You are on page 1of 13

GEOPHYSICS, VOL. 77, NO. 5 (SEPTEMBER-OCTOBER 2012); P. T187–T199, 12 FIGS., 1 TABLE.

10.1190/GEO2011-0345.1

A new family of finite-difference schemes to solve


the heterogeneous acoustic wave equation

Leandro Di Bartolo1, Cleberson Dors1, and Webe J. Mansur1

2007). Boore (1972) is given credit for introducing heterogeneous


ABSTRACT schemes, and Kelly et al. (1976) for defining them as approaches
where the heterogeneities are considered implicitly by means of
Equivalent staggered grid scheme (ESG) is a new family the variation of material properties from point to point. On the other
of schemes based on the finite-difference method (FDM). hand, homogeneous schemes use the homogeneous wave equation in
The method is applied to acoustic wave propagation in vari- each layer and apply explicit boundary conditions at the interfaces.
able density media and the results are compared with those We should point out that the scheme proposed in Kelly et al. (1976) is
from some classic FDM approaches. The main feature of still widely used for modeling elastic media in onshore regions, be-
this new family is that it is designed to generate results nu- cause in offshore regions such a scheme does not produce good re-
merically equivalent to those using the standard staggered sults due to the presence of water. In fact, materials with Poisson’s
grid formulations (SSG), but with the same memory require- ratio greater than ν ¼ 0.25 are not modeled accurately by this scheme
ments of simple grid schemes. Hence, it results in a reduc-
and frequently the associated results present numerical instability
tion of memory usage by 33% in 2D and 50% in 3D
(Marfurt, 1984; Levander, 1988). Other schemes based on FDM have
problems, compared to the memory usage of SSG. The
the same difficulty in modeling regions with acoustic-elastic coupling
first-order system of equations in terms of pressure and ve-
(Stephen, 1988).
locity is not used here. Instead, the formulation is based on
The seminal paper of Alford et al. (1974) discusses precision and
applying new central difference operators to the second-
stability of acoustic FDM schemes applied to acoustic wave equa-
order acoustic wave equation in terms of pressure, obtaining
the same level of accuracy and stability as the SSG schemes. tion in comparison with analytical solutions, showing that the FDM
The equivalence between the ESG and SSG is mathemati- provides reliable results for wave propagation modeling. At the end
cally demonstrated and issues concerning the application of of the 1970s and beginning of the 1980s, the FDM applied to acous-
seismic sources and the boundary conditions are addressed. tics was used for seismic migration by means of the reverse time
migration (RTM) approach (Hemon, 1978; Baysal et al., 1983; Loe-
wenthal and Mufti, 1983). More expensive, as well as more accurate
than Kirchhoff migration, RTM has increasingly been used by the
INTRODUCTION industry to overcome the new migration challenges. Furthermore,
full waveform inversion (FWI) requires efficient and accurate meth-
The finite-difference method (FDM) remains the most popular nu- ods for seismic modeling. These facts renewed the interest in
merical method for modeling the propagation of seismic waves numerical methods for wave propagation in geophysics.
mainly due to its computational efficiency. Since the early papers An alternative and more accurate approach to solve elastic and
on the subject concerned primarily with seismic wave propagation acoustic wave equation is through the use of staggered grids, where
in elastic, homogeneous, and isotropic media (Alterman and Korn- fields of propagation and properties are considered at intermediate
feld, 1968; Alterman and Rotenberg, 1969; Alterman and Loe- positions of the simple grid. Madariaga (1976) was the first to
wenthal, 1970; Ottaviani, 1971), a great variety of numerical apply a scheme based on staggered grids to solve the system of
schemes for heterogeneous media (here named heterogeneous equations of first order in geophysics for elastic and isotropic media,
schemes) have been proposed (e.g., Levander, 1989; Moczo et al., which circumvents the aforementioned problems for materials with

Manuscript received by the Editor 16 September 2011; revised manuscript received 4 March 2012; published online 21 August 2012; corrected version
published online 28 August 2012.
1
Federal University of Rio de Janeiro, COPPE/Department of Civil Engineering, Brazil. E-mail: leandrodibartolo@gmail.com; clebersondors@gmail.com;
webe@coc.ufrj.br.
© 2012 Society of Exploration Geophysicists. All rights reserved.

T187
T188 Di Bartolo et al.

high Poisson’s ratio. Actually, the use of staggered grids was other situations, because it is general. Hence, this procedure
introduced by Yee (1966) to resolve problems of electrodynamics allows the construction of schemes equivalent to any known stag-
and by Harlow and Welch (1965) for problems of fluid dynamics. gered grid scheme, but it directly applies second-order wave equa-
The subject of numerical wave propagation is still a very active tions in terms of only one field.
area of research, and several papers have proposed staggered grid In the first section of this paper, we present the mathematical de-
schemes or partially staggered grid schemes (e.g., Magnier et al., scription of acoustic wave propagation used to formulate the differ-
1994; Saenger et al., 2000). There are several other examples of ent FDM schemes. In the next section, we review the classic FDM
important developments on FDM schemes as advanced and implicit schemes for homogeneous density media and for general heteroge-
methods (see Liu and Sen, 2009a, 2009b), and the nearly analytic neous media. Following that, we develop the ESG and, in the sub-
integration discrete method (NAID), which was fully mathemati- sequent section, we show its equivalence with SSG. Issues
cally revised for solving the 2D acoustic wave equation in Tong concerning the source application, boundary conditions, and stabi-
et al. (2011). lity are treated. At the end, the numerical simulations and our
Currently the scheme of Virieux (1986) and their extension to conclusions are presented.
fourth-order approximation in space (Levander, 1989) are very pop-
ular. They are applied for elastic and isotropic media and also make
use of staggered grids. The great advantage of schemes based on ACOUSTIC WAVE PROPAGATION THEORY
staggered grids is that, unlike many available schemes based on The phenomena of acoustic wave propagation in linear regime
the second-order wave equation, they are able to model regions with can be formulated using two different but equivalent mathematical
a water-rock interface, allowing application in the presence of ma- descriptions: One based on a second-order wave equation, and other
terials with any Poisson’s ratio (including water), even when den- based on a first-order system of equations. In this section, we pre-
sity contrasts are considered. Thus, the Virieux type schemes can sent these two mathematical formulations necessary to develop the
also be used for acoustics. numerical schemes discussed afterward.
A disadvantage of staggered grid schemes is that they require
more computational memory than simple grid schemes. In this
First-order system
sense, Luo and Schuster (1990) proposed a “parsimonious” displa-
cement-stress staggered scheme, which solves the drawback The description of acoustic wave propagation, in terms of velo-
of large memory requirement for the case of elastic waves in iso- city and pressure, is obtained using Newton’s Second Law together
tropic media. The parsimonious staggered grid scheme (PSG) with the continuity equation, and considering adiabatic propagation
can be applied to the acoustic case leading to a displacement- (Wapenaar and Berkhout, 1989). The vectorial linear equation of
pressure scheme where displacements are stored rather than the motion for the particle velocity field v and acoustic pressure field
pressure. The acoustic version of PSG is presented in Appendix A. p is
Such a scheme needs to store two panels for each displacement
component, which means four panels for 2D and six for 3D; ∂vðr; tÞ
ρðrÞ þ ∇pðr; tÞ ¼ fðr; tÞ; (1)
i.e., requiring more memory than a standard staggered grid scheme ∂t
(SSG). Because the PSG is based on approximating the first-order
spatial derivatives of displacements and stresses using stable Vir- where r is the position vector, t is the time, ρ is the density of the
ieux type FDM operators, as presented in Luo and Schuster medium, f is the density of external body force applied, ∂∕∂t is the
(1990), it only can be applied to systems of coupled equations time derivative and the nabla symbol ∇ is the vector differential
(displacement-stress, displacement-pressure, etc.). In other words, operator known as del. In equation 1, the del is applied to the acous-
it cannot be applied directly to second-order wave equations invol- tic field (pressure) and results in a vectorial field known as gradient
ving only one field (displacements, stresses, or pressure), because in field. The continuity equation together with the linear adiabatic
this case one needs to develop new stable FDM operators to approx- equation of state results in the equation
imate the second-order spatial derivatives.
In numerical simulations with the FDM, the second-order acous- 1 ∂pðr; tÞ ∂i ðr; tÞ
þ ∇ · vðr; tÞ ¼ V ; (2)
tic wave equation is traditionally used to formulate single-grid κðrÞ ∂t ∂t
schemes and the system of equations of first order in terms of ve-
locity and pressure (for acoustics) is used to formulate staggered where κ is the adiabatic compression modulus of the medium and
grid schemes. However, there is no physical reason for this, because iV , which represents a source, is a volume density of volume injec-
the different mathematical descriptions must be completely equiva- tion (an air gun, for example). The operator del, in this equation, is
lent, in the sense that they must give the same results for the same applied via the scalar product to a vectorial field (particle velocity)
problem. In this context, in the present paper, a new family of FDM and results in a scalar field known as divergence field.
schemes termed “equivalent staggered grid” (ESG) is developed. The set of equations 1 and 2 describes the acoustic wave propa-
The ESG applies new FDM approximations (operators) to the sec- gation in a medium with properties κ and ρ submitted to external
ond-order acoustic wave equation to make it numerically equivalent perturbations iV (more specifically its first time derivative) and f.
to the SSG, the Virieux type schemes for acoustics, but with the Specifically, these equations describe how the velocity field (vec-
advantage of using the same computational memory as the nonstag- torial) and the acoustic pressure field (scalar) change in a general
gered grid schemes (NSG) and with similar computational cost heterogeneous medium. This system of equations is hyperbolic of
(after optimizations). The proposed ESG operators are constructed first order, where the velocity of wave propagation c depends on the
by means of a systematic procedure which can be applied to medium’s parameters κ and ρ, according to equation 5 below.
New finite-difference schemes for acoustics T189

Second-order equations derivative term, a second-order approximation of the following


form is used
By eliminating the velocity field from the set of equations 1 and
2, the acoustic wave equation for pressure only can be obtained as ∂2 pni;k pnþ1 − 2pni;k þ pn−1
i;k
  2
¼ i;k ; (7)
1 1 ∂2 pðr; tÞ ∂t Δt2
ρðrÞ∇ · ∇pðr; tÞ − 2 ¼ −sðr; tÞ; (3)
ρðrÞ c ðrÞ ∂t2 where pni;k is the pressure at discrete positions ðx; zÞ ¼ ðih; khÞ in
discrete times t ¼ nΔt. We emphasize that this approximation leads
where s is the source and c is the velocity of propagation given, to a two-step time marching scheme in which it is necessary to store
respectively, by two pressure panels at any time step. It also is important to note that
  high-order time approximations based on central difference lead to
∂2 i ðr; tÞ 1 unstable numerical schemes (Anne et al., 2000).
sðr; tÞ ¼ ρðrÞ V 2 − ρðrÞ∇ · fðr; tÞ ; (4)
∂t ρðrÞ For space derivatives, the most common approximation is the
fourth-order one. In the x direction, it is given by
sffiffiffiffiffiffiffiffiffi
κðrÞ ∂2 pni;k 4 pniþ1;k − 2pni;k þ pni−1;k
cðrÞ ¼ : (5) ¼
ρðrÞ ∂x2 3 h2
1 iþ2;k 2pni;k þ pni−2;k
p n

− (8)
Considering media with low-density gradients, equation 3 can be 3 ð2hÞ2
simplified, resulting in the classic acoustic wave equation
or, equivalently,
1 ∂2 pðr; tÞ
∇2 pðr; tÞ − 2 ¼ −sðr; tÞ; (6)
c ðrÞ ∂t2 ∂2 pni;k −pni−2;k þ 16pni−1;k − 30pni;k þ 16pniþ1;k − pniþ2;k
¼ :
∂x2 12h2
where the operator ∇2 is known as the Laplacian operator (9)
(∇2 ¼ ∇ · ∇), and is applied to a scalar field resulting in another
scalar field. Like the previous equations, equation 6 is written with In the z-direction, a similar expression is used. Note that higher-
differential operators and holds for any system of coordinates in order approximations are possible for spatial terms, obtained in a
one, two, or three dimensions, it only being necessary to use correct straightforward way.
expressions for the operators in each case. Considering equation 6 and using the approximations given
above, the NSG scheme is obtained. The stencil of the fourth-order
REVIEW OF SIMPLE FDM SCHEMES
spatial approximation considers five points for each direction (see
In this section, some of the classic schemes based on the FDM are equation 9). Nevertheless, the memory requirement of NSG is two
reviewed, because they are related to the approach to be presented in panels for the acoustic pressure field plus one panel for the velocity
the next section. Characteristics of accuracy, stability, and memory of propagation.
requirement for these schemes are outlined. As will be seen, the
new scheme presented later is an alternative formulation that takes Standard staggered grid scheme (SSG)
advantage of the best characteristics of accuracy, stability, and
memory requirements of each classical scheme. For simplicity, In the present paper, the staggered approach proposed by Virieux
all schemes presented below are developed for 2D media and (1986) is applied to the acoustic case. The resulting scheme is the
adopt approximations of fourth order in space and second order SSG. The staggered grid used here (Figure 1) is simpler than the
in time. original one, because there is no shear stress. Moreover, the normal
stresses in all directions are equal to the negative of the acoustic
Nonstaggered grid scheme (NSG) pressure and the properties of the medium reduce to one (plus
the density), i.e.,
The most simple and popular finite-difference numerical scheme
for acoustic wave propagation is based on the application of central
τxz → 0; τxx ¼ τzz → −p; (10)
difference approximations to the homogeneous density equation,
given by equation 6. Clearly, such a scheme can give good results
only for media with very low-density gradients or when only kine-
μ → 0; λ → κ; (11)
matic responses are considered (e.g., traveltime analysis).
This simple finite-difference approach for the acoustic wave where μ and λ are, as usual, Lamé parameters and τij represents the
equation is developed making use of a simple (nonstaggered) fi- different components of the stress tensor. In Figure 1, the position of
nite-difference grid, resulting in a numerical conditionally stable the acoustic fields and properties are shown for a 2D grid, the ver-
scheme. This approximation is the NSG. In the NSG, the pressure tical direction being used to represent the time axis.
is considered only at discrete points equally spaced in each spatial This staggered grid depicted in Figure 1 can be used to solve
direction (hereafter h) and in time (hereafter Δt). Then, each deri- equations 1 and 2 numerically, using central difference approxima-
vative of equation 6 can be approximated using known central dif- tions, as proposed by Virieux. The resulting scheme is exactly the
ference operators obtained by Taylor series expansions. For the time Virieux scheme simplified with equations 10 and 11. As is well
T190 Di Bartolo et al.

known, the original scheme is independent of the Poisson’s ratio, nþ1 n−1 Δt n
1 ¼ v − bi;kþ12 ðp − pni;k Þ (14)
2 2
vðzÞi;kþ ðzÞi;kþ1
which in consequence means that its acoustic version is stable too 2 2 h i;kþ1
and gives good results. So, to obtain the second order in space and
time, the central difference operator of the form 
Δt nþ12 nþ12
i;k ¼ pi;k − κ i;k
pnþ1 vðxÞiþ1;k − vðxÞi−
n
1
dgni;k giþ12;k − gi−12;k h 2;k
n n 2

¼ (12) 
dx h nþ12 nþ12
þ vðzÞi;kþ 1 − v
ðzÞi;k−1
; (15)
2 2

is applied to approximate the derivatives in equations 1 and 2, at the


correct positions shown in Figure 2, where g is the field of propaga- where vðxÞ and vðzÞ are, respectively, the components of particle ve-
tion (or its components) in the medium (p, vx , or vz ). In Figure 2, locity in each direction (x and z) and b is the buoyancy (inverse of
the arrows pointing to the right and front face of the cube show, the density 1∕ρ) at the staggered grid positions. These buoyancy
respectively, the location where equation 1 is evaluated for the x values are approximated using simple interpolations (for example,
and z-components. On the other hand, the arrow pointing to the linear interpolation). The scheme of fourth order in space (Levan-
upper face shows the location where equation 2 is applied. There- der, 1989) is obtained using the approximation
fore, the second-order SSG can be obtained as below
dgniþ1;k    
9 gniþ1;k − gni;k 1 gniþ2;k − gni−1;k
nþ1 n−1 Δt n
2
¼ −
2
vðxÞiþ 1 ¼ vðxÞiþ
2
1 − biþ12;k ðp − pni;k Þ (13) dx 8 h 8 3h
2 ;k 2;k h iþ1;k
−gniþ2;k þ 27gniþ1;k − 27gni;k þ gni−1;k
¼ (16)
24h
in the x-direction and, an equivalent expression in
the z-direction.
The SSG is an alternative scheme to solve the
acoustic wave equation in terms of pressure and
velocities. The memory required corresponds to
three panels for the propagation fields in 2D (one
for each field: p, vx , and vz ), plus two panels for
medium properties (velocity and density). It is
50% more than the NSG, disregarding the density
ρ and velocity c panels. In 3D, the SSG requires
twice as much memory as NSG (100%). As is
well known, the accuracy of the SSG is greater
than that of the NSG, although both are
fourth-order accurate in space. In addition, the
SSG is suitable to model coupling between elas-
Figure 1. Standard staggered grid for acoustics. The left panel shows the positions of the
pressure field p and of the adiabatic compression modulus κ of the medium, considered tic and acoustic media.
at integer indices of time steps (n). The right panel shows the staggered grid considered
at intermediate time steps, for the velocity components in the x-direction (circles) and Compact staggered grid scheme (CSG)
z-direction (squares) and buoyancy b of the medium.
Another possibility for the discretization of
completely heterogeneous media is the CSG,
proposed by Cohen and Joly (1990). The CSG considers equation 3
in terms of pressure only and is based on the following second-order
spatial approximation
  bniþ1;k
∂ ∂p n
b ¼ 22 ðpniþ1;k − pni;k Þ
∂x ∂x i;k h
bni−1;k
− 2
ðpni;k − pni−1;k Þ; (17)
h2
where b is the buoyancy parameter. It is important to note that equa-
Figure 2. A cube showing the equilibrium position for applying tion 17 formally generates a staggered grid scheme, because the b
SSG and ESG equations. The point P at the center of the cube parameter is taken at intermediate positions, although the pressure
has coordinate Pni;k , the horizontal and vertical edges being given,
respectively, by h and Δt. The ESG equation is evaluated at the positions are always in the simple grid. In fact, this approach can be
cube’s center, whereas the SSG equations are evaluated at the loca- visualized as the application of second-order approximation twice
tions indicated by the arrows. in succession, given by equation 12, to the derivatives in equation 3.
New finite-difference schemes for acoustics T191

The fourth-order approximation proposed by the same authors is second-order equations can be equation 3 or equation 6, according
given by to the case of interest.
  The finite-difference formulas used in the ESG are obtained by
∂ ∂p n twice applying the approximations used to derive the SSG, resulting
b
∂x ∂x i;k in new central difference approximations. The location where the
     equations are evaluated are shown schematically in Figure 2 for
4 λþ1 p − pi λ−1 p − pi−1
¼ biþ12 iþ1 − bi−12 i the ESG (center of the cube) in comparison with the SSG (arrows).
3h 2 h 2 h Therefore, in the next topic, the procedure to obtain the ESG ap-
    
1 p − pi p − pi−2 proximations is explained by deriving the simplest case, i.e., the
− biþ1 iþ2 − bi−1 i ; (18)
3ð2hÞ 2h 2h second-order approximation for the time derivative of the wave
equation. Subsequently, the spatial approximations for the case
λ1 of heterogeneous density media, given by equation 3, are presented,
where bi12 are smoothing buoyancy introduced to stabilize the
2
scheme, given by along with their simplification for the homogeneous case.

λiþ1    
Temporal approximation
biþ12 ¼ 1 − 2λiþ12 biþ12 þ λiþ12 bi−12 þ biþ32 ; (19)
2
The way to obtain the temporal approximation is to apply the
λi−1     central difference formula twice in succession for the first-order de-
bi−12 ¼ 1 − 2λi−12 bi−12 þ λi−12 bi−32 þ biþ12 ; (20) rivative to generate the second-order derivative, as follows
2
   
pnþ1 −pni;k pni;k −pn−1
i;k
− i;k
with λ being the smoothing factor. Note that the form of the approx- ∂2 pni;k Δt Δt
¼ DðDp n Þ ¼
imation given by equation 18 is an attempt to extrapolate equation 8: ∂t2 i;k
Δt
The fourth order is obtained by means of a linear combination of
two second-order approximations with grid space h and 2h, respec-
pi;k − 2pni;k þ pn−1
nþ1
i;k
¼ ; (21)
tively, with the corresponding constants of multiplication 3∕4 and Δt2
−1∕3. For this reason, the stencil of the CSG is the same at that of
where D is the second-order operator for the first-order derivative
the NSG. Hence, five grid points in each direction are considered. In
(similar to equation 12, for the time), used in the SSG scheme. It is
fact, for homogeneous density media, the CSG reduces to the NSG
important to emphasize that only for this case, where simple sec-
scheme, giving the same numerical responses.
ond-order derivatives are approximated by a second-order operator,
For nonhomogeneous density media, it is necessary to consider a
smoothing in the buoyancy model, as presented in equations 19 and does the ESG approximation result in the same formula used for the
20, to keep the scheme stable. This is done by introducing the para- simple finite-difference scheme (NSG).
meter λ as indicated. The optimal λ proposed for regions with dis- Equation 21 can be used either in equations 3 or 6 to approximate
continuous density is λ ¼ 0.75∕16, obtained from a reflection and the time derivatives. The fourth-order approximation for the time
transmission analysis (Cohen and Joly, 1990). Nevertheless, the va- derivative, obtained using the same procedure, leads to an uncon-
lue of λ is out of the range for which the numerical stability has ditionally unstable scheme, exactly as occurs with other central-
been demonstrated formally, and there is no guarantee that the finite-difference schemes, like the NSG and SSG.
scheme maintains the stability for any density contrast.
For the temporal approximation, Cohen and Joly (1990) devel- Spatial approximation
oped a scheme that increases the time approximation to fourth order
The approximations for spatial derivatives, for the case of hetero-
(using the Lax-Wendroff approach), which is not included in the
geneous density media, are similar to the procedure used for the
discussion presented here. So, the temporal approximation consid-
temporal derivatives. Below, we develop the discretization of the
ered here is the second-order central difference given by equation 7.
equation 3 for the second and fourth orders of approximation.
Applying the operator given by equation 12 twice to the spatial
EQUIVALENT STAGGERED GRID SCHEME (ESG) term of equation 3, the desired second-order approximation is
obtained as
This section reports a new family of schemes, the ESG, which are
    n 
constructed to give the same numerical result as the SSG, but using ∂ ∂p n 1 piþ1;k − pni;k
the second-order heterogeneous wave equation in terms of acoustic b ¼ b 1
∂x ∂x i;k h iþ2;k h
pressure only. As a consequence, this new family preserves the  n 
same precision and stability properties of the SSG, but with the pi;k − pni−1;k
− bi−12;k ; (22)
same storage requirement of the NSG (two panels of pressure), h
as will be shown. For simplicity’s sake, we will develop the 2D
case and spatial approximations of fourth order. The 3D and where the buoyancy b must be approximated using the same inter-
high-order extensions are obtained in a straightforward way. polation adopted for the SSG, so that the ESG and SSG present the
To formulate the ESG, the same grid of the SSG is considered same response (as will be shown in the next section). This approx-
(Figure 1). However, instead of discretizing the first-order system of imation is similar to the CSG of second order in space.
equations, given by equations 1 and 2, the second-order acoustic Similarly, using equation 16 for the fourth-order approximation,
wave equations in terms of pressure only are discretized. These the following expression is obtained
T192 Di Bartolo et al.
 
∂ ∂p n (in space and time). The extensions to fourth- and higher-order
b
∂x ∂x i;k schemes, as well as to 3D formulations are straightforward.
   Rewriting equation 3 for 2D media, neglecting the source term,
9 biþ12;k 9 pniþ1;k − pni;k 1 pniþ2;k − pni−1;k results in
¼ −
8 h 8 h 8 3h
 n      
bi−1;k 9 pi;k − pi−1;k 1 piþ1;k − pni−2;k
n n
∂2 p ∂ 1 ∂p ∂ 1 ∂p
− 2 − ¼ κ þ : (26)
h 8 h 8 3h ∂t2 ∂x ρ ∂x ∂z ρ ∂z
  n 
1 iþ2;k 9 piþ2;k − piþ1;k 1 pniþ3;k − pni;k
b 3 n
− −
8 3h 8 h 8 3h Applying now the ESG second-order approximations, given by
 n 
bi−3;k 9 pi−1;k − pi−2;k 1 pi;k − pi−3;k
n n n equation 21 for time and by equation 22 for the x-derivative,
− 2 − : (23) and a similar one for the z-derivative, leads to
3h 8 h 8 3h

Because the same interpolation for b parameters is used, the approx- Δt2
i;k − 2pi;k þ pi;k ¼
pnþ1 κ i;k
n n−1
imation given by equation 23 produces a scheme with the same ac- h2
curacy and stability as the SSG for a medium with general 
heterogeneous density. The complete scheme of fourth order is pre- × biþ12;k ðpniþ1;k − pni;k Þ − bi−12;k ðpni;k − pni−1;k Þ (27)
sented in Appendix B. This scheme is conditionally stable exactly
as the SSG, consequently being relatively insensitive to the density
contrasts. For example, it is different from the CSG, which uses a
þ bi;kþ12 ðpni;kþ1 − pni;k Þ − bi;k−12 ðpni;k − pni;k−1 Þ ;
complicated process of stabilization and can lose stability for some
density contrasts.
To derive the case of constant density media, one can consider the where, as before, the grid spacing is h.
buoyancy constant in equations 22 and 23 which results, respec- On the other hand, the corresponding SSG formulation, as dis-
tively, in cussed before, is represented by

∂2 pni;k pniþ1;k − 2pni;k þ pni−1;k nþ1 n−1 Δt n


¼ ; (24) 2
vðxÞi− 1 ¼ vðxÞi−
2
1 − bi−12;k ðp − pni−1;k Þ; (28)
∂x2 h2 2;k ;k 2 h i;k

for second-order approximation and


nþ1 n−1 Δt n
1 ¼ v − bi;kþ12 ðp − pni;k Þ; (29)
2 2
vðzÞi;kþ ðzÞi;kþ1
∂2 pni;k
2 2 h i;kþ1
1
2
¼ fðpniþ3;k þ pni−3;k Þ − 54ðpniþ2;k þ pni−2;k Þ
∂x 576h2 
Δt nþ12 nþ12
þ 783ðpniþ1;k þ pni−1;k Þ − 1460pni;k g (25) pnþ1
i;k ¼ − κi;k
pni;k vðxÞiþ1;k − vðxÞi− 1
h 2 2;k

for fourth-order approximation. For homogeneous density media, nþ12 nþ12
þ vðzÞi;kþ 1 − v
ðzÞi;k−1
: (30)
the velocity of propagation is considered at the grid nodes (integer 2 2

index) instead of at intermediate points. This means that the ESG is


completely nonstaggered in this case. Actually, this scheme is Substituting equation 28 and equation 29 into equation 30, shifting
equivalent to the NSG for the second order of approximation be- the space indexes when necessary, and regrouping the result by the
cause the spatial and temporal approximations are equal to one an- time index, results in
other, which is no longer the case for the fourth- and high-order
schemes. For both cases, homogeneous and heterogeneous densities 
Δt n−12 n−12
in fourth order of approximation, the stencil of ESG involves seven pnþ1
i;k ¼ − κ i;k
pni;k vðxÞiþ1;k þ vðxÞi− 1
h 2 2;k
points in each direction against five points for the NSG and CSG. 
This explains the greater accuracy of the ESG, and by consequence, n−12 n−12 Δt
of the SSG in relation to the other mentioned schemes. For higher þ vðzÞi;kþ 1 þ v
ðzÞi;k− 1 − κ i;k
2 2 h
orders of approximation, the ESG is entirely new, as in the case of 
fourth order, and provides solutions numerically equivalent to Δt Δt
× biþ12;k ðpniþ1;k − pni;k Þ − bi−12;k ðpni;k − pni−1;k Þ
respective SSG. h h

Δt n Δt n
PROOF OF NUMERICAL EQUIVALENCE þ bi;kþ12 ðpi;kþ1 − pi;k Þ − bi;k−12 ðpi;k − pi;k−1 Þ .
n n
h h
BETWEEN ESG AND SSG
(31)
In this section, we mathematically demonstrate the equivalence
between the ESG and SSG, for the 2D formulation involving Finally, rewriting the second term on the right side of equation 31,
heterogeneous density scheme in the second order of approximation using equation 30 shifted in time and regrouping the result, yields
New finite-difference schemes for acoustics T193

Δt2 Figure 3b, and for wave equation 3, or equation 6, in Figure 3c.
i;k − 2pi;k þ pi;k ¼ −κ i;k 2
pnþ1 n n−1
h The two cases correspond to the same wavelet. The Fourier trans-
 form of this last source term is depicted in Figure 3d.
× biþ12;k ðpniþ1;k − pni;k Þ − bi−12;k ðpni;k − pni−1;k Þ (32) To complete the numerical equivalence between the SSG and
ESG, the additional time derivative of the ESG source term must
be taken numerically (by central differences, using the approxima-
þ bi;kþ12 ðpni;kþ1 − pni;k Þ − bi;k−12 ðpni;k − pni;k−1 Þ : tion given by equation 12 for time) as can be demonstrated by in-
cluding the source term in the demonstration of the preceding
This equation is exactly the same as that of the ESG (equation 27), section. In other words, the second time derivative of the iV para-
as long as one uses the same expressions for buoyancy b at inter- meter must be obtained by applying central differences to the first
mediate positions in both cases. As can be demonstrated in a similar derivative of iV (the source in the SSG case). Despite this, however,
way, the ESG of fourth order is also equivalent to the corresponding if one applies the analytical expression discretized, the result is
SSG. This demonstration is similar to the previous case, as well as more accurate than taking the corresponding numerical derivative.
for the higher-order and 3D schemes. Actually the results with these two approaches are very similar.

SOURCE, BOUNDARY CONDITIONS, Absorbing boundary conditions (ABS)


AND STABILITY
In geophysical simulations of acoustic waves, often a combina-
In this section, we discuss the source term and the correct math- tion of two techniques is used to simulate infinite media. One is
ematical expressions which make the ESG and SSG numerically absorbing regions where the wave amplitudes are attenuated using
equivalent. Some important numerical issues are also discussed methods like that presented in Cerjan et al. (1985) combined with
concerning boundary conditions, stability, and computational cost. nonreflexive conditions on the boundaries, based on the 1D one-
way wave equation (Reynolds, 1978), as done in the present paper.
The source: Ricker pulse and Another possibility is to apply the perfectly matched layer ap-
equivalence between ESG and SSG proach, originally developed by Berenger (1994), or hybrid ABS
(Liu and Sen, 2010).
The source adopted in this paper is a punctual injection of mass
equivalent to a Ricker pulse, with null density of force. In this case,
Numerical stability and computational aspects
the source is characterized only by the injection parameter iV , with
Gaussian functional form (Figure 3a) given by The properties of stability, dispersion, and accuracy of the new
ESG are exactly the same as those of the SSG because both ap-
 
A proaches are numerically equivalent, as shown previously. There-
iV ðtÞ ¼ exp −πðπf t Þ 2 (33)
2πðπf c Þ2
c d fore, all the characteristics of the SSG scheme, which have been
discussed thoroughly in the literature, apply to the ESG scheme.
fðtÞ ¼ 0; (34)
a) 0.00014 b)
0.008
where A is the source amplitude, fc is a para-
0.00012 0.006
meter related to cutoff frequency f cut, and td is
0.0001 0.004
a shifted time whose expressions are given by
0.002
8e-05
t iV
iV

0
6e-05
f cut –0.002
fc ¼ p ffiffiffi (35) 4e-05
3 π –0.004
2e-05 –0.006
pffiffiffi 0 –0.008
2 π 0 0.02 0.04 0.06 0.08 0.1 0.12 0 0.02 0.04 0.06 0.08 0.1 0.12
td ¼ t − : (36) Time (s) Time (s)
f cut
c) d)
0.6 0.025
These expressions completely specify the source 0.4
for all the wave equations presented. As can be 0.2 0.02
seen comparing the first-order system given by 0 0.015
t iV

|IV|

equations 1 and 2 with the second-order equa- –0.2


2

tions 3 or 6, the source has different forms in –0.4 0.01

each mathematical description. Thus, the SSG –0.6


0.005
–0.8
and ESG must use different source terms, respec-
–1 0
tively, the first time derivative of mass injection 0 0.02 0.04 0.06 0.08 0.1 0.12 0 10 20 30 40 50 60 70
parameter iV and the negative of the second time Time (s) Frequency (Hz)
derivative of this parameter multiplied by the
Figure 3. Gaussian density of volume injection iV (setting f cut ¼ 60 Hz and A ¼ 1) and
density of the medium, to be entirely equivalent. source terms of the first-order system of equations and of the second-order equation:
The source term for the first-order system, (a) iV ; (b) first derivative of iV ; (c) second derivative of iV ; and (d) the Fourier transform
given by equations 1 and 2, is presented in of (c).
T194 Di Bartolo et al.

In this section, we discuss some characteristics of the ESG in rela- a kinematic result and poorly representing the amplitudes in this
tion to the other schemes (NSG and CSG). case. The CSG, which is able to model nonhomogeneous density
One of the most important characteristics of a numerical scheme media (like SSG and ESG), requires the introduction of a smoothing
is its stability property. Commonly, stability properties are assessed coefficient λ to circumvent stability problems for some values of the
by means of the Von Neumann type analysis for homogeneous med- density contrasts. In fact, the stability analysis carried out in the
ia. As a result of such analysis, it can be shown that, to avoid in- original paper (Cohen and Joly, 1990) shows that the scheme
stability, the SSG and ESG maximal time interval (critical) is can be unstable for the optimal λ suggested. In the case of homo-
(Saenger et al., 2000) geneous density media, the CSG is equivalent to the NSG, and
hence it has the same stability properties.
 −1
h pffiffiffiffi Xn Regarding memory requirements, Table 1 compares the storage
Δtc ¼ N di ; (37) requirements for the wavefields (Mf ), model parameters (Mm ), and
c i¼1 their sum (Mt ) for 2D and 3D schemes. We can see that the ESG
needs to store only two panels of pressure at any time (equal to the
where h is the grid spacing, c is the highest velocity of propagation NSG), in 2D or 3D problems, whereas the SSG needs one panel for
in the medium, N is the space dimension, and di are the finite- each component of velocity and for pressure (for 2D modeling the
difference coefficients of spatial approximation of the 2n order total is three panels and for 3D modeling the total is four panels). On
for staggered grid. The equation above is valid for the approxima- the other hand, the scheme proposed by Kelly et al. (1976) and the
tion of second order in time. parsimonious one applied to the acoustic case (Appendix A) needs
For fourth-order spatial approximation, in 2D simulations to store two panels for each component of displacement, which is
(N ¼ 2), the coefficients of equation 37 are d1 ¼ 9∕8 and even more than SSG (remembering that the Kelly’s scheme gives no
d2 ¼ 1∕24, obtained from the first line of equation 16. Thus, the good results for acoustics).
critical time step is given by Δtc ¼ 0.606h∕c. As is well known, Regarding the efficiency, Table 1 compares the number of float-
this criterion is more restrictive than the usual nonstaggered ing point operations of the numerical schemes, considering one
finite-difference schemes applied to elastic wave propagation (Kelly grid point and one time step. To compute the number of floating
et al., 1976), in which Δtc ¼ 0.866h∕c for Poisson’s ratio ν ¼ 0.25. point operations, we used the discrete equations in raw form (as
It is important to recall that the Kelly’s scheme is for elastic media obtained from the direct application of the approximation and
and presents unacceptable responses to acoustic media. Hence, it without joining the similar terms) for all the schemes. Table 1
does not apply to the present case. shows the number of products N p and sums N s for 2D and 3D
On the other hand, the critical time step for the NSG is schemes. It is possible to simplify the equations to reduce the num-
Δtc ¼ 0.612h∕c, slightly less restrictive than the staggered grid ber of operations and increase the computational efficiency, par-
one, but equivalent in practice. The NSG scheme, however, is ticularly in the case of the ESG, which presents several repeated
not able to model heterogeneous density media, thus only giving terms. However, issues of optimization are out of the scope of
this paper.

Table 1. Computational efficiency and memory requirement NUMERICAL SIMULATIONS


for 2D and 3D fourth-order schemes: N p and N s are, respec-
tively, the number of product and sum operations; M f , M m This section presents the simulations performed for 2D media.
and M t are, respectively, the number of panels related to The main objective of these simulations is to illustrate the equiva-
wavefields (components), model parameters and total of lence between the SSG and ESG. The order of approximation used
panels that needs to be stored in the memory. in all the implemented schemes is fourth-order in space and second-
order in time. For comparison, the results of the NSG and CSG are
2D Np Ns Mf Mm Mt presented as well. In all the examples, we used a Ricker wavelet
with a cutoff frequency of 60 Hz, amplitude A ¼ 1 and the numer-
NSG 10 13 2 1 3 ical parameters (h and Δt) were chosen to be near the critical ones.
SSG 14 16 3 2 5 In all the simulations, we considered the previously mentioned ab-
CSG 26 29 2 2 4 sorbing regions and nonreflexive boundary conditions to avoid
undesirable reflections from the boundaries.
ESG ρ het. 31 33 2 2 4
ESG ρ hom. 12 15 2 1 3 Homogeneous model
PSG 50 74 4 2 6
A homogeneous medium of dimension 3; 000 × 3; 000 m, com-
3D Np Ns Mf Mm Mt posed of water (c ¼ 1500 m∕s and ρ ¼ 1000 kg∕m3 ), was mod-
NSG 14 19 2 1 3 eled using the numerical parameters h ¼ 5 m and Δt ¼
2.0 × 10−3 s. The shot position is exactly at the center of the
SSG 20 24 4 2 6
model. Figure 4 shows the time response at the point ðx; zÞ ¼
CSG 38 45 2 2 4 ð1.5; 2.0Þ km. As can be seen in Figure 4a, all the responses are
ESG ρ het. 44 49 2 2 4 very similar. Figure 4b shows a zoom where the differences can
ESG ρ hom. 17 22 2 1 3 be better visualized: The response using ESG is numerically equal
PSG 99 159 6 2 8 to that of SSG, because equivalent sources are applied (as explained
before). The other numerical schemes implemented show different
New finite-difference schemes for acoustics T195

responses, although close. With respect to the NSG and CSG, the ρ4 ¼ 2140 kg∕m3 ; c5 ¼ 2500 m∕s, ρ5 ¼ 2300 kg∕m3 ; c6 ¼
results are the same, because the schemes are equivalent for homo- 3000 m∕s, ρ6 ¼ 2400 kg∕m3 . The time interval used is
geneous density media, as emphasized previously. Δt ¼ 6.7 × 10−4 s, near the stability limit. The position of the shot
To illustrate that the ESG gives the same response as SSG with is x0 ¼ 1500 m and z0 ¼ 50 m.
exactly the same numerical characteristics, the situation above was Figure 6 shows the results of a zero offset trace for the different
modeled again with near-critical parameters to generate numerical numerical schemes. In Figure 6a, the direct wave and the five
dispersion (Figure 5). The grid spacing was double (h ¼ 10 m) and reflections from the interfaces can be seen, including the multiple
the time interval was Δt ¼ 1.5 × 10−3 s. The responses are given in waves. Eliminating the direct wave and adjusting the scale allows
Figure 5a. A detail is shown in Figure 5b. As can be seen, the SSG better visualization of the reflections and multiples (Figure 6b). The
and ESG have exactly the same numerical responses, just as is first four events in this figure are the reflections at the first four
verified between the NSG and CSG. interfaces; the fifth event that can be distinguished is related to
the multiple reflections, and the sixth event is the last reflection
Layered model (at the time t ¼ 2.25 s). The last part of the trace is associated with
multiples. It is important to emphasize that there are no multiples
In this example, a medium composed of six horizontal homoge- from the surface of the model because “transparent” boundaries
neous layers is modeled, each one with homogeneous velocity and were used there.
density. This model represents a simplified offshore region with ty- Figure 7a shows a zoom of the late time of the trace, which is
pical rock parameters and with an intermediate salt layer. The do- composed of the last reflection (first wavefront in the picture)
main is a square with sides of 3000 m composed of equally thick and by the multiple waves. The results of the SSG and ESG are
layers (thickness of 500 m). It was discretized with a grid of 601 × exactly equal, whereas that from the CSG is quite similar to the
601 points; i.e., the grid spacing is h ¼ 5 m. The parameters of each staggered grid one but not equal, as can be see in Figure 7b, which
layer are c1 ¼ 1500 m∕s, ρ1 ¼ 1010 kg∕m3 ; c2 ¼ 2000 m∕s, ρ2 ¼
1950 kg∕m3 ; c3 ¼ 2300 m∕s, ρ3 ¼ 2200 kg∕m3 ; c4 ¼ 4500 m∕s,

a)
30
NSG
a) 30 25 CNSG
NSG SSG
25 CNSG 20 ESG
SSG
20 ESG 15
15
Amplitude

10
10
Amplitude

5
5
0
0
–5
–5
–10
–10
–15 –15

–20 –20
0.3 0.35 0.4 0.45 0.5
–25 Time (s)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Time (s)
b) 2
b) NSG
30 1.5 CNSG
NSG SSG
CNSG 1 ESG
29 SSG
ESG 0.5
28 0
Amplitude

27 –0.5
Amplitude

–1
26
–1.5
25 –2

24 –2.5

–3
23
–3.5
0.45 0.455 0.46 0.465 0.47 0.475 0.48 0.485 0.49 0.495 0.5
22
0.388 0.39 0.392 0.394 0.396 0.398 0.4 Time (s)
Time (s)
Figure 5. Responses of a homogeneous medium for different
Figure 4. Responses of a homogeneous medium for the different schemes implemented using a greater grid spacing (h ¼ 10), a si-
schemes implemented, showing that the responses of ESG and tuation where the numerical dispersion is present, but the responses
SSG are the same: (a) The complete response; (b) a zoom version of ESG and SSG remain the same: (a) The complete response; (b) a
of (a) from time 0.388 s to 0.4 s. zoom version of (a) from time 0.45 s to 0.5 s.
T196 Di Bartolo et al.

is an expansion of Figure 7a. On the other hand, the NSG result is Figure 8. The grid size is 13601 × 2801 points. The numerical para-
quite different because it ignores the density contrasts. meters used are h ¼ 3 m and Δt ¼ 3.8 × 10−4 s. The seismic
source is located at ði; kÞ ¼ ð6801; 31Þ.
Marmousi model Regarding memory requirements, it is interesting to note that in
Marmousi 2 model the number of grid points is exactly
In this example, we modeled the wave propagation in a general N ¼ 38; 096; 401, i.e., 145 MB in single precision are necessary
complex heterogeneous medium, the Marmousi 2 (Martin et al., to store any field of propagation (and model earth parameters) or
2006), which is originally composed of a model where density, 290 MB in double precision. Therefore, the Marmousi 2 simula-
compressional velocity, and shear velocity all are heterogeneous. tions using ESG implemented in this paper required 290 MB less
So, the acoustic modeling presented here uses the first two models,
related to the density and compressional velocity, as shown in

a) 0.6
a) 100 NSG
0.5 CNSG
NSG SSG
CNSG ESG
80 SSG 0.4
ESG
60 0.3

Amplitude
0.2
40
Amplitude

0.1
20 0

0 –0.1

–0.2
–20
–0.3
–40
–0.4
2.2 2.25 2.3 2.35 2.4 2.45 2.5
–60 Time (s)
0 0.5 1 1.5 2 2.5
Time (s)
b)
0.16
b) CNSG
SSG
NSG 0.155 ESG
CNSG
10 SSG
ESG 0.15

0.145
Amplitude

5
Amplitude

0.14

0.135
0
0.13

0.125
–5
0.12
2.344 2.346 2.348 2.35 2.352 2.354 2.356
0.5 1 1.5 2 2.5 Time (s)
Time (s)
Figure 7. Late arrivals in the zero offset seismogram trace in
Figure 6. Zero-offset trace of seismogram in a layered medium for layered media, i.e., a zoomed version of Figure 6: (a) from time
the different schemes implemented: (a) Complete response from 0 s 2.2 s to 2.5 s; (b) a zoom version of (a) from time 2.3435 s to
to 2.5 s; (b) a zoom version of (a) without the direct wave. 2.356 s, showing a low-amplitude multiple wave.

a) b)
Distance (km) x10 Distance (km) x10
3 6 9 1.2 1.5 1.8 2.1 2.4 2.7 3.0 3.3 3.6 3.9 3 6 9 1.2 1.5 1.8 2.1 2.4 2.7 3.0 3.3 3.6 3.9

1.5 1.5
Depth (km)

Depth (km)

3.0 4000 3.0


2000
4.5 4.5
2000
6.0 6.0 1000
c
7.5 7.5 ρ (kg/m 3)

Figure 8. Marmousi 2 model: (a) Velocity model in m∕s; (b) Density model in kg∕m3 .
New finite-difference schemes for acoustics T197

memory than did the SSG simulations (we have used double
precision). This, of course, is not much, but in 3D simulations this
issue becomes very important. For example, considering one di-
mension more in the Marmousi 2 model, e.g., 1,024 grid points
in the y-direction, each field of propagation would need 290 GB
to be stored in the memory. So the memory saved using the
ESG in this situation would be 580 GB (two panels).
In the previous figures, we plotted only seismogram traces be-
cause our main interest is to show the equivalence between the
ESG and SSG. In the next figures, we show the results in a more
traditional form. Figure 9 shows the snapshots of the wave propa-
gation using ESG, and Figure 10 shows the corresponding complete
seismogram.
Next, we present figures to compare the results between the dif-
ferent numerical formulations. Figure 11 shows the zero offset trace
of the seismogram for different schemes. Differences in responses
between the NSG and the other schemes can be seen (Figure 11a),
particularly in the first reflection associated with the water sediment
interface. Figure 11b shows a detail of the zero-offset trace of the
seismogram, highlighting the differences between the NSG and
the other schemes. Because the NSG does not take into account
the density contrasts, the observed amplitudes through the time
are, in general, lower than those of the other schemes.

a) 8
NSG
Figure 9. Snapshots of the wave propagation using the new CNSG
6 SSG
ESG for the times: (a) t ¼ 1.9 s, (b) t ¼ 3.8 s, (c) t ¼ 5.7 s, and ESG
(d) t ¼ 7.6 s. 4
Amplitude

Distance (km) x10 4


–2
0 1 2 3 4
0
–4

–6
0.5 1 1.5 2 2.5 3 3.5 4 4.5
Time (s)

5 b)
0.6 NSG
CNSG
SSG
0.4 ESG
Time (s)

0.2
10
Amplitude

–0.2

–0.4

15
–0.6

–0.8
3.4 3.5 3.6 3.7 3.8 3.9 4
Time (s)

Figure 10. Complete seismogram until time t ¼ 19 s with 90% of Figure 11. Zero offset trace of Marmousi 2 model: (a) With all
clip level in amplitude. The x axis shows the offset and the z axis events up to time 4.6 s; (b) a zoom version of (a) from time
shows time of arrival. 3.4 s to 4 s.
T198 Di Bartolo et al.
0.05
NSG alternative to the classic NSG and presenting similar computa-
CNSG
SSG tional cost (after optimization).
0.045 ESG

ACKNOWLEDGMENTS
Amplitude

0.04
We thank André Bulcão and Djalma M. S. Filho of Petrobras for
0.035
the long-term collaboration and FAPERJ, CNPq, CAPES, and PET-
ROBRAS for the financial support.

0.03
APPENDIX A
0.025
3.972 3.974 3.976 3.978 3.98 3.982 3.984 3.986 3.988
ACOUSTIC VERSION OF PSG SCHEME
Time (s)
Applying the parsimonious algorithm (second-order accurate) to
Figure 12. Zero offset trace of Marmousi 2 model: Zoom of a low the acoustic case, using equation 1 (taking into account the relation
amplitude event, showing that the ESG and SSG responses are the v ¼ ∂u∕∂t between velocity and displacement), the displacements
same, unlike for the other schemes. are computed with the expression (Luo and Schuster, 1990)

Δt2 ∂pni;k
ðxÞi;k ¼ 2uðxÞi;k − uðxÞi;k −
unþ1 (A-1)
n n−1
Figure 12 depicts a detail of the same trace shown in the preced- ;
ρ ∂x
ing figures. It can be seen that there are no differences between the
ESG and SSG responses. On the other hand, the CSG result is
slightly different from that of the SSG.
unþ1
ðzÞiþ1;kþ1
¼ 2unðzÞiþ1;kþ1 − un−1
ðzÞiþ1;kþ1
2 2 2 2 2 2
CONCLUSION
Δt2 ∂pniþ1;kþ1
− 2 2
; (A-2)
In this paper, a new family of FDM operators (ESG) was devel- ρ ∂z
oped to solve the heterogeneous acoustic wave equation in terms of
pressure only. Unlike SSG schemes, which were formulated using
where the derivatives are computed with
the system of equations in terms of velocity and pressure, ESG
schemes only consider the pressure field at nodes of a simple grid,
although the physical properties of the medium are taken at inter-
∂pni;k piþ12;k − pi−12;k
n n
mediate positions (formally on a staggered grid). Taking these in- ¼ ; (A-3)
termediate values of the properties in the same way (same ∂x Δx
interpolation) as in the SSG, the ESG responses are numerically
equivalent to those of the SSG of the same order, as long as
the source is applied in a correct way, as explained previously.
∂pniþ1;kþ1 pniþ1;kþ1 − pniþ1;k
2 2
¼ 2 2
: (A-4)
The equivalence between the ESG and SSG was mathematically ∂z Δz
demonstrated for a simple case, 2D second-order in time and space
approximations. However, the fourth and higher orders can be de-
monstrated in a similar way with a little more algebra. Therefore, The pressure is computed from (equation 2 integrated)
the main feature of this new family of schemes is that they preserve

the same precision and stability properties of the SSG, but with the unðxÞiþ1;k − unðxÞi;k
same memory requirements of NSG. This results in a reduction of pniþ1;k ¼ −κ niþ1;k
memory usage by 33% in 2D and 50% in 3D problems, compared to 2 2 Δx
the memory usage of the SSG. unðzÞiþ1;kþ1 − unðzÞiþ1;k−1 
The new proposed ESG was developed and implemented for 2D þ 2 2 2 2
: (A-5)
Δz
media and fourth-order approximation for the sake of simplicity and
several numerical examples were presented to show the numerical
equivalence between ESG and SSG. Nevertheless, 3D extensions of Therefore, the acoustic PSG, given by equations A-1–A-5 is a stag-
the ESG can be carried out in a straightforward way. gered finite-difference scheme in displacement and pressure, where
For second-order approximation in space, the ESG coincides the pressure is not stored.
with the CSG of the same order. However, the ESG of fourth order
in space is entirely different from the fourth-order CSG, because
the last one forces the use of a smaller number of points (stencil) APPENDIX B
than the ESG. This is the reason why the CSG often is referred to 4S-2T ESG EXPRESSION
as a compact scheme, being less accurate and stable than the ESG.
In the special case of homogeneous density media, the ESG op- The ESG expression for fourth order of approximation in space
erator of fourth and higher orders also constitute new schemes, and second order in time is
New finite-difference schemes for acoustics T199

Δt2 Harlow, F., and J. Welch, 1965, Numerical calculation of time-dependent


i;k − 2pi;k þ pi;k ¼
pnþ1 κi;k
n n−1 viscous incompressible flow of fluid with free surface: Physics of Fluids,
h2 8, 2182–2189, doi: 10.1063/1.1761178.
    Hemon, C., 1978, Equations d'onde et modeles: Geophysical Prospecting,
9 9 n 1 n 26, 790–821, doi: 10.1111/gpr.1978.26.issue-4.
× b 1 ðp − pi;k Þ − ðpiþ2;k − pi−1;k Þ
n n
Kelly, K. R., R. W. Ward, S. Treitel, and R. M. Alford, 1976, Synthetic
8 iþ2;k 8 iþ1;k 24 seismograms: A finite-difference approach: Geophysics, 41, 2–27, doi:
  10.1190/1.1440605.
9 1
− bi−12;k ðpni;k − pni−1;k Þ − ðpniþ1;k − pni−2;k Þ Levander, A. R., 1988, Fourth-order finite-difference P-SV seismograms:
8 24 Geophysics, 53, 1425–1436, doi: 10.1190/1.1442422.
   Levander, A. R., 1989, Finite-difference forward modeling in seismology:
1 9 n 1 n Van Nostrand Reinhold, 410–430.
− b 3 ðp − piþ1;k Þ − ðpiþ3;k − pi;k Þ
n n Liu, Y., and M. Sen, 2009a, Advanced finite-difference methods for seismic
24 iþ2;k 8 iþ2;k 24 modeling: Geohorizons, 5–16.
  Liu, Y., and M. Sen, 2009b, An implicit staggered-grid finite-difference
9 1 method for seismic modelling: Geophysical Journal International, 179,
− bi−32;k ðpni−1;k − pni−2;k Þ − ðpni;k − pni−3;k Þ 459–474, doi: 10.1111/gji.2009.179.issue-1.
8 24 Liu, Y., and M. Sen, 2010, A hybrid scheme for absorbing edge reflections in
   numerical modeling of wave propagation: Geophysics, 75, no. 2, A1–A6.
9 9 1 Loewenthal, D., and I. Mufti, 1983, Reversed time migration in spatial
þ bi;kþ12 ðpni;kþ1 − pni;k Þ − ðpni;kþ2 − pni;k−1 Þ
8 8 24 frequency domain: Geophysics, 48, 627–635, doi: 10.1190/1.1441493.
  Luo, Y., and G. Schuster, 1990, Parsimonious staggered grid finite-differen-
9 n 1 n cing of the wave equation: Geophysical Research Letters, 17, 155–158,
− bi;k−12 ðpi;k − pi;k−1 Þ − ðpi;kþ1 − pi;k−2 Þ
n n doi: 10.1029/GL017i002p00155.
8 24 Madariaga, R., 1976, Dynamics of an expanding circular fault: Bulletin of
   the Seismological Society of America, 66, 639.
1 9 1 Magnier, S., P. Mora, and A. Tarantola, 1994, Finite differences on minimal
− bi;kþ32 ðpni;kþ2 − pni;kþ1 Þ − ðpni;kþ3 − pni;k Þ grids: Geophysics, 59, 1435–1443, doi: 10.1190/1.1443700.
24 8 24 Marfurt, K., 1984, Accuracy of finite-difference and finite-element modeling
   of the scalar and elastic wave equations: Geophysics, 49, 533–549, doi: 10
9 1 .1190/1.1441689.
− bi;k−32 ðpni;k−1 − pni;k−2 Þ − ðpni;k − pni;k−3 Þ :
8 24 Martin, G., R. Wiley, and K. Marfurt, 2006, Marmousi2: An elastic up-
grade for Marmousi: The Leading Edge, 25, 156, doi: 10.1190/1
(B-1) .2172306.
Moczo, P., J. Robertsson, and L. Eisner, 2007, The finite-difference time-
domain method for modeling of seismic wave propagation: Ad-
vances in Geophysics, 48, 421–516, doi: 10.1016/S0065-2687(06)
REFERENCES 48008-0.
Ottaviani, M., 1971, Elastic-wave propagation in two evenly-welded quar-
Alford, R. M., K. R. Kelly, and D. M. Boore, 1974, Accuracy of finite- ter-spaces: Bulletin of the Seismological Society of America, 61, 1119.
difference modeling of the acoustic wave equation: Geophysics, 39, Reynolds, A. C., 1978, Boundary conditions for the numerical solution of
834–842, doi: 10.1190/1.1440470. wave propagation problems: Geophysics, 43, 1099–1110, doi: 10.1190/1
Alterman, Z. S., and P. Kornfeld, 1968, Finite difference solution for pulse .1440881.
propagation in a sphere: Israel Journal of Technology, 6, 138–149. Saenger, E. H., N. Gold, and S. A. Shapiro, 2000, Modeling the propagation
Alterman, Z. S., and D. Loewenthal, 1970, Seismic waves in a quarter of elastic waves using a modified finite-difference grid: Wave Motion, 31,
and three-quarter plane: Geophysical Journal of the Royal Astronomical 77–92, doi: 10.1016/S0165-2125(99)00023-2.
Society, 20, 101–126, doi: 10.1111/gji.1970.20.issue-2. Stephen, R., 1988, A review of finite difference methods for seismoacoustics
Alterman, Z. S., and A. Rotenberg, 1969, Seismic waves in a quarter plane: problems at the seafloor: Reviews of Geophysics, 26, 445–458, doi: 10
Bulletin of the Seismological Society of America, 59, 347–368. .1029/RG026i003p00445.
Anne, L., P. Joly, and Q. H. Tran, 2000, Construction and analysis of higher Tong, P., D. Yang, and B. Hua, 2011, High accuracy wave simulation-
order finite difference schemes for 1D wave equation: Computational revised derivation, numerical analysis and testing of a nearly analytic
Geosciences, 4, 207–249, doi: 10.1023/A:1011520202197. integration discrete method for solving acoustic wave equation: Interna-
Baysal, E., D. D. Kosloff, and J. W. C. Sherwood, 1983, Reverse time tional Journal of Solids and Structures, 48, 56–70, doi: 10.1016/j.ijsolstr
migration: Geophysics, 48, 1514–1524, doi: 10.1190/1.1441434. .2010.09.003.
Berenger, J., 1994, A perfectly matched layer for the absorption of electro- Virieux, J., 1986, P-sv wave propagation in heterogeneous media: Velocity-
magnetic waves: Journal of Computational Physics, 114, 185–200, doi: stress finite difference method: Geophysics, 51, 889–901, doi: 10.1190/1
10.1006/jcph.1994.1159. .1442147.
Boore, D., 1972, Finite difference methods for seismic wave propagation in Wapenaar, C. P. A., and A. J. Berkhout, 1989, Elastic wave field extrapola-
heterogeneous materials: Methods in Computational Physics, 11, 1–37. tion: Redatuming of single and multicomponent seismic data, 2: Elsevier
Cerjan, C., D. Kosloff, R. Kosloff, and M. Reshef, 1985, A nonreflecting Science Ltd.
boundary condition for discrete acoustic and elastic wave equations: Yee, K., 1966, Numerical solution of initial boundary value problems
Geophysics, 50, 705–708, doi: 10.1190/1.1441945. involving Maxwell’s equations in isotropic media: IEEE Transactions
Cohen, G., and P. Joly, 1990, Fourth order schemes for the heterogeneous on Antennas and Propagation, 14, 302–307, doi: 10.1109/TAP.1966
acoustics equation: Computer Methods in Applied Mechanics and Engi- .1138693.
neering, 80, 397–407, doi: 10.1016/0045-7825(90)90044-M.

You might also like