Soil Dynamics and Earthquake Engineering 161 (2022) 107423

Contents lists available at ScienceDirect

Soil Dynamics and Earthquake Engineering

journal homepage: www.elsevier.com/locate/soildyn

An efficient full-wavefield computational approach for seismic testing in a

layered half-space
Chih-Ping Lin a, *, Tsai-Jung Wu a, Ernian Pan b
a
Department of Civil Engineering, National Yang Ming Chiao Tung University, 1001 University Rd., Hsinchu, 300, Taiwan
b
Disaster Prevention & Water Environment Research Center, National Yang Ming Chiao Tung University, 1001 University Rd., Hsinchu, 300, Taiwan

A R T I C L E I N F O A B S T R A C T

Keywords: Inversion of subsurface shear-wave velocity profile based on effective dispersion curve or full wavefield is still
Fast computation rare due to the lack of efficient forward computation kernel for the full wavefield in a layered medium. To
Elastic layered medium facilitate exploiting the complete signal content in seismic testing, an efficient new computational framework for
Dynamic response
modeling synthetic full wavefields in horizontally layered systems is proposed. Dynamic response of an isotropic
Seismic testing
Surface wave method
elastic layered system subjected to vertical loading is implemented and thoroughly validated. The new procedure
is based the Fourier-Bessel series expansion (not the traditional Fourier-Bessel transform), and it is at least two-
orders faster than the current full wavefield computation method based on integral transformation. It considers
survey parameters and models all wave phenomena, including near-field effect and leaky waves, which cannot be
captured by the modal summation method. Examples of full wavefield simulation are presented to demonstrate
different levels of information provided by the experimentally available fundamental mode, broadband effective
dispersion curve, and the full velocity spectrum. Different earth models may have similar fundamental modes or
even effective dispersion curves; however, their difference manifests in different higher-mode contributions and
can only be exploited by utilizing the entire velocity spectra including the vertical and radial components. The
coupling between the vertical and radial components in terms of their spectral ratio and phase difference can
provide additional constraints (data) in seismic testing.

1. Introduction dimensional (2D) or three-dimensional (3D) velocity models, espe­

cially in the oil and gas industry, without relying on the specific seismic
Seismic velocity, S-wave velocity (vs) in particular, is highly correlated characteristics (e.g., travel-time curve of the first arrivals or the
with the engineering parameters of the medium, such as liquefaction po­ dispersion curve). Numerous studies (e.g. Refs. [4–6], demonstrated that
tential [1], pre-consolidation stress [2], and shear strength [3]. S-wave FWI is promising in quantitatively imaging near-surface structures with
velocity is an important soil dynamic property, which also plays an high resolution. FWI requires high computational time and needs to deal
important role for predicting the ground deformation subject to loading. with the unknown source wavelet. While waveform fitting could provide
As a result, accurate vs-profile with high-spatial resolution is much desired much stronger constraints on velocity and attenuation models, it is
for geotechnical site characterization. Surface seismic methods, such as highly non-linear and very sensitive to the initial model. The refraction
seismic refraction, seismic reflection, and seismic surface wave methods, or surface wave method can be used to provide an appropriate initial
are attractive because they are non-destructive, cost effective, and model where the non-linearity from local minima can be reduced by
distributive. In particular, the surface wave method has become an using a smart workflow or by alternative measure of the misfit. Related
important tool for geotechnical site investigation due to its higher approaches include multiscale FWI [7], envelop-based FWI [8],
signal-to-noise ratio along with automation of data processing. The amplitude-spectrum-based FWI [9–11], and inversion of skeletonized
equipment and field-testing procedure used are very similar among these data [12–14]. Still, the complexity is formidable when variations of
methods, especially between seismic refraction and surface wave methods. S-wave velocity, P-wave velocity, quality factor, or even their anisot­
Due to the rapid increase in computational power, there are growing ropies are considered simultaneously in 2D/3D FWI. To the other
developments looking into the full-waveform inversion (FWI) for two- extreme contrary to the 2D or 3D FWI approach, the conventional

* Corresponding author. Department of Civil Engineering, National Yang Ming Chiao Tung University, 1001 Ta-Hsueh Rd., Hsinchu, 300, Taiwan.
E-mail address: [email protected] (C.-P. Lin).

https://doi.org/10.1016/j.soildyn.2022.107423
Received 19 April 2022; Received in revised form 20 June 2022; Accepted 2 July 2022
Available online 15 July 2022
0267-7261/© 2022 Elsevier Ltd. All rights reserved.
C.-P. Lin et al. Soil Dynamics and Earthquake Engineering 161 (2022) 107423

surface wave method in which only the S-wave velocity is estimated, not interpreted as the fundamental mode dispersion curve; however, compli­
only assumes 1D layered structure but also considers only the normal cations due to higher modes are later recognized [35–37]. The “effective”
modes in the plane-wave condition. Thus, there is a gap between the (or apparent) dispersion curves of SASW can be modeled by
complex 2D/3D FWI and the oversimplified 1D 1-variable surface wave multiple-mode stochastic waves [37] or more rigorously by theoretical
method. Under the assumption of horizontally layered media, the modeling of the SASW procedure [38,39]. The MASW method, on the
inversion non-linearity in the FWI could be greatly reduced. This is other hand, takes the advantages of the multi-station data and wavefield
particularly true when data misfit is in the transformed domain (i.e., the transformation techniques to identify different propagation modes. While
wavefield spectrum or spectrum of the Green’s function in the it is very common to pick the fundamental-mode dispersion curve for
frequency-wavenumber, frequency-slowness, or frequency-velocity do­ inversion, including higher modes for inversion has been also exploited
mains). Furthermore, such horizontally layered models could facilitate [40–43]. However, picking the dispersion curve accurately is hindered by
the construction of better initial models for more complex 2D or 3D survey/testing condition (i.e., near field and truncation effects) as well as
inversion when necessary for higher lateral resolution [15,16]. the higher-mode effect. The latter effects include mode kissing [44–46],
It should be emphasized again that, the simplified horizontally-layered mode jumping [46–48], and leaky wave [48,49]. In light of the potential
system is quite reasonable in many engineering applications, such as pitfalls in inversion based on the normal modes, inversion based on the
pavement testing, characterization of soil deposits, and non-destructive “effective” dispersion curve was also proposed for MASW (i.e., [50]. The
testing of laminar structures. While various analytical Green’s functions theoretical “effective” dispersion curve is modeled using the forward
(in the transformed domain) in layered systems are available [17], nu­ calculation with full-waveform P-SV reflectivity synthetic shot gathers for
merical integration is generally needed in order to compute the reproducing all wavefields. The inversion based on the “effective”
frequency-domain response or time-domain waveforms from the Green’s dispersion curve, either in SASW or MASW, can be regarded as a semi-full
function. To avoid the computationally intensive integration involved wavefield inversion or skeletonized inversion [43]. In the MASW disper­
[15], proposed an empirical transformation method to estimate the sion analysis, the entire dispersion spectrum can be used as input data for
wavefield spectrum (i.e., Fourier-Bessel expansion coefficients) from the inversion [10,15,16,21,29].
recorded multi-station wavefields using a discrete approximation to the In the inversion in terms of either skeletonized data (e.g., effective
Bessel transformation. This experimental wavefield spectrum is then fitted dispersion curve) or the full dispersion spectra, the forward computation
by the theoretical one calculated by the reflectivity method [16]. How­ kernel of the full wavefield in a layered system is required. Unfortu­
ever, such a wavefield transformation to approximate the Green’s function nately, current methods which can accurately generate synthetic full
in the transformed domain imposes requirements on the receiver array. As wavefield are time consuming. This is particularly true when the actual
a matter of fact, the actual wavefield spectrum can never be precisely loading condition (i.e., not point source) and source-to-receiver distance
obtained due to the spatial truncation of the limited size of the receiver are considered. Thus, in this study, we introduce an efficient computa­
array. Alternatively, the common plane-wave transformation methods, tional framework for modeling the synthetic full wavefields in layered
such as frequency-wavenumber (f-k) transform [18], slowness-frequency systems. An isotropic elastic layered system which is subjected to a
(p-f) transform [19], or frequency-velocity transform by the phase-shift vertical load is implemented to demonstrate its potential applications in
method [20], can be simply used to transform the seismic data to the common seismic testing.
experimental dispersion spectrum. In a more rigorous but computationally
intensive manner, the same field-setup geometry and wavefield trans­ 2. Methods
formation could be used in a theoretical layered model to fit the experi­
mental dispersion spectrum. For example, Ryden & Park [21] used the To illustrate the framework for computing the full wavefield in the
stiffness matrix approach [22] to perform inversion of the frequency domain, an isotropic elastic layered system subjected to a ver­
frequency-velocity spectrum of the pavement. The thin-layer method [23], tical loading is assumed. Fig. 1 shows the layered half space in a cylindrical
a discrete formulation of the stiffness matrix approach by subdivision of coordinate system. A uniform (axial symmetric) time-harmonic vertical
layers with linear interpolation between the layer interfaces, was recently load of frequency ω is applied within radius a centered at (z, r) = (0, 0).
used for full-waveform inversion of frequency response to estimate the This is to emulate a common seismic survey with a vertical impact source
material properties of layered half-spaces [6]. Notice that the thin-layer on a circular testing plate. It should be pointed out that while a uniform
method is similar to the discrete-layer method (i.e. [24], finite-layer load over a circle can be easily reduced to the concentrated one (by
method (i.e. [25], or precise integral method (i.e., [26]. While its final decreasing the area of the circle and increasing the load density), the other
matrix format is a linear system of eigenequations, a fine discretization way is not simple. To find the response due to a uniform load from the
with a large system of equations may be required in order to achieve an given concentrated force solution, one has to carry out the integral of the
accurate solution [24]. Dou & Ajo-Franklin [10] used the reflectivity
method and frequency-velocity spectrum to perform full-wavefield
inversion for velocity models in permafrost. To reduce the computation
time, the modal summation method [27,28] based on normal modes was
adopted by Dal Moro [29] and Song et al. [30] to compute the steady-state
surface waves in layered media. While the modal summation fails to
capture the transient response in the near field, other more complete so­
lutions (such as the reflectivity method and stiffness matrix method)
require intensive computation time. As such, a forward modeling kernel,
which is complete and efficient in computing the theoretical dispersion
spectrum, is very desirable in the 1D full-wavefield inversion.
The above-mentioned studies are mostly motivated by the difficulties
in modal identification for conventional dispersion curve inversion due to
interference of higher modes and leaky modes in complex field conditions.
There are currently two main tracks in shallow surface wave testing and
analysis. Namely, the spectral analysis of surface waves (SASW [31,32],
and the multi-channel (multi-station) analysis of surface waves (MASW,
[33,34]. The SASW method measures the frequency-dependent phase Fig. 1. The coordinates and variables in an l-layered half space where a time-
velocity from the phase differences between two receivers. Originally, it is harmonic load over a surface circle is applied.

2
C.-P. Lin et al. Soil Dynamics and Earthquake Engineering 161 (2022) 107423

Green’s function over the circular area. While the Green’s functions of the
concentrated force are singular (near the source point), those of the uni­
∂ ψ̂
u = − kφ
̂ ̂− (5b)
form load are regular and behavior well (everywhere). As such, the re­ ∂z
sponses from these two sources are different mathematically and in which the hat ˆ denotes the displacement and potential variables in
fundamentally. In this study, we apply a uniform load over a surface circle the transformed domain (ω, k).
to mimic a hammer impact on a circular plate which can be very conve­ The wave equations (eq. (1) in terms of the Helmholtz potentials) in
nient in surface seismic testing/analysis. the transformed domain become

2.1. Governing equations and general solutions ∂2 φ

̂
+ kp2 φ
̂ =0 (6a)
∂z2
The governing equation for the elastic wave propagation is summa­
rized in Appendix A in terms of the displacement field (u), which is further ∂2 ψ̂
+ ks2 ψ
̂ =0 (6b)
decomposed as the sum of the gradient of a scalar potential (φ) and the ∂z2
curl of a vector potential (ψ ) [51]. For a vertical load applied uniformly
over a circular area on the surface, the problem becomes axially sym­ where
metric. As a result, the vector potential ψ has only a nonzero component in ω2
θ-direction, ψ θ (denoted as ψ for convenience). Thus, the Navier equation kp2 = − k2 (7a)
v2p
(wave equation) in terms of the potentials can be written as

∂2 φ 1 ∂φ ∂2 φ 1 ∂2 φ ω2
+ + = (1a) ks2 = − k2 (7b)
∂r2 r ∂r ∂z2 v2p ∂t2 v2s

Finally, applying the Hankel transform properties (eqs (B3) and B6 in

∂2 ψ 1 ∂ψ ∂2 ψ ψ 1 ∂2 ψ Appendix B), the stress-displacement relations (eq. (4)) in the trans­
+ + − = (1b)
∂r2 r ∂r ∂z2 r2 v2s ∂t2 formed domain become

where vp is the P-wave velocity and vs the S-wave velocity, defined as ∂w

̂
σ zz = (λ + 2μ)
̂ + λk̂
u (8a)
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ ∂z
vp = (λ + 2μ)/ρ (2a)
∂̂u
√̅̅̅̅̅̅̅̅ ̂τ zr = μ − μk w
̂ (8b)
vs = μ/ρ (2b) ∂z
Notice that a hat is used to represent the quantities in the trans­
in which λ and μ are the Lame constants and ρ is the density of the formed domain as functions of (ω, k, z) of the original field quantities in
medium. For the present axis-symmetric case, the displacement the physical domain which are functions of (t, r, z).
component in θ-direction is zero, and the displacements in the r- and z- The general solutions to the wave equation in terms of the Helmholtz
directions (denoted as u and w, respectively) in terms of the potentials potentials (eq. (6)) can be easily found as
become
̂ = A+ e−
φ ikp z
+ A− eikp z (9a)
∂φ 1 ∂(rψ )
w= + (3a)
∂z r ∂r ψ̂ = B+ e− iks z
+ B− eiks z (9b)

where A+, A-, B+, and B− are unknown constants to be determined by the
∂φ ∂ψ
u= − (3b)
∂r ∂z boundary conditions. The plus and minus symbols are used to represent
The z-direction stress-displacement relations under the axis- the forward (downward) propagating and backward (upward) propa­
symmetric condition are simplified to gating wave mode, respectively.
∂w λ ∂(ru)
σ zz = (λ + 2μ) + (4a)
∂z r ∂r 2.2. Green’s functions in the layered system
( )
∂u ∂w
τzr = μ + (4b) There are different approaches to calculate the Green’s functions or
∂z ∂r
the transfer functions in a layered system. These include the transfer
The above equations are in the physical domain (t-r-z domain). We matrix method [52,53], the reflectivity method, also called the reflec­
now transform the time to frequency (ω) by the Fourier transform and tion and transmission (RT) method [54–57], the global stiffness matrix
the radius r to wavenumber k by the Hankel transform. After applying method [22], and the hybrid compliance-stiffness matrix method [58] or
the Fourier transform to the time variable (in terms of eiωt), we apply the dual variable and position (DVP) method [59]. By taking the advantages
Hankel transform to the spatial variable r. Hankel transform of order 0 is of its numerical stability and simplicity, the global stiffness matrix
applied to eqs (1a), (3a) and (4a), while Hankel transform of order 1 is method is adopted in this paper.
applied to eqs (1b), (3b) and (4b), so that u(r = 0) = 0, ψ (r = 0) = 0, For any layer (e.g., the jth layer) above the lower homogenous half
and τzr (r = 0) = 0 (i.e., zero radial displacement, vector potential, and space with finite thickness h, the general solution can be modified from
shear stress at the origin r = 0 under axial symmetric vertical load). eq. (9) as (for convenient use later on),
Referring to Appendix B, noting the derivative property (eq. B3) and
property of the Bessel differential operator (eq. B4), the displacement ̂ = A+ e−
φ ikp z
+ A− e − ikp (h− z)
(10a)
field (eq. (3)) in the transformed domain becomes
ψ̂ = B+ e− iks z
+ B− e − iks (h− z)
(10b)
∂φ
̂
w
̂= + kψ
̂ (5a) -
It is noted that A and B in eq. (10) are not the same as those in eq.
−
∂z
(9), even though the same notations are used (they are unknown co­
efficients and can be easily connected with each other). Since the layer
thickness h is constant, the effect of this modification can be absorbed in

3
C.-P. Lin et al. Soil Dynamics and Earthquake Engineering 161 (2022) 107423

the new unknowns A- and B− . Substituting eq. (10) into (5) gives Making use of eqs (12) and (14), and eliminating the involved co­
efficient vector, we obtain the relationship between the tractions and
w
̂ = − iA kp e+ − ikp z −
+ iA kp e − ikp (h− z)
+ B ke + − iks z
+ B ke − − iks (h− z)
(11a) displacements on both interfaces of the jth layer.
[ ] (j)
u = − A+ ke−
̂ ikp z
− A− ke− ikp (h− z)
+ iB+ ks e− iks z
− iB− ks e− iks (h− z)
(11b) ̂ (j) = K (j) U
T ̂ (16)
Rewriting eq. (11) in a matrix form, the displacements on the upper
(z = 0) interface (denoted by subscript 1) and lower (z = h) interface where the local stiffness matrix of the jth layer, [K (j) ], is symmetric and
(denoted by subscript 2) of the jth layer become determined by
[ ] [ ][ ]− 1
[ ] K(j) = H(j) E(j)
̂ (j) = E(j) C(j)
U (12)
For the lower homogeneous half space (i.e. the lth layer in Fig. 1),
where there is only the top interface, and the backward (upward) reflection
[ ]T term can be dropped in eqs (12) and (14). Then, the displacement,
̂ (j) = ̂u (j) ̂ (j) u (j) ̂ (j)
U 1 w 1 ̂ 2 w 2 traction, and unknown coefficient vectors are reduced to
[ ]T [ ]T
is the displacement vector, ̂ (l) = ̂
U u (l)
1 w ̂ (l)
1
̂ (l) = − ̂τ (l) − ̂
, T zr1 σ (l)
zz1 ,
[ ]T
C(j) = A+(j) B+(j) A− (j) B− (j)
and
is the unknown coefficient vector, and [ ]T
C(l) = A+(l) B+(l) ,
⎡ ⎤
− k iks − ke− ikp h − iks e− iks h

[ (j) ] ⎢ − ikp k ik − ikp h

ke− iks h ⎥ respectively. Hence, the local stiffness matrix for the lower homoge­
pe
E =⎢ ⎣ − ke− ikp h − iks h
⎥
⎦ (13) neous half space (the lth layer) is only 2 × 2 and can be explicitly derived
− iks e − k iks
− ikp e− ikp h ke− iks h ikp k as
[ ( 2 ) ( ) ]
A superscript (j) is added in eqs (12) and (13) to indicate that these [ (l) ] μ ikp ks + k2 2kp ks − ks2 + k2 k
relations are for the jth layer. The superscript is omitted in the matrix K = 2 ( ) ( ) (17)
k + kp ks 2kp ks − ks2 + k2 k iks ks2 + k2
elements for simplicity, with understanding that the parameters in eq.
(13) are those of the jth layer. Applying the constitutive relation (eq. Similar to the finite-element formulation, the global stiffness matrix
(8)), we obtain the following equation for tractions on both interfaces of [K] which links the nodal (i.e., being at the layer interface) displace­
the jth layer. ments and nodal forces (i.e., the external load on the layer interfaces and
[ ] surface of the layered system) can be written as
̂ (j) = H(j) C(j)
T (14)
̂ = [K] U
P ̂ (18)
where
[ ]T where the nodal displacement vector is
̂ (j) = − ̂τ (j) − ̂
T σ (j) τ (j) σ (j)
zz1 ̂ zr2 ̂
zr1 zz2 ̂ = [̂
U u1 w u2 w
̂1 ̂ ̂ l ]T ,
ul w
̂ 2 …̂

is the traction vector, and the nodal force (traction) vector in global coordinates is
⎡ ⎤
( ) ( ) ̂ = [P
̂ r1 P
̂ z1 P
̂ r2 P
̂ z2 … P ̂ zl ]T ,
̂ rl P
− ks2 − k2 2ikkp e− ikp h − ks2 − k2 e− iks h P
⎢ −( 2ikkp ) ( 2 2) − ⎥
⎢ 2 2 ikp h ⎥
[ (j) ] ⎢ ks − k 2ikks ks − k e − 2ikks e− iks h ⎥
H = μ⎢ ⎢ 2ikk e− ikp h ( 2 2 ) − iks h ( 2 2) ⎥
⎥ and the global stiffness matrix assembled from all local stiffness matrix is
⎢ p k − k e − 2ikkp ks − k ⎥
⎣ ( 2 2 ) − ik h s ( ) ⎦
− ks − k e p − 2ikks e− iks h − ks2 − k2 2ikks

(15)

⎡ ⎤
⎢K̂ (1) ̂ (1)
K ̂ (1)
K ̂ (1)
K ⎥
⎢ 11 12 13 14 ⎥
⎢ (1) (1) (1) ⎥
⎢ ̂
K ̂
K ̂
K ⎥
⎢ 22 23 24 ⎥
⎢ ⎥
⎢ (1)
̂ +K
K ̂ (2)
̂ (1) + K
K ̂ (2) ̂ (2)
K ̂ (2)
K ⎥
⎢ 33 11 34 12 13 14 ⎥
⎢ ⎥
⎢ (1)
̂ 22
̂ 44 + K (2) (2)
̂ 23 ̂ (2) ⎥
⎢ K K K 24 ⎥
⎢ ⎥
⎢ ̂ (2) ̂ (3) ̂ (2) ̂ (3) ⎥
⎢ K 33 + K 11 K 34 + K 12
⎥
⎢ ⎥
[K] = ⎢
⎢ ̂ (2) ̂ (3)
⎥
⎥ (19)
⎢ K 44 + K 22 ⎥
⎢ ⋱ ⎥
⎢ ⎥
⎢ 1) 1) 1) (l− 1) ⎥
⎢
⎢ ̂ (l−
K 11
̂ (l−
K 12
̂ (l−
K 13
̂ 14
K ⎥
⎥
⎢ 1) 1) (l− 1) ⎥
⎢
⎢ ̂ (l−
K 22
̂ (l−
K 23
̂ 24
K ⎥
⎥
⎢ ⎥
⎢ ̂ (l− 1)
̂ (l) ̂ (l− 1)
̂ (l) ⎥
⎢ sym. K 33 +K 11 K 34 + K 12 ⎥
⎢ ⎥
⎣ ̂ (l) ⎦
̂ (l− 1) + K
K 44 22

4
C.-P. Lin et al. Soil Dynamics and Earthquake Engineering 161 (2022) 107423

∑
M (α )
In the seismic testing with vertical impact, all the tractions (or nodal u1 (r, ω) = u 1m J
̂
m
r (23a)
1
forces) in the transformed domain are zeros except P ̂ z1 , which is the m=1
R
Hankel transform of the unit vertical load on the surface Pz1 (r) =
(α )
1 for r ≤ a. Solving eq. (18) gives the displacements on all the layer in­ ∑
M
(23b)
m
w1 (r, ω) = w
̂ 1m J r
terfaces in the transformed domain.
0
m=1
R

It is observed again from eq. (23) that different orders of Bessel

2.3. Computation from transformed domain to physical domain functions are needed. However, they can be truncated at the same M
during the forward and inverse calculations. Synthetic waveforms can
Computation of the solution in the transformed domain does not cost be retrieved by multiplying eq. (23) with the Fourier series coefficient of
much time. What is computationally intensive is the integral trans­ the time-varying loading and taking the Fourier series expansion (or
formation back to the physical domain. Hence, this study mainly focused inverse FFT). We focus mainly on the synthetic dispersion spectra in the
on the computation from the transformed domain to the physical rest of the paper. The wavenumber or velocity resolution in dispersion
domain. Although the solutions above are derived by the stiffness matrix spectra is controled by the receiver spread length; the receiver spacing
method, the computation framework introduced in the following is very only affects the aliasing criterion. The same aliasing is actually modeled
general where the Green’s functions can be solved by other layer matrix in the synthetic dispersion spectra. To generate synthetic dispersion
method (e.g., the DVP method) and for other loading conditions. We do spectra, only a dozen of receiver locations (r) over the interested fre­
not need to transform the frequency-domain solution to the time domain quency range (ω) are needed in eq. (23). To obtain accurate synthetic
for computing the dispersion spectrum, but we do need to transform the seismograms, however, much more frequency points are required in the
wavenumber to the physical space (r) to obtain the spectral wavefield by Fourier series expansion.
eq. (B2). Numerical integration of eq. (B2) for obtaining the physical-
domain wavefield is time consuming. Thus, we propose the Fourier- 3. Results and discussion
Bessel series to expand any function of r, g(r), as
∑
∞ (α ) 3.1. Model verification
(20)
m
g(r) = gm J
̂ υ r
R
m=1 To verify the synthetic wavefield computed by the proposed
approach, the dynamic version of a layered elastic analysis code called
where m in ̂ g m denotes the mth Fourier-Bessel coefficient. The discrete
MultiSmart3D (e.g. Refs. [59,62], is used. It is based on a novel cylin­
wavenumber points are defined as km = αm /R in which αm is the mth drical system of vector functions combined with the new DVP method.
zero of the Bessel function J υ (αm ) = 0 and R is chosen to be large MultiSmart3D has been thoroughly validated for its accuracy and effi­
enough so that function g(r) is very small or equal to zero when r > R. ciency in both engineering and science fields [13,14,59,63]. Table 1 lists
Compared to the integral transform method, the series approach is the different earth models used to verify the proposed stiffness matrix
computationally more efficient for a given accuracy. This is due to the formulation and the Fourier-Bessel series expansion (FBS) approach
fact that, the Fourier-Bessel series method requires only simple sum­ (denoted as FBS in figure legends): two homogeneous models (one with
mation over discrete zero points of the Bessel functions, whereas nu­ Poisson’s ratio 0.33 and the other 0.48), a three-layered model (velocity
merical integration has to be carried out over each interval between
increases with depth), and a four-layered model (with an embedded soft
these zero points in the traditional integral transform method. layer).
Based on the orthogonal relations among different eigenfunctions
Figs. 2–4 present the spectral dynamic responses at some selected
[60], the expansion coefficient in eq. (20) can be expressed as frequencies for the earth models listed in Table 1. The radius of the
∫R (α ) surface loading circle (a) and the uniform load density used are 0.12 m
gm =
̂
2
g(r)J υ
m
r rdr (21) and 280 N/m2, respectively, in both MultiSmart3D and FBS. The vertical
2
R2 J υ+1 ( m ) α
0
R displacement on the ground surface (w1) is shown from 5 m to 28 m (i.e.
the distance from the source to receivers) at 1 m interval, mimicking a
Notice that while the selected time window (T) determines the fre­ 24-channel survey with 5 m near offset and 1 m receiver spacing. Both
quency resolution, the selected R value, i.e., the selected spatial window, the real and imaginary parts of the frequency response oscillate and
determines the wavenumber resolution or the smallest wavenumber attenuate with the distance from the source, clearly showing different
(corresponding to the largest wavelength). Hence, selection of R can be wavelengths (spatial periods) at different frequencies. The results of FBS
decided according to the maximum possible wavelength in the wave­ match the MultiSmart3D benchmark extremely well for all the four cases
field. The summation in eq. (20) can be truncated at M if the smallest listed in Table 1. For the homogeneous model, the comparison with
wavelength in the wavefield is equal to or greater than 2R/M (i.e.,
maximum wavenumber is less than km = α2RM ).
Table 1
The Fourier-Bessel coefficient of P
̂ z1 in eq. (18) for the unit vertical
Earth models used for verification.
load over a circle on the surface has a close-form expression when
substituting the load distribution (Pz1 (r) = 1 for r ≤ a) into eq. (21) Layer Thickness VS VP Density

(order 0) [60,61]. (m) (m/ (m/ (kg/

s) s) m3)
∫R (α ) (α )
2 2a Homogeneous-1 1 Inf. 300 600 2000
(22)
̂ z1m = m m
P J 0 r rdr = J 1 a
R2 J 2υ+1 ( m ) α R 2
αm RJ 1 (αm ) R Homogeneous-2 1 Inf. 300 1500 2000
0
Three-layered (regular 1 2 200 400 1700
where, in this case, αm is the mth zero of Bessel function J 0 (αm ) = 0. profile) 2 4 240 480 1800
Substituting eq. (22) into (18), we can solve the Fourier-Bessel co­ 3 Inf. 320 640 2000
efficients for all the interface displacements in the transformed domain Four-layered (irregular 1 2 200 400 1700
from the global stiffness equation. In our present application, we are profile) 2 4 150 300 1650
particularly interested in u1 and w1 on the ground surface. These dis­ 3 5 240 480 1800
4 Inf. 320 640 2000
placements in the frequency domain can be expressed as

5
C.-P. Lin et al. Soil Dynamics and Earthquake Engineering 161 (2022) 107423

Fig. 2. Comparison of the normalized vertical displacements on the ground surface of the homogeneous half space (Table 1) for fixed frequency at (a) f = 5 Hz, (b) f
= 30 Hz, and (c) f = 70 Hz (FBS vs. MultiSmart3D). Top row is the real part and bottom row the imaginary part of the displacement.

Fig. 3. Comparison of the normalized vertical displacements on the ground surface of the 3-layered model (Table 1) for fixed frequency at (a) f = 5 Hz, (b) f = 30 Hz,
and (c) f = 70 Hz (FBS vs. MultiSmart3D). Top row is the real part and bottom row the imaginary part of the displacement.

Fig. 4. Comparison of the normalized vertical displacements on the ground surface of the 4-layered model (Table 1) for fixed frequency at (a) f = 5 Hz, (b) f = 30 Hz,
and (c) f = 70 Hz (FBS vs. MultiSmart3D). Top row is the real part and bottom row the imaginary part of the displacement.

MultiSmart3D is only shown for the case of Poisson’s ratio (ν = 0.33) in The computation time and accuracy of FBS depend on the parameters
Fig. 2. The FBS results of Poisson’s ratio (ν = 0.48) is also plotted in R and M used. In Figs. 2–4, R = 1500 m and M = 9000 are used. The FBS
Fig. 2 to illustrate the notable effect of Poisson’s ratio on the dynamic approach is very efficient. Taking the 3-layered model for example, it is
response. about 500 times faster than the integral-transform based MultiSmart3D,

6
C.-P. Lin et al. Soil Dynamics and Earthquake Engineering 161 (2022) 107423

not to mention that smaller R and M could be used to further speed up field data (experimental wavefield).
the computation without losing much accuracy. The appropriate R and Applying eq. (24), we obtain the dispersion spectra of the four syn­
M depend on the minimum and maximum wavelengths of interest, thetic wavefields as shown in Figs. 5 and 6. While Figs. 2–4 show the
which can be selected easily and intuitively. vertical components for only three selected frequencies, Figs. 5 and 6
To compare the entire wavefield, the wavefield in eq. (23) is trans­ effectively present the dispersion spectra of both the vertical and radial
formed to obtain the velocity spectrum by the following equations, components, respectively, for all frequencies of interest. It is noted that
equivalent to the f-k transform or phase shift method [64,65]. these dispersion spectra (also referred to as velocity spectra since the
( ) results are presented in frequency-velocity domain) are receiver array
∑
N
iω
u *1 (ω, v) =
̂ u1 (rn , ω)exp − rn (24a) dependent. The resolution of a velocity spectrum can be enhanced by
n=1
v increasing the spread length of a receiver array. The velocity spectrum
allows more comprehensive verification of the FBS results (bottom plots
( )
∑ in Figs. 5 and 6) by comparing them with the MultilSmart3D results (top
N
iω
̂ *1 (ω, v) =
w w1 (rn , ω)exp − rn (24b)
n=1
v plots in Figs. 5 and 6). The peaks of the velocity spectrum are considered
as the “effective” dispersion curves. The theoretical modal dispersion
where v is the phase velocity and subscript n denotes the nth receiver curves, including fundamental mode and higher modes, are also shown
location (with N = 24). We use (ω, v) instead of (ω, k) in the theoretical in Figs. 5 and 6 for comparison (using the present method for the case
derivation as it is more frequently adopted in processing and presenting with zero applied force). The velocity spectra obtained by the proposed
data. One can set a physically reasonable boundary for the phase ve­ FBS approach are basically identical to the results of MultiSmart3D.
locity, and the sampling of reasonable phase velocity is better preserved Both reveal the near-field effect in all the three cases and the mode jump
in the f-v spectrum [11]. Other representation of the transformed spec­ in the 4-layered irregular profile. The near-field effect at lower fre­
trum, such as (ω, p) with p being the slowness, has also been used (e.g., quencies (longer wavelengths) may be mitigated by using a larger near
Refs. [15,16]. We use the superscript * to differentiate ̂
*
u 1 (ω, v) and offset or/and spread length of the receiver array. However, the signal
attenuates rapidly as the offset increases. Full wavefield analysis can
̂ *1 (ω, v) from the true theoretical values (̂
w u 1 and w
̂ 1 ) in eq. (18). It
fully consider the near-field effect, making field survey more flexible.
should be pointed out that it is impossible to recover the true wavefield
Contribution of higher modes in the wavefield is intrinsic to the velocity
in the transformed domain, ̂ u 1 (ω, v) and w
̂ 1 (ω,v), from a limited number
profile. However, the apparent dispersion curve is affected not only by
of receivers, not to mention the factor that the plane-wave trans­
the velocity profile but also by the receiver array configuration. Higher
formation is used here. Instead, eq. (24) uses the synthetic wavefield to
*
modes may not be separated by a short receiver array due to spatial
u 1 (ω, v) and
theoretically simulate the experimental dispersion spectra, ̂ truncation. Effect of higher modes and array size is also fully taken into
̂ *1 (ω, v). For inversion, these results are then used to match the field
w account by eq. (24). While the velocity spectrum obtained from the
experimental dispersion spectra, also obtained by eq. (24) but with the vertical component (Fig. 5) is similar to that of the radial component

Fig. 5. Dispersion images and dispersion curves of the vertical component signal from different models: (a) homogeneous half space (ν = 0.33); (b) homogeneous
half space (ν = 0.48); (c) 3-layered model; (d) 4-layered model. Results from MultiSmart3D are shown on the top whilst those from FBS on the bottom. The solid lines
in the spectra are the theoretical modal dispersion curves.

7
C.-P. Lin et al. Soil Dynamics and Earthquake Engineering 161 (2022) 107423

Fig. 6. Dispersion images and dispersion curves of the radial component signal from different models: (a) homogeneous half space (ν = 0.33); (b) homogeneous half
space (ν = 0.48); (c) 3-layered model; (d) 4-layered model. Results from MultiSmart3D are shown on the top whilst those from FBS on the bottom. The solid lines in
the spectra are the theoretical modal dispersion curves.

(Fig. 6), the near-field effect in the radial component is significantly possible difference between vertical and radial velocity spectra (or
different from that in the vertical component. Both vertical and radial effective dispersion curves) in the full wavefield analysis could provide
components can be easily measured by vertical and horizontal geo­ more constraints on the velocity model that can fit the multi-component
phones. While normal modes are theoretically invariant to direction, the data.

Fig. 7. Effective dispersion curves from FBS and modal summation for different earth models: (a) homogeneous half space with ν = 0.33; (b) homogeneous half space
with ν = 0.48; (c) 3-layered regular model; (d) 4-layered irregular model.

8
C.-P. Lin et al. Soil Dynamics and Earthquake Engineering 161 (2022) 107423

3.2. Comparison with modal summation and finite element method velocity seismograms (normalized by the maximum value of each trace)
of the two homogeneous models using the proposed FBS algorithm. They
The proposed FBS algorithm can significantly save the computa­ are compared to the spectral-element method (SEM, a high-order finite
tional time as compared to the conventional integration method (e.g., element technique) using the open-source package SPECFEM3D [69,70]
MultiSmar3D). For a given accuracy, it is at least two-orders faster than in Fig. 8(b) (SEM in the middle row), and the modal summation method
the latter one! However, compared to the approximate modal summa­ in Fig. 8(c) (in the bottom row). To model the dynamic response subject
tion method, the FBS still requires more computation time. In this sec­ to a point source or circular load, the computationally intensive 3D
tion, we compare our FBS results with those obtained by the modal model SPECFEM3D has to be used. The results shown in Fig. 8(b) (in the
summation method using the computer programs developed by Herr­ middle) are obtained by setting the mesh size to 0.7 m under a point
mann [28]. We also compare our results with those obtained by the source of Gaussian function with 0.01 s half duration, instead of the
finite element method. These comparisons will demonstrate the effi­ actual circular load in the FBS. The computation time of SPECFEM3D is
ciency and accuracy of the proposed FBS method. The receiver config­ roughly 5000 times longer than the FBS! Though the results of SPEC­
uration remains the same as in the previous examples (5 m near offset, FEM3D are slightly different from those based on the FBS due to the
24 channels and 1 m receiver spacing). To better illustrate the difference difference in source condition and numerical errors, higher-speed waves
between our FBS-based full wavefield simulation and the modal sum­ simulated by the FBS are clearly verified by this finite-element code
mation, only the peaks of the dispersion spectrum are plotted in Fig. 7, SPECFEM3D. However, for the finite element method to achieve the
along with the modal dispersion curves. It is observed that both methods same accuracy of FBS, it would require finer mesh with smaller time
predict similar results at high frequencies. However, the dispersion interval, resulting in much longer computation time. Since the Poisson’s
curve from the modal summation starts to deviate from the correct one ratio is significantly larger than 0.263 [68], the faster leaky waves and
(i.e., the FBS) as frequency decreases. This can be attributed to the lateral waves become pronounced. These features are clearly captured in
near-field effect and leaky wave, which cannot be correctly modeled by the results of the FBS method and finite element method, but are missing
the modal summation method. It is worth noticing that the effective in those of the modal summation method. This is particularly important
dispersion curve of the vertical component tends to underestimate the when the Poisson’s ratio is larger, where the leaky wave can be mis­
fundamental mode at lower frequencies [66,67], whilst the opposite interpreted as the high velocity fundamental mode [49]. Since recog­
trend may occur in the horizontal component. We analyze these features nizing and filtering out the leaky surface waves could be tricky, it is
below, based now on the finite element method. hence better to be able to include them in the theoretical forward model.
We point out that, in addition to the near-field effect, the proposed One of the major advantages of using full wavefield instead of modal
FBS method can also completely model other possible waves. These dispersion curve is that the directional particle motion characteristics
include the leaky wave that travels along the surface with a wave speed provide additional constraints to the earth model. In addition to the
smaller than the compressional wave but larger than the shear wave, vertical and radial velocity spectra, the radial to vertical spectral ratio
and the lateral wave (induced by the compressional wave at the surface) (RVSR) and radial to vertical phase difference (RVPD) can be examined
[68]. These higher-speed waves are absent in the modal summation and utilized in the inversion. Fig. 9 shows the RVSR and RVPD of the last
approach. As an example, Fig. 8(a) (in the top row) shows the synthetic receiver (28 m offset) of the earth models, calculated by the FBS method,

Fig. 8. Synthetic seismograms of the homogeneous model with different Poisson’s ratio ν using the FBS approach in 8a (top row), the SEM approach in 8b (middle
row), and the modal summation method in 8c (bottom row). Left two columns are for the vertical component and right two columns for the radial component.

9
C.-P. Lin et al. Soil Dynamics and Earthquake Engineering 161 (2022) 107423

Fig. 9. The RVSR (top row) and RVPD (bottom row) frequency curves for different earth models based on both FBS and modal summation: (a) homogeneous half
space with ν = 0.33; (b) homogeneous half space with ν = 0.48; (c) 3-layered regular model; (d) 4-layered irregular model.

as compared to those by the modal summation method. It is obvious that 3.3. Implications on seismic testing
significant difference exits between the rigorous full wavefield solution
by the FBS algorithm and the result of modal summation method. Some In this section, we present three pairs of cases to demonstrate the
interesting behaviors in the RVSR and RVPD for the 4-layered irregular capability of the proposed FBS method in simulating completely and
model should be noticed. As shown in Fig. 9(d1), the FBS method ap­ efficiently the wavefield in the context of seismic testing. The wavefields
pears to reveal a high frequency resonance in RVSR due to the embedded are simulated using 48 receiver locations with 5 m near offset and 1 m
shallow soft layer. Furthermore, Fig. 9(d2) reveals that there are sudden receiver spacing. The models differ mainly in the shear-wave velocity
jumps in the RVPD curve at the same frequencies of mode jumping as profile, fixing the Poisson’s ratio at 1/3 (except for the last pair of cases,
indicated in Figs. 5(d) and 6(d). Furthermore, due to the embedded soft in which the effect of Poisson’s ratio is examined) and with density
layer, the particle motion at some frequencies turns from retrograde to slightly change in proportion to shear-wave velocity. Modal dispersion
prograde (i.e., RVPD from negative to positive). These phenomena imply curves, effective dispersion curves, and dispersion spectra are presented
that the RVSR and RVPD could provide strong constraints on the elastic to show different levels of information for uniquely characterizing the
earth model. shear-wave velocity profiles.
Inversion limitation by considering only the fundamental mode is
demonstrated in the first pair of cases, as shown in Fig. 10. Both shear

Fig. 10. (a) The vs-profiles; (b) the velocity spectra and effective dispersion curves of vertical component (top) and radial component (bottom) for Profile A; (c) the
velocity spectra and effective dispersion curves of vertical component (top) and radial component (bottom) for Profile B; (d) comparison of the two fundamental-
mode dispersion curves within the same frequency range for vertical component (top) and radial component (bottom). Lines in (b) and (c) are normal modes.

10
C.-P. Lin et al. Soil Dynamics and Earthquake Engineering 161 (2022) 107423

velocity Profile A and Profile B (Fig. 10(a)) have an embedded soft layer It is typically assumed that the surface wave dispersion is mainly
at shallow depth; the difference is that the soft layer in Profile A has a governed by the shear-wave velocity. The third pair of cases demon­
lower shear-wave velocity than the surface layer, resulting in higher strates a scenario where such an assumption is not entirely true. Fig. 12
modes dominant at higher frequencies. Because of the longer receiver (a) shows two vs-profiles, one of which (Profile E) has lower vs but larger
array used in this section, the fundamental mode can be clearly identi­ Poisson’s ratio (0.45 instead of 0.33 for other cases). It is noted that a
fied and separated from the higher modes. However, within the fre­ large Poisson’s ratio is not uncommon in saturated sediments. The ve­
quency range (5–38 Hz) where the fundamental mode is experimentally locity spectra and effective dispersion curves of these two vs-profiles are
available, the fundamental mode dispersion curve of Profile A is very shown in Fig. 12(b) and (c). The two dispersion curves are almost the
close to that of Profile B as shown in Fig. 10(d), although their vs-profiles same as shown in data points of Fig. 12(d). It is generally considered that
are significantly different. This indicates that discarding high modes in fundamental modes dominate in regular vs-profiles where vs increases
the fundamental-mode inversion would aggravate the non-uniqueness. monotonically with depth, as also the case in Fig. 12. However, as
In a wider frequency range, Fig. 10(b) and (c) show that the two ve­ Poisson’s ratio becomes larger, the observed dispersion curve at the
locity profiles have distinct effective dispersion curves and velocity lower end of frequencies actually corresponds on the leaky wave. As a
spectra. Thus, including high modes in terms of effective dispersion result, the lower-frequency effective dispersion curve is due to the leaky
curve or considering the entire velocity spectrum provides additional mode in Fig. 12(b) (Profile E), while it is predominantly fundamental
features, which can help distinguish one velocity profile from the other. mode in Fig. 12(c) (Profile F). Both may be conventionally interpreted as
It can be observed from Fig. 10(b) that what appears as the first higher the fundamental mode of a regular vs-profile. As such, performing
mode is in fact a smooth transition from the 1st to 2nd higher mode in fundamental-mode inversion in the case of Profile E would result in
40–63 Hz, and the seemingly 2nd higher mode is actually a smooth serious overestimation of vs in deep layers. Effective-mode inversion is
transition from the 3rd to 4th higher mode in 64–100 Hz. Hence, multi- deemed more appropriate than fundamental-mode inversion in all cases.
mode inversion of normal modes is not feasible in this case. Yet, the effect of Poisson’s ratio (or compressive wave velocity vp)
The second pair of cases depicted in Fig. 11 are used to illustrate the cannot be neglected when leaky waves are pronounced. Simultaneous
reduced sensitivity with increasing depth. It is a common understanding inversion of vp requires more information than just the effective
that part of the non-uniqueness in surface wave inversion arises from dispersion curve. One possibility is to simultaneously consider the first-
rapid reduction of sensitivity with increasing depth. In Fig. 11(a), the arrival travel time curve for P-wave refraction tomography. However,
two vs-profiles have the same shear-wave velocity up to 7 m depth, but reliable picking of the first break could be tricky. Fig. 12(d) shows that
show significant difference below 7 m. Their velocity spectra and the two vs-profiles are significantly different in the velocity spectra
dispersion curves are shown in Fig. 11(b) and (c), respectively. The beyond the peak effective dispersion curves, and the difference is also
fundamental mode dominates in all frequencies for both cases, and their quite different in the vertical and radial components. Thus, simulta­
peak effective dispersion curves are almost identical as shown in the neous inversion of vs and vp based on the full-velocity spectrum can be
data points of Fig. 11(d), even though there is a significant difference in further investigated.
vs-profiles below 7 m. The same level of inversion uncertainty is ex­ Up to this point, the vertical and radial components of the six vs-
pected below 7 m regardless if it is fundamental-mode or effective-mode profiles are considered separately by their respective velocity spectra.
inversion. Higher modes have longer wavelength, and hence may help Examining how these two components are coupled in the same survey
improve the sensitivity to the velocity at larger depths. However, pick­ may provide more constraints, especially when both vs and vp are to be
ing the effective dispersion curves due to higher modes is not an easy simultaneously inverted. Fig. 13 shows the RVSR and RVPD curves at
task as can be seen from Fig. 11(b) and (c). Fig. 11(d) shows the velocity the middle receiver (offset = 28 m) for the six cases discussed in this
spectrum difference between the two vs-profiles where significant dif­ section. In each pair of the cases, significant differences in the RVSR and
ference in the velocity spectrum beyond the fundamental mode can be RVPD curves can be observed even though they may have similar
observed. In this case, different information contained in higher modes effective dispersion curves as shown in Figs. 11(d), 12(d) and 13(d). In
can only be exploited by utilizing the entire velocity spectrum (within particular, Profile A (with an embedded layer softer than surface layer)
proper frequency and velocity range). and Profile E (with a large Poisson’s ratio) exhibit positive RVPD

Fig. 11. (a) The vs-profiles; (b) the velocity spectra and effective dispersion curves of vertical component (top) and radial component (bottom) for Profile C; (c) the
velocity spectra and effective dispersion curves of vertical component (top) and radial component (bottom) for Profile D; (d) comparison of the two effective
dispersion curves and their velocity spectrum difference for vertical component (top) and radial component (bottom). Lines in (b) and (c) are normal modes.

11
C.-P. Lin et al. Soil Dynamics and Earthquake Engineering 161 (2022) 107423

Fig. 12. (a) The vs-profiles; (b) the velocity spectra and effective dispersion curves of vertical component (top) and radial component (bottom) for Profile E; (c) the
velocity spectra and effective dispersion curves of vertical component (top) and radial component (bottom) for Profile F; (d) comparison of the two effective
dispersion curves and their velocity spectrum difference for vertical component (top) and radial component (bottom). Lines in (b) and (c) are normal modes.

Fig. 13. RVSR (top) and RVPD (bottom) for the three pairs of cases at the middle receiver (offset = 28 m): (a) Profiles A and B; (b) Profiles C and D; (c) Profiles E
and F.

(prograde motion) at certain frequencies. Although the velocity spectra, the source). Its efficiency is achieved by replacing the computationally
RVSR, and RVPD are all derived from the same dynamic response, it is intensive integral transform with the Fourier-Bessel series expansion
believed that the RVSR and RVPD can provide complementary infor­ (FBS) in the radial direction (not the traditional Fourier-Bessel trans­
mation to the velocity spectra as they reveal the coupling between form). Dynamic response of an isotropic elastic layered half-space sub­
vertical and radial components, which is also a function of the elastic jected to vertical loading is implemented and thoroughly validated. The
earth model. proposed algorithm is at least two-orders faster than the conventional
integral transform method and three-orders faster than the finite
4. Conclusion element code. Comparing to the fast modal summation method, the
proposed framework preserves the complete wavefield, including the
To accurately exploit the complete signal content in seismic testing, near-field effect, leaky waves, and lateral waves.
an efficient computational framework for modeling synthetic full Three pairs of velocity earth models are further used to demonstrate
wavefields (mainly in the frequency domain) in layered systems has how to make use of this fast FBS method in order to reveal the important
been proposed considering the survey parameters (i.e., area and offset of features in the full wavefield. Experimentally available fundamental

12
C.-P. Lin et al. Soil Dynamics and Earthquake Engineering 161 (2022) 107423

mode, broadband effective dispersion curve, and the full-velocity spec­ Author statement
trum represent different levels of information on the elastic earth model.
Different earth models may have similar fundamental modes or even Chih-Ping Lin: Conceptualization, Methodology, Software, Writing -
effective dispersion curves; yet, their difference manifests in different Original Draft, Supervision, Funding acquisition. Tsai-Jung Wu: Meth­
higher-mode contributions (including leaky mode) and these can only be odology, Software, Formal analysis, Writing - Review & Editing, Visu­
exploited by utilizing the entire velocity spectra including the vertical alization. Ernian Pan: Writing - Review & Editing, Supervision.
and radial components. There are cases where compressive wave ve­
locity vp cannot be neglected in the surface wave dispersion character­
istics when leaky waves are pronounced. Examining how vertical and Declaration of competing interest
radial components are coupled in the same survey (e.g., in terms of
RVSR and RVPD) may provide more constraints, especially when both vs The authors declare that they have no known competing financial
and vp are to be simultaneously inverted. Velocity spectra, RVSR, and interests or personal relationships that could have appeared to influence
RVPD are all source-independent attributes that can be readily used for the work reported in this paper.
inversion.
This study focuses on the forward computational kernel and intends Acknowledgements
to shed some lights on why and how full wavefield should be considered.
Field application of the proposed FBS approach and other loading types The research is funded by Ministry of Science and Technology,
and related Green’s functions will be considered in the future. Taiwan, R.O.C. under project number 110-2625-M-A49.

Appendix A. Wave equation of axis-symmetric motion in cylindrical coordinates

The wave equation (without body force) for the elastic wave propagation in an isotropic solid can be expressed in vector calculus form as
(λ + 2μ)∇∇ ⋅ u − μ∇ × ∇ × u = ρü (A1)
According to the Helmholtz decomposition, the displacement field (u) can be expressed as the sum of the gradient of a scalar potential (φ) and the
curl of a vector potential (ψ ) whose divergence vanishes (i.e., ∇⋅ψ ) [51].
u = ∇φ + ∇ × ψ (A2)
Taking the divergence and curl of eq. (A2) leads to the following relations,
∇ ⋅ u = ∇2 φ (A3a)

∇×u = ∇ × ∇ × ψ (A3b)
Substituting eq (A2) and (A3) into eq. (A1) gives two independent wave equations in terms of Helmholtz potentials:
(λ + 2μ)∇2 φ = ρφ̈ (A4a)

− μ∇ × ∇ × ψ = ρψ̈ (A4b)

In cylindrical coordinates (r, θ, z) and applying the condition of axis-symmetric motion in Fig. 1, we have ∂φ
∂θ = 0, ψ = ψ θ iθ (where iθ is the unit
vector in θ-direction, and ψ θ is denoted as ψ hereafter for convenience), and ∂ψ
∂θ = 0. Then, the scalar and vector differential operations on the left-hand
sides of eq. (A4) are reduced to
∂2 φ 1 ∂φ ∂2 φ
∇2 φ = + + (A5a)
∂r2 r ∂r ∂z2

∂2 ψ 1 ∂ψ ∂2 ψ ψ
− ∇ × ∇ × ψ iθ = + + − (A5b)
∂r2 r ∂r ∂z2 r2
Substituting eq. (A5) into (A4) leads to eq. (1) in the main text, i.e., the wave equation in terms of Helmholtz potentials in cylindrical coordinates.

Appendix B. Hankel transform and its properties

− 1
The Hankel transform H n {} and its inverse transform H n {} of order n are defined as
∫∞
g (k) = H n {g(r)} =
̂ rg(r)J n (kr)dr (B1)
0

∫∞
g(r) = H − 1
n {̂g (k)} = k̂
g (k)J n (kr)dk (B2)
0

where g(r) represents any function in the physical domain r (e.g., the vertical and radial displacements on the surface w1 (r) ​ and ​ u1 (r) in Fig. 1); ̂
g (k)
denotes the corresponding function in the transformed domain k (hereafter, a hat is used to denote any function in the transformed domain); J n () is
the Bessel function of the first kind of order n. Only orders 0 and 1 are used in this study. The Hankel transform has several elementary properties [71]
as follows:

13
C.-P. Lin et al. Soil Dynamics and Earthquake Engineering 161 (2022) 107423

1. Derivative property:

(B3)
′
H 1 {g (r)} = − kH 0 {g(r)}

2. The Hankel transform of the Bessel differential operator:

For the Bessel differential operator Δn,

d2 1 d (n)2
Δn = 2 + −
dr r dr r

H n {Δn g(r)} = − k2 H n {g(r)} (B4)

3. Another important relation

{ }
d [ n+1 ]
H n r− n− 1 r g(r) = kH n+1 {g(r)} (B5)
dr
For the special case n = 0, eq. (B5) becomes
{ }
1 d
H 0 [rg(r)] = kH 1 {g(r)} (B6)
r dr

References [26] Gao Q, Lin JH, Zhong WX, Howson WP, Williams FW. A precise numerical method
for Rayleigh waves in a stratified half space. Int J Numer Methods Eng 2006;67:
771–86.
[1] Zhou YG, Chen YM. Laboratory investigation on assessing liquefaction resistance of
[27] Panza GF. Synthetic seismograms: the Rayleigh waves modal summation.
sandy soils by shear wave velocity. J Geotech Geoenviron Eng 2007;133:959–72.
J Geophys 1985;58:125–45.
[2] L’Heureux JS, Long M. Relation between shear wave velocity and geotechnical
[28] Herrmann RB. Computer programs in seismology: an evolving tool for instruction
parameters for Norwegian clays. J Geotech Geoenviron Eng 2017;143:1–20.
and research. Seismol Res Lett 2013;84:1081–8.
[3] Trafford A, Long M. Relationship between shear wave velocity and undrained shear
[29] Dal Moro GD. Surface wave analysis: improving the accuracy of the shear-wave
strength of peat. J Geotech Geoenviron Eng 2020;146:1–10.
velocity profile through the efficient joint acquisition and Full Velocity Spectrum
[4] Pan Y, Gao L, Bohlen T. Time-domain full-waveform inversion of Rayleigh and
(FVS) analysis of Rayleigh and Love waves. Explor Geophys 2019;50:408–19.
Love waves in presence of free-surface topography. J Appl Geophys 2018;152:
[30] Song Z, Zeng X, Thurber CH. Surface-wave dispersion spectrum inversion method
77–85.
applied to Love and Rayleigh waves recorded by distributed acoustic sensing.
[5] Pan Y, Gao L, Bohlen T. High-resolution characterization of near-surface structures
Geophysics 2021;86:EN1–12.
by surface-wave inversions: from dispersion curve to full waveform. Surv Geophys
[31] Stokoe II KH, Nazarian S. Effectiveness of ground improvement from spectral
2019;40:167–95.
analysis of surface waves. In: Proceeding of the 8th European conference on soil
[6] Lee JH, Lee SH. Full waveform inversion to estimate the material properties of a
mechanics and foundation engineering; 1983. Helsinki, Finland.
layered half-space. Soil Dynam Earthq Eng 2021;151:106956.
[32] Nazarian S. In situ determination of elastic moduli of soil deposits and pavement
[7] Liu Z, Huang L. Multiscale and layer-stripping wave-equation dispersion inversion
systems by spectral-analysis-of-surface-waves method. Ph.D. thesis. Austin:
of Rayleigh waves. Geophys J Int 2019;218:1807–21.
University of Texas; 1984.
[8] Wu RS, Luo J, Wu B. Seismic envelope inversion and modulation signal model.
[33] Park CB, Miller RD, Xia J. Multichannel analysis of surface waves. Geophysics
Geophysics 2014;79:WA13–24.
1999;64:800–8.
[9] Pérez Solano C, Donno D, Chauris H. Alternative waveform inversion for surface
[34] Xia J, Miller RD, Park CB. Estimation of near-surface shear wave velocity by
wave analysis in 2-D media. Geophys J Int 2014;198:1359–72.
inversion of Rayleigh waves. Geophysics 1999;64:691–700.
[10] Dou S, Ajo-Franklin JB. Full-wavefield inversion of surface waves for mapping
[35] Nazarian S, Stokoe II KH. Use of surface waves in pavement evaluation:
embedded low-velocity zones in permafrost. Geophysics 2014;79:107–24.
transportation Research Record. Transport Res Rec 1986;1070:132–44.
[11] Zhang Z, Alkhalifah T. Wave-equation Rayleigh-wave dispersion inversion using
[36] Sanchez-Salinero I, Roesset JM, Shao KY, Stokoe II KH, Rix GJ. Analytical
fundamental and higher modes. Geophysics 2019;84:EN57–65.
evaluation of variables affecting surface wave testing of pavements. Transport Res
[12] Li J, Feng Z, Schuster GT. Wave-equation dispersion inversion. Geophys J Int 2016;
Rec 1987;1136:86–95.
208:1567–78.
[37] Tokimatsu K, Tamura S, Kojima H. Effective of multiple modes on Rayleigh wave
[13] Liu H, Pan E, Cai Y. General surface loading over layered transversely isotropic
dispersion characteristics. J Geotech Eng 1992;118:1529–43.
pavements with imperfect interfaces. Adv Eng Software 2018;115:268–82.
[38] Gucunski N, Woods RD. Numerical simulation of the SASW test. Soil Dynam Earthq
[14] Liu Z, Li J, Hanafy SM, Schuster GT. 3D wave-equation dispersion inversion of
Eng 1992;11:213–27.
Rayleigh waves. Geophysics 2018;84:1–127.
[39] Ganji V, Gucunski N, Nazaraian S. Automated inversion procedure for spectral
[15] Forbriger T. Inversion of shallow-seismic wavefields: I. Wavefield transformation.
analysis of surface waves. J Geotech Geoenviron Eng 1998;124:757–70.
Geophys J Int 2003;153:719–34.
[40] Beaty KS, Schmitt DR, Sacchi M. Simulated annealing inversion of multimode
[16] Forbriger T. Inversion of shallow-seismic wavefields: II. Inferring subsurface
Rayleigh wave dispersion curves for geological structure. Geophys J Int 2002;151:
properties from wavefield transforms. Geophys J Int 2003;153:735–52.
622–31.
[17] Pan E. Green’s functions for geophysics: a review. Rep Prog Phys 2019;82:106801.
[41] Song X, Gu H, Liu J, Zhang X. Estimation of shallow subsurface shear-wave velocity
[18] Gabriels P, Snieder R, Nolet G. In situ measurements of shear-wave velocity in
by inverting fundamental and higher-mode Rayleigh waves. Soil Dynam Earthq
sediments with higher mode Rayleigh waves. Geophys Prospect 1987;35:187–96.
Eng 2007;27:599–607.
[19] McMechan GA, Yedlin MJ. Analysis of dispersive waves by wave-field
[42] Ivanov J, Miller RD, Xia J, Peterie S. Multi-mode inversion of multi-channel
transformation. Geophysics 1984;46:869–74.
analysis of surface waves (MASW) dispersion curves and high-resolution linear
[20] Park CB, Miller RD, Xia J. Imaging dispersion curves of surface waves on multi-
radon transform (HRLRT). SEG Technical Program Expanded Abstract; 2010.
channel record. SEG Technical Program Expanded Abstracts; 1998. p. 1377–80.
p. 1902–7.
[21] Ryden N, Park CB. Fast simulated annealing inversion of surface waves on
[43] Lu K, Li J, Guo B, Fu L, Schuster GT. Tutorial for wave-equation inversion of
pavement using phase-velocity spectra. Geophysics 2006;71:49–58.
skeletonized data. Interpretation 2017;5:SS1–SO10.
[22] Kausel E, Roësset JM. Stiffness matrices for layered soils. Bull Seismol Soc Am
[44] Xia J, Xu Y, Chen C, Kaufmann RD, Luo Y. Simple equations guide high-frequency
1981;71:1743–61.
surface-wave investigation techniques. Soil Dynam Earthq Eng 2006;26:395–403.
[23] Kausel E. An explicit solution for the Green’s functions for dynamic loads in
[45] Gao L, Xia J, Pan Y, Xu Y. Reason and condition for mode kissing in MASW method.
layered media. MIT Research Report; 1981. R81-13.
Pure Appl Geophys 2016;173:1627–38.
[24] Ramirez F, Heyliger PR, Pan E. Free vibration response of two-dimensional
[46] Lin CP, Sufwandini A, Wu TJ, Pan E. Behavior of apparent dispersion curve and its
magneto-electro-elastic laminated plates. J Sound Vib 2006;292:626–44.
implication to MASW testing. J GeoEng 2021;16:121–31.
[25] Cheung YK, Chakrabarti S. Free vibration of thick layered rectangular plates by a
finite layer method. J Sound Vib 1972;21:277–84.

14
C.-P. Lin et al. Soil Dynamics and Earthquake Engineering 161 (2022) 107423

[47] Zhang SX, Chan LS. Possible effects of misidentified mode number on Rayleigh [61] Al-Khoury R, Scarpas A, Kasbergen C, Blaauwendraad J. Spectral element
wave inversion. J Appl Geophys 2003;53:17–29. technique for efficient parameter identification of layered media. I. Forward
[48] O’Neill A, Matsuoka T. Dominant higher surface-wave modes and possible calculation. Int J Solid Struct 2001;38:1605–23.
inversion pitfalls. J Environ Eng Geophys 2005;10:185–201. [62] Alkasawneh W, Pan E, Green R. The effect of loading configuration and footprint
[49] Gao L, Xia J, Pan Y. Misidentification caused by leaky surface wave in high- geometry on flexible pavement response based on linear elastic theory. Road Mater
frequency surface wave method. Geophys J Int 2014;199:1452–62. Pavement Des 2008;9:159–79.
[50] O’Neill A, Dentith M, List R. Full-waveform P-SV reflectivity inversion of surface [63] Zhou J, Pan E, Bevis M. A point dislocation in a layered, transversely isotropic and
waves for shallow engineering applications. Explor Geophys 2003;34:158–73. self-gravitating Earth. Part IV: exact asymptotic solutions of dislocation Love
[51] Aki K, Richards GP. Quantitative seismology. second ed. University Science Books; numbers for the special case of isotropy. Geophys J Int 2021;225:664–83.
2002. [64] Lin CP, Chang TS. Multi-station analysis of surface wave dispersion. Soil Dynam
[52] Thomson WT. Transmission of elastic waves through a stratified solid media. Earthq Eng 2004;24:877–86.
J Appl Phys 1950;21:89–93. [65] Lin CP, Chang CC, Chang TS. The use of MASW method in the assessment of soil
[53] Haskell NA. The dispersion of surface waves on multilayered media. Bull Seismol liquefaction potential. Soil Dynam Earthq Eng 2004;24:689–98.
Soc Am 1953;43:17–34. [66] Yoon S, Rix GJ. Near-field effects on array-based surface wave methods with active
[54] Kennett BLN. Reflection, rays and reverberations. Bull Seismol Soc Am 1974;64: sources. J Geotech Geoenviron Eng 2009;135:399–406.
1685–96. [67] Roy N, Jakka RS. Near-field effects on site characterization using MASW technique.
[55] Kennett BLN, Kerry NJ. Seismic waves in a stratified half-space. Geophys J Int Soil Dynam Earthq Eng 2017;97:289–303.
1979;57:557–83. [68] Schröder CT, Scott Jr WR. On the complex conjugate roots of the Rayleigh
[56] Kennett BLN. Seismic wave propagation in stratified media. Australian National equation: the leaky surface wave. J Acoust Soc Am 2001;110:2867–77.
University E Press; 1983. [69] Komatitsch D, Vilotte JP. The spectral element method: an efficient tool to simulate
[57] Müller G. The reflectivity method: a tutorial. J Geophys 1985;58:153–74. the seismic response of 2D and 3D geological structures. Bull Seismol Soc Am 1998;
[58] Tan EL. Hybrid compliance-stiffness matrix method for stable analysis of elastic 88:368–92.
wave propagation in multilayered anisotropic media. J Acoust Soc Am 2006;119: [70] Komatitsch D, Tromp J. Introduction to the spectral element method for three-
45–53. dimensional seismic wave propagation. Geophys J Int 1999;139:806–22.
[59] Liu H, Pan E. Time-harmonic loading over transversely isotropic and layered elastic [71] Piessens R. The Hankel transform. In: Poularikas AD, editor. The transforms and
half spaces with imperfect interfaces. Soil Dynam Earthq Eng 2018;107:35–47. applications handbook. second ed. CRC Press LLC; 2000.
[60] Kreyszig E. Advanced engineering mathematics. tenth ed. John Wiley & Sons, Inc;
2011.

15