STOCHASTIC MODELS TO REPRESENT SOIL SPATIAL

VARIABILITY AND IMPACT ON SSI ANALYSIS

Irmela ZENTNER1, Georges DEVESA2, Didrik VANDEPUTTE3

ABSTRACT

It has been shown by several authors that the spatial variability of seismic ground motion may have a
major impact on extended and multi-supported structures such as dams, bridges or power plants.
Concurrently, the spatial variation of the ground motion has been evidenced by field experiments such
as the Pinyon Flat array (California) and the Argostoli array (Greece). In engineering applications, the
spatial variability is sometimes introduced by means of a coherency function that expresses the degree
of correlation of seismic motion on the ground surface. A very handy and convenient way to model
random variability of material properties consists in a representation by the Karhunen-Loève
decomposition. Analytical solutions of the latter can be deduced for most of the common correlation
kernels, in particular the exponential correlation function. This methodology has been implemented in
code_aster software and can be used to conduct full FEM SSI analysis accounting for spatial
variability of soil properties. The implementation is validated against surface coherency functions
available in the literature. We study the impact of the correlation length of soil parameters on the
coherency of the wave field and assess configurations were soil variability can play a major role on
structural response dynamics wavelength. We then consider a large foundation of an industrial
structure resting on the soil domain. We perform Monte Carlo simulation combined with LHS in order
to sample the random fields and propagate uncertainties. The structural response statistics obtained by
the full FEM model are then compared to the simplified approach.

Keywords: spatial variability, soil-structure interaction, random field, Karhunen-Loève, soil

parameters

1. INTRODUCTION

It has been shown by several authors that the spatial variability of seismic ground motion may have a
major impact on extended and multi-supported structures such as dams, bridges or power plants.
Concurrently, the spatial variation of the ground motion has been evidenced by field experiments such
as the Pinyon Flat array (California) and the Argostoli array (Greece). In engineering practice, the
spatial variability of the free field ground motion is often introduced by means of a coherency
function. The spatial variability of seismic ground motion has various origins such as wave passage
effects (oblique wave incidence), site effects and wave scattering in the spatially variable soil medium.
This paper focuses on the mathematical modelling of spatial variability of soil properties by means of
random fields. The Karhunen-Loeve (KL) expansion constitutes an efficient way to represent and
simulate such random fields defined by their correlation function. It allows to represent the continuous
random fields by means of a set of basis functions and a finite number of random variables.

1
Research engineer, EDF R&D, Lab Paris Saclay, France, [email protected]
2
Research engineer, EDF R&D, Lab Paris Saclay, France, [email protected]
3
Earthquake Geotechnical Engineer, EDF CEIDRE/TEGG, Aix-en-Provence, France, [email protected]
Nevertheless, for use in the framework of computational mechanics, the random fields have to be
discretized. The direct discretization and numerical computation of the multidimensional Karhunen-
Loeve expansion involves the solution of the Fredholm integral eigenvalue problem which may not be
feasible for large domains or very fine discretization. In this work, we adopt a non-intrusive approach
where the soil random fields are expressed by their KL expansion. In contrast to spectral
representation based methods, it does not require a regular mesh with equally distanced points.
Moreover, for common correlation kernels, there is an analytical solution to the Fredholm integral
equation allowing for an analytical expression of the random field (Ghanem & Spanos, 1991, Lemaitre
et al. 2010).

The methodology has been implemented in code_aster (Finite Element Method) code and allows for
assessing the impact of soil spatial variability on the structural response at affordable cost.

In what follows we:

1) first give some theoretical elements on the representation and simulation of soil properties
random fields and its numerical implementation.
2) then study the propagation of a vertical Ricker pulse in order to illustrate the phenomena and
gain further insights in the spatial variability of seismic motion observed on ground surface.
3) and compare the results of the full FEM approach to the simplified approach where the soil
variability is not explicitly modeled but where, instead, the ground motion spatial variability is
introduced by means of a coherency function.

2. STOCHASTIC SOIL MODEL

The soil random fields are generally supposed to be homogeneous (properties depend only on
separation distance and not on the spatial coordinate) and described by their correlation function.
The Markov and the more general von Karman model are the most popular for modelling random soil
properties in the geophysics literature, e.g. Sato et al (2012). The Gaussian correlation kernel has
however some advantages from a mathematical point of view, which is why it is often used for the
derivation of analytical models. The expressions of the different correlation models are shown in
Figure 1. Both the Markov and the Gaussian correlation function depend on the correlation length Lc.
In this work, we retain the Markov kernel.

Figure 1. Common correlation kernels for .

For the 3D random fields, defined by the coordinates we adopt the following expression
(Rackwitz, 2000) for the correlation function of the soil’s Young’s modulus (Equation 1):

(1)

This simple model allows for anisotropic behavior. Indeed, it is possible to consider correlation length
for the different spatial directions. The effect of the horizontal and the vertical correlation length are
illustrated in Figure 2.

2
 Lcx = Lcy Lcx >> Lcy

Figure 2. Examples of 2D soil random fields

2.1 Karhunen Loève decomposition

The Karhunen-Loève (KL) expansion is probably the most general approach for the representation non
Gaussian and non-stationary processes and random fields. In what follows we first give some general
elements on the KL expansion of spatial random fields. We then introduce the expression of the 3D
random fields retained in this work. The auto-correlation function, ,
defines a continuous self-adjoint Hilbert-Schmidt linear operator which has a countable number of
eigenvalues . The associated eigenfunctions are the solutions of the Fredholm
integral equation (Equation 2):

(2)

Then the random process U(x) has the following expansion (Equation 3):

(3)

Where the random variables are centered and normalized (zero mean and unit variance). They are
uncorrelated but possibly mutually dependent. In the case of a Gaussian process, the are standard
Gaussian uncorrelated random variables, and thus independent.

Obviously, a Gaussian distribution is not adequate for modeling physical quantities such as Young's
modulus, shear wave velocity or mass density that only take positive values. The lognormal
distribution is more adequate for representing physical quantities such as soil's Young's modulus.
Indeed, the lognormal law is the maximum entropy distribution that is the distribution that uses only
given information and adds no further assumption, if the median and the log-standard deviation of the
distribution are known. We consider lognormal target random fields G(x) with constant log standard
deviation G that can be obtained by transformation of the Gaussian random field. For simplicity’s
sake, we consider constant (not space dependent) median Gm. It is acknowledged, that a spatial trend
could be introduced by making the median depend on the spatial coordinate (increase with depth for
example).

For the 3D random process U(x,y,z) defined by its multiplicative correlation kernel, it can be shown
that the stochastic process U(x,y,z) can be expressed as (Equation 4):

(4)

where are independent random variables. By virtue of the Mercer's theorem it can be easily

3
verified that E(U(x,y,z)U(x',y',z’) yields the given product-correlation kernel. The construction of the
3D KL expansion requires the solution of the Fredholm integral equations. In the following section,
we show how this can be achieved in a very handy way by means of analytical solutions that can be
derived for the 1D KL expansion.

2.2 Numerical simulation

For use in the framework of computational mechanics, the random fields have to be discretized and
realizations of the latter have to be simulated. We adopt analytical solution of the Fredholm integral
equation based on the Markov kernel as derived in Ghanem & Spanos (1991) and Lemaitre et al.
(2001). The analytical solutions are developed for unitary domains (bounding boxes of the considered
2D or 3D domains) and then mapped to the FEM mesh (Gauss points). As an example, the first 10
highest eigenvalues and the first 5 eigenfunctions are shown in the Figure 3 below.

Figure 3. Example of eigenvalues and eigenfunctions computed with the analytical solution.

3. APPLICATION

In what follows, we first provide some qualitative analysis of the impact of soil spatial variability on
seismic ground motion by means of the full FEM approach accounting for soil variability as described
previously. We then quantitatively assess the resulting free-field ground motion incoherency by
comparison to empirical and analytical coherency functions available in the literature. Moreover, the
full FEM approach with soil variability is compared to the simplified methodology where the ground
motion variability is introduced by a coherency function. Eventually, the methodology is applied to
SSI analysis and the impact of soil variability on floor response spectra is studied.

The numerical analysis are performed with code_aster open source software using a full FEM
approach for SSI. This means that both the soil and the structure are modeled by the FEM. Absorbing
boundary conditions are applied to the lateral boundaries of the soil domain and to its bottom. The
paraxial elements available with code_aster (dashpots and additional stiffening elements) are applied.
The input motion is a vertically propagating plane SV wave. The seismic input is defined (i) at the
base as the incoming wave, (ii) at the lateral boundaries as the corresponding 1D propagated incoming
wave. When considering spatial variability, then the median Vs value was assigned to the stripes, that
is small portions of the soil domain, in the vicinity of the boundaries in order to comply with the
boundary conditions. For the soil domain, equivalent Rayleigh damping is used, where the damping
ratio is calibrated in the range 5Hz − 30Hz at a value of 25%. The Young’s modulus of the soil
domain is modeled as a lognormal spatial random field with the Markov correlation function. In order
to assure compatibility with the boundary conditions, uniform soil properties, assuming the median
value of the Young’s modulus are considered on small stripes next to the boundary. For the KL

4
decomposition, 60 terms are retained for each direction for the 2D analysis while 30 terms are retained
for each direction in the 3D analysis.

3.1 2D simulation

We first consider a rectangular soil domain with dimension 400m x 200m. Firstly, the response of the
soil domain to a vertically propagating SV-wave pulse is studied. The sine pulse is shown in Figure 4
(left) and has a fundamental frequency of 10 Hz. Assuming a median shear wave velocity V S=578m/s
leads to a dominant wavelength of around =60m. The coefficient of variation (COV) of the shear
wave velocity was assumed as 0.2. We furthermore assume isotropic soil where the correlation length
I horizontal and vertical direction are the same. The right hand side of Figure 4 shows the acceleration
of the soil domain (incoming wave) in horizontal direction for variable soil properties.

Figure 4. Vertically propagating SV-wave (sine pulse) in a 2D model: one example of variable soil domain (left
bottom) and two examples of resulting wave front incoherency (right). The lower right figure corresponds to the
soil domain shown on the lower left figure.

The simulations for different correlation lengths and fundamental frequencies of the incoming wave
highlight that the scattering is more important when the wavelength is close to the correlation length.

We now compare the ground surface coherency functions determined from the simulations to the
models in the literature, in particular the Abrahamson, the Mita & Luco and the Uscinki-Sato
coherency functions. In agreement with theory, a broad band stochastic excitation is used. The results
are shown in Figure 5 for a median Vs=578m/s and Lc= 60m.
.

5
 Figure 5. Coherency functions for different distances D

The coherency functions determined from the 50 simulations are quite close to the Sato-Uscinski and
the Mita & Luco model.

3.2 3D SSI model

We study the effect of spatial variability of the soil on the structural response. For this purpose we
consider a 3D soil domain with a superficial rigid rectangular foundation made of concrete as shown
in Figure 6:
 rectangular foundation : 48m x 48m x 3m
 soil domain : 200m x 200m x 150m
 median shear wave velocity: Vs = 566m/s, Lc = 60m, COV =0.2

A vertically propagating SV wave is considered. The waveform introduced at bedrock level (central
frequency on bedrock is 8Hz) is shown on Figure 6 (bottom left). We compute the in-structure
response spectra (ISRS) for the foundation. The response spectra determined for the 50 analysis
(yellow) are shown together with the median (solid blue) and the +1  spectrum (dashed blue) on the
bottom right of Figure 6. The response spectrum determined for the uniform soil profile is shown in
red in the same figure.
In Figure 7, the results are compared with the simplified approach where the spatial variability is
accounted for by the Mita & Luco coherency function in the framework of FEM-BEM analysis. The
coherency function is parameterized by the soil’s median Vs value and =0.4 which is in agreement
with the chosen soil variability. The FEM-BEM result is close to full FEM simulations (median ISRS)
with spatial soil variability. The free-field curve (magenta) has been obtained by 1D soil column
analysis and constitutes the input for the FEM-BEM analysis.

6
 Figure 6. Mesh, example of random field, time history (acceleration in m/s) and response spectra at foundation
level

The direct random simulation of the random fields might lead to some “unrealistic” soil profiles which
would not have been accepted by the geotechnical engineer. This is also visible in Figure 6 (bottom
right) where some soil samples lead to extremely high amplification around 10Hz. This problem could
be overcome by conditioning the random fields on values obtained for example by in situ geotechnical
investigations. Moreover, a trend representing the increase of stiffness of the soil with depth should be
implemented in order to improve the modelling.

Figure 7. Comparison to FEM-BEM approach (ISRS)

7
4. CONCLUSION

A stochastic spatial ground motion model based on an analytical KL-decomposition has been
implemented in code_aster. The methodology adopted makes use of the analytical solution of the
Karhunen- Loeve (KL) expansion. In consequence, it is not necessary to fully discretize the problem
and construct the correlation matrices. In particular, there is no need to numerically solve the
eigenvalue problem related to the KL decomposition. The random fields are evaluated at the Gauss
points through code_aster operators already available. The 2D studies could be ran on a PC without
considerably increasing the resources with respect to the case without variability (but more than one
simulation is needed, of course to compute the response statistics). The coherency functions
determined by numerical simulation were in good agreement with the models available in the
literature. The application to a simple SSI analysis showed, as expected, a reduction of the ISRS in the
higher frequency range, comparable with what is obtained with FEM-BEM analysis using coherency
function.

5. ACKNOWLEDGMENTS

The financial support from FP7 Nugenia+ is gratefully acknowledged.

6. REFERENCES

Ghanem R, Spanos Pol D. (1991). Stochastic finite elements: A spectral approach. Springer.
Le Maitre O, Knio OM (2000). Spectral Methods for Uncertainty Quantification. With Applications to
Computational Fluid Dynamics. Springer.
Mita A, Luco JE (1987). Response of structures to a spatially random ground motion.. ASCE J Engineering
Mechanics.113(1):1-15.
Rackwitz (2000). Reviewing probabilistic soils modelling. Computers and Geotechnics 3-4(26):199-223.
Salloum N, Jongmans D, Cornou C (2014). The shear wave velocity structure of the heterogeneous alluvial plain
of Beirut (Lebanon): combined analysis of geophysical and geotechnical data. Geophysical Journal International.
199(2):894-913.
Sato H, Fehler MC, Maeda T (2012). Seismic Wave Propagation and Scattering in the Heterogeneous Earth :
Second Edition. Springer Berlin Heidelberg.
Svay, Imtiaz, Perron, Zentner, Cottereau, Clouteau, Bard, Hollender, Lopez-Caballero (2017). Spatial
incoherency analysis of seismic ground motions from 2014-Argostoli earthquake dense array. Earthquake
Engineering & Structural Dynamics.
Thompson EM, Baise LG, Kayen RE, Guzina BB (2009). Impediments to Predicting Site Response: Seismic
Property Estimation and Modeling Simplifications. Bulletin of the Seismological Society of America. 99(5):,
927–2949.
Uscinski BJ (1977). The elements of wave propagation in random media. McGraw-Hill, New York.
Zentner I (2017). Génération de signaux sismiques Code_Aster documentation http://www.code-
aster.org/V2/doc/default/fr/man_r/r4/r4.05.05.pdf last accessed.
Zerva A (2009). Spatial Variation of Seismic Ground motion. Boca Raton: CRC Press.