You are on page 1of 10

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/301198940

T313.full

Data · April 2016

CITATIONS READS

0 107

3 authors, including:

Hanming Chen Hui Zhou


China University of Petroleum-Beijing China University of Petroleum-Beijing
69 PUBLICATIONS   680 CITATIONS    153 PUBLICATIONS   1,322 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Full waveform inversion View project

time-frequency View project

All content following this page was uploaded by Hanming Chen on 11 April 2016.

The user has requested enhancement of the downloaded file.


GEOPHYSICS, VOL. 79, NO. 6 (NOVEMBER-DECEMBER 2014); P. T313–T321, 6 FIGS., 2 TABLES.
10.1190/GEO2014-0103.1

Application of unsplit convolutional perfectly matched layer


for scalar arbitrarily wide-angle wave equation

Hanming Chen1, Hui Zhou1, and Yanqi Li1

split PML (e.g., Chew and Weedon, 1994; Collino and Monk, 1998;
ABSTRACT Collino and Tsogka, 2001), nonsplitting PML (Wang and Tang,
2003), and unsplit convolutional PML (CPML) with a general com-
A classical split perfectly matched layer (PML) method plex frequency shifted (CFS) stretching operator (Kuzuoglu and
has recently been applied to the scalar arbitrarily wide-angle Mittra, 1996), referred to as CPML (Komatitsch and Martin,
wave equation (AWWE) in terms of displacement. However, 2007; Martin and Komatitsch, 2009) or CFS-CPML (e.g., Roden
the classical split PML obviously increases computational and Gedney, 2000; Drossaert and Giannopoulos, 2007a, 2007b).
cost and cannot efficiently absorb waves propagating into The unsplit CPML with a general CFS factor has been proven
the absorbing layer at grazing incidence. Our goal was to to be more effective in absorbing the evanescent wave mode
improve the computational efficiency of AWWE and to en- (e.g., Roden and Gedney, 2000; Bérenger, 2002a, 2002b; Drossaert
hance the suppression of edge reflections by applying a and Giannopoulos, 2007b) and the propagating wave mode at graz-
convolutional PML (CPML). We reformulated the original ing incidence (e.g., Komatitsch and Martin, 2007; Martin and Ko-
AWWE as a first-order formulation and incorporated the matitsch, 2009; Shi et al., 2012) than the PMLs with a special CFS
CPML with a general complex frequency shifted stretching factor used in the split PML and nonsplitting PML. The relations
operator into the renewed formulation. A staggered-grid fi- among different PML formulations are summarized by Kristek et al.
nite-difference (FD) method was adopted to discretize the (2009). Additionally, several PMLs have also been applied to the
first-order equation system. For wavefield depth continu- second-order wave equation (e.g., Komatitsch and Tromp, 2003;
ation, the first-order AWWE with the CPML saved memory Basu and Chopra, 2004; Yuan et al., 2014).
compared with the original second-order AWWE with the Despite wide usage of PML in the full (double-way) wave equa-
conventional split PML. With the help of numerical exam- tion as mentioned above, the applications of PML for one-way wave
ples, we verified the correctness of the staggered-grid FD equations (OWWEs) are rare and are limited to the frequency do-
method and concluded that the CPML can efficiently absorb main (e.g., Collino, 1997; Levy, 1997; Yevick and Thomson, 1999).
evanescent and propagating waves. OWWEs have been used in many areas such as seismic imaging
(Claerbout, 1976), unbounded domain modeling (e.g., Lindman,
1975; Guddati and Tassoulas, 2000; Thirunavukkarasu and Gud-
dati, 2011), underwater acoustics (e.g., Bamberger et al., 1988; Col-
INTRODUCTION lins, 1989), and nondestructive testing and evaluation (e.g., Chang
and Chir-Cherng, 2000).
Since its initial presentation (Bérenger, 1994), the perfectly In the context of seismic imaging, the development and applica-
matched layer (PML) method has been widely used as an absorbing tions of OWWEs have spanned several decades. The early OWWE
boundary condition (ABC) for modeling acoustic (e.g., Liu and in the wavenumber domain (Gazdag, 1978) is only applicable to
Tao, 1997; Qi and Geers, 1998; Guo et al., 2012) and elastic vertically layered media. To consider a laterally varying velocity
(e.g., Chew and Liu, 1996; Hastings et al., 1996; Collino and Monk, model, some corrections have been introduced, e.g., phase shift plus
1998; Collino and Tsogka, 2001; Festa and Vilotte, 2005) wave interpolation (PSPI) (Gazdag and Sguazzero, 1984), split-step Fou-
propagation in unbounded domains. Currently, various PML formu- rier method (SSF) (Stoffa et al., 1990), Fourier finite difference
lations have been published in the literature, mainly including the (FFD) (Ristow and Rühl, 1994), and an implicit scheme proposed

Manuscript received by the Editor 28 February 2014; revised manuscript received 2 July 2014; published online 10 October 2014.
1
China University of Petroleum, State Key Laboratory of Petroleum Resources and Prospecting, CNPC Key Lab of Geophysical Exploration, Beijing, China.
E-mail: huizhou@cup.edu.cn; huichanming@126.com; li_chelsea@yeah.net.
© 2014 Society of Exploration Geophysicists. All rights reserved.

T313
T314 Chen et al.

2 3
by Biondi (2002). The OWWEs represented in the space domain are 1
−1
more favored due to the ability to handle an arbitrarily varying 6 c11 1 c11 7
6−c c þ c − c12 7
velocity (e.g., Berkhout, 1979; Bamberger et al., 1988; Collins 6 1 1 2 7
6 .. .. 7
c6 7
and Westwood, 1991). Currently, OWWE imaging methods are still
6 − c12 . . 7
in use (e.g., Bakker, 2009; Jia and Wu, 2009; Zhang et al., 2010). Λ1 ¼ 6 7 ; (2)
The arbitrarily wide-angle wave equation (AWWE) proposed by 26 .. .. 7
6 . . 7
Guddati (2006) is an outstanding OWWE in terms of accuracy and 6 7
6 7
cn−2 þ cn−1 − cn−1
1 1 1
simplicity. AWWE is a highly accurate space domain OWWE and 4 5
hence is applicable to laterally inhomogeneous media. Besides, − cn−1 cn−1 þ c1n
1 1
n×n
AWWE is formulated as a second-order differential equation that
ensures a convenient numerical implementation. An efficient dou-
ble-marching FD scheme has been proposed for AWWE discretiza-
tion and calculation (Guddati and Heidari, 2005). Due to its 2 3
c1 c 1
advantages, AWWE has been successfully used for geophysical im- 6 c1 c1 þc2 c2 7
aging (e.g., Guddati and Heidari, 2005, 2007; He et al., 2008; Sun 6 7
6 .. .. 7
and Du, 2011; Zhou et al., 2012). However, the discretized AWWE 16 c2 . . 7
is actually not very computationally efficient due to its vector Λ2 ¼ 6 6 .. ..
7
7 ;
2c 6 7
formulation containing auxiliary variables. When the number of 6 . . 7
4 cn−2 þcn−1 cn−1 5
auxiliary variables increases, the computational cost increases
dramatically. cn−1 cn−1 þcn n×n
Chen et al. (2013) apply a split PML to the scalar AWWE to (3)
suppress artificial edge reflections. The split PML obviously further
increases the elapsed time and memory consumption. Meanwhile, it and
fails to fully absorb the outgoing waves with large incident angles.
In this paper, we slightly modify the original second-order AWWE d ¼ ð1;0; : : : ; 0Þ1×n ; and u ¼ ðu; u1 ; : : : ; un−1 ÞT : (4)
formulation. Application of the existing double-marching FD
scheme (Guddati and Heidari, 2005) to the revised second-order The superscript T denotes transposition, ρ is density, c is real veloc-
formulation leads to much higher computational efficiency than ity, and ci ði ¼ 1;2; : : : ; nÞ represents reference velocities. Equa-
the previous AWWE formulation. Based on the revised formulation, tion 1 is a vector formulation, and n is the number of elements
we further rewrite it as a first-order equation system and then apply in the displacement vector u. The first component of u (the scalar
an unsplit CPML for the boundary condition. The use of the CPML quantity u) is the main wavefield component, and the other com-
further decreases the overall computational cost. For clarity, we re- ponents are auxiliary variables. The quantities ρ, c, and ci are the
fer to the original AWWE (Guddati, 2006) as the second-order functions of x, and u is a function of x and t.
AWWE and the presented AWWE as the first-order AWWE The accuracy of AWWE can be increased greatly by increasing
throughout this paper. The staggered-grid FD method widely used the number of auxiliary variables and distributing the reference
in seismic modeling (e.g., Virieux, 1984, 1986; Levander, 1988; velocities properly. The tridiagonal matrices Λ1 and Λ2 , represented
Graves, 1996; Robertsson, 1996; Kristek and Moczo, 2003) is by equations 2 and 3, are actually constant matrices when setting the
adopted for numerical calculation. reference velocities as

ci ¼ λi c ðλi ≡ constantÞ: (5)


METHOD
Each reference velocity ci corresponds to a specific propagating an-
The first-order scalar arbitrarily wide-angle wave gle θi (cos θi ¼ cci ) measured from the vertical axis, at which the
equation dispersion curve of AWWE accurately matches that of the ex-
In a 3D Cartesian system with horizontal x-, y-, and z-axis act OWWE.
pointing downward, the scalar downward-propagating AWWE with Because the matrices Λ1 and Λ2 are constant matrices, equation 1
spatially varying velocity and density takes the form (Guddati, can be further simplified as
2006),     
∂2 u ∂2 u 1 ∂ 3 ∂u þ ∂ 3 ∂u
þcH 1 − H ρc ρc ¼ 0;
∂t2 ∂z∂t ρc 2 ∂x ∂x ∂y ∂y
∂2 u ∂2 u (6)
−ρc2 d − ρcðΛ1 þ Λ2 Þ 2
∂z∂t ∂t
     where
∂ ∂u ∂ ∂u
þ Λ2 ρc 3 þ ρc 3 ¼ 0; (1) 
∂x ∂x ∂y ∂y H1 ¼ LD; H2 ¼ LΛ2 ;
(7)
D ¼ dT d; L ¼ ðΛ1 þ Λ2 Þ−1 :

where By defining 1ρ^ ¼ ρc3 , we can write equation 6 as


Unsplit CPML for AWWE T315

     8 h  i
∂2 u ∂2 u ∂ 1 ∂u ∂ 1 ∂u > ψvx ;x ¼ − dk2x exp − dkxx þ αx t  ∂v
> ;
x
þ cH1 − ρ
^ c 2
H2 þ ¼ 0: >
> h   i ∂x
∂t2 ∂z∂t ∂x ρ^ ∂x ∂y ρ^ ∂y >
x
>
< ψv ;y ¼ − d2y exp − dy þ αy t  ∂vy ;
(8) y ky
h  ky  i ∂y (12)
>
> ψp;x ¼ − k2 exp − kx þ αx t  ∂p
dx dx
;
>
> h   i ∂x
Equation 8 is the revised second-order AWWE mentioned in the >
>
x
: ψ ¼ − dy exp − dy þ α t  ∂p :
introduction. The double-marching FD scheme developed by Gud- p;y k2y ky y ∂y
dati and Heidari (2005) for equation 1 can be applied to discretize
equation 8. The matrices defined in equation 7 are constant and The operator  stands for convolution that can be approximated by a
need to be calculated only once. Therefore, the revised second-order recursive technique (Wang and Tang, 2003). The parameters di , αi ,
AWWE has a much higher computational efficiency than the origi- and ki ði ¼ x; yÞ constitute a general CFS stretching factor, di is the
nal formulation 1, in which repeated matrix inversions and multi- damping profile in the PML layers, whereas αi and ki are expected
plications are required. to take effect in absorbing propagating waves with large incident
To write equation 8 as a first-order system, we redefine three vec- angles and evanescent waves. A special CFS operator correspond-
tors as ing to αx ðxÞ ¼ 0 and kx ðxÞ ¼ 1 underlies the existing split PML for
  the second-order AWWE (Chen et al., 2013). The staggered-grid
∂u 1 ∂u 1 ∂u FD method (Virieux, 1984) is adopted to discretize equation 11,
ðp; vx ; vy Þ ¼ ; ; . (9)
∂t ρ^ ∂x ρ^ ∂y and the details are presented in Appendix A.

Substitution of equation 9 into equation 8 yields the downward- Computational cost analysis
propagating first-order AWWE as
For simplicity, AWWE with n − 1 auxiliary variables is symbol-
8    ized as AWWEn . According to the analysis of Guddati and Heidari
>
> ∂
þ ∂
¼ ρ ∂vx
þ
∂vy
> ∂t
< cH 1 ∂z p ^ c 2
H 2 ∂x ∂y ;
(2005), the computational time of the second-order AWWE2,
∂vx 1 ∂p AWWE3 , AWWE4 , and AWWE5 without using the ABC is 2.3,
∂t ¼ ρ^ ∂x ;
(10)
>
> 4.89, 8.45, and 12.96 times that of the 15° wave equation (Claerb-
>
: ∂vy ¼ 1 ∂p :
∂t ρ^ ∂y out, 1976), respectively. Apparently, the computational time of the
second-order AWWE increases dramatically with increasing the
number of auxiliary variables. However, the problem eases remark-
We refer to equation 10 as the first-order quasi-pressure-velocity
ably for the revised second-order AWWE 8. Based on extensive
equation system because the vectors p and ðvx ; vy Þ are not really
numerical tests, it is concluded that the revised second-order
the pressure and velocities. The form of equation 10 is similar
AWWE2 , AWWE3 , AWWE4 , and AWWE5 without using the
to the first-order acoustic-wave equation (Virieux, 1984) except that
ABC take a computational time about 1.4, 2.0, 2.6, and 3.5 times
the physical quantities are vectors containing auxiliary variables.
that of the 15° wave equation. On the other hand, the required
Another difference is the absence of quasi-vertical-velocity compo-
memory storage of AWWE is almost equivalent to that of the
nent vz in the AWWE. Actually, the partial derivative of the
15° wave equation.
quasi-pressure p with respect to z is kept in the first formula of equa-
Next, we analyze the computational cost of the first-order
tion 10, indicating that it is an OWWE instead of a double-way
AWWE and the increased computational cost due to the usage
wave equation. The upward-propagating AWWE is obtained by
of the CPML. The computational cost of the revised second-order
replacing z with −z.
AWWE and the increased computational cost due to the usage of the
split PML are regarded as references correspondingly.
Convolutional perfectly matched layer for the first-or- For wavefield depth continuation using equations A-6, A-7, and
der arbitrarily wide-angle wave equation A-9, yet without considering the PML regions, the required memory
Because equation 10 represents downward-propagating waves, storage in the 3D case is
attenuation only along the x- and y-axes is required. We refer to
Komatitsch and Martin (2007) and Drossaert and Giannopoulos 3D ¼ ð3n þ 2N t þ 2ÞN x N y ;
S1st (13)
(2007a, 2007b) to derive a CPML for the first-order equa-
tions. The CPML formulation of the first-order AWWE 10 can and in the 2D case, it is
be written as
8 2D ¼ ð2n þ 2N t þ 1ÞN x ;
S1st (14)
  
>
> ∂
þ cH1 ∂z∂ p ¼ ρ^ c2 H2 ∂v
∂v
þ ∂yy
>
x
>
> ∂t
h ∂x
i where N x , N y , and N t are the numbers of grid points in the x-, y-,
>
>
< þ^ρc2 H2 1−kx ∂vx þ ψ v ;x þ 1−ky ∂vy þ ψ v ;y ; and t-axes, respectively, and n is the number of auxiliary variables.
kx ∂x ky ∂y
h x
i y
(11) After discretization, the revised second-order AWWE 8 requires
>
> ∂vx
¼ 1 ∂p
þ 1 1−kx ∂p
þ ψ ;
> ∂t
> ρ^ ∂x ρ^
h kx ∂x
p;x
i
memory storage of,
>
>
> ∂v
: ∂ty ¼ 1ρ^ ∂p 1−ky ∂p
∂y þ ρ^ k ∂y þ ψ p;y ;
1
y
3D ¼ ð3n þ 2N t ÞN x N y ;
S2nd (15)

where in the 3D case and


T316 Chen et al.

2D ¼ ð3n þ 2N t ÞN x ;
S2nd (16) in the 3D case, and

in the 2D case. ΔS2nd


2D ¼ 14 mðn þ 1Þ þ 2 mð3n þ 2N t Þ (20)
Therefore, in the 3D case without considering ABC, the first-or-
der AWWE requires a little more memory storage than the revised in the 2D case.
second-order AWWE, whereas in the 2D case, the first-order Because m ≥ 1 and n ≥ 1, it is easy to prove that ΔS2nd 3D > ΔS3D
1st

AWWE does not need more memory storage than the revised sec- and ΔS2nd2D > ΔS 1st . Therefore, taking the total computational do-
2D
ond-order AWWE for n ≥ 1. mains into account, the first-order AWWE with the unsplit CPML
When the CPML is incorporated into the first-order AWWE, the requires less memory storage than the revised second-order AWWE
increased memory is with the split PML. The total memory requirement of the first-order
AWWE with the unsplit CPML is OðN t N x N y Þ þ mO½N t ðN x þ
ΔS1st N y Þ in the 3D case and OðN t N x Þ þ mOðN t Þ in the 2D case.
3D ¼ 4 mðn þ 1ÞðN x þ N y þ 4 mÞ
The computational efficiency of the revised second-order AWWE
þ 2 mðN x þ N y þ 2 mÞð2N t þ 3n þ 2Þ (17) 8 and the first-order AWWE 10 is compared by running a 2D
numerical simulation using the Marmousi model (Bourgeois et al.,
in the 3D case, and 1991) on a dual-core E5700 CPU system. The number of grid points
are N x ¼ 1737, N z ¼ 750, and N t ¼ 5001. The wavefield gener-
ΔS1st
2D ¼ 4 mðn þ 1Þ þ 2 mð2N t þ 2n þ 1Þ (18) ated with a Ricker wavelet with a dominant frequency of 20 Hz
is extrapolated downward from the top of the model to its bottom.
in the 2D case, where m is the number of PML layers on each side No ABC is adopted in the simulation. The elapsed times of the re-
of the model. vised second-order AWWE and the first-order AWWE with differ-
When the split PML is incorporated into the revised second-order ent numbers of auxiliary variables are listed in columns 2 and 3 in
AWWE, the increased memory is Table 1. All the values are averaged results by repeating the com-
putations 20 times. Actually, the first-order AWWE takes a little
ΔS2nd
3D ¼ 4 mðn þ 1Þ½5ðN x þ N y Þ þ 11 m
more computing time than the revised second-order AWWE, but
the increased percentage is within 10%, as listed in column 4 in
þ 2 mðN x þ N y þ 2 mÞð2N t þ 3nÞ (19) Table 1.
The increased elapsed time due to the usage of ABC is also pre-
sented. We use a 20-layer split PML in the revised second-order
Table 1. The computational efficiency comparison between AWWE simulation and a 10-layer CPML in the first-order AWWE
the first-order AWWE and the revised second-order simulation. The reason for using more layers of PML for the sec-
AWWE. The comparison is based on a 2D benchmark
calculation with 737 × 750 discrete spatial grid points and ond-order AWWE is that even a 20-layer split PML cannot achieve
5001 temporal grid points. No ABC is used. Column 4 lists the same absorbing level as a 10-layer CPML for the first-order
the increased percentages of column 3 compared with AWWE. With these two different PMLs, we repeat the computa-
column 2. tions conducted for Table 1. Columns 2 and 3 in Table 2 list the
increased percentage of the elapsed time due to the applications
Elapsed time of Elapsed time of Increased of the split PML and the CPML, respectively. Column 4 in Table 2
second order (s) first order (s) percentage (%) shows the result of the subtraction of column 2 from column 3. It is
easy to find that the percentage listed in column 4 in Table 2 is larger
AWWE2 126.94 134.74 6.14 than that listed in column 4 in Table 1, suggesting an overall effi-
AWWE3 179.27 195.55 9.08 ciency gain of the first-order AWWE with the CPML as compared
AWWE4 236.66 253.78 7.23 with the revised second-order AWWE with the split PML.
In general, the first-order AWWE with the CPML saves computa-
AWWE5 315.50 323.05 2.39
tional memory and has a higher computational efficiency than the
second-order AWWE with the split PML.

Table 2. The increased percentages of elapsed time due to NUMERICAL EXAMPLES


the usage of PML. A 20-layer split PML is used for the
revised second-order AWWE, whereas a 10-layer CPML is We first conduct a poststack impulse migration test using the pre-
used for the first-order AWWE. Column 4 shows the result sented first-order AWWE and staggered-grid FD method. The
of the subtraction of column 2 from column 3.
medium has a constant velocity of 2000 m∕s and a constant density
of 1.5 g∕cm3 . The impulse response of the 15° wave equation is
Split PML for CPML for Decreased displayed as a reference in Figure 1a. Figure 1b–1e shows, respec-
second order (%) first order (%) percentage (%) tively, the responses of the first-order AWWE2, AWWE3 , AWWE4 ,
and AWWE5 . The exact semicircle is depicted by white asterisks. In
AWWE2 28.34 10.93 17.41 Figure 1b, the downward-propagating AWWE2 with two reference
AWWE3 25.84 13.48 12.36 velocities ðc; 4cÞ is used, and in Figure 1c, AWWE3 with three
AWWE4 28.92 12.55 16.37 reference velocities ðc; 2c; 4cÞ is adopted. In Figure 1d and 1e,
AWWE5 26.72 11.51 15.21 all the reference velocities applied in the AWWE propagators equal
the real velocity. It can be shown that the first-order AWWE with
Unsplit CPML for AWWE T317

different numbers of reference velocities can image a much steeper for the second-order AWWE (Chen et al., 2013). For a clear com-
dip angle than the 15° wave equation. Note that the Y-shaped wave- parison, the wavefield snapshots at the same time are displayed in
forms represent the evanescent waves, which are common in the the same amplitude range (amplified by a factor of 10). In the
OWWE numerical simulation (Graves and Clayton, 1990). CPML (case C) and the second-order split PML formulations (cases
The absorption effect of CPML is illustrated by Figure 2. The P1 and P2), the same damping function,
snapshots in a homogeneous medium with a constant velocity of
2500 m∕s and a constant density of 1.5 g∕cm3 are calculated with   3
3cmax 1 x
the rightward-propagating AWWE2 with two reference velocities dx ðxÞ ¼ log ; (21)
ðc; 4cÞ. A square mesh with a grid spacing of 4 m and a time step
2δ R δ
of 0.5 ms are used. A Ricker wavelet with a dominant frequency of
20 Hz excites rightward-propagating waves. The ABC is used on is used, where cmax is the maximum velocity of the medium, δ is the
the upper and lower edges. Four different settings of ABCs are used width of PML layers, and the reflection coefficient R is set as 10−6 .
for comparison, and they are, respectively, denoted as C, H, P1, and In the CPML case, the additional parameter kx equals 1. The param-
P2; C represents a 10-layer CPML for the first-order AWWE and H eter αx decays linearly from the innermost layer to the outermost
represents a three-layer, highly accurate ABC (HABC) proposed by and the maximum αmax ¼ πf o , where f o is the dominant frequency
Heidari and Guddati (2006) for the second-order AWWE. The P1 of the source and equals 20 Hz. The setting of kx and αx has been
and P2 stand for, respectively, a 10-layer and a 20-layer split PML previously adopted by Komatitsch and Martin (2007).

Figure 1. Poststack impulse responses of the first-


order AWWE with various numbers of auxiliary
variables. (a) The impulse response of the 15°
wave equation; (b-e) display, respectively, the re-
sponses of the first-order AWWE2, AWWE3 ,
AWWE4 , and AWWE5 . The exact semicircle is
plotted by the white asterisks.

Figure 2. Absorbing effectiveness comparison. Four different settings of ABCs, denoted by C, H, P2, and P1, are, respectively, applied; C
denotes a 10-layer CPML for the first-order AWWE, H denotes a three-layer HABC for the second-order AWWE, and P2 and P1 denote a 20-
layer and a 10-layer split PMLs for the second-order AWWE. (a-c) Display, respectively, the snapshots at 210, 500, and 725 ms. The white
arrows indicate edge reflections.
T318 Chen et al.

Figure 2a displays the snapshot at 210 ms, and the waves have boundaries. Some visible reflections of the propagating waves can
not yet reached the boundaries. Figure 2b displays the snapshot at be observed in case P1, as indicated by the white arrow, whereas no
500 ms, showing that the propagating waves have entered into the obvious reflections appear in the other cases. Figure 2c displays the
boundary layers, whereas the evanescent waves are just reaching the snapshot at 725 ms. The propagating waves nearly slide along the
upper and lower boundary layers, with the incident angle approach-
ing 90°. Under this situation, the CPML still fully absorbs the
propagating and evanescent waves, whereas in the other cases dif-
ferent levels of reflections linger. It is seen that the HABC absorbs
the propagating waves a little more than the 20-layer split PML, but
the latter absorbs the evanescent waves more efficiently (see the
white arrows for a comparison). The result agrees with previous
numerical observations (Chen et al., 2013). Unfortunately, in case
P1, the absorption effect is unacceptable.
To check the performance of the CPML in heterogeneous media
and the stability of the presented staggered-grid FD scheme, we
simulate the wavefields using the first-order downward-propagating
Figure 3. A 2D Marmousi velocity model. AWWE2 in the 2D Marmousi model (Bourgeois et al., 1991). Fig-
ure 3 displays the 2D Marmousi velocity model with a relatively
fine grid of ðΔx; ΔzÞ ¼ ð6 m; 3 mÞ. A time step
of 1 ms is used. The model has a constant density
of 1.5 g∕cm3 .
Figure 4a displays the snapshot at 250 ms,
showing that the waves are reaching the left
boundary. Figure 4b–4d show the snapshots at
900 ms and corresponds to a 10-layer split
PML, a three-layer HABC, and a 10-layer CPML
applied on the left edge, respectively. The snap-
shots are displayed in the same amplitude range
with a magnification factor of 20. The white ar-
rows point out different levels of edge reflections
due to incomplete absorption. Obviously, the 10-
layer CPML best absorbs evanescent waves
and propagating waves. To check the long-term
stability of the staggered-grid FD method and the
CPML, the calculation is conducted for 100,000
time steps. No unreasonable increase of the
wavefield energy is observed in the calculation
process all the time.
A 3D model is also used to simulate the wave
Figure 4. Snapshots calculated by the first-order downward-propagating AWWE2 in the propagation. The 3D inhomogeneous velocity
2D Marmousi model. (a) Snapshot at 250 ms; (b-d) snapshots at 900 ms, respectively, model (Figure 5) is generated by truncating
corresponding to the usage of a 10-layer split PML, a three-layer HABC, and (d) a 10-
layer CPML. The white arrows represent artificial edge reflections. the 2D Marmousi velocity model (Figure 3)
and then stretching it along the x-axis. The
density is obtained by using an empirical
formula (Gardner et al., 1974). A grid of ðΔx; Δy; ΔzÞ ¼
ð8 m; 6 m; 4 mÞ and a time step of 0.5 ms are used for calculation.
The downward-propagating waves are simulated by the first-order
AWWE2 with two reference velocities ðc; 4cÞ. The excitation
source is located at the center of the upper surface. Figure 6a dis-
plays the snapshot at 300 ms, and the waves have not arrived at the
boundaries. Figure 6b and 6c represents the snapshots at 750 ms,
and they are displayed in the same amplitude range. In Figure 6b,
no ABC is applied; as a result, the outgoing waves are reflected
back to the interior computational domain (marked by the white
arrows). In Figure 6c, a 10-layer CPML is applied, and the edge
reflections seen in Figure 6b are suppressed very well. Note that
in Figure 6c the weak reflections on the upper surface are caused
by the laterally varying velocity and density, not by the artifi-
Figure 5. A 3D inhomogeneous velocity model. cial PML.
Unsplit CPML for AWWE T319

Figure 6. Three-dimensional snapshots propagated by the first-order downward-propagating AWWE2 . (a) Snapshot at 300 ms; (b and c)
snapshots at 750 ms, respectively, corresponding to no ABC and a 10-layer CPML applied. The white arrows indicate edge reflections.

CONCLUSIONS where t is the index for time discretization; i, j, and k are, respec-
tively, indexes for x-, y-, and z-axis discretizations; Δt is the time
The scalar AWWE is reformulated as the first-order quasi- step; and ðΔx; Δy; ΔzÞ is the grid spacing. Using the operators to
velocity-pressure equation system just like the acoustic and elastic discretize equation 11 and adopting the Crank-Nicolson scheme for
wave equations. The first-order AWWE with the CPML boundary marching in depth, we have
condition is an appealing substitution of the existing second-order
AWWE with the split PML. The widely used staggered-grid FD 8
method is applied to discretize the first-order AWWE. An empiri- >
>
tþ1 tþ1
p 2 1 −pi;j;k2
>
> þ
> Dt pi;j;kþ12 þ 2ci;j;kþ12 H1 ¼
i;j;kþ
2
cal restrictive condition can ensure numerical stability. The advan- >
>  Δz 
tages of CPML, saving memory storage and effectively absorbing >
< tþ 1 1
ρ^ i;j;kþ12 c2i;j;kþ1 H2 Dhx vx;j;kþ h tþ2
1 þ Dy v þ G1 ;
2
1
evanescent waves and propagating waves with large incident an- y;i;kþ2
>
2 2
> Dt v þ
1 ¼ bx;iþ1;j;kþ1 Dx p þ G2 ;
h t
gles, are demonstrated again in the AWWE case. The required >
> x; i þ 2 ; j; k þ 2
1
j; k þ 12
>
>
2 2
memory storage of the first-order AWWE with the CPML is >
>
actually less than the second-order AWWE with the split PML. : Dht vy; i; j þ 1 ; k þ 1 ¼ by;i;jþ12;kþ12 Dþ yp
t
þ G3 :
2 2 i; k þ 12
Regarding the computational efficiency, the presented first-order
AWWE with the CPML reduces the overall elapsed time. There- (A-2)
fore, the first-order AWWE with the CPML is preferable for
2 ðp þ pt Þ, equation A-2 is further
1
seismic imaging to the existing second-order AWWE with the Approximating ptþ2 by 1 tþ1
split PML. written as

8
ACKNOWLEDGMENTS >
> ptþ1 1 −ptþ1 þpt −pti;j;k
>
> Dþ pi;j;kþ12 þ ci;j;kþ12 H1
i;j;kþ i;j;k i;j;kþ1
¼
>
2 2
>
>
t
 Δz 
This work is supported in part by the 973 Program of China >
< tþ12 tþ12
(2013CB228603), National Science and Technology Program ρ^ i;j;kþ12 ci;j;kþ1 H2 Dx vx;j;kþ1 þ Dy vy;i;kþ1 þ G1 ;
2 h h
2 2 2
(2011ZX05010-001 and 2011ZX05024-001), National Natural Sci- > Dht v
> 1 ¼ bx;iþ1;j;kþ1 Dx p
þ t þ G2 ;
>
> x; i þ 1
; j; k þ j; k þ 12
>
2 2
ence Foundation of China (41174119), and the Research of Novel 2 2
>
> þ t
Method and Technology of Geophysical Prospecting, CNPC >
: Dt vy; i; j þ 1 ; k þ 1 ¼ bz;i;jþ12;kþ12 Dy p i; k þ 1 þ G3 ;
ho
2 2 2
(2011A-3602). H. Chen is also grateful for Shell’s scholarship.
(A-3)

APPENDIX A
where
STAGGERED-GRID FINITE-DIFFERENCE
DISCRETIZATION 2 1
  3
1−kx h tþ2 tþ1
Dx vx;j;kþ1 þψ vx ;x2
i;j;kþ 12
6 k x 7
G1 ¼ ð^ρc2 Þi;j;kþ12 H2 4   5;
2
Several difference operators are defined as
1−ky h tþ12 tþ12
þ ky Dy vy;i;kþ1 þψ vy ;y i;j;kþ 2 1
f tþ1 − f t f iþ1 − f i 
2
 

t f ¼ ; Dþ
xf ¼ ; 1−kx þ t 1 1
Δt Δx G2 ¼ bx;iþ12;j;kþ12 Dx p þψ t iþ ;j;kþ ;
p;x
f jþ1 − f j f kþ1 − f k kx j;kþ 12 2 2

yf ¼ ; Dþ
z f ¼ ;   
Δy Δz 1−ky þ t 1 1
G3 ¼ by;i;jþ12;kþ12 Dy p þψ p;y i;jþ ;kþ
t
;
f tþ2 − f t−2
1 1
f iþ12 − f i−12 ky i;kþ 12 2 2
Dht f ¼ ; Dhx f ¼ ; and
Δt Δx (A-4)
f jþ12 − f j−12
Dhy f ¼ ; (A-1)
Δy and
T320 Chen et al.

2 2 partial derivatives with respect to x, y, and t is possible, but it is


bx;iþ12 ¼ ; by;jþ12 ¼ ;
ρ^ iþ1 þ ρ^ i ρ^ jþ1 þ ρ^ j not discussed here.
ρ^ kþ1 þ ρ^ k ckþ1 þ ck
ρ^ kþ12 ¼ ; and ckþ12 ¼ : (A-5)
2 2 REFERENCES
Bakker, P. M., 2009, A stable one-way wave propagator for VTI media:
Geophysics, 74, no. 5, WB3–WB17, doi: 10.1190/1.3196818.
After some arithmetic operations, equation A-3 can be further Bamberger, A., B. Engquist, L. Halpern, and P. Joly, 1988, Higher-order
simplified as paraxial wave-equation approximations in heterogeneous media: SIAM
Journal on Applied Mathematics, 48, 129–154, doi: 10.1137/0148006.
Basu, U., and A. K. Chopra, 2004, Perfectly matched layers for transient
ptþ1
i;j;kþ1
þ αz ptþ1
i;j;kþ1
s ¼ F þ ΔtG1 ; elastodynamics of unbounded domains: International Journal for Numeri-
2 2 cal Methods in Engineering, 59, 1039–1074, doi: 10.1002/nme.896.
tþ12 t−12
Bérenger, J. P., 1994, A perfectly matched layer for the absorption of electro-
v ¼v magnetic waves: Journal of Computational Physics, 114, 185–200, doi:
x; i þ 1
2 ; j; k þ þ 12 ; j; k þ 12
1
2 x; i 10.1006/jcph.1994.1159.
  Bérenger, J. P., 2002a, Application of the CFS PML to the absorption of
Δt evanescent waves in waveguides: IEEE Microwave and Wireless Compo-
þ bx; i þ 1 ; j; k þ 1 p t −pt
nents Letters, 12, 218–220, doi: 10.1109/LMWC.2002.1010000.
2 2 Δx i þ 1; j; k þ 12 i; j; k þ 12 Bérenger, J. P., 2002b, Numerical reflection from FDTD-PMLs: A compari-
son of the split PML with the unsplit and CFS PMLs: IEEE Transactions
þ ΔtG2 ; and on Antennas and Propagation, 50, 258–265, doi: 10.1109/8.999615.
Berkhout, A. J., 1979, Steep dip finite-difference migration: Geophysical
tþ12 t−12
v ¼v Prospecting, 27, 196–213, doi: 10.1111/j.1365-2478.1979.tb00965.x.
y ; i; j þ 1
2;k þ 1
2þ 12 ; k þ 12
y; i; j Biondi, B., 2002, Stable wide-angle Fourier finite-difference downward
  extrapolation of 3-D wavefields: Geophysics, 67, 872–882, doi: 10
Δt .1190/1.1484530.
þ by; i; j þ 1 ; k þ 1 p t
−pt
Bourgeois, A., M. Bourget, P. Lailly, M. Poulet, P. Ricarte, and R. Versteeg,
2 2 Δy i; j þ 1; k þ 12 i; j; k þ 12 1991, Marmousi model and data, in R. Versteeg, and G. Grau, eds., the
Marmousi experience, proceedings of the 1990 EAGE workshop on prac-
þ ΔtG3 ; (A-6) tical aspects of seismic data inversion, EAGE, Zeist, 5–16.
Chang, Y.-F., and C. Chir-Cherng, 2000, Frequency-wavenumber migration
of ultrasonic data: Journal of Nondestructive Evaluation, 19, 1–10, doi: 10
where .1023/A:1006671706818.
Chen, H., H. Zhou, H. Lin, and S. Wang, 2013, Application of perfectly
  matched layer for scalar arbitrarily wide-angle wave equations: Geophys-
ics, 78, no. 1, T29–T39, doi: 10.1190/geo2012-0062.1.
F ¼ pti;j;kþ1 þαz ptþ1 i;j;k þp t
i;j;k −p t
1
i;j;kþ2
s Chew, W. C., and Q. Liu, 1996, Perfectly matched layers for elastodynamics:
2
A new absorbing boundary condition: Journal of Computational Acous-
   tics, 4, 341–359, doi: 10.1142/S0218396X96000118.
tþ1 tþ12
þ ρ^ i;j;kþ12 ci;j;kþ12 H2 αx v 2 1 1 −v 1 1
Chew, W. C., and W. H. Weedon, 1994, A 3D perfectly matched medium
x; iþ 2 ;j;kþ 2 x; i− 2 ;j;kþ 2 from modified Maxwell’s equations with stretched coordinates: Micro-
  wave and Optical Technology Letters, 7, 599–604, doi: 10.1002/mop
tþ1 tþ12 .4650071304.
þαy v 2 1 1 −v 1 1 ; Claerbout, J. F., 1976, Fundamentals of geophysical data processing: Black-
y; i;jþ 2 ;kþ 2 y ;i;j− 2 ;kþ 2 well Scientific Publications.
Collino, F., 1997, Perfectly matched absorbing layers for the paraxial equa-
(A-7) tions: Journal of Computational Physics, 131, 164–180, doi: 10.1006/jcph
.1996.5594.
Collino, F., and P. Monk, 1998, Optimizing the perfectly matched layer:
and the vector s is the first column of the matrix L defined in equa- Computer Methods in Applied Mechanics and Engineering, 164, 157–
tion 7. The other variables are defined as 171, doi: 10.1016/S0045-7825(98)00052-8.
Collino, F., and C. Tsogka, 2001, Application of the PML absorbing layer
model to the linear elastodynamic problem in anisotropic heterogeneous
cΔt cΔt cΔt media: Geophysics, 66, 294–307, doi: 10.1190/1.1444908.
αx ¼ ; αy ¼ ; and αz ¼ : (A-8) Collins, M. D., 1989, Applications and time-domain solution of higher-order
Δx Δy Δz parabolic equations in underwater acoustics: Journal of the Acoustical So-
ciety of America, 86, 1097–1102, doi: 10.1121/1.398101.
Collins, M. D., and E. K. Westwood, 1991, A higher-order energy-conserving
parabolic equation for range-dependent ocean depth, sound speed, and den-
The staggered scheme is only applied in the t-, x-, and y-axes, not sity: Journal of the Acoustical Society of America, 89, 1068–1175, doi: 10
in the z-axis in equation A-6. Because the Crank-Nicolson method .1121/1.400526.
Drossaert, F. H., and A. Giannopoulos, 2007a, A nonsplit complex fre-
is used for marching in depth, the values of the quasi pressure vector quency-shifted PML based on recursive integration for FDTD modeling
pkþ1 at zkþ1 can be calculated by of elastic waves: Geophysics, 72, no. 2, T9–T17, doi: 10.1190/1.2424888.
Drossaert, F. H., and A. Giannopoulos, 2007b, Complex frequency shifted
convolution PML for FDTD modelling of elastic waves: Wave Motion,
pkþ1 ¼ 2pkþ12 − pk : (A-9) 44, 593–604, doi: 10.1016/j.wavemoti.2007.03.003.
Festa, G., and J. P. Vilotte, 2005, The Newmark scheme as velocity-stress
time-staggering: An efficient PML implementation for spectral-element
simulations of elastodynamics: Geophysical Journal International, 161,
789–812, doi: 10.1111/j.1365-246X.2005.02601.x.
Equations A-6, A-7, and A-9 can be used for wavefield depth Gardner, G. H. F., L. W. Gardner, and A. R. Gregory, 1974, Formulation
extrapolation or time extrapolation. Eliminating the terms G1 , velocity and density — The diagnostic basics for stratigraphic traps: Geo-
physics, 39, 770–780, doi: 10.1190/1.1440465.
G2 , and G3 , equation A-6 degrades to the case without CPML. Gazdag, J., 1978, Wave equation migration with the phase-shift method:
ffi empirical restrictive condition cΔt ≤ Δx for the 2D case and
pffiffiAn Geophysics, 43, 1342–1351, doi: 10.1190/1.1440899.
2cΔt ≤ minðΔx; ΔyÞ for the 3D case can ensure the numerical Gazdag, J., and P. Sguazzero, 1984, Migration of seismic data by phase-shift
plus interpolation: Geophysics, 49, 124–131, doi: 10.1190/1.1441643.
stability. Applying high-order FD operators to approximate the
Unsplit CPML for AWWE T321

Graves, R. W., 1996, Simulating seismic wave propagation in 3D elastic Lindman, E. L., 1975, “Free Space” boundary conditions for the time de-
media using staggered-grid finite differences: Bulletin of Seismological pendent wave equation: Journal of Computational Physics, 18, 66–78,
Society of America, 86, 1091–1106. doi: 10.1016/0021-9991(75)90102-3.
Graves, R. W., and R. W. Clayton, 1990, Modeling acoustic waves with Liu, Q., and J. Tao, 1997, The perfectly matched layer for acoustic waves in
paraxial extrapolators: Geophysics, 55, 306–319, doi: 10.1190/1 absorptive media: Journal of the Acoustical Society of America, 102,
.1442838. 2072–2082, doi: 10.1121/1.419657.
Guddati, M. N., 2006, Arbitrarily wide-angle wave equations for complex Martin, R., and D. Komatitsch, 2009, An unsplit convolutional perfectly
media: Computer Methods in Applied Mechanics and Engineering, 195, matched layer technique improved at grazing incidence for the viscoelas-
65–93, doi: 10.1016/j.cma.2005.01.006. tic wave equation: Geophysical Journal International, 179, 333–344, doi:
Guddati, M. N., and A. H. Heidari, 2005, Migration with arbitrarily wide- 10.1111/j.1365-246X.2009.04278.x.
angle wave equations: Geophysics, 70, no. 3, S61–S70, doi: 10.1190/1 Qi, Q., and T. L. Geers, 1998, Evaluation of the perfectly matched layer for
.1925747. computational acoustics: Journal of Computational Physics, 139, 166–
Guddati, M. N., and A. H. Heidari, 2007, Subsurface imaging via fully 183, doi: 10.1006/jcph.1997.5868.
coupled elastic wavefield extrapolation: Inverse Problems, 23, 73–98, Ristow, D., and T. Rühl, 1994, Fourier finite-difference migration: Geophys-
doi: 10.1088/0266-5611/23/1/004. ics, 59, 1882–1893, doi: 10.1190/1.1443575.
Guddati, M. N., and J. L. Tassoulas, 2000, Continued-fraction absorbing Robertsson, J. O. A., 1996, A numerical free-surface condition for elastic/
boundary conditions for the wave equation: Journal of Computational viscoelastic finite-difference modeling in the presence of topography:
Acoustics, 8, 139–156, doi: 10.1142/S0218396X00000091. Geophysics, 61, 1921–1934, doi: 10.1190/1.1444107.
Guo, H., S. Wang, N. Guo, and W. Chen, 2012, Wave equation simulation by Roden, J. A., and S. D. Gedney, 2000, Convolution PML (CPML): An
finite-element method with perfectly matched layer: Advanced Materials efficient FDTD implementation of the CFS-PML for arbitrary media: Mi-
Research, 524, 96–100, doi: 10.4028/www.scientific.net/AMR.524- crowave and Optical Technology Letters, 27, 334–339, doi: 10.1002/
527.96. 1098-2760(20001205)27:5<334::AID-MOP14>3.0.CO;2-A.
Hastings, F. D., J. B. Schneider, and S. L. Broschat, 1996, Application of the Shi, R., S. Wang, and J. Zhao, 2012, An unsplit complex-frequency-shifted
perfectly matched layer (PML) absorbing boundary condition to elastic PML based on matched Z-transform for FDTD modelling of seismic wave
wave propagation: Journal of the Acoustical Society of America, 100, equations: Journal of Geophysical Engineering, 9, 218–229.
3061–3069, doi: 10.1121/1.417118. Stoffa, P. L., J. T. Fokkema, R. M. de Luna Freire, and W. P. Kessinger, 1990,
He, B., H. Zhang, and J. Zhang, 2008, Prestack reverse-time depth migration Split-step Fourier migration: Geophysics, 55, 410–421, doi: 10.1190/1
of arbitrarily wide-angle wave equations: Acta Seismologica Sinica, 21, .1442850.
492–501, doi: 10.1007/s11589-008-0492-x. Sun, Q., and Q. Du, 2011, Study on depth migration with arbitrarily wide-
Heidari, A. H., and M. N. Guddati, 2006, Highly accurate absorbing boun- angle wave equation: Presented at 3rd IEEE International Conference on
dary conditions for wide-angle wave equations: Geophysics, 71, no. 3, Communication Software and Networks, 230–233.
S85–S97, doi: 10.1190/1.2192914. Thirunavukkarasu, T., and M. N. Guddati, 2011, Absorbing boundary
Jia, X., and R. Wu, 2009, Super wide-angle one-way wave propagator and its conditions for time harmonic wave propagation in discretized domains:
application in imaging steep salt flanks: Geophysics, 74, no. 4, S75–S83, Computer Methods in Applied Mechanics and Engineering, 200,
doi: 10.1190/1.3124686. 2483–2497, doi: 10.1016/j.cma.2011.04.021.
Komatitsch, D., and R. Martin, 2007, An unsplit convolutional perfectly Virieux, J., 1984, SH-wave propagation in heterogeneous media: Velocity-
matched layer improved at grazing incidence for the seismic wave equa- stress finite-difference method: Geophysics, 49, 1933–1942, doi: 10
tion: Geophysics, 72, no. 5, SM155–SM167, doi: 10.1190/1.2757586. .1190/1.1441605.
Komatitsch, D., and J. Tromp, 2003, A perfectly matched layer absorbing Virieux, J., 1986, P-SV wave propagation in heterogeneous media: Velocity-
boundary condition for the second-order seismic wave equation: Geo- stress finite-difference method: Geophysics, 51, 889–901, doi: 10.1190/1
physical Journal International, 154, 146–153, doi: 10.1046/j.1365- .1442147.
246X.2003.01950.x. Wang, T., and X. Tang, 2003, Finite-difference modeling of elastic wave
Kristek, J., and P. Moczo, 2003, Seismic-wave propagation in viscoelastic propagation: A nonsplitting perfectly matched layer approach: Geophys-
media with material discontinuities: A 3D fourth-order staggered-grid fi- ics, 68, 1749–1755, doi: 10.1190/1.1620648.
nite-difference modeling: Bulletin of the Seismological Society of Yevick, D., and D. J. Thomson, 1999, Nonlocal boundary conditions for
America, 93, 2273–2280, doi: 10.1785/0120030023. finite-difference parabolic equation solvers: Journal of the Acoustical So-
Kristek, J., P. Moczo, and M. Galis, 2009, A brief summary of some PML ciety of America, 106, 143–150, doi: 10.1121/1.427043.
formulations and discretizations for the velocity-stress equation of seismic Yuan, S., S. Wang, W. Sun, L. Miao, and Z. Li, 2014, Perfectly matched
motion: Studia Geophysica et Geodaetica, 53, 459–474, doi: 10.1007/ layer on curvilinear grid for the second-order seismic acoustic wave equa-
s11200-009-0034-6. tion: Exploration Geophysics, 45, 94–104.
Kuzuoglu, M., and R. Mittra, 1996, Frequency dependence of the constitutive Zhang, J., W. Wang, S. Wang, and Z. Yao, 2010, Optimized Chebyshev Fou-
parameters of causal perfectly matched anisotropic absorbers: IEEE Micro- rier migration: A wide-angle dual-domain method for media with strong
wave and Guided Wave Letters, 6, 447–449, doi: 10.1109/75.544545. velocity contrasts: Geophysics, 75, no. 2, S23–S34, doi: 10.1190/1
Levander, A. R., 1988, Fourth-order finite-difference P-SV seismograms: .3350861.
Geophysics, 53, 1425–1436, doi: 10.1190/1.1442422. Zhou, H., H. Lin, S. Sheng, H. Chen, and Y. Wang, 2012, High angle pre-
Levy, M. F., 1997, Transparent boundary conditions for parabolic equation stack depth migration with absorption compensation: Applied Geophys-
solution of radio wave propagation problems: IEEE Transactions on An- ics, 9, 293–300, doi: 10.1007/s11770-012-0339-z.
tennas and Propagation, 45, 66–72, doi: 10.1109/8.554242.

View publication stats

You might also like