Efficient computation of complete synthetic seismograms for ...

Geophys. J. Int. (2000) 142, 000–000

Efficient computation of complete synthetic seismograms for transversely isotropic spherically symmetric media using optimally accurate DSM operators Kenji Kawai ½ , Nozomu Takeuchi¾ and Robert J. Geller ½

Department of Earth and Planetary Science, Graduate School of Science, Tokyo University, Hongo 7-3-1, Bunkyo-ku, Tokyo 113-0033, Japan. E-mail: (K.K.) [email protected]; (R.J.G.) [email protected]

Earthquake Research Institute, Tokyo University, Yayoi 1-1-1, Bunkyo-ku, Tokyo 113-0032, Japan. E-mail: [email protected]

Received 2004 July xx; in original form 2004 July 30

SUMMARY We use the DSM (Direct Solution Method) with optimally accurate numerical operators to make accurate and efficient computation of complete (including both body- and surface-waves) 3-component synthetic seismograms for transversely isotropic (TI), spherically symmetric media, up to 2 Hz. Such synthetics should be useful in forward and inverse studies of the Earth’s anisotropic structure. We conducted numerical tests to confirm the accuracy of the synthetic seismograms by examining the dependence of the numerical error and the CPU time on the number of vertical grid intervals and the maximum angular order. Key words: transverse isotropy, seismic anisotropy, synthetic seismograms, Direct Solution Method, waveform inversion

2 1

K. Kawai, N. Takeuchi and R. J. Geller INTRODUCTION

Analysis of seismic data on a global scale is one of the most important tools for inferring the structure of the Earth’s interior. Seismic tomography has greatly advanced our knowledge of the structure of the Earth’s deep interior, but almost all studies to date (especially body-wave studies) have assumed isotropic models (or assumed appropriate scaling relations to reduce the number of unknown elastic constants in TI media from 5 to 2). On the other hand, it is well-known that there is significant anisotropy of seismic structure in the mantle (e.g. Montagner & Kennett 1996), and that information on the anisotropic structure is important for understanding flow and deformation processes in the Earth (e.g. Karato 1998). Travel time tomography for an isotropic model is greatly simplified because the group velocity,

, (which determines seismic travel times) is equal to the phase velocity, , and neither depends on propagation direction. (This is of course the definition of isotropy.) On the other hand, both the group velocity and phase velocity for an anisotropic medium are functionals of the wavenumber

; both depend on the propagation direction, and the two velocities are in general not equal (e.g. Helbig, 1994). For a wave in Cartesian coordinates we have

(1)

as the three components of the slowness vector and the components of the group velocity are given by

(2)

Explicit expressions in spherical coordinates for the above quantities for a laterally homogeneous transversely isotropic (TI) medium are given by Woodhouse (1981). The complexity of these expressions and the dependence of the velocities on propagation direction means that travel time tomography for global TI anisotropic structure is effectively unfeasible. (Travel time inversion for a full anisotropic model is even more intractable.) Thus inversion of seismic waveform data appears to be the most promising approach to determining the Earth’s anisotropic structure (e.g. Gung et al. 2003). This means that we require accurate and efficient methods for computing synthetic seismograms and their partial derivatives in anisotropic media.

DSM synthetics for 1-D TI

3

Detailed comparison of observed seismic waveform data (as observed at an array of stations in a narrow azimuth range) and synthetic seismograms is a powerful method for resolving the fine structure of the Earth’s seismic velocities. For example, the existence of the D” discontinuity (Lay & Helmberger 1983), the existence of an ultra-low velocity zone at the base of the mantle (Garnero et al. 1993), the existence of the inner core transition zone (Song & Helmberger 1998) were detected by such analyses. It is desirable to be able to use complete synthetic seismograms (which include all frequency components and all phases in the data) in such studies, because we could then rigorously account for the effect of finite wavelength and the possibility of contamination by phases traveling outside the regions of interest. When we analyze only the lower frequency components of the data, complete synthetic seismograms can be computed by using, for example, the modal summation method (e.g., Gilbert 1971). Such computations are impractical for analyses of broadband data even for spherically symmetric Earth models, so short wavelength approximations such as generalized ray theory (e.g., Helmberger 1974) or the combined use of reflectivity and ray theory (e.g., Kennett 1988) have generally been employed. However, the methods presented in this paper allow complete synthetic seismograms to be computed up to, say, 2 Hz, without the need for such approximations. The following methods have been used to compute synthetic seismograms for anisotropic media: modal summation (Woodhouse & Dziewonski 1984; Yu & Park 1993; Li & Romanowicz 1995; Park 1997), the reflectivity method (Keith & Crampin 1977; Fryer & Frazer 1987), and the finite difference method (Rumpker & Ryberg 2000). However, although the above methods are useful for waveform modeling for shallow regions, they have some difficulties (e.g. computationally time consuming; not formulated for spherical coordinates) for global-scale applications. For example, the large number of overtones that must be summed seems to make it infeasible to compute synthetics at sufficiently high frequencies using modal summation. The Direct Solution Method (DSM, Geller & Ohminato 1994) has previously been used (e.g. Cummins et al. 1994a; Takeuchi et al. 1996) to compute synthetic seismograms for spherically symmetric isotropic media. In this paper, we present a method for calculating synthetics for laterally homogeneous TI media. Our methods can be extended to the laterally heterogeneous TI case

4

K. Kawai, N. Takeuchi and R. J. Geller

in a straightforward way, following the work of Cummins et al. (1997) and Takeuchi et al. (2000) for a laterally heterogeneous isotropic model.

2

DIRECT SOLUTION METHOD

In this paper, lower case roman subscripts

denote cartesian vector components, and sub-

scripts denote vector components in locally Cartesian spherical coordinates. We use Greek subscripts and superscripts to denote components in the abstract vector space of trial functions. In this paper, for the solid part of the medium these Greek indices correspond to a quadruplet of indices and for the index of the radial trial function, the angular order, the azimuthal order, and the spherical harmonic component, respectively. In the fluid part of the medium the Greek indices correspond to a triplet of indices , defined as above for the solid case. (See Geller & Ohminato 1994 and Cummins et al. 1997 for details.) We use vector trial functions and scalar trial functions

to represent the displacement in the solid part of the medium,

to represent the dependent variable, , where

is the

change in the pressure, in the fluid part of the medium:

(3)

As shown by Geller & Ohminato (1994), the DSM transforms the weak form of the elastic

equation of motion to the following system of discretized linear equations

(4)

is the mass (kinetic energy) matrix, is the stiffness (potential energy) matrix, is the excitation vector, and , which enforces continuity conditions at fluid-solid boundaries, is

where

non-zero only for the spheroidal (P-SV) case. In the DSM, as the name suggests, we compute the synthetic seismogram by directly solving eq. (4). In contrast, modal superposition methods would first set the r.h.s. of eq. (4) to zero, and would solve this equation once for each angular order to compute the eigenfrequencies and the


5

vertically dependent part of the eigenfunctions (which must then be stored). Then each time a synthetic is computed for a given source the modes must be summed. In order to compute complete (including body- and surface-waves) synthetics, all of the modes in the frequency band in question (up to very high overtones) must be summed, so modal methods are disadvantageous. (A detailed discussion of the relative computational costs of the DSM and modal methods is beyond the scope of this paper.)

solid fluid otherwise solid fluid otherwise solid and fluid fluid and solid otherwise solid

fluid

The matrix elements and vector elements in eq. (4) are

(5)

(6)

(7)

(8)

where denotes complex conjugation, is the density, is the elastic tensor in the solid, is

the elastic modulus in the fluid, and

is the outward unit normal to the solid regions at the fluid-

solid boundaries. Anelastic attenuation is included in

and , which are in general complex

and frequency-dependent. Note that the superscripts on the right-hand side of eqs. (5)-(8) refer to the abstract vector space of trial functions, while the subscripts on the right-hand side of these equations refer to the physical space.

6 3

K. Kawai, N. Takeuchi and R. J. Geller DSM FORMULATION FOR TI LATERALLY HOMOGENEOUS MODEL

Cummins et al. (1994a, b) and Takeuchi et al. (1996) presented numerical methods for using the DSM to compute complete synthetics in a spherically symmetric isotropic medium. In this paper we extend the above results to the calculation of complete synthetic seismograms for a spherically symmetric TI medium.

3.1

Transverse isotropy

TI, the simplest anisotropy, is symmetric about the vertical axis. In spherical coordinates the con-

stitutive equation is given by

#

#

#

#

# # "

!

!

(9)

where and # are the stress and strain tensors respectively, the blank spaces in the matrix denote zeros, and

! " are the independent 5 elastic constants in TI, as defined by Love (1927).

The dependent elastic constant is given by

"

(10)

In the isotropic case the elastic constants are as follows:

$ $ !"

where $ and are the Lamé constants.

(11)

DSM synthetics for 1-D TI 3.2

7

Trial functions

All computations are carried out in spherical coordinates. The trial functions used in the expansions of the wavefield are given by Eq. (3). Here we show the explicit form of the vector trial functions for the solid part of the medium. We use linear spline functions % for the vertically dependent part of the trial functions and vector spherical harmonics

, and , defined as follows,

for the laterally dependent part of the trial functions:

& & & &

&

where & is a (fully normalized) surface spherical harmonic and

(12)

. See Press

et al. (1986, p.181) for a complete description of the spherical harmonics and associated Legendre polynomials. The explicit form of the trial functions is as follows:

% % %

(13)

The scalar trial functions in the fluid regions are

% &

(14)

We choose linear splines as the radially dependent part of the trial functions. Their explicit

%

form is

where

'

'

(15)

otherwise

' ' . The first and second lines of (15) are ignored for and " ,

respectively.

8


3.3

Matrix elements

Due to the degeneracy of the spherically symmetric problem, the matrix elements depend only on and are independent of . The following discussion therefore considers the matrix elements and DSM equation of motion for some particular and ; for simplicity we henceforth omit the indices and . We thus denote the elements of the matrices as ( and for the solid part ¼

¼

of the medium, and ( and for the fluid part of the medium. ¼

¼

¼

¼

and corresponds

to the spheroidal case, and corresponds to the toroidal case.

! % % !% % !% %

% % % % "% %

We define the following intermediate integrals for the solid part of the medium:

) % % ) % %

¼

¼

)

¼

)

¼

)

¼

¼

¼

)

¼

)

¼

¼

¼

¼

)

¼

¼

¼

(16)

¼

where the integrals (16) are non-zero only if

, is the density and dots indicate

differentiation with respect to . Using the above intermediate expressions, the matrix elements for the solid part of the medium for the spheroidal case ( or

) are:

( ( )

¼

¼

¼

( ( ¼

(17)

¼

and

) ) )

) ) )

¼

¼

¼

¼

) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ¼

¼

¼

¼

¼

¼

¼

¼

¼

¼

¼

¼

¼

¼

¼

¼

¼

¼

¼

¼

¼

¼

(18)

The explicit form of the matrix elements for the toroidal case is:

( )

¼

¼

) ) ) )

¼

¼

¼

¼

¼

(19)


9

The integrals are taken from the CMB to the Earth’s surface or from the Earth’s center to the ICB for the respective solid regions. (We assume there is no ocean layer at the surface. If an ocean layer is present at the surface, it would be handled in the same way as the outer core.)

% % % % % %

We define the following intermediate integrals for the fluid part of the medium:

) ¼

) ¼

) ¼

¼

¼

¼

(20)

The matrix elements for the fluid part of the medium are

) )

( ) ¼

¼

¼

¼

¼

(21)

At the boundary between the solid and fluid media, there are two nodes corresponding to the for the solid same depth. At the ICB (Inner Core Boundary) the indices for these nodes are for medium and for the fluid, while for the CMB (Core Mantle Boundary) we have

the fluid and for the solid. As shown by Geller & Ohminato (1994, 50-51),

enforces the

continuity of displacement and traction at fluid-solid boundaries. The only non-zero elements of

are:

* *

* *

(22)

where and are the radius at the ICB and CMB, respectively. The essential boundary condition + 3.4

+ + is imposed at .

Optimally accurate operators

The displacement in the previous section is represented using spherical harmonics for the lateral dependence and linear spline functions for the vertical dependence of the trial functions. The numerical operators derived using these trial functions are then replaced by optimally accurate operators (Geller & Takeuchi 1995; Takeuchi et al. 1996). Following Takeuchi et al. (1996) the sub-matrices in eq. (16) are replaced by the corresponding optimally accurate matrix operators.

10


This can be done in a straightforward fashion, because the elastic moduli in eq. (16) of this paper are

!, and " , while the elastic moduli in eq. (3) of Takeuchi et al. (1996) are and $,

but the functional forms of the various integrals in this paper are otherwise identical to those of Takeuchi et al. (1996). Similarly, the functional form of the toroidal integrals in eq. (19) is also identical to functional forms used in Takeuchi et al. (1996).

3.5

Excitation

In the vector spherical harmonics basis, the displacement can be written as

,

(23)

In terms of the trial functions used in this paper we can express and , as follows:

,

%

%

%

(24)

Even if optimally accurate numerical operators are used, the expansion coefficients still give the value of , , and , at the nodes:

(25)

, In this section we drop the subscripts and

on , and , . The excitation is obtained from

, in which case the right-hand side of eq. (8) is zero except for . The excitation vector eq. (8). We consider the case of a point moment tensor on the - -axis

can be determined by straightforward application of eq. (8). However, the body-force equivalents for certain moment tensor terms are kinematically equivalent to requiring

or ,

to be discontinuous at the source depth (e.g. Takeuchi & Saito 1972, p. 290). This is an essential boundary condition that will be violated by a straightforward application of eq. (8), thereby leading


11

to suboptimal convergence (Strang & Fix 1973). It is preferable to incorporate the displacement discontinuities directly into the definitions of and , . The discontinuities are given by Takeuchi and Saito (1972, p.289).

.

/Æ0 . /Æ 0 0 ! . , /Æ 0 0 ! where ! ! and / 1 .

(26)

The remaining source terms involve discontinuities in the radial traction at the source depth, which is a natural boundary condition.

2 / Æ0

0 0 % 2 / Æ0

0 0 % 2

/Æ0

0 0 % /Æ 0

0 0 %

where /

(27)

1

The above discussion applies to the case of a source exactly at a node. However, in most cases the source will be between nodes. For such cases we use the source representation of Takeuchi & Geller (2003), which allows us to calculate synthetics with the same accuracy as for a source at a node.

4

NUMERICAL EXAMPLES

We compute 3-component (vertical, radial, transverse) synthetic seismograms for the anisotropic (TI) earth model, anisotropic PREM (Dziewonski & Anderson 1981). (Anelastic attenuation is also included in the elastic constants using the above methods without using perturbation approximations.) As shown by Figs. 1 and 2, the accuracy of the synthetic seismograms is improved by a factor of about 30 without increasing the CPU time by using optimally accurate numerical operators. The synthetics in Figs. 1 and 2 are calculated up to 0.05 Hz (20 s period) and a maximum angular order of 2000. Note that the synthetics in Figs. 1 and 2 are not the physical displacement

12


"Exact"

Optimally accurate

Residual * 20 (R.m.s. error 0.026%)

Conventional


500 s

Figure 1. Toroidal (SH) velocity synthetic seismograms for the anisotropic PREM model (transverse component) computed up to 0.05Hz and a maximum angular order of 2000 using the source representation of Takeuchi & Geller (2003). The source is a 600 km deep point double-couple source, with

and all other components of the moment tensor zero. The time dependence of the moment tensor for all

Æ

of the calculations in this paper is a -function rather than a step-function. A low-pass 6-pole filter with corners at 0.02 Hz is applied. The top trace is an essentially exact solution computed using the optimally accurate operators with 25,600 vertical intervals. The second and the fourth traces show the numerical solutions computed using optimally accurate (Geller & Takeuchi 1995; Takeuchi et al. 1996) and conventional numerical operators respectively, with 3,000 vertical intervals for each case. The third and fifth traces show the residuals for the second and fourth traces. The accuracy of the synthetics is improved by a factor about 30 by using optimally accurate operators without increasing the CPU time.

components, as the toroidal and spheroidal wavefields have not yet been summed (see below). Although the examples in Figs. 1 and 2 are relatively long-period synthetics, the ratio of performance improvement obtained by using the optimally accurate operators is independent of the frequency band. We thus use only optimally accurate operators in the following computational examples. As discussed in the previous section, for the TI laterally homogeneous case we decompose eq. (4) into separate systems for each angular order. We then further decompose these systems into separate sets of linear equations for the toroidal (SH) and spheroidal (P-SV) displacement. After these displacement fields are separately computed, we then sum them to obtain the synthetics. As shown by eq. (12), both the spheroidal and toroidal wavefields have non-zero - and

-components, so it is necessary to sum both to obtain accurate synthetics for the horizontal components of displacement. Next, the synthetic seismograms were compared to theoretical travel times to further confirm


13

"Exact"

Optimally accurate


Conventional


500 s

Figure 2. Spheroidal (P-SV) velocity synthetic seismograms for anisotropic PREM model (radial component) computed using the source representation of Takeuchi & Geller (2003). Details are the same as Fig. 1, except that 51,200 vertical intervals were used to compute the “exact” trace and 6,000 vertical intervals were used for the other traces. A larger number of vertical intervals (as compared to the toroidal case) is required because the grid extends to the Earth’s center.

their validity. The arrival times of the synthetic seismograms agree with the travel times (Figs. 3a, 3b and 3c), which were computed independently following Woodhouse (1981). The existence of anisotropy in the upper mantle affects long period surface waves (Rayleigh and Love waves). For the PREM structure, Rayleigh waves are slower and Love waves are faster than for an isotropic Earth model. To study the effects of anisotropic structure in the upper mantle, we calculated synthetic seismograms for anisotropic PREM, which includes TI structure between depths of 24.4 and 220 km. (Figs. 4a and 4b). In these calculations the source is at a depth of 50 km, where TI structure exists in PREM. Next we consider the P waves for a deep focus earthquakes (Fig. 5). For these calculations we use a mesh of 51,200 nodes from the Earth’s surface to the center, which is an average vertical grid interval of about 100 m. The P-wave computed up to 2 Hz for anisotropic PREM using the above gridding is shown in Fig. 5(a). Note the sharp arrival for this case. Next we compute a synthetic for the same fine gridding as Fig. 5(a), but for isotropic PREM rather than anisotropic PREM. The result is shown in Fig. 5(b). The right hand column of Fig. 5 shows the difference between each synthetic and the reference synthetic in Fig. 5(a). In the case of Fig. 5(b) the residual is not a computational error but rather is the physical difference between the anisotropic model and isotropic model due primarily to the difference in arrival times. Fig. 5(c) shows a synthetic for

14


anisotropic PREM computed only up to 1 Hz rather than 2 Hz, for the same gridding as Figs. 5(a) and 5(b). Note the much more emergent character of the arrival. We find sharp onsets in the synthetics computed up to 2 Hz (Figs. 5a and 5b), while there is ringing before P-arrival in the synthetic which was computed up only to 1 Hz (Fig. 5c). This means that the P-wave has significant energy between 1 and 2 Hz. Note that synthetics in Fig. 5 are unfiltered. As a Æ -function source time function is used, truncation of the calculation at a given frequency is equivalent to applying uncausal filter to the synthetics. This is why truncation at a frequency (1 Hz in this case) at which waves have a significant amplitude causes ringing. In order to obtain a sharp onset, we should calculate synthetics up to a sufficiently high frequency. The frequency at which ringing will vanish is determined by the anelasticity of the Earth. Seismic waves are attenuated along the propagation from the source to the receiver. As synthetics computed for an epicentral distance of 45.2 Æ up to 2 Hz (Figs. 5a and 5b) show sharp onsets, this implies that the waves at frequencies higher than 2 Hz are largely attenuated along the travel path for this case. Thus we can tentatively conclude that the synthetics computed up to 2 Hz are effectively ‘complete’ synthetic seismograms that are equivalent in frequency content to the observed waveforms although for further tests should be conducted to confirm this. As a general rule of thumb if the gridding per wavelength is kept constant, the CPU time increases as the cube of the maximum frequency, because when the maximum frequency is doubled the number of frequencies, number of grid points, and maximum angular order all approximately double. Thus a calculation up to only 1 Hz requires only about 1/8 of the CPU time for a calculation up to 2 Hz for the same number of grids per vertical wavelength. In our calculations we use a High Performance Computer (HPC) which has a Pentium 4 3.2 GHz processor with the Intel Fortran Compiler (ifc). The CPU times for the synthetics in Fig. 5 whose waveforms are computed using with an automatic cutoff algorithm and regridding algorithm (discussed below) are as follows: for the toroidal case up to 2 Hz (not shown in Fig. 5): 869 min; for the spheroidal case up to 2 Hz: 4962 min. Most of the CPU time is devoted to solving eq. (4) for each frequency and angular order. For this reason the incremental CPU time required to compute synthetics for each additional receiver for the same earthquake is much smaller than the above times. For example,


15

for the toroidal case (up to 2 Hz) the incremental time for each additional receiver is 9.9 min, or approximately 1 per cent of the cost of the first synthetic. The incremental cost of computing the first synthetic for each additional earthquake is 245 min for the toroidal case (up to 2 Hz), as one additional forward- and back-substitution is also required, but this is only about 25 per cent of the cost of the first synthetic for the first earthquake.

5

DISCUSSION

Many workers use modal superposition (e.g., Woodhouse 1988) to compute synthetics in a spherically symmetric TI Earth model. The most widely used software for computing the eigensolutions appears to be the free software available from UC San Diego (http://mahi.ucsd.edu/Gabi/ rem.dir/surface/minos.html). If modal superposition is used, the calculation of synthetics involves two steps: (1) computing the eigenfrequencies and eigenvalues, and (2) superposing the modes to compute synthetics. We have not made a detailed comparative evaluation, but it appears that several days are required to compute all the modes up to 0.2 Hz (a period of 5 s) with sufficient accuracy for body wave studies. A calculation of the modes up to, say, 2 Hz might thus require on the order of 1 yr or more. An extremely large amount of memory would also be required to store the eigenfunctions. Thus it seems reasonable to say that the methods presented in this paper allow practical calculations of complete synthetic seismograms up to frequencies of several Hertz, whereas modal superposition methods do not. Synthetic seismograms must be computed with sufficient accuracy to allow comparison to observed data. The acceptable error must be determined by each user, but the range between 1 per cent error and 0.1 per cent error is probably a typical accuracy requirement. We now consider the effect of the choice of vertical gridding and maximum angular order on the accuracy and CPU time of the synthetics. We begin by considering the relation between vertical gridding and accuracy (Fig. 6). All calculations in Fig. 6 were carried out up to angular order 8000, to eliminate possible errors due to truncation. The regridding and automatic angular order cutoff algorithms (see Fig. 8) were disabled for the calculations presented in Figs. 6 and 7. Fig. 6 shows error and CPU time estimates for spheroidal synthetics for three passbands. For the passband up to 2 sec period (0.5

16


Hz), a vertical grid with 6,000 nodes gives an error of 1 per cent and a vertical grid with 16,000 nodes gives an error of about 0.1 per cent. The respective CPU times are about 12,000 s and 34,000 s. Next we consider the effect of maximum angular order on the solution error. Figure 7 shows examples in which spheroidal expansions truncated at various angular orders are compared to a reference solution. All solutions used a vertical grid with 51,200 intervals to essentially eliminate errors due to vertical discretization. As shown in Fig. 7, a truncation of the expansion at overly low angular orders leads to large errors, but once a critical threshold is passed the solution rapidly becomes highly accurate. (Note that this applies equally to modal superposition calculations, or to any other method using a spherical harmonic expansion.) Figure 7 shows that the accuracy threshold for the initial P-wave is marginally lower than that for the complete synthetic, but that the difference is not significant. We seek a physical interpretation of the angular order accuracy threshold. We denote the angular order where the drop in r.m.s. error occurs as . approximately depends on frequency, (in Hz), as follows:

(28)

The phase velocity is defined asymptotically given by:

3 13 where 3 is the Earth’s radius. We

(29) obtain a phase velocity of about 5.2 km/s corresponding to

, which means that this is the slowest phase velocity component with significant amplitude in the high-frequency synthetics. Because the source is at a depth of 600 km, the excitation of high frequency Rayleigh and Love waves is essentially nil. In contrast, the accuracy threshold for a shallow source would be much higher than that shown in Fig. 7 or given by eq. (28). The purpose of Figs. 6 and 7 is to demonstrate the effect of grid spacing and maximum angular order on accuracy. However, in any practical computations minimization of CPU time is also an important consideration. In practical calculations our program uses an automatic angular order cutoff algorithm (essentially equivalent to cutting off the expansion at a given phase velocity), so


17

that unnecessary calculations are not made. In actual calculations our program also automatically uses a sparser vertical grid in the deeper part of the Earth at lower frequencies and higher angular orders (“regridding”). The regridding is conducted so that the number of elements per vertical wavelength is roughly constant, as this quantity controls the relative error (see Geller & Takeuchi 1995, eq. 6.2). The above two features were disabled when producing Figs. 6 and 7, but Fig. 8 shows that the CPU time is reduced by a factor of about 3.5 by using regridding and the automatic angular order cutoff following eq. (28). The automatic cutoff can also be used for shallow events, except that the numerical factor in eq. (28) must be greatly increased so that surface waves are included. The methods presented in this paper can also be used for forward and inverse calculations for 3-D TI structure models. To do this we would treat the laterally heterogeneous part of the structure as a perturbation to the laterally homogeneous part of the model and use the Born expansion (Takeuchi et al. 2000; Igel et al. 2000). If we use only the first-term (the first order Born expansion) accuracy is limited, but by taking the expansion to third- or fourth-order, accuracy can be greatly improved.

ACKNOWLEDGMENTS We thank Daisuke Suetsugu and Tatsuhiko Hara for critically reading the manuscript. This research was partly supported by a grant from the Japanese Ministry of Education, Science and Culture (Nos. 14540393 and 16740249). K.K. was supported by a JSPS Fellowship for Young Scientists.

REFERENCES Cummins, P.R., Geller, R.J., Hatori, T. & Takeuchi, N., 1994a. DSM complete synthetic seismograms: SH, spherically symmetric, case, Geophys. Res. Lett., 21, 533–536. Cummins, P.R., Geller, R.J. & Takeuchi, N., 1994b. DSM complete synthetic seismograms: P-SV, spherically symmetric, case, Geophys. Res. Lett., 21, 1663–1666. Cummins, P.R., Takeuchi, N. & Geller, R.J., 1997. Computation of complete synthetic seismograms for laterally heterogeneous models using the Direct Solution Method, Geophys. J. Int., 130, 1–16.

18


Dziewonski, A.M. & Anderson, D.L., 1981. Preliminary reference Earth model, Phys. Earth planet. Inter., 25, 297–356. Freyer, G.J. & Frazer, L.N., 1987. Seismic waves in stratified anisotropic media - II. Elastodynamic eigensolutions for some anisotropic systems, Geophys. J. R. astr. Soc., 91, 73–101. Garnero, E.J., Grand, S.P. & Helmberger, D.V., 1993. Low P-wave velocity at the base of the mantle, Geophys. Res. Lett., 20, 1843–1846. Geller, R.J. & Hara, T., 1993. Two efficient algorithms for iterative linearized inversion of seismic waveform data, Geophys. J. Int., 115, 699–710. Geller, R.J. & Ohminato, T., 1994. Computation of synthetic seismograms and their partial derivatives for heterogeneous media with arbitrary natural boundary conditions using the Direct Solution Method, Geophys. J. Int., 116, 421–446. Geller, R.J. & Takeuchi, N., 1995. A new method for computing highly accurate DSM synthetic seismograms, Geophys. J. Int., 123, 449–470. Gilbert F., 1971. Excitation of normal modes of the Earth by earthquake source, Geophys. J. R. astr. Soc., 22, 223–226. Gung, Y., Panning, M. & Romanowicz, B., 2003. Global anisotropy and the thickness of continents, Nature, 422, 707–711. Helbig, K., 1994. Foundations of Anisotropy for Exploration Seismics, Pergamon. Helmberger, D.V., 1974. Generalized ray theory for shear dislocations, Bull. seism. Soc. Am., 64, 45–64. Igel H., Takeuchi, N., Geller, R.J., Megnin C., Bunge H.-P., Clévédé E., Dalkolmo J., & Romanowicz B., 2000. The COSY Project: verification of global seismic modeling algorithms, Phys. Earth planet. Inter., 119, 3–23. Karato, S., 1998. Some remarks on the origin of seismic anisotropy in the D” layer, Earth Planets Space, 50, 1019–1028. Keith, C.M. & Crampin, S., 1977. Seismic body waves in anisotropic media: propagation through a layer, Geophys. J. R. astr. Soc., 49, 209–223. Kennett, B.L.N., 1988. Systematic approximations to the seismic wavefield, in Seismological Algorithms, Computational Methods and Computer Programs, pp. 237–259, ed. Doornbos, D.J., Academic Press. Lay, T. & Helmberger, D.V., 1983. A lower mantle S-wave triplication and the shear velocity structure of D”, Geophys. J. R. astr. Soc., 75, 799–837. Li X.-D. & Romanowicz, B., 1995. Comparison of global waveform inversions with and without considering cross-branch modal coupling, Geophys. J. Int., 121, 695–709. Love, A.E.H., 1927. A Treatise on the Mathematical Theory of Elasticity, Cambridge University Press. Montagner, J.-P. & Kennett, B.L.N., 1996. How to reconcile body-wave and normal-mode reference earth models, Geophys. J. Int., 125, 229–248.


19

Park J., 1997. Free oscillations in an anisotropic earth: Path-integral asymptotics, Geophys. J. Int., 129, 399–411. Press, W.H., Flannery, S.A., Teukolsky, S.A. & Vetterling, W.T., 1986. Numerical Recipes, Cambridge University Press. Rumpker, G. & Ryberg, T., 2000. New “Fresnel-zone” estimates for shear-wave splitting observations from finite-difference modeling, Geophys. Res. Lett., 27, 2005–2008. Song, X.D. & Helmberger, D.V., 1998. Seismic evidence for an inner core transition zone, Science, 282, 924–927. Strang, G. & Fix, G.J., 1973. An Analysis of the Finite Element Method, Prentice Hall. Takeuchi, H. & Saito, M., 1972. Seismic surface waves, Meth. comp. Phys., 11, 217–295. Takeuchi, N. & Geller, R.J., 2003. Accurate numerical methods for solving the elastic equation of motion for arbitrary source locations, Geophys. J. Int., 154, 852–866. Takeuchi, N., Geller, R.J. & Cummins, P.R., 1996. Highly accurate P-SV complete synthetic seismograms using modified DSM operators, Geophys. Res. Lett., 23, 1175–1178. Takeuchi, N., Geller, R.J. & Cummins, P.R., 2000. Complete synthetic seismograms for 3-D heterogeneous Earth models computed using modified DSM operators and their applicability to inversion for Earth structure, Phys. Earth planet. Inter., 119, 25–36. Woodhouse, J.H., 1981. A note on the calculation of travel times in a transversely isotropic Earth model, Phys. Earth planet. Inter., 25, 357–359. Woodhouse, J.H., 1988. The calculation of eigenfrequencies and eigenfunctions of the free oscillations of the Earth and the Sun, in Seismological Algorithms, Computational Methods and Computer Programs, pp. 321–370, ed. Doornbos, D.J., Academic Press. Woodhouse, J.H. & Dziewonski, A.M., 1984. Mapping the upper mantle: three-dimensional modeling of Earth structure by inversion of seismic waveforms, J. geophys. Res., 89, 5953–5986. Yu Y. & Park J., 1993. Upper mantle anisotropy and coupled-mode long-period surface waves, Geophys. J. Int., 114, 473–489.

20

K. Kawai, N. Takeuchi and R. J. Geller Radial component (Toroidal + Spheroidal)

2000

Time (s)

1000

0 0

30 60 Epicentral Distance (Degrees)

90

Transverse component (Toroidal + Spheroidal) 2000

Time (s)

1000

0 0


90

Vertical component (Spheroidal) 2000

Time (s)

1000

0 0


90

Figure 3. 3-component record section (velocity, rather than displacement) including both toroidal and spheroidal contributions (Top: radial component, Middle: transverse component, Bottom: vertical component) for the anisotropic PREM model. Travel times (computed independently) agree excellently with the synthetics.


21

isotropic anisotropic

Rayleigh

100 s

isotropic anisotropic

Love

100 s

Figure 4. Synthetic seismograms (velocity, rather than displacement) for radial (top) and transverse (bottom) components for isotropic PREM and anisotropic PREM. A band-pass 4-pole filter with corners at 0.01 and 0.02 Hz is applied. The source depth is 50 km and the epicentral distance is Æ. The anisotropic synthetics are delayed relative to the isotropic synthetics for Rayleigh waves, and advanced for Love waves, as expected for PREM.

22

K. Kawai, N. Takeuchi and R. J. Geller (a) 2Hz anisotropic

(b) 2Hz isotropic

Residual

(c) 1Hz anisotropic

Residual (R.m.s. error 32.2%)

2s

Figure 5. Examples of displacement synthetic seismograms (vertical component) for the initial P-wave calculated using the methods of this study. (a) 2 Hz synthetic for anisotropic PREM; (b-left) 2 Hz synthetic for isotropic PREM; (b-right) The differences of traces b-left and a; (c-left) 1 Hz synthetic for anisotropic PREM; (c-right) The differences of traces c-left and a. All synthetics are calculated up to a maximum

angular order of 18,000 for an epicentral distance of Æ and for a point moment tensor source at a depth

Æ

of 600 km depth, with a -function source time function. with

and all other components

of the moment tensor zero. The above traces are unfiltered. 106

100

2s 4s 8s

CPU time (s)

R.m.s. error (%)

2s 4s 8s

1

0.001 2 10

105

n (number of vertical grids)

104

102 10

105

n (number of vertical grids)

Figure 6. These figures show the dependence of r.m.s. error (left) and CPU time (right) on the number of vertical grid points, with angular order fixed to 8,000. Synthetics were computed up to 2, 4 and 8 s (0.5, 0.25 and 0.125 Hz). The “exact” synthetics used as the reference solution were computed using 51,200 vertical intervals, while the others were computed with fewer intervals.


23

106

100

2s 4s 8s

CPU time (s)

R.m.s. error (%)

1

2s 4s 8s 2 s (P only) 4 s (P only) 8 s (P only)

10-10 2 10

l (angular order)

104

102 2 10

l (angular order)

104

Figure 7. The dependence of r.m.s. error (left) and CPU time (right) on maximum angular order. The number of vertical grid intervals is fixed to 51,200. Variance reduction is shown for both the whole synthetic seismogram and also for only the portion including the initial P wave. Synthetics were computed for bands up to 2, 4 and 8 s (0.5, 0.25 and 0.125 Hz) using with 51,200 vertical intervals. The “exact” solutions for each band were computed including angular orders up to 8,000.

24

K. Kawai, N. Takeuchi and R. J. Geller 105

100

4s 4 s (practical use)

CPU time (s)

R.m.s. error (%)

1

10-2

4s 4 s (practical use)

10-8 2 10

l (angular order)

104

103 2 10

l (angular order)

104

Figure 8. The r.m.s. error (left) and CPU time (right) with (dashed line) and without (solid line) regridding and an angular order cutoff. Synthetics were computed for the band up to 4 s (0.25 Hz) using 51,200 vertical intervals. These values were fixed in the calculation shown by the solid line but were reduced as appropriate by regridding and the automatic angular order cutoff in the calculation shown by the dashed line. In the latter calculation the cutoff was the lesser of the value on the horizontal axis and that given by eq. (28). Regridding and the automatic cutoff make the calculation about 3.6 times more efficient.