SELECTION OF PHYSICAL 2D PROBABILISTIC REALIZATIONS OF

SHEAR WAVE VELOCITY RANDOM FIELD

Eliane YOUSSEF1, Elias EL HABER2, Dalia ABDEL MASSIH3, Cécile CORNOU4, Tamara AL BITTAR5,
Fernando LOPEZ-CABALLERO6

ABSTRACT

Soil spatial variability in terms of Vs variability is modeled using random fields (RF) theory. Discretization is generally
adopted to evaluate the RF of a soil property at any position of a finite element or difference mesh. Several realizations of RF
are generated randomly to represent the statistical distribution of the soil parameter. However, not all the generated realizations
will be consistent with the real physical variability, classically expressed as a standard deviation on measured dispersion curve
at the surface by using active or passive surface waves methods. Thus, this paper suggests a numerical methodology for the
selection of the shear wave velocity (Vs) RF realizations that are “physical” and could be further used for quantifying the
effects of soil spatial variability on seismic response. The Expansion Optimal Linear Estimation (EOLE) method is used to
discretize the RF of Vs and several realizations for different statistical cases are generated. In order to determine the dispersion
curve of a heterogeneous soil representing one RF realization, the MASW experiment is modeled using finite difference
method incorporated in FLAC2D software. Derived dispersion curves are compared to the one of the mean homogeneous
deterministic model. All curves located outside the range of +/-3 standard deviation of the mean curve are rejected and the
realization is considered as nonphysical. The effect of the lateral soil variability, i.e. the horizontal autocorrelation distance
(θx) and coefficient of variation (COV) on the selection of the physical realizations is presented and analyzed.

Keywords: Spatial variability; Random field; Shear-wave velocity; MASW; Numerical modeling

1. INTRODUCTION

“Natural soil is never uniform. Its properties change from point to point while our knowledge of its properties is
limited to those few spots at which the samples have been collected…” (Karl Terzaghi).

However, estimation of seismic response most often relies on mean values of soil elastic properties while
neglecting the effect of near-surface spatial heterogeneities of the subsurface structure caused by natural processes
of erosion, weathering and deposition and anthropogenic activities (Einsele, 2000). Accounting for variable soil
properties in seismic response assessment would require knowledge of statistical parameters of the variable soil

1
Scientific Research Center in Engineering, (CSRI), Faculty of Engineering, Geotechnical Engineering, Lebanese University,
Lebanon; e-mail: eliane-youssef@hotmail.com
2
ISTerre, UJF/CNRS/IRD, Université Grenoble Alpes, ISTerre CS 40700 38058 Grenoble Cedex 9, France; e-mail: elias.el-
haber@univ-grenoble-alpes.fr
3
Scientific Research Center in Engineering, (CSRI), Faculty of Engineering, Geotechnical Engineering, Lebanese University,
Lebanon; e-mail: dalia.abdelmassih@ul.edu.lb
4
ISTerre, UJF/CNRS/IRD, Université Grenoble Alpes, ISTerre CS 40700 38058 Grenoble Cedex 9, France; e-
mail: cecile.cornou@univ-grenoble-alpes.fr
5
Scientific Research Center in Engineering, (CSRI), Faculty of Engineering, Geotechnical Engineering, Lebanese University,
Lebanon; e-mail: tamara.albittar@gmail.com
6Laboratoire MSSMat, CNRS UMR 8579, CentraleSupélec, 8/10 Rue Joliot-Curie, 91190 Gif-Sur-Yvette, France; e-

mail: fernando.lopez-caballero@ecp.fr
 2

properties and their probability distributions which are not easy to get because of the high-cost of traditional
geotechnical invasive tests when lateral soil variability is targeted. On the other hand, geophysical tests based on
seismic surface waves methods, are becoming more commonly used in geotechnical engineering (e.g. Foti, 2000)
to define the elastic properties of the subsoil.
Random fields (RF) theory is usually used to model the variable property allowing thus to introduce the
quantified uncertainty related to soil spatial variability in any analysis. Discretization based on spatial averaging
or series expansion methods is generally adopted to evaluate the RF at any position of a finite element or difference
mesh (Sudret, 2000). Several realizations of RF are generated randomly to represent the statistical distribution of
the soil parameter in question. Ground motion simulation at surface receivers is then performed in order to evaluate
the impact of spatial variability of the soil properties on surface ground motion indicators (Thompson et al., (2009);
Pagliaroli et al., (2014)).
In order to select the physical 2D shear-wave (Vs) profile (Griffiths et al., 2016), this paper suggests a
numerical methodology for choosing 2D Vs RF realizations that lead to dispersion curves consistent within a
specified uncertainty bound with the experimental dispersion curve that is extracted from surface wave methods.
In this study, the random fields of VS are discretized using the Expansion Optimal Linear Estimation method
(EOLE) and generated using Monte Carlo simulations.
Furthermore, MASW (multi-channel analysis of surface waves) experiment is modeled using finite difference
method incorporated in FLAC2D software and the dispersion curve of each RF realization is evaluated and
compared to the deterministic dispersion curve derived when considering a homogeneous velocity structure. All
dispersion curves, located outside the range of +/- 3 standard deviation of the deterministic dispersion curve, are
rejected, and the corresponding realization is considered nonphysical. The effect of the lateral soil variability
properties on the selection of the physical realizations is presented and analyzed.

2. PROBABILISTIC MODEL

2.1 Soil Data statistical parameters and discretization method

The soil model used in this paper has been derived from the simplification of the subsoil of the alluvial plain of
Nahr Beirut region characterized by Salloum et al. (2014). The sedimentary layers of this region were homogenized
into a clayey-gravel unique sediment layer of 15.5 m depth overlaying a seismic bedrock (Figure 1).
In the probabilistic analysis, Vs is modeled as a RF following a well-known probability distribution that
has a mean value (µVs), a coefficient of variation (COVVs), and an autocorrelation function which is defined by
vertical and horizontal autocorrelation distances (θy and θx, respectively). The autocorrelation distance indicates
the lag in space beyond which Vs values are considered as statistically uncorrelated in a certain direction.
The quantified uncertainties related to the spatial variability of soil shear wave velocity have shown an
average µVs of 220m/s and a coefficient of variation COVVs ranging between 20% and 40% for the considered
sediment layer. The corresponding density ρ is taken equal to 1600 kg/m3. A squared exponential autocorrelation
function given by Equation 1 is used:

 
2

    exp   
  (Eq. 1)

Where  is the lag (data interval) distance and  is the autocorrelation distance. Concerning the values of
the autocorrelation distances, y was fixed to 2 m, and two values were considered for x: 5 m and 10 m. For the
bottom bedrock layer assumed as a deterministic layer, Vs=1000 m/s, Vp=3000 m/s and =2500 kg/m3.
 3

Figure 1. 2D soil structure and mechanical properties of the homogenized sediment layer of Nahr-Beirut alluvial plain.

To discretize Vs RF, the Expansion Optimal Linear Estimation (EOLE) (Sudret & Der Kiureghian, 2000)
is used in this study. This method is an extension of the Optimal Linear Estimation (OLE) developed by Li and
Der Kiureghian (1993) and is based on the kriging method. The Monte Carlo method is then applied to generate
several RFs’ discretization. Vs is considered to follow a lognormal distribution. By using the EOLE method, VS at
each point of the stochastic domain is estimated from the computed autocorrelation matrix and its eigen values
and vectors (Sudret & Der Kiureghian, 2000).

2.2 Finite Difference model

In order to model the MASW experiment while considering the spatial variability of soil shear-wave structure, the
finite difference method was adopted using the software FLAC2D. A 2D plane strain model is used. The mesh size
Δl of the model is chosen less than one tenth of the minimum wavelength to remove numerical dispersion
phenomena (Eq. 2) (Bourdeau C., 2005).

λmin VS
Δlmax ≤ 10
≤ 10.fmin (Eq. 2)
max

Where ∆lmax is the maximum size of the finite difference mesh, λmin is the minimum wave length, Vsmin is
the minimum shear wave velocity, and fmax is the maximum wave frequency. In all the discretized random fields
used in FLAC2D simulations of this paper, the minimum value of Vs was found to be 50 m/s. Provided that the
maximum computed frequency is 25 Hz, Δl should be less than 0.2 m as per Eq. 2. The minimum computational
frequency is 1 Hz. Regarding the boundary conditions, along the lateral left boundary of the model, zero horizontal
displacements are applied and at its base the vertical movements are fixed allowing to reach the initial equilibrium
condition of the model. In a second step, the initial stresses due to the weight of soil are calculated in FLAC 2D by
applying a gravity acceleration (i.e. 9.81 m/s2) in the negative y-direction. When the boundary conditions and
initial constraints have been defined, FLAC2D ensures that the steady state is reached. As for the shear-wave
structure, the generated Vs field will be incorporated in FLAC2D model by data exchange with Matlab according
to the equation 𝐺 = 𝜌 × 𝑉𝑠 2 where G is the shear modulus of the soil. Once the stresses are initialized, a seismic
vertical acceleration excitation of a plane wave of type SV is applied at the surface of the model. The source time
function is a pseudo-Dirac having a flat Fourier amplitude spectrum equals to 1 between 1 Hz and 25 Hz and it is
applied as a history input (Figure 2). Also, a quiet boundary is applied to the model base and to the right-side edge
of the model so that the outward waves propagating from inside the model can be appropriately absorbed by the
 4

side boundaries and avoiding their reflection back into the model which can lead to an inadequate energy radiation
modeling. (Figure 3).
60 1.4

50
1.2
40

30 1

Fourier Spectrum [1]

20
Velocity [1]

0.8

10

0.6
0

-10 0.4

-20
0.2
-30

-40 0
0 1 2 3 4 5 0.1 1 10 25 50
Time [s] Frequency [Hz]

Figure 2. Input Signal and Fourier amplitude spectrum, [1] refers to no unit.

Figure 3. FLAC2D sketch model (Not to scale).

3. ACCURACY OF FLAC2D MODEL FOR THE MASW METHOD

3.1 Half Space model

In order to check the precision of the active surface wave method (MASW) simulation in FLAC 2D, a dispersion
curve obtained from FLAC2D model assuming an elastic half space with Vs=220m/s, Vp=1500 m/s and a
density=1600 kg/m3 is compared to the theoretical curve having a constant phase velocity of 209.5 m/s. In this
comparison, several models with different mesh sizes, horizontal and vertical boundary limits and types were
tested and the dispersion curve noted in the following “DC” of each case was derived using Geopsy.org software
and then compared to the theoretical one. These trials revealed that the vertical boundary limits have significant
effect on the accuracy of the DC while the horizontal limits showed no evident impact. Figure 5 illustrates the high
agreement between the theoretical dispersion curve and the one extracted from frequency-wavenumber (FK)
 5

spectrum for frequencies above 3 Hz, the lowest frequencies being affected by near-field effects (Socco &
Strobbia, 2004; O’Neil, 2005).

Time (s)
(a)

Time (s)
(b)

m/s

(c)

Figure 4. Vertical velocities simulated at the surface for (a) homogeneous sediment layer with Vs=220 m/s and (b) a Vs
realization with COV= 40%, θx= 5 m and θy = 2 m as displayed in (c). (c) Spatial distribution of Vs value.
 6

Figure 5. (Left) Half space FLAC2D model showing lateral and vertical limits and boundaries and (right) the FK spectrum
with the theoretical DC indicated by the horizontal black line and the O’Neil (2005) near-field wavelength criteria. The
magenta color indicate location of maximum FK spectrum.

3.2 Sediment Layer on bedrock - Deterministic case

Once the MASW model was validated in the software FLAC2D for the case of a halfspace model, the case of a
homogeneous layer over a bedrock is considered. In Figure 6, the FK spectrum obtained by analyzing FLAC 2D
surface velocities fits well with the theoretical dispersion curves for the fundamental and harmonic modes. The
computed frequency ranges between 1 and 25 Hz, however the minimum reliable frequency depends on the near-
field wavelength criteria. The strong decrease of phase velocity below 3.5 Hz is associated with the lack of energy
in the simulated velocities due to the filtering effects of the sediment structure (Scherbaum et al., 2003).

Figure 6. (Left) The FLAC2D sketch model with the same boundary conditions as the half space model and (right) the FK
spectrum of the deterministic model with the theoretical DC curves (black line) for the fundamental and higher modes and
the O’Neil (2005) near-field wavelength criteria (blue line). The magenta color indicate location of maximum FK spectrum.

3.3 Effect of the soil variability on the dispersion curves

The different parametric studies with varying values of COV, θx and θy are illustrated in this section by displaying
results obtained for one realization of the Vs random field.
Figure 7a and 7b show the VS random field distribution for a) of COVVs=5%, θx= 5m, θy=2m and b) of
COVVs=40%, θx=5m, θy=2m and FK spectrum computed from synthetic velocities computed at the surface when
mimicking MASW experiment. In case of 5% COVVs, Vs fluctuations do not produce any significant changes in
FK dispersion spectra, except larger excitation of higher modes (Figure 7a). Same observations were noticed for
 7

other values of the autocorrelation distances. However, in case of high COVVs values (and/or small autocorrelation
distances), Vs fluctuations lead to an FK dispersion spectrum being largely scattered (Figure 7b).

(a)

(b)

Figure 7. (a) (top) VS Random field distribution for COV 5%, θx= 5 m and θy=2 m and (bottom) FK dispersion spectrum.
The magenta color indicate location of maximum FK spectrum, the black line the O’Neil (2005) near-field wavelength
criteria and the black dots phase velocity estimates corresponding to maximum FK spectrum, (b) (top) VS random field
distribution with COV 40%, θx= 5 m and θy=2 m and (bottom) FK dispersion spectrum. The magenta color indicate location
of maximum FK spectrum, the black line the O’Neil (2005) near-field wavelength criteria and the black dots phase
velocity estimates corresponding to maximum FK spectrum.
 8

3.4 Selection of the physical Vs soil realizations

As previously mentioned, the generated random field realizations of Vs should represent the physical randomness
of the soil elastic variable properties. However, not all the realizations will provide dispersion curves that fit within
a given uncertainty bound of dispersion curves that one would observe when doing field measurement even though
they are statistically satisfactory with respect to the modeled probability distribution and autocorrelation function.
The main reason is that some of the generated Vs realization do not provide travel time close to the “deterministic”
travel time. While selection of physics based random profiles is easy in 1D situation (e.g. Teague and Cox, 2016),
it is more difficult when considering 2D spatial variability of Vs. In order to capture only those generated Vs
random fields that provides dispersion curves within +/- 3 standard deviation of the deterministic dispersion curve
(which in practice could be the experimental dispersion curve if available) are selected and the corresponding
realization is considered being “physical”.

Figure 8. Probabilistic dispersion curves for the 100 realizations (black lines) for COVVs = 40%, θx= 5 m and θy= 2 m, the
dispersion curve (red) of the deterministic ground model and the upper and lower limits +/- 3 standard deviation of the
deterministic dispersion curve (dotted red).

Figure 8 displays the DCs of 100 Vs RF realizations considering a COVVs = 40%, θx= 5 m and θy= 2 m. This figure
clearly reveals numerous dispersion curves outside the boundary limits which indicate that not all RF VS
realizations correspond to the real physical model. By assuming the frequency limit at 7 Hz (resulting from the
wavelength limit provided by O’Neil, 2005), only 51 curves (i.e. 51%) are located in the “physical” standard
deviation range.

4. CONCLUSION

This research concentrates on investigating the effect of soil spatial variability on the dispersion curves by
modeling the MASW procedure in the software FLAC2D. Considering variable soil properties (i.e. variable
autocorrelation distances and coefficient of variation of Vs), multiple realizations of Vs random field were
generated and utilized in MASW modeling. Derived dispersion curves at the surface of the ground are then
compared to the dispersion curve inferred from a deterministic Vs model (or the dispersion measured in the field).
 9

Such comparison provides a way to select the Vs RF realizations that provide dispersion curves in agreement with
the deterministic dispersion curve within a given uncertainty bound, which can be further used to study the realistic
impact of 2D soil spatial variability on seismic response. In this paper, the suggested methodology for the selection
of the physical dispersion curves is of high interest when real experimental dispersion curve exists for a given site.
For instance, the Vs RF realizations are generated following the statistical parameters of the site properties.
Conversely, no criteria usually exists to evaluate the validity of the generated RF (i.e. respect the soil real physical
behavior). Thus, by comparing the dispersion curves of the modeled RF realizations to the real experimental
dispersion curve acquired for example from a MASW experiment on a real soil, one can select the realizations
leading to a DC close to the experimental curve within a certain deviation interval. A realization producing DC
that do not fit well in the uncertainty bound of the experimental curve is rejected, reducing in this way the response
models computational time and conserving only the physical realizations which could lead to a realistic assessment
of the soil response.

5. ACKNOWLEDGMENTS

This work was realized in the framework of the JEAI – SAMMoVA research group funded by the “Institut de
Recherche pour le Développement (IRD) - France” and a research project on “the effect of soil spatial variability
on soils responses” financed by the Lebanese University.

6. REFERENCES

Bourdeau C (2005). Effets de site et mouvements de versant en zones sismiques: apport de la modélisation numérique. Ph.D
thesis. École Nationale Supérieure des Mines de Paris.

Einsele G (2000). Sedimentary basins: evolution, facies, and sediment budget. Springer.4.

Foti S (2000). Multistation methods for geotechnical characterization using surface waves: Ph.D Thesis, Politecnico di
Torino p. 229.

Griffiths S. C, Cox B. R, Rathje E. M and Teague D. P.(2016). Surface-wave dispersion approach for evaluating statistical
models that account for shear-wave velocity uncertainty: Journal of Geotechnical and Geoenvironmental Engineering,
142(11), 04016061.

Li C.-C and Der Kiureghian A (1993). Optimal discretization of random fields. Journal of Engineering Mechanics, 119 (6),
1136-1154.

O’Neill, Adam and Toshifumi, Matsuoka (2005). Dominant higher surface-wave modes and possible inversion pitfalls:
Journal of Environmental & Engineering Geophysics, 10(2):185–201.

Salloum N, Jongmans D, Cornou C, Youssef Abdel Massih D, HageChehade F, Voisin C and Mariscal A (2014). The shear
wave velocity structure of the heterogeneous alluvial plain of Beirut (Lebanon): Combined analysis of geophysical and
geotechnical data. Geophys. J. Int., vol.199, 2014, p.894–913.

Salloum N (2015). Evaluation de la variabilité spatiale des paramètres géotechniques du sol à partir de mesures géophysiques:
application à la plaine alluviale de Nahr-Beyrouth (Liban). Phd thesis, Universite Joseph Fourier, Grenoble.

Scherbaum F, Hinzen K. G and Ohrnberger M (2003). Determination of shallow shear wave velocity profiles in the Cologne,
Germany area using ambient vibrations: Geophysical Journal International, 152(3), 597-612.

Socco L and, Strobbia C (2004). Surface-wave method for near-surface characterization: a tutorial: Near Surface Geophysics,
2(4):165–185.
 10

Sudret B and Der Kiureghian A (2000). Stochastic finite element methods and reliability: a state-of-the-art report. Department
of Civil and Environmental Engineering, University of California.

Teague D. P, and Cox B. R (2016). Site response implications associated with using non-unique vs profiles from surface wave
inversion in comparison with other commonly used methods of accounting for vs uncertainty. Soil Dynamics and Earthquake
Engineering, 91, 87-103.

Thompson E, Baise L, Kayen R and Guzina B (2009). Impediments to predict in site response: Seismic property estimation
and modeling simplifications. Bull. seism. Soc. Am., 99 (5), 2927–2949.

Vanmarcke E (1983). Random fields: analysis & synthesis. MIT Press. Cambridge.