Frequency domain seismic forward modelling: A tool for waveform ...
Frequency domain seismic forward modelling: A tool for waveform ...
Frequency domain seismic forward modelling: A tool for waveform ...
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
<strong>Frequency</strong> <strong>domain</strong> <strong>seismic</strong> <strong><strong>for</strong>ward</strong> <strong>modelling</strong>:<br />
A <strong>tool</strong> <strong>for</strong> wave<strong>for</strong>m inversion<br />
I Stekl<br />
Department of Geology, Royal School of Mines, Imperial College London<br />
Submitted <strong>for</strong> a degree of Doctor of Philosophy and<br />
Diploma of Imperial College<br />
September 22, 1997
Abstract<br />
Modelling the propagation of <strong>seismic</strong> waves, and thereby predicting the response<br />
at <strong>seismic</strong> receivers is crucial in order to interpret, or <strong>for</strong>mally invert data<br />
from <strong>seismic</strong> experiments. Commonly used <strong>seismic</strong> wave<strong>for</strong>m <strong>modelling</strong> techniques<br />
become impractical when one has to simulate datasets involving a large number of<br />
sources.<br />
The multiple source problem can be eciently solved by frequency <strong>domain</strong><br />
<strong><strong>for</strong>ward</strong> <strong>modelling</strong>.<br />
Futhermore, viscous attenuation is easy to incorporate into<br />
frequency-<strong>domain</strong> methods. Once the frequency <strong>domain</strong> equations are discretized,<br />
the solution (at each given frequency) is implicit in the solution of an extremely<br />
large matrix equation. The essential problem is to ensure the structural sparsity of<br />
the matrix and to take full advantage of this. The sparsity of the matrix is best<br />
handled by the nested-dissection method described by George and Liu (1981).<br />
Ihave analysed and extended the visco-acoustic rotated nite dierence scheme<br />
developed by Joetal.(1996). Ihave shown that these operators are optimal: if the<br />
nested dissection method is used, nothing can be gained by higher order operators.<br />
Several <strong>modelling</strong> and waveeld inversion examples using the scheme are desribed<br />
that demonstrate the eciency of optimised frequency <strong>domain</strong> <strong>modelling</strong> scheme. A<br />
waveeld inversion example proves that frequency <strong>domain</strong> <strong>modelling</strong>, when used as<br />
an integral part of the inversion procedure, can generate an accurate, high quality<br />
image quickly and eciently. A pre-processing technique <strong>for</strong> waveeld inversion is<br />
developed and the eects of the pre-processing on the image and on the convergence<br />
are analyzed. The need <strong>for</strong> an elastic scheme is identied.<br />
To meet the need <strong>for</strong> an elastic sheme, I have further extended the rotated<br />
operator method to the visco-elastic case. This extension leads to a high accuracy<br />
sheme. The visco-elastic scheme is used to predict and identify the presence of shear<br />
waves on a real data example.<br />
1
Acknowledgements<br />
I would like to gratefully acknowledge the assistance and encouragement of<br />
my supervisor, Gerhard Pratt, during the course of this work. Gerhard's suggestions<br />
and inuence have taken the nal result of this thesis at least one step further than<br />
Ihaveinitially expected.<br />
I am grateful to Prof. M Worthington and to the Imperial College borehole<br />
consortium <strong>for</strong> the founds and the data provided. Also I would like toacknowledge<br />
Dr Albert and NAGRA <strong>for</strong> founding of the inversion part of the research and the<br />
supply of the appropriate data set. I would like to acknowledge the help from the<br />
Overseas Research Award Scheme. I would also like to give special thanks to the<br />
members of the geophysics group at Imperial College <strong>for</strong> their comradeship and<br />
encouragement. Thanks to Paul Williamson, Claudia, Zhong-Min, Graham, Mike,<br />
Peter R., Claire, John, Michel, Kevin, Miguel, Martijn, Jo, Hamish, Pui, Paul D.,<br />
Kerry, Anna, Marcus, Richard, Eric, Yanghua, Jeremy, Patricia, Richard, Heraldo<br />
and George <strong>for</strong> creating an amiable and supportive working environment at Imperial<br />
College.<br />
Many thanks to my friends Momo, Neven, Branka, Sandra and all others <strong>for</strong><br />
the great time we had together. I am grateful to my relatives here in London <strong>for</strong><br />
the support in the last ve years I have had from them. I would like to dedicate<br />
this thesis to my parents.<br />
2
Contents<br />
Abstract 1<br />
Acknowledgements 2<br />
List of Figures 7<br />
Chapter 1 Introduction 16<br />
1.1 Historical overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18<br />
1.2 The signicance of the frequency-space <strong>domain</strong> . . . . . . . . . . . . . 22<br />
1.3 Forward <strong>modelling</strong> in the frequency-space <strong>domain</strong> . . . . . . . . . . 24<br />
1.4 Fourier trans<strong>for</strong>ms and frequency <strong>domain</strong> <strong>modelling</strong> . . . . . . . . . 28<br />
1.4.1 Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28<br />
1.4.2 Sampling and the Sampling Theorem . . . . . . . . . . . . . . 29<br />
1.4.3 Anti time-aliasing . . . . . . . . . . . . . . . . . . . . . . . . . 30<br />
1.4.4 Reduced time and the Fourier trans<strong>for</strong>m shifting property . . 31<br />
1.5 Overview of chapters in this thesis . . . . . . . . . . . . . . . . . . . . 33<br />
Chapter 2 Solving frequency <strong>domain</strong> wave equations: Numerical Considerations<br />
35<br />
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35<br />
2.2 Solving linear equation systems: bottlenecks . . . . . . . . . . . . . . 36<br />
2.3 Solving linear equation systems with multiple right hand sides . . . . 38<br />
2.4 Matrix \ll in" and ordering schemes . . . . . . . . . . . . . . . . . . 41<br />
3
2.5 Nested dissection ordering . . . . . . . . . . . . . . . . . . . . . . . . 42<br />
2.6 Operators and memory requirements . . . . . . . . . . . . . . . . . . 49<br />
2.7 Comparison of band and nested dissection ordering . . . . . . . . . . 52<br />
2.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56<br />
Chapter 3 Visco-acoustic frequency <strong>domain</strong> acoustic <strong><strong>for</strong>ward</strong> <strong>modelling</strong><br />
using rotated nite dierence operators 58<br />
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58<br />
3.2 Forward <strong>modelling</strong> using rotated nite dierence operators . . . . . . 60<br />
3.2.1 Second order frequency-<strong>domain</strong> <strong>seismic</strong> <strong>modelling</strong> . . . . . . . 60<br />
3.2.2 The rotated operator concept . . . . . . . . . . . . . . . . . . 61<br />
3.2.3 Finite dierence scheme in homogeneous media . . . . . . . . 63<br />
3.2.4 Lumped and consistent matrix terms . . . . . . . . . . . . . . 64<br />
3.2.5 Determination of optimal coecients . . . . . . . . . . . . . . 64<br />
3.2.6 Discussion of savings with rotated operators . . . . . . . . . . 67<br />
3.2.7 Extension to the heterogenous case . . . . . . . . . . . . . . . 68<br />
3.3 Improvements acheived by rotated nite dierence operators . . . . . 70<br />
3.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76<br />
Chapter 4 <strong>Frequency</strong> <strong>domain</strong> waveeld inversion example 77<br />
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77<br />
4.2 Site description: Grimsel Rock Labaratory . . . . . . . . . . . . . . . 79<br />
4.3 Waveeld inversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84<br />
4.4 Waveeld inversion theory . . . . . . . . . . . . . . . . . . . . . . . . 85<br />
4.4.1 Ecient calculation of the gradient direction . . . . . . . . . 89<br />
4.5 Processing of third party synthetic data . . . . . . . . . . . . . . . . . 91<br />
4.5.1 Travel time tomography . . . . . . . . . . . . . . . . . . . . . 94<br />
4.5.2 Full waveeld inversion . . . . . . . . . . . . . . . . . . . . . 94<br />
4.5.3 Full waveeld inversion of trace-normalised data . . . . . . . . 95<br />
4
4.5.4 Comparison of travel time and full waveeld inversion methods 96<br />
4.6 Inversion of real eld data . . . . . . . . . . . . . . . . . . . . . . . . 97<br />
4.6.1 Initial full waveeld inversion . . . . . . . . . . . . . . . . . . 98<br />
4.6.2 Regularization tests . . . . . . . . . . . . . . . . . . . . . . . . 102<br />
4.7 Isotropic results: Evaluation and verication . . . . . . . . . . . . . 102<br />
4.7.1 Discussion of isotropic results . . . . . . . . . . . . . . . . . . 107<br />
4.8 Anisotropic inversion of the eld data . . . . . . . . . . . . . . . . . . 110<br />
4.9 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113<br />
Chapter 5<br />
Visco-elastic frequency <strong>domain</strong> <strong>seismic</strong> <strong><strong>for</strong>ward</strong> <strong>modelling</strong>119<br />
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119<br />
5.2 Visco-elastic <strong>modelling</strong> . . . . . . . . . . . . . . . . . . . . . . . . . . 121<br />
5.2.1 Rotated nite dierences: Computational stars . . . . . . . . 122<br />
5.2.2 Rotated nite dierences: Operators . . . . . . . . . . . . . . 125<br />
5.2.3 Consistent and lumped mass terms . . . . . . . . . . . . . . . 128<br />
5.2.4 Heterogeneous <strong>for</strong>mulation . . . . . . . . . . . . . . . . . . . . 129<br />
5.3 Numerical errors and optimization . . . . . . . . . . . . . . . . . . . . 132<br />
5.3.1 Determination of optimal coecients . . . . . . . . . . . . . . 132<br />
5.3.2 Numerical dispersion . . . . . . . . . . . . . . . . . . . . . . . 134<br />
5.3.3 Modelling in uids . . . . . . . . . . . . . . . . . . . . . . . . 137<br />
5.3.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139<br />
5.4 Elastic <strong>modelling</strong> example . . . . . . . . . . . . . . . . . . . . . . . 141<br />
5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147<br />
Chapter 6 Conclusions and further work 148<br />
6.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148<br />
6.1.1 Matrix solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . 149<br />
6.1.2 Rotated nite dierence operators . . . . . . . . . . . . . . . . 150<br />
6.1.3 Visco-elastic <strong><strong>for</strong>ward</strong> <strong>modelling</strong> . . . . . . . . . . . . . . . . . 150<br />
5
6.1.4 Waveeld inversion . . . . . . . . . . . . . . . . . . . . . . . . 151<br />
6.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152<br />
6.2.1 Developments in <strong>seismic</strong> <strong>modelling</strong> . . . . . . . . . . . . . . . 152<br />
6.2.2 Developments in waveeld inversion . . . . . . . . . . . . . . . 155<br />
Bibliography 159<br />
Appendix A Dispersion analysis <strong>for</strong> visco-elastic <strong>modelling</strong> 170<br />
6
List of Figures<br />
1.1 A Discrete representation of the <strong><strong>for</strong>ward</strong> <strong>modelling</strong> problem. The<br />
representation is schematic; the assumption of two dimensions is not<br />
required at this stage, nor is this ordering of the node points necessary.<br />
The waveeld (either a scalar or a vector quantity) is sampled at each<br />
of the n x n z node points. . . . . . . . . . . . . . . . . . . . . . . . 27<br />
2.1 Nested dissection versus sequentially ordered matrix a),b) be<strong>for</strong>e LU<br />
decomposition, and c),d) equivalent L matrix after LU decomposition<br />
(George and Liu,1981). Only non-zero elements are shown in each<br />
case. a) Matrix S <strong>for</strong> a sequentially ordered grid. b) Matrix S <strong>for</strong> a<br />
grid ordered using nested dissection. c) L part of the LU decomposed<br />
matrix S <strong>for</strong> case a) (memory required is O(n 3 )). d) L part of the LU<br />
decomposed matrix S <strong>for</strong> case b) (memory required is O(n 2 log(n))).<br />
The memory required to store matrix <strong>for</strong> a realistic value of n on<br />
gure d) is signicantly lower than the one required <strong>for</strong> the matrix<br />
on gure c). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43<br />
2.2 Two-way dissected nite dierence grid. The two way dissector, 5 (in<br />
black) is the last part of the grid to be ordered. . . . . . . . . . . . . 44<br />
7
2.3 Two waydissected matrix S ~<br />
= L ~<br />
U ~<br />
. During LU decomposition the<br />
values <strong>for</strong> L i;j and U i;j are lled in at the corresponding locations used<br />
by S i;j . L i;j and U i;j denotes possible non-zero elements in matrices<br />
L ~<br />
and U ~<br />
respectively after LU decomposition while 0 denotes zero<br />
elements. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45<br />
2.4 All possible subgrid (S(n; 2);S(n; 3) and S(n; 4)) situations arising<br />
during nested dissection.<br />
The thick black borders represent neighbouring<br />
dissectors from previous dissections in the recursion. . . . . . 45<br />
2.5 Enlarged L 5;5 part of the twoway dissected matrix. Non zero elements<br />
are in grey. White space represents logical zero elements. . . . . . . . 46<br />
2.6 Fourth order nite dierence computational star. The symbol identies<br />
those grid points coupled to the central grid point. . . . . . . . 49<br />
2.7 Memory requirements comparison <strong>for</strong> n x = 6:25 n z in case of band<br />
and nested dissection ordering.<br />
The required mesh size represents<br />
the model size necessary to per<strong>for</strong>m acoustic <strong>modelling</strong> of a wide<br />
angle experiment with 10 Hz data and a model 350 km by 48 km.<br />
The minimum P wave velocity is 2.8 km/s. . . . . . . . . . . . . . . 55<br />
2.8 CPU time versus number of grid points <strong>for</strong> the case in which n x = n z ,<br />
computed on Digital Alpha 3000/300 workstation. . . . . . . . . . . 56<br />
3.1 Finite dierence operators <strong>for</strong> acoustic frequency <strong>domain</strong> <strong>seismic</strong> <strong>modelling</strong><br />
in two coordinate systems.<br />
The symbol indicates that the<br />
model parameter is used at the corresponding grid point. a) Finite<br />
dierence operator in the original coordinate system. b) Finite dierence<br />
operator in the rotated coordinate system. c) The combination<br />
of both schemes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60<br />
8
3.2 Numerical dispersion curves <strong>for</strong> frequency <strong>domain</strong> acoustic <strong><strong>for</strong>ward</strong><br />
<strong>modelling</strong> using ordinary second order nite dierence operators. a)<br />
Phase velocity dispersion. b) Group velocity dispersion. . . . . . . . 61<br />
3.3 Numerical dispersion curves <strong>for</strong> frequency <strong>domain</strong> acoustic <strong><strong>for</strong>ward</strong><br />
<strong>modelling</strong> using rotated nite dierence operators. a) Phase velocity<br />
dispersion. b) Group velocity dispersion. . . . . . . . . . . . . . . . 65<br />
3.4 Dierence between the numerical velocity produced with and without<br />
the additonal coecient, d. a) Dierence in group velocity. b)<br />
Dierence in phase velocity. See text <strong>for</strong> detail explanation. . . . . . 67<br />
3.5 a) Model used <strong>for</strong> wide angle <strong><strong>for</strong>ward</strong> <strong>modelling</strong>, from McCarthy et al.<br />
(1991). b) c) and d) The shaded regions depict the size of the models<br />
that one could simulate without nested dissection and/or rotated<br />
nite dierences if the same equipment were used. . . . . . . . . . . 71<br />
3.6 a) Synthetic data section from the model on gure 3.5. b) Common<br />
shot gather from the eld data. c) One of the models suggested by<br />
McCarthy et al. (1991) showing the ray paths used in their <strong>modelling</strong><br />
approach. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73<br />
3.7 Time slices generated by <strong><strong>for</strong>ward</strong> <strong>modelling</strong> true the model on Figure<br />
3.5(a) at 5, 10, 15, 20, 25 and 30 seconds. . . . . . . . . . . . . . . . 74<br />
4.1 Grimsel Pass areal photo. . . . . . . . . . . . . . . . . . . . . . . . . 80<br />
4.2 Inside of the Grimsel Rock labaratory tunnel. . . . . . . . . . . . . . 80<br />
9
4.3 Two representative source gathers of VSP data from Field 2, astrue<br />
amplitude displays. a) A VSP source gather with large oset. The<br />
spurious variation of amplitude from trace to trace is evident, as<br />
is the consistency of alternate traces.<br />
The data were recorded in<br />
two passes, with intermediate traces recorded during a later, \in-ll"<br />
survey. b) A near oset VSP source gather, on which the dramatic<br />
change in amplitude with receiver depth is evident. These variations<br />
in amplitudes cannot be modelled using the 2D acoustic method.<br />
In order to invert these data I apply a normalization to each trace<br />
separately. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82<br />
4.4 A representative common receiver gather of the Field 2 data, following<br />
windowing and trace normalization. The receiver was in borehole 3.<br />
The rst portion of the gather was recorded with sources in borehole 2,<br />
and thus represents a portion of the cross borehole data. The second<br />
section was recorded with sources in the tunnel, and thus represents<br />
a portion of the VSP data. The data have been windowed and tracenormalized.<br />
The random static shifts in the cross borehole data, and<br />
the systematic static shifts in the VSP data are evident. The labels<br />
indicate the VSP source groups that were identied, in order to solve<br />
<strong>for</strong> the source consistent static shifts. . . . . . . . . . . . . . . . . . 83<br />
4.5 Map of the Field 2 study area at the Grimsel Test Site. The <strong>seismic</strong><br />
data were acquired using the tunnel and boreholes BOUS85.002 and<br />
BOUS85.003 (\boreholes 2 and 3").<br />
The remaining boreholes are<br />
exploratory boreholes in which velocity in<strong>for</strong>mation is available and<br />
is used to test the wave<strong>for</strong>m images. The scale of this map is 1:1000,<br />
a representative square area 160m 160m is shown <strong>for</strong> reference. . . 84<br />
10
4.6 Comparison of the travel time tomography result and the full wave-<br />
eld inversion from the third party synthetic elastic wave data. a)<br />
True velocity model used in elastic <strong><strong>for</strong>ward</strong> wave<strong>for</strong>m <strong>modelling</strong>, b)<br />
traveltime tomographic image <strong>for</strong>med from the picked synthetic data,<br />
c) acoustic waveeld inversion of the elastic synthetic data, without<br />
trace normalization, d) acoustic waveeld inversion with tracenormalization.<br />
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92<br />
4.7 Starting model <strong>for</strong> waveeld inversions of the eld data (from anisotropic<br />
velocity tomography). . . . . . . . . . . . . . . . . . . . . . . . . . . 99<br />
4.8 Preliminary full waveeld inversion image using non normalized crosshole<br />
part of the data only. . . . . . . . . . . . . . . . . . . . . . . . . 100<br />
4.9 Preliminary full waveeld inversion image using non normalized VSP<br />
part of the data only. Short oset VSP data are excluded due to large<br />
amplitude variations. . . . . . . . . . . . . . . . . . . . . . . . . . . 100<br />
4.10 Preliminary full waveeld inversion image using non normalized Field<br />
2 data, including both crosshole and VSP sections of the data. Short<br />
oset data are excluded due to large amplitude variations. . . . . . . 101<br />
4.11 Isotropic full waveeld inversion results with various values of smoothing<br />
parameter increasing from 0 (top left corner) to 100 (bottom right<br />
corner). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103<br />
4.12 Trade o curve showing RMS roughness vs RMS residuals <strong>for</strong> a suite<br />
of smoothing parameters. . . . . . . . . . . . . . . . . . . . . . . . . 104<br />
4.13 Final isotropic full waveeld inversion result. . . . . . . . . . . . . . 104<br />
4.14 <strong>Frequency</strong> <strong>domain</strong> eld data at 800Hz. Please see the text <strong>for</strong> a<br />
full description of this gure. The grey scale is a relative amplitude,<br />
from the maximum negative values through to the maximum positive<br />
values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105<br />
11
4.15 <strong>Frequency</strong> <strong>domain</strong> modelled (predicted) data at 800Hz. See text <strong>for</strong><br />
a full description of this gure. The grey scale is a relative amplitude,<br />
from the maximum negative values through to the maximum positive<br />
values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106<br />
4.16 Dierence between eld and modelled data at 800Hz. See text <strong>for</strong> a<br />
full description of this gure. The grey scale is a relative amplitude,<br />
from the maximum negative values through to the maximum positive<br />
values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106<br />
4.17 Inverted source signatures. These signatures were extracted as an<br />
integral part of the waveeld inversion scheme.<br />
The similarity of<br />
the VSP source signatures, apart from the known static shifts, gives<br />
credence to the robustness of the inversion scheme. . . . . . . . . . . 107<br />
4.18 Isotropic inversion of synthetic data set from a homogeneous, anisotropic<br />
model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109<br />
4.19 Data residuals <strong>for</strong> the wave<strong>for</strong>m inversion runs on the acoustic syntetic<br />
elliptically anisotropic (2 percent) data by assuming: a) Isotropic<br />
data (underestimated level of anisotropy) b) 2 percent elliptical anisotropy<br />
(correct value) c) 4 percent eliptical anisotropy (overestimated value). 110<br />
4.20 Anisotropic full waveeld inversion results with 0, 1, 2 and 3% elliptical<br />
anisotropy. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112<br />
4.21 RMS residuals <strong>for</strong> each test anisotropy level. . . . . . . . . . . . . . 113<br />
4.22 Final full waveeld inversion image using 2% elliptical anisotropy. . . 114<br />
4.23 <strong>Frequency</strong> <strong>domain</strong> dierence eld (i.e., data residuals) at 800 Hz from<br />
the anisotropic inversion. See text <strong>for</strong> a full description of this gure.<br />
The grey scale is a relative amplitude, from the maximum negative<br />
values through to the maximum positive values. . . . . . . . . . . . 115<br />
4.24 Final waveeld inversion images from both Fields 1 and 2, using 2%<br />
elliptical anisotropy. . . . . . . . . . . . . . . . . . . . . . . . . . . . 116<br />
12
4.25 Final waveeld inversion images from both Fields 1 and 2, using 2%<br />
elliptical anisotropy (colour version). . . . . . . . . . . . . . . . . . . 117<br />
5.1 Computational stars <strong>for</strong> frequency <strong>domain</strong> elastic <strong>modelling</strong>. These<br />
stars indicate the coupling of the components of the displacement<br />
eld at the central node to displacements at the nearest neighbors.<br />
a) The ordinary, second order computational star, b) a possible rotated<br />
computational star, and c) a minimal, rotated computational<br />
star. The symbol, represents the coupling of the same displacement<br />
components, and also represents the only non-zero terms required in<br />
acoustic <strong>modelling</strong>. The symbol, symbol represents the coupling between<br />
perpendicular displacement components. The star in c) does<br />
not use additional points over the star in a), but introduces additional<br />
coupling between components not present in the original star. . . . . 121<br />
5.2 Optimal values of coecients, a (the fraction of the ordinary second<br />
order scheme) and b (the fraction of the consistent mass matrix),<br />
plotted as a function of the Poisson's ratio, . The optimal value of<br />
coecient b is relatively insensitive to the value of . The optimal<br />
value of coecient a decreases <strong>for</strong> high values of , and becomes 0<br />
<strong>for</strong> the uid case, in which case only the rotated scheme is used. . . 134<br />
13
5.3 Numerical dispersion of the new scheme <strong>for</strong> a Poisson ratio = 0:33,<br />
depicting normalized numerical velocity curves <strong>for</strong> compressional and<br />
shear phase velocities (top tworows) and group velocities (bottom two<br />
rows). Results are presented <strong>for</strong> the standard second order scheme<br />
(left column) and the new, combined scheme (right column).<br />
The<br />
dispersion curves are plotted against the shear wavenumber in grid<br />
point units, i.e., the reciprocal of the number of grid points per shear<br />
wavelength, G s . See text <strong>for</strong> the meaning of the symbols used on the<br />
vertical axes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135<br />
5.4 Numerical dispersion <strong>for</strong> a Poisson ratio =0:4, depicting normalized<br />
numerical velocity curves <strong>for</strong> compressional and shear phase velocities<br />
(top two rows) and group velocities (bottom two rows). Results are<br />
presented <strong>for</strong> both the standard second order scheme (left column)<br />
and the new, combined scheme (right column). The dispersion curves<br />
are plotted against the shear wavenumber in grid point units, i.e., the<br />
reciprocal of the number of grid points per shear wavelength, G s . See<br />
text <strong>for</strong> the meaning of the symbols used on the vertical axes. . . . . 136<br />
5.5 Compressional wave dispersion in uids <strong>for</strong> the new, rotated scheme.<br />
In the uid case I use only the rotated scheme, with no component of<br />
the original, unrotated scheme (a = 0). a) Normalized compressional<br />
phase velocities. b) Normalized compressional group velocities. . . . 139<br />
5.6 P-wave velocity model <strong>for</strong> the Imperial College crosshole experiment.<br />
The model was obtained using acoustic fullwave inversion (Pratt at<br />
al. 1995). Data from the experiment, and modelled data <strong>for</strong> this<br />
velocity structure, are shown in Figures 5.7 and 5.8 . . . . . . . . . . 143<br />
14
5.7 a) A representative common source gather from the crosshole data<br />
collected at the Imperial College test site. The signal to noise ratio is<br />
high, and the rst arrival wave<strong>for</strong>ms are clear and coherent. At late<br />
times, incoherent, large amplitude arrivals dominate. b) Predicted<br />
common source data using acoustic <strong><strong>for</strong>ward</strong> <strong>modelling</strong> in the velocity<br />
structure shown in Figure 5.6. The rst arrival traveltimes and wave<strong>for</strong>ms<br />
match well with the observed data, but the large amplitude,<br />
late arrivals are not predicted with the acoustic method. . . . . . . . 144<br />
5.8 Predicted common source data using the new visco-elastic <strong>modelling</strong><br />
results. a) Horizontal displacement component. b) Vertical displacement<br />
component. The horizontal component shows rst arrival times<br />
and wave<strong>for</strong>ms that are similar to the acoustic <strong>modelling</strong> results, and<br />
some high amplitude arrivals at late times. The vertical component<br />
shows high amplitude arrivals similar to those observed on the real<br />
data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145<br />
15
Chapter 1<br />
Introduction<br />
Modelling the propagation of <strong>seismic</strong> waves, and thereby predicting the response<br />
at <strong>seismic</strong> receivers is a crucial step in the interpretion, or the <strong>for</strong>mal inversion<br />
of data from <strong>seismic</strong> experiments. Seismic <strong>modelling</strong> is thus an important <strong>tool</strong> in<br />
geological hypothesis testing.<br />
As <strong>seismic</strong> experiments become increasingly more<br />
sophisticated and complete, we naturally seek to model the <strong>seismic</strong> response of increasingly<br />
realistic media. To model complete, wide band, <strong>seismic</strong> wave behaviour,<br />
in a heterogeneous, porous, and visco-elastic medium, numerical modeling of the full<br />
visco-elastic wave equation would seem desirable. Ideally one would like to include,<br />
if possible, 3 dimensions (3-D), general anisotropy and arbitrary visco-elasticity.<br />
While <strong>for</strong>mulations <strong>for</strong> 3-D anisotropic media are possible (Mora, 1989a; Carcione<br />
et al., 1992), the memory and cpu time requirements <strong>for</strong> realistic model sizes currently<br />
still prevents the production use of such methods, especially <strong>for</strong> multi-source<br />
problems. Nevertheless, there has been a historical progression toward the practical<br />
use of ever more general methods (Alterman and Karal Jr, 1968; Kelly et al., 1975;<br />
Gazdag, 1981; Dablain, 1986; Holdberg, 1987; Virieux, 1986a; Dai et al., 1995),<br />
both <strong>for</strong> one-dimensional (1-D) two-dimensional (2-D) earth models.<br />
There are two major approaches used in <strong>seismic</strong> <strong>modelling</strong>: The rst of these<br />
is asymptotic ray theory (e.g. Cerveny et al. (1982) or Chapman (1985)), a technique<br />
that can oer insight into the nature of various arrivals in the <strong>seismic</strong> record but<br />
16
that may fail to adequately modelthewave<strong>for</strong>ms in complex media. Asymptotic ray<br />
theory assumes a high-frequency wave behaviour; this puts certain constraints on the<br />
model complexity asa function of the lowest wavelength. If velocity discontinuities<br />
are reached, explicit boundary conditions must be applied in order to divide the ray<br />
into reected and transmitted (i.e., refracted) rays, each of which are further traced<br />
through the model. The high frequency restriction limits the use of the technique<br />
to simple models with relatively few data phases, usually specied in advance. The<br />
second group of <strong>modelling</strong> methods comprise the numerical methods based on partial<br />
dierential or integro-dierential wave equations, without the use of a high frequency<br />
approximation. These methods are usually <strong>for</strong>mulated as nite dierence or nite<br />
element problems. Such wave equation methods equation guarantee the simulation<br />
of all possible phases (within the assumptions built into the initial wave equation).<br />
The generation of mode conversions, reections and refractions is not determined<br />
by the choice of input parameters (as in asymptotic ray theory), but is instead<br />
an integral feature of the <strong>modelling</strong> itself. [An exception to this are the numerical<br />
methods of (Madariaga, 1984), based on matrix propagator methods. However these<br />
methods are usually only available <strong>for</strong> 1-D models]. As a result, relating the phases<br />
in the <strong>seismic</strong> record to individual features in the model may not be straight<strong><strong>for</strong>ward</strong><br />
in complex models.<br />
Wave equation methods can be further sub-divided intoanumber of classes,<br />
depending on the <strong>domain</strong> in which the initial wave equation is solved.<br />
Possible<br />
choices of <strong>domain</strong> include any combination of time/frequency, space/wavenumber<br />
or other <strong>domain</strong>s, such as the , p trans<strong>for</strong>m <strong>domain</strong>. Each <strong>domain</strong> has its own<br />
advantages and disadvantages. For 2-D earth models, time <strong>domain</strong> methods have<br />
dominated the literature.<br />
In contrast, this thesis will be largely concerned with<br />
numerical <strong>modelling</strong> in the frequency-space <strong>domain</strong>. The primary reason <strong>for</strong> this is<br />
that the <strong>modelling</strong> algorithm is tightly coupled to a <strong>for</strong>mal method <strong>for</strong> the automatic,<br />
frequency-space <strong>domain</strong> inversion of <strong>seismic</strong> wave<strong>for</strong>m data. The results obtained by<br />
17
Pratt (1995) and Song (1995) showed the great potential of frequency-space <strong>domain</strong><br />
waveeld inversion. Un<strong>for</strong>tunately, the size of the geological experiments in which<br />
these techniques could potentially be applied were limited by the ineciency of the<br />
<strong><strong>for</strong>ward</strong> <strong>modelling</strong> technique. The overall objective of the project described in this<br />
thesis was to develop improved <strong><strong>for</strong>ward</strong> <strong>modelling</strong> methods and to incorporate these<br />
into frequency-space inverse methods, thereby increasing the maximum model size<br />
that can be handled in these methods.<br />
1.1 Historical overview<br />
Seismologists began using wave equation based numerical methods in the late<br />
1960's. Most of this inital work was based on nite dierence techniques. Numerous<br />
discrete solutions <strong>for</strong> the second order wave equation in homogeneous regions by use<br />
of explicit time integration were published (Alterman and Kornfeld, 1968; Alterman<br />
and Karal Jr, 1968; Alterman and Rotenberg, 1969; Alterman and Loewenthal,<br />
1970). In Alterman's work, a homogeneous wave equation <strong>for</strong>mulation was used and<br />
interfaces were treated using explicit boundary conditions. This early work was only<br />
of limited value due to the limited computational resources available at the time,<br />
and due to the limitations on model complexity due to the necessity of treating<br />
each interlayer boundary explicitly.<br />
Nevertheless, these experiments produced a<br />
deep theoretical understanding of wave propagation in homogeneous and two layer<br />
media, and proved that a numerical methods were useful in the innitely many earth<br />
modesl <strong>for</strong> which no analytical solution is available.<br />
Today exploration geophysicists attempt to model much more complex, realworld<br />
media that include irregular boundaries and laterally varying model parameters<br />
in all directions. In order to predict the response in such cases, the interlayer<br />
boundary conditions had to be built implicitly into the <strong>modelling</strong> scheme. It become<br />
common practice in the mid 1970's to use a heterogeneous wave equation <strong>for</strong>mulation<br />
18
(Boore, 1972; Kelly et al., 1975). With these <strong>modelling</strong> techniques, the simulation<br />
of complex media became possible, although due to the simple, low accuracy nite<br />
dierence <strong>for</strong>mulations and the still limited computational resources, realistically<br />
sized models were still out of reach.<br />
As methods became more capable, a number of other development directions<br />
were explored. These included a switch from the earliest, 1 - D earth models<br />
(Abramovici and Alterman, 1965), to 2-D models (Alterman and Karal Jr, 1968),<br />
and nally to 3-D models (Johnson, 1984; Reshef et al., 1988a; Reshef et al., 1988b;<br />
Mora, 1989a). A fundamental limitation of Cartesian 2-D methods is that they do<br />
not accurately simulate the phase and amplitude of eld <strong>seismic</strong> data, even within<br />
the assumption of a 2-D earth model. Bleistein (1986) suggested a \2.5-D" method<br />
<strong>for</strong> correcting <strong>for</strong> phase and amplitude data from point sources using ray trace parameters<br />
later re<strong>for</strong>mulated by Randall (1991) as a nite dierence <strong>for</strong>mulation;<br />
Song and Williamson (1995) suggested a wavenumber trans<strong>for</strong>m approach <strong>for</strong> accounting<br />
<strong>for</strong> these corrections with nite dierence <strong>modelling</strong> methods in frequency<br />
<strong>domain</strong>.<br />
In addition to the progress made in the last decades in extending the dimensionality<br />
of numerical wave equation <strong>modelling</strong> methods, researchers have also<br />
attempted to model increasingly general physical phenomena.<br />
Methods <strong>for</strong> the<br />
acoustic wave equation (Michell, 1969; Gazdag, 1981; Virieux, 1986b; Reshef et<br />
al., 1988a; Song and Williamson, 1995), the elastic wave equation (Alterman and<br />
Karal Jr, 1968; Virieux, 1986a; Pratt, 1990a), the visco-elastic wave equation (Kjartansson,<br />
1979; Emmerich and Korn, 1987; Robertsson et al., 1994), the anistropic<br />
wave equation (Mora, 1989a; Carcione et al., 1992; Carcione, 1995) and the poroelastic<br />
wave equation (Zhu and McMechan, 1991; Dai et al., 1995) have all been<br />
developed. While one knows that the earth is 3-D, porous and anisotropic, in production<br />
<strong>modelling</strong> and inversion choices and compromises must be made. Even if<br />
it were possible to incorporate all these physical eects, dening appropriate model<br />
19
parameters is a daunting task.<br />
Extracting a detailed P-wave velocity eld from<br />
reection <strong>seismic</strong> data is already dicult (Al-Chalabi, 1994); the full extraction of<br />
heterogeneous, complex-valued, anisotropic visco-elastic parameters in detail would<br />
seem impossible. It is perhaps obvious that one must always simplify the model in<br />
order to be able to represent the essence of the recorded data without unnecessary<br />
overparameterization. This decision naturally depends on the <strong>modelling</strong> objectives;<br />
in some cases the arrival times may be a sucient data representation. In other case<br />
wave<strong>for</strong>m data will be required. If the data do not contain a signicant amount of<br />
the S-wave energy, or if the S-wave phases are not used in the interpretation, the<br />
acoustic assumption may be sucient. However, if S-wave phases are important,<br />
then additional considerations regarding the physical parameters (e.g., the source<br />
mechanism, anisotropy, polarization and borehole eects) often become important.<br />
The second order elastic wave equation is analytically equivalent to the coupled,<br />
rst order, elasto-dynamic equations. However, the two <strong>for</strong>mulations of the<br />
wave equation lead to dierent numerical solutions. Numerically stable, dierencing<br />
expressions are much easier to <strong>for</strong>mulate <strong>for</strong> rst order partial dierential equations<br />
than <strong>for</strong> second order equations.<br />
However, the model parameters must then be<br />
dened on two, separate, \staggered" grids. The denition of the model itself becomes<br />
ambiguous at intermediate points on the grid. Madariaga (1976) developed<br />
the rst, staggered grid, nite-dierence method <strong>for</strong> the elasto-dynamic wave equation<br />
<strong>for</strong>mulation. This <strong>for</strong>mulation became dominant (Virieux, 1986a; Dai et al.,<br />
1995) <strong>for</strong> time <strong>domain</strong> schemes, due to the fact that it was the only known scheme<br />
which enabled the simulation of elastic waves in models with liquid-solid interfaces<br />
(obviously a critical facility in exploration studies (see, <strong>for</strong> example Kerner (1990))).<br />
Some simplications are possible <strong>for</strong> certain kind of experiments by using a<br />
one-way wave equation (Claerbout, 1970). The one-way wave equation is primarily<br />
used due to the high computational cost of simulating the full wave equation. The<br />
method can predict a transmitted wave eld; it is possible to simulate scattered wave<br />
20
elds by explicitly dening each back scattered event, but reverberations and surface<br />
waves travelling perpendicular to the paraxial direction cannot be modelled at all.<br />
Even with this serious disadvantage, the approach has been very popular as a migration<br />
algorithm (Claerbout and Doherty, 1972; Loewenthal et al., 1976; Berkhout<br />
and Van Wulten Palthe, 1979; Berkhout, 1985), since in post-stack migration the<br />
propagation is required in only one direction (down), and the computational costs<br />
are much lower than full wave equation <strong>modelling</strong>. Full wave equation methods,<br />
however have been used extensively in migration from the late 70's (Hemon, 1978;<br />
Beysal et al., 1983; Loewenthal and Mufti, 1983).<br />
In many disciplines the nite element <strong>for</strong>mulation is the primary choice of<br />
numerical method. However, <strong>seismic</strong> <strong>modelling</strong> the nite element method has never<br />
taken over from nite dierences as a main stream technique. Although the earliest<br />
papers on <strong>seismic</strong> <strong>modelling</strong> used the nite element method, (Smith, 1974),<br />
the essential diculty remains with us today: The lack of a mesh generator which<br />
will utilise the full advantage of nite elements, distorting the grid where possible<br />
and still providing a sucient number of node points <strong>for</strong> accurate wave equation<br />
<strong>modelling</strong>. There is another reason why nite element <strong>seismic</strong> <strong>modelling</strong> is not used<br />
more often: Since wave propagation problems demand that the model be sampled<br />
at a very ne scale, using exact, irregular boundaries will not signicantly aect the<br />
result. In most practical cases the knowledge of the model itself is only known at a<br />
relatively long scale length, much coarser than the model parametrization, so that<br />
exact boundaries cannot be dened. Finite elements may have certain advantages<br />
in the case of theoretical, simple models in which only a limited number of homogenous<br />
regions represent the model and an exact solution is required, but in applied<br />
cases where the model is highly heterogeneous and the shape of the \homogeneous"<br />
elements is not known the main advantages of nite elements appear to be of little<br />
use. The main nite element work is still on square (rectangular) grids. In this<br />
case there is no particular advantage of using either the nite dierence or the nite<br />
21
element methods.<br />
1.2 The signicance of the frequency-space <strong>domain</strong><br />
It is clear from the review given above that frequency <strong>domain</strong> methods are<br />
less common than time <strong>domain</strong> methods in <strong>seismic</strong> wave propagation <strong>modelling</strong>. An<br />
early exception was (Lysmer and Drake, 1972), and the fundamental advantages<br />
of frequency <strong>domain</strong> <strong>modelling</strong> (especially <strong>for</strong> multi-source inverse problems) was<br />
clearly pointed out by Marfurt (1984a; 1984b), by Marfurt and Shin (1989) and by<br />
Pratt (1989a).<br />
Time <strong>domain</strong> methods are suitable if the full time <strong>domain</strong> <strong>seismic</strong> section<br />
<strong>for</strong> a single source, or <strong>for</strong> a small number of sources is required.<br />
On the other<br />
hand, frequency <strong>domain</strong> methods are ecient in cases in which a limited number of<br />
single frequency data are required, or in cases in which a time response <strong>for</strong> a large<br />
number of sources is required. As I will show in this thesis, in these circumstances a<br />
frequency <strong>domain</strong> approach can produce results at a fraction of the computational<br />
costs required by time <strong>domain</strong> schemes.<br />
Recently, much attention has been focussed on the <strong>modelling</strong> of <strong>seismic</strong> waves<br />
in models that include visco-elastic losses. Because the attenuation due to viscoelastic<br />
losses is thought to be related to time dependent creep and relaxation effects<br />
(see <strong>for</strong> example Kjartansson (1979)) that lead to integral terms in the wave<br />
equation, special techniques are required to include these eects into time <strong>domain</strong><br />
numerical simulations (Emmerich and Korn, 1987; Carcione et al., 1988; Robertsson<br />
et al., 1994). A solution to the diculty of the representation of the integral<br />
terms in the time dependent wave equation is to trans<strong>for</strong>m the equations into the<br />
frequency <strong>domain</strong>, and model the resultant Helmholtz type equations, in which case<br />
frequency dependent attenuation can be easily represented by complex valued elastic<br />
parameters (Muller, 1983), without any additional computational eort.<br />
22
The advantages of frequency <strong>domain</strong> schemes are utilised to the full extent<br />
in waveeld inversion approaches, in which only a few frequencies may be required<br />
(Pratt et al., 1995; Song et al., 1995); in this thesis I will deal exclusively with<br />
frequency-space <strong>domain</strong> methods and their application to inverse problems | hence<br />
the sub-title of this thesis, \A <strong>tool</strong> <strong>for</strong> waveeld inversion". In order to eciently<br />
invert eld data one needs an accurate and fast <strong><strong>for</strong>ward</strong> <strong>modelling</strong> technique. Computation<br />
time and accuracy normally tradeo against each other: Accuracy can<br />
usually be achieved by using very ne discretization grids, but this is costly in terms<br />
of computational resources. Sophisticated numerical methods achieve accuracy by<br />
optimized design of the numerical method, allowing the number of grid points to be<br />
reduced (<strong>for</strong> a given level of accuracy).<br />
In order to improve the per<strong>for</strong>mance of any <strong><strong>for</strong>ward</strong> <strong>modelling</strong> approach, the<br />
limitations of the particular scheme must be well understood. These limitations include<br />
those introduced by the original choice of the underlying wave equation (i.e.,<br />
are we <strong>modelling</strong> acoustic or elastic waves, are we using one, two or three dimensions,<br />
are we accounting <strong>for</strong> viscous eects, etc), and those limitations caused by<br />
the numerical approximations (i.e., are the numerical methods suciently accurate).<br />
This understanding of the <strong>modelling</strong> method is important in order not to misinterpret<br />
the results, in order to build better models, and in order to be able to choose<br />
appropriate <strong>modelling</strong> techniques and <strong>modelling</strong> parameters. Simple frequency <strong>domain</strong><br />
<strong>modelling</strong> codes, such as the one developed by Pratt (1989b) (in use when I<br />
started the project), are not suciently accurate to be able to handle large surveys.<br />
The software engineering problems associated with frequency <strong>domain</strong> <strong><strong>for</strong>ward</strong><br />
<strong>modelling</strong> are rather dierent from those associated with time <strong>domain</strong> methods.<br />
Once the frequency <strong>domain</strong> equations are discretized, the solution (at a given<br />
frequency) is implicit in the solution of an extremely large matrix equation. The<br />
essential problem is to control the sparsity pattern of the matrix (itself controlled<br />
by the spatial extent of the dierencing operators), and to take full advantage of<br />
23
this sparsity. As in time <strong>domain</strong> methods, however, the overiding concern is to limit<br />
the number of grid points per wavelength that are required. Thus, <strong>for</strong> frequency<br />
<strong>domain</strong> methods, it is critical to optimize the accuracy of the numerical operators,<br />
while also minimizing the spatial extent of these operators.<br />
In time <strong>domain</strong> <strong><strong>for</strong>ward</strong> <strong>modelling</strong>, accuracy can be achieved through the use<br />
of high order spatial operators; <strong>for</strong> large problems this is crucial. In frequency <strong>domain</strong><br />
<strong>modelling</strong>, high order spatial nite dierence operators lead to large increases<br />
in computational costs that are not compensated <strong>for</strong> by the gain in accuracy (see<br />
chapters 2 and 3 <strong>for</strong> a full explanation). Instead we must seek other methods by<br />
which to increase the accuracy (and thereby reduce the costs).<br />
Two major improvements were implemented: The rst of these involved modications<br />
to speed up the numerical aspects of the matrix solver; the second involved<br />
modications to the numerical operators themselves. Together these two modications<br />
lead to a dramatic improvement in the computation times <strong>for</strong> acoustic wave<br />
equation <strong>modelling</strong> and inversion.<br />
These improvements were made use of in the<br />
inversion of a large, transmission <strong>seismic</strong> survey in which the use of extended parameter<br />
tests, involving a large number of <strong>modelling</strong> runs, was made possible by<br />
the improvements. During the inversion of the eld data, it became clear that the<br />
acoustic method needed to be replaced by an elastic method; in the nal stage of<br />
this project I developed extensions of the <strong>modelling</strong> methods to the elastic wave<br />
equation, paving the way <strong>for</strong> future elastic waveeld inversion methods.<br />
1.3 Forward <strong>modelling</strong> in the frequency-space <strong>domain</strong><br />
In this section the fundamental equations <strong>for</strong> <strong>seismic</strong> <strong>modelling</strong> in the frequencyspace<br />
<strong>domain</strong> are presented, and some of the basic considerations <strong>for</strong> frequency<strong>domain</strong><br />
methods are reviewed. I begin with the assumption that a particular wave<br />
equation has been selected <strong>for</strong> <strong>modelling</strong> purposes, and we have already discretized<br />
24
the partial dierential equations <strong>for</strong> numerical <strong>modelling</strong>. The discretized equations<br />
<strong>for</strong> the time <strong>domain</strong> acoustic or elastic wave equations using either a nite dierence<br />
or a nite element approach can be written as<br />
M ~<br />
~u(t)+ K ~<br />
~u(t)= ~ f(t) (1.1)<br />
(see <strong>for</strong> example Marfurt, 1984a), where ~u(t) is the discretized waveeld (i.e., the<br />
pressure, or the displacement) arranged as a column vector, M is the mass matrix,<br />
~<br />
K is the stiness matrix and ~ f(t) are the source terms, also arranged as a column<br />
~<br />
vector.<br />
Equation (1.1) can be approached in either the time <strong>domain</strong> or in the frequency<br />
<strong>domain</strong>.<br />
From this point on, this thesis is concerned with the frequency<br />
<strong>domain</strong> method <strong>for</strong> solving these problems. Taking the temporal Fourier trans<strong>for</strong>m<br />
of equation (1.1) yields<br />
where<br />
u(!) =<br />
Z 1<br />
,1<br />
K ~<br />
u(!) , ! 2 M ~<br />
u(!) =f(!) (1.2)<br />
~u(t)e ,i!t dt and f(!) =<br />
Z 1<br />
,1<br />
~f(t)e ,i!t dt (1.3)<br />
are Fourier trans<strong>for</strong>ms. If viscous damping is included, equation (1.2) becomes<br />
K (!) u(!)+i! C (!) u(!) , ! 2 M (!) u(!) =f(!) (1.4)<br />
~ ~ ~<br />
where C (!) is the damping matrix. Details of the nite-element and nite-dierence<br />
~<br />
approaches can be found in many textbooks (Zienkijevic, 1977; Bathe and Wilson,<br />
1976). In Chapter 3 and Chapter 5 I will give explicit <strong>for</strong>mulas <strong>for</strong> the matrix coecients<br />
<strong>for</strong> both the acoustic and elastic wave equations. The mass, stiness and<br />
damping matrices are computed by <strong>for</strong>ming a discrete representation of the underlying<br />
partial dierential equations and the physical parameters (<strong>for</strong> example, the<br />
<strong>seismic</strong> velocities, the bulk density and the attenuation parameters). For simplicity<br />
I rewrite equation (1.4) as<br />
S ~<br />
(!) u = f or u = S ~ ,1 (!) f (1.5)<br />
25
where the complex \impedance" matrix, S ~<br />
, is given by S ~<br />
(!) = K ~<br />
(!),! 2 M ~<br />
(!)+<br />
i! C ~<br />
(!). I shall refer to any <strong>modelling</strong> approach based on equation (1.5) as \<strong>Frequency</strong><br />
<strong>domain</strong> <strong>modelling</strong>". <strong>Frequency</strong> <strong>domain</strong> <strong>modelling</strong> is an implicit nite dierence<br />
method (Marfurt, 1984a); the second, explicit <strong>for</strong>m shown in equation (1.5) is<br />
only representational, as it is not generally possible (or desirable) to actually invert<br />
the very large impedance matrix S ~<br />
. Equation (1.5) is often solved using matrix<br />
factorisation methods, such as LU decomposition (Press et al., 1992; George and<br />
Liu, 1981; Pratt and Worthington, 1990).<br />
If LU decomposition is used to solve<br />
equation (1.5), the matrix factors can be re-used to solve the <strong><strong>for</strong>ward</strong> problem <strong>for</strong><br />
any new source vector, f extremely eciently. This point is especially important in<br />
the iterative solution of the inverse problem, in which many <strong><strong>for</strong>ward</strong> solutions, <strong>for</strong><br />
real sources and \virtual" sources will be required at each iteration. It is critical<br />
to use ordering schemes that allow maximum advantage to be taken of the sparsity<br />
of both S ~<br />
and its LU factorisation; nested dissection (George and Liu, 1981)<br />
is such a method. Later in this thesis I will explain this method and discuss the<br />
computational aspects that may aect the eciency.<br />
Inowintroduce a specic discretization, depicted in Figure 1.1, in which the<br />
waveeld is to be computed at n x n z nodal points on a regular grid (the grid is<br />
2 dimensional <strong>for</strong> illustration purposes, but could be 1, 2 or 3 dimensional). The<br />
model can be thought of as being specied at each of these node points.<br />
The waveeld vector, u and the source vector, f are (n x n z ) 1 column<br />
vectors; the complex impedance matrix, S is an (n x n z ) (n x n z ) matrix. All<br />
~<br />
quantities except the model parameters can take on complex values.<br />
Note that,<br />
although we will treat equation (1.5) as if it describes <strong><strong>for</strong>ward</strong> <strong>modelling</strong> <strong>for</strong> a single<br />
source position, additional source locations can be incorporated simply by increasing<br />
the number of elements in u by n x n z <strong>for</strong> each additional source; S ~<br />
and S ~<br />
,1<br />
then have block diagonal structures, in which each diagonal block is an identical<br />
submatrix.<br />
We could also feed in additional frequency components in the same<br />
26
Figure 1.1: A Discrete representation of the <strong><strong>for</strong>ward</strong> <strong>modelling</strong> problem. The representation<br />
is schematic; the assumption of two dimensions is not required at this<br />
stage, nor is this ordering of the node points necessary. The waveeld (either a<br />
scalar or a vector quantity) is sampled at each of the n x n z node points.<br />
manner, although the diagonal block submatrices of S ~<br />
are then no longer identical.<br />
The same comment applies to the 2:5 , D method of Song and Williamson (1995),<br />
in which a new diagonal block would be generated <strong>for</strong> each wavenumber considered.<br />
By examining the solutions to equation (1.5) when the components of the<br />
source vector, f i are replaced by a Kronecker delta, ij , it is clear that the columns<br />
of S ~ ,1 must contain the discrete approximations to the Green's functions. Thus,<br />
h<br />
S ,1 = g<br />
(1)<br />
g (2)<br />
~<br />
::: g (nxnz) i<br />
; (1.6)<br />
where the column vectors g (j) approximate the discretized Green's function <strong>for</strong> an<br />
impulse at the jth node. If the original physical problem is exactly reciprocal with<br />
,1<br />
respect to an interchange of source and receiver elements, then both S and S ~ ~<br />
must be symmetric (not Hermitian) matrices. [In implementation S is often not<br />
~<br />
perfectly symmetric when certain (unphysical) absorbing boundary conditions are<br />
implemented (Pratt and Worthington, 1990). This does not cause any problems].<br />
27
1.4 Fourier trans<strong>for</strong>ms and frequency <strong>domain</strong> <strong>modelling</strong><br />
An understanding of Fourier trans<strong>for</strong>ms and their properties is important <strong>for</strong><br />
frequency <strong>domain</strong> <strong>modelling</strong>. The continuous Fourier trans<strong>for</strong>m is dened in equation<br />
(1.3); <strong>for</strong> numerical computations this trans<strong>for</strong>m and its inverse are discretized,<br />
leading to a Discrete Fourier Trans<strong>for</strong>m (DFT), and its optimized implementation,<br />
the Fast Fourier Trans<strong>for</strong>m (FFT). The Fourier trans<strong>for</strong>m in exploration seismology<br />
is most commonly used to trans<strong>for</strong>m time <strong>domain</strong> data into the frequency <strong>domain</strong>,<br />
in order to apply a particular lter, following which an inverse Fourier trans<strong>for</strong>m is<br />
applied to bring the data back into the time <strong>domain</strong>. However, in frequency <strong>domain</strong><br />
<strong>modelling</strong> the rst step is not needed. We generate the components of the DFT of<br />
the data directly; if time <strong>domain</strong> results are required we obtain these by the inverse<br />
DFT, in which case sucient sampling in the frequency <strong>domain</strong> is required. Often,<br />
as we shall see, when solving the inverse problem we never need the time <strong>domain</strong><br />
data, and we need not be as concerned with sampling criteria. The following subsections<br />
will show the Fourier trans<strong>for</strong>m properties and explain the implications <strong>for</strong><br />
frequency <strong>domain</strong> <strong>modelling</strong>.<br />
1.4.1 Theory<br />
The Fourier trans<strong>for</strong>m, in essence, decomposes or separates a wave<strong>for</strong>m or<br />
function into sinusoids of dierent frequencies, which sum to yield the original wave<strong>for</strong>m.<br />
In frequency <strong>domain</strong> <strong>modelling</strong> we use a monofrequency component of the<br />
source to produce a monofrequency response at the receiver points. By per<strong>for</strong>ming<br />
an inverse Fourier trans<strong>for</strong>m of the monofrequency responses we are able to produce<br />
a required time <strong>domain</strong> response at the receiver locations. The <strong><strong>for</strong>ward</strong> and inverse<br />
DFT pairs <strong>for</strong> a time series h and a frequency series H are dened as (Hatton et al.,<br />
1986)<br />
H k = 1 N<br />
N,1<br />
X<br />
r=0<br />
h r e ,i2kr=N ; (1.7)<br />
28
<strong>for</strong> k = 0; 1;:::;N ,1 and<br />
h r =<br />
N,1<br />
X<br />
k=0<br />
H k e i2kr=N ; (1.8)<br />
<strong>for</strong> r =0;1;:::;N,1, where r is a time sample index, k is a frequency <strong>domain</strong> sample<br />
index, H k is the k-th Fourier trans<strong>for</strong>m coecient, h r is the time series. Provided<br />
each representation is complete (the time series or the frequency components), each<br />
series can be uniquely recovered from the other, using these <strong>for</strong>mulas.<br />
1.4.2 Sampling and the Sampling Theorem<br />
As we actually work with a discrete representation h n = h(t n ). The function<br />
h(t) is said to be band limited if its Fourier trans<strong>for</strong>m H(f) = 0 <strong>for</strong> jfj > f c<br />
where f c is a nite \critical" frequency. In <strong>seismic</strong> case all signals are band limited<br />
due to a limited source spectrum, and are almost always treated with an analogue<br />
\anti-alias" lter to ensure this property be<strong>for</strong>e sampling.<br />
The sampling theorem states that a band-limited function h(t) is completely<br />
specied by the sampled values f n (t n ), provided that the sampling interval, t<br />
satises<br />
f c 1 N y<br />
= 1<br />
2t<br />
(1.9)<br />
The frequency N y is known as the Nyquist frequency <strong>for</strong> the given sampling interval<br />
t.<br />
Beyond the Nyquist frequency, the periodicity and the conjugate symmetry<br />
of the DFT (<strong>for</strong> a real valued time series) causes the highest frequencies to be<br />
\wrapped" around the frequency axis and to be aliased as lower frequencies. This<br />
theorem is important if we are attempting to construct an un-aliased frequency<br />
spectrum from a time series.<br />
A similar sampling theorem is relevant ifweare trying to reconstruct a time<br />
series from a limited number of samples of the frequency spectrum, as is the case <strong>for</strong><br />
frequency <strong>domain</strong> <strong>modelling</strong>. We must sample the frequency spectrum suciently<br />
29
in order to unambiguously reconstruct the time series <strong>for</strong> the required length of<br />
time. If we again assume the time series is real valued, and make use of the resultant<br />
conjugate symmetry in the frequency spectrum, then the time series up to a<br />
maximum time, t max is completely specied by the sampled values, provided that<br />
the frequency sampling interval, f satises<br />
f 1<br />
t max<br />
: (1.10)<br />
This <strong>for</strong>mula assumes that the time series is completely causal, i.e., that the time<br />
series is equal to zero <strong>for</strong> all negative values of time. If this is not the case, then<br />
an additional factor of two must incorporated into the denominator. Provided the<br />
frequency sampling criteria is met, the time function may be reconstructed with any<br />
desired time sampling, t until the maximum time, t max .<br />
If the frequency sampling criteria above is not satised, then the periodicityof<br />
the inverse DFT causes the time samples <strong>for</strong> times greater than t max to be wrapped<br />
around the time axis, and to appear as if these were early time samples (i.e., these<br />
are aliased in time). Thus it is important that the model be designed in such a<br />
fashion as to prevent the simulation of any arrivals later than t max . Naturally this<br />
is not always possible; <strong>for</strong>tunately a trick exists that can be made use of to inhibit<br />
time aliased signals.<br />
1.4.3 Anti time-aliasing<br />
The technique <strong>for</strong> anti time-aliasing frequency <strong>domain</strong> <strong>modelling</strong> results has<br />
been described by (Subhashis and Frazer, 1987).<br />
Due to the periodicity in any<br />
Fourier series, the inverse DFT returns not a non periodic h(t), but periodic<br />
1X<br />
n=,1<br />
h(t + nt max ): (1.11)<br />
Thus, a time series which is non-zero <strong>for</strong> times greater than t max will be corrupted.<br />
To prevent this we can compute F (!+i) instead of F (!) where is an appropriate,<br />
30
small real number. This computation is easily implemented in frequency <strong>domain</strong><br />
<strong>modelling</strong>, and it has the advantage of yielding, after the inverse DFT, the time<br />
function<br />
1X<br />
n=,1<br />
h(t + nt max )e ,(t+ntmax) : (1.12)<br />
Thus, by using a complex value <strong>for</strong> the frequency, the time function has been eectively<br />
multiplied by a decaying exponential function. Each successive alias component<br />
of the time function is multiplied by a smaller value. To recover an approximation<br />
of the original, desired function we multiply this result by e t and produce<br />
the result<br />
1X<br />
n=,1<br />
h(t + nt max )e ,ntmax : (1.13)<br />
For the orginal, unaliased component (n = 0), the original signal is recovered. For<br />
all positive values of n, the signal is attenuated by an ever smaller factor { the<br />
aliased signal is still there, but it is attenuated. The method fails if the time series<br />
has non-zero values <strong>for</strong> negative times (n
If the <strong><strong>for</strong>ward</strong> Fourier trans<strong>for</strong>m is dened by H(!) =<br />
1 ,1 h(t)e i!t dt, then<br />
Z 1<br />
,1 h(t , x=c)e,i!t dt =<br />
Z 1<br />
,1 h()e,i!(t+x=c) d<br />
= e ,i!x=c Z 1<br />
,1<br />
h()e ,i! d<br />
= e ,i!x=c H(!): (1.15)<br />
The quantity x=c is a simple time shift, and thus we make use of the shifting property<br />
of the Fourier trans<strong>for</strong>m (in which the dummy integration variable is changed from<br />
t to ). Equation (1.15) shows that we can move the time window of interest by<br />
multiplying the frequency <strong>domain</strong> result by e ,i!x=c be<strong>for</strong>e per<strong>for</strong>ming the inverse<br />
DFT. This is an important result in frequency <strong>domain</strong> <strong>modelling</strong>, since we often<br />
need to simulate the <strong>seismic</strong> time <strong>domain</strong> response at far oset receivers in large<br />
experiments.<br />
Let us take <strong>for</strong> example a source-receiver oset of 300 km. In this case the<br />
full wave propagation time can be up to 30 to 40 seconds, while the required signal<br />
may be only a few seconds long. Due to the shifting property above, we have the<br />
option in frequency <strong>domain</strong> <strong>modelling</strong> to compute only a ve second time window,<br />
from, e.g., 35 to 40 seconds. We do this by setting the time shift equal to 35 seconds<br />
in equation (1.15). A ve second window can be completely represented using a<br />
frequency sample interval of f =1=5=0:2 Hz. This should be compared with the<br />
frequency sample interval required <strong>for</strong> the full 40 second record, f =1=40 = 0:025<br />
Hz. The number of samples required <strong>for</strong> a given maximum frequency is reduced by<br />
87.5%. In the time <strong>domain</strong>, it would be necessary to simulate the full 40 seconds in<br />
order to generate the same, nal, ve seconds of useful data.<br />
In this way we may directly calculate the time <strong>domain</strong> data in reduced time,<br />
in order to decrease the number of frequencies required <strong>for</strong> <strong><strong>for</strong>ward</strong> <strong>modelling</strong>. If<br />
reduced time output is used in conjunction with anti time-aliasing, it is important<br />
to ensure that the rst output time sample occurs be<strong>for</strong>e the rst data arrival. This<br />
is becaues non-zero signal arriving be<strong>for</strong>e the desired time window begins will be<br />
32
amplied instead of attenuated.<br />
1.5 Overview of chapters in this thesis<br />
This thesis begins with a discussion of the matrix solver used to generate<br />
solutions to the frequency <strong>domain</strong> nite dierence matrix, equation (1.5). If the<br />
matrix solver is inecient, no matter how good the nite dierence <strong>for</strong>mulation<br />
is, the costs involved will be prohibitive.<br />
In Chapter 2 I will initially dene the<br />
requirements expected from ecient matrix solvers. I will provide an analysis of<br />
the structure of the matrix, and the manner in which this structure aects the<br />
matrix solver in general, and the eects that various nite dierence operators will<br />
have on this structure.<br />
This will enable guidelines to be set <strong>for</strong> the appropriate<br />
nite dierence operators in order that the computational costs can be kept low.<br />
The technique of nested dissection, which optimises the initial sparsity pattern will<br />
be described and quantitative estimates of the computation times and the storage<br />
requirements will be given.<br />
After analysing the general problem of the matrix solver, I will move on to<br />
a specic nite dierence technique <strong>for</strong> visco-acoustic media in Chapter 3. I will<br />
use the rotated nite dierence operators suggested by Jo et al. (1996). I will<br />
present and analyze these operators, and then extend them to the heterogeneous<br />
case. I will then discuss the combined use of the nested dissection method and the<br />
implementation of the rotated nite dierence operators and prove that the scheme<br />
is optimal, and that no improvement will be achieved by the use of higher order<br />
spatial operators. Chapter 3 concludes with an example (based on a real wide-angle<br />
experiment) that demonstrates the visco-acoustic <strong>modelling</strong> described and analyzed<br />
in these initial chapters.<br />
In Chapter 4 the application of the frequency <strong>domain</strong> visco-acoustic <strong>modelling</strong><br />
scheme as a <strong>tool</strong> <strong>for</strong> wave<strong>for</strong>m inversion will be presented. The example presented is<br />
33
ased on data setfrom an underground laboratory in a crystalline rocks. The data<br />
suer from signicant noise problems. I will describe a pre-processing ow used to<br />
deal with these data problems. A way of determining the correct parameters based<br />
on the level of data residuals to obtain an optimal image will be presented. Potential<br />
anisotropy eect on the image will be discussed with the procedure, base on data<br />
residuals, <strong>for</strong> minimizing the imaging artefacts when present.<br />
A complete set of<br />
tests <strong>for</strong> selecting the inversion parameters is made possible by the improvements in<br />
eciency presented in Chapters 2 and 3. I will show evidence <strong>for</strong> spatial variation<br />
of the anisotropy; an eect that cannot be properly modelled or inverted using the<br />
visco-acoustic method.<br />
As a result of the conclusions of Chapter 4, in Chapter 5 I develop a viscoelastic<br />
<strong>modelling</strong> scheme, as a rst step toward the development of a fully anisotropic,<br />
visco-elastic <strong>modelling</strong> and inversion scheme. In developing these scheme, I begin<br />
by dening the rotated nite dierence operators required <strong>for</strong> the visco-elastic wave<br />
equation. A full description of the visco-elastic scheme, including a dispersion analysis,<br />
will be presented. An analytical proof that the scheme can work in the uid case<br />
will be given. As an example a cross-borehole data set from the Imperial College<br />
test site will be shown and compared with the visco-acoustic <strong>modelling</strong> results.<br />
In Chapter 6 I summarize the developments presented in the thesis and<br />
present my conclusions. For some of the models I present in the thesis, a reduction<br />
of over 90% in computational requirements, in comparison with the original simple<br />
<strong>modelling</strong> techniques, have been acheived, in both the visco-acoustic and the<br />
visco-elastic <strong>modelling</strong> cases. This has been achieved through the use of a fully integrated<br />
approach, in which I concentrated on all aspects of the <strong>modelling</strong> procedure<br />
| optimization of each individual aspect of <strong>modelling</strong> technique separately is not<br />
enough. The thesis concludes with several indications as to where possible further<br />
work could be concentrated, to allow the extension of these results to more complex<br />
data examples.<br />
34
Chapter 2<br />
Solving frequency <strong>domain</strong> wave equations:<br />
Numerical Considerations<br />
2.1 Introduction<br />
Seismic <strong><strong>for</strong>ward</strong> <strong>modelling</strong> can be <strong>for</strong>mulated as a time <strong>domain</strong> inital value<br />
problem or as a frequency <strong>domain</strong> boundary value problem (see Chapter 1 equation<br />
(1.5)). Explicit initial value problems do not require a large amount of memory to<br />
run, however the amount of computational time can be signicant if the number of<br />
time steps or the number of sources is large. The numerical solution of boundary<br />
value problems involve solving a large (usually sparse) system of linear equations<br />
(i.e., the matrix S ~<br />
in equation (1.5)). The cost of solving the system increases dramatically<br />
as the number of equations increases. To per<strong>for</strong>m full matrix inversion,<br />
or Gaussian elimination on a large system of linear equations requires a signicant<br />
amount of memory and CPU time. However, <strong>for</strong> sparse systems, savings can be<br />
obtained by exploiting the sparsity, and further savings are realized when a large<br />
number of right hand sides are involved (representing additional sources in the <strong>seismic</strong><br />
<strong>modelling</strong> case). The utility of dealing with multiple right hand sides is critical<br />
in <strong>seismic</strong> inverse problems, in which only a limited number of frequencies <strong>for</strong> a<br />
large number of sources may be required (Pratt and Worthington, 1990). This is<br />
35
there<strong>for</strong>e one of the main applications of frequency-<strong>domain</strong> <strong>seismic</strong> <strong><strong>for</strong>ward</strong> <strong>modelling</strong>.<br />
To solve a large system of linear equations eciently one has to consider the<br />
detailed numerical properties of the problem and use them to the full extent tokeep<br />
overheads as low aspossible.<br />
In this chapter I will consider the characteristics of the frequency <strong>domain</strong><br />
<strong>seismic</strong> <strong><strong>for</strong>ward</strong> <strong>modelling</strong> problem in the case of multiple source experiments, and<br />
develop the appropriate matrix description. I will then move on to a consideration<br />
of the characteristics of the nite dierence operator required to generate solutions<br />
at minimum computational cost. These characteristics will be utilized in chapters 3<br />
and and 5todevelop optimal operators.<br />
2.2 Solving linear equation systems: bottlenecks<br />
As shown in Chapter 1, frequency <strong>domain</strong> <strong><strong>for</strong>ward</strong> <strong>modelling</strong> requires a solution<br />
to a system of linear equations (see equation 1.5). In 2-D, to nd the solution<br />
(the eld vector u), one has to solve a linear system of n x n z<br />
equations with<br />
n x n z unknowns (where n x and n z are the number of grid points in the x and<br />
z directions in the model).<br />
If the problem is elastic, each eld component is a<br />
two-component vector, and the total number of equations is doubled.<br />
Although<br />
conceptually straight<strong><strong>for</strong>ward</strong>, the computational costs involved in Gaussian elimination<br />
or matrix inversion can become prohibitive when the problem size (n x<br />
n z<br />
)<br />
increases due to a cubic (O(n x n z min(n x ;n z ))) growth in memory requirements.<br />
In order to decrease the computational costs involved one has to consider the properties<br />
of the matrix<br />
S ~<br />
and of the underlying physical problem, be<strong>for</strong>e developing<br />
an appropriate matrix solver.<br />
The requirements I am going to consider in this chapter include the following:<br />
The problem must be ecently solved <strong>for</strong> mutiple right hand sides (multiple sources),<br />
the matrix solver must be computationally ecient and use a minimum of physical<br />
36
(RAM) memory, andtheunderlying numerical approximation must betuned tothe<br />
matrix solver <strong>for</strong> minimum overall computational costs. These problems must all be<br />
considered together, as the each choice at each stage aects the choice, at the next<br />
stage.<br />
Iterative solvers are usually considered to be the best way of solving positive<br />
denite linear systems (<strong>for</strong> a denition of positive denite equations see <strong>for</strong> example<br />
George and Liu (1981)). The main advantage of iterative matrix solvers is that full<br />
advantage can be taken of the initial matrix sparsity. As a result, the amount of<br />
memory required is small (of the order of n x n z ). The problem with frequency<strong>domain</strong><br />
<strong><strong>for</strong>ward</strong> <strong>modelling</strong> of the <strong>seismic</strong> problem is that the matrices arising from<br />
the nite dierence equations are not always positive denite.<br />
For example, the<br />
absorbing boundary conditions often used are not physical - they are used only<br />
because we are attempting to model innite media by using a nite model. As a<br />
result, due to ill conditioning, iterative methods either do not converge, or converge<br />
too slowly to be considered appropriate to solve the system. The other problem with<br />
iterative solvers is that they are not suitable <strong>for</strong> systems with multiple right hand<br />
sides. The computational costs <strong>for</strong> iterative solvers increases in linear proportion to<br />
the number of right hand sides. In the problems I am going consider, the number<br />
of right-hand sides can be signicant. Direct methods, which are able to solve the<br />
problem eciently <strong>for</strong> multiple right hand sides, are there<strong>for</strong>e more eective than<br />
the iterative ones in this case.<br />
Direct methods <strong>for</strong> solving linear equation systems require signicantphysical<br />
memory.<br />
Matrices produced by nite-dierence (or nite-elements) methods are<br />
always sparse, but the sparsity pattern is not preserved by most direct matrix solvers.<br />
The sparsity pattern of the initial matrix depends on the nite dierence operator<br />
used and on the grid ordering used. In this chapter I will initially concentrate on<br />
the eect of grid ordering, and then move on to consider the eect of the size (i.e.,<br />
the order) of the nite dierence operator on the memory requirements.<br />
37
2.3 Solving linear equation systems with multiple right hand<br />
sides<br />
An eective direct method <strong>for</strong> solving a system of linear equations with the<br />
multiple right hand sides is LU decomposition, which trans<strong>for</strong>ms a system :<br />
S ~<br />
u = f (2.1)<br />
into the system<br />
L ~<br />
U ~<br />
u = f; (2.2)<br />
where matrices L ~<br />
and<br />
U ~<br />
are lower and upper triangular matrices. LU decomposition<br />
inevitably destroys some of the sparsity of the original sparse matrix through<br />
matrix \ll in" (not a big problem if the matrix is dense); in section 2.5 I discuss how<br />
this ll in is minimised. The solution can then be eciently obtained by per<strong>for</strong>ming<br />
the following set of Gaussian eliminations:<br />
L ~<br />
u 0 = f (2.3)<br />
(<strong><strong>for</strong>ward</strong> reduction) and<br />
U ~<br />
u = u 0 (2.4)<br />
(back substitution). Due to the fact that L ~<br />
and U ~<br />
are triangular this procedure is<br />
simple and there is no additional ll in suered by these eliminations. The number<br />
of operations is in direct proportion to the number of non-zero elements in L and<br />
~<br />
U . If an additional result is required <strong>for</strong> a new right hand side vector, f 0 , then the<br />
~<br />
same cheap <strong><strong>for</strong>ward</strong> and back substitution procedure can be repeated with f 0<br />
as the<br />
right hand side in the equation (2.3), using the original LU factors.<br />
Matrices generated from nite dierence (or nite element) equations are<br />
usually well structured if simple grid ordering is used. I will concentrate initially on<br />
the simple row ordering of the nodes shown on Figure (1.1). I will refer to this later<br />
38
as sequential grid ordering. Sequential ordering just involves starting <strong>for</strong> example,<br />
in the top left corner of the grid and numbering the grid points in the rst row<br />
sequentially up to n x (where n x is number of grid points in x direction). We then<br />
move to the next row, repeat the procedure, and continue in this manner until we<br />
run out of grid points. Imagine our problem is dened on a grid of n x by n z nodes:<br />
If each node is coupled only to it's immediate neighbours (as in nite dierence<br />
equations arising from second order nite dierence operators), the initial matrix<br />
S ~<br />
will only have non-zero elements on the main diagonal, on the two neighbouring<br />
sub-diagonals, and on two sub-diagonal bands at a distance of n z diagonals away<br />
from the main diagonal.<br />
In general any nite dierence operator will produce a<br />
symmetric sparsity pattern in the initial matrix<br />
S ~<br />
. This does not imply that the<br />
matrix itself will be symmetric. This depends on the boundary conditions and on<br />
the type of nite dierence operators used.<br />
Now let us examine the way in which LU decomposition can be per<strong>for</strong>med.<br />
The algorithm is relatively simple (more details can be found, <strong>for</strong> example, in Peres<br />
et al. (1992)): Let i;j be elements of the starting matrix S ~<br />
, i;j be elements of<br />
matrix L ~<br />
and i;j be elements of matrix U ~<br />
. The algorithm proceeds as follows.<br />
Set i;i =1<br />
For each j =1;2;:::;N carry out the following two procedures:<br />
First, <strong>for</strong> i =1;:::;j<br />
i;j = i;j ,<br />
Second, <strong>for</strong> i = j +1;:::;N<br />
Xi,1<br />
k=1<br />
i;k k;j : (2.5)<br />
0<br />
1<br />
i;j = 1<br />
j,1<br />
X<br />
@ i;j , i;k k;j<br />
A : (2.6)<br />
j;j<br />
k=1<br />
Once the element a i;j is used the value is not required any more, so the same memory<br />
location can be used to store the corresponding i;j or i;j . Values i;j and i;j<br />
39
are always calculated by the time they are needed to calculate next values. The<br />
diagonal unity elements i;i =1need not be stored at all. From this description of<br />
the algorithm one can see that all the elements between the rst physical non-zero in<br />
the lower triangular part of S and the main diagonal on the same row will become<br />
~<br />
non-zero elements in L , while all the elements from the rst physical non-zero in the<br />
~<br />
upper triangular part of S and the main diagonal on the same column will become<br />
~<br />
non-zero elements in<br />
U ~<br />
matrix. Elements outside this band remain logically zero<br />
and need not be stored.<br />
Sequential ordering of the grid is the most natural way of grid ordering and<br />
it is the easiest to implement. The simple matrix structure ts well into ordinary<br />
array variables available in almost any programming language, and no overhead is<br />
needed to describe the matrix structure. However, as I will show sequential grid<br />
ordering requires too much memory in comparison with alternative grid ordering<br />
schemes.<br />
It is relatively easy to predict the number of non-zero elements in the matrices<br />
L ~<br />
and U ~<br />
in this case. They have approximately rectangular regions which are lled<br />
in with non-zero elements. The number of matrix rows is n x n z and the bandwidth is<br />
n z . Thus the number of elements in L ~<br />
and U ~<br />
matrix is approximately 2n x n z (n z +1)<br />
in the case of sequential row ordering. In the case of n x = n z = n the memory<br />
required is the order of n 3 , or O(n 3 ), where n is number of grid points along one<br />
edge.<br />
One can thus see that care has to be taken whether the row or column<br />
ordering is used, due to the fact that one of the grid dimensions inuences the<br />
memory required by O(n 2 ) while the other is only O(n). The memory capacity of<br />
commonly available systems is of the order of 1GB 10 9 B (where B means bytes).<br />
If one can store a complex number in 8B then the maximum (square) problem size<br />
will be of order 400 by 400 grid points. If one assumes 10 grid points per wavelength<br />
this will imply that 40 wavelengths in both directions will be the maximum model<br />
size.<br />
40
2.4 Matrix \ll in" and ordering schemes<br />
Here I will show that the same system of linear equations can produce high ll<br />
in or no ll in at all depending on the grid ordering used. An example (see George<br />
and Liu, 1981 ) will be used to show this extreme case. Consider the following two<br />
matrices, both representing the same equation system:<br />
Case 1<br />
2<br />
S =<br />
~ 6<br />
4<br />
4 1 2 .5 2<br />
1 .5 0 0 0<br />
2 0 3 0 0<br />
.5 0 0 .625 0<br />
2 0 0 0 16<br />
3<br />
7<br />
5<br />
Case 2<br />
2<br />
S =<br />
~ 6<br />
4<br />
16 0 0 0 2<br />
0 .625 0 0 .5<br />
0 0 3 0 2<br />
0 0 0 .5 1<br />
2 .5 2 1 4<br />
3<br />
7<br />
5<br />
After per<strong>for</strong>ming LU decomposition on these matrices the following matrices are<br />
obtained:<br />
Case 1<br />
2<br />
S =<br />
~ 6<br />
4<br />
2 .5 1 .25 1<br />
.5 .5 -1 -.25 -1<br />
1 -1 1 -.5 -2<br />
.25 -.25 -.5 .5 -3<br />
1 -1 -2 -3 1<br />
3<br />
7<br />
5<br />
41
Case 2<br />
2<br />
S =<br />
~ 6<br />
4<br />
4 0 0 0 .5<br />
0 .791 0 0 .632<br />
0 0 1.73 0 1.15<br />
0 0 0 .707 1.41<br />
.5 .632 1.15 1.41 .129<br />
3<br />
7<br />
5<br />
Case 2 needs 13 memory locations to store the non-zero results of LU decomposition,<br />
while case 1 needs 25 memory locations (almost twice as much). The<br />
linear systems are exactly the same, except that the variables have been reordered.<br />
The clear conclusion from this example is that care in the ordering of equations can<br />
keep the computational costs and memory requirements low. Now I will move on<br />
to show the way in which I will reorder the nite dierence grid nodes so that the<br />
resulting matrix suers the minimal ll in.<br />
2.5 Nested dissection ordering<br />
In this section I will discuss the optimal way of trans<strong>for</strong>ming the system<br />
into a system<br />
S ~<br />
u = f (2.7)<br />
( P ~<br />
S ~<br />
P ~<br />
t<br />
)( P ~<br />
u)= P ~<br />
f; (2.8)<br />
where the matrix P ~<br />
is a permutation operator which will trans<strong>for</strong>m the matrix S ~<br />
in such a manner as to ensure that the L ~<br />
and U ~<br />
matrices have the lowest number<br />
of non-zero elements. The grid reordering I will use is known as \nested dissection"<br />
and is explained in detail by George and Liu (1981). The equivalent matrices (be<strong>for</strong>e<br />
and after LU decomposition) <strong>for</strong> a sequentially ordered grid, and <strong>for</strong> a grid ordered<br />
using nested dissection are shown on Figure 2.1. The same approach <strong>for</strong> grid ordering<br />
is used by Marfurt et al. (1987) to decrease memory requirements <strong>for</strong> frequency<strong>domain</strong><br />
<strong>seismic</strong> <strong><strong>for</strong>ward</strong> <strong>modelling</strong>.<br />
42
(a)<br />
(b)<br />
(c)<br />
(d)<br />
Figure 2.1: Nested dissection versus sequentially ordered matrix a),b) be<strong>for</strong>e LU<br />
decomposition, and c),d) equivalent L matrix after LU decomposition (George and<br />
Liu,1981). Only non-zero elements are shown in each case. a) Matrix S <strong>for</strong> a sequentially<br />
ordered grid. b) Matrix S <strong>for</strong> a grid ordered using nested dissection. c) L part<br />
of the LU decomposed matrix S <strong>for</strong> case a) (memory required is O(n 3 )). d) L part<br />
of the LU decomposed matrix S <strong>for</strong> case b) (memory required is O(n 2 log(n))). The<br />
memory required to store matrix <strong>for</strong> a realistic value of n on gure d) is signicantly<br />
lower than the one required <strong>for</strong> the matrix on gure c).<br />
43
Figure 2.2: Two-way dissected nite dierence grid. The two way dissector, 5 (in<br />
black) is the last part of the grid to be ordered.<br />
Let us assume initially that n x = n z = n (i.e., that the grid is square).<br />
The grid is then dissected into four quarters so that there are approximately n 2 =4<br />
elements in each part of the dissected grid. Each of the four sections are are coupled<br />
only through the dissectors and within themself (see Figure 2.2). The minimal twoway<br />
dissector has to have at least approximately 2n elements. For the moment I<br />
will assume that it is possible to nd a dissector of this size (this is actually the case<br />
if second order nite dierences are used). The nested dissection recipe <strong>for</strong> ordering<br />
the elements is to rst number the elements in each block of the nite dierence<br />
grid, and then to number the elements in the dissector. The matrices S ~<br />
, L ~<br />
and U ~<br />
are shown schematically on Figure 2.3. This procedure is called two-way dissection.<br />
If one continues with the procedure recursively on all parts of the dissected matrix<br />
the result is called \nested dissection".<br />
Now let us consider the memory requirements necessary to store the non-zero<br />
elements of the matrices when per<strong>for</strong>ming LU decomposition on the n by n grid<br />
ordered by nested dissection. From Figure 2.3 one can see that the memory can be<br />
divided into ve parts. The nal part, L 5;5 is the memory necessary to per<strong>for</strong>m LU<br />
decomposition on the dissector itself, while the remaining four parts L 5;i and L i;i are<br />
44
U 11<br />
0<br />
U 15<br />
L 11<br />
U 55<br />
U 22<br />
U 25<br />
L 22<br />
L 33<br />
L 44<br />
U 33<br />
0<br />
U 35<br />
U 44<br />
L 15 L 25 L 35 L 45<br />
L 55<br />
U 45<br />
Figure 2.3: Two way dissected matrix S ~<br />
= L ~<br />
U ~<br />
. During LU decomposition the<br />
values <strong>for</strong> L i;j and U i;j are lled in at the corresponding locations used by S i;j . L i;j<br />
and U i;j denotes possible non-zero elements in matrices L ~<br />
and U ~<br />
respectively after<br />
LU decomposition while 0 denotes zero elements.<br />
the amounts necessary to per<strong>for</strong>m LU decomposition on the n 2<br />
by n 2 grids (L i;i), plus<br />
L 5;i which comes from the coupling between the elements within each subgrid and<br />
the elements within the dissector. In the rst dissection one can write the memory<br />
requirements as:<br />
S(n; 0) = 4S(n=2; 2) + D(n; 0) (2.9)<br />
where S(i; j) represents memory requirement <strong>for</strong> the subgrid of size i bordered by<br />
n<br />
n<br />
n<br />
n<br />
n<br />
n<br />
S(n,2) S(n,3) S(n,4)<br />
Figure 2.4: All possible subgrid (S(n; 2);S(n; 3) and S(n; 4)) situations arising during<br />
nested dissection. The thick black borders represent neighbouring dissectors<br />
from previous dissections in the recursion.<br />
45
(n/2*n/2)/2<br />
L 55<br />
(n/2*n/2)/2<br />
n*n/2<br />
n*n/2<br />
(n*n)/2<br />
Figure 2.5: Enlarged L 5;5 part of the two way dissected matrix. Non zero elements<br />
are in grey. White space represents logical zero elements.<br />
dissectors at j sides (Figure 2.4) (L i;i +L 5;i in Figure 2.3), while D(i; j) is the memory<br />
required to per<strong>for</strong>m LU decomposition on the dissector itself, which is coupled to<br />
j parts of the other dissectors. By continuing the dissection one will nd that only<br />
two more situations can occur: S(n; 3) and S(n; 4) as shown on Figure 2.4. So we<br />
obtain the following equations, together with equation 2.9:<br />
S(n; 2) = S(n=2; 2)+2S(n=2; 3) + S(n=2; 4) + D(n; 2) (2.10)<br />
S(n; 3) = 2S(n=2; 3) + S(n=2; 4) + D(n; 3) (2.11)<br />
S(n; 4) = 4S(n=2; 4) + D(n; 4) (2.12)<br />
From here on I will concentrate on the memory necessary to store only the L ~<br />
part of the matrix; the full amount is just twice the values I will derive. I will start<br />
with D(n; 0) = L 5;5 from Figure 2.3. If the dissectors are ordered sequentially the<br />
enlarged part L 5;5 of the matrix L ~<br />
from Figure 2.3 will look like the one shown on<br />
Figure 2.5. Here I consider the worst possible case in which all the last n elements<br />
inatwoway dissector are coupled to each other, and that both n=2 sized dissectors<br />
are related to all n elements in the n sized dissector. With these considerations one<br />
46
can write directly from the Figure 2.5:<br />
<br />
D(n; 0) n 2 =2+2(n=2) 2 =2+2 n n 1<br />
<br />
=n 2<br />
2 2 +1 4 +1 = 7 4 n2 : (2.13)<br />
In a similar manner the following equations can be derived:<br />
D(n; 2) 19 4 n2 (2.14)<br />
D(n; 3) 25 4 n2 (2.15)<br />
D(n; 4) 31 4 n2 (2.16)<br />
and equation 2.12 can be expanded in the following <strong>for</strong>m using 2.16:<br />
S(n; 4) 31 31<br />
<br />
4 n2 +4<br />
4 (n=2)2 +4S(n=4; 4) =<br />
31<br />
4 n2 (1+1)+16S(n=4; 4) =<br />
:::= 31<br />
X<br />
4 n2 log 2 (n)<br />
i=1<br />
1 = 31<br />
4 n2 log 2 (n): (2.17)<br />
Substituting this into the equations (2.9) to (2.11) and using (2.13) to (2.15) the<br />
following expressions can be obtained:<br />
S(n; 3) 31 4 n2 log 2<br />
(n)+O(n 2 ) (2.18)<br />
S(n; 2) 31 4 n2 log 2<br />
(n)+O(n 2 ) (2.19)<br />
S(n; 0) 31 4 n2 log 2<br />
(n)+O(n 2 ) (2.20)<br />
which gives us a total memory requirementof 31<br />
2 n2 log 2<br />
(n) <strong>for</strong> the matrices L ~<br />
and U ~<br />
together. George and Liu (1981) have shown that the theoretical minimal memory<br />
requirements to per<strong>for</strong>m the LU decomposition on an n by n grid is of the same<br />
order of magnitude, so that nested dissection can there<strong>for</strong>e be assumed to be an<br />
\optimal" grid ordering to within, at least, an order of magnitude. George and Liu<br />
(1981) also showed that nested dissection gives an optimal number of operations<br />
((n; 0))<br />
(n; 0) 829<br />
84 n3 ; (2.21)<br />
47
(to within anorder of magnitude) necessary to per<strong>for</strong>m the LU decomposition.<br />
The amount of CPU time required to solve the system <strong>for</strong> each right hand<br />
side is again of order of n 2 log 2<br />
n (i.e. of the order of the number of elements in the<br />
LU decomposed matrix). This amount of CPU time can easily be less then <strong>for</strong> the<br />
iterative matrix solver where one needs at least n 2 operations per iteration and <strong>for</strong> a<br />
large n the number of iterations will almost certainly be greater than log 2<br />
(n). From<br />
this observation we see that <strong>for</strong> the numerical problems with the multiple right hand<br />
sides the nely tuned direct matrix solver may per<strong>for</strong>m better than the iterative one.<br />
There is an additional computational cost <strong>for</strong> LU decomposition that has not<br />
yet been mentioned. A certain amount of CPU time (a signicant one) is needed<br />
to generate the nested dissection ordering. The algorithm complexity necessary to<br />
dissect the matrix is of the order of O(n 4 ), however it is well worth the eort, as<br />
I will show in the following chapters, to generate nested dissection ordering. The<br />
same ordering can of course be used <strong>for</strong> all runs with any model of the same size. A<br />
second hidden cost is that the sparsity pattern of the LU decomposed matrix is far<br />
from simple, and a suitable pointing algorithm is required to track this sparsity. The<br />
memory requirements <strong>for</strong> this algorithm are the same as <strong>for</strong> the non-zero elements of<br />
the matrix, so that there is a linear increase in memory required. A certain amount<br />
of computing time is also lost during factorization on searching through the matrix<br />
structure to nd given matrix locations. In the case of sequential row and column<br />
access (as in LU decomposition) this is negligible.<br />
It is important to point out that I have not been limited to a particular partial<br />
dierential equation while working with nested dissection: The implementation<br />
depends only on the structure of the matrix. From now on all the developments<br />
on nite dierence methods <strong>for</strong> frequency <strong>domain</strong> <strong>seismic</strong> <strong><strong>for</strong>ward</strong> <strong>modelling</strong> will assume<br />
that the LU decomposition will be per<strong>for</strong>med on the grid ordered by the nested<br />
dissection and that the properties of the nite-dierence scheme will be adjusted to<br />
take the full advantage of the nested dissection ordering.<br />
48
Figure 2.6: Fourth order nite dierence computational star. The symbol identies<br />
those grid points coupled to the central grid point.<br />
2.6 Operators and memory requirements<br />
The memory requirement predictions in the previous section have assumed<br />
that one can nd a two way dissector of size 2n on an n by n grid. This assumption<br />
is valid provided second order nite dierence operators are used. If a fourth order<br />
nite dierence operator is used, this involves coupling of grid points at distances<br />
of 2 x and 2 z as shown on Figure 2.6. The minimal two way dissector size then<br />
increases to 4n. At rst sight this does not look like a big increase. However the<br />
memory requirements <strong>for</strong> nested dissection are highly dependent on the size of this<br />
dissector.<br />
If the dissector size is increased to 4n from 2n, how big will the impact be on<br />
the required memory? Let us return to Figure 2.5. In this case n becomes 2n so:<br />
49
D 4 (n; 0) (2n) 2 =2 + 2(2n=2) 2 =2+2 2n 2n 2<br />
= n 2 (2+1+4)<br />
= 7n 2 (2.22)<br />
<br />
In a similar manner the following equations can be derived:<br />
D 4 (n; 2)=23n 2 (2.23)<br />
D 4 (n; 3)=31n 2 (2.24)<br />
D 4 (n; 4) = 39n 2 : (2.25)<br />
Substituting these into the equations (2.9) to (2.12) <strong>for</strong> S(n; i) the required memory<br />
to per<strong>for</strong>m LU decomposition in this case will be:<br />
39<br />
<br />
S 4 (n; 4) = 4S 4 (n=2; 4) + D 4 (n; 4) = 39n +4 2 4 n2 +4S(n=4; 4)<br />
(2.26)<br />
39n 2 log 2 n (2.27)<br />
and similarly<br />
S 4 (n; 3) = 39n 2 log 2 n + O(n 2 ) (2.28)<br />
S 4 (n; 2) = 39n 2 log 2 n + O(n 2 ) (2.29)<br />
S 4 (n; 0) = 39n 2 log 2 n + O(n 2 ) (2.30)<br />
where S 4 (n; i) is the equivalent ofS(n; i) if the fourth order nite dierence scheme<br />
is used.<br />
Equations (2.20) and (2.30) show the memory required to per<strong>for</strong>m LU decomposition<br />
on an n by n grid if second and fourth order nite dierence operators<br />
are used, respectively. However, the use of higher order nite dierence operators<br />
reduces the required grid size (<strong>for</strong> a given accuracy).<br />
I will now show what the<br />
decrease in the number of grid points in one direction would have to be in order to<br />
50
educe the memory required to per<strong>for</strong>m LU decomposition. To show this one has<br />
to solve the following equation<br />
S(n 2 ; 0) = S 4 (n 4 ; 0); (2.31)<br />
<strong>for</strong> n 4 = kn 2 where k is the factor by which we have to reduce the number of grid<br />
points in one direction in order to at least equal the second order scheme with respect<br />
to the required memory. Here S 4 (n 4 ; 0) represents the memory required to per<strong>for</strong>m<br />
LU decomposition on the n 2 by 4 n2 (or 4 k2 n 2 by 2 k2 n 2 2<br />
) matrix generated by<br />
using 4th order nite dierence operator and S(n 2 ; 0) (as dened in equation 2.20)<br />
represents the memory required to per<strong>for</strong>m the LU decomposition of the n 2 2<br />
by n 2 2<br />
matrix generated by using second order nite dierence operators.<br />
If we equate<br />
equations 2.20 and 2.30:<br />
39n 2 4<br />
log 2<br />
(n 4 )+O(n 2 4)= 31 4 n2 2<br />
log 2<br />
(n 2 )+O(n 2 2);<br />
then<br />
39k 2 n 2 2 log 2(n 2 )+O(n 2 2 )=31 4 n2 2 log 2(n 2 )+O(n 2 2 ):<br />
This equality can be approximately expressed by discarding O(n 2 2) terms as:<br />
39k 2 n 2 log 2 2(n 2 ) = 31<br />
4 n2 log 2 2(n 2 )<br />
31<br />
k 2 n 2 2<br />
log 2<br />
(n 2 ) =<br />
39 4 n2 2<br />
log 2<br />
(n 2 )<br />
k 2 =<br />
31<br />
39 4<br />
k = :445 (2.32)<br />
This result shows that one would need to reduce the number of grid points per<br />
wavelength by more than 50% in order to justify the use of higher order nite<br />
dierence operators in a nested dissection ordered grid.<br />
In the case of sequential ordering a much smaller improvement will justify<br />
the higher degree operators due to the n 3 dependency of the memory requirements:<br />
51
The memory required to per<strong>for</strong>m LU decomposition on sequential n 4 by n 4 grid if<br />
the 4th order nite dierence operators are used is:<br />
S 4 (n) =4n 3 4<br />
(2.33)<br />
If this is compared with the second order scheme there is only a linear increase so:<br />
4n 3 4 =2(n 2) 3 (2.34)<br />
n 4<br />
n 2<br />
=( 1 2 )1 3 =:7937: (2.35)<br />
This shows that, <strong>for</strong> the sequential ordering scheme, a reduction of only 21% in<br />
number of grid points per wavelength will justify the use of a higher order nite<br />
dierence scheme. Nevertheless, the overall cost will be much higher than <strong>for</strong> the<br />
equivalent nested dissection scheme.<br />
As a comparison, <strong>for</strong> time <strong>domain</strong> schemes the best results are expected by<br />
using a staggered grid, with a fourth order nite dierence operator in space and<br />
a second order operator in time, as pointed out by Sei (1994a). This choice comes<br />
from the CPU time requirements versus accuracy <strong>for</strong> the time <strong>domain</strong> approach.<br />
The reason <strong>for</strong> this comparison between the second and the fourth order<br />
nite dierence schemes will become clear in the following chapters, when I will<br />
show that it is possible to develop second order nite dierence operators which will<br />
require only 4 grid points per wavelength to achieve high accuracy. Due to the fact<br />
that the theoretical limit <strong>for</strong> any nite dierence operator is two grid points per<br />
wavelength, I consider that no gains will be achieved if higher order nite dierence<br />
approximations are used.<br />
In any case, the use of local nite dierence operators<br />
improves per<strong>for</strong>mance in heterogenous media ( Ozdenvar and McMechan, 1996).<br />
2.7 Comparison of band and nested dissection ordering<br />
In this section the advantage of using nested dissection to per<strong>for</strong>m frequency<br />
<strong>domain</strong> <strong><strong>for</strong>ward</strong> <strong>modelling</strong> with realistic models will be demonstrated. Here I will<br />
52
consider rst a set of parameters <strong>for</strong> a realistic crosshole data set, using as an<br />
example a cross-borehole experiment described by Pratt and Sams (1996). In that<br />
experiment the following parameters apply:<br />
Source and receiver array length: 100 m<br />
Borehole separation: 100 m<br />
Minimum P-wave velocity: 2.5 km/s<br />
Data frequency: 1 kHz<br />
If one assumes that the required accuracy can be achieved by using 10 grid points per<br />
wavelength (which is consistent with the ordinary second order frequency <strong>domain</strong><br />
<strong>seismic</strong> <strong>modelling</strong> scheme accuracy), then the required nite dierence grid would<br />
have to be 400 grid points by 400 grid points, with x = z = :25m. If we need<br />
8 bytes to store a complex number, then a band ordered scheme will require in<br />
order of 1000 MB to store the LU decomposed matrix, whereas a nested dissection<br />
scheme will require 100 MB. This demonstrates dramatically the need <strong>for</strong> nested<br />
dissection methods. The situation becomes even more critical with larger and more<br />
general <strong>seismic</strong> experiments. In the following chapter I will show that far less than<br />
10 grid points per wavelength are actually required. If we use 4 grid points per<br />
wavelength, we will require only 13.5 MB to store the matrix. This represents an<br />
overall reduction of 98:65% from the initial gure of 1000 MB.<br />
The previous considerations were based on square models with n x = n z .<br />
However, useful geological models are not always square. Geophysical experiments<br />
usually have larger distances in one direction. The savings introduced by nested<br />
dissection are the highest in the n x = n z<br />
case and much less <strong>for</strong> a models where<br />
n x >> n z or n x
are of theorder of hundereds of kilometers (see (Holbrook et al., 1992)). In the case<br />
of wide angle experiments one records not only the reections from the impedance<br />
contrasts beneath the source, but also refracted arrivals which travel deep into the<br />
earth (up to 30 to 50 km) and turn back to the surface. The results <strong>for</strong> this numerical<br />
experiment are shown on Figure 2.7.<br />
For this aspect ratio, we found out that a<br />
nested dissection ordering allowed a grid with 5000 by 800 grid points to t within<br />
2 GB of memory. A sequentially ordered grid of the same size would require 50<br />
GB. All predictions assume that acoustic frequency <strong>domain</strong> <strong><strong>for</strong>ward</strong> <strong>modelling</strong> can<br />
be done with four grid points per wavelength, as will be explained in the following<br />
chapter. The frequency <strong>domain</strong> numerical simulation of such models (with hundreds<br />
of wavelengths propagation distance between the sources and receivers) would not be<br />
possible if simple grid ordering were used. Simulations of such experiments require<br />
huge computational resources (mainly CPU time) even <strong>for</strong> the time <strong>domain</strong> based<br />
schemes. This kind of experiment involves a large number of sources, and late arrival<br />
times, which consequently makes a time <strong>domain</strong> approach too expensive even with<br />
the fastest available computer resources.<br />
For example, if we assume a machine capable of one gigaop (where one<br />
gigaop is equal to one billion oating point operations per second) the following<br />
prediction is obtained: Let a model be dened by six hundred thousand grid points,<br />
and let ten oating point operations be required per grid point <strong>for</strong> one time step.<br />
We further assume that the maximal time step is 0:0001 second, that the maximal<br />
required simulation time is 40 seconds and that the number of sources is 150. The<br />
approximate CPU time under these conditions will be four days. For comparison,<br />
a similar computation in the frequency <strong>domain</strong> can be carried out within ten hours<br />
on Digital alpha 600/333 workstation (104 megaops) with 256 MB of RAM, if the<br />
data are generated in reduced time (see the example in the following chapter). On a<br />
one gigaop machine this calculation would take only one hour. Most importantly,<br />
if additional sources responses were required, these could be computed in a trivial<br />
54
1x10 11<br />
Memory (BYTES)<br />
1x10 10<br />
1x10 9<br />
1x10 8<br />
Sequential Ordering<br />
Nested dissection<br />
2GB<br />
Actual Mesh Size 5000X800<br />
50GB<br />
1x10 7<br />
500 1000 2000 5000 10000<br />
n x<br />
Figure 2.7: Memory requirements comparison <strong>for</strong> n x = 6:25 n z in case of band<br />
and nested dissection ordering. The required mesh size represents the model size<br />
necessary to per<strong>for</strong>m acoustic <strong>modelling</strong> of a wide angle experiment with 10 Hz data<br />
and a model 350 km by 48 km. The minimum P wave velocity is 2.8 km/s.<br />
amount of extra time.<br />
The diagram on Figure 2.7 shows that in order to work with 10 Hz data with<br />
a 350 km wide model and the depths in order of 48km (using 4 grid points per<br />
wavelength) and minimum P wave velocity of 2:8 km/s one would need a machine<br />
with approximately 2 GB of memory. Such machines are available these days at<br />
the top end of the workstation market. One can see that without frequency <strong>domain</strong><br />
methods, nested dissection, current workstation resources cannot tackle experiments<br />
of this size in production time scales.<br />
Figure 2.8 illustrates the CPU times on Digital Alpha 3000/300 workstation<br />
<strong>for</strong> the two ordering schemes. For small models, nested dissection per<strong>for</strong>ms worse<br />
(due to a computation overhead imposed by an irregular matrix structure), but when<br />
n x is greater than 200, CPU times are lower <strong>for</strong> the nested dissection case. However,<br />
it is important to point out that the main consideration in frequency <strong>domain</strong> <strong><strong>for</strong>ward</strong><br />
<strong>modelling</strong> is the memory requirements; the CPU time is usually low. The elapsed<br />
time is dominated by a disk input and output due to the large amount of <strong>seismic</strong><br />
data being computed. In our numerical tests, the nested dissection matrix solver<br />
did not need more than 15 minutes per frequency, even <strong>for</strong> models with the grid<br />
55
10000<br />
1000<br />
Time (s)<br />
Sequential ordering<br />
Nested dissection ordering<br />
100<br />
10<br />
100 200 500<br />
Grid points (n x or n z )<br />
Figure 2.8: CPU time versus number of grid points <strong>for</strong> the case in which n x = n z ,<br />
computed on Digital Alpha 3000/300 workstation.<br />
approaching 1,000,000 grid points. To per<strong>for</strong>m nested dissection ordering on a grid<br />
with 1; 000; 000 grid points requires two days.<br />
2.8 Conclusions<br />
For frequency <strong>domain</strong> <strong>seismic</strong> <strong>modelling</strong> using nite dierences, direct matrix<br />
solvers are the method of choice, due in part to poor conditioning of the matrices.<br />
Direct solvers have further advantages over iterative solvers if the linear systems are<br />
to be solved <strong>for</strong> multiple right hand sides. The LU decomposition matrix solver is<br />
the most apropriate. Ihaveshown that the amount of ll in suered by the matrix<br />
during LU decomposition depends strongly on the grid ordering. Nested dissection<br />
is an optimal grid ordering, but requires that the nite dierence operator be as<br />
local as possible in order to keep the ll in as small as posible.<br />
In order to justify using higher order operators, one would have to achieve<br />
an improvement in accuracy sucient to allow a greater than 50% reduction in the<br />
number of grid points per wavelength. As I will show in the following chapter it<br />
is more eective to keep the nite dierence operator small and accurate by using<br />
56
otated nite dierence operators.<br />
57
Chapter 3<br />
Visco-acoustic frequency <strong>domain</strong> acoustic <strong><strong>for</strong>ward</strong><br />
<strong>modelling</strong> using rotated nite dierence operators<br />
3.1 Introduction<br />
Forward <strong>modelling</strong> of the scalar wave equation in the frequency <strong>domain</strong> was<br />
introduced by Lysmer and Drake (1972), extended by Marfurt (1984b), and applied<br />
to <strong>seismic</strong> imaging by Pratt (1989b; 1990). Modelling in the frequency <strong>domain</strong> is<br />
computationally more demanding than time <strong>domain</strong> based schemes if a time <strong>domain</strong><br />
result is required <strong>for</strong> only a limited number of sources. The advantage of frequency<br />
<strong>domain</strong> <strong>seismic</strong> <strong>modelling</strong> is realized in multi-source experiments, and in frequency<strong>domain</strong><br />
waveeld inversion in particular, in which only limited number of frequencies<br />
from a large number of sources are needed.<br />
In Chapter 2 I concentrated on the minimization of the computational costs<br />
<strong>for</strong> a xed matrix size (and a xed dierence operator) on the ll in suered by the<br />
matrix. In this chapter I will concentrate on the minimizing the size of the initial<br />
matrix (a function of the grid size), by improving the accuracy of the nite dierence<br />
operators.<br />
This chapter begins with an overview of the nite dierence scheme developed<br />
by Jo et al. (1996) (itself an extension of a result by Cole (1994)) based on rotated<br />
58
nite dierence operators. As pointed out in the previous chapter, the size of the<br />
nite dierence operator is one of the crucial factors inuencing the total memory<br />
required to per<strong>for</strong>m frequency-<strong>domain</strong> <strong><strong>for</strong>ward</strong> <strong>modelling</strong>. If higher order (larger)<br />
nite dierence operators are used, the result is more accurate and thus a smaller<br />
grid is needed. However, a direct matrix solver then becames more expensive. The<br />
main problem is to nd a balance between the following two objectives:<br />
i) to use as small an operator as possible, and<br />
ii) to obtain as accurate result as possible.<br />
Both these objectives must be balanced to minimize the overall cost. Although high<br />
order nite dierence operators can be easily implemented in the frequency <strong>domain</strong>,<br />
I will show that this leads to an unacceptable increase in computational costs ( in<br />
particular, in memory requirements).<br />
Jo et al. (1996) showed that by using more than one second order nite difference<br />
operator <strong>for</strong> the same partial derivatives it is possible to develop a scheme<br />
which is comparable in accuracy to higher order schemes without signicantly increasing<br />
computational costs. In this chapter I will review the scheme proposed by<br />
Jo et al. (1996), and extend it to the heterogenous case. I will further discuss some<br />
of the parameters introduced by Jo et al. and evaluate their eect on the overall<br />
scheme.<br />
As I am not working on the boundary conditions a <strong>for</strong>mulation is from<br />
Pratt (1989b) in all examples.<br />
59
Figure 3.1: Finite dierence operators <strong>for</strong> acoustic frequency <strong>domain</strong> <strong>seismic</strong> <strong>modelling</strong><br />
in two coordinate systems. The symbol indicates that the model parameter<br />
is used at the corresponding grid point. a) Finite dierence operator in the original<br />
coordinate system. b) Finite dierence operator in the rotated coordinate system.<br />
c) The combination of both schemes.<br />
3.2 Forward <strong>modelling</strong> using rotated nite dierence operators<br />
3.2.1 Second order frequency-<strong>domain</strong> <strong>seismic</strong> <strong>modelling</strong><br />
The visco-acoustic, constant density frequency-<strong>domain</strong> wave equation in homogeneous<br />
isotropic (source-free) media can be written in the following <strong>for</strong>m:<br />
r 2 P + !2<br />
P =0; (3.1)<br />
2<br />
v<br />
where P is the pressure wave eld, ! is the angular frequency and v is the velocity.<br />
Because we pose the problem in the frequency <strong>domain</strong>, we may allow <strong>for</strong> viscous<br />
eects by using complex valued velocities if we wish. By using second order nitedierence<br />
approximations one can obtain the following nite dierence equation:<br />
P m+1;n , 2P m;n + P m,1;n<br />
2 x<br />
+ P m;n+1 , 2P m;n + P m;n,1<br />
2 z<br />
+ !2<br />
v 2 P m;n =0; (3.2)<br />
where P m;n represents the pressure of waveeld at the discrete location (m; n) (see<br />
Figure 3.1(a)) within the grid while x = z = is the grid spacing (grid point<br />
interval in x and z direction). Using this simple equation, one can solve the wave<br />
60
V /V ph<br />
1.03<br />
V/V gr<br />
1.03<br />
1.02<br />
1.02<br />
1.01<br />
1.01<br />
1.00<br />
1.00<br />
0.99<br />
0.99<br />
0.98<br />
0.97<br />
1/G<br />
0.05 0.1 0.15 0.2 0.25<br />
0.98<br />
0.97<br />
1/G<br />
0.05 0.1 0.15 0.2 0.25<br />
(a)<br />
(b)<br />
Figure 3.2: Numerical dispersion curves <strong>for</strong> frequency <strong>domain</strong> acoustic <strong><strong>for</strong>ward</strong> <strong>modelling</strong><br />
using ordinary second order nite dierence operators. a) Phase velocity dispersion.<br />
b) Group velocity dispersion.<br />
propagation problem numerically. However, as in Chapter 2, we will see that the<br />
simplest solution is not always the best.<br />
The usual way of describing the numerical accuracy of a particular scheme is<br />
to plot the normalized velocity as a function of number of grid points per wavelength.<br />
The normalized velocity isusually expressed by the ratio of the numerical velocity,<br />
bv, over the analytical velocity, v. The numerical result can be derived by applying a<br />
plane wave solution into the nite dierence equation 3.2 (see <strong>for</strong> example (Marfurt,<br />
1984a)). Figure 3.2 shows that <strong>for</strong> this simple second order scheme one needs over<br />
ten grid points per wavelength in order to keep dispersion errors small (under 3%).<br />
3.2.2 The rotated operator concept<br />
For a particular physical problem we will normally choose an orthogonal coordinate<br />
system in which to pose the equations and solve the problem. If the physical<br />
problem is described by a partial dierential equation in a Cartesian coordinate system,<br />
then the same solution should be obtained in all Cartesian coordinate systems.<br />
In the analytical case the solution will not depend on the coordinate system used.<br />
61
However, analytical solutions do not exist <strong>for</strong> most realistic cases, so numerical solutions<br />
are required. In this case the choice of the coordinate system will aect the<br />
solution. A numerical solution is only an approximation, and the accuracy of the<br />
approximation usually has an angular dependence, so that the result depends on<br />
the orientation of the coordinate system.<br />
In the case of plane wave propagation through homogeneous media, one would<br />
usually choose a coordinate system congruent with the direction of the wave propagation.<br />
For a single plane wave one can always develop a numerical scheme (or<br />
adjust the coordinate system) to produce an accurate result using low order nite<br />
dierence operators and a low number of grid points per wavelength. However, if<br />
waves can propagate in all directions (in complex models) how can one minimise<br />
the errors that arise due to the choice of the coordinate system?<br />
The solution utilized by Jo et al. (1996) is to use more than one Cartesian<br />
coordinate system, without including any points except nearest neighbours. In the<br />
2D case there are two possible coordinate systems (see Figure 3.1). We may pose<br />
the numerical problem in each of these coordinate systems and attempt <strong>for</strong>m a<br />
combined solution. On Figure 3.1(a) the nite dierence operator used in original<br />
coordinate system <strong>for</strong> the 2D acoustic wave equation is shown. In this operator,<br />
values from only ve grid points are used. Figure 3.1(b) shows the same operator<br />
in the rotated coordinate system.<br />
This operator uses values from four new grid<br />
points.<br />
A combination of the two operators (Figure 3.1(c)) uses values from all<br />
nine neighbouring grid points. In terms of memory requirements <strong>for</strong> direct matrix<br />
solvers (including a nested dissection one), there is virtually no extra cost associated<br />
with using the additional four grid points in the operator. The same is true <strong>for</strong> the<br />
CPU time involved. The main advantage of this approach is that it is possible to<br />
solve only one combined linear system of the same size and average the solutions<br />
implicitly during the calculation.<br />
Note that the procedure is specied <strong>for</strong> a grid with square elements ( x =<br />
62
z ).<br />
A similar procedure can be applied in the case of a rectangular grid, but<br />
the rotated coordinate system is then no longer Cartesian, and the appropriate<br />
wave equation <strong>for</strong>mulation must be used. Furthermore, a scheme developed <strong>for</strong> a<br />
rectangular grid would work correctly only <strong>for</strong> the x = z ratio <strong>for</strong> which the scheme<br />
is developed.<br />
3.2.3 Finite dierence scheme in homogeneous media<br />
Here I will apply the rotated nite dierence operators concept to equation<br />
3.1. If a second order nite dierence <strong>for</strong>mula is developed using a rotated grid (see<br />
Figure 3.1(b)) one can write:<br />
P m+1;n+1 , 2P m;n + P m,1;n,1<br />
2 2 x<br />
+ P m,1;n+1 , 2P m;n + P m+1;n,1<br />
2 2 z<br />
+ !2<br />
v 2 P m;n =0: (3.3)<br />
The factor 2 2 (as opposed to 2 in equation 3.2) comes from the increase in the<br />
grid point distance, in this case, from to p 2. A linear combination of the two<br />
schemes can be expressed by simple addition and multiplication of (3.2) by a and<br />
(3.3) by (1 , a) as:<br />
aA +(1,a)B+ !2<br />
v 2P m;n =0 (3.4)<br />
where A is the part of (3.2) consisting of nite dierence approximations <strong>for</strong> the<br />
Laplacian term in equation (3.1) while B is the equivalent part of equation (3.3).<br />
If equation (3.4) is expressed in a nite dierence <strong>for</strong>m it is easy to see that in fact<br />
only one system of linear equations need be solved. The the size of the resulting<br />
matrix is almost the same as that required <strong>for</strong> the single coordinate system alone<br />
(see Chapter 2).<br />
For example, if sequential grid ordering is used, the additional<br />
points used as shown on Figure 3.1(c) will generate four additional diagonals in the<br />
matrix next to exsisting diagonals, and the LU decomposed matrix will have only<br />
two more diagonals. This will add 2 n x n z elements (negligible in comparison with<br />
the total number of elements, 2 n x n z min(n x ;n z ) <strong>for</strong> realistic n x and n z ).<br />
63
3.2.4 Lumped and consistent matrix terms<br />
The second improvement introduced by Joetal.(1996) focused on the algebraic<br />
part of the acoustic wave equation, and is based on an approach used in the<br />
nite element method (Zienkijevic, 1977): The algebraic part of equation (3.1) is<br />
approximated by averaging the solution from neighbouring points. In nite-element<br />
terminology this is called a lumped matrix approach. For homogeneous media this<br />
approach results in the following replacement in the equations:<br />
! 2<br />
v 2 P m;n ) !2<br />
v 2 bP m;n + !2<br />
v 2 c(P m+1;n + P m,1;n + P m;n+1 + P m;n,1 )+<br />
! 2<br />
v 2d(P m+1;n+1 + P m+1;n,1 + P m,1;n+1 + P m,1;n,1 ); (3.5)<br />
where<br />
b +4c+4d=1: (3.6)<br />
This approach then combined with the approach indicated in the equation (3.4).<br />
3.2.5 Determination of optimal coecients<br />
Although any choice of values <strong>for</strong> the coecients a, b, c and d (satisfying<br />
equation (3.6)) will produce a possible numerical solution <strong>for</strong> the acoustic wave<br />
propagation problem in homogenous media, to obtain the most accurate solution<br />
<strong>for</strong> the problem optimal values <strong>for</strong> coecients a, b, c and d must be found. Note<br />
that due to equation (3.6) only two of the cocients b, c and d are independent.<br />
This can be posed as a minimization problem in which the errors in the<br />
solution are minimized as a function of the coecients. The minimization problem<br />
can be set up in more than one way, depending on what is actually minimized. The<br />
function to be minimized chosen by Jo et al. (1996) is :<br />
F (a; b; c) =<br />
Z :5<br />
Z =4<br />
0 0<br />
2<br />
bv ph (a; b; c; g; )<br />
, 1!<br />
d dg (3.7)<br />
v<br />
64
V/V ph<br />
1.03<br />
V/V gr<br />
1.03<br />
1.02<br />
1.02<br />
1.01<br />
1.01<br />
1.00<br />
1.00<br />
0.99<br />
0.99<br />
0.98<br />
0.97<br />
0.0<br />
1/G<br />
0.05 0.1 0.15 0.2 0.25<br />
0.98<br />
0.97<br />
1/G<br />
0.05 0.1 0.15 0.2 0.25<br />
(a)<br />
(b)<br />
Figure 3.3: Numerical dispersion curves <strong>for</strong> frequency <strong>domain</strong> acoustic <strong><strong>for</strong>ward</strong> <strong>modelling</strong><br />
using rotated nite dierence operators. a) Phase velocity dispersion. b)<br />
Group velocity dispersion.<br />
where g = 1=G, G is the number of grid points per wavelength, is the wave<br />
propagation angle, bv ph (a; b; c; g) is the numerical phase velocity while v is the exact<br />
velocity. The coecient d is not used since d = (1 , b , 4c)=4 (see equation<br />
3.6). Expressions <strong>for</strong> the numerical phase velocity bv ph (a; b; c; ; g) and group velocity<br />
bv gr (a; b; c; ; g) can be derived by applying a plane wave solution into the nal<br />
nite dierence equation (<strong>for</strong> details see Jo et al. (1996)). Jo et al. suggested the<br />
following optimal values:<br />
a = :5461<br />
b = :6248<br />
c = :09381<br />
d = :1297 10 ,5 (3.8)<br />
The value <strong>for</strong> the coecient d, in equation (3.5), suggested by Jo et al. (1996) is<br />
negligible. Thus it would appear that the following equation can be used <strong>for</strong> the<br />
parameter estimation:<br />
65
+4c=1 (3.9)<br />
which makes the minimization problem one dimension smaller. The slight non-zero<br />
value obtained by Jo et al. (1996) is quite likely due to the minimization procedure;<br />
the coecient d may be set to zero without any noticable deteriation of the result. I<br />
found that if d is increased by any signicant amount the normalized velocity starts<br />
to oscillate strongly as a function of G. For some extreme values of the coecient d<br />
the resulting velocity becomes complex valued. To demonstrate that d can eectively<br />
be set equal to zero, I have reproduced group and phase velocity dispersion curves<br />
<strong>for</strong> the case:<br />
a = :5461<br />
b = :6248<br />
c = 1 (1 , b)<br />
4<br />
d = 0: (3.10)<br />
Figure 3.4 shows the functions:<br />
D gr;ph (%) = 1 ,<br />
!<br />
bv(a; b; c; d; ; g)<br />
100<br />
bv(a; b; (1 , b)=4; 0;;g)<br />
where bv(a; b; c; d; ; g) is the numerical group or phase velocity asa function of Jo's<br />
coecients a, b, c and d and propagation angle . The dierent curves depict various<br />
wave propagation angles in isotropic homogenous media. The maximal dierence in<br />
numerical velocity introduced by setting coecient d equal to 0 is less than :004%,<br />
which is negligible in comparison with the errors we are dealing with. A similar<br />
low level of discrepancy is found if the coecient c is kept with the value suggested<br />
by Jo et al., and the coecient b set equal to b = 1 , 4c.<br />
For this reason the<br />
additional coecent, d will not be used here, nor will I use the equivalent parameter<br />
in the elastic case (see Chapter 5 <strong>for</strong> the visco-elastic <strong><strong>for</strong>ward</strong> <strong>modelling</strong> scheme).<br />
66
D gr(%)<br />
D<br />
ph(%)<br />
0.004<br />
0.003<br />
0.002<br />
0.001<br />
0.000<br />
-0.001<br />
-0.002<br />
-0.003<br />
-0.004<br />
0.05 0.1 0.15 0.2 0.25<br />
(a)<br />
1/G<br />
0.004<br />
0.003<br />
0.002<br />
0.001<br />
0.000<br />
-0.001<br />
-0.002<br />
-0.003<br />
-0.004<br />
1/G<br />
0.05 0.1 0.15 0.2 0.25<br />
(b)<br />
Figure 3.4: Dierence between the numerical velocity produced with and without<br />
the additonal coecient, d. a) Dierence in group velocity. b) Dierence in phase<br />
velocity. See text <strong>for</strong> detail explanation.<br />
The minimization problem thus reduces to a problem with two unknowns a and b.<br />
This reduces the search space, and it is possible to plot the minimzation result as<br />
a surface <strong>for</strong> various values of coecients a and b (while coecient c is a function<br />
of a and b see equation 3.10). This helps avoid local minima in the optimisation<br />
problem.<br />
3.2.6 Discussion of savings with rotated operators<br />
The dispersion curves <strong>for</strong> the set of parameters dened in the equation 3.8<br />
are shown on Figure 3.3.<br />
The results show that numerical errors in the phase<br />
velocity of less than 1% can be acheived with 4 grid points per wavelength, with<br />
errors of up to 3% <strong>for</strong> the group velocity <strong>for</strong> the same value of G. In comparison<br />
with the ordinary second order nite dierence schemes (see Figure 3.2) <strong>for</strong> the<br />
same problem, this shows a saving of more than 60% in the number of grid points<br />
required per wavelength <strong>for</strong> the same accuracy. The computational costs are almost<br />
the same as <strong>for</strong> the ordinary scheme (<strong>for</strong> the same number of grid points). Recalling<br />
67
that the required memory is afunction of n 2 log 2<br />
(n) (from equation (2.20), where n<br />
is the number of nodes on one side of a square grid), we can produce the following<br />
exact results <strong>for</strong> the savings in memory obtained if the same accuracy is required in<br />
both cases, by using :4n instead of n <strong>for</strong> the rotated operators case:<br />
M new<br />
M old<br />
=<br />
<br />
<br />
4<br />
n 2<br />
<br />
4n log2<br />
10 10<br />
n 2 log 2 n<br />
<br />
< 16 2+log 2 n , 3<br />
100 log 2 n<br />
!<br />
= :16 1 , 1<br />
log 2 n<br />
< :16; (3.11)<br />
where M is required memory. This shows that the saving in memory obtained <strong>for</strong><br />
the square model is at least 84%. The savings are slightly more <strong>for</strong> smaller grids<br />
than <strong>for</strong> larger grids. With regard to CPU time, one can use the following equation<br />
(from the equation 2.21):<br />
which shows a CPU time saving of over 90%.<br />
CPU new<br />
CPU old<br />
= :43 n 3<br />
n 3 = :4 3 =0:064; (3.12)<br />
3.2.7 Extension to the heterogenous case<br />
written:<br />
The 2-D visco-acoustic wave equation in heterogenous isotropic media can be<br />
@<br />
@x<br />
!<br />
1 @P<br />
+ @ (x; z) @x @z<br />
!<br />
1 @P 1<br />
+<br />
(x; z) @z K(x; z) !2 P =0; (3.13)<br />
where (x; z) is the 2D density function and K(x; z) is the bulk modules (in general<br />
complex valued). In this case one can still apply the rotated nite dierence <strong>for</strong>mulation,<br />
but the appropriate partial derivatives <strong>for</strong> the model parameters (K and )<br />
will have to be used. There is only one missing nite dierence operator required:<br />
68
@<br />
@u<br />
!<br />
1 @P<br />
(u; v) @u<br />
(3.14)<br />
This problem was solved by Kelly (1975) in the case of the original coordinate system<br />
using the operator:<br />
@<br />
@x<br />
!<br />
1 @P<br />
<br />
(x; z) @x<br />
1<br />
m+<br />
1<br />
2<br />
;n<br />
where m<br />
1<br />
2 ;n = 1 2 ( m;n + m1;n )<br />
[P m+1;n , P m;n ] , 1<br />
m, 1<br />
2<br />
;n<br />
2 x<br />
[P m;n , P m,1;n ]<br />
; (3.15)<br />
The same approach can be re<strong>for</strong>mulated in the rotated coordinate system by<br />
substituting x = x 0 , m = m 0 and n = n 0 :<br />
@ 1<br />
@x 0 (x 0 ;z 0 )<br />
where m<br />
1<br />
2 ;n 1 2<br />
@=@z 0<br />
!<br />
@P<br />
<br />
@x 0<br />
=<br />
partial derivatives.<br />
1<br />
m<br />
0 +<br />
1<br />
2<br />
;n 0 hP m<br />
0 +1;n<br />
0 ,P m<br />
0 ;n<br />
0<br />
1<br />
m+<br />
1<br />
2<br />
;n+ 1 2<br />
i<br />
,<br />
1<br />
m<br />
0 , 1<br />
2<br />
;n 0<br />
2 x 0<br />
[P m+1;n+1 ,P m;n ], 1<br />
m,<br />
1 ;n, 1 2 2<br />
2 2 x<br />
h<br />
P m<br />
0 ;n<br />
0 ,P m 0 ,1;n 0 i<br />
[P m;n ,P m,1;n,1 ]<br />
; (3.16)<br />
= 1 2 ( m;n + m1;n1 ). Equivalent equations can be derived <strong>for</strong> the<br />
In the case of lumped and consistent mass matrix terms equation (3.5) can<br />
be used, but the bulk modulus has to be distributed as well. If the coecient d is<br />
set to zero the one obtains the replacement <strong>for</strong>mula:<br />
1<br />
1<br />
1<br />
K !2 P m;n ) ! 2 [b P m;n + c( P m+1;n<br />
K m;n K m+1;n<br />
+ 1<br />
K m,1;n<br />
P m,1;n +<br />
1<br />
K m;n+1<br />
P m;n+1 +<br />
1<br />
K m;n,1<br />
P m;n,1 )]: (3.17)<br />
Substituting equations (3.15), (3.16) and (3.17) into (3.13), together with the equivalent<br />
equations <strong>for</strong> @=@z and @=@z 0<br />
denes the heterogenous nite dierence <strong>for</strong>mulation.<br />
The use of the heterogenous wave equation gives much more accurate results<br />
(when compared with use of the homogeneous wave equation with explicit boundary<br />
conditions) in the case of realistic geological models ( Ozdenvar and McMechan,<br />
69
1996). However quantitativeevaluation ofthe accuracy is dicult, although possible<br />
(Sei and Symes, 1994b). Tests we have run on a number of models have shown that<br />
the values <strong>for</strong> the coecients a, b and c obtained <strong>for</strong> the homogenous case may be<br />
used in the heterogenous <strong>for</strong>mulation sucessfully, even in highly heterogenous media<br />
(see Pratt et al. (1995)).<br />
3.3 Improvements acheived by rotated nite dierence operators<br />
In this section I will show the real improvements produced by the introduction<br />
of the nested dissection grid ordering and rotated nite dierence operators.<br />
Dealing with orders of magnitude and numbers of grid points per wavelength does<br />
not depict the achievements visually. Here I show the frequency <strong>domain</strong>, <strong>seismic</strong><br />
<strong><strong>for</strong>ward</strong> <strong>modelling</strong> of a realistic wide angle experiment. The model used is taken<br />
from McCarthy et al. (1991), simplied as in Pratt et al. (1996).<br />
The metamorphic core complex belt in southeast Cali<strong>for</strong>nia and western Arizona<br />
is a NW-SE trending zone of unusually large Tertiary extension and uplift.<br />
Three <strong>seismic</strong> refraction/wide angle reection proles were acquired and analyzed<br />
by McCarthy et al. (1991) as a part of of the U.S. Geological Survey's Pacic to<br />
Arizona Crustal Experiment. The <strong>seismic</strong> data were of excellent quality, and a large<br />
number of phases were observed and interpreted. A prominent midcrustal reection<br />
was indentied between 10 and 20km depth. Some non-horizontal features on the<br />
crust-mantle boundary can be observed on the data. The acqusition geometry consists<br />
of hundreds of sources on all proles spaced at 500 m intervals; the data were<br />
recorded at 250 m intervals. The proles are between 250 km and 400 km long. The<br />
data recorded evidence of structures at more that 30 km depths.<br />
The model used in this synthetic study (see Figure 3.5(a)) consists of most<br />
of the features observed on the three data sections presented by McCarthy et al.<br />
70
Figure 3.5: a) Model used <strong>for</strong> wide angle <strong><strong>for</strong>ward</strong> <strong>modelling</strong>, from McCarthy et<br />
al. (1991). b) c) and d) The shaded regions depict the size of the models that one<br />
could simulate without nested dissection and/or rotated nite dierences if the same<br />
equipment were used.<br />
71
(1991). The low velocity regions (1.5 km/s)from the top 500 mofthemodelare not<br />
incorporated, in order to make the simulation easier. The model topography is at,<br />
although the actual site is in a mountainous region. The dominant frequencies in the<br />
real data are as large as 10 Hz, however I have used a maximum frequency of 10 Hz<br />
(with a dominant frequency of 3.3 Hz), due to the lack of processing power available.<br />
The grid used represents 250 km by 38 km (2000 by 320 grid points) with a grid<br />
spacing of 125 m. Expressed in wavelengths this is 500 80 minimal wavelengths.<br />
This results in a linear system with 640 ; 000 complex (or 1 ; 280 ; 000 real) linear<br />
equations. The whole computation was carried out on a DEC Alpha 600/333 with<br />
512 MB of RAM. This workstation conguration will be standard very shortly and<br />
more powerfull equipment is already available on the market. The model is close<br />
to our current limit <strong>for</strong> frequency <strong>domain</strong> <strong><strong>for</strong>ward</strong> <strong>modelling</strong>, but memory prices<br />
continue to be reduced and machines are increasingly congured with more and more<br />
memory. The computational times are acceptable <strong>for</strong> this model size, approximately<br />
30 minutes per frequency using 240 sources, which would result in a total time of less<br />
than one day to invert the data set of this size in the frequency <strong>domain</strong> (assuming<br />
four iterations per frequency, <strong>for</strong> four frequencies). The time required to produce<br />
the time <strong>domain</strong> response <strong>for</strong> all 240 sources (128 frequencies <strong>for</strong> 256 time samples)<br />
was under two days.<br />
The main portion of the time was spent in the disk input<br />
and output. This is largely due to an inecient implementation <strong>for</strong> time <strong>domain</strong><br />
output, since we read and write all the time <strong>domain</strong> data after each frequency, which<br />
required a signicant amount of the total time. We utilized the ability of frequency<br />
<strong>domain</strong> <strong><strong>for</strong>ward</strong> <strong>modelling</strong> to produce the data directly in reduced time (see Section<br />
1.4 <strong>for</strong> an explanation), so that less frequencies were required.<br />
Figure 3.6 shows a resulting synthetic common shot gather in reduced time<br />
<strong>for</strong> a shot located at the top left corner of the model, and a similar section of real<br />
data from McCarthy et al. (1991). Many phases from the eld data, such as midcrustal<br />
reections, moho reections, and the head wave, can be observed on the<br />
72
Reduced time (T-x/6.0) Reduced time (T-x/6.0)<br />
4<br />
2<br />
0<br />
4<br />
2<br />
0<br />
0<br />
P P mc<br />
50<br />
PP m<br />
100<br />
OFFSET (km)<br />
PP lc<br />
(a)<br />
P g<br />
P mc<br />
P n<br />
150<br />
4<br />
2<br />
0<br />
4<br />
2<br />
0<br />
0<br />
50<br />
100<br />
OFFSET (km)<br />
(b)<br />
150<br />
0<br />
Depth (km)<br />
10<br />
20<br />
30<br />
mc<br />
lc<br />
Moho<br />
50 100 150 200<br />
OFFSET (km)<br />
(c)<br />
Figure 3.6: a) Synthetic data section from the model on gure 3.5. b) Common<br />
shot gather from the eld data. c) One of the models suggested by McCarthy et al.<br />
(1991) showing the ray paths used in their <strong>modelling</strong> approach.<br />
73
a) Time slice at 5s<br />
b) Time slice at 10s<br />
c) Time slice at 15s<br />
d) Time slice at 20s<br />
e) Time slice at 25s<br />
f) Time slice at 30s<br />
Figure 3.7: Time slices generated by <strong><strong>for</strong>ward</strong> <strong>modelling</strong> true the model on Figure<br />
3.5(a) at 5, 10, 15, 20, 25 and 30 seconds.<br />
74
synthetic section. It is also possible to see weak phases in the synthetic data that<br />
are not visible on the eld data. Those phases are diractions from discontinuities<br />
on the reectors. The time slices at 5, 10, 15, 20, 25 and 30 s (Figure 3.7) produced<br />
by the <strong><strong>for</strong>ward</strong> <strong>modelling</strong> code clearly show the <strong>for</strong>mation of a head wave on the<br />
Moho. It is possible to see the diractions from the model discontinuities on some<br />
of these time slices.<br />
In order to depict the improvements acheived by using nested dissection<br />
method (see Chapter 2) and the rotated nite dierence operators (this chapter),<br />
Figures 3.5(b), (c) and (d) show the model from the Figure 3.5(a) with rectangles<br />
covering the size of the regions that could be modelled using the frequency <strong>domain</strong><br />
technique, using only some or none of these improvements.<br />
It is clear that it is<br />
not feasible to predict the response of a realisticly sized wide angle model without<br />
nested dissection and without rotated nite dierences. Without our improvements,<br />
the largest acceptable model will corespond to a maximal source receiver distance<br />
of 50 wavelengths. Introducing either nested dissection or rotated nite dierence<br />
operators increases this to 100 or 150 wavelengths.<br />
The model used here represents<br />
500 wavelengths in oset and 80 wavelengths in depth. The total increase in<br />
the size of the model in Figure 3.5 a) is not just the sum of the improvements on<br />
Figures b) and c). A certain improvement comes from the interaction between the<br />
two techniques. This shows the importance of simultaneously developing both the<br />
nite dierence operator and the matrix solver. If we were to try to simulate the<br />
smallest model (gure 3.5(d)), but with our improvements, the required memory is<br />
reduced to 25 MB, which represents savings of over 95% (from 512 MB). Seen from<br />
this perspective, the improvements have reduced the memory requirements down to<br />
that normally available on a small personal computer.<br />
By generating the time <strong>domain</strong> data in reduced time directly it is possible<br />
to minimise the number of frequencies required to produce time <strong>domain</strong> output: I<br />
needed to model only 5 seconds of reduced time output.<br />
In comparison, a time<br />
75
<strong>domain</strong> approach would have to generate the full time <strong>domain</strong> simulation <strong>for</strong> 35 to<br />
40 seconds, using a small time A step (due to the highest velocity of 8.5 km/s in<br />
the model). If one multiplies this eort by the number of sources involved (240) the<br />
simulation is seen to be completely impractical.<br />
This same model was used earlier by Pratt et al. (1996) to show the feasibillity<br />
of the waveeld inversion on wide angle synthetic data. Although the frequencies<br />
used in that simulation (up to 2 Hz) were less than realistic wide angle data frequencies,<br />
machines available on the market today will be able to per<strong>for</strong>m frequency<br />
<strong>domain</strong> <strong><strong>for</strong>ward</strong> <strong>modelling</strong> at the more realistic frequencies used in this chapter.<br />
3.4 Conclusions<br />
In this chapter I have reviewed the development of the rotated nite dierence<br />
operators that allow one to signicantly reduce the number of grid points per<br />
wavelength <strong>for</strong> second order schemes. Ihave pointed out that not all the coecients<br />
introduced by Jo et al (1996) are useful and that the elimination of one of them<br />
does not aect the result in a measurable way. If this coecient is not used, the<br />
minimization problem becomes a 2D search, and can be carried out graphically.<br />
This approach will be of use in a later chapter in which the rotated nite dierence<br />
operators are developed <strong>for</strong> elastic <strong><strong>for</strong>ward</strong> <strong>modelling</strong>.<br />
Ihave further shown the extension of the rotated nite dierence operators to<br />
the heterogenous case and I have shown that by using both rotated nite dierence<br />
and nested dissection grid ordering it is possible to solve a realistic, large scale<br />
problem. Taking into account the whole solution procedure while working on the<br />
matrix solver puts certain constraints on the method in question.<br />
76
Chapter 4<br />
<strong>Frequency</strong> <strong>domain</strong> waveeld inversion example<br />
4.1 Introduction<br />
Computer <strong>modelling</strong> is used in many engineering disciplines <strong>for</strong> product development<br />
and testing. However, in exploration geophysics the main problem in not<br />
to model the data but to try to nd the model which \ts" the data collected at the<br />
site. This is a reverse engineering problem and in geophysics it is usually referred<br />
as inversion. By trans<strong>for</strong>ming the data into a geological model the target area can<br />
be better understood and exploited. Ideally one would like to determine the exact<br />
position, size and geometry of the target. This is not an easy problem. In order<br />
to trans<strong>for</strong>m from the data space into the appropriate model space it is necessary<br />
to have a good and fast <strong>seismic</strong> <strong>modelling</strong> algorithm with which the comparison<br />
between the real data and the synthetic data can be made, and with which updates<br />
to the model can be computed. It is critical that the main data phases from the<br />
eld data can be reproduced. In this sense <strong>seismic</strong> inversion is closely dependent on<br />
<strong>seismic</strong> <strong>modelling</strong>.<br />
Traveltime tomography, is a standard processing technique <strong>for</strong> certain kinds<br />
of <strong>seismic</strong> experiments due to its eciency and robustness. Tomographic approaches<br />
using <strong>seismic</strong> travel times have been used <strong>for</strong> a long time to generate images of geological<br />
regions (Dines and Lytle, 1979; Peterson et al., 1985; Dyer and Worthington,<br />
77
1988). Reviews have been provided by Worthington (1984), Bording et al. (1987)<br />
and Wong et al. (1987). In traveltime tomography, ray based methods are usually<br />
used to predict the travel times, and <strong>for</strong>m the required matrices. Waveeld inversion,<br />
as opposed to travel time tomography, attempts to t the wave<strong>for</strong>m data instead of<br />
the travel times only. Waveeld inversion is a computationally expensive procedure.<br />
It relies on ecency of <strong><strong>for</strong>ward</strong> <strong>modelling</strong> to quickly predict the synthetic responses<br />
through the model. Simulating wave<strong>for</strong>ms requires more resources than simulating<br />
the arrival times only. Tomographic datasets require a large number of sources in<br />
order to acheive the required data coverage. As pointed out earlier the frequency<br />
<strong>domain</strong> <strong><strong>for</strong>ward</strong> <strong>modelling</strong> can deal with large number of sources eciently.<br />
The improvements described in the previous chapters have been incorporated<br />
into the <strong><strong>for</strong>ward</strong> <strong>modelling</strong> part of a waveeld inversion routine in order to<br />
signicantly increase the speed of the procedure. This enables multiple runs with<br />
weighting constraint parameters to be evaluated and the correct constraints selected<br />
and used to produce the optimal output result. In this Chapter I will present the<br />
results obtained by waveeld inversion of a transmission data set recorded at the<br />
Grimsel Rock Laboratoty in Switzerland. The data set is an unusual one from the<br />
acquisition point of view. The full data set can be devided into the almost horizontal<br />
cross bore-hole data set and the two almost horizontal multiple oset VSP data<br />
sets recorded by applying the sources in between the two bore-holes used <strong>for</strong> the<br />
cross bore-hole survey and applying the receivers into the bore-holes. This acquisition<br />
geometry enabled excellent data coverage in a large part of the area. We see<br />
waveeld inversion as one of the main applications of the frequency <strong>domain</strong> <strong><strong>for</strong>ward</strong><br />
<strong>modelling</strong>.<br />
Resolution limitations of the traveltime methods (Williamson and Worthington,<br />
1993) have lead to attempts to t not only the arrival times but the <strong>seismic</strong><br />
wave<strong>for</strong>ms (eg, Devaney, (1984)). In the last decade, due to an increase in computational<br />
power, <strong>seismic</strong> waveeld inversion has become a feasible approach (Gauthier<br />
78
et al., 1986; Kolb et al., 1986; Zhou et al., 1985; Song et al., 1995; Pratt et al.,<br />
1995).<br />
Waveeld inversion was introduced by Lailly (1984) and is a non linear problem.<br />
The aim is to build the model which will t the wave <strong>for</strong>ms in the data. This<br />
approach is much more physical then the travel time tomography due to the fact<br />
that the travel times can be over ted (with a rough model with a lot of nodes it may<br />
be possible to reduce travel time residuals to zero), while it is imposible to t certain<br />
<strong>for</strong>ms of noise in the wave <strong>for</strong>ms, as those are a physical phenomena. For example<br />
if the random noise is present in the data and one attempts to invert the data using<br />
the waveeld inversion the underlying wave equation can never reproduce rapidly<br />
varying noise in the data however it can t some \source generated noise". The main<br />
dierence between waveeld inversion and traveltime tomography is in the nature<br />
of the data. The travel-times are not a directly recorded parameter: They include<br />
subjective in<strong>for</strong>mation introduced during traveltime picking. In some cases it may<br />
be dicult to pick consistent rst break travel time due to a signicant amount of<br />
noise in the data. Even in the cases where there is no signicant noise problem, the<br />
consistency of the picks may be systematically aected by human factors. On the<br />
other hand, the data used in waveeld inversion are a directly measured physical<br />
property. There is no subjective trans<strong>for</strong>mation involved in the processing which<br />
will aect the data. The only error in the input data is the error introduced by the<br />
eld equipment. Provided we can simulate the right waveeld we should be able to<br />
use the full undistorted data in inversion.<br />
4.2 Site description: Grimsel Rock Labaratory<br />
The Grimsel Rock Laboratory is located in SW Switzerland in the Aar Massif.<br />
The site is owned and operated by NAGRA, the Swiss national cooperative <strong>for</strong><br />
the disposal of radioactive waste. The laboratory is an underground test site located<br />
79
Figure 4.1: Grimsel Pass areal photo.<br />
Figure 4.2: Inside of the Grimsel Rock labaratory tunnel.<br />
80
eneath the Grimsel pass (see Figure 4.1) in anunderground tunnel (aphotograph<br />
of the tunnel interior is shown on Figure 4.2). The purpose of the site is to provide<br />
an in-situ, controlled location <strong>for</strong> the testing of rock characterization methods, with<br />
the ultimate objective being the application of techniques at a long term site <strong>for</strong><br />
the storage of radioactive waste. In this chapter I will present the re-processing of a<br />
tomographic data set acquired in 1985 (Gelbke et al., 1989). The test site is located<br />
within granitic rocks with a few mac dike intrusions, and a number of predominantly<br />
vertical fracture zones. A series of approximatly horizontal boreholes where<br />
used to deploy sources and receivers in the conguration shown on Figure 4.5. The<br />
data quality was quite high, with noise-free records and clean rst arrivals, however<br />
the data set suers from relatively large static shifts and signicant, unexplained,<br />
amplitude variations (representative data sections are shown on Figure 4.4 and Figure<br />
4.3). Similar data problems are observed by Gelbke etal.(1989) and Song and<br />
Worthington (Song and Worthington, 1995).<br />
The project aim was to test the utility of tomographic images as a <strong>tool</strong> <strong>for</strong><br />
detection of fractures capable of transmitting uids in nuclear waste depositories.<br />
Various tomographics techniques were tested at the site and compared. The techniques<br />
included radar tomography, dierential radar tomography and traveltime<br />
tomography (isotropic and anisotropic). Here I investigate the waveeld inversion<br />
approach to the tomographics data.<br />
The Field 2 region at the Grimsel Test Site is shown schematically in Figure<br />
4.5. It comprises a horizontal panel, bounded on two sides by horizontal boreholes,<br />
and on a third side by the underground access tunnel (the bottom of the Figure<br />
4.5). The boreholes dip approximately 15 degrees downwards from the tunnel. A<br />
number of other small boreholes traverse the region, the projection of onto the<br />
source-receiver plane of those boreholes is also shown in Figure 4.5. The Field 2<br />
<strong>seismic</strong> survey consisted of locating sources in the tunnel and recording two \oset<br />
VSP" datasets with receivers in both boreholes, and locating sources in one of the<br />
81
5 10 15 20 25 30 35 40 45 50 55 60<br />
Wave<strong>for</strong>m problem due to "in fill" survey<br />
0.0<br />
0.0<br />
0.01<br />
0.01<br />
0.02<br />
0.02<br />
0.03<br />
0.03<br />
Time (s)<br />
0.04<br />
0.05<br />
0.06<br />
0.04<br />
0.05<br />
0.06<br />
0.07<br />
0.07<br />
0.08<br />
0.08<br />
0.09<br />
0.09<br />
0.1<br />
0.1<br />
Receiver no<br />
(a)<br />
0.0<br />
0.0<br />
0.01<br />
0.01<br />
0.02<br />
0.02<br />
0.03<br />
0.03<br />
Time (s)<br />
0.04<br />
0.05<br />
0.06<br />
0.04<br />
0.05<br />
0.06<br />
0.07<br />
0.07<br />
0.08<br />
0.08<br />
0.09<br />
0.09<br />
0.1<br />
10 20 30 40 50 60 70 80 90 100 110 120<br />
Receiver no<br />
(b)<br />
0.1<br />
Figure 4.3: Two representative source gathers of VSP data from Field 2, as true<br />
amplitude displays. a) A VSP source gather with large oset. The spurious variation<br />
of amplitude from trace to trace is evident, as is the consistency of alternate traces.<br />
The data were recorded in two passes, with intermediate traces recorded during a<br />
later, \in-ll" survey. b) A near oset VSP source gather, on which the dramatic<br />
change in amplitude with receiver depth is evident. These variations in amplitudes<br />
cannot be modelled using the 2D acoustic method. In order to invert these data I<br />
apply a normalization to each trace separately.<br />
82
Sources from 1 to 121<br />
0<br />
VSP1<br />
5<br />
10<br />
VSP2<br />
15<br />
VSP3<br />
Time (ms)<br />
20<br />
25<br />
Crosshole<br />
VSP4<br />
30<br />
35<br />
40<br />
45<br />
50<br />
Figure 4.4: A representative common receiver gather of the Field 2 data, following<br />
windowing and trace normalization. The receiver was in borehole 3. The rst<br />
portion of the gather was recorded with sources in borehole 2, and thus represents<br />
a portion of the cross borehole data. The second section was recorded with sources<br />
in the tunnel, and thus represents a portion of the VSP data. The data have been<br />
windowed and trace-normalized. The random static shifts in the cross borehole data,<br />
and the systematic static shifts in the VSP data are evident. The labels indicate the<br />
VSP source groups that were identied, in order to solve <strong>for</strong> the source consistent<br />
static shifts.<br />
83
160m<br />
BOUS 85.003<br />
BOBK 85.004<br />
BOBK 85.008<br />
FBX 95.002<br />
N<br />
BOUS 85.002<br />
Tunnel<br />
160m<br />
Figure 4.5: Map of the Field 2 study area at the Grimsel Test Site. The <strong>seismic</strong><br />
data were acquired using the tunnel and boreholes BOUS85.002 and BOUS85.003<br />
(\boreholes 2 and 3"). The remaining boreholes are exploratory boreholes in which<br />
velocity in<strong>for</strong>mation is available and is used to test the wave<strong>for</strong>m images. The<br />
scale of this map is 1:1000, a representative square area 160m 160m is shown <strong>for</strong><br />
reference.<br />
boreholes and recording cross-borehole data in the other borehole.<br />
A number of<br />
other small boreholes traverse the region, the projection of onto the source-receiver<br />
plane of those boreholes is also shown in Figure 4.5.<br />
4.3 Waveeld inversion<br />
The idea of waveeld inversion (which attempts to t the complete arrival<br />
waveeld) follows on from the results obtained by tting the travel times through<br />
tomography. Initially, waveeld research was focused on development of migration<br />
84
algorithms. Conventional migration techniques attempt tofocus scattered waves at<br />
their point of origin (McMechan, 1983; Hu et al., 1988). From this starting point,<br />
work was extended to inversion techniques which produce quantitative in<strong>for</strong>mation<br />
on the physical parameter of the medium (Devaney, 1984; Gauthier et al., 1986).<br />
Lailly (1984) and Tarantola (1984) laid the foundations <strong>for</strong> wave<strong>for</strong>m inversion<br />
by posing the problem as a least-squares optimisation and showing how to<br />
eciently calculate the gradient of the objective function. The analytic <strong>for</strong>m of the<br />
Frechet derivative of waveeld data with respect to changes in the model parameters<br />
is given by the Born approximation, <strong>for</strong>mulated as an integral solution to the<br />
wave equation. This method attracted a lot of interest (Mora, 1987; Mora, 1989b;<br />
Beydoun and Mendes, 1989). The nonlinearity of the problem can be overcome by<br />
iterative procedures. The general nature of the approach enabled its implementation<br />
with various <strong><strong>for</strong>ward</strong> modeling approaches. Gauthier et al. (1986) demonstrated the<br />
application of Tarantola's idea to synthetic acoustic data using a time-<strong>domain</strong> nite<br />
dierence <strong>modelling</strong> algorithm. Gauthier et al. commented on the computational<br />
complexity of the problem due to slow convergence and the expense in a multisource<br />
conguration. Pratt and Worthington (1990) applied Tarantola's idea using<br />
frequency <strong>domain</strong> nite dierence <strong>modelling</strong> in 2D and overcame the problem of multiple<br />
sources. They showed that only a limited number of frequencies are required<br />
in some experimental geometries, particularly <strong>for</strong> the cross-borehole conguration.<br />
4.4 Waveeld inversion theory<br />
Wave<strong>for</strong>m inversion in general will require many solutions to systems of equations<br />
of the <strong>for</strong>m of equation (1.5). The iterative approaches to solving the non-linear<br />
problem assumes the following:<br />
One has access to n experimental observations, u (0)<br />
at a subset of grid<br />
points corresponding to receiver locations (<strong>for</strong> convinience of problem <strong>for</strong>mulation I<br />
85
will assume that rst n r n x n z (where n x n z isnumber of grid points in the model)<br />
grid points are receiver locations)<br />
An initial model exists that lies within range of a global minimum in the<br />
objective function<br />
Synthetic data u generated using this initial model that are representative<br />
of the real data<br />
For the purpose of the inversion the solution to the <strong><strong>for</strong>ward</strong> problem (equation<br />
(1.5)) can be written schematically as:<br />
u = S ~<br />
,1<br />
f (4.1)<br />
where S ~<br />
is in general a complex \impedance matrix", f is a source term and u is a<br />
column vector of length n representing the eld variable. The residual error, u is<br />
dened as the dierence between the initial model response and the observed data<br />
at the receiver locations. Thus<br />
u i = u i , u (0)<br />
i ; i =(1;2;:::;n r ) (4.2)<br />
where the subscript i represents the receiver number, n r is a number of receivers<br />
and the subscripted quantities are the individual components of u; u (0) , and u.<br />
As is common in many inverse problems, we seek to minimize the l 2 norm of<br />
the data residuals. Thus we minimize the \objective" function<br />
E(p) = 1 2 ut u ; (4.3)<br />
where p is the vector corresponding to the discretization of the physical parameters.<br />
In equation (4.3) the superscript t represents the ordinary matrix transpose and the<br />
superscript represents the complex conjugete, introduced to ensure the objective<br />
function is a true (real valued) norm <strong>for</strong> complex valued data.<br />
One method which may be used to calculate the update of the model at each<br />
iteration is the gradient method. The gradient method is a recipe <strong>for</strong> reducing the<br />
86
l 2 norm (8) by iteratively updating the parameter vector according to<br />
p (k+1) = p (k) , (k) r p E (k) ; (4.4)<br />
where k is an iteration number, and is a scalar step length chosen to minimize the<br />
l 2 norm in the direction given by the gradient ofE(p). The gradient of the objective<br />
function represents the direction in which the objective function is changing fastest.<br />
Thus, the objective function can always be reduced by pursuing such a direction.<br />
Although the optimal step length can be computed <strong>for</strong> linear problems, the<br />
step length in non-linear problems must generally be sought using line search techniques.<br />
The iteration in equation (4.4) is per<strong>for</strong>med until some suitable stopping<br />
criteria is reached. The convergence rate of the gradient method is generally quite<br />
slow, especially in the early iterations. Convergence can be improved by adopting a<br />
conjugate gradient approach (see <strong>for</strong> example Mora, 1988), which does not require<br />
any signicant additional computations.<br />
One may evaluate the gradient direction by taking partial derivatives of equation<br />
(4.3) with respect to the inversion parameters, p<br />
r p E = @E<br />
@p = Re n J~ t u o (4.5)<br />
where Re fxg denotes the real part of x. I assume there are m model parameters,<br />
so that p is a column vector of length m, and<br />
J ~ t is the transpose of the n r m<br />
Frechet derivative matrix, J ~<br />
, the elements of which are given by<br />
J ij = @u i<br />
@p j<br />
i =(1;2;:::;n r ); j =(1;2;:::;m): (4.6)<br />
One can see from equation (4.5) that the elements of J ~<br />
are not explicitly required<br />
in the gradient method, all that is required is to be able to compute the action of<br />
J ~ t on the vector u .<br />
Computation of the step length, required in the equation 4.4, is straight<strong><strong>for</strong>ward</strong>.<br />
For linear <strong><strong>for</strong>ward</strong> problems the step length is given by the followin equation:<br />
= jr pEj 2<br />
J ~<br />
r p Ej 2 (4.7)<br />
87
where jxj represents represents the Euclidean lenth of the vector x. For non-linear<br />
<strong><strong>for</strong>ward</strong> problems, the step length must be found using line search techniques along<br />
the direction opposite to the gradient (this is the case <strong>for</strong> <strong>seismic</strong> wave<strong>for</strong>m inversion).<br />
The gradient vector, required in euation (4.5), can be eciently computed<br />
through additional frequency <strong>domain</strong> <strong><strong>for</strong>ward</strong> <strong>modelling</strong> steps. To show this, I rst<br />
augment the m n r matrix J ~<br />
with the additional terms required to dene partial<br />
derivatives at all node points, not just at the receiver locations, to obtain a new<br />
m n x n z matrix c J ~<br />
. One can write an equation similar to the equation (4.5)<br />
r p E = Re<br />
cJ~t c u<br />
; (4.8)<br />
where c u is the data residual vector, of length n r , augmented with n x n z , n r zero<br />
values to produce a new vector of length n x n z . Explicitly, equation (4.8) represents<br />
2<br />
6<br />
4<br />
3<br />
@E<br />
@p 1<br />
@E<br />
@p 2<br />
.<br />
7<br />
5<br />
@E<br />
@p m<br />
=<br />
=<br />
2<br />
6<br />
4<br />
<br />
@u 1<br />
@p 1<br />
@u 1<br />
@p 2<br />
:::<br />
.<br />
@u n<br />
@p 1<br />
.<br />
@u n<br />
@p 2<br />
:::<br />
@u n+1<br />
@p 1<br />
@u n+1<br />
@p 2<br />
:::<br />
.<br />
@u nxnz<br />
@p 1<br />
@u<br />
@p 1<br />
.<br />
@u nxnz<br />
@p 2<br />
:::<br />
@u<br />
@p 2<br />
:::<br />
. .. .<br />
. .. .<br />
3t 2<br />
@u 1<br />
@p m<br />
@u n<br />
@p m<br />
@u n+1<br />
@p m<br />
7<br />
5<br />
6<br />
4<br />
@u nxnz<br />
@p m<br />
2<br />
t @u<br />
@p m<br />
6<br />
6<br />
4<br />
u 1<br />
.<br />
u n<br />
0<br />
.<br />
0<br />
3<br />
7<br />
5<br />
u 1<br />
.<br />
u n<br />
0<br />
.<br />
0<br />
3<br />
7<br />
5<br />
: (4.9)<br />
An expression <strong>for</strong> any of the partial derivatives in equation (4.9) in terms<br />
of the <strong><strong>for</strong>ward</strong> <strong>modelling</strong> matrix equation (1.5) can now be obtained by taking the<br />
88
partial derivative of both sides of equation (1.5) with respect to the ith parameter<br />
p i :<br />
or<br />
where<br />
@ u<br />
S = , @ S ~ u<br />
~ @p i @p i<br />
@ u<br />
@p i<br />
= S ~<br />
,1<br />
g (i) (4.10)<br />
g (i) = , @ S ~<br />
@p i<br />
u: (4.11)<br />
By analogy with equation (1.5), the partial derivatives in equation (4.10) are<br />
the solution to a new <strong><strong>for</strong>ward</strong> <strong>modelling</strong> problem, one in which the term on the right<br />
hand side plays the part of a \virtual" n x n z 1 source vector, g (i) . Perturbing the ith<br />
parameter by an amount p i will yield a perturbation in the <strong>seismic</strong> waveeld with<br />
values given by the solution to the <strong><strong>for</strong>ward</strong> problem in equation (4.10) multiplied<br />
by p i . The virtual source represents the interaction (or scattering) of the predicted<br />
(or background) waveeld, u with the parameter p i . I will there<strong>for</strong>e refer to @u=@p i<br />
as the \partial derivative waveeld from the ith node". As shown in equation (4.9),<br />
each column of<br />
J ~<br />
contains a partial derivative waveeld from a single physical<br />
parameter; there are m such columns. Where the inversion parameters consist of<br />
the values of a single physical parameter at the node points (the \point collocation"<br />
scheme), there will be m = n x n z columns and J ~<br />
is a square matrix.<br />
4.4.1 Ecient calculation of the gradient direction<br />
Since I could generate an equation similar to equation (4.10) <strong>for</strong> any choice of<br />
i, I can represent all the partial derivatives simultaneously by the matrix equation<br />
<br />
c J~ =<br />
@u<br />
@p 1<br />
@u<br />
@p 2<br />
:::<br />
@u<br />
@p m<br />
<br />
<br />
,1<br />
= S ~<br />
g (1) g (2) ::: g (m) <br />
(4.12)<br />
89
or<br />
c J~ = S ~<br />
,1<br />
G ~<br />
(4.13)<br />
where<br />
F ~<br />
is a n x n z m matrix, the columns of which are the virtual source terms<br />
<strong>for</strong> each of the m physical parameters.Equation (4.13) gives an explicit <strong>for</strong>mula <strong>for</strong><br />
the Frechet derivative matrix, J ~<br />
(being the rst n n x n z rows of b J). Computation<br />
of the elements of J using equation (4.13) would require m <strong><strong>for</strong>ward</strong> propagation<br />
problems to be solved, in addition to the one required to compute the virtual sources<br />
using equation (4.11). However, in order to compute the gradient using equations<br />
(4.5) or (4.8) it is not necessary to compute the elements of J explicitly. Substituting<br />
(4.13) into (4.8) I obtain<br />
<br />
r p E = Re cJ~<br />
t<br />
n<br />
bu = Re G~<br />
t vo ; (4.14)<br />
where<br />
v =<br />
h i t<br />
S~<br />
,1<br />
bu (4.15)<br />
or<br />
v = S ~<br />
,1<br />
bu (4.16)<br />
(by symmetry of the impedence matrix), which only requires one additional <strong><strong>for</strong>ward</strong><br />
problem to be solved. Thus the gradient is calculated in two steps: i) The \backpropagated"<br />
eld, v, is computed by solving a <strong><strong>for</strong>ward</strong> problem with the source terms<br />
replaced by the conjugate predicted waveeld (time reversed) and ii) The backpropagated<br />
eld is multiplied by the conjugate (time reversed) sources generated by the<br />
original predicted waveeld u.<br />
It is in<strong>for</strong>mative to use equations (4.11) and (4.14) to express the i-th component<br />
of the gradient vector as<br />
(r p E) i<br />
= Re<br />
(u (i)t "<br />
@ S~<br />
t<br />
@p i<br />
#<br />
)<br />
v<br />
(4.17)<br />
90
from which itisevident thatwhere @S<br />
@p i<br />
consists ofhighly local non-zero values near<br />
or at the ith row, as it will <strong>for</strong> the point collocation scheme, the gradient can be<br />
computed by a scaled multiplication of <strong><strong>for</strong>ward</strong> and backpropagated waveelds. This<br />
is the description usually given <strong>for</strong> the computation of the gradient vector, and it is<br />
clearly closely related to some reverse time migration algorithms, and to Claerbout's<br />
(1976) U/D imaging principle.<br />
4.5 Processing of third party synthetic data<br />
In this section I show the application of frequency <strong>domain</strong> <strong>modelling</strong> as a part<br />
of the frequency <strong>domain</strong> wave<strong>for</strong>m inversion technique. Be<strong>for</strong>e inverting the eld<br />
data an extensive study was carried out using a full elastic, 2D synthetic dataset<br />
generated by Prof. Korn of Leipzig University using a time <strong>domain</strong> nite dierence<br />
method. The velocity model used <strong>for</strong> this numerical experiment was provided by<br />
NAGRA, and is shown in Figure 4.6 (a). This model is intended to represent some<br />
of the expected geological features at the Grimsel Test Site, and the source-receiver<br />
geometry mimics that of Field 2. The large, low velocity zone in the lower right<br />
hand section of the model represents the known presence of lampophyre dykes that<br />
intersect the tunnel wall, and the thin, dipping features represent the known fracture<br />
directions at the site. There is a low velocity zone situated at the top of the model.<br />
This zone lies within a region with poor coverage, and will serve to illustrate the<br />
image degradation of features not well covered by the data. The eld geometry is<br />
bounded by borehole 2 at the right side of the gure, borehole 3 on the left hand<br />
side of Figure 4.6 (a) and the access tunnel at the bottom of the low velocity zone<br />
close to the bottom of the gure.<br />
In order to process these synthetic data, in preparation <strong>for</strong> the processing to<br />
be used <strong>for</strong> the real Field 2 data, the following pre-processing steps were undertaken:<br />
i) Project the two-component geophone data onto a local coordinate system de-<br />
91
Figure 4.6: Comparison of the travel time tomography result and the full wave-<br />
eld inversion from the third party synthetic elastic wave data. a) True velocity<br />
model used in elastic <strong><strong>for</strong>ward</strong> wave<strong>for</strong>m <strong>modelling</strong>, b) traveltime tomographic image<br />
<strong>for</strong>med from the picked synthetic data, c) acoustic waveeld inversion of the elastic<br />
synthetic data, without trace normalization, d) acoustic waveeld inversion with<br />
trace-normalization.<br />
92
ned by straight ray paths.<br />
ii) Window the projected wave<strong>for</strong>m rst arrivals in time using an exponentially<br />
tapered time window 15 ms wide, starting 5 ms be<strong>for</strong>e the picked arrival time.<br />
iii) Trace normalise the windowed data to remove spurious trace-to-trace amplitude<br />
variations.<br />
iv) Use travel-time tomography to produce a starting model <strong>for</strong> waveeld inversion.<br />
In the following paragraphs I summarize the reasoning behind the application of<br />
each of these steps:<br />
Data projection is used to trans<strong>for</strong>m the two component displacement data<br />
into single component data. This is required since the inversion software models only<br />
acoustic, compressional waves, and hence requires data that represent equivalent<br />
pressure eld variations. By geometrically projecting the two components onto the<br />
straight ray direction I enhance the compressional waves and partly eliminate the<br />
shear waves. This step was largely successful in eliminating most of the shear wave<br />
energy on the synthetic elastic data.<br />
Data windowing should ensure that only the rst arrival, transmission wave<strong>for</strong>ms<br />
are in the data. Transmission data are more suitable <strong>for</strong> waveeld inversion<br />
than the reections. Windowing also serves to exclude remaining shear wave energy<br />
from the data.<br />
Trace normalisation is not generally necessary, however, the amplitude variations<br />
in the eld data make this step essential when processing the real Field 2<br />
data. I there<strong>for</strong>e include this step with the synthetic data in order to assess the<br />
eect, detrimental or otherwise, on the inversion scheme. The data were collected<br />
in two passes, the original survey and the inll survey. The trace to trace consistency<br />
is not high. The traces seem to be consistently oset by a small time shift<br />
and the amplitude diers <strong>for</strong> more than an order of magnitude from trace to trace<br />
(see Figure 4.3 and Figure 4.4).<br />
93
In order to initialize the waveeld inversion scheme, it is necessary to begin<br />
with an adequate starting model. This model should be capable of describing the<br />
time <strong>domain</strong> data to within a half of the dominant period, in order to avoid tting<br />
the wrong cycle of the wave<strong>for</strong>ms. The lower the frequency, the less accurate the<br />
starting model need be, however all real data are band limited, and thus a certain<br />
accuracy is required of the starting model. In the real data the lowest frequencies<br />
are corrupted by an unacceptable amount of noise. I there<strong>for</strong>e choose to generate an<br />
accurate initial model using traveltime tomography, and proceed with the waveeld<br />
inversion using the higher frequencies.<br />
4.5.1 Travel time tomography<br />
All arrival times in the full synthetic dataset were picked, and used to <strong>for</strong>m<br />
a velocity image using travel-time tomography. The procedure used <strong>for</strong> travel-time<br />
tomography has been described Pratt and Chapman (1992) and Chapman and Pratt<br />
(1992). The anisotropic travel time tomography at the Grimsel test site is decribed<br />
in a report by Pessoa and Worthington (1995). Although that report describes the<br />
use of anisotropic velocity tomography, on the synthetic dataset here I have used<br />
only isotropic travel-time tomography, since an isotropic <strong><strong>for</strong>ward</strong> <strong>modelling</strong> scheme<br />
is used to generate the data. The tomographic result is shown on Figure 4.6 (b).<br />
Some of the features are recovered, but the thin low velocity layers do not appear<br />
in the image. There is a severe imaging problem with the low velocity layer on the<br />
top of the model due to poor coverage.<br />
4.5.2 Full waveeld inversion<br />
I carried out waveeld inversion of the projected synthetic data to show the<br />
advantage of using the waveeld in<strong>for</strong>mation, instead of travel-times only, and to<br />
verify the processing approach <strong>for</strong> the eld data. The result of waveeld inversion<br />
<strong>for</strong> the synthetic data is shown on Figure 4.6 (c). These images were <strong>for</strong>med using<br />
94
6frequency components ofthedata: 200, 300, 500, 700, 800 and 1000 Hz. Each frequency<br />
componentwas used <strong>for</strong> a maximum of 5 iterations be<strong>for</strong>e moving to the next<br />
frequency, using the current image as a starting model <strong>for</strong> the next frequency. The -<br />
nal frequency was used <strong>for</strong> 10 iterations. The individual frequency components were<br />
iterated upon until convergence, dened as the point beyond which the algorithm<br />
could no longer reduce the mist function. Following the amplitude normalization<br />
of the data (see next section), this occurred typically within 2 or 3 iterations. The<br />
same iteration strategy was followed <strong>for</strong> all subsequent images. In Figure 4.6 (c) it is<br />
evident that there is some improvement with respect to the traveltime image shown<br />
in Figure 4.6 (b). In particular, the exact geometry of the low velocity \dyke" at<br />
the bottom right is better resolved, and there is a subtle improvement in the geometry<br />
of most of the features. Moreover, the magnitudes of the velocity values are<br />
closer to the \true" velocity values. Nevertheless, the image is largely comparable in<br />
resolution to the traveltime image, although it is true that the use of full wave<strong>for</strong>m<br />
data excludes systematic errors introduced by manual travel-time picking.<br />
4.5.3 Full waveeld inversion of trace-normalised data<br />
In order to investigate whether the inaccurate amplitude simulation of the<br />
acoustic inversion method is adversely aected by the elastic wave amplitudes in<br />
the synthetic data, and furthermore to verify completely the approach <strong>for</strong> processing<br />
eld data (see below), I carried out waveeld inversion of trace-normalised synthetic<br />
data. This was necessary on the eld data due to high amplitude variations {<br />
here I attempt to verify the normalization as a pre-processing step. As the image in<br />
Figure 4.6 (d) shows some important features are better recovered than from the nonnormalised<br />
data. While inverting these data I found that the convergence rate was<br />
higher <strong>for</strong> normalised data set. This shows that the trace-normalisation can be used<br />
as a preconditioning technique in a wave<strong>for</strong>m inversion. The result conrms that the<br />
main source of in<strong>for</strong>mation in transmission data is in the wave<strong>for</strong>m itself and not in<br />
95
the trace-to-trace amplitude variations. It will be appreciated that amplitude preprocessing<br />
was important even in this synthetic example as elastic wave amplitudes<br />
are aected in a dierent manner from acoustic amplitudes.<br />
4.5.4 Comparison of travel time and full waveeld inversion methods<br />
In this section I will compare the results from the travel time and the full<br />
waveeld inversions carried out on the synthetic elastic data. On Figure 4.6 I depict<br />
the model used <strong>for</strong> <strong><strong>for</strong>ward</strong> <strong>modelling</strong>, the traveltime result and the two full waveeld<br />
results, on raw synthetic data and on trace normalised synthetic data. This gure is<br />
presented to show on a single gure the advantages of using the high resolution of the<br />
waveeld inversion technique. Smaller anomalies, completely overlooked by traveltime<br />
tomography, are completely recovered by waveeld inversion. All the anomalies<br />
lie at the correct positions in the region with good coverage. It is, however, possible<br />
to obtain false anomalies in the regions with poor coverage (at the top of the model),<br />
where the traveltime result has generated a low velocity anomaly. This anomaly is<br />
transfered into the full waveeld result by using the traveltime tomogram as an<br />
initial guess.<br />
This can be avoided in synthetic studies, in which, in most cases,<br />
one can start the waveeld inversion from a homogeneous model. However, when<br />
working with real data it is usually impossible to use suciently low frequency data,<br />
so that a better initial guess is required.<br />
In conclusion, once the data have been trace-normalized, waveeld inversion<br />
produces images in which the low velocity anomalies are much better resolved than<br />
on the traveltime tomographic image, and the velocity values are closer to the ones<br />
in the model shown on Figure 4.6 (a). It is important to point out that this is a<br />
signicant test of the method, as the data were generated by a third party, using<br />
an elastic wave simulation. The inversion software uses an acoustic wave method,<br />
which ignores elastic eects, but the images justify the use of this approximation.<br />
96
4.6 Inversion of real eld data<br />
Having successfully demonstrated the waveeld inversion technique on third<br />
party elastic synthetic data, and having veried much of the pre-processing techniques<br />
required, I now turn my attention to the real Field 2 data. The most signicant<br />
problem with the real data (originally identied in the report by Song and<br />
Worthington (1995)) was the trace-to-trace amplitude variation. An example of this<br />
amplitude variation is shown in Figure 4.3. As I shall show in this section, this problem<br />
has been entirely solved by trace-normalisation of the data. The other problem<br />
visible on gure 4.3 (a) is a trace to trace wave<strong>for</strong>m change. This is due to the data<br />
acqusition. The data were collected in two attempts: The original survey and the<br />
inll survey. The acquisition equipment had a dierent characteristic so the trace to<br />
trace consistancy is not high. The traces seem to consistently oset by a small time<br />
shift. The same behaviour can be observed on the trace normalised VSP common<br />
shot gathers. I have decided not to try to account <strong>for</strong> this problem.<br />
The pre-processing ow <strong>for</strong> the real data, with one exception, was identical<br />
to the pre-processing used <strong>for</strong> the elastic synthetic data. The full procedure was:<br />
i) Project the two-component geophone data onto a local coordinate system de-<br />
ned by straight raypaths.<br />
ii) Window the projected wave<strong>for</strong>m rst arrivals in time using an exponentially<br />
tapered time window 15 ms wide, starting 5 ms be<strong>for</strong>e the picked arrival time.<br />
iii) Trace normalise the windowed data to remove spurious trace-to-trace amplitude<br />
variations.<br />
iv) Use travel-time tomography to produce a starting model <strong>for</strong> waveeld inversion.<br />
v) Separate the unknown source behaviour into ve distinct physical \groups".<br />
Four individual groups were used <strong>for</strong> the VSP data, and one additional group<br />
was used to represent all of the crosshole data.<br />
Figure 4.4 shows in which<br />
97
manner these groups were identied.<br />
The additional step here, not used with the synthetic data, was the manner<br />
in which the unknown source behaviour was separated into ve distinct groups and<br />
solved <strong>for</strong>. For the synthetic data I solved <strong>for</strong> the source behaviour, but I treated the<br />
entire data as if it came from a single physical source. The eld data are known to<br />
contain signicant source-consistent static time shifts (as commented on by Gelbke<br />
et al, (1989)). An example of these static time shifts is shown in Figure 4.4.<br />
The source-consistent static time shifts were included into the inverse problem<br />
by using 4 separate VSP source \groups" <strong>for</strong> the eld data, and solving <strong>for</strong> 4 separate<br />
source functions. Using more than one source group does not signicantly eect the<br />
uniqueness of the inversion approach, but it is essential that these source-consistent<br />
errors are accounted <strong>for</strong>. There are also random source and receiver static shifts on<br />
the cross-borehole data, that I do not account <strong>for</strong>. The random nature of these latter<br />
problems causes a decrease in the signal to noise level of the nal images (see next<br />
section), but does not cause a signicant systematic deterioration of the images.<br />
4.6.1 Initial full waveeld inversion<br />
I begin the discussion of the results from the eld data by showing the initial<br />
results that were obtained be<strong>for</strong>e the complete pre-processing ow described in the<br />
previous section was worked out. In this section I will also study the cross-borehole<br />
and VSP components of the data separately. In all cases I begin from a starting<br />
model obtained from anisotropic velocity tomography, as described by M. Pessoa<br />
and M.H. Worthington in their 1995 report.<br />
This tomogram, after some simple<br />
smoothing, is shown in Figure 4.7.<br />
Ihave carried out tests to study the image quality if only a subsection of the<br />
data set is used. The result if only the cross-borehole component of the data is used<br />
is shown on Figure 4.8. The result is contaminated by strong velocity variations<br />
apparently originating at the borehole source-receiver locations. From this result<br />
98
km/s<br />
4.80<br />
4.85<br />
4.90<br />
4.95<br />
5.00<br />
5.05<br />
5.10<br />
5.15<br />
5.20<br />
5.25<br />
5.30<br />
5.35<br />
5.40<br />
Figure 4.7: Starting model <strong>for</strong> waveeld inversions of the eld data (from anisotropic<br />
velocity tomography).<br />
I may conclude that condence in these cross-borehole data cannot be high. The<br />
problem appears to be linked with the inconsistent source coupling in the borehole<br />
and the random static shifts described in the previous section.<br />
This section of<br />
the data is much noisier than the VSP section. In contrast, the result from VSP<br />
component of the data is shown on Figure 4.9. This result is less contaminated, and<br />
much closer to the expected geology at the site. These results show that imaging<br />
each subset of the data is not sucient on its own. However, the use of the whole<br />
data set should improve the result considerably.<br />
Figure 4.10 shows the inversion result using both cross borehole and VSP<br />
sections of the Field 2 data. The image shows some signicant improvements when<br />
compared with the individual images in Figures 4.8 and 4.9. However, there is still<br />
a strong noise component to these images that is apparently related to individual<br />
source and receiver locations. These noise patterns seem to propagate into the image<br />
and obscure the geological features. Ihave traced these noise features to the strong<br />
99
km/s<br />
4.80<br />
4.85<br />
4.90<br />
4.95<br />
5.00<br />
5.05<br />
5.10<br />
5.15<br />
5.20<br />
5.25<br />
5.30<br />
5.35<br />
5.40<br />
Figure 4.8: Preliminary full waveeld inversion image using non normalized crosshole<br />
part of the data only.<br />
km/s<br />
4.80<br />
4.85<br />
4.90<br />
4.95<br />
5.00<br />
5.05<br />
5.10<br />
5.15<br />
5.20<br />
5.25<br />
5.30<br />
5.35<br />
5.40<br />
Figure 4.9: Preliminary full waveeld inversion image using non normalized VSP<br />
part of the data only. Short oset VSP data are excluded due to large amplitude<br />
variations.<br />
100
km/s<br />
4.80<br />
4.85<br />
4.90<br />
4.95<br />
5.00<br />
5.05<br />
5.10<br />
5.15<br />
5.20<br />
5.25<br />
5.30<br />
5.35<br />
5.40<br />
Figure 4.10: Preliminary full waveeld inversion image using non normalized Field<br />
2 data, including both crosshole and VSP sections of the data. Short oset data are<br />
excluded due to large amplitude variations.<br />
and spurious trace-to-trace amplitude variations pointed out in Figure 4.3. This led<br />
to the decision to apply a trace normalization factor to each time <strong>domain</strong> trace after<br />
windowing and be<strong>for</strong>e extracting the various frequency components.<br />
A further decision was made to attempt to control remaining noise in the<br />
images by applying a constraint on the roughness of the solutions. This constraint<br />
is similar to the constraint used by Pessoa and Worthington in their 1995 report on<br />
traveltime tomography. The objective is to <strong>for</strong>m images that contain no unnecessary<br />
structure | the only structure that should appear in the images is structure is<br />
required to t the data. The eect of this additional constraint is explored in the<br />
next section.<br />
101
4.6.2 Regularization tests<br />
From this point on I depict images obtained from the data following the<br />
full pre-processing scheme, including the trace-normalization of the data.<br />
In the<br />
previous section I described the use of an additional constraint ontheroughness of<br />
the solution (I term this a \smoothing constraint"). From the pre-processed data<br />
I have generated a series of full waveeld inversion results with various levels of<br />
smoothing parameters. The resulting images are shown on Figure 4.11. In order<br />
to select an appropriate regularization level, I also computed the RMS residuals,<br />
and RMS roughness <strong>for</strong> each of these images, and plotted these against each other<br />
(Figure 4.12). As Pratt and Chapman have advocated <strong>for</strong> travel-time tomography<br />
in the past, I select an image that simultaneously ts the data as well as possible<br />
(low residuals) and is as smooth as possible (low roughness). I seek a \knee point"<br />
on the tradeo curve, which, in this case indicates a smoothing parameter of close to<br />
15. The full waveeld image shown on Figure 4.13 is my nal isotropic result, using<br />
a regularization level of 15, as determined from the previous gures. This image<br />
is already an important improvement on the starting model, however it appears to<br />
suer from a strong variation in background velocities from the left side of the image<br />
to the right. In the next sections I will further evaluate this image by studying the<br />
residuals, and I propose that this eect is caused by the low level anisotropy present<br />
at the test site.<br />
4.7 Isotropic results: Evaluation and verication<br />
In order to evaluate the isotropic result, I produced Figures 4.14 to 4.16,<br />
which represent respectively the eld data, the predicted data following the waveeld<br />
inversion and nally, the dierences, or residuals following the inversion.<br />
These<br />
plots are somewhat unconventional: Each pixel in these gures represents the real<br />
part of the complex-valued, single frequency waveeld at 800 Hz, recorded by a<br />
102
0 5 10 15<br />
20 25 30 35<br />
40 45 50 100<br />
km/s<br />
4.80<br />
4.85<br />
4.90<br />
4.95<br />
5.00<br />
5.05<br />
5.10<br />
5.15<br />
5.20<br />
5.25<br />
5.30<br />
5.35<br />
5.40<br />
Figure 4.11: Isotropic full waveeld inversion results with various values of smoothing<br />
parameter increasing from 0 (top left corner) to 100 (bottom right corner).<br />
103
RMS Roughness<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
5<br />
10<br />
15<br />
20<br />
25<br />
30<br />
3540<br />
4550<br />
0.4<br />
0.3<br />
0.0002 0.0004 0.0006 0.0008 0.0010 0.0012<br />
RMS Residuals<br />
Figure 4.12: Trade o curve showing RMS roughness vs RMS residuals <strong>for</strong> a suite<br />
of smoothing parameters.<br />
km/s<br />
4.80<br />
4.85<br />
4.90<br />
4.95<br />
5.00<br />
5.05<br />
5.10<br />
5.15<br />
5.20<br />
5.25<br />
5.30<br />
5.35<br />
5.40<br />
Figure 4.13: Final isotropic full waveeld inversion result.<br />
104
Figure 4.14: <strong>Frequency</strong> <strong>domain</strong> eld data at 800Hz. Please see the text <strong>for</strong> a full<br />
description of this gure. The grey scale is a relative amplitude, from the maximum<br />
negative values through to the maximum positive values.<br />
single source-receiver pair. The horizontal axis represents the receiver number (with<br />
receiver 1 at the left hand edge), the vertical axis represents the source number (with<br />
source 61 at the top edge). The data divides naturally into 3 sections: The crosshole<br />
data (top left quadrant), and the two VSP datasets (bottom two quadrants).<br />
As this is not a common representation, it is useful to explain the regular features<br />
one can (and should) observe: If this were a homogeneous media one would expect<br />
to see a set of linear features parallel to the main diagonal in the cross hole part of<br />
the survey, and circular patterns in the two quadrants representing two VSP data<br />
sets. These patterns are indeed visible in the data, Figure 4.14, in spite of the fact<br />
that this is not a perfectly homogeneous region. The patterns may also be compared<br />
with the synthetic waveelds predicted in the nal isotropic image, Figure 4.15. On<br />
both these gures the source-consistent static shifts can be observed in the VSP data<br />
sets as horizontal lines on the gures. If I had been successful in predicting the data<br />
with the inversion result, the dierences between these gures would be small, but<br />
more importantly would not show any systematic patterns. However, Figure 4.16<br />
shows that much of the systematic patterns in the data remain unaccounted <strong>for</strong>.<br />
This indicates a failure to explain some of the main features in the data. As I will<br />
show in the following sections of this chapter, this is likely to be due to anisotropy.<br />
It is also importanttoverify the method used to account <strong>for</strong> source-consistent<br />
105
Figure 4.15: <strong>Frequency</strong> <strong>domain</strong> modelled (predicted) data at 800Hz. See text <strong>for</strong><br />
a full description of this gure. The grey scale is a relative amplitude, from the<br />
maximum negative values through to the maximum positive values.<br />
Figure 4.16: Dierence between eld and modelled data at 800Hz. See text <strong>for</strong> a full<br />
description of this gure. The grey scale is a relative amplitude, from the maximum<br />
negative values through to the maximum positive values.<br />
106
-5<br />
-2.5<br />
0<br />
2.5<br />
5<br />
7.5<br />
10<br />
12.5<br />
15.<br />
Crosshole VSP1 VSP2 VSP3 VSP4<br />
-5<br />
-2.5<br />
0<br />
2.5<br />
Time (ms)<br />
5.<br />
7.5<br />
10.<br />
12.5<br />
15.<br />
17.5<br />
20.<br />
22.5<br />
Time (ms)<br />
17.5<br />
20.<br />
22.5<br />
Figure 4.17: Inverted source signatures. These signatures were extracted as an<br />
integral part of the waveeld inversion scheme. The similarity of the VSP source<br />
signatures, apart from the known static shifts, gives credence to the robustness of<br />
the inversion scheme.<br />
static shifts. As described above, to account <strong>for</strong> these static shifts, I divided the<br />
VSP data into 4 source \groups", each assumed to have a separate source behaviour<br />
(recall, these groups were identied on Figure 4.4). I also included a fth group<br />
to collectively represent all crosshole sources.<br />
In order to evaluate this approach<br />
I display the resultant (inverted) time <strong>domain</strong> source signatures (shown on Figure<br />
4.17). Each of these signatures was estimated independently from the data alone {<br />
it is reassuring that the wave<strong>for</strong>ms of the VSP source signatures are consistent, and<br />
that most of the dierences are due only to time shifts. This consistency tends to<br />
verify the approach.<br />
4.7.1 Discussion of isotropic results<br />
We have seen that the isotropic results show a large variation in velocities<br />
from the left hand edge of the images to the right hand edge. We have also seen<br />
that the data residuals show that much of the data variation remains unexplained<br />
107
y the best isotropic results. In all studies of Field 2 using travel-time tomography<br />
it has proven necessary to account <strong>for</strong> a small level of anisotropy (Gelbke et al.,<br />
1989; Pessoa and Worthington, 1995). I believe that the variation in velocities in<br />
the images and the remaining residual levels in the data are both best explained by<br />
the <strong>seismic</strong> anisotropy of the rocks.<br />
The anisotropy at the Grimsel Test Site is expected to be relatively low.<br />
Previous estimates (Pessoa and Worthington, 1995) from the <strong>seismic</strong> traveltimes<br />
have shown an overall level from 1% to 3 %, with a slow axis dipping 45 o from the<br />
top right corner to the bottom left corner. From the results of Chapter 3, I would<br />
expect that the velocity errors in the <strong>modelling</strong> code are of the order of 1%, and<br />
that the inversion errors will be at least an additional few percent. If the errors of<br />
the method are of the same order as the anisotropy level, can the anisotropy aect<br />
the images so strongly? The answer may lie in the systematic distribution of the<br />
ray directions in the data. The main ray directions in the VSP data sets are, in this<br />
case, almost exactly matched with slow and fast velocity axes. As the VSP data<br />
primarily recorded low and high velocities this had to be compensated in the image<br />
regions which where covered by a single part of the VSP data.<br />
The eect of anisotropy on the wave <strong>for</strong>m images has not been examined in<br />
detail primarily due to the expense of anisotropic <strong><strong>for</strong>ward</strong> <strong>modelling</strong>. However some<br />
experiments with homogenous elliptical anisotropy have been published (Pratt et<br />
al., 1995) but only if the amount of anisotropy is high (in the example used by Pratt<br />
et al. (1995) the amount of anisotropy was of the order of 20%, much larger than the<br />
maximum expected numerical errors). At Grimsel, in homogenous crystalline rocks,<br />
the anisotropy level is expected to be low and we did not expect any signicant<br />
artifacts on the image from anisotropy. However as shown in previous section the<br />
data residuals <strong>for</strong> the nal image are coherent and the image suers from signicant<br />
left right velocity distribution.<br />
In order to test the possible eect of low anisotropy, using the acqusition<br />
108
km/s<br />
4.80<br />
4.85<br />
4.90<br />
4.95<br />
5.00<br />
5.05<br />
5.10<br />
5.15<br />
5.20<br />
5.25<br />
5.30<br />
5.35<br />
5.40<br />
Figure 4.18: Isotropic inversion of synthetic data set from a homogeneous,<br />
anisotropic model.<br />
geometry at Grimsel, I generated a synthetic, homogeneous, elliptically anisotropic<br />
model (with 3% anisotropy and the slow axis dipping 45 o from the top right corner<br />
to the bottom left one) by shrinking the model in the fast velocity direction by<br />
3% and using the exact Field 2 source receiver conguration. Using this anisotropic<br />
model, I generated a full waveeld dataset using the isotropic frequency <strong>domain</strong> nite<br />
dierence <strong>modelling</strong> as described in previous chapter. The homogeneous (isotropic)<br />
velocity that was perturbed was V p = 5:2 km=s. I then inverted these data using<br />
the isotropic inversion scheme. The result, shown on Figure 4.18, suers from the<br />
same left-rightvelocity distribution problem as the isotropic images computed using<br />
the real data. The synthetic inversion result is correct in the central region where<br />
I have coverage from both the VSP datasets and from the cross-hole data sets. In<br />
the regions covered by only a single VSP data set the image compensates <strong>for</strong> the<br />
mismatch bycreating a alow (or high) velocity anomaly.<br />
As an additional test I have modelled and inverted the 2% elliptically anisotropic<br />
109
a) b) c)<br />
Figure 4.19: Data residuals <strong>for</strong> the wave<strong>for</strong>m inversion runs on the acoustic syntetic<br />
elliptically anisotropic (2 percent) data by assuming: a) Isotropic data (underestimated<br />
level of anisotropy) b) 2 percent elliptical anisotropy (correct value) c) 4<br />
percent eliptical anisotropy (overestimated value).<br />
synthetic data from the test model (Figure 4.7) and examined the data residuals <strong>for</strong><br />
various levels of assumed anisotropy. The data residuals <strong>for</strong> the isotropic assumption,<br />
the correct 2 percent elliptical anisotropy result and the overestimated elliptical<br />
anisotropy of 4 percent are shown on Figure 4.19. The gure shows that i) The data<br />
residuals are coherent when the incorrect amount of anisotropy is used and ii) The<br />
amplitude of data residuals in the correct case is the smallest (Thus an objective<br />
determination of the correct image is to use the level of data residuals) In the cases<br />
where incorrect anisotropic assumptions are made (the isotropic case and the 3%<br />
anisotropic case) the residuals are similar in apperance to residuals from the isotropic<br />
wave<strong>for</strong>m inversion of the Field 2 data on Figure 4.16. This tends to conrm that<br />
the nal isotropic image suers from unacounted anisotropy.<br />
The amplitude of data residuals in the correct case is the smallest. Thus<br />
an objective determination of the correct image is to use the level of data residuals.<br />
4.8 Anisotropic inversion of the eld data<br />
In order to compensate <strong>for</strong> the strong anisotropy eect evident from the<br />
initial isotropic inversions, I have carried out inversion of the eld data by assuming<br />
constant level of elliptical anisotropy of 1, 2 and 3% by shrinking the model (and<br />
110
the acquisition geometry) bythe same percentage in the high velocity direction. In<br />
each case the slow axis was chosen as in the previous synthetic study and consistent<br />
with the orientation used in most of the traveltime tomography studies at the site,<br />
i.e., dipping 45 o from the top right corner to the bottom left corner. The images are<br />
shown on Figure 4.20. There is a signicant dierence between the images (especially<br />
in the top corners). A high velocity at the top left corner of the Figure 4.20 (a) has<br />
become the low velocity zone on the gure 4.20 (d). The opposite trans<strong>for</strong>mation<br />
has occurred in the top right corner. The top corners are the main regions covered<br />
by a single VSP data set only. However, it is not clear from these images which<br />
is the correct background level of anisotropy. In order to aid the selection of this<br />
parameter, I also computed the RMS residuals <strong>for</strong> each of these images. The result<br />
is shown on Figure 4.21. The diagram shows that a level of 3% anisotropy gives<br />
residuals that are as far from the solution as the isotropic result is, and that the<br />
optimal result will have 1:8 , 1:9% anisotropy.<br />
Figure 4.22 show the nal anisotropic result, obtained by assuming 2% elliptical<br />
anisotropy. The left-right velocity distribution has largely disappeared. In<br />
order to verify this image I also show the data residual eld from this image on Figure<br />
4.23. The data residuals no longer display the strong systematic distributions<br />
observed in the isotropic case (see gure 4.16). Instead the data residuals are more<br />
nearly randomly distributed.<br />
Finally, I now include the result from the area directly to the right of Field<br />
2 (known as Field 1). The acqusition geometry is similar to the Field 2 geometry,<br />
although the boreholes are only 70m accros. The data were inverted independantly<br />
of the eld 2 data, using the same processing sequence. The nal image is shown<br />
next to the nal result <strong>for</strong> eld 2 on gure 4.25 and shows agreement on the common<br />
borehole (borehole 2). The level of anisotropy found from the Field 1 data was the<br />
same as <strong>for</strong> Field 2 (i.e. approximately 2 percent). This seems to verify the Field 2<br />
result and the approach used <strong>for</strong> data processing. The consistency of the anisotropy<br />
111
Figure 4.20: Anisotropic full waveeld inversion results with 0, 1, 2 and 3% elliptical<br />
anisotropy.<br />
112
-8<br />
7.0x10<br />
Data residuals<br />
-8<br />
6.8x10<br />
-8<br />
6.6x10<br />
-8<br />
6.4x10<br />
-8<br />
6.2x10<br />
0 1 2 3<br />
% Anisotropy<br />
Figure 4.21: RMS residuals <strong>for</strong> each test anisotropy level.<br />
estimation achieved by the waveeld inversion points out that it may be possible<br />
to give anestimate of the anisotropy level by waveeld inversion or even invert the<br />
data by using the anisotropic waveeld inversion to obtain a detailed anisotropic<br />
model.<br />
4.9 Conclusions<br />
In this Chapter I have shown the potential of frequency <strong>domain</strong> <strong>modelling</strong> as<br />
a<strong>tool</strong>inwaveeld inversion, and I have demonstrated the ability ofwaveeld inversion<br />
to yield high resolution images. A speedup of several orders of magnitude has<br />
been acheived during the course of the project. The whole computation takes approximately<br />
10 minutes per frequency, including ve iterations on a Digital 600/333<br />
workstation and requires only 40MB of RAM. Five to six frequencies are usually suf-<br />
cient so the full computation takes about 60 minutes (in comparison with 700MB<br />
of RAM and about ve days required be<strong>for</strong>e). The speed increase enabled multiple<br />
runs with various smoothing values and anisotropy levels. If less ecient <strong>modelling</strong><br />
113
km/s<br />
4.80<br />
4.85<br />
4.90<br />
4.95<br />
5.00<br />
5.05<br />
5.10<br />
5.15<br />
5.20<br />
5.25<br />
5.30<br />
5.35<br />
5.40<br />
Figure 4.22: Final full waveeld inversion image using 2% elliptical anisotropy.<br />
114
Figure 4.23: <strong>Frequency</strong> <strong>domain</strong> dierence eld (i.e., data residuals) at 800 Hz from<br />
the anisotropic inversion. See text <strong>for</strong> a full description of this gure. The grey scale<br />
is a relative amplitude, from the maximum negative values through to the maximum<br />
positive values.<br />
techniques were used this amount of testing would not be possible. From a computational<br />
point of view this problem size (14; 400 traces, 40 by 40wavelengths across,<br />
grid size of 160 by 160 grid points, 120 sources and 120 receivers) may be solved on<br />
a fast pentium based personal computer with enugh RAM (40MB) in a reasonable<br />
time (under 1day).<br />
From the inversion point of view the following conclusions may be drawn.<br />
High resolution waveeld images in a controlled test using synthetic elastic data can<br />
be achieved. This proves that if the underlying physics is sucient representation of<br />
the data one can expect the correct result. It is possible to pre-process the data to<br />
cope with large amplitude variations in a data, inconsistent trace to trace variations<br />
and signicant time static shift problems. This shows that the common eld data<br />
problems can be overcame. It is possible to produce high resolution reliable and<br />
interpretable images from the eld data. The following problems have been veried:<br />
we are unable to work directly with the two component displacement data, it is<br />
necessary to use only rst arrival waveeld in order to overcame the S wave arrivals<br />
in the data (the underlying physics is not good enough) and even low level anisotropy<br />
115
Figure 4.24: Final waveeld inversion images from both Fields 1 and 2, using 2%<br />
elliptical anisotropy.<br />
116
Figure 4.25: Final waveeld inversion images from both Fields 1 and 2, using 2%<br />
elliptical anisotropy (colour version).<br />
117
can eect theimages (the wrong theory once more). The problems accounted <strong>for</strong> in<br />
this Chapter have lead to the development of the elastic frequency <strong>domain</strong> <strong>modelling</strong><br />
scheme in the next Chapter in order to build the waveeld inversion procedure on<br />
top of it which may overcame some of the problems seen in this example.<br />
118
Chapter 5<br />
Visco-elastic frequency <strong>domain</strong> <strong>seismic</strong> <strong><strong>for</strong>ward</strong><br />
<strong>modelling</strong><br />
5.1 Introduction<br />
In this Chapter I will extend the rotated nite dierence operators introduced<br />
in Chapter 3 to the visco-elastic wave equation. As seen in Chapter 4, waveeld<br />
inversion based on the acoustic <strong><strong>for</strong>ward</strong> <strong>modelling</strong> can not work with the eld data<br />
directly.<br />
A certain degree of preprocessing was required, including projection of<br />
the two component data and amplitude normalisation.<br />
In order to improve on<br />
generality and accuracy of the model (and the resulting data), I have extended the<br />
improvements of the <strong>modelling</strong> scheme to the elastic wave equation.<br />
This chapter will describe the method I have developed and implemented to<br />
improve the accuracy and eciency of two-dimensional isotropic, frequency <strong>domain</strong><br />
visco-elastic <strong>seismic</strong> <strong>modelling</strong>. The new scheme uses the same grid points in the<br />
computational star as a standard second order dierencing scheme, thus conserving<br />
numerical bandwidth and sparsity. The new scheme allows a signicant reduction<br />
in the size of the numerical mesh, from 15 grid points per smallest wavelength in<br />
the model to 4 grid points per wavelength, dramatically reducing the computational<br />
costs.<br />
119
In essence, our new scheme is the extension to the visco-elastic case of the<br />
ideas discussed in Chapter 3 <strong>for</strong> visco-acoustic <strong>modelling</strong>. The method involved<br />
the introduction of two new numerical operators, the rst in a rotated coordinate<br />
frame, and the second using a lumped mass term, both of which are combined<br />
with standard second order numerical operators in an optimal manner to minimize<br />
numerical errors.<br />
I will begin with a development of the required dierencing operators. Because<br />
the visco-elastic waveeld (i.e., the displacement) is a vector quantity, the<br />
rotation of the coordinate frame presents additional diculties (when compared<br />
with the visco-acoustic case).<br />
Moreover, since the 2-D dierencing operators are<br />
nine point stars (as opposed to the ve point stars required <strong>for</strong> acoustic <strong>modelling</strong>),<br />
the rotated operators must be modied to minimize their spatial extent. I show how<br />
the required rotation can be achieved on the original nine point star. The second<br />
modication discussed in Chapter 3, the use of a lumped mass term, is dealt with<br />
in a straight<strong><strong>for</strong>ward</strong> manner.<br />
Following the development of the required operators, I then show how the new<br />
schemes are to be combined with the original, ordinary second order operators, to<br />
yield a scheme which is optimized to have minimum numerical errors. The dispersion<br />
analysis required <strong>for</strong> this optimization is given in the appendix of this thesis. I then<br />
analyze the numerical errors in the proposed scheme, and compare these with the<br />
errors <strong>for</strong> the standard second order scheme. I also show that, in contrast to the<br />
standard second order scheme, the new scheme correctly predicts a zero numerical<br />
shear velocity in uids.<br />
Finally, the <strong>modelling</strong> scheme is used to generate synthetic crosshole data<br />
from a model representative of the geological section at a near-surface test site.<br />
Boundary condition <strong>for</strong>mulation is the same as in Pratt (1990b). The <strong>modelling</strong><br />
results are used to demonstrate a possible relationship between strong, late arrivals<br />
in these crosshole data and the generation of mode converted shear waves in the<br />
120
(a) (b) (c)<br />
Figure 5.1: Computational stars <strong>for</strong> frequency <strong>domain</strong> elastic <strong>modelling</strong>. These stars<br />
indicate the coupling of the components of the displacement eld at the central node<br />
to displacements at the nearest neighbors. a) The ordinary, second order computational<br />
star, b) a possible rotated computational star, and c) a minimal, rotated<br />
computational star. The symbol, represents the coupling of the same displacement<br />
components, and also represents the only non-zero terms required in acoustic<br />
<strong>modelling</strong>. The symbol, symbol represents the coupling between perpendicular<br />
displacement components. The star in c) does not use additional points over the<br />
star in a), but introduces additional coupling between components not present in<br />
the original star.<br />
highly layered, attenuating sediments at the site.<br />
5.2 Visco-elastic <strong>modelling</strong><br />
In this section I will fully develop the method I have developed <strong>for</strong> nite<br />
dierencing the 2-D, frequency <strong>domain</strong>, homogeneous, elastic wave equation. I will<br />
comment at the end of this section on the extension to the heterogeneous wave<br />
equation.<br />
As with most other works in nite dierence methods, I will use the<br />
homogeneous <strong>for</strong>mulation to analyze the numerical errors, and I obtain a scheme<br />
that minimizes these errors.<br />
The 2-D, second order, frequency <strong>domain</strong>, visco-elastic wave equation in a<br />
homogeneous, isotropic and source free media consists of the two coupled equations:<br />
! 2 u +(+2) @2 u<br />
@x + u<br />
2 @2 @z +(+) @2 v<br />
=0 2 @x@z<br />
(5.1)<br />
! 2 v +(+2) @2 v<br />
@z + v<br />
2 @2 @x +(+) @2 u<br />
=0; 2 @x@z<br />
(5.2)<br />
121
where ! = 2f is the angular frequency, is the density, and and are Lame<br />
parameters. In order to be able to simulate visco-elasticity these Lame parameters<br />
will, in general, be frequency dependent and complex-valued. The waveeld<br />
variables, u and v are, respectively, the horizontal and vertical components of the<br />
Fourier trans<strong>for</strong>med displacements.<br />
5.2.1 Rotated nite dierences: Computational stars<br />
The numerical error of a regular grid nite dierencing scheme <strong>for</strong> equations<br />
(5.1) and (5.2) will depend on the wave propagation angle (an eect termed \numerical<br />
anisotropy"). This is because the distance between two discrete grid points<br />
is not the same in every direction. Usually, propagation will be most accurate in<br />
directions parallel to the coordinate axes. The solution suggested by Cole (1994)<br />
and by Jo et al.<br />
(1996) <strong>for</strong> scalar wave equations is to use two separate coordinate<br />
systems, one rotated with respect to the other, on the same discrete numerical<br />
mesh. A linear combination of the two results will, we hope, compensate <strong>for</strong> some<br />
of the numerical anisotropy. The aim is to minimize the numerical anisotropy, while<br />
retaining the existing grid and keeping the computational star as small as possible.<br />
The basic approach <strong>for</strong> developing a rotated nite dierence coordinate system<br />
is best understood with reference to Figure 5.1, in which the computational<br />
dierencing stars used to approximate the local partial derivatives on the grid are<br />
depicted. In the scheme devised by Jo et al., (1996) <strong>for</strong> acoustic wave propagation,<br />
the ve point dierence star <strong>for</strong> second order nite dierencing of the acoustic wave<br />
equation was rotated by 45 , expanded, and overlayed on the original grid (see Figure<br />
5.1(a) and 5.1(b)). This introduces four additional node points into the star,<br />
turning the combined computational star from a ve point star into a nine point star.<br />
For the elastic wave equation, applying standard second order nite dierencing to<br />
equations (5.1) and (5.2) results in a nine point computational star (e.g., Pratt,<br />
1989) (Figure 5.1(a)). At rst sight, it would appear that the same technique can<br />
122
e simply extended to the elastic waveequation by using a rotation and expansion<br />
of the computational star, resulting in the new star seen in Figure 5.1(b). Un<strong>for</strong>tunately,<br />
this star is not useful <strong>for</strong> <strong><strong>for</strong>ward</strong> <strong>modelling</strong>. In order to understand this, it<br />
is necessary to understand the manner in which the nite dierence approximation<br />
of the continuous equations (5.1) and (5.2) is actually solved.<br />
In general, wave equations such as (5.1) and (5.2) can be represented by:<br />
L(!) u(r) =f(r) (5.3)<br />
where L(!) is the appropriate, frequency dependent, linear partial dierential operator,<br />
u(r) is the eld variable (in this case the displacement, a continuous, 2<br />
component, vector eld), and f(r) is a source term (zero everywhere except at the<br />
location of the source). This equation, together with the boundary conditions must<br />
be satised everywhere. In 2-D, when equation (5.3) is approximated numerically,<br />
by nite dierences using a grid of n n nodes, this yields the matrix equation<br />
S ~<br />
u = f; (5.4)<br />
where S ~<br />
is a 2n 2 2n 2 complex valued matrix approximating the partial dierential<br />
operator L(!), u is now a 2n 2 -vector representing the two components of the<br />
displacement eld at all n 2 node points, and f is a similar 2n 2 -vector representing<br />
the source terms (the equation (5.4 is the same as the equation (1.5) but we just<br />
have two times as much eld variables).<br />
The matrix S represents a signicant storage requirement. The requirements<br />
~<br />
are largely determined by the sparsityof S , and by the manner in which this sparsity<br />
~<br />
is maintained in any solution method. In order to take advantage of the fact that<br />
additional sources involve only a change in the right hand side vector, s, I use direct<br />
solution methods as described in Chapter 2. Direct solvers are also required because,<br />
if articial absorbing boundary conditions are used, S ~<br />
will be non-symmetric and<br />
non-denite (precluding iterative solvers that require positive deniteness).<br />
It is<br />
123
dicult to <strong>for</strong>mulate direct solvers <strong>for</strong> arbitrarily sparse matrices, however it is<br />
simple to restrict computations to only those matrix elements which lie within the<br />
numerical bandwidth of the matrix. Better schemes can be developed by making<br />
use of optimal ordering schemes; I have discussed these in Chapter 2.<br />
The size of the computational star directly determines the eective numerical<br />
bandwidth of the dierencing matrix<br />
S ~<br />
. The bigger the star, the wider the<br />
bandwidth of non-zero elements in the matrix.<br />
The optimal bandwidth <strong>for</strong> the<br />
visco-elastic case is obtained by using a nine point star. The inclusion of any additional<br />
points will increase the bandwidth severely. If this increase is balanced by a<br />
corresponding decrease in the number of grid points required per wavelength, then<br />
this is acceptable. When one uses an optimal storage scheme (nested dissection from<br />
Chapter 2) <strong>for</strong> the matrix, the 13 point dierencing stars must be accurate enough<br />
to allow more than a 50% reduction in the number of grid points per wavelength in<br />
comparison with the 9 point dierencing star (see Chapter 2 <strong>for</strong> details). As I shall<br />
show, the new scheme I present requires of the order of 4 grid points per wavelength<br />
using a 9 point dierencing star <strong>for</strong> reasonable accuracy. Since one can never subsample<br />
the waveeld below the Nyquist criterion of two grid points per wavelength,<br />
there is no benet in using a dierencing star larger than 9 points. The exception<br />
to this may be<strong>for</strong> cases in which one requires an extremely accurate scheme.<br />
There<strong>for</strong>e, the rotated star in Figure 5.1(b) is not an acceptable computational<br />
scheme. These considerations lead to the new choice of a nite dierence star<br />
in the rotated coordinate frame, shown in Figure 5.1(c). This computational star<br />
does not require the use of any new grid nodes. The implication of this is that there<br />
will be no increase in the bandwidth of the dierencing matrix, and the increase<br />
in computational cost and in storage requirements over the ordinary second order<br />
scheme will be negligible. Having described the design of the optimal dierencing<br />
star, I now proceed in the next section to derive the exact <strong>for</strong>m of the required<br />
operators.<br />
124
5.2.2 Rotated nite dierences: Operators<br />
Solutions to the 2-D, visco-elastic wave equation, represented by the partial<br />
dierential equations (5.1) and (5.2) should naturally be independent ofany rotation<br />
of the coordinate system in which they are expressed. However, numerical solutions<br />
are approximations of the exact solution, and usually diverge from the actual solution<br />
in a manner that depends on the coordinate system.<br />
If there is more than one<br />
approximate solution <strong>for</strong> a particular problem, then a linear combination of them<br />
may bea more accurate approximate solution <strong>for</strong> the same problem.<br />
Inowintroduce a scheme which works with the original Cartesian coordinate<br />
system (x; z) and a new system (x 0 ;z 0 ) rotated by 45 o (see Figure (5.1)). I will<br />
assume from hereon that the equations <strong>for</strong> both coordinate systems will will be<br />
discretized on the same discretization mesh, with sample intervals x = z .<br />
The relationship between the displacements u,v in the original coordinate system,<br />
and u 0 ,v 0 in the new coordinate system is given by:<br />
u = 1 p<br />
2<br />
(u 0 , v 0 ) v = 1 p<br />
2<br />
(u 0 + v 0 ) (5.5)<br />
u 0 = 1 p<br />
2<br />
(u + v) v 0 = 1 p<br />
2<br />
(v , u): (5.6)<br />
Equations (5.1) and (5.2) in the new coordinate system are:<br />
! 2 u 0 +(+2) @2 u 0<br />
@x + u 0<br />
0 2 @2 @z +(+) @2 v 0<br />
0 2<br />
@x 0 @z 0 =0 (5.7)<br />
! 2 v 0 +(+2) @2 v 0<br />
@z + v 0<br />
0 2 @2 @x +(+) @2 u 0<br />
=0; 0 2<br />
@x 0 @z 0 (5.8)<br />
where x 0 and z 0 are the new coordinate directions. If we subtract and add equations<br />
(5.7) and (5.8), we obtain:<br />
<br />
! 2 u 0 , v 0 +<br />
( +2)<br />
+ ( + )<br />
! !<br />
@ 2 u 0<br />
@x , @2 v 0<br />
+ @2 u 0<br />
0 2<br />
@z 0 2<br />
@z , @2 v 0<br />
0 2<br />
@x 0 2<br />
!<br />
@ 2 v 0<br />
, @2 u 0<br />
=0 (5.9)<br />
@x 0 @z 0 @x 0 @z 0<br />
125
! 2 u 0 + v 0 +<br />
( +2)<br />
+ ( + )<br />
@ u<br />
@x + @ v<br />
0 2<br />
@z 0 2<br />
+ @ u<br />
@z 0 2<br />
+ @ v<br />
@x 0 2<br />
!<br />
@ 2 v 0<br />
+ @2 u 0<br />
=0: (5.10)<br />
@x 0 @z 0 @x 0 @z 0<br />
Dividing by p 2, and recalling the trans<strong>for</strong>mations (5.5) and (5.6), the resulting<br />
system is:<br />
! 2 u + 1 2<br />
! 2 v + 1 2<br />
"<br />
( +2)<br />
+ ( + )<br />
"<br />
( +2)<br />
+ ( + )<br />
@ 2 u<br />
@x , 2 @2 u<br />
0 2<br />
@x 0 @z 0<br />
@ 2 v<br />
@x 0 2 , @2 v<br />
@z 0 2<br />
!#<br />
@ 2 v<br />
@x +2<br />
@2 v<br />
0 2<br />
@x 0 @z 0<br />
@ 2 u<br />
@x , @2 u<br />
0 2<br />
@z 0 2<br />
!#<br />
!<br />
+ @2 u<br />
+ @2 u<br />
@z 0 2<br />
@z +2 @2 u<br />
0 2<br />
@x 0 @z 0<br />
!<br />
+ @2 u<br />
@x 0 2<br />
=0 (5.11)<br />
+ @2 v<br />
@z 0 2<br />
!<br />
+ @2 v<br />
@z 0 2 , 2 @2 v<br />
@x 0 @z 0<br />
+ @2 v<br />
@x 0 2<br />
=0: (5.12)<br />
This procedure trans<strong>for</strong>ms the eld variables from the rotated coordinate system<br />
into the original coordinate system, but leaves the coordinate axes themselves in the<br />
rotated frame of reference. The equations (5.11) and (5.12) are the elastic equivalent<br />
of equation (3.2) from Chapter 3. They represent the wave equation expressed as a<br />
nite dierence equation in the rotated coordinate system, using the original eld<br />
variables. This is required in order to be able to combine the resulting numerical<br />
solutions with numerical solutions to the original system.<br />
We now have two partial dierential equation systems: In the original coordinate<br />
system<br />
!<br />
! 2 u + A 1 =0 (5.13)<br />
! 2 v + B 1 =0; (5.14)<br />
(where A 1 and B 1 are the partial dierential parts of equations (5.1) and (5.2)). In<br />
the new (rotated) coordinate system<br />
! 2 u + A 2 =0 (5.15)<br />
! 2 v + B 2 =0; (5.16)<br />
126
(where A 2 and B 2 are the partial dierential parts ofequations (5.11) and (5.12)). I<br />
also have described the dierencing operators that will be used to approximate each<br />
of these two systems. The resulting two numerical systems will each havenumerical<br />
errors, but these errors will dier, and the numerical anisotropy <strong>for</strong> the two systems<br />
will be dierent.<br />
We can write a linear combination of the two systems as:<br />
! 2 u + aA 1 +(1,a)A 2 =0 (5.17)<br />
! 2 v + aB 1 +(1,a)B 2 =0; (5.18)<br />
and, by varying the coecient a, we obtain a whole family of results. Once again the<br />
exuations (5.17) and (5.18) are the elastic equivalent of the equation (3.4) <strong>for</strong> the<br />
elastic case. There are no limitations in the selection of the value of the coecient,<br />
a, as long as the value is real, although Jo at al. 1996 suggest a search in the region<br />
0 a 1 <strong>for</strong> practical purposes. The optimal value of coecient a must then be<br />
sought to maximize the accuracy of the solution, <strong>for</strong> all propagation directions. In<br />
other words, we seek to combine the two solutions in order to minimize the numerical<br />
anisotropy.<br />
Adequate second order nite dierence approximations <strong>for</strong> partial derivatives<br />
<strong>for</strong> equations (5.1), (5.2), (5.11) and (5.12) in each coordinate system can be found<br />
in Kelly at al. (1975), and are unchanged in this approach. For completeness I shall<br />
give the dierence <strong>for</strong>mulas required <strong>for</strong> the rotated scheme. The approximations<br />
used <strong>for</strong> the non-mixed partial derivatives in the 45 o coordinate system are:<br />
@ 2 v<br />
@x 0 2<br />
@ 2 v<br />
@z 0 2<br />
!<br />
!<br />
m;n<br />
m;n<br />
v m+1;n+1 , 2v m;n + v m,1;n,1<br />
2 2 (5.19)<br />
v m,1;n+1 , 2v m;n + v m+1;n,1<br />
2 2 ; (5.20)<br />
where is the grid spacing in x and z directions, and m, n are discrete grid point coordinates.<br />
In order to better visualize the computations implied by equations (5.19)<br />
and (5.20), and similar equations to follow, I will present these as computational<br />
127
\stars":<br />
!<br />
@ 2<br />
1 0 0 1<br />
0 -2 0<br />
@x 02 2 2 1 0 0<br />
!<br />
@ 2<br />
1 1 0 0<br />
0 -2 0<br />
@z 02 2 2 0 0 1<br />
; (5.21)<br />
The mixed nite dierence term in the rotated frame of reference, using the<br />
star shown in Figure 5.1(c), is given by<br />
or<br />
@ 2 v<br />
@x 0 @z 0 !m;n<br />
v m,1;n + v m+1;n , v m;n+1 , v m;n,1<br />
2 2 (5.22)<br />
!<br />
@ 2<br />
1 0 -1 0<br />
1 0 1<br />
@x 0 @z 0 2 2 0 -1 0<br />
: (5.23)<br />
5.2.3 Consistent and lumped mass terms<br />
The previous discussion targeted the dierential parts of the equations. This<br />
led to a scheme to minimize the amount of numerical anisotropy. In order to minimize<br />
the overall numerical dispersion, I now concentrate on the algebraic terms,<br />
! 2 v and ! 2 u in equations (5.17) and (5.18). These terms are normally approximated<br />
by using the value of the density, and the eld variable u or v at each local<br />
node point. This is known as a consistent <strong>for</strong>mulation. An alternative <strong>for</strong>mulation,<br />
known in nite element methods as a lumped <strong>for</strong>mulation, is obtained by using an<br />
interpolation of the eld values from the nearest node points, where the interpolation<br />
is weighted by the local mass (density) (Zienkijevic, 1977). If we combine the consistent<br />
and lumped mass methods by aweighted average, the required replacement<br />
terms <strong>for</strong> homogeneous media (with constant ) become<br />
and<br />
! 2 v m;n ) ! 2 2<br />
(1 , b)<br />
b v m;n + ! (v m+1;n + v m,1;n + v m;n+1 + v m;n,1 ) ; (5.24)<br />
4<br />
(1 , b)<br />
!u m;n ) ! b u m;n + ! (u m+1;n + u m,1;n + u m;n+1 + u m;n,1 ) ; (5.25)<br />
4<br />
128
where the coecient b, as with the combined rotated schemes, is chosen to minimize<br />
the numerical errors (the equations (5.24) and (5.25) are the equivalent of the<br />
equation (3.5) in the acoustic case). Here I have only used the values from the ve<br />
point star, and I have ignored the values from the corners of the nine point star. As<br />
I have shown in Chapter 3, the value of the third coecient (related to the corner<br />
points of the 9 point star) is always close to zero and can be set to zero without<br />
any visible eect on the nal result. At the same time this makes the minimisation<br />
problem 2-D and helps avoid local minima.<br />
The nal dierencing scheme <strong>for</strong> the 2-D, homogeneous, visco-elastic wave<br />
equation is obtained by combining the nite dierence approximations <strong>for</strong> equations<br />
(5.17) and (5.18), with equations (5.24) and (5.25). The complete scheme is given<br />
in the appendix, as equations (A-1) and (A-2).<br />
We now have a total scheme that i) minimizes numerical anisotropy (by an<br />
appropriate choice of the weighting factor, a) and ii) minimizes overall numerical<br />
dispersion (by an appropriate choice of the weighting factor b). All that remains is<br />
to determine the optimal values of the two weighting parameters, a and b. These<br />
parameters are determined by searching <strong>for</strong> values that provide a minimum of numerical<br />
anisotropy and numerical dispersion over the range of expected values of<br />
velocity. In the next section I describe the manner in which the optimal selection<br />
of parameters a and b is made. It should be noted that the two coecients, a and<br />
b, are not independent, and must determined simultaneously. Be<strong>for</strong>e proceeding to<br />
the optimization scheme, I will comment briey on the scheme <strong>for</strong> the heterogeneous<br />
wave equation.<br />
5.2.4 Heterogeneous <strong>for</strong>mulation<br />
The approach used in the previous three sections <strong>for</strong> the homogeneous viscoelastic<br />
wave equation can also be applied to the equivalent wave equation <strong>for</strong> heterogeneous<br />
media, in which the elastic constants, and , and the density are<br />
129
free to vary from one node point tothe next.<br />
The partial dierential equations <strong>for</strong> visco-elastic wave propagation in a heterogeneous,<br />
2-D media are:<br />
! 2 u + @<br />
@x<br />
"<br />
<br />
@u<br />
@x + @v<br />
@z<br />
!<br />
# "<br />
+2 @u + @ <br />
@x @z<br />
!#<br />
@v<br />
@x + @u =0 (5.26)<br />
@z<br />
and<br />
! 2 v + @ @z<br />
"<br />
<br />
! # "<br />
@u<br />
@x + @v +2 @v + @ <br />
@z @z @x<br />
!#<br />
@v<br />
@x + @u =0: (5.27)<br />
@z<br />
In an analogous manner to the approach used <strong>for</strong> homogeneous media, I<br />
substitute u,v,x and z with u 0 ,v 0 ,x 0 and z 0 to obtain equations in a 45 o rotated<br />
coordinate system, following which I apply similar manipulations to those used in<br />
equations (5.7), (5.8), (5.9) and (5.10), to obtain a new, mixed system of partial<br />
dierential equations in heterogeneous media:<br />
( "<br />
@ @u<br />
! 2 u + a <br />
@x @x + @v <br />
#<br />
@u +2 + @ @z @x @z<br />
( "<br />
(1 , a) @ @u<br />
+ @v + @v , @u<br />
2 @x 0 @x 0<br />
@x 0<br />
@z<br />
"<br />
0<br />
@ @v<br />
<br />
@z 0 @x 0<br />
"<br />
@ @u<br />
<br />
@z 0 @x 0<br />
+ @v<br />
@x 0<br />
+ @v<br />
@z 0<br />
@v<br />
"<br />
@<br />
<br />
@x 0<br />
@x 0<br />
"<br />
@v<br />
<br />
@x + @u<br />
@z<br />
@u +2<br />
@z 0 @x 0<br />
, @u<br />
@x 0<br />
+ @u<br />
@z 0<br />
, @u<br />
@z 0 <br />
+2<br />
@v<br />
@z 0<br />
, @u<br />
@x 0<br />
+ @u<br />
@z 0<br />
#) +<br />
+ @v<br />
@x 0 # +<br />
+ @v<br />
@z 0 # ,<br />
, @u<br />
@z 0 # ,<br />
+ @v<br />
@z 0 #) =0 (5.28)<br />
and<br />
(<br />
@<br />
! 2 v + a<br />
@z<br />
(1 , a)<br />
2<br />
"<br />
(<br />
@<br />
@x 0 "<br />
<br />
"<br />
@ @u<br />
<br />
@z 0 @x 0<br />
#<br />
"<br />
+ @ @v<br />
<br />
@x<br />
@u<br />
<br />
@x + @v <br />
+2 @v<br />
@z @z<br />
@x + @u<br />
@z<br />
@u<br />
+ @v + @v , @u @u<br />
+2<br />
@x 0<br />
@x 0<br />
@z 0<br />
@z 0 @x<br />
"<br />
0<br />
@ @v<br />
, @u + @u<br />
@z 0 @x 0<br />
@x 0<br />
@z 0<br />
+ @v + @v , @u @v<br />
+2<br />
@x 0<br />
@z 0<br />
@z 0 @z<br />
"<br />
0<br />
@ @v<br />
, @u + @u<br />
@x 0 @x 0<br />
@x 0<br />
@z 0<br />
#) +<br />
+ @v<br />
@x 0 # +<br />
+ @v<br />
@z 0 # +<br />
, @u<br />
@z 0 # +<br />
+ @v<br />
@z 0 #) =0: (5.29)<br />
130
in which, asbe<strong>for</strong>e, a is aweighting term used to control therelative importance of<br />
the two coordinate systems used in this mixed equation.<br />
Equations (5.28) and (5.29) must then be nite dierenced.<br />
Once again,<br />
I use the dierencing stars given by (Kelly et al., 1975) to produce the required<br />
operators. As an illustration I present the four required dierence operators <strong>for</strong> the<br />
rotated coordinate system as nite dierence stars, using the symbol to represent<br />
either of the Lame parameters, or :<br />
!<br />
0 0 + +<br />
@<br />
@x 0 @<br />
@x 0 1<br />
0 ,(<br />
2 + + 2 + , , ) 0<br />
, , 0 0<br />
; (5.30)<br />
!<br />
, + 0 0<br />
@<br />
@z 0 @<br />
@z 0 1<br />
0 ,(<br />
2 + 2 , + , ) 0<br />
+<br />
; (5.31)<br />
0 0 + ,<br />
and<br />
!<br />
0 , + +<br />
0<br />
@<br />
@x 0 @<br />
@z 0 1<br />
<br />
2 , 2 , 0 + +<br />
0 , , , 0<br />
!<br />
0 ,, + 0<br />
@<br />
@z 0 @<br />
@x 0 1<br />
<br />
2 + 2 , 0 , +<br />
0 , + , 0<br />
; (5.32)<br />
; (5.33)<br />
where the parameters () at intermediate grid points are given by<br />
= m<br />
1<br />
;n 1 : (5.34)<br />
2 2<br />
These four stars specify all required nite dierence operators <strong>for</strong> the rotated coordinate<br />
system; the remaining operators <strong>for</strong> the original coordinate system are<br />
unchanged from Pratt (1990a). The approach used <strong>for</strong> the lumped and consistent<br />
mass terms is introduced in exactly the same manner as <strong>for</strong> the homogeneous case<br />
131
(see equations 5.24 and 5.25), except that the density value must also be averaged<br />
from the neighboring node points, along with the eld variables.<br />
The nal system <strong>for</strong> <strong>modelling</strong> the 2-D, heterogeneous, visco-elastic wave<br />
equation is thus fully specied, apart from the unspecied weighting parameters,<br />
a, the relative amount of the original, unrotated second order scheme, and b, the<br />
relative amount of the consistent mass term with respect to the lumped mass term.<br />
These weighting parameters are obtained by returning to the homogeneous <strong>for</strong>mulation,<br />
and choosing values that provide minimum numerical errors.<br />
5.3 Numerical errors and optimization<br />
5.3.1 Determination of optimal coecients<br />
As discussed in the previous section, in order to fully specify the new differencing<br />
scheme, I now must determine values <strong>for</strong> the weighting coecients a in<br />
equations (5.17), (5.18), (5.28) and (5.29), and b in equations (5.24) and (5.25). In<br />
order to minimize the errors, I must be able to predict the numerical errors <strong>for</strong> a<br />
particular choice of a and b. The numerical errors are are predicted in a standard<br />
fashion by assuming a plane wave solution <strong>for</strong> the homogeneous scheme (given in the<br />
appendix as equations (A-1) and (A-2)), and solving the resultant system <strong>for</strong> the<br />
numerical compressional and shear wave velocities. The required analysis is given<br />
fully in the appendix. The nal equations depend on a, b, , K and , where is<br />
the Poisson ratio of the elastic medium, K is the wavenumber in grid point units<br />
(i.e., K =1=G where G is the number of grid points per wavelength), and is the<br />
propagation angle relative tothegrid axes.<br />
The method applied <strong>for</strong> determining the coecients is as follows: I search <strong>for</strong><br />
a set of values <strong>for</strong> a and b, using a representative value of <strong>for</strong> the elastic medium,<br />
such that a given mist function is minimized. The mist function is designed to<br />
measure the aggregate mist of the error in the numerical velocities over a range of<br />
132
possible values of K (governed by the range in true velocities in the medium), and<br />
over a range of propagation angles, . Formally I minimize<br />
where<br />
and<br />
F (a; b; ) =<br />
Z :5<br />
Z =4<br />
0<br />
0<br />
max n F p (a; b; ;K;);F s (a; b; ;K;) o d dK (5.35)<br />
<br />
F p (a; b; ;K;)=<br />
1,bv p g<br />
(a; b; ;R K; )<br />
<br />
v pg<br />
<br />
2<br />
F s (a; b; ;K;)=<br />
1,bv s g<br />
(a; b; ;K;)<br />
v sg<br />
<br />
2<br />
(5.36)<br />
: (5.37)<br />
In equations (5.36) and (5.37), K is the number of grid points per shear wavelength,<br />
q<br />
R = (0:5 , )=(1 , ) isthev s =v p ratio in the medium, v pg and v sg are the (true)<br />
compressional and shear wave group velocities, and bv pg and bv sg are the numerical<br />
group velocities, <strong>for</strong> which explicit expressions in terms of the variables (a; b; ;K;)<br />
are given in the appendix.<br />
For a given value of there are only 2 unknown parameters. It is there<strong>for</strong>e<br />
possible to evaluate the function, F (a; b; ), <strong>for</strong> a reasonable range of values of a and<br />
b, and plot this function as a surface. The optimal values can then be estimated,<br />
and the procedure can be repeated on a tighter interval near the optimal point.<br />
This was the procedure used to determine a and b <strong>for</strong> the examples that follow.<br />
The coecients a and b will in general depend on the Poisson ratio used in the<br />
model. If this is expected to vary widely, one could include an integral over possible<br />
Poisson ratios in the mist function.<br />
In Figure (5.2) the optimum values of the<br />
parameters as a function of the Possion ratio are shown. While the optimal value of<br />
the ratio between consistent and lumped mass matrix methods, b is relatively stable,<br />
it is evident that <strong>for</strong> large Poisson ratios (i.e., <strong>for</strong> near uids) the weighting of the<br />
unrotated scheme, a required <strong>for</strong> minimal numerical errors approaches zero. This<br />
is consistent with the expectation that the ordinary second order scheme cannot<br />
handle near uids (Stephen, 1983; Virieux, 1986a; Kerner, 1990). I shall return to<br />
the uid case in a later section, when I show that the scheme predicts no numerical<br />
133
1<br />
1<br />
0.8<br />
0.8<br />
a<br />
0.6<br />
0.4<br />
b<br />
0.6<br />
0.4<br />
0.2<br />
0.2<br />
0 0.1 0.2 0.3 0.4 0.5<br />
σ<br />
0 0.1 0.2 0.3 0.4 0.5<br />
σ<br />
Figure 5.2: Optimal values of coecients, a (the fraction of the ordinary second order<br />
scheme) and b (the fraction of the consistent mass matrix), plotted as a function of<br />
the Poisson's ratio, . The optimal value of coecient b is relatively insensitive to<br />
the value of . The optimal value of coecient a decreases <strong>for</strong> high values of , and<br />
becomes 0 <strong>for</strong> the uid case, in which case only the rotated scheme is used.<br />
dispersion <strong>for</strong> shear waves in uids, a necessary condition <strong>for</strong> being able to model<br />
liquid-solid interfaces.<br />
5.3.2 Numerical dispersion<br />
In this section I present some representative dispersion analyses <strong>for</strong> the combined<br />
scheme presented in this Chapter. The analysis is generated using the choice<br />
of weighting parameters a and b depicted in Figure 5.2, and the dispersion analysis<br />
given in the Appendix A. Figures 5.3 and 5.4 depict the normalized numerical phase<br />
and group velocities (<strong>for</strong> both compressional and <strong>for</strong> shear waves), using the standard<br />
second order scheme (left column), and the new, optimally combined scheme<br />
(right column). A value of 1.0 <strong>for</strong> the normalized velocity represents an error free<br />
numerical result; in all gures this is achieved when K =1=G, the wavenumber in<br />
grid point units, is zero.<br />
From Figure 5.3 it is evident that the original second order scheme yielded<br />
good, isotropic and undispersed results <strong>for</strong> the compressional wave phase velocities,<br />
and very poor phase and group results <strong>for</strong> shear waves. The original scheme also<br />
134
Numerical dispersion curves <strong>for</strong> σ=.33<br />
Old scheme<br />
Combined scheme<br />
1.1<br />
1.05<br />
v Pph /v Pph<br />
0.95<br />
0.9<br />
1.1<br />
1.05<br />
v Sph /v Sph<br />
1.0<br />
0.95<br />
0.9<br />
1.1<br />
1.05<br />
v Pgr /v Pgr<br />
1.0<br />
0.95<br />
0.9<br />
1.1<br />
1.05<br />
1.0<br />
0.95<br />
0.9<br />
v Sgr /v Sgr<br />
1.0<br />
0.05 0.1 0.15 0.2 0.25<br />
1/Gs<br />
0.05 0.1 0.15 0.2 0.25<br />
1/Gs<br />
0.05 0.1 0.15 0.2 0.25<br />
1/Gs<br />
0.05 0.1 0.15 0.2 0.25<br />
1/Gs<br />
1.1<br />
1.05<br />
v Pph /v Pph<br />
1.0<br />
v Sph /v Sph<br />
v Pgr /v Pgr<br />
0.95<br />
0.9<br />
1.1<br />
1.05<br />
1.0<br />
0.95<br />
0.9<br />
1.1<br />
1.05<br />
1.0<br />
0.95<br />
0.9<br />
1.1<br />
1.05<br />
v Sgr /v Sgr<br />
1.0<br />
0.95<br />
0.9<br />
0.05 0.1 0.15 0.2 0.25<br />
1/Gs<br />
0.05 0.1 0.15 0.2 0.25<br />
1/Gs<br />
0.05 0.1 0.15 0.2 0.25<br />
1/Gs<br />
0.05 0.1 0.15 0.2 0.25<br />
1/Gs<br />
Legend: Propagation angle (degrees)<br />
0 11.25 22.5 45<br />
Figure 5.3: Numerical dispersion of the new scheme <strong>for</strong> a Poisson ratio = 0:33,<br />
depicting normalized numerical velocity curves <strong>for</strong> compressional and shear phase velocities<br />
(top tworows) and group velocities (bottom tworows). Results are presented<br />
<strong>for</strong> the standard second order scheme (left column) and the new, combined scheme<br />
(right column). The dispersion curves are plotted against the shear wavenumber in<br />
grid point units, i.e., the reciprocal of the number of grid points per shear wavelength,<br />
G s . See text <strong>for</strong> the meaning of the symbols used on the vertical axes.<br />
135
Numerical dispersion curves <strong>for</strong> σ=.4<br />
Old scheme<br />
Combined scheme<br />
1.1<br />
1.1<br />
1.05<br />
1.05<br />
v Pph /v Pph<br />
1.0<br />
v Pph /v Pph<br />
1.0<br />
0.95<br />
0.95<br />
0.9<br />
0.05 0.1 0.15 0.2 0.25<br />
1/Gs<br />
0.9<br />
0.05 0.1 0.15 0.2 0.25<br />
1/Gs<br />
1.1<br />
1.1<br />
1.05<br />
1.05<br />
v Sph /v Sph<br />
1.0<br />
0.95<br />
0.9<br />
0.05 0.1 0.15 0.2 0.25<br />
1/Gs<br />
v Sph /v Sph<br />
1.0<br />
0.95<br />
0.9<br />
0.05 0.1 0.15 0.2 0.25<br />
1/Gs<br />
1.1<br />
1.05<br />
1.1<br />
1.05<br />
v Pgr /v Pgr<br />
1.0<br />
0.95<br />
v Pgr /v Pgr<br />
1.0<br />
0.95<br />
0.9<br />
0.05 0.1 0.15 0.2 0.25<br />
1/Gs<br />
0.9<br />
0.05 0.1 0.15 0.2 0.25<br />
1/Gs<br />
1.1<br />
1.05<br />
v Sgr /v Sgr<br />
1.0<br />
0.95<br />
0.9<br />
1.1<br />
1.05<br />
v Sgr /v Sgr<br />
1.0<br />
0.95<br />
0.9<br />
0.05 0.1 0.15 0.2 0.25<br />
0.05 0.1 0.15 0.2 0.25<br />
1/Gs<br />
1/Gs<br />
Legend: Propagation angle (degrees)<br />
0 11.25 22.5 45<br />
Figure 5.4: Numerical dispersion <strong>for</strong> a Poisson ratio =0:4, depicting normalized<br />
numerical velocity curves <strong>for</strong> compressional and shear phase velocities (top tworows)<br />
and group velocities (bottom two rows). Results are presented <strong>for</strong> both the standard<br />
second order scheme (left column) and the new, combined scheme (right column).<br />
The dispersion curves are plotted against the shear wavenumber in grid point units,<br />
i.e., the reciprocal of the number of grid points per shear wavelength, G s . See text<br />
<strong>for</strong> the meaning of the symbols used on the vertical axes.<br />
136
yielded a nearly isotropic result <strong>for</strong> compressional wave phase and group velocities,<br />
but contained errors of more than 8% at 4 grid points per wavelength (0.25 on<br />
the horizontal axes) <strong>for</strong> the compressional wave group velocity. The new combined<br />
scheme introduces a small amount of anisotropy into the compressional wave results,<br />
and slightly decreases the group velocity dispersion <strong>for</strong> P waves, but yields a<br />
much improved shear wave per<strong>for</strong>mance. Figure 5.4, in which the Poisson ratio has<br />
been increased to 0:4, is similar to Figure 5.3, except that the shear wave dispersion<br />
is much more severe <strong>for</strong> the standard second order scheme. Again, the combined<br />
scheme introduces a small amount of anisotropy into the compressional wave velocities,<br />
but decreases the overall velocity dispersion <strong>for</strong> compressional waves and yields<br />
a dramatically improved shear wave per<strong>for</strong>mance. In contrast to the standard second<br />
order scheme, the new scheme is thus able to cope well with a range of Poisson<br />
ratios.<br />
5.3.3 Modelling in uids<br />
It is well known that standard second order schemes generate innite dispersion<br />
<strong>for</strong> shear waves when used to simulate propagation in liquid layers (Bamberger<br />
et al., 1980; Stephen, 1983). This problem has eectively been solved <strong>for</strong><br />
time <strong>domain</strong> <strong>modelling</strong> by Virieux (1986a) and others (Kerner, 1990), in which the<br />
elasto-dynamic system, rather than the wave equation, is simulated on a staggered<br />
numerical grid.<br />
Let us consider the per<strong>for</strong>mance of the new scheme in uid layers. In Figure<br />
5.2 I show that the optimal value of the parameter a (the fraction of the standard<br />
scheme) approaches zero as the Poisson ratio approaches 0:5. This is a direct consequence<br />
of the behaviour of the standard scheme at large Poisson ratios. There<strong>for</strong>e,<br />
in pure uids, I will have to use the rotated scheme only (a=0). I now show analytically<br />
that the rotated scheme predicts the true shear wave behaviour in uids<br />
(v s =0:0):<br />
137
Ifollow closely the dispersion analysis given in the appendix. Equation (A-5)<br />
predicts the normalized, numerical shear wave group velocity. This equation cannot<br />
be used where the true shear velocity, v s is zero, as v s appears on the denominator<br />
of equation (A-5). However, a similar expression <strong>for</strong> the non-normalized, numerical<br />
shear velocity can be obtained:<br />
bv Sp =<br />
v p<br />
2K<br />
vu q<br />
u<br />
t 1 , 2 1 , 4 2 3<br />
; (5.38)<br />
2 3<br />
where the coecients are the same as those dened <strong>for</strong> equation (A-5). Inserting<br />
a =0into the coecients I nd that, <strong>for</strong> R = 0 (because v s = 0),<br />
1 =[,1 + cos x<br />
, cos z<br />
+ cos x<br />
cos z<br />
] ;<br />
2 =[,1,cos x<br />
+ cos z<br />
+ cos x<br />
cos z<br />
],<br />
3 = b + (1,b)<br />
2<br />
(cos x<br />
+ cos z<br />
), 1 = 3 ( 1 + 2 ),<br />
and 2 = 2 1 , sin 2 x<br />
sin 2 z<br />
where x = 2K cos , z = 2K sin and K =<br />
=2 (the wavenumber in gridpoint units). Inserting these simplied coecients<br />
back into equation(5.38) yields<br />
bv Sp =<br />
v p<br />
2K<br />
vu<br />
u<br />
q t( 1 + 2 ) , ( 1 + 2 ) 2 , 4 2<br />
; (5.39)<br />
2<br />
Further algebraic manipulation reveals that 2 = 0, from which it is then obvious<br />
that v s = 0 (<strong>for</strong> all values of K).<br />
Thus the new, rotated scheme, used on its own gives the exact shear wave<br />
group velocity (v s = 0) in a uid. An exact, constant, numerical phase velocity<br />
<strong>for</strong> all wavenumbers implies that the numerical group velocity (bv g =@!=@) is also<br />
exact.<br />
It is interesting to note that the rotated scheme generates a dierencing<br />
scheme that has certain similarities with the staggered grid dierencing schemes<br />
used by Virieux, (1986a) and Kerner, (1990) to solve the uid layer problem.<br />
The compressional wave group and phase dispersion in a uid <strong>for</strong> the new<br />
scheme are depicted in Figure 5.5. These results show that the numerical dispersion<br />
<strong>for</strong> compressional wave in uids can be signicant with the new scheme. However,<br />
138
(a)<br />
(b)<br />
1.1<br />
1.1<br />
1.05<br />
1.05<br />
v Pph /v Pph 1.0<br />
v Pgr /v Pgr 1.0<br />
0.95<br />
0.95<br />
0.9<br />
0.9<br />
0.05 0.1 0.15 0.2 0.25<br />
0.05 0.1 0.15 0.2 0.25<br />
1/G P<br />
1/G P<br />
Legend: Propagation angle (degrees)<br />
0 11.25 22.5 45<br />
Figure 5.5: Compressional wave dispersion in uids <strong>for</strong> the new, rotated scheme.<br />
In the uid case I use only the rotated scheme, with no component of the original,<br />
unrotated scheme (a = 0). a) Normalized compressional phase velocities. b)<br />
Normalized compressional group velocities.<br />
errors of less than 5% can be achieved with 7-8 grid points per wave length.<br />
5.3.4 Discussion<br />
From Figures 5.3 to 5.5 it is evident that the new, combined scheme per<strong>for</strong>ms<br />
better than the standard second order scheme <strong>for</strong> a wide range of Poisson ratios<br />
(naturally, the optimized scheme can be no worse than a single scheme).<br />
When<br />
compared with the optimal frequency <strong>domain</strong> (nite element) scheme given by Marfurt<br />
(1984a), the new combined scheme provides better accuracy in cases where is<br />
greater than approximately 0.3. For values of less than 0:3 the combined scheme<br />
gives comparable results to the optimal nite element scheme, with a slightly higher<br />
numerical anisotropy.<br />
By using a linear combination of two schemes I distorted the isotropic nature<br />
of the original scheme <strong>for</strong> <strong>modelling</strong> compressional waves, and thereby gained<br />
a higher accuracy <strong>for</strong> shear waves. For most Poisson ratios, the new rotated scheme<br />
models shear waves better than compressional waves, while the original scheme gives<br />
139
the opposite results. The fraction of each scheme (the a coecient) appears to control<br />
the relative accuracy <strong>for</strong> compressional and shear waves, and thus acts as a<br />
\tradeo factor". The other parameter (the b coecient) controls the overall dispersion<br />
<strong>for</strong> a given linear combination of both the standard and rotated schemes. It<br />
is there<strong>for</strong>e clear that there is room <strong>for</strong> customizing the scheme in certain situations:<br />
By choosing critical data phases, and choosing alternative values of the parameters<br />
a and b, itwould be possible to obtain a scheme which requires even less grid points<br />
per wavelength and obtain the same, or better accuracy <strong>for</strong> the particular wave type.<br />
If the model contains a range of Poisson ratios, then there may be inaccuracies<br />
in regions in which the Poisson ratio diers largely from that used in the selection of<br />
the parameters a and b. Fortunately, from Figure 5.2, the value of the parameters are<br />
relatively stable over a range of Poisson ratios. Only at large values of Poisson ratios<br />
is there an indication of a need to adjust these parameters { in heterogeneous models<br />
it may then be necessary to adjust these locally. A local variation of coecients <strong>for</strong><br />
variable Poisson ratios was also suggested <strong>for</strong> the nite element method by Marfurt<br />
(1984a). The eect of using variable coecients within the same model is not clear,<br />
although initial numerical tests Ihaverun look promising.<br />
One eect of the use of a rotated dierencing scheme is that any interface,<br />
dipping or at, will be treated in the model as a \staircase", due to the fact that<br />
at least one scheme is not aligned with the interface.<br />
This leads to regular, low<br />
intensity grid diractions on the interfaces. Such grid diractions are common <strong>for</strong><br />
dipping layers with standard schemes, and I do not consider the presence of these<br />
diractions <strong>for</strong> at interfaces to be a serious drawback. Muir et al. (1992) and<br />
Zeng and West (1996) pointed out a simple schemes based on eective media <strong>for</strong><br />
minimizing these eects.<br />
140
5.4 Elastic <strong>modelling</strong> example<br />
Verication of the new <strong>modelling</strong> scheme has been carried out in several standard<br />
models. Rather than showing these obvious results here, I now demonstrate<br />
the new scheme using the class of complex medium <strong>for</strong> which the scheme was designed:<br />
I use the scheme to study crosshole <strong>seismic</strong> data from a layered and faulted<br />
sedimentary sequence, in which a high level of attenuation is known to exist. The<br />
data come from the Imperial College test site at Whitchester in Northern England.<br />
The site consists of cyclically layered, interbedded mudstones, sandstones and carbonates<br />
(as described in more detail elsewhere (Pratt and Sams, 1996; Neep et al.,<br />
1996)).<br />
The crosshole data were acquired between two boreholes at the test site penetrating<br />
the top 220 m of the sequence, and separated by 75m. The acquisition was<br />
carried out using a clamped piezo-electric transmitter and hydrophone receivers.<br />
The transmitter was driven with a pseudo-random binary signal at a central frequency<br />
of 400Hz.<br />
The source was positioned in the rst borehole (the left hand<br />
side of the following gures), and recordings were made from each source position<br />
at receivers positioned every 2m between 17m and 217m in the second borehole (on<br />
the right hand side of the gures).<br />
The only structural feature in the section is a small, steep, right-dipping normal<br />
fault crossing Borehole 1 at 120m with a vertical displacement of approximately<br />
10m. The largest velocity contrasts are provided by two high velocity carbonate layers<br />
beds at depths of 145m and 170m, with thickness of 5m and 10m respectively.<br />
Figure 5.6 is a grey scale representation of the P-wave velocity variations used<br />
in a <strong><strong>for</strong>ward</strong> model of wave propagation at the site showing the location of the normal<br />
fault. This model is the result of an earlier study in full waveeld inversion (Pratt<br />
et al., 1995), using an acoustic inversion routine as explained in Chapter 4, after the<br />
method described by Song et al (1994). The high velocity carbonates can be easily<br />
141
identied; it is also possible to identify the location of the normal fault from the<br />
truncations of these carbonates towards the left of the image. Figure 5.7(a) depicts<br />
the observed <strong>seismic</strong> data from a representative common source gather, and Figure<br />
5.7(a) depicts the result of acoustic <strong><strong>for</strong>ward</strong> <strong>modelling</strong> in the velocity model shown in<br />
Figure 5.6. The acoustic <strong>modelling</strong> succeeds in reproducing the arrival times nearly<br />
exactly, and in predicting much of the character of the wave<strong>for</strong>m within the rst 5<br />
ms of the rst arrival. There is, however, little correspondence between the predicted<br />
and observed waveelds at late time. The <strong>modelling</strong> has failed to generate the large<br />
amplitude, incoherent events observed 10 to 20 ms following the rst arrivals.<br />
In order to study the remaining discrepancies between the observed data and<br />
the predicted data, I built a fully visco-elastic model from the P wave velocities in<br />
Figure 5.6 using the following assumptions: First, the S-wave velocities are assumed<br />
to be everywhere 50% of the P-wave velocities (i.e., I assumed a Poisson ratio of<br />
= 0:33). Next, since the rocks at the site are known to be highly attenuating<br />
(Neep et al., 1996), I incorporated inelastic attenuation by assuming that the P and<br />
S quality factors were each constant over frequencys, and homogeneous. I selected<br />
a quality factor of Q p =50<strong>for</strong> the P waves and Q s =20<strong>for</strong> the S waves (Neep et<br />
al., 1996). Appropriate complex valued Lame parameters <strong>for</strong> this elastic model were<br />
separately computed at each frequency, after Muller (1983). Finally, I modelled the<br />
source by using a horizontal point <strong>for</strong>ce introduced into the numerical mesh at the<br />
source location.<br />
It should be noted that the site exhibits signicant elastic anisotropy (as<br />
reported by Pratt and Sams (1996)), with P-wave velocities 20% faster in the horizontal<br />
direction than in the vertical direction. Although the <strong>modelling</strong> scheme has<br />
not been extended to the anisotropic case (an extension to simple transverse isotropy<br />
would be feasible but has not yet been carried out), I eected a simulation of the<br />
anisotropy by compressing the horizontal distances in the model by 20%, thus creating<br />
the kinematic equivalent of an elliptically anisotropic media with a vertical<br />
142
Distance<br />
from BH1 (m)<br />
Depth (m)<br />
0 15 30 45 60 75<br />
130<br />
140<br />
150<br />
160<br />
170<br />
180<br />
190<br />
200<br />
210<br />
Fault<br />
km/s<br />
4.4<br />
4.2<br />
4.0<br />
3.8<br />
3.6<br />
3.4<br />
3.2<br />
3.0<br />
2.8<br />
2.6<br />
Figure 5.6: P-wave velocity model <strong>for</strong> the Imperial College crosshole experiment.<br />
The model was obtained using acoustic fullwave inversion (Pratt at al. 1995). Data<br />
from the experiment, and modelled data <strong>for</strong> this velocity structure, are shown in<br />
Figures 5.7 and 5.8 .<br />
symmetry axis.<br />
This is consistent with the manner in which the anisotropy was<br />
simulated by Shipp and Pratt (1995), and most importantly, predicts the correct<br />
traveltimes.<br />
The results of visco-elastic <strong>modelling</strong> using the new scheme are shown on<br />
Figures 5.8(a), and 5.8(b)). The <strong>modelling</strong> synthesizes the horizontal and vertical<br />
components of displacement.<br />
No direct comparison of these data with the borehole<br />
pressure measured by the hydrophones in the eld can easily be made: The<br />
relationship is complicated and highly dependent on a number of poorly controlled<br />
variables (Peng et al., 1993). However, a qualitative comparison can be made: The<br />
horizontal component shows rst arrival times and wave<strong>for</strong>ms that are similar to<br />
the acoustic <strong>modelling</strong> results, and some high amplitude arrivals at late times. The<br />
vertical component shows high amplitude, incoherent arrivals similar to those observed<br />
on the real data. There is no exact match between these late arrivals and the<br />
143
Real data<br />
Acoustic <strong>modelling</strong> results<br />
0.22 0.21 0.2 0.19 0.18 0.17 0.16 0.15 0.14 0.13 0.12 0.11 0.1<br />
Receiver depth (km)<br />
0.0<br />
0.0<br />
0.01<br />
0.01<br />
0.02<br />
0.03<br />
0.04<br />
0.05<br />
Source<br />
0.22 0.21 0.2 0.19 0.18 0.17 0.16 0.15 0.14 0.13 0.12 0.11 0.1<br />
Receiver depth (km)<br />
0.02<br />
Time (s)<br />
0.03<br />
0.04<br />
0.05<br />
0.0<br />
0.01<br />
0.02<br />
0.03<br />
0.0<br />
0.04<br />
0.05<br />
0.01<br />
0.02<br />
Time (s)<br />
0.03<br />
0.04<br />
0.05<br />
(a)<br />
(b)<br />
Figure 5.7: a) A representative common source gather from the crosshole data collected<br />
at the Imperial College test site. The signal to noise ratio is high, and the<br />
rst arrival wave<strong>for</strong>ms are clear and coherent. At late times, incoherent, large amplitude<br />
arrivals dominate. b) Predicted common source data using acoustic <strong><strong>for</strong>ward</strong><br />
<strong>modelling</strong> in the velocity structure shown in Figure 5.6. The rst arrival traveltimes<br />
and wave<strong>for</strong>ms match well with the observed data, but the large amplitude, late<br />
arrivals are not predicted with the acoustic method.<br />
144
Elastic <strong>modelling</strong> horizontal component<br />
Elastic <strong>modelling</strong> vertical component<br />
Source<br />
0.22 0.21 0.2 0.19 0.18 0.17 0.16 0.15 0.14 0.13 0.12 0.11 0.1<br />
Receiver depth (km)<br />
0.0<br />
0.0<br />
0.01<br />
0.01<br />
0.02<br />
0.03<br />
0.04<br />
0.22 0.21 0.2 0.19 0.18 0.17 0.16 0.15 0.14 0.13 0.12 0.11 0.1<br />
Receiver depth (km)<br />
0.02<br />
0.05<br />
Time (s)<br />
0.03<br />
0.04<br />
0.05<br />
0.0<br />
0.0<br />
0.01<br />
0.02<br />
0.03<br />
0.01<br />
0.04<br />
0.05<br />
0.02<br />
Time (s)<br />
0.03<br />
0.04<br />
0.05<br />
(a)<br />
(b)<br />
Figure 5.8: Predicted common source data using the new visco-elastic <strong>modelling</strong> results.<br />
a) Horizontal displacement component. b) Vertical displacement component.<br />
The horizontal component shows rst arrival times and wave<strong>for</strong>ms that are similar<br />
to the acoustic <strong>modelling</strong> results, and some high amplitude arrivals at late times.<br />
The vertical component shows high amplitude arrivals similar to those observed on<br />
the real data.<br />
145
observed data. However, even with the simple assumptions I have made in building<br />
the visco-elastic model from the P-wave velocity model, I managed to create<br />
synthetic visco-elastic data that look more like the data collected at the site than<br />
the synthetic acoustic wave data. The late arrivals thus appear to be related to the<br />
mode conversion of P wave energy into shear wave energy within the heterogeneous<br />
model. One may speculate as to whether a better match in the times of the late<br />
arrivals could be achieved by adjusting the Poisson ratios in the model, or even to<br />
ask whether a <strong>for</strong>mal visco-elastic inversion of the data from this experiment could<br />
be attempted.<br />
In comparison with the visco-acoustic <strong>modelling</strong> (see Chapter 3) <strong>for</strong> the viscoelastic<br />
one needs twice as big grid (due to = :33). If the required memory <strong>for</strong> the<br />
visco-acoustic case is n 2 log 2<br />
n (see Chapter 2) the memory required in visco-elastic<br />
case, <strong>for</strong> = :33, can be written as 4(2n) 2 log 2<br />
(2n) 16n 2 log 2<br />
n. The rst factor<br />
4 comes from the fact that we need a 2 2 matrix to solve atwo component vector<br />
at each point instead of a single value in the scalar visco-acoustic case. The factor<br />
2, in 2n instead of n, comes from the double grid size. The calculation shows that<br />
16 times more memory is required in the visco-elastic case, although some small<br />
extra overhead in memory will e required.<br />
The actual required memory in this<br />
case was 250 MB <strong>for</strong> a grid size of 310 258 grid points (the linear system with<br />
approximately 160,000 variables) as opposed to 15 MB in the visco-acoustic case<br />
(on a 155 129 grid and a linear system with approximately 20,000 variables). If<br />
the old second order scheme were used, without nested dissection, the required grid<br />
size would be 1162 967, which would require 130 GB of RAM; the memory saving<br />
<strong>for</strong> this case is thus about 99.9%. For the Whitchester model, including 2 sources<br />
and 51 frequencies, the total CPU time <strong>for</strong> the visco-elastic scheme was about two<br />
hours and <strong>for</strong>ty minutes. The total CPU time <strong>for</strong> the visco-acoustic scheme is about<br />
6 minutes. This increase in CPU time of 32 times can be calculated theoretically<br />
146
using the equation (2.21) and assuming a double grid size ( = :33):<br />
CPU elastic =CP U acoustic = k 4(2n)3<br />
kn 3 = 4 (2) 3 =32 (5.40)<br />
The factor 4 in the elastic CPU time comes from the fact that a 2 2 matrix is a<br />
single element in the matrix (vector instead of scalar value).<br />
5.5 Conclusion<br />
In this Chapter I have shown that it is possible to dramatically improve on<br />
standard second order nite dierence schemes <strong>for</strong> visco-elasticity without increasing<br />
computational costs. It would appear that <strong>for</strong>mer limitations on second order<br />
schemes were due to the shape of the dierencing operators; by reshaping these<br />
operators one can use models with high values of Poisson's ratio in a manner not<br />
previously possible with frequency <strong>domain</strong> schemes. This has been achieved by extending<br />
the grid rotation technique proposed by Cole (1994) and Jo et al. (1996) to<br />
the visco-elastic case. The technique would appear to be quite generally useful, and<br />
worthy of testing in other applications of the nite dierence method. A substantial<br />
increase in accuracy is achieved with little or no increase in computational costs.<br />
I would expect signicant improvements in 3-D, due to possibility of combining<br />
rotation in each Cartesian plane with the original scheme.<br />
Ihave shown analytically the improvements in accuracy <strong>for</strong> homogeneous media,<br />
and I have <strong>for</strong>mally proven that the scheme predicts the correct shear wave behaviour<br />
in uid layers. Using my numerical scheme I was able to successfully model<br />
crosshole eld data from a highly heterogeneous sedimentary environment known<br />
to be anisotropic and strongly attenuating. To do this I made several simplifying<br />
assumptions (a constant Poisson's ratio, a homogeneous, constant Q attenuation,<br />
a homogeneous, elliptical anisotropy, and simple, point <strong>for</strong>ce source mechanisms).<br />
Nevertheless, I was able to generate a synthetic data set qualitatively consistent<br />
with the eld data.<br />
147
Chapter 6<br />
Conclusions and further work<br />
6.1 Conclusions<br />
The primary objective of the research described in this thesis was to develop<br />
and implement a sequence of improvements in numerical <strong>seismic</strong> <strong>modelling</strong> that<br />
would allow ecient simulation of large scale, multi-source <strong>seismic</strong> surveys, and to<br />
apply the resultant method to a number of test problems. A secondary objective<br />
of the research was to use the resulting <strong>modelling</strong> code as the basis <strong>for</strong> a waveeld<br />
inversion method, and to test the inversion method on a representative data set.<br />
It was decided early on that the method of choice to meet these objectives was<br />
the frequency <strong>domain</strong> nite dierence method. Although Marfurt (1984) pointed<br />
out the potential of frequency <strong>domain</strong> nite dierences more than a decade ago,<br />
little use has been made of his suggestion since, although an elementary version<br />
of this approach has been used successfully <strong>for</strong> waveeld inversion <strong>for</strong> several years<br />
(Pratt et al., 1995; Pratt et al., 1996; Song et al., 1995). The details of the <strong>modelling</strong><br />
method used in these studies were given by Pratt (1990); the method was a relatively<br />
unsophisticated implementation of simple, second order approach and was not useful<br />
<strong>for</strong> large problems.<br />
The research proceeded by rst developing and implementing a nested dissection<br />
method <strong>for</strong> solving the matrix equations in frequency <strong>domain</strong> nite dierences,<br />
148
then developing and implementing a rotated operator approach <strong>for</strong> reducing the<br />
number of grid points required <strong>for</strong> visco-acoustic <strong>modelling</strong>. The combination of<br />
these two techniques led to signicant increases in numerical eciency. Once these<br />
improvements were in place, the waveeld inversion software was updated to include<br />
these and an extensive study of a real data set was carried out using the new methods.<br />
This led to the conclusion that a visco-elastic approach was required. The nal<br />
chapter of this thesis describes the development of the nested dissection and rotated<br />
operator approaches to the visco-elastic <strong><strong>for</strong>ward</strong> <strong>modelling</strong> problem. This makes the<br />
future development of a visco-elastic waveeld inversion procedure possible.<br />
6.1.1 Matrix solvers<br />
It is a primary conclusion of this project that in order to retain the potential<br />
advantages of frequency <strong>domain</strong> <strong>seismic</strong> <strong>modelling</strong> (to eciently solve the multiple<br />
source problem), one has to use a direct matrix solver. Although there it may be<br />
possible to solve single source, monofrequency problems by using an iterative matrix<br />
solver, the computational cost involved in solving realistic, multiple source problems<br />
will inevitably, Ibelieve, involve the use of optimised direct matrix solvers.<br />
Large computational savings can be achieved if appropriate care is taken with<br />
the initial grid ordering. Nested dissection is an optimal solution to the grid ordering<br />
problem. The memory requirements can be cut down from an n 3 requirement (<strong>for</strong><br />
sequential ordering) to an n 2 log 2<br />
(n) requirement, where n is the number of grid<br />
points in one direction in a square model.<br />
If a realistic value of n is used (n <br />
300), the savings in memory requirements can be over 70% just by using the nested<br />
dissection instead of the ordinary grid ordering.<br />
Regardless of the method used to solve the matrix equations, the size of the<br />
dierence operator has to be kept as small as possible, <strong>for</strong> frequency <strong>domain</strong> nite<br />
dierence methods. This is because, when nested dissection is used, the increases<br />
in memory requirements due to the larger dierence operators are unlikely to be<br />
149
compensated <strong>for</strong> by the accuracy gained by using a higher order of dierence operator.<br />
Ihave shown that the number of grid points per wavelength would have tobe<br />
decreased by more than 50% in order to justify the use of the higher order dierence<br />
operators if the grid is ordered by the nested dissection. Although in some cases this<br />
may beachieved, <strong>for</strong> the level of accuracy one would normally require this actually<br />
would involve sampling at less than the Nyquist criterion.<br />
6.1.2 Rotated nite dierence operators<br />
I have shown that the introduction of rotated nite dierence operators and<br />
lumped mass terms can increase accuracy without any signicant increase in computing<br />
costs <strong>for</strong> both acoustic and elastic methods. This is a technique that has<br />
very general potential application, to a wide range of nite dierence methods. For<br />
the acoustic scheme, this step, in conjunction with the nested dissection method <strong>for</strong><br />
grid ordering, has reduced the memory requirements by 96:4%, <strong>for</strong> a given velocity<br />
model of a realistic size (30 30 wavelengths).<br />
6.1.3 Visco-elastic <strong><strong>for</strong>ward</strong> <strong>modelling</strong><br />
By developing the rotated operator and lumped mass methods and applying<br />
them to the visco-elastic problem, I was able to achieve even greater increases in<br />
computational eciency. Due to a reduction from 15 grid points to 4 grid points per<br />
wavelength, and the use of the nested dissection implementation, the full memory<br />
saving (<strong>for</strong> the elastic scheme) is 99%, <strong>for</strong> a given velocity model of a realistic size<br />
(50 50 wavelengths). I anticipate that this development, implemented on the<br />
appropriate hardware, will allow the routine production of time <strong>domain</strong> 2D full<br />
multiple source pre-stack data <strong>for</strong> realistic, 2D data problems in the near future.<br />
Currently we can solve a visco-elastic model with 250 50 S-wavelengths using a<br />
machine with 512 MB of RAM, and produce results <strong>for</strong> hundreds of sources within<br />
three to four days.<br />
150
The visco-elasticscheme Ihave presented istheoretically capableof<strong>modelling</strong><br />
the eect of uid layers on <strong>seismic</strong> wave propagation. I was not jet able to solve<br />
all the aspects of the problem, which require that the boundary conditions and the<br />
source denition be properly handled <strong>for</strong> uid layers, however it is anticipated that<br />
these problems can also be overcome.<br />
6.1.4 Waveeld inversion<br />
Having developed the necessary <strong><strong>for</strong>ward</strong> <strong>modelling</strong> code, the routines were<br />
used to improve the eciency of an existing <strong>seismic</strong> waveeld inversion method<br />
(Song, 1994). This allowed a large tomographic data set (from the Grimsel Rock<br />
Laboratory) to be eectively modelled and inverted within a fraction of the time<br />
required using the original code, and using a fraction of the memory requirements<br />
(5% of the originally required memory).<br />
The application of waveeld inversion to the data from the Grimsel Rock Laboratory<br />
showed clearly the advantages of waveeld inversion over simple, traveltime<br />
methods. These advantages were rst conrmed on a synthetic data set (generated<br />
by a third party), demonstrating the potential resolution advantages of the waveeld<br />
approach. In order to invert the eld data, a large number of tests with variable<br />
smoothing constraints and variable levels of anisotropy were run. It was in the ecient<br />
computation of these test results that the fast <strong><strong>for</strong>ward</strong> <strong>modelling</strong> routines were<br />
particularly useful. Such tests would have been too expensive without the improvements<br />
introduced by a nested dissection and the rotated nite dierence operators.<br />
The results eectively prove the utility of the frequency <strong>domain</strong> approach as a basis<br />
<strong>for</strong> the production waveeld inversion of multiple source <strong>seismic</strong> transmission data.<br />
The data example in Chapter 4 showed the manner in which the inversion<br />
parameters, specically the smoothing constraints and the anisotropy level, can be<br />
estimated from the results of a set of parameter tests, using the level of data mist<br />
and the solution roughness as a guide in the selection of the parameters (after Pratt<br />
151
(1992)).<br />
I have also shown how sensitive the nal images can be to even a low level<br />
of <strong>seismic</strong> anisotropy. I there<strong>for</strong>e conclude that the inclusion of anisotropy in the<br />
waveeld inversion may beextremely important step to take in the near future.<br />
6.2 Future work<br />
There are clear avenues <strong>for</strong> future research into the techniques that have<br />
been developed in this thesis. These topics can be divided into two main topics:<br />
i) Developments in the <strong>modelling</strong> methods and ii) developments in the waveeld<br />
inversion techniques.<br />
6.2.1 Developments in <strong>seismic</strong> <strong>modelling</strong><br />
Simple improvements on the existing codes<br />
There are a number of possible simple improvements in the existing <strong>modelling</strong><br />
codes that will improve the <strong>modelling</strong> speed. Currently we require that the<br />
sources and the receivers be located exactly at grid point locations. Instead we can<br />
interpolate the waveeld in order to nd the receiver responses at intermediate grid<br />
positions. The source can be described over a small region, allowing the eective<br />
location of the source to also be interpolated to intermediate grid points. To do this<br />
we may use the <strong>for</strong>mulation suggested by Alterman and Aboudi (1970). This will<br />
enable us to progressively increase the grid size with the increase in frequency and<br />
reduce the computational time even further.<br />
A second important improvement will be to optimise the generation of the<br />
time <strong>domain</strong> output in the codes. The current codes update the time <strong>domain</strong> output<br />
trace by trace after each <strong><strong>for</strong>ward</strong> <strong>modelling</strong> step is nished. Although this does not<br />
initially seem to be a time consuming task it can represent a bottleneck in the<br />
computations. For example if we generate 3 GB of synthetic <strong>seismic</strong> data, we may<br />
152
need to run 200 frequencies during the <strong>modelling</strong>. This implies that we have to read<br />
and write 1200 GB of data during the computation. The maximal disk I/O speed on<br />
fast wide SCSI II disks is 20 MB/s. Thus 17 hours would be wasted on unnecessary<br />
disk I/O operations. We should just save the required frequency <strong>domain</strong> data at<br />
the receiver positions after each frequency/source step and per<strong>for</strong>m the inverse FFT<br />
at the end of the <strong>modelling</strong> run. This would signicantly improve per<strong>for</strong>mance if a<br />
large amount of time <strong>domain</strong> data is required as output. There is a demand <strong>for</strong> a<br />
<strong>modelling</strong> code which can generate realistic 2D (or 2.5D) pre-stack data sets. The<br />
code is well suited <strong>for</strong> simulating full 2D eld experiments <strong>for</strong> a variety of purposes,<br />
such as processing and acquisition testing.<br />
Currently, the main problem with the elastic scheme is to redene the source<br />
description mechanism, which eectively prevents us from positioning sources within<br />
uid layers.<br />
The solution to this problem is still under investigation. The other<br />
necessary improvement in the elastic code is to improve on the current absorbing<br />
boundary conditions. Currently we use the one way wave equation (Clayton and<br />
Enquist, 1985; Pratt, 1990b) which cannot cope with high values of Poisson ratio.<br />
One easy way to improve this is to use sponge boundary conditions (Cerjan et al.,<br />
1985; Shin, 1995). However this is not an ideal solution. Sponge boundary conditions<br />
require an increase in the model size to accommodate the absorbing boundary. As we<br />
have seen in the case of frequency <strong>domain</strong> <strong>modelling</strong>, the main problem is to reduce<br />
the model size as much as possible so this increase will not be welcome. There is<br />
a possibility that the rotated nite dierence based approach could be extended to<br />
the boundary conditions. This may improve the absorbing boundary, since schemes<br />
based on the rotated operators are more stable <strong>for</strong> the high Poisson ratios. A linear<br />
combination of the absorbing boundary conditions based on rotated operators and<br />
ordinary operators (as we use <strong>for</strong> the full wave equation) may reduce reections from<br />
the edges of the model.<br />
153
Extensions to more complex cases<br />
The extensions of the rotated nite dierence frequency <strong>domain</strong> techniques<br />
to anisotropic media is the next step to be taken. We have seen the eect of the<br />
low level anisotropy on the waveeld images in Chapter 4.<br />
In order to improve<br />
the quality of the synthetics and the waveeld images we will have to simulate<br />
anisotropy. Extension to TI anisotropic case (<strong>for</strong> example like Tsingas et al. (1990))<br />
can be relatively easily implemented, but the required accuracy will depend on the<br />
nature and the level of the anisotropy. In order to simulate a low level of anisotropy<br />
we will have to take great care of numerical accuracy and numerical anisotropy.<br />
Higher order nite dierence operators may in fact per<strong>for</strong>m better in this case, as<br />
we will require extremely low numerical anisotropy and high accuracy; I do expect<br />
that fourth order in space will be sucient. If we use fourth order operators, we<br />
will be able to use at least four second order and two fourth order schemes in a<br />
combined operator. There is a possibility that we may be able to dene additional<br />
second order operators.<br />
With more degrees of freedom in search of the optimal<br />
coecients, one could hope to nd the scheme which will need not more than ve<br />
grid points per shortest wavelength. This accuracy would be sucient to enable us<br />
to run realistic 2D exploration models on existing top-range workstations.<br />
Full 3D, production anisotropic <strong>modelling</strong> is still beyond us. The <strong>for</strong>mulation<br />
of a frequency <strong>domain</strong> 3D scheme is however straight<strong><strong>for</strong>ward</strong>, and I would expect<br />
to be able to run small 3D examples (of the order of tens of wavelengths in all<br />
directions) within two to three years. However we will have to wait <strong>for</strong> about ve<br />
years from then to model full, realistic 3D surveys (using the acoustic wave equation<br />
to begin with). These predictions assume that the amount of available memory on<br />
workstations eectively doubles every year or two (as it has <strong>for</strong> last fteen years).<br />
Accuracy in 3D <strong>modelling</strong> should not be a problem since we can utilise at least<br />
four rotated coordinate systems without increasing the nite dierence operator<br />
154
size. With such a number of possible second order schemes one would expect to<br />
achieve high accuracy. In the meantime I would expect that the implementation<br />
of the rotated nite dierence techniques to 3D time <strong>domain</strong> <strong>seismic</strong> <strong>modelling</strong> will<br />
produce a computationally inexpensive solution (in comparison with the existing<br />
schemes). Low order, high accuracy operators will enable the use of coarse grids<br />
with large time steps required while the CPU time per grid point/time step will<br />
be low. If the time <strong>domain</strong> nite dierence computation is per<strong>for</strong>med only in the<br />
regions close to the wave fronts the CPU time can be further reduced (this is similar<br />
to the reduced time idea in frequency <strong>domain</strong>).<br />
In this way we may be able to<br />
improve the speed <strong>for</strong> 3D <strong>modelling</strong>. The small spatial extent of nite dierence<br />
operators will enable easy utilisation of parallel computer architectures.<br />
6.2.2 Developments in waveeld inversion<br />
In Chapter 5 I have presented an ecient visco-elastic <strong>modelling</strong> technique.<br />
The next step will be to implement aninversion algorithm which will use it. In principle<br />
the existing inversion code can be extended to the elastic case. The potential<br />
benets are improved imaging and the recovery of additional elastic parameters. We<br />
mayeven be able to obtain high resolution images of the Poisson ratio, an interesting<br />
parameter <strong>for</strong> the oil industry. The problem in elastic inversion will be to simulate<br />
(and invert) the correct source signature together with the source mechanism. In<br />
the acoustic case the only available source mechanism is a P-wave source, with a<br />
circular radiation pattern. In the elastic case the wave<strong>for</strong>ms can vary dramatically<br />
as a result of the source type used. We haveto adjust the source amplitudes of the<br />
source generated P and Swaves, and we may haveto use complex synthetic source<br />
mechanisms to reproduce the observed far eld source behaviour. This may be the<br />
main secret of a successful elastic inversion of a eld data. In the elastic waveeld<br />
inversion case the correct 3D source behaviour may bemore important than in the<br />
acoustic case so the extension of the elastic scheme to 2.5D may be required.<br />
155
Although waveeld inversions have been used <strong>for</strong> quite some time, little is<br />
known about the appropriate data processing sequences. The data processing required<br />
<strong>for</strong> waveeld inversion is dierent from the processing required <strong>for</strong> conventional<br />
purposes. Any processing step which may inuence the wave<strong>for</strong>ms (even a<br />
simple bandpass lter) may eect on the nal result. The processing example shown<br />
in Chapter 4 may not work on other datasets. We have had success with rst arrival<br />
windowing on many eld data sets. This is due to the relatively simple acoustic assumption<br />
used in the inversion procedure, and to the fact that the main diraction<br />
in<strong>for</strong>mation is contained in the rst arrival. The longer the time window the more<br />
likely it is that important S-wave phases and conversions may be included in the<br />
data; windowing eectively excluded non P-wave events. When we extend the inversion<br />
algorithm to the elastic case the images will be improved as will resolution but<br />
we will not wish to use restrictive windowing as a pre-processing approach. Elastic<br />
inversion may cope better with complex wave<strong>for</strong>ms, but other signal generated noise<br />
(<strong>for</strong> example tube waves) can generate the undesired image artefacts. The use of a<br />
longer time window may imply the use more frequencies in the inversion (to improve<br />
frequency <strong>domain</strong> sampling) but there is a possibility of more local minima.<br />
We will have to nd the appropriate processing which will remove such signal<br />
generated noise from the data, without adversely aecting the wave<strong>for</strong>ms. Additional<br />
problems can be expected if a recorded amplitudes are aected by eg coupling<br />
problems (as in the example in Chapter 4).<br />
There is an outstanding question of appropriate data weighting. Ihave shown<br />
in Chapter 4 that a distortion of the recorded signal amplitudes can additionally<br />
help with convergence and the resolution of the images.<br />
It remains to nd an<br />
appropriate way of working with the data amplitudes in productive way.<br />
In the<br />
transmission surveys we have had usually only small amplitude ranges in the data<br />
(with the exception of the example from Chapter 4). If the technique evolves into<br />
one which will also work on reection data sets, the main data amplitudes will be<br />
156
in the direct arrivals. We are usually very interested in the reections from the oil<br />
reservoirs, which are much weaker. The deeper the reector from which the data<br />
comes from the weaker is the signal going to be. Thus the dierence between the<br />
modelled signal and the eld signal will be small in comparison with the direct<br />
arrival. If we do not take the amplitude decay with depth into account the resulting<br />
image will be dominated by the in<strong>for</strong>mation from the part of the signal with the<br />
highest amplitude. To prevent this we will have to think the way of scaling the<br />
gradient vector with the depth in order to enhance the in<strong>for</strong>mation from the deep<br />
weak arrivals. One way will be to use the reciprocal of the waveeld to multiply the<br />
gradient vector so we can enhance the deep signal.<br />
Although we need large grids to model the data eectively in order to prevent<br />
numerical artefacts in the inversion procedure, we cannot resolve the model at very<br />
ne scales (below <strong>seismic</strong> resolution). Inverting <strong>for</strong> all parameters in the model is<br />
not necessary nor desirable, as it involves higher computational costs and additional<br />
potential convergence to local minima. Alternatively, we can use a more sparsely<br />
varying (inversion) model parameters at the level of the <strong>seismic</strong> resolution (or even<br />
more sparsely initially to prevent convergence to a local minima). The idea comes<br />
from Williamson (1990) and Bunks et al. (1995). The ideas of using certain parts of<br />
the signal spectrum at the time are inherent part of the frequency <strong>domain</strong> waveeld<br />
inversion; the only problem is to use the correct model parametrisation at each<br />
frequency.<br />
This approach will reduce the computational costs of the frequency<br />
<strong>domain</strong> waveeld inversion even further and reduce the possibility of convergence to<br />
the local minimum of the mist function.<br />
Although the methods described in the thesis are all 2D, the results show how<br />
little in<strong>for</strong>mation we still use from the data with conventional techniques, and the<br />
improvements we can expect once we start using more in<strong>for</strong>mation from the data<br />
in the industry. As exact reservoir positioning becomes more and more important,<br />
more accurate (but expensive) techniques may be considered, even in 2D, in order<br />
157
to nd more dicult targets.<br />
An increase in the data quantity will not help if<br />
the data processing is too simplied. With extensions to the more complex cases<br />
(eg elastic, anisotropic) we may expect to produce detailed, and quantitative depth<br />
images which may be related to the site geology and which will help predicting<br />
parameters of great importance, such as the Poisson ratio, the fracture orientation,<br />
etc.<br />
Hopefully the improvements in acquisition such as the development of new<br />
sources capable of generating higher frequency data, the development of see bottom<br />
cables (recording shear waves), and recording longer osets will increase the data<br />
resolution and quality and will produce data which are more suited to the inversion<br />
techniques.<br />
The application of the methods I have shown are not limited to the examples<br />
covered in this thesis. The implementation of the rotated nite dierence operators<br />
<strong>for</strong> the time <strong>domain</strong> based nite dierence methods may prove to be the easiest way<br />
to reduce the cost of 3D <strong>seismic</strong> <strong>modelling</strong>. Historically, geophysicist have learned<br />
from medical science how to per<strong>for</strong>m tomography. Similarly, awider audience may<br />
discover applications of the <strong>modelling</strong> and inversion techniques described in this<br />
thesis to similar problems in other disciplines.<br />
158
Bibliography<br />
Abramovici, F., and Alterman, Z., 1965, Computations pertaining to the problem<br />
of propagation of a <strong>seismic</strong> pulse in a layered solid: Methods in Computational<br />
Physics, 4, 349{379.<br />
Al-Chalabi, M., 1994, Seismic velocities-a critique: First Break, 12, 589{596.<br />
Alterman, Z. S., and Aboudi, J., 1970, Source of Finite Extent, Applied Force and<br />
Couple in an Elastic Half-Space: Geophys. J. R. astr. Soc., 21, 47{64.<br />
Alterman, Z., and Karal Jr, F. C., 1968, Propagation of elastic waves in layered media<br />
by nite dierence methods: Bulletin of the Seismological Society of America,<br />
58:1, 367{398.<br />
Alterman, Z., and Kornfeld, P., 1968, Finite dierence solution <strong>for</strong> pulse propagation<br />
in a sphere: Israel Journal of Technology, 6, 138{149.<br />
Alterman, Z., and Loewenthal, D., 1970, Seismic waves in a quarter and threequarter<br />
plane: Geophys. J. Roy. Astr. Soc., 20, 101{126.<br />
Alterman, Z., and Rotenberg, A., 1969, Seismic waves in quarter planes: Bull.<br />
Seism. Soc. Am., 59, 347{368.<br />
Bamberger, A., Chavent, G., and Lailly, P., 1980, Etude de schemas numeriques<br />
pour les equations de l'elastodynamique lineaire:<br />
INRIA (Institut National de<br />
Recherche en In<strong>for</strong>matique et en Automatique), 032-79.<br />
159
Bathe, K. J., and Wilson, E. L., 1976, Numerical methods in niteelement analysis:<br />
Prentice-Hall Inc.<br />
Berkhout, A. J., and Van Wulten Palthe, D. W., 1979, Migration in terms of<br />
spatial deconvolution: Geophys. Prosp., 27, 261{291.<br />
Berkhout, A. J., 1985, Seismic migration: Imaging of acoustic energy by wave eld<br />
extrapolation. A. Theoretical aspects: Elsevier Science Publ. Co., Inc.<br />
Beydoun, W. B., and Mendes, M., 1989, Elastic ray-born l 2<br />
migration/inversion:<br />
Geophysics Journal, 97, 151{160.<br />
Beysal, E., Koslo, D. D., and Sherwood, J. W. C., 1983, Reverse time migration:<br />
Geophysics, 48, 1514{1524.<br />
Bleistein, N., 1986, Two-and-one-half dimensional in-plane wave propagation: Geophys.<br />
Prosp., 34, 686{703.<br />
Boore, D. M., 1972, Finite-dierence methods <strong>for</strong> <strong>seismic</strong> wave propagation in<br />
heterogeneous materials in Bolt, B. A., Ed., Methods in computational physics,<br />
11:: Academic Press.<br />
Bording, R. P., Gersztekorn, A., Lines, L. R., Scales, J. A., and Tritel, S., 1987,<br />
Applications of <strong>seismic</strong> traveltime tomography: Geophys. J. R. astr. Soc., 90, 285{<br />
303.<br />
Bunks, C., Saleck, F. M., Zaleski, S., and Chavent, G., 1995, Multiscale <strong>seismic</strong><br />
wave<strong>for</strong>m inversion: Geophysics, 60, no. 5, 1457{1473.<br />
Carcione, J. M., Koslo, D., and Koslo, R., 1988, Visco-acoustic wave propagation<br />
simulation in the earth: Geophysics, 53, 769{777.<br />
160
Carcione, J. M., Koslo, D., Behle, A., and Seriani, G., 1992, A spectral scheme<br />
<strong>for</strong> wave propagation simulation in 3-D elastic-anisotropic media: Geophysics, 57,<br />
no. 12, 1593{1607.<br />
Carcione, J. M., 1995, Constitutive model and wave equations <strong>for</strong> linear, viscoelastic,<br />
anisotropic media: Geophysics, 60, no. 2, 537{548.<br />
Cerjan, C., Kosolo, D., Kosolo, R., and Rashef, M., 1985, A nonreecting boundary<br />
condition <strong>for</strong> discrete acoustic and elastic wave equation: Geophysics, 50, 705{<br />
708.<br />
Cerveny, V., Popov, M. M., and Psencik, I., 1982, Computation of waveelds in<br />
inhomogenous media - Gaussian beam approach: Geophys. J. R. astr. Soc., 70,<br />
109{128.<br />
Chapman, C. H., and Pratt, R. G., 1992, Traveltime tomography in anisotropic<br />
media{i. theory: Geophys. J. Intl., 109, 1{19.<br />
Chapman, C. H., 1985, Ray theory and its extensions: WKBJ and Maslov seismograms:<br />
J. Geophys., 58, 27{43.<br />
Claerbout, J. F., and Doherty, S. M., 1972, Downward continuation of move-outcorrected<br />
seismograms: Geophysics, 37, 741{768.<br />
Claerbout, J. F., 1970, Coarse grid calculations of waves in inhomogeneous media<br />
with applications to delineation of complicated <strong>seismic</strong> structure: Geophysics, 35,<br />
407{418.<br />
Claerbout, J. F., 1976, Fundamentals of geophysical data processing: McGraw-Hill<br />
Book Co.<br />
Clayton, R. W., and Enquist, B., 1985, Absorbing boundary conditions <strong>for</strong> acoustic<br />
and elastic wave equation: Geophysics, 50, 705{708.<br />
161
Cole, J. B., 1994, A nearly exact second-order nite-dierence time-<strong>domain</strong> wave<br />
propagation algorithm on a coarse grid: Computers in Physics, 8, 730{734.<br />
Dablain, M. A., 1986, The application of high-order dierencing to the scalar wave<br />
equation: Geophysics, 51, 54{66.<br />
Dai, N., Vadis, A., and Kanasewich, E. R., 1995, Wave propagation in heterogeneous,<br />
porous media: A velocity-stress, nite-dierence method: Geophysics, 60,<br />
327{340.<br />
Devaney, A. J., 1984, Geophysical diraction tomography: Inst. Electr. Electron.<br />
Eng., GE-22, 3{13.<br />
Dines, K. A., and Lytle, R. J., 1979, Computerised geophysical tomography: Proc.<br />
Inst. Electr. Electron. Eng., 67, 471{480.<br />
Dyer, B. C., and Worthington, M. H., 1988, Seismic reection tomography: A case<br />
study: First Break, 6, 354{366.<br />
Emmerich, H., and Korn, M., 1987, Incorporation of attenuation into time-<strong>domain</strong><br />
computations of <strong>seismic</strong> wave elds: Geophysics, 52, 1252{1264.<br />
Gauthier, O., Virieux, J., and Tarantola, A., 1986, Two-dimensional nonlinear<br />
inversion of <strong>seismic</strong> wave<strong>for</strong>ms: Geophysics, 51, 1387{1403.<br />
Gazdag, J., 1981, Modeling of the acoustic wave equation with trans<strong>for</strong>m methods:<br />
Geophysics, 46, 845{859.<br />
Gelbke, C., Miranda, F., and Sattel, G., 1989, Results of a <strong>seismic</strong> transmission<br />
tomography survey at the grimsel rock laboratory: The Log Analyst, July-August,<br />
243{260.<br />
George, A., and Liu, J. W., 1981, Computer solution of large sparse positive denite<br />
systems: Prentice-Hall.<br />
162
Hatton, L., Worthington, M. H., and Makin, J., 1986, Seismic data processing:<br />
Theory and practice: Blackwell Scientic Publications.<br />
Hemon, C., 1978, Equations d'onde et modeles: Geophys. Pros., 26, 790{821.<br />
Holbrook, W. S., Mooney, W. D., and Christensen, N. I., 1992, The <strong>seismic</strong> velocity<br />
structure of the deep continental crust: Developments in Geotectonics 23,<br />
Continental Lower Crust.<br />
Holdberg, O., 1987, Computational aspects of the choice of operator and sampling<br />
interval <strong>for</strong> numerical dierentiation in large-scale simulation of wave phenomena:<br />
Geophysical Prospecting, 35, 629{655.<br />
Hu, L. Z., McMechan, G. A., and Harris, J. M., 1988, Acoustic prestack migration<br />
of cross-hole data: Geophysics, 53, 1015{1023.<br />
Jo, C. H., Shin, C. S., and Suh, J. H., 1996, An optimal 9 point, nite dierence,<br />
frequency-space, 2-d scalar wave extrapolator: Geophysics, 61, 529{537.<br />
Johnson, O. G., 1984, Three-dimensional wave equation computations on vector<br />
computers: Proc. of the Inst. Elect. and Electron Eng., 72, 90{95.<br />
Kelly, K.R., Ward, R. W., Treitel, S., and Al<strong>for</strong>d, R. M., 1975, Synthetic seismograms:<br />
A nite dierence approach: Geophysics, 41, 2{27.<br />
Kerner, C., 1990, Modelling of soft sediments and liquid-solid interfaces: modied<br />
wavenumber summation method and application: Geophysical Prospecting, 38,<br />
111{137.<br />
Kjartansson, E., 1979, Constant Q-wave propagation and attenuation: Journal<br />
Geophysical Research, 84.(B9), 4737{4748.<br />
Kolb, P., Collino, F., and Lailly, P., 1986, Pre-stack inversion of a 1-D medium:<br />
74, Proceedings of the IEEE, 498{508.<br />
163
Lailly, P., 1984, Migration methods: partial but ecient solutions to the <strong>seismic</strong><br />
inverse problem in Santosa, F., Pao, Y. H., Symes, W., and Holland, C., Eds.,<br />
Inverse problems of acoustic and elastic waves:: Soc. Industr. Appl. Math.<br />
Loewenthal, D., and Mufti, I. R., 1983, Reverse time migration in spatial frequency<br />
<strong>domain</strong>: Geophysics, 48, 627{635.<br />
Loewenthal, D., Roberson, L., Lu, R., and Sherwood, J. W. C., 1976, The wave<br />
equation applied to migration: Geophys. Prosp., 24, 380{399.<br />
Lysmer, and Drake, 1972, A nite element method <strong>for</strong> seismology: Methods in computational<br />
physics: Volume 11: Seismology: Surface waves and earth oscillations,<br />
181{216.<br />
Madariaga, R., 1976, Dynamics of an expanding circular fault: Bulletin of the<br />
Seismological Society of America, 66:3, 639{666.<br />
Madariaga, R., 1984, Gaussian beam synthetic seismograms in a vertically varying<br />
medium: Geophys. J. Roy. Astr. Soc., 79, 589{612.<br />
Marfurt, K. J., and Shin, C. S., 1989, The future of iterative modeling of geophysical<br />
exploration in E, E., Ed., Supercomputers in <strong>seismic</strong> exploration:: Pergamon Press,<br />
203{228.<br />
Marfurt, K. J., Sandschraper, J. R., and Corvin, J. B., 1987, Seismic modeling on<br />
the cray-2 supercomputers: Proceedings of the Third International Symposium,<br />
Minneapolis, Minesota, Science and Engineering on Cray Supercomputers.<br />
Marfurt, K. J., 1984a, Accuracy of nite-dierence and nite-elements modeling of<br />
the scalar and elastic wave equation: Geophysics, 49, 533{549.<br />
||{ 1984b, Seismic modeling: A frequency-<strong>domain</strong> nite ellement approach: 54th<br />
Annual International Meeting, Expanded abstracts.<br />
164
McCarthy, J., Larkin, S. P., Fuis, G. S., Simpson, R. W., and Howard, K. A., 1991,<br />
Anatomy of a metamorphic core complex: Seismic refraction/wide-angle reection<br />
proling in Southeastern Cali<strong>for</strong>nia and Western Arizona: Journal of Geophysical<br />
research, 96, 12259{12291.<br />
McMechan, A., 1983, Seismic tomography in boreholes: Geophys. J. R. A. S., 74,<br />
601{612.<br />
Michell, A. R., 1969, Computational methods in partial dierential equations: New<br />
York: John Wiley & Sons.<br />
Mora, P. R., 1987, Nonlinear two-dimensional elastic inversion of multioset <strong>seismic</strong><br />
data: Geophysics, 52, 1211{1228.<br />
Mora, P., 1989a, Modeling anisotropic <strong>seismic</strong> waves in 3-D: 59th Annual International<br />
Meeting, Expanded Abstracts, 1039{1043.<br />
||{ 1989b, Inversion= migration + tomography: Geophysics, 12, 1575{1586.<br />
Muir, F., Dellinger, J., Etgen, J., and Nichols, D., 1992, Modeling elastic elds<br />
across irregular boundaries (short note): Geophysics, 57, no. 9, 1189{1193.<br />
Muller, G., 1983, Rheological properties and velocity dispersion of a medium with<br />
power-law dependence of Q on frequency: Geophysics, 54, 20{29.<br />
Neep, J. P., Sams, M. S., Worthington, M. H., and O'Hara-Dhand, K. A., 1996,<br />
Measurement of <strong>seismic</strong> attenuation from high-resolution crosshole data:<br />
Geophysics,<br />
61, 1175{1188.<br />
Ozdenvar, T., and McMechan, G. A., 1996, Causes and reduction of numerical<br />
artefacts in pseudospectral waveeld extrapolation: Geophys. J. Int., 126, 819{<br />
828.<br />
165
Peng, C., Cheng, C. H., and Toksoz, M. N., 1993, Borehole eects on downhole<br />
<strong>seismic</strong> measurements: Geophys. Prosp., 41, no. 7, 883{912.<br />
Pessoa, J. M., and Worthington, M. H., Seismic traveltime tomography at the<br />
Grimsel Rock Laboratory, Annual Report 95-05, NAGRA internal report, 1995.<br />
Peterson, J. E., Paulsson, B. N. P., and McEvilly, T. V., 1985, Applications of<br />
algebraic reconstruction techniques to cross-hole <strong>seismic</strong> data: Geophysics, 50,<br />
1566{1580.<br />
Pratt, R. G., and Chapman, C. H., 1992, Traveltime tomography in anisotropic<br />
media{ii. application: Geophys. J. Intl., 109, 20{37.<br />
Pratt, R. G., and Sams, M. S., 1996, Reconciliation of crosshole <strong>seismic</strong> velocities<br />
with well in<strong>for</strong>mation in a layered sedimentary environment: Geophysics, 61, 549{<br />
560.<br />
Pratt, R. G., and Worthington, M. H., 1990, Inverse theory applied to multisource<br />
cross-hole tomography: Part-I: Acoustic wave-equation method: Geophysical<br />
Prospecting, 38, 287{310.<br />
Pratt, R. G., Shipp, R. M., Song, Z. M., and Williamson, P. R., 1995, Fault<br />
delineation by waveeld inversion of cross-borehole <strong>seismic</strong> data: 57th Conference,<br />
Technical Programme and Abstracts of Papers European Association of Explration<br />
Geophysicists: Extended Abstriacts vol 1.<br />
Pratt, R. G., Song, Z. M., Williamson, P. R., and Warner, M. R., 1996, Twodimensional<br />
velocity models from wide angle <strong>seismic</strong> data by wave<strong>for</strong>m inversion:<br />
Geophysical Journal International, 124, 323{340.<br />
Pratt, R. G., 1989a, An improved <strong>seismic</strong> tomogram using the elastic wave equation:<br />
51st annual meeting, Technical Programme and Abstracts of Papers European<br />
Association of Exploration Geophysicists, 173{174.<br />
166
||{ 1989b, Wave equation methods in cross-hole <strong>seismic</strong> imaging: Ph.D. thesis,<br />
Imperial college of Science and Technology.<br />
Pratt, R. G., 1990a, <strong>Frequency</strong>-<strong>domain</strong> elastic wave modeling by nite dierences:<br />
A <strong>tool</strong> <strong>for</strong> crosshole <strong>seismic</strong> imaging: Geophysics, 55, 626{632.<br />
||{ 1990b, <strong>Frequency</strong> <strong>domain</strong> elastic wave modeling by nite dierences: A<strong>tool</strong><br />
<strong>for</strong> cross-hole <strong>seismic</strong> imaging: Geophysics, 55, 626{632.<br />
Press, W. H., Teukolsky, S. A., Vettering, W. T., and Flannery, B. P., 1992, Numerical<br />
recipes in FORTRAN: The art of scientic computing, second edition:<br />
Cambridge University Press.<br />
Randall, C. J., 1991, Multiple acoustic wave<strong>for</strong>ms in nonaxisymmetric boreholes<br />
and <strong>for</strong>mations: The Journal of the Acoustical Society of America, 90, 1620{1631.<br />
Reshef, M., Koslo, D., Edvards, M., and Hsiung, C., 1988a, Three dimensional<br />
acoustic modeling by the Fourier method: Geophysics, 53, 1175{1183.<br />
||{ 1988b, Three dimensional elastic modeling by the Fourier method: Geophysics,<br />
53, 1184{1193.<br />
Robertsson, J. . A., Blanch, J. O., and Symes, W. W., 1994, Viscoelastic nite<br />
dierence modeling: Geophysics, 59, 1444{1456.<br />
Sei, A., and Symes, W. W., 1994a, Dispersion analysis of numerical wave propagation<br />
and its computational consequences: Journal of Scientic Computing, 10,<br />
1{27.<br />
||{ 1994b, Error analysis of numerical schemes <strong>for</strong> the wave equation in heterogeneous<br />
media: Appl.-Numer.-Math., 15, no. 4, 465{480.<br />
Shin, C., 1995, Sponge boundary condition <strong>for</strong> frequency-<strong>domain</strong> modeling: Geophysics,<br />
60, no. 6, 1870{1874.<br />
167
Smith, W. D., 1974, The application of nite element analysis to body wavepropagation<br />
problem: Geophys. J. R. astr. Soc., 42, 747{768.<br />
Song, Z.-M., and Williamson, P. R., 1995, <strong>Frequency</strong>-<strong>domain</strong> acoustic-wave modeling<br />
and inversion of crosshole data: Part I- 2.5-d modeling method: Geophysics,<br />
60, no. 3, 784{795.<br />
Song, Z. M., and Worthington, M. H., Full Waveeld Inversion <strong>for</strong> the Grimsel<br />
Test Site, Annual Report 95-76, NAGRA internal report, 1995.<br />
Song, Z. M., Williamson, P. R., and Pratt, R. G., 1995, <strong>Frequency</strong>-<strong>domain</strong> acousticwave<br />
modeling and inversion of crosshole data; part ii, inversion method, synthetic<br />
experiments and real data results: Geophysics, 60, 796{809.<br />
Song, Z. M., 1994, <strong>Frequency</strong> <strong>domain</strong> wave equation crosshole <strong>modelling</strong> and inversion:<br />
Ph.D. thesis, Imperial college of Science and Technology.<br />
Stephen, R. A., 1983, A comparison of nite dierence and reectivity seismograms:<br />
Geophys. J. R. Astr. Soc., 72, 39{57.<br />
Subhashis, M., and Frazer, L. N., 1987, Practical aspects of reectivity modeling:<br />
Geophysics, 52, 1355{1364.<br />
Tarantola, A., 1984, Inversion of <strong>seismic</strong> reection data in the acoustic approximation.:<br />
Geophysics, 49, 1259{1266.<br />
Tsingas, C., Vadis, A., and Kanasewich, E. R., 1990, Elastic wave propagation in<br />
transversely isotropic media using nite dierence: Geophysical Prospecting, 38,<br />
933{949.<br />
Virieux, J., 1986a, P-SV wave propagation in heterogeneous media: Velocity-stress<br />
nite-dierence method: Geophysics, 51, 889{901.<br />
168
||{ 1986b, SH-wave propagation in heterogenous media: Velocity-stress nitedierence<br />
method: Geophysics, 49, 1933{1957.<br />
Williamson, P. R., and Worthington, M. H., 1993, Resolution limits in ray tomography<br />
due to wave behavior: Numerical experiments: Geophysics, 58, no. 5,<br />
727{735.<br />
Williamson, P. R., 1990, Tomographic inversion in reection seismology: Geophysical<br />
Journal International, 100, 255{274.<br />
Wong, J., Bergman, N., West, G., and Hurley, P., 1987, Crosshole <strong>seismic</strong> scanning<br />
and tomography: Geophysics: The Leading Edge of Exploration, 36{41.<br />
Worthington, M. H., 1984, An introduction to geophysical tomography:<br />
First<br />
Break, 2, 20{25.<br />
Zeng, X., and West, G. F., 1996, Reducing spurious diractions in elastic waveeld<br />
calculations: Geophysics, 61, 1436{1439.<br />
Zhou, C., Cai, W., Luo, Y., Schuster, G. T., and Hassanzadeh, S., 1985, Acoustic<br />
wave-equation traveltime and wave<strong>for</strong>m inversion of crosshole <strong>seismic</strong> data: Geophysics,<br />
60, 765{773.<br />
Zhu, X., and McMechan, G. A., 1991, Finite-dierence modeling of the <strong>seismic</strong><br />
response of uid-saturated, porous, elastic media using Biot theory: Geophysics,<br />
56, 328{339.<br />
Zienkijevic, O. C., 1977, The nite element method: 3-rd edition: McGraw-Hill.<br />
169
Appendix A<br />
Dispersion analysis <strong>for</strong> visco-elastic <strong>modelling</strong><br />
If we take the second order nite dierence equations generated from equations<br />
(5.17) and (5.18), and use the combined consistent and lumped mass <strong>for</strong>mulations<br />
<strong>for</strong> the density weighted term in the visco-elastic wave equation (5.24 and<br />
5.25), We obtain the following scheme <strong>for</strong> homogeneous media:<br />
"<br />
(1 , b) <br />
#<br />
! 2 b u + u + + u , + u + + u , +<br />
4<br />
"( +2) u+ ,2u+u ,<br />
a<br />
(1 , a) 1 2<br />
"<br />
( +2)<br />
"<br />
! 2 b v +<br />
a<br />
(1 , a) 1 2<br />
"<br />
"<br />
#<br />
+ u +,2u+u ,<br />
+(+) v+ +,v ,+v + ,,v , , +<br />
+<br />
2 2 4 2<br />
!<br />
u + +,2u+u , ,<br />
, u+ , u , , u + + u ,<br />
+ u+ , , 2u + u , +<br />
+<br />
2 2 2 2 2 2<br />
!<br />
u+ +<br />
, 2u + u , ,<br />
+ u+ , u , , u + + u ,<br />
+ u+ , , 2u + u , +<br />
+<br />
2 2 2 2 2<br />
!#<br />
2<br />
v + +<br />
, v , + ( + )<br />
+ v, , , v , +<br />
=0; (A-1)<br />
2 2<br />
(1 , b)<br />
4<br />
<br />
v + + v , + v + + v ,<br />
# +<br />
#<br />
( +2) v +,2v+v ,<br />
+ v+ ,2v+v ,<br />
+(+) u+ +,u + , +u, ,,u , +<br />
+<br />
2 2 4 2 !<br />
v +<br />
( +2)<br />
+,2v+v ,<br />
, + v+ , v , , v + + v ,<br />
+ v+ , , 2v + v , +<br />
+<br />
2 2 2 2 2 2<br />
( + )<br />
!<br />
v+ +<br />
, 2v + v ,<br />
, , v+ , v , , v + + v ,<br />
+ v+ , , 2v + v , +<br />
+<br />
2 2 2 2 2 2<br />
!#<br />
u + +<br />
, u + , + u, , , u , +<br />
=0; (A-2)<br />
2 2<br />
170
where isthegrid pointinterval, u = u m;n , u + = u m+1;n , u = u m,1;n , u + = u m;n+1 ,<br />
u , = u m:n,1 , u + + = u m+1;n+1, u , , = u m,1;n,1, u + , = u m+1;n,1, u , + = u m,1;n+1 and the<br />
equivalent <strong>for</strong> v +;,<br />
+;,.<br />
By substituting a vector plane wave solution<br />
0 1 0<br />
B<br />
@ u C<br />
A = B<br />
v<br />
@ U V<br />
1<br />
C<br />
A e,i r ;<br />
(A-3)<br />
where = ( x ; z ) is the wave vector and r = (x; z) is the position vector, into<br />
equations (A-1) and (A-2), one obtains a homogeneous linear system of two equations<br />
with two unknowns (U and V ). The determinant of this homogeneous system<br />
must equal zero, leading to a quadratic equation in ! in terms of = jj. The<br />
two solutions of this determinant represent the numerical compressional and shear<br />
wave modes. By using the relations <strong>for</strong> the group velocity, v g = ! <br />
and <strong>for</strong> the phase<br />
velocity, v p = @!<br />
@<br />
I obtain the numerical group and phase velocities, bv Pp, bv Pg , bv Sp<br />
and bv Sg . Finally, normalized numerical velocities are obtained by dividing by the<br />
exact values.<br />
The nal expressions depend on K = =2 (the wavenumber in<br />
gridpoint units, i.e., the inverse of G, the number of gridpoints per wavelength), <br />
(the propagation angle), R (the v s =v p ratio in the homogeneous medium), and a and<br />
b (the weighting factors of the rotated and lumped mass schemes):<br />
vu q<br />
bv Pp<br />
= 1 u<br />
t 1 + 2 1 , 4 2 3<br />
; (A-4)<br />
v Pp 2K 2 3<br />
bv Sp<br />
v Sp<br />
=<br />
1<br />
R 2K<br />
bv Pg<br />
v Pg<br />
= 1<br />
2<br />
bv Sg<br />
v Sg<br />
= 1<br />
R 2<br />
vu<br />
u<br />
t 1 ,<br />
vu<br />
u<br />
@ t 1 +<br />
@K<br />
vu<br />
u<br />
@ t 1 ,<br />
@K<br />
q<br />
2 1 , 4 2 3<br />
; (A-5)<br />
2 3<br />
q<br />
2 1 , 4 2 3<br />
; (A-6)<br />
2 3<br />
q<br />
2 1 , 4 2 3<br />
2 3<br />
; (A-7)<br />
where 1 = a [,2+2cos x<br />
]+(1,a)[,1 + cos x<br />
, cos z<br />
+ cos x<br />
cos z<br />
];<br />
2 = a [,2+2cos z<br />
]+(1,a)[,1,cos x<br />
+ cos z<br />
+ cos x<br />
cos z<br />
],<br />
3 = b + (1,b)<br />
2<br />
(cos x<br />
+ cos z<br />
), 1 = 3 ( 1 + 2 )(1 + R 2 ),<br />
171
and 2 = ( 1 R 2 + 2 )( 1 + 2 R 2 ) , (R 4 , 2R 2 +1)sin 2 x<br />
sin 2 z<br />
. Inthe computation<br />
of these coecients, x = cos =2K cos and z = sin =2K sin <br />
are the wavevector components in grid point units. The v s =v p ratio, R is related to<br />
the Poisson ratio, by<br />
R 2 =<br />
<br />
+2 = 0:5,<br />
1,<br />
(A-8)<br />
172