness of the mixing process,
which adds important variations in the soil. This heterogeneity depends on the mitigation technique but
will in general have a spatial correlation on the vertical and horizontal direction. Studies on this topic
usually apply homogeneous random field techniques
to account for the spatial variation. Properties are
described by a probabilistic function where a mean
value and a coefficient of variation is provided. In
this analysis, the heterogeneous deposit is composed
of two materials: a treated-soil and the remaining soft
soil. The interaction of these soils is important as the
loose sand induces an increase in excess pore-water
pressure in the surrounding dense sand. Chakrabortty
& Popescu (2012) analyzed centrifuge experiments of
heterogeneous sand deposits and compared with an
average-density homogeneous sand. Results show an

Figure 6. Heterogeneous zone for 2 distributions with

M 1 = 0.4 and 1 = 2 = 0.4.

Figure 7. Relative settlement (uz ) of the structure with

respect to free-field for a) EQ1 and b) EQ3.

increase in excess pore water pressure for the heterogeneous case, although the analysis was deterministic and the use of a homogeneous equivalent model
could be inappropriate. The spatial discretization
described in section 2.4 is used to analyze the heterogeneous deposits. The spatial correlation on both directions is assumed to be the same and equal to 0.4. The
spatial fraction (LMS ) is varied and 20 distributions
per value were realized. Two distributions for the same
spatial fraction (LMS = 0.4) are shown in figure 6.
Figure 7 shows the relative settlement (uz ) of the
structure with respect to free-field for EQ1 and EQ3.
The box-and-whiskers plot is used to show scalar-value
statistics for the different spatial fractions. The mean
values are in red and joined by the curve. The data
dispersion increases with the settlement and the spatial fraction. The variation on the box and whiskers
size suggests that the discrete model is highly dependent on the earthquake motion and on the interaction
between the two materials.
5.2 Generalized Effective Medium (GEM)
The general behavior of complex heterogeneous
materials is rarely the average of their constituent

1280

behavior. The effective medium depends on the geometry (e.g. shape and size of particles) and the topology
or connectivity among particles. Homogenization theories, mostly based on the mean field theory, provide
analytical solutions for the effective properties. These
models require many high-order terms to account for
micro-structural information, thus, there is no guarantee that they can capture the proper connectivity.
However, percolation theories as statistical tools have
been developed to describe the topology of heterogeneous materials. Traditional homogenization theories
are based on the geometric arrangements among the
phases, e.g. parallel and series (Wiener 1912) or concentric (Hashin & Shtrikman 1962). Effective medium
theories (EMT) are derived on the exact solution for
a single spherical inclusion in an infinite medium and
therefore neglect interactions among particles. The
percolation theory is focused on the existence of a
percolating cluster that connects opposite sides of a
system. The first emergence of the percolation cluster
corresponds to the critical fraction pc . This threshold
depends on the system size, dimensionality and correlation, among others. McLachlan et al. (1990) derived
a semi-empirical correlation that includes the effective
diffusivity in the EMT.The resulting equation is known
as the generalized effective medium (GEM) equation:

where s and t are the scaling exponents for p < pc

and p > pc , respectively. They are universal percolation quantities and are independent on the discrete
or percolation type, the size nor the correlation of
the system. Their values can be found via numerical
simulations (Sahimi 2003). In the present study, p is
the spatial fraction LMS and D is the relative settlement uz . So, D1 and D2 are the relative settlements
when LMS = 0 and LMS = 1, respectively.
Figure 8 shows the mean relative settlement (uz )
of the structure with respect to free-field for EQ2
and the traditional homogenization theories. For spatial fractions less than 20%, data can be found between
the bounds presented; while for higher values, the
effective settlement was found to be higher than the
mean and therefore over the parallel Wiener+ and
HS+ bounds. The use of these homogenization techniques will underestimate the maximum settlement of
the structure for LMS >20%.
Figure 9 shows the GEM equation for the three
earthquake motions. The percolation threshold pc is
around 0.4 for the three cases which is in accordance
with the literature. For p > pc , the behavior is almost
the same for all cases; while for p < pc , the response
varies greatly. The GEM equation gives a well suited
fit for the effective settlement of the structure under
the three earthquake motions tested. However, this

Figure 8. Normalized relative settlement (uz ) of the structure with respect to free-field and traditional homogenization
theories for EQ2.

Figure 9. Normalized mean relative settlement (uz ) of the

structure with respect to free-field and GEM equation for
EQ1, EQ2 and EQ3.

behavior is likely to depend on the input motion and

will be further investigated. By the moment, it appears
that the efficiency of the soil treatment is only important for a spatial fraction (LMS ) between 20 and 60%.
While for LMS above this range, the settlement is
reduced by less than 20%. Similarly, for LMS less than
20%, the efficiency is not greatly reduced (less than
15%). The latter can be explained by the theory of the
percolation threshold, i.e. that after a certain value (in
this case LMS = 0.6), there is a well-connected cluster
that connects opposite sides of the system.

CONCLUSIONS

A numerical model of discrete spatial heterogeneity

was used to assess the liquefaction induced settlement of an inelastic soil-structure system. Correlation
in both directions was introduced by the nearestneighbor model. The effect of the spatial fraction ()
of the treated-untreated soil mixture was analyzed.
Twenty simulations were used for each of the nine different and for each of the three unscaled recorded

1281

ground motions. The main conclusions of the present

study are:

For the three earthquake motions tested, the case

of the fully-treated soil reduces the relative settlement of the structure with respect to free-field but
increases the liquefaction ratio under and around the
treated area.
Because of excess pore-pressure migration, treated
soil surrounded by loose sand can present liquefaction and therefore low fractions of soil treatment
have very low efficiency. Similarly, once the loose
sand is well surrounded by treated soil, improvement in the relative settlement is small.
The generalized-effective-medium (GEM) equation
is well fitted for the effective mean settlement as
a function of the spatial fraction; while, the traditional homogenization techniques will overestimate the efficiency of the inclusions for small
treated-untreated fractions.
For high treated-untreated fractions, the effective
settlement is highly dependent on the earthquake
motion and their relation should be further investigated.
Furthermore, the effect of the correlation in the
horizontal and vertical direction should be investigated and, if possible, related to real in-situ measurements. The numerical model of discrete spatial
variation applied in a probabilistic framework appears
to properly include heterogeneities on the soil. Cases of
inherent spatial variability or other soil improvement
techniques could be also analyzed with this model.

REFERENCES
Aubry, D., J. Hujeux, F. Lassoudire, and Y. Meimon (1982).
A double memory model with multiple mechanisms for
cyclic soil behavior. In International Symposium Num.
Mod. Geomech., Balkema, pp. 313.
Bartlett, M. S. and J. E. Besag (1969). Correlation Properties
of Some Nearest-Neighbor Models. Bull. International
Statistical Institute 43(2), 191193.
Byrne, P., S.-S. Park, M. Beaty, M. Sharp, L. Gonzalez,
and T. Abdoun (2004). Numerical modeling of liquefaction and comparison with centrifuge tests. Canadian
Geotechnical Journal 41(2), 193211.
Chakrabortty, P. and R. Popescu (2012). Numerical simulation of centrifuge tests on homogenous and heterogeneous soil models. Computers and Geotechnics 41(1),
95105.
Coelho, P., S. Haigh, S. Madabhushi, and T. OBrien (2004).
Centrifuge modeling of the use of densification as a
liquefaction resistance measure for bridge foundations.
In 13th World Conference on Earthquake Engineering,
Paper 210, Vancouver, B.C., Canada.

DAguiar, S. C., A. Modaressi-Farahmand-Razavi, J. A.

Dos Santos, and F. Lopez-Caballero (2011). Elastoplastic constitutive modelling of soil structure interfaces
under monotonic and cyclic loading. Computers and
Geotechnics 38(4), 430447.
Hashin, Z. and S. Shtrikman (1962). A variational approach
to the theory of the effective magnetic permeability of
multiphase materials. Journal of Applied Physics 33,
3125.
Hujeux, J.-C. (1985). Gnie Parasismique: Une loi de comportement pour le chargement cyclique des sols (V.
Davidovici ed.)., pp. 278302. France: Presses ENPC.
Kasama, K., A. Whittle, and K. Zen (2012). Effect of spatial variability on the bearing capacity of cement-treated
ground. Soils and Foundations 52(4), 600619.
Koutsourelakis, S., J. Prevost, and G. Deodatis (2002). Risk
assessment of an interacting structure-soil system due
to liquefaction. Earthquake Engineering and Structural
Dynamics 31(4), 851879.
Kuhl, D. and M. A. Crisfield (1999). Energy-conserving
and decaying algorithms in non-linear structural dynamics. International Journal for Numerical Methods in
Engineering 45, 569599.
Lopez-Caballero, F. and Modaressi-Farahmand-Razavi
(2013). Numerical simulation of mitigation of liquefaction seismic risk by preloading and its effects on the
performance of structures. Soil dynamics and Earthquake
engineering 49(1), 2738.
Lopez-Caballero, F., A. Modaressi-Farahmand-Razavi, and
H. Modaressi (2007). Nonlinear numerical method for
earthquake site response analysis i- elastoplastic cyclic
model and parameter identification strategy. Bulletin of
Earthquake Engineering 5(3), 303323.
McLachlan, D., M. Blaskzkiewics, and R. Newnham (1990).
Electrical resistivity of composites. Journal of the American Ceramics Society 73(8), 21872203.
Modaressi, H. and I. Benzenati (1994). Paraxial approximation for poroelastic media. Soil Dynamics and Earthquake Engineering 13(2), 117129.
Prakash, V., G. Powel, and S. Campbell (1993). DRAIN 2DX,
Base program description and user guide.
Saez, E., F. Lopez-Caballero, & A. Modaressi-FarahmandRazavi (2013). Inelastic dynamic soil-structure interaction effects on moment-resisting frame buildings. Engineering structures 51(1), 166177.
Sahimi, M. (2003). Heterogeneous materials 1: Linear transport and optical properties. New York: Springer.
Vechio, F. and M. Emara (1992). Shear deformation in reinforced concrete frames. ACI Structural journal 89(1),
4656.
Wiener, O. (1912). Die theorie des Mischkorpers fur
das feld des stationaren Stromung. Erste Abhandlung
die Mittelswertsatze fur Kraft, Polarisation un Energie.
Abhandlungen der mathematisch-physischen Klasse der
Kniglich Schsischen Gesellschaft der Wizzenschaften
32(6), 509604.
Zienkiewicz, O. and R. Taylor (1991). The Finite element
method, solid and fluid mechanics, dynamics and nonlinearity (4th ed.), Volume 2. London: McGraw-Hill Book
Company.

1282