Bulletin of the Seismological Society of America, Vol. 102, No. 2, pp. 802–811, April 2012, doi: 10.

1785/0120110091

Geometric Seismic-Wave Inversion by the Boundary Element Method

by Samuel Bignardi, Francesco Fedele,* Anthony Yezzi, Glenn Rix, and Giovanni Santarato

Abstract Surface-wave methods are widely used in earth sciences and engineering
for the geometric characterization of geological bodies and tectonic structures of
the subsurface. These techniques exploit the dispersive nature of Rayleigh waves
to indirectly estimate shear wave velocity profiles from surface-wave measurements;
however, they are limited to parallel-layered geometries. To overcome such limita-
tions, we present a new class of geometric inverse models for a full waveform inver-
sion (FWI) based on the boundary element method (BEM). The proposed approach
enables an effective identification of two dimensional (2D) subsurface geometries
by directly estimating the shape of laterally varying interfaces from raw measure-
ments. It thus aims at filling the gap between the standard simplistic parallel-
layered-based inversion and that of more complex three-dimensional (3D) geometries
based on finite element methods (FEMs). Numerical tests on synthetic data unveil the
effectiveness of the inverse algorithm, and its applicability to field measurements is
finally presented.

Introduction
The geometric characterization of geological bodies and Tselentis and Delis, 1998; Park et al., 1999; Strobbia and
tectonic structures of the subsurface is a crucial issue in many Foti, 2006). Other techniques utilize passively generated sur-
fields of earth sciences and engineering, and particularly for face waves to increase penetration depth (Louie, 2001; Rix
seismic hazard assessment. Indeed, a good geological knowl- et al., 2002; Park et al., 2005) and to enable inversions of soil
edge of the subsurface allows researchers to predict local seis- damping ratios via measurements of wave attenuation (Rix
mic effects for microzonation at large scales ∼O
103  m et al., 2000, 2001; Lai et al., 2002).
(sedimentary basins or alluvial valleys) or to locate with suf- The inverse problem at the root of SWMs requires a for-
ficient precision a fault trace and provide useful information ward model to solve the propagation, scattering, and dissipa-
for antiseismic management plans at smaller scales. Direct in- tion of seismic waves through the medium. This is idealized as
vestigations, such as drilling or excavations, yield pointwise a sequence of flat, horizontal, isotropic, and homogeneous
information of the subsurface, but they are very expensive, layers. Such models are capable of capturing only a discrete
even at the metric-to-decametric scale. Common alternative vertical variation in elastic properties, that is, they are one
methods are based on the analysis of passive (natural) or active dimensional (1D). One of their attractive features is that they
(artificial) seismic waves. In particular, in recent years surface are computationally efficient to calculate the modal displace-
wave methods (SWMs) have been widely used to reveal sub- ments and tractions using matrix operations (Kausel and
surface properties, especially those of S waves in shallow Roesset, 1981; Aki and Richards, 2002), which makes such
depths (∼30 m), which are relevant for planning any construc- models ideal for inverse problems employing iterative solu-
tion in areas prone to seismic risk. tions. Clearly, however, this model is only an approximation
Surface wave methods utilize the dispersive nature of of the actual subsurface conditions for most sites, and it may
Rayleigh waves in a vertically heterogeneous half-space to yield misleading results if the actual soil profile is far from the
obtain the shear wave velocity profile from the wave particle assumed geometry. Indeed, the major drawback is that wave
motions excited and recorded on the ground surface. To date, propagation is modeled as the linear superposition of disper-
many variations of the basic test and inversion protocol exist, sive Rayleigh modes that do not interfere or diffract as they
such as the two-receiver approach based on the spectral would if the profile were not horizontally layered. Observa-
analysis of surface waves (Nazarian et al., 1984; Stokoe et al., tions and modeling of earthquakes confirm, for example, that
1994) and the multichannel method that exploits arrays of seismic waves can be amplified at alluvial valley edges, and
receivers (e.g., Gabriels et al., 1987; Tokimatsu, 1995; diffraction effects can be observed because the interface be-
tween soft sediments and bedrock is far from being horizontal
*Also at School of Civil and Environmental Engineering, Georgia Insti- (Bard and Gabriel, 1986; Aki, 1988; Kawase, 1988; Kawase
tute of Technology, 225 North Avenue NW, Atlanta, Georgia 30332. and Aki, 1989). Furthermore, surface topography can alter

802
Geometric Seismic-Wave Inversion by the Boundary Element Method 803

seismic wave propagation (Bard, 1982; Raptakis et al., 2000; which soil properties of the medium may be considered as
Assimaki and Kausel, 2007). In particular, amplification is regionally homogeneous (Beskos, Leung and Vardoulakis,
usually expected at hilltops (Bard, 1982), and complex 1986; Beskos, Dasgupta and Vardoulakis, 1986; Manolis
amplified and deamplified patterns occur at hill flanks and Beskos, 1988; Xu, 2001; Katsikadelis, 2002). Indeed,
(Savage, 2004). BEM was previously exploited to solve problems concerning
Several attempts have been made to model lateral structural vibration analysis, transient waves, and dynamics
variations via the so-called pseudo-2D inversion (e.g., Luo of cavities under the influence of body or surface waves
et al., 2008), in which successive 1D inversions along the (Manolis and Beskos, 1981, 1983; Beskos, Leung and Var-
length of the survey are combined via interpolation or other doulakis, 1986; Beskos, Dasgupta, and Vardoulakis, 1986;
type of smoothing kernel; however, this approach is inher- Manolis and Beskos, 1988) and to investigate the propaga-
ently limited. Analytical solutions of forward models for true tion of seismic waves in laterally varying layered media (Di-
2D and 3D media are available, but only for a very restricted neva and Manolis, 2001a, 2001b).
class of weakly varying layered geometries (Gjevik, 1973; The proposed FWI approach is geometric because it
Maupin and Kennett, 1987; Dravinski and Mossessian infers the shape of 2D unknown irregular interfaces
1997; Maupin, 2007). As such, these solutions are of limited between regionally homogeneous soil layers directly from
practical value. single-frequency measurements at the receivers, instead
Thus, the further maturation of SWMs has been limited of an indirect inference from the observed dispersion curve,
by the lack of effective and efficient computational models or an interpretation from maps of elastic properties esti-
that enable more realistic representations of 2D/3D subsur- mated via FEM-based inversions. The elastic parameters
face media. Only recently, for crustal and lithospheric scale are assumed to be given, but they can be easily estimated
structures have such limitations been overcome by a full seis- together with the geometric properties of the subsurface.
mic waveform inversion (FWI) methodology that combines This joint inversion is fairly straightforward and will be dis-
an effective forward model based on the finite element meth- cussed elsewhere.
od (FEM) to model seismic wave propagation and adjoint The paper is structured as follows: we first introduce
techniques to efficiently compute the gradient of the data the analytical formulation of the forward model for the
misfit functional to perform minimization (e.g., Virieux and wave propagation through 2D laterally varying layers and
Operto, 2009; Fichtner, 2010). When an FEM forward model the associated numerical solution via BEM. Then a general
is used, complex media are finely discretized into a large inverse problem is introduced for the inference of subsur-
number of nodes or 3D volume elements. Each node is as- face interface geometries. To illustrate the potentialities
sociated with an unknown parameter, which characterizes of the geometric wave inversion, applications to benchmark
the soil seismic property at that node location. Typically, problems and to a set of experimental data are finally pre-
the number of parameters to be estimated largely exceeds the sented.
number of measurements available by several orders of
magnitude; thus, the inverse problem is severely ill-posed.
Ill-posedness is usually treated by regularization procedures BEM-Based Forward Model
(Tikhonov and Arsenin, 1977; Tarantola, 1987) by including Consider monochromatic waves at frequency ω propa-
additional constraints that yield well-posed inverse algo- gating through a 2D subsurface made of N layers, as shown
rithms. The shortcoming is that, as with most nonlinear in Figure 1a. Let Ωj be the generic layer bounded by the
inverse methods, these approaches suffer from slow conver- curve Γ  Γj1 ∪Γj , and x and z are the horizontal and ver-
gence and instability. Indeed, it is well known that a good tical coordinates, respectively. In the frequency domain, the
starting model is needed in FWI, and generally in inverse wave propagation in the generic j-th layer is governed by the
methods, to avoid false positives (i.e., false result indica-
tions) that may arise because local optimization does not pre-
vent convergence of the misfit function toward local minima. (i)
In this paper, we propose an alternative surface-wave (a) Receiver x k Γ0 Source x 0 Γs (b) 0
inversion formulation based on the boundary element meth- Ωj Target Γj Ω1
u+j t+j
od (BEM) (see, for example, Brebbia and Dominguez, 1992) -10 It. 500
Ω j+1 u-j
z [m]

t-j
that overcomes the previously mentioned limitations of u+j+1
It. 100
t+j+1 -20
FEM-based FWIs, and at the same time still exploits a far u-j+1 t-j+1
Ω j+2
z[m]

more realistic modeling of the subsurface than the current Initial guess Ω2
x[m] -30
state-of the-art SWMs. The proposed BEM-based FWI aims -30 -25 -20 -15 -10 -5 0
x [m]
at filling the gap between the standard simplistic parallel-
layered-based SWM and the more complex 3D FEM-based
Figure 1. (a) Model setup and (b) BEM-based inversion of syn-
FWI, enabling an effective and better geometric characteriza- thetic data on a two-layered media. Intermediate iterations show
tion at any geological scale. The associated forward model is convergence to the target interface. No noise added to simulated
solved by the BEM, which is very effective for applications in measurements at the receivers.
804 S. Bignardi, F. Fedele, A. Yezzi, G. Rix, and G. Santarato

following integral equation (Beskos, Leung and Vardoulakis, Beskos, Leung, and Vardoulakis (1986). The resulting linear
1986): system can be written in compact form as
Z Kug  tg ; (5)
U
j
x; s; ωtj
N
x; ωdΓ
Γ
Z where ug and tg are global degrees of freedom with respect to
 T
j
x; s; ωuj
x; ωdΓ  Cuj
x; ω; x∈Γ: (1) the fixed coordinate system as shown in Figure 1. The solu-
Γ tion of equation (5) follows after imposing loads and zero
stresses at the free surface. Weight drop or sledgehammer
The fundamental tensors U
j and T
j are given in Domin-
are modeled in Fourier space as a nodal stress vector of
iguez et al. (1984), and they depend upon the elastic para-
intensity f0
ω and perpendicular to the free surface. For
meters of the layer j, viz. density ρj, shear modulus μj , and
massive sources, such as an electromechanical shaker on a
Poisson’s ratio ν j . The
2 × 2 tensor C accounts for singular
portion Γs of the free surface, the generated stress is given
contributions of the Green tensor U
j, when x  s, as well as
by ts  mω2 ds  F, where the vectors F and ds represent the
for energy conservation in infinite domains (Dominiguez,
force produced by the shaker and the displacements of Γs ,
1993). Continuity of displacements and tractions at each in-
respectively, with m being the shaker’s mass. To test the nu-
terface requires that
merical implementation of equation (4), we have considered
the elastic problem of solving for the radial displacements on
u 
j
x; ω  uj
x; ω;
an infinitely long cylinder subject to a normal stress p. This
t
j
N 

x; ω  t
j
N 
x; ω; x∈Γj ; (2) admits an exact analytical solution given by Kitahara (1984).
We set the elastic characteristics of the cylinder as ρ 
for j  1; …N, where  and  superscripts refer to the va- 100 kg=m3 , μ  106 N=m2 , ν  0:25, damping ratio β r 
lues of stresses or displacements at the upper and lower 0:05, and the normal load f0  1002 N=m2 . Figure 2 shows
layers adjacent to the interface Γj , and N is the outward nor- the BEM numerical stresses at two different resolutions and
mal. Further, vanishing normal stress t 0
N must be enforced those expected from theory for 10 < ω < 150. The agree-
at the free surface Γ0 , except at the source location x0 ∈Γs ment with theory is remarkable (see also Dominguez, 1993).
where t0
N  f0
ωδ
x
x0 ^ z, f0 being the Fourier ampli- We point out that BEM can be easily generalized to multi-
tude of the load along the normal z^ to Γ0 , and δ
x is the frequency environments and 3D geometries. Further, viscoe-
Dirac function. Hereafter, we drop denoting the dependence lasticity can be accounted for by the elastic-viscoelastic
on ω, and the numerical approximation of equation (1) is correspondence principle of linear viscoelasticity (Christen-
obtained by discretizing each interface Γj in ej quadratic iso- sen, 1971; Manolis and Beskos, 1981).
parametric elements, along which the values of displace-
ments and tractions are given, respectively, by
Geometric BEM-Based Wave Inversion
X
3
uj
x  Pn
ηu
j
k;n ; Assume that the Fourier amplitudes udk
ω of the vector
n1 displacements are known from measurements using N rec
X
3 receivers at x
r
k ∈Γs and k  1; …N rec . Further, we focus
tj
N
x  Pn
ηt
j
k;n ; x∈Γj;k ; (3)
n1
12 8 Elements
where ujk;n and tjk;n are displacements and tractions vectors at 32 Elements
the nth node on the kth element Γj;k , and Pn
η are interpo- 10 Analytic
Displacements (1e-4 m)

lating functions with η∈1; 1. Using equation (3) and cy-
8
cling the point of application of the fundamental tensors
6m
through all the nodes, a system of linear equations for the
6
layer j follows as
4
K
j fu
j g  G
j 1 H
j fu
j g  ft
j g; j  1; …N; 6m

2
(4)
0
where the matrices G
j  and H
j  are given in Beskos, 0 0.5 1 1.5 2
ω/ω1
Leung, and Vardoulakis (1986), and the vectors fu
j g and
ft
j g list displacements and tractions at the boundary nodes,
Figure 2. Radial displacements of an infinitely long cylinder
respectively. These are degrees of freedom in a coordinate subject to normal stress as function of the frequency ratio ω=ω1 :
system local to the layer. Equation (4) is assembled by BEM versus analytical solution (Kitahara, 1984; Dominguez,
imposing the matching conditions of equation (2) following 1993) (resonance frequency ω1  59:73 rad=s).
Geometric Seismic-Wave Inversion by the Boundary Element Method 805

only on identifying the geometry of the boundary interfaces ments at the receivers due to perturbations of the interfaces
Γj and consider the elastic properties of the layers to be Γ
0
j . The matrix J is computed numerically via standard
known a priori. We point out that these parameters could second order finite-differencing from equation (4). We antici-
be easily estimated together with the geometry, but this pate that an adjoint solution (Eppstein et al., 2003; Fedele
will be discussed elsewhere. We consider an initially uniform et al., 2003, 2005) was formulated for Helmholtz-type equa-
discretization along x^ using nj nodes for each interface Γj , tions, and we are extending it to the Navier equations in order
located at x
j
j
j
k 
xk ; zk , k  1; …; nj . We calculate the to efficiently compute J for ndof ≫ 2N rec using active sur-
optimal iterative displacement of these nodes by computing faces (Yezzi et al., 2002, 2003).
the first variation of the following energy functional: From equation (6),
   
1X
N
Rec ∂E ∂uk T ∂uk
E ku  udk k2  uk0  Δp  udk
2 r1 k ∂Δp ∂Δp ∂Δp

N int X
X n 
e
j Z 1
 2  2 1 j
X
N int X
α rpi Δi  qp ; (10)
∂x
j
j 2
α k =∂η  ∂zk =∂η dη: (6) n1 i1
j1 k1 1

where T denotes Hermitian transpose, and

Here, the first term represents the mismatch between mea-
sured data at the receivers and the displacements uk calcu- X 3 Z
ne X 1
lated from the forward model in equation (4), hereafter rpi 
γ k  γ 3k ϕ2zk P0a P0b δ bl dη; (11)
referred to as the misfit E. The second term is an arc-length k1 a;b1 1
regularizer that ensures smoothness of the boundary inter-
face. This helps us avoid local minima of the energy repre- and
sented in equation (6) due to measurement noise as well as X 3 Z
ne X 1
avoid the ill-posedness of theP problem when the number of qp  γ k ϕzk P0i δ il dη; l  p  2
k  1:
degrees of freedom, ndof  N int 1
j1 nj , exceeds the number of
k1 i1
available measurements (2N rec ). N int is the number of inter- (12)
faces to be estimated, and e
j is the number of elements on
the jth interface. Here, δ is the standard Kronecker delta,

qnm
The energy functional in equation (6) depends nonli- γ k  1= ϕ2xk  ϕ2zk , and
nearly on the interfaces and is therefore difficult to minimize
directly. We apply an iterative deformation of an initial
ϕxk  P01
ηx01;k  P02
ηx02;k  P03
ηx03;k ;
guessed interface by discretizing the continuous gradient
descent flow obtained from the first variation of our energy ϕzk  P01
ηz01;k  P02
ηz02;k  P03
ηz03;k ; (13)
functional E. Let Γ
0
j be an initial guess for the unknown
(target) interface Γj , with nodal coordinates where P0 denotes a derivative of P with respect to η. The
index k labels the interface elements, and l accounts for
xi 
x
j
j
i ; z0;i : (7) corresponding node elements. The optimal corrections Δp to
reduce our energy functional from equation (6) are obtained
To minimize E, we restrict the deformation of the curves to by setting equation (10) to zero. This yields the optimal de-
small perturbations along the vertical direction only by formation Δz as a solution of the linear system of equations,
setting whose matrix form is given by Bignardi (2011):

z
j
j
j

i  z0;i  Δz;i Δ
j
j
z;i ≪ z0;i : (8) fJJT   αRgΔz  Jfuk0  udk g  αfqg; (14)

Further, we indicate with Δp the inversion parameters cor- where Δz is a column vector that lists the vertical perturba-
responding to the Δ
j
z;i perturbations. The Taylor expansion tions Δp , and the matrix entries in R and fqg are rpi and qp ,
of the displacements to the first order with respect to Δp is respectively.
Thus, starting from an initial guess Γ
0
j for the interfaces,
∂uk the energy represented in equation (6) is minimized iteratively
uk  u
0
k  Δ  O
Δ2p ; p  1; …; n; (9)
∂Δp p by evolving the interfaces according to equation (14), until the
error er  max
jΔp j is smaller than a chosen tolerance ϵ.
where O
Δ2p  denotes the neglected higher-order terms of Relaxation is introduced to limit the increments Δp at each
the series. Note that u
0
k is the BEM-solution for a geometry iteration by penalizing the squared norm of Δz . This adds
with interfaces Γ
0
j , and ∂uk
∂Δp is the pth row of the Jacobian an extra term βI to the left side of equation (14), where
matrix J, which measures the changes of the wave displace- I is the identity matrix and β is a fixed constant. The weight
806 S. Bignardi, F. Fedele, A. Yezzi, G. Rix, and G. Santarato

α of the regularizer is initially chosen large enough so that the version is under development to speed up the algorithm (see,
algorithm reconstructs the coarser scale features of the curve; for example, Fedele et al., 2003 and Eppstein et al., 2003 for
as the error er reduces, α is decreased to allow reconstruction an application of adjoint methods in optical tomography).
of the finer scale features. Further, implementation of the algorithm in a C environ-
ment will also reduce the computational burden.
Applications To investigate the performance of the algorithm under
noisy data, we added a Gaussian noise to the simulated
In the following we will test consistency and effective- measurements (both x and y components) with standard de-
ness of the geometric algorithm from equation (14) using viation σ 
P=100Qmax =3, with P  5% as the noise per-
synthetic data for the simple case of a one-layered geometry. centage, and Qmax as the maximum value of the data (phase
Further, as a field-case study we present the inversion or amplitude) to be perturbed, which usually corresponds to
of an experimental data set collected at a site in Alabama the value recorded at the receiver with minimum offset. Note
in 2004. that the noise level of Fourier amplitudes corresponds to a
larger error in the associated time-series displacements
Synthetic Data due to the nature of the Fourier transform. In particular,
the algorithm still provides a good estimate of the true inter-
Consider a 2D laterally varying layer on a half-space such face as shown in Figures 3 and 4, for the two previously men-
as Figure 1b and a typical instrumentation setup for surface-
tioned cases of low and high impedance jumps. In the same
wave testing at the geotechnical scale ∼O
10 m. In particu-
figures, in order to appreciate convergence we also report the
lar, the receivers are equally spaced every 4 meters, and the
normalized misfit E=E0 as a function of the number of itera-
source is located 3 meters away from the closest receiver.
tions, with E0 being the relative misfit of the initial guess.
We assume to know the source, and the two different elastic
The inversion results are more sensitive, and thus less accu-
characteristics of the two layers are considered. The first case
rate, when noise is added to phases only (curve P) than when
represents a high acoustic impedance jump scenario with
parameters V S  150
800 m=s, V P  500
2000 m=s, and noise is added to amplitudes only (curve A), as clearly seen
ρ  1600
2200 kg=m3 for the upper (lower) layer. The sec- from the plots of the misfit errors in Figures 3–6. The largest
ond case, on the other hand, represents a low impedance jump separation along z between the estimated and target inter-
configuration with parameters V S  150
250 m=s, V P  faces ranges from 1.2% to 8% with respect to the vertical
500
1000 m=s, and ρ  1600
2000 kg=m3 . Further, the extent of the dipping part of the layer, viz. Δh  15 m.
source intensity f0
ω  1000 N at frequency ω  1 rad=s. On the other hand, in field surveys only the vertical compo-
The laterally varying interface Γ should be inferred correctly nent of displacements is usually recorded. Thus, it is relevant
(within given numerical accuracy) from both zero-noise to check the performance of the geometric algorithm when
horizontal and vertical displacements (data) collected at the horizontal measurements are discarded in the inversion. Fig-
seven receivers. Corrupted data may lead to estimates that ure 5 shows the converged interfaces when noise is added
deviate from the true interface, and such deviations will also only to phases (curve P) or amplitudes (curve P). Clearly, the
be quantified. The unknown interface is discretized using 33
nodes. For the inversion, data at the receivers are simulated
using the forward model from equation (4) and consist of (a) 5
(b)
1

complex Fourier amplitudes of both vertical and horizontal 0

displacements. At first, no noise is added to data in order
Normalized misfit E/E0

It. 1500 10-1

-5
to test the convergence properties of the algorithm. The num- It. 1500 A
It. 1500 P
ber of model parameters (33) is slightly larger than that of the -10 Target
z [m]

available data (2 × 7). The arc-length regularizer in equa- 10-2

-15
tion (6) guarantees a course-to-fine scale reconstruction of
the unknown interface. The initial guessed interface Γ is -20
10-3
set as horizontal at z  25 m. Some of the intermediate iter- -25
ates are shown in Figure 1 to illustrate the convergence toward
the target interface. Observe that the iterated curve translates -30 10-4
-30 -25 -20 -15 -10 -5 0 5 15 25
rigidly at the first few iterations as a result of the initially large x [m] 10 2 Iterations
arc-length regularization. As the iterations increase, the reg-
ularizer is reduced, and the curve deforms accordingly to Figure 3. Sensitivity of the algorithm to noise level (5%): con-
adapt to the shape of the target interface in roughly 1500 verged interfaces for unperturbed data, noisy amplitudes (A) or
iterations. phases (P) only; (a) case of high acoustic impedance jump and
(b) associated normalized misfits E=E0 as functions of the number
The algorithm runs in a MATLAB environment on a of iterations. Elastic parameters: V S  150
800 m=s, V P 
MacBook with an Intel dual core at 2.4 Ghz, and the com- 500
2000 m=s, and ρ  1600
2200 kg=m3 for the upper (lower)
putational time is about 40 s=iteration. An adjoint-based in- layer.
Geometric Seismic-Wave Inversion by the Boundary Element Method 807

5 1 5 1
(a) (b) (a) (b)
0 0

Normalized misfit E/E0

Normalized misfit E/E0

-5 It. 1500 10-1 -5 10-1
It. 1500 A It. 1500 A
It. 1500 P It. 1500 P
-10 Target -10 Target

z [m]
z [m]

10-2 10-2
-15 -15

-20 -20
10-3 10-3
-25 -25

-30 10-4 -30 10-4

-30 -25 -20 -15 -10 -5 0 5 15 25 -30 -25 -20 -15 -10 -5 0 5 15 25
2
x [m] 10 Iterations x [m] 10 2 Iterations

Figure 4. (a) Geometric inversion using vertical displacements Figure 6. Boundary element method-based geometric inversion
only for unperturbed data, noisy amplitudes (A) and phases (P), re- for geometries with nonflat free surfaces: reconstructed interfaces
spectively; (b) associated normalized misfit E=E0 as a function of for corrupted amplitudes (A) and phases (P), respectively; (a) high
the number of iterations. Noise level 5%, high acoustic impedance acoustic impedance jump case and (b) associated normalized
jump case with elastic parameters: V S  150
800 m=s, misfits E=E0 as functions of the number of iterations. Elastic
V P  500
2000 m=s, and ρ  1600
2200 kg=m3 for the upper parameters: V S  150
800 m=s, V P  500
2000 m=s, and
(lower) layer. ρ  1600
2200 kg=m3 for the upper (lower) layer.

geometric algorithm identifies the major features of the values (u), within the accuracy of measurements. For exam-
target interface, even if half of the data are discarded. In such ple, Figure 8 reports Fourier amplitudes and phases of mea-
conditions, the arc-length parameter α is decreased at a surements at the receivers in comparison with the associated
slower rate to yield a smoother course-to-fine evolution of estimated values for the inversion of Figure 3. Clearly, the
the iterated curve. Moreover, a larger value of the Tikhonov BEM geometric algorithm is effective in reconstructing the
parameter β is required to limit the correction of the curve at interface Γ from noisy data.
each iteration. As a result, the number of iterations has
increased in comparison to the cases in which the entire set Field-Case Study
of data is exploited. The geometric algorithm can also handle We point out that the geometric algorithm is effective in
irregular free surfaces in a natural manner, as clearly illu- estimating lateral interfaces, in particular, flat interfaces as a
strated in Figures 6 and 7, where we report the reconstructed special case. To prove this, we considered an experimental
curves for data with noise level P  10%. data set collected at a site in Alabama in 2004. The experi-
Observe that in all the inversions, we saw a remarkable mental campaign was carried out using an electromechanical
agreement between data (ud ) and the associated modeled shaker equipped with an additional reactive mass that acts as

5 1
(a) (b) (a) 5
(b) 1
0 0
Normalized misfit E/E0

10-1
Normalized misfit E/E0

-5 It. 1500 10-1

It. 1500 A -5
It. 1500 A
It. 1500 P It. 1500 P
-10 Target -10 Target
z [m]

z [m]

10-2 10-2
-15 -15

-20 -20
10-3 10-3
-25 -25

-30 10-4 -30 10-4

-30 -25 -20 -15 -10 -5 0 5 15 25 -30 -25 -20 -15 -10 -5 0 5 15 25
2
x [m] 10 Iterations x [m] 10 2 Iterations

Figure 5. Sensitivity of the algorithm to noise level (5%): con- Figure 7. Boundary element method-based geometric inversion
verged interfaces for unperturbed data, noisy amplitudes (A) and for geometries with nonflat free surfaces: reconstructed interfaces
phases (P), respectively; (a) case of low acoustic impedance jump for corrupted amplitudes (A) and phases (P) respectively; (a) low
and (b) corresponding normalized misfits E=E0 as functions of the acoustic impedance jump case and (b) corresponding normalized
number of iterations. Elastic parameters: V S  150
250 m=s, misfits E=E0 as functions of the number of iterations. Elastic
V P  500
1000 m=s, and ρ  1600
2000 kg=m3 for the upper parameters: V S  150
250 m=s, V P  500
1000 m=s, and ρ 
(lower) layer. 1600
2000 kg=m3 for the upper (lower) layer.
808 S. Bignardi, F. Fedele, A. Yezzi, G. Rix, and G. Santarato

π
(a) (b) Source Receiver
1
2/3π ux 0
0.8
Normalized Amplitude

π/3
uy

Phase (rad)
0.6 -10 Iter. 1
0
0.4
Iter. 3
-20
−π/3 uy
0.2
Estimated interface, Iter. 27

z [m]
−2/3π -30
0
ux Iter. 2
−π
1 2 3 4 5 6 7 1 2 3 4 5 6 7 -40
Sensor number

Figure 8. (a) Modeled normalized Fourier amplitudes of hori- -50

zontal (vertical) displacements ux (uy ) against data at the receivers
for the case of noisy data with corrupted amplitudes only (see curve Initial guess
A of Fig. 3 for the associated reconstructed interface); (b) modeled -60
phases against data at the receivers for noisy data with corrupted
phases only (see curve P of Fig. 3 for the associated reconstructed
interface). Noise level P  5% and error bars denote stability bands -80
0 5 10 15 20 25 30 35
of measurements.
x [m]

a harmonic source at frequencies ranging from ∼3 to 100 Hz; Figure 10. Geometric inversion of experimental data collected
therefore, these data are particularly indicated for one- at the Alabama site. Elastic parameters: V S  355
1370 m=s,
V P  664
2000 m=s, and ρ  1500
2400 kg=m3 for the upper
frequency inversion. Vertical particle accelerations were (lower) layer.
measured by a linear array of 15 low-frequency acceler-
ometers located at distances ranging from 2.4 to 32 meters
from the source. Figure 9 shows the typical shear wave such geometry with that of a soft upper layer with thickness
velocity profile estimated from conventional surface-wave ∼30 m laying on top of an half-space. An estimate of the
shear velocity V S30 of the top layer is obtained from the pro-
testing on the field setup of Figure 10. We can approximate
file of Figure 9 as

Xnl Xnl
0 di
V S30  di ; (15)
i1
V
i
i1 S
-10

where di and V S
i are thickness and shear wave velocity of

-20 the ith layer, respectively, and similarly for that of the half-
space. Further, remaining elastic parameters are inferred
-30 from some geological knowledge of the region. As a result,
Depth [m]

the elastic properties of the top layer (half-space) are set

-40 as V S  335
1369 m=s, V P  664
2001 m=s, and ρ 
1500
2400 kg=m3 , respectively. To proceed with the inver-
-50
sion at the excitation frequency ω  10 Hz, we first com-
puted the Fourier spectrum S
w of the measured vertical
displacements, which is very narrow. Integrating the spec-
-60
trum around ω over a window Δw=ω ∼ 0:01 (to filter out
noise at lower and higher w’s) provides the complex Fourier
-70 amplitudes at the receivers to be used as data in the BEM
inversion. Here, the shaker force is modeled at the source
-80 node as ts  mω2 ds  Fg z^ , where Fg is the maximum force
0 500 1000 1500
Shear wave velocity Vs [m/s] magnitude generated by the shaker, and m is the sum of re-
active and device masses. The free surface Γs is discretized
Figure 9. Local shear wave velocity profile estimated using using 67 nodes, and the unknown interface Γ1 is discretized
standard SWMs at the Alabama site. with 41 nodes. The inverse problem is ill-posed because we
Geometric Seismic-Wave Inversion by the Boundary Element Method 809

have 15 receivers and 41 unknowns, so the arc-length reg- horizontal scales with steep interfaces, irregular free sur-
ularizer is needed. As shown in Figure 10, the initial guess faces, and subsurfaces with low acoustic impedance ratios.
of the unknown interface Γ1 is set as horizontal at z  The method is easily generalizable to 3D geometries,
60 m. The algorithm converged in 27 iterations, and some and the estimated geometry provides a good initial guess
intermediates are shown in the same figure. Observe that in for FEM-based FWIs in order to reconstruct higher-order
few iterations the iterate rigidly translates to nearby structured heterogeneities. Thus, BEM-based inverse models
z ∼ 24 m, then it shapes in to a very mild laterally varying have the potential to improve the effectiveness and accuracy
interface, in good agreement with the local geology, which of FEM-based inversions. Further, the BEM forward model
was previously known to be almost 1D. Shaping requires can be accelerated using fast multipole techniques (Gumerov
more iterations than translation as in the inversion of syn- and Duraiswami, 2004; Liu, 2009), in combination with the
thetic data (see Fig. 1). use of adjoint methods to speed up the computation of the
The single-layer inversion provides an estimate of the Jacobian, thus drastically reducing the computational burden
location of the interface between the upper soft layer and of the algorithm.
the hard bottom at approximately z ∼ 24 m. The model fit We believe that the application of BEM-based inverse
the data, as illustrated in Figure 11, where we report the mod- models to geological problems from small to large scales will
eled Fourier amplitudes and measurements at the receivers. enable researchers and other stakeholders to investigate the
Better agreement with data (especially for amplitudes) could underground geological complexities with a low cost-to-
be attained by performing the inversion with a multilayer benefits ratio method. Our preliminary results are very pro-
geometry to have a better resolution of the softer, shallow mising, with successbeing achieved in nicely reproducing an
layers of Figure 9. This is currently under investigation. irregular interface in a 2D layered model. As far as the
approach is in principle scale-invariant and has no space lim-
itations, work is in progress to extend it for the inversion of
Conclusions multifrequency data and 3D multilayered geometries with in-
We have introduced a geometric inversion formulation clusions or cavities. This will empower complete reconstruc-
that overcomes the stiffness of classic surface-wave methods tions of the subsurface in complex geological and tectonic
in dealing with nonparallel subsurface layers. The inverse structures at any scale.
model exploits the BEM to infer the shape of 2D laterally
varying interfaces directly from measurements at the recei-
vers. No indirect inference is observed from the dispersion
Data and Resources
curve. Indeed, the proposed inverse algorithm exploits the Available data were collected by G. J. Rix, and they can-
ability of the BEM to model wave propagation through a not be shared publically.
medium whose soil properties may be approximated as re-
gionally homogeneous. Tests on both synthetic and experi-
mental data on 2D geometries provide evidence that the References
BEM-based geometric inversion is effective even in the
Aki, K. (1988). Local site effects on strong ground motion, in Geotechnical
presence of noise and is applicable for cases over large
Special Publication, N. 20, Earthquake Engineering and Soil
Dynamics II: Recent Advances in Ground Motion Evaluation,
J. L. Von Thun (Editor), American Society Civil Engineers,
-1.2 Park City, Utah.
(a) 10-7 (b) Aki, K., and P. G. Richards (2002). Quantitative Seismology, Theory and
-1.4
Methods, Second Ed., University Science Books, Sausalito, California,
-1.6 249–322.
10-8
Assimaki, D., and E. Kausel (2007). Modified topographic amplification
-1.8
factors for a single-faced slope due to kinematic soil-structure interac-
Phase [rad]
|| u || [m]

-2 tion, J. Geotech. Geoenviron. Eng. 133, no. 11, 1414–1431.

Bard, P. Y. (1982). Diffracted waves and displacement field over two-
−2.2 dimensional elevated topographies, Geophys. J. R. Astr. Soc. 71,
10-9
−2.4
731–760.
Bard, P. Y., and J. C. Gabriel (1986). The seismic response of two-
-2.6 dimensional sedimentary deposits with large vertical velocity gradi-
ents, Bull. Seismol. Soc. Am. 76, 343–366.
0 −2.8
5 10 15 5 10 15 Beskos, D. E., B. Dasgupta, and I. G. Vardoulakis (1986). Vibration isolation
# Receiver using open or filled trenches, Comput. Mech. 1, no. 1, 43–63.
Beskos, D. E., K. L. Leung, and I. G. Vardoulakis (1986). Vibration isolation
Figure 11. Field-case study: comparisons between data and of structures from surface waves in layered soil, Rec. Appl. Comput.
model estimates at the receivers for the inversion of Figure 10. Mech., 125–140.
(a) Estimated normalized Fourier amplitudes and (b) associated Bignardi, S. (2011). Complete Waveform Inversion Approach to Seismic
phases against data (error bars denote stability bands of measure- Surface Waves and Adjoint Active Surfaces, Ph.D. Thesis, University
ments). of Ferrara, Ferrara, Italy.
810 S. Bignardi, F. Fedele, A. Yezzi, G. Rix, and G. Santarato

Brebbia, C. A., and J. Dominguez (1992). Boundary Elements, An Introduc- Manolis, G. D., and D. E. Beskos (1981). Dynamic stress concentration
tory Course, WIT Press/Computational Mechanics Publications, studies by boundary integrals and Laplace transform, Int. J. Numer.
Southampton, UK, 314 pp. Meth. Eng. 17, no. 4, 573–599.
Christensen, R. M. (1971). Theory of Viscoelasticity, Academic Press, Manolis, G. D., and D. E. Beskos (1983). Dynamic response of lined tunnels
New York, 364 pp. by an isoparametric boundary element method, Comput. Meth. Appl.
Dineva, P., and G. Manolis (2001a). Scattering of seismic waves by cracks in Mech. Eng. 36, no. 3, 291–307.
multi-layered geological regions: I. Mechanical model, Soil Dynam. Manolis, G. D., and D. E. Beskos (1988). Boundary Element Methods in
Earthq. Eng. 21, 615–625. Elastodynamics, Unwin Hyman Ltd., London, UK, 288 pp.
Dineva, P., and G. Manolis (2001b). Scattering of seismic waves by cracks in Maupin, V. (2007). Introduction to mode coupling methods for surface
multi-layered geological regions: II. Numerical results, Soil Dynam. waves, Adv. Geophys. 48, 127–155.
Earthq. Eng. 21, 627–641. Maupin, V., and B. L. N. Kennett (1987). On the use of truncated
Dominguez, J. (1993). Boundary Elements in Dynamics, Computational modal expansions in laterally varying media, Geophys. J. Int. 91,
Mechanics Publications, Elsevier, Southampton, UK, 165–207. 837–851.
Dominguez, J., and R. Abascal (1984). On fundamental solutions for the Nazarian, S., and K. H. Stokoe (1984). In situ shear wave velocity
boundary integral equations method in static and dynamic elasticity, from spectral analysis of surface waves, Proc. 8th Conf. on Earthquake
Eng. Anal. 1, no. 3, 128–134. Engineering, Vol. 3, Prentice Hall, San Francisco, California,
Dravinski, M., and T. K. Mossessian (1987). Scattering of plane harmonic 31–38.
P, SV, and Rayleigh waves by dipping layers of arbitrary shape, Bull. Park, C. B., R. D. Miller, and J. Xia (1999). Multichannel analysis of surface
Seismol. Soc. Am. 77, no. 1, 212–235. waves, Geophysics 64, 800–808.
Eppstein, M., F. Fedele, J. P. Laible, C. Zhang, A. Godavarty, and Park, C. B., R. D. Miller, N. Ryden, J. Xia, and J. Ivanov (2005). Combined
E. M. Sevick-Muraca (2003). A comparison of exact and approximate use of active and passive surface waves, J. Environ. Eng. Geophys. 10,
adjoint sensitivities in fluorescence tomography, IEEE Trans. Med. 323–334.
Imag. 22, no. 10, 1215–1222. Raptakis, D., F. J. Chavez-Garcia, K. Makra, and K. Pitilakis (2000). Site
Fedele, F., M. Eppstein, J. P. Laible, A. Godavarty, and E. M. Sevick-Muraca effects at Euroseistest-I. Determination, of the valley structure and
(2005). Fluorescence photon migration by the Boundary Element confrontation of observations with 1D analysis, Soil Dynam. Earthq.
Method, J. Comput. Phys. 210, no. 1, 109–132. Eng. 19, no. 1, 1–22.
Fedele, F., J. P. Laible, and M. Eppstein (2003). Coupled complex Rix, G. J., G. L. Hebeler, and M. C. Orozco (2002). Near-surface Vs profil-
adjoint sensitivities for frequency-domain fluorescence tomography: ing in the New Madrid seismic zone using surface wave methods,
Theory and vectorized implementation, J. Comput. Phys. 187, Seismol. Res. Lett. 73, no. 3, 380–392.
no. 2, 597–619. Rix, G. J., C. G. Lai, and A. W. Spang (2000). In situ measurement of
Fichtner, A. (2011). Full Seismic Waveform Modelling and Inversion damping ratio using surface waves, J. Geotech. Geoenviron. Eng.
Springer-Verlag, Heidelberg, Germany, 356 pp. 126, no. 5, 472–480.
Gabriels, P., R. Snieder, and G. Nolet (1987). In situ measurements of Rix, G. J., C. G. Lai, and S. Foti (2001). Simultaneous measurement of
shear-wave velocity in sediments with higher-mode Rayleigh waves, surface wave dispersion and attenuation curves, Geotech. Test. J.
Geophys. Prospect. 35, 187–196. 24, no. 4, 350–358.
Gjevik, B. (1973). A variational method for Love waves in nonhorizontally Savage, W. Z. (2004). An exact solution for effects of topography
layered structures, Bull. Seismol. Soc. Am. 63, no. 3, 1013–1023. on free Rayleigh waves, Bull. Seismol. Soc. Am. 94, no. 5,
Gumerov, N. A., and R. Duraiswami (2004). Fast Multipole Methods for the 1706–1727.
Helmholtz Equation in Three Dimensions, Elsevier Ltd., Oxford, UK, Stokoe, K. H., S. G. Wright, J. A. Bay, and J. M. Roesset (1994). Charac-
520 pp. terization of geotechnical sites by SASW method, Proc. XIII Int. Conf.
Katsikadelis, J. T. (2002). Boundary Elements: Theory and Applications on Soil Mechanics and Foundation Engineering, Oxford & IBH
Elsevier Science, Oxford, UK, 201–234. Publishing, New Delhi, India, 15–25.
Kausel, E., and J. M. Roesset (1981). Stiffness matrices for layered soils, Strobbia, C., and S. Foti (2006). Multi-offset phase analysis of surface wave
Bull. Seismol. Soc. Am. 71, 1743–1761. data (MOPA), J. Appl. Geophys. 59, 300–313.
Kawase, H. (1988). Time-domain response of a semi-circular canyon for Tarantola, A. (1987). Inverse Problem Theory: Methods for Data Fitting and
incident SV, P and Rayleigh waves calculated by the discrete wave- Model Parameter Estimation, Elsevier, New York, 356 pp.
number boundary element method, Bull. Seismol. Soc. Am. 78, Tikhonov, A., and V. Arsenin (1977). Solution of Ill-Posed Problems, Wiley,
1415–1437. New York.
Kawase, H., and K. Aki (1989). A study of the response of a soft basin for Tokimatsu, K. (1995). Geotechnical site characterization using
incident S, P, and Rayleigh waves with special reference to the long surface waves, Proc. 1st Int. Conf. on Earthquake Geotechnical
duration observed in Mexico City, Bull. Seismol. Soc. Am. 79, 1361– Engineering (IS-Tokyo ’95), Balkema, Rotterdam, The Netherlands,
1382. 1333–1368.
Kitahara, M. (1984). Applications of boundary integral equation methods to Tselentis, G. A., and G. Delis (1998). Rapid assessment of S-wave profiles
eigenvalue problems of elastodynamic and thin plates, Research from the inversion of multichannel surface wave dispersion data, Ann.
Report, University of Kyoto, Kyoto, Japan. Geofisc. 41, 1–15.
Lai, C. G., G. J. Rix, S. Foti, and V. Roma (2002). Simultaneous measure- Virieux, J., and S. Operto (2009). An overview of full-waveform
ment and inversion of surface wave dispersion and attenuation curves, inversion in exploration geophysics, Geophysics 74, no. 6,
Soil Dynam. Earthq. Eng. 22, no. 9–12, 923–930. 127–152.
Liu, Y. (2009). Fast Multipole Boundary Element Method: Theory and Xu, S. Z. (2001). The Boundary Element Method in Geophysics, in Geophy-
Applications in Engineering, Cambridge University Press 8, New sical Monograph Series, Society of Exploration Geophysicists, Tulsa,
York, 5–118. Oklahoma, 169–203.
Louie, J. N. (2001). Faster, better: Shear-wave velocity to 100 meters depth Yezzi, A., A. Tsai, and A. Willsky (2002). A fully global approach to image
from refraction microtremor arrays, Bull. Seismol. Soc. Am. 91, no. 2, segmentation via coupled curve evolution equations, J. Vis. Comm. Im-
347–364. age Represent. 13, 195–216.
Luo, Y. H., J. H. Xia, J. P. Liu, Y. X. Xu, and Q. S. Liu (2008). Generation of Yezzi, A., L. Zollei, and T. Kapur (2003). A variational framework for in-
a pseudo-2D shear-wave velocity section by inversion of a series of 1D tegrating segmentation and registration through active contours, Med.
dispersion curves, J. Appl. Geophys. 64, 115–124. Image Anal. 7, 171–185.
Geometric Seismic-Wave Inversion by the Boundary Element Method 811

v
u 3
Appendix u X dPi 2 X 3
dPi
2
Jt xi  zi ; (A5)
i1
dη i1
dη
BEM Matrices
with
xi ; zi  as the nodal coordinates.

8 9
>
> T 111 T 112 T 113  T 121 T 122  >
>
>
> >
>
>
> T 2 T 212 T 213  T 221 T 222   >
>
>
> T 11 >
>
< 3
T 3 T 3  T 3 T 3 
11 12 13 21 22
=
H  T 11 T 12 T 13  T 21 T 22 
4 4 4 4 4 ; (A1)
>
> >
>
>
> . . >
>
>
> : .. : .. >
>
>
> >
>
: T
nj nj n n
 T ne 1 T nje 2 T nje 3 ;

ne 13

and
8 9
>
> U 111 U 112 U 113  U 121 U 122  >
>
>
> >
>
>
> U 211 U 212 U 213  U 221 U 222   >
>
>
> >
>
< U 3 U 3 U 3  U 3 U 3 
11 12 13 21 22
=
G  U 11 U 12 U 13  U 21 U 22 
4 4 4 4 4 : (A2)
>
> >
>
>
> . .. >
>
>
> : .. : . >
>
>
> >
>
: n n n n
U
nj e 13  U nje 1 U nje 2 U nje 3 ;

Further, Department of Earth Science

University of Ferrara
Z 1 Università degli Studi di Ferrara
U skn ji  Uji xk
η; xs ; ωPn
ηJ
ηdη; Dipartimento di Scienze della Terra
1 Via Saragat 1 (Edicio B), 44122 Ferrara, Italy
Z 1
(A3) (S.B., G.S.)
T skn ji  T ji xk
η; xs ; ωPn
ηJ
ηdη
1

Here, Uji
xk
η; xs ; ω is the 2 × 2 fundamental tensor School of Electrical and Computer Engineering
whose entries represent the i component of displacements Georgia Institute of Technology
at xk
η on the element k due to a Dirac source at xs directed 225 North Avenue NW
along j where i, j  1 or 2 (x or z directions). Further, U skn is Atlanta, Georgia 30332
(F.F., A.Y.)
the fundamental tensor Uji projected onto the element
k∈1; ne  at the node n∈f1; 2; 3g with a Dirac source at
the node s∈1; nj . Finally, the Lagrangian interpolation
functions and Jacobian are defined, respectively, as School of Civil and Environmental Engineering
Georgia Institute of Technology
P1
η  1=2η  1=2η2 ; 225 North Avenue NW
Atlanta, Georgia 30332
P2
η  1  η2 ; (G.R.)

P3
η  1=2η  1=2η2 ; (A4)

and Manuscript received 29 March 2011