You are on page 1of 8

Downloaded 08/01/16 to 14.139.116.34. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.

org/

GEOPHYSICS, VOL. 81, NO. 4 (JULY-AUGUST 2016); P. T155T162, 9 FIGS., 2 TABLES.


10.1190/GEO2015-0550.1

A 2.5D finite-difference algorithm for elastic wave


modeling using near-optimal quadratures

Maokun Li1, Vladimir Druskin2, Aria Abubakar2, and Tarek M. Habashy2

use a 2D model to approximate the earth while maintaining the 3D


acquisition layout. This defines the settings of the 2.5D modeling
schemes. Due to its good balance between modeling accuracy and
computational complexity, 2.5D schemes have been applied to
process seismic data sets by many researchers, such as Igel et al.
(1996), Charara et al. (1996), Cao et al. (2008), Liang et al.
(2013), and Owusu et al. (2016).
Modeling seismic wave propagation using 2.5D models can be
more efficient than solving the same problem using 3D models. Using the properties of translational symmetry along the invariant direction, we can apply Fourier or Laplace transform on this axis and
convert a 3D problem into a set of independent 2D problems. The
3D elastic fields can then be computed from these 2D solutions
through an inverse transform. In the modeling algorithm, these
transforms are implemented using numerical integrations. Because
the 2D problems contain less unknowns than the original 3D ones
and can be solved in parallel, 2.5D modeling algorithms can be very
efficient in simulating inline acquisitions, such as crosswell seismic
surveys, survey lines in land and marine acquisition, etc. Many algorithms have been developed for 2.5D acoustic and elastic wave
modeling (see Williamson and Pratt, 1995; Narayan, 2001; Cao
et al., 2008; Zhou et al., 2012; Doyon and Giroux, 2014).
The computational complexity of the 2.5D modeling algorithms
depends on the computational complexity of the 2D modeling algorithm and the total number of quadrature points for numerical
integration in the spectral domain. Many researchers have investigated this problem and developed quadrature rules (see Zhou et al.,
2012) and truncation criteria for these numerical integrations (see
Doyon and Giroux, 2014). In these works, the integration points are
usually located along the real axis in the spectral domain of the
invariant axis (usually represented as ky ). The number of sampling
points is proportional to the wavenumber in the medium and increases with frequency. We need many quadrature points to achieve
accurate results. In the worst scenario, the computational cost of a
2.5D problem can be comparable with a full 3D one. This bottleneck limits the applicability of 2.5D modeling algorithms.

ABSTRACT
We have developed a 2.5D finite-difference algorithm to
model the elastic wave propagation in heterogeneous media.
In 2.5D problems, it is assumed that the elastic properties of
models are invariant along a certain direction. Therefore, we
can convert the 3D problem into a set of 2D problems in the
spectral domain. The 3D solutions are then obtained by applying a numerical integration in the spectral domain. Usually, the quadrature points used in the numerical integration
scheme are sampled from the real axis in the spectral domain. The convergence of this quadrature can be very slow
especially at high frequencies. We have applied the optimal
quadrature scheme for the spectral integration. This is equivalent to transforming the contour of integration from the real
axis into a path in the complex plane. Our numerical studies
have indicated more than 10 times of reduction in the number of quadrature points compared with sampling along the
real axis in the spectral domain. This scheme alleviates the
bottleneck of computing speed in the 2.5D elastic wave
modeling. Furthermore, it can improve the computational
efficiency of 2.5D elastic full-waveform inversion algorithms.

INTRODUCTION
Numerical simulation of elastic wave propagation in heterogeneous medium provides a fundamental tool in seismic data
processing. Although real data are acquired in the 3D world, we
can still apply some approximation to simplify the model and to
accelerate the computation. For example, when the data are acquired along a straight line, they are mainly sensitive to the formation in the vertical plane of the acquisition line. Therefore, we can

Manuscript received by the Editor 25 October 2015; revised manuscript received 12 January 2016; published online 31 May 2016.
1
Tsinghua University, Beijing, China. E-mail: maokunli@tsinghua.edu.cn.
2
Schlumberger, Cambridge, Massachusetts, USA. E-mail: druskin1@slb.com; aria.abubakar@gmail.com; habashy1@slb.com.
2016 Society of Exploration Geophysicists. All rights reserved.
T155

Li et al.

Downloaded 08/01/16 to 14.139.116.34. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

T156

It is known that modifying the integration path in the complex


plane can change the convergence rate of the integral (see Keener,
2000). In this paper, we extend the frequency-domain 2.5D finitedifference modeling algorithm (see Xiong et al. [2013] for the algorithm and Pan and Abubakar [2013] for the absorbing boundary
conditions) with a near-optimal quadrature for spectral domain integration. This quadrature is derived from an optimized integration
path in the complex plane (see Druskin et al., 2016). The location of
the quadrature points and their weights are numerically optimized to
achieve optimal convergence. Using this scheme, the number of
quadrature points is greatly reduced. Numerical results showed that
the number of sampling points reduces by more than one order of
magnitude. Moreover, it is also observed that the quadrature is stable for high-frequency operations. Therefore, this new quadrature
can improve the computational efficiency by more than 10 times,
especially for higher frequencies. In this paper, we will discuss the
derivation of the near-optimal quadrature for the 2.5D elastic modeling algorithm and validate the performance of this algorithm using
numerical examples.

FORMULATION
2.5D finite-difference modeling of elastic wave
propagation



2
ux uz
2

y
y
y x
z
y2




uy
uy

x
z
x
z





ux

uz

fy ;
x
z
y
y

 

2

ux uy uz
2

z
z
z
x
y
z
y2







uz

uz

ux

x
z
x
x
z
z


2

u
uy
z fz :

z
yz
z

2
u T u f;
y2

(1)

The details of this equation can be found in Aki and Richards


(1980). Here, r x; y; zT represents the location vector in 3D
and ur ux ; uy ; uz T represents the displacement field at r induced by excitation fr; rS f x ; f y ; f z T at rS . The function
represents the angular frequency, and represent the Lam parameters, and represents the density.
Here, we consider a 2.5D elastic model in which the material parameters , , and are invariant along one direction. Then, we have

(2)

0;
y

(3)

0:
y

(4)

31

6
7
7
T6
4 2 5
2
6
6
6
4


2
x y yx



x z z x

yx
x y

2 x x z z

yz
z y



z x x z


2
z y yz

2 z 2 z x x

2 x 2 x z z
2

7
7
7;
5

(9)
and f is the force vector. The Fourier transform along the y-direction
is given by

1
ux; y; z

~ ky ; zeiky y dky
ux;

(10)

ux; y; zeiky y dy:

(11)

and

Applying the above equations, we can write equation 1 explicitly


in Cartesian coordinates as

 

2

ux uy uz
2

2 ux ux

x
x
y
z
y







ux

ux

ux

x
z
x
x
z
x


2

u
u
z fx ;

yx y z
x

(8)

where T is a tensor given by

r ur rfur urT g

0;
y

(7)

In these equations, we leave the second-order derivatives of u


with respect to y on one side of the equation for future derivations. Considering u as a vector in Cartesian coordinate, i.e.,
u ux ; uy ; uz T , we can write the above equations as follows:

The 3D isotropic elastic wave equation in frequency domain can


be written as

2 rur fr; rS :

(6)

~ ky ; z
ux;

1
2

Equation 8 can also be transformed into Fourier domain as

~ ky ; z;
~ y ux;
~ ky ; z fx;
~ ky ; z Tk
k2y ux;

(5)

(12)

~ ky ; z are the Fourier transforms of T and f in


~ y and fx;
where Tk
the ky domain, respectively. We can observe that the derivatives related to y are replaced by ky. The above equation becomes a 2D
~ ky ; z. After ux;
~ ky ; z is solved for every ky,
equation of ux;
we can apply equation 11 to compute the displacement field
ux; y; z in spatial domain.

Downloaded 08/01/16 to 14.139.116.34. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

2.5D elastic wave modeling


In this work, we applied the second-order finite-difference
scheme with a Cartesian Yee grid (for details, see Yee, 1966) to
discretize the 2D solution domain. Figure 1 shows the definition
of the displacement field on the Cartesian grid. We also applied
the perfectly matched layer as boundary condition (see Berenger
[1994] and Pan and Abubakar [2013] for implementation details).
The discrete field values on the grid can be computed by solving a
sparse matrix equation using multifrontal LU decomposition (see
~ ky ; z is solved, we can compute
Davis and Duff, 1997). Once ux;
the ux; y; z using equation 10.

Application of quasi-optimal quadrature to spectral


domain integration

T157

i
2

ux; y; z

At y 0, we have

ux; 0; z

i
2

N
1X
~ ky;i ; zeiky;i y ky;i ;
a ux;
i1 i

(13)

where ky;i is the length of the ith interval and ai is the weighting
~ ky;i ; z is computed by solvcoefficient. In this equation, every ux;
ing a 2D problem as in equation 12. Therefore, the computational
complexity of the 2.5D problem is significantly affected by the
number of quadrature points N in equation 13. Many numerical
quadratures have been developed, such as the work by Zhou et al.
(2012) and Xiong et al. (2013). These studies usually follow the
definition of Fourier transform and the numerical quadrature points
are sampled along the real axis of ky . Because of the oscillatory
nature of eiky y , ky needs very dense samples. Moreover, the domain
of integration ky;max ; ky;max is proportional to the wavenumber k.
Therefore, the total number of integration points increases with frequency. Usually the number of samples for accurate computation
(relative error <0.01) is around hundreds to even thousands for
some applications (Xiong et al., 2013). In this paper, we apply
the quasi-optimal quadrature developed by Druskin et al. (2016)
to reduce the number of quadrature points. This method is based
on conversion of the Fourier integral from real ky axis to contours
in complex ky domain (see, e.g., Trefethen et al., 2006; Trefethen
and Weideman, 2014). The derivation can be explained as follows.
Substituting ux; ky ; z in equation 10 with equation 12, we obtain

ux; y; z

Assuming iky
lows:

~ y eiky y dky :
k2y I Tky 1 fk

(14)

p
s, the above equation can be written as fol-

Figure 1. The 2.5D finite-difference grid.

p
p
1
~
p sI Ti s1 fi
sds;
s

(16)

p
where 1 s can be approximated using rational functions (see Ingerman et al., 2000); i.e.
N
X
1
wn
p
;
s n1 s sn

To compute ux; y; z, equation 10 can be approximated using


numerical integration as

ux; y; z

p
p
p
e sy
~
p sI Ti s1 fi
sds:
s

(15)

(17)

where sn represents the sampling points and wn represents the


weight at each point. The values sn and wn are complex numbers.
Substituting equation 17 into equation 16, we obtain

ux; 0; z

i
2

Z X
N
p
p
wn
~
sI Ti s1 fi
sds:
s sn
n1
(18)

Applying the residue theorem, we have

ux; 0; z

N
X

p
~ p
wn sn I Ti sn 1 fi
sn

n1

N
X

p
~ i sn ; z;
wn ux;

(19)

n1

p
~ i sn ; z represents the solution of a 2D problem in
where ux;
p
equation 12 with ky i sn .
From the above derivations, we can see that an efficient approxip
mation of 1 s using rational functions is crucial to the efficiency
of the 2.5D elastic modeling algorithm. The details of computing
the poles si and residues wi in equation 17 can be found in Druskin
et al. (2016). This derivation is valid if sn is not located in the
p
same half-plane with the poles of sn I Ti sn 1 . The above
assumption was experimentally verified in Druskin et al. (2016)
for the cases in which absorbing boundary conditions such as stable
perfectly matched layer (PML) are in use for the 2D problem. However, to our knowledge, rigorous proof is currently absent.
Moreover, the convergence of equation 19 is bounded by the error of the rational approximant on the eigenvalues of matrix T. The
negative and positive eigenvalues correspond to propagating and
evanescent modes in elastic waves, respectively, represented by real
and imaginary ky values. We will use the two interval Zolotarev
approximation of the square root developed in Druskin et al.
(2016). In the original formulation, it requires an estimate of positive and negative intervals of Ts spectrum.
We start with the propagative (negative s) interval. Its left bound
can be estimated as k2max , where kmax is the maximal wavenumber
in the model. Normally, in elastic model, this is the wavenumber of
the slowest S-wave of the model (Lisitsa, 2008). The right bound of

Li et al.
the algorithms. Figure 3 compares the displacement field along
the line of z 100 m and from x 110 to 201 m. We use three
methods to compute the displacement field including analytical
formulation, finite-difference modeling using uniform sampling
along the real axis, or quasi-optimal sampling in complex spectral
plane. The integration interval for quasi-optimal sampling is
1; 5 109 5 109 ; 1. We observe that all three results
agree with each other well. This validates the new quadrature
scheme. Figure 4 compares the convergence of the two integration
schemes. We observe that the quasi-optimal quadrature converges
much faster than the uniform sampling along the real axis of ky . To
achieve the same level of error (<0.01), we only need 20 points in
optimal quadrature compared with 100 points using uniform sampling. Because we need to solve a 2D problem for each sampling
point of ky , reduction in the number of quadrature points means a
great improvement in the total computational efficiency.

a)
4

x 10

11

ux by xdipole
Analytic
Uniform sampling
Optimal sampling

abs (u )

the negative interval, i.e., the negative value of s closest to the origin, is rather difficult to estimate.
Estimation of right bound of the interval can be quite laborious,
and one approach is to extend the methodology of Asvadurov et al.
(2003). They use for this bound k2min 2 , where kmin is the minimal
wavenumber in the model (normally of the fastest P-wave) and is
the cosine of the minimal incidence number. Our quadrature corresponds to the equivalent PML boundary at y 0 (Druskin et al.,
2016), and assuming that the first grid step of the equivalent optimal
PML is close to the minimal grid step of the XZ grid, we can estimate as the condition number of the latter grid (ratio of the minimal step to the size of computational domain); i.e., we estimate the
propagative interval as k2max ; k2min 2 .
Another approach to set this parameter is to use an optimized
formula from Ingerman et al. (2000),
p which yields the negative interval as k2max ; k2max exp2 N . As it was shown in the
above paper, such a choice yields consistent approximation error
on the interval with the same left bound but including the origin.
In other words, if
even there would be eigenvalues of T in interval
p
k2max exp2 N ; 0, their contribution would not increase the
error significantly. In the numerical examples of this paper we used
the first approach, but the second one is preferable for grids with a
very large condition number.
Even though usually the positive (evanescent) spectral interval is
larger than the negative one, the density of the evanescent spectrum
of the solution decays rather fast, and we always obtain good results
if we choose positive interval symmetrically to the negative ones
described above.

0
100

120

140

120

140

160

180

200

160

180

200

In this example, we compute the displacement field radiated by


an x-directed dipole source in a homogeneous background as shown
in Figure 2. The dipole is located at x; z 100;100 m, and its
frequency is 5 Hz. The computational domain ranges from 0 to
300 m and partitioned using 1 m grid in both directions. The background is homogeneous with P-wave velocity V P 1 kms, Swave velocity V S 0.5 kms, and density 1 kgm3. The
model is kept simple to conveniently validate the performance of

5
100

x (m)

b)

uz by xdipole

12

abs (uz )

Homogeneous model

Angle (ux )

NUMERICAL EXAMPLES

x 10

Analytic
Uniform sampling
Optimal sampling

0
100

120

140

160

180

200

Angle (uz )

Downloaded 08/01/16 to 14.139.116.34. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

T158

Analytic
Uniform sampling
Optimal sampling

5
100

120

140

160

180

200

x (m)

Figure 2. Plot of the homogeneous model.

Figure 3. Plot of displacement field (ux and uz ) from x 1.1 to 2 m


along the line of z 1 m. The field is excited by a x-directional
dipole source at (1,1) m. The blue solid lines represent the field
computed analytically, the red dashed lines represent the field
computed using finite-difference method with uniform sampling
scheme, and the green dashed-dotted lines represent the field
computed using finite-difference method with the quasi-optimal
scheme.

2.5D elastic wave modeling

the two schemes agree with each other very well. Table 1 compares
the computational cost of the two schemes. We observe a significant
speed up using the quasi-optimal quadrature. Moreover, the number
of quadrature points in the uniform ky sampling scheme increases
with frequency, whereas this number does not change in the quasioptimal ky sampling. This is because in uniform ky sampling along
the real axis, each ky corresponds to a propagating plane wave
mode. The higher the frequency, the more plane wave modes exist.
Therefore, we need more sampling points. On the other hand, in the

In this example, we model a synthetic crosswell seismic survey.


The P-wave, S-wave, and density models are shown in Figure 5.
Two wells are located at x 20 and 20 m. Monopole transmitters
and dipole receivers in the x- and z-directions are deployed along
each well separately from 2.5 to 118.5 m. The transmitter spacing is
4 m, and the receiver spacing is 2 m. The simulation domain is partitioned using a grid of 1 m in both directions. We choose three
frequencies at 50, 75, and 100 Hz and compute the displacement
field at the receivers using uniform and quasi-optimal ky sampling
scheme. The results are plotted in Figures 6 and 7. Figure 6 plots the
receiving displacement field ux and uz along the well excited by the
source at z 58.5 m at 75 Hz. Figure 7 plots the receiving field for
every source and receiver at 75 Hz. We observe that the results from

a)

10

Depth (m)

Numerical error

Uniform sampling
Optimal sampling

Real (ux)

Imag (ux)

20

20

40

40

60

60

80

80

100

100

10

120

0.5

b)

10

50

100

150

200

120

0.5

x 10

0.5

14

0.5

x 10

Real (ux)

14

Imag (ux)
0

Number of quadrature sampling points

Depth (m)

Figure 4. Field error comparison between uniform sampling (red


dashed lines) and quasi-optimal sampling (solid blue lines).

Table 1. Comparison of computational efficiency between


uniform and quasi-optimal sampling schemes.
Quadrature points

50 Hz
75 Hz
100 Hz

a)

CPU time (s)

Optimal

Uniform

Optimal

400
560
1600

20
20
20

95.7
128.3
369.6

5.47
5.47
5.47

m/s

b)

VP

4200

20

3600

100

3400
20

20

Distance (m)

60

60

80

80

100

100

kg/m3
2300

0
20

40
60

1900

80

1800

100

1700
20

20

Distance (m)

Depth (m)

80

40

2000
Depth (m)

3800

40

0.5

0.5
14
x 10

120

0.5

0.5
14
x 10

Figure 6. Displacement field at receivers by the monopole source at


58.5 m. The red dashed-dotted lines represent the field computed
using uniform sampling scheme, and the blue dashed lines represent
the field computed using the quasi-optimal sampling scheme.

c)

2100

20

60

20

120

17.5
23.4
67.6

m/s

4000
40

20

Speed up

Uniform

VP

Depth (m)

Downloaded 08/01/16 to 14.139.116.34. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Heterogeneous model

T159

2250

40
60

2200

80
100
20

20

Distance (m)

2150

Figure 5. Crosswell seismic model: (a) P-wave


velocity V P, (b) S-wave velocity V S, and (c) density
.

Li et al.

T160

to 6830 m in horizontal direction with spacing 80 m at depth of


150 m. The receiving field is computed from 1 to 5 Hz. Figure 9
compares the receiving field at each frequency computed by uniform or quasi-optimal ky sampling, from which we observe a good

VP

m/s

3500
3000
2500
2000
1500

500
1000
3000

4000

5000

6000

Distance (m)

b)

VS

m/s

Depth (m)

In this example, we simulate a truncated Marmousi model. The


settings in this simulation are the same as the ones in Xiong et al.
(2013). The simulation domain is 4600 m horizontally (2300
6900 m) and 1500 m (01500 m) in depth. The values of density,
P- and S-wave velocities are shown in Figure 8. The domain is partitioned using a grid size of 20 20 m. The source in the simulation
is an x-directed dipole located at x 4250; z 50 m. The receivers are also x-directed dipole and distributed uniformly from 2350

Depth (m)

a)

Elastic wave modeling using Marmousi model

Table 2. Comparison of computational efficiency between


uniform and quasi-optimal ky sampling schemes for the
Marmousi model.

2500

500

2000

1000

1500
3000

4000

5000

6000

Distance (m)

109
109
211
211
211

20
20
20
20
20

106
106
193
193
193

Optimal
20.3
20.3
20.3
20.3
20.3

Figure 7. Plot of the receiving field for the synthetic crosswell survey. The horizontal and vertical
axes represent the locations of the sources and
receivers, respectively. The color at each pixel
maps to the field amplitude.

kg/m

5.2
5.2
9.5
9.5
9.5

a)

3000
500

2500

1000

2000
3000

4000

6000

Figure 8. A truncated Marmousi model: (a) P-wave velocity V P,


(b) S-wave velocity V S, and (c) density .

Uniform sampling
20

x 10

15

40

60
4
80
2

100

c)

Optimal sampling
20

Uniform sampling

x 10
8
6

60

80
2

100
20 40 60 80 100
Depth of receivers (m)

60
4
80
2

100
20 40 60 80 100
Depth of receivers (m)

20
40

15

x 10
8

40

20 40 60 80 100
Depth of receivers (m)

b)

5000

Distance (m)

Depth of sources (m)

Hz
Hz
Hz
Hz
Hz

Uniform

c)

15

d)
Depth of sources (m)

1
2
3
4
5

Optimal

Speed up

Depth (m)

Uniform

CPU time (s)

Depth of sources (m)

Quadrature points

Depth of sources (m)

Downloaded 08/01/16 to 14.139.116.34. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

quasi-optimal sampling scheme, each quadrature point corresponds


to an evanescent ky mode that decays with y. At higher frequencies,
the modes with large ky decay exponentially with y. Therefore,
the quasi-optimal sampling maintains its accuracy at higher
frequencies.

Optimal sampling

15

x 10
8

20
6

40
60

80
2

100
20 40 60 80 100
Depth of receivers (m)

2.5D elastic wave modeling

0
1
2000

3000

4000

5000

6000

7000

x 10

0
3000 4000 5000 6000
Horizontal distance (m)

d)

7000

1
2000
x 10

3000

4000

0
5
2000

3000 4000 5000 6000


Horizontal distance (m)

e)

Real (ux)

0
3000

4000

5000

6000

2
2000

7000

Imag (ux)

Imag (ux)

x 10

0
2
2000

3000 4000 5000 6000


Horizontal distance (m)

7000

x 10

7000

13

1
2000
x 10

3000

4000

5000

6000

7000

3000 4000 5000 6000


Horizontal distance (m)

7000

13

0
1
2000

13

13

7000

x 10

1
2000

6000

x 10

14

13

5000

Real (ux)

2
2000

c)

13

Imag (ux)

x 10

Real (ux)

Real (ux)

Real (ux)

x 10

14

Imag (ux)

Downloaded 08/01/16 to 14.139.116.34. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

b)

13

IImag (ux)

a)

T161

x 10

3000

4000

5000

6000

7000

3000 4000 5000 6000


Horizontal distance (m)

7000

13

0
2
2000

Figure 9. Comparison of computed data using uniform and quasi-optimal quadrature. The red dashed-dotted lines represent the receiving field
computed using uniform ky sampling, and the blue dashed lines represent the receiving field computed using quasi-optimal ky sampling.
agreement between these two methods. To achieve accurate results,
109 uniform quadrature points are used at 1 and 2 Hz, and 211
points are needed at higher frequencies of 3, 4, and 5 Hz. But only
20 quasi-optimal quadrature points are needed to achieve the same
level of accuracy. The acceleration is approximately 5 or 10 times in
computational time. Table 2 shows a detailed comparison of computational cost between these two methods. Furthermore, similar as
the previous example, we also observe that the quasi-optimal quadrature is more stable with frequency than the uniform quadrature.

CONCLUSIONS
We presented a new quadrature for numerical integration in the ky
domain for the frequency-domain 2.5D finite-difference modeling
algorithm. This scheme replaces the sampling scheme along the real
ky axis with a quasi-optimal sampling scheme in the complex ky
domain. It can significantly reduce the number of quadrature points
for spectral integration. Therefore, the new scheme has a better
computational efficiency. Moreover, unlike the uniform sampling
scheme, the new quadrature maintains its accuracy at higher
frequencies with no need to increase the number of quadrature
points. This makes the new scheme more efficient for high
frequencies.

ACKNOWLEDGMENTS
The authors would like to thank L. Knizhnerman from Central
Geophysical Expedition in Moscow, Russia, for his help in the
quasi-optimal quadrature and G. Pan from Schlumberger-Doll Research for his help in the absorbing boundary condition and parallelization of the algorithm. The first author would also like to

acknowledge the support from the National Science Foundation


of China under contract no. 61571264.

REFERENCES
Aki, K., and P. Richards, 1980, Quantitative seismology Theory and
methods: Freeman and Company.
Asvadurov, S., V. Druskin, M. N. Guddati, and L. Knizhnerman, 2003, On
optimal finite difference approximation of PML: SIAM Journal on
Numerical Analysis, 41, 287305, doi: 10.1137/S0036142901391451.
Berenger, J.-P., 1994, A perfectly matched layer for the absorption of
electromagnetic waves: Journal of Computational Physics, 114, 185
200, doi: 10.1006/jcph.1994.1159.
Cao, H., G.-J. Chen, J. Guo, Y.-S. Wu, and Z.-X. Yao, 2008, The 2.5-D
cross-well waveform tomography in the frequent domain and its practical
application: Chinese Journal of Geophysics, 51, 654663, doi: 10.1002/
cjg2.v51.3.
Charara, M., C. Barnes, and A. Tarantola, 1996, The state of affairs in inversion of seismic data: An OVSP example: 66th Annual International
Meeting, SEG, Expanded Abstracts, 19992002.
Davis, T. A., and L. S. Duff, 1997, An unsymmetric-pattern multifrontal
method for sparse LU factorization: SIAM Journal on Matrix Analysis
and Applications, 18, 140158, doi: 10.1137/S0895479894246905.
Doyon, B., and B. Giroux, 2014, Practical aspects of 2.5D frequency-domain finite-difference modelling of viscoelastic waves: 84th Annual
International Meeting, SEG, Expanded Abstracts, 34823486.
Druskin, V., S. Gttel, and L. Knizhnerman, 2016, Near-optimal perfectly
matched layers for indefinite Helmholtz problems: SIAM Review, 58,
90116, doi: 10.1137/140966927.
Igel, H., H. Djikps, and A. Tarantola, 1996, Waveform inversion of marine
reflection seismograms for P impedance and Poissons ratio: Geophysical
Journal International, 124, 363371, doi: 10.1111/j.1365-246X.1996.
tb07026.x.
Ingerman, D., V. Druskin, and L. Knizhnerman, 2000, Optimal finite
difference grids and rational approximations of the square root I. Elliptic
problems: Communications on Pure and Applied Mathematics, 53, 1039
1066, doi: 10.1002/(ISSN)1097-0312.
Keener, J., 2000, Principles of applied mathematics: Transformation and
approximation: Westview Press.
Liang, L., M. Li, R. Rufino, A. Abubakar, L. Nutt, H. Menkiti, S. Dummong,
and R. Tndel, 2013, Application of frequency-domain full-waveform in-

Downloaded 08/01/16 to 14.139.116.34. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

T162

Li et al.

version for time-lapse 3D VSP data interpretation: 83rd Annual


International Meeting, SEG, Expanded Abstracts, 51075112.
Lisitsa, V., 2008, Optimal discretization of PML for elasticity problems:
Electronic Transactions on Numerical Analysis (ETNA), 30, 258277.
Narayan, J. P., 2001, Site-specific strong ground motion prediction using
2.5-D modelling: Geophysical Journal International, 146, 269281,
doi: 10.1046/j.0956-540x.2001.01424.x.
Owusu, J. C., O. Podgornova, M. Charara, S. Leaney, A. Campbell, S. Ali, I.
Borodin, L. Nutt, and H. Menkiti, 2016, Anisotropic elastic full-waveform inversion of walkaway vertical seismic profiling data from the Arabian Gulf: Geophysical Prospecting, 64, 3853, doi: 10.1111/1365-2478
.12227.
Pan, G., and A. Abubakar, 2013, Iterative solution of 3D acoustic wave
equation with perfectly matched layer boundary condition and multigrid
preconditioner: Geophysics, 78, no. 5, T133T140, doi: 10.1190/
geo2012-0287.1.
Trefethen, L. N., and J. A. C. Weideman, 2014, The exponentially convergent trapezoidal rule: SIAM Review, 56, 385458, doi: 10.1137/
130932132.

Trefethen, L. N., J. A. C. Weideman, and T. Schmelzer, 2006, Talbot quadratures and rational approximations: BIT Numerical Mathematics, 46,
653670, doi: 10.1007/s10543-006-0077-9.
Williamson, P. R., and R. G. Pratt, 1995, A critical review of acoustic wave
modeling procedures in 2.5 dimensions: Geophysics, 60, 591595, doi:
10.1190/1.1443798.
Xiong, J. L., Y. Lin, A. Abubakar, and T. M. Habashy, 2013, 2.5-D forward
and inverse modelling of full-waveform elastic seismic survey: Geophysical Journal International, 193, 938948, doi: 10.1093/gji/ggt013.
Yee, K., 1966, Numerical solution of initial boundary value problems
involving Maxwells equations in isotropic media: IEEE Transactions
on Antennas and Propagation, 14, 302307, doi: 10.1109/TAP.1966
.1138693.
Zhou, B., S. Greenhalgh, and M. Greenhalgh, 2012, Wavenumber sampling
strategies for 2.5-D frequency-domain seismic wave modelling in general
anisotropic media: Geophysical Journal International, 188, 223238, doi:
10.1111/j.1365-246X.2011.05246.x.

You might also like