AN HELMHOLTZ ITERATIVE SOLVER FOR THE THREE-DIMENSIONAL SEISMIC IMAGING PROBLEMS?

Similar documents
Migration with Implicit Solvers for the Time-harmonic Helmholtz

A PRECONDITIONER FOR THE HELMHOLTZ EQUATION WITH PERFECTLY MATCHED LAYER

Spectral analysis of complex shifted-laplace preconditioners for the Helmholtz equation

On complex shifted Laplace preconditioners for the vector Helmholtz equation

ON A ROBUST ITERATIVE METHOD FOR HETEROGENEOUS HELMHOLTZ PROBLEMS FOR GEOPHYSICS APPLICATIONS

A decade of fast and robust Helmholtz solvers

Comparison between least-squares reverse time migration and full-waveform inversion

1. Fast Iterative Solvers of SLE

A MULTIGRID-BASED SHIFTED-LAPLACIAN PRECONDITIONER FOR A FOURTH-ORDER HELMHOLTZ DISCRETIZATION.

Kasetsart University Workshop. Multigrid methods: An introduction

Multigrid based preconditioners for the numerical solution of two-dimensional heterogeneous problems in geophysics

EFFICIENT MULTIGRID BASED SOLVERS FOR ISOGEOMETRIC ANALYSIS

W011 Full Waveform Inversion for Detailed Velocity Model Building

AMS526: Numerical Analysis I (Numerical Linear Algebra)

Elliptic Problems / Multigrid. PHY 604: Computational Methods for Physics and Astrophysics II

Solving Symmetric Indefinite Systems with Symmetric Positive Definite Preconditioners

Full-waveform inversion application in different geological settings Denes Vigh*, Jerry Kapoor and Hongyan Li, WesternGeco

Computational Linear Algebra

CONVERGENCE BOUNDS FOR PRECONDITIONED GMRES USING ELEMENT-BY-ELEMENT ESTIMATES OF THE FIELD OF VALUES

Full waveform inversion in the Laplace and Laplace-Fourier domains

MULTIGRID METHODS FOR NONLINEAR PROBLEMS: AN OVERVIEW

Algebraic Multigrid as Solvers and as Preconditioner

Solving PDEs with Multigrid Methods p.1

SUMMARY REVIEW OF THE FREQUENCY DOMAIN L2 FWI-HESSIAN

University of Illinois at Urbana-Champaign. Multigrid (MG) methods are used to approximate solutions to elliptic partial differential

Towards full waveform inversion: A torturous path

Multigrid Methods and their application in CFD

Compressive sampling meets seismic imaging

The solution of the discretized incompressible Navier-Stokes equations with iterative methods

Full-Waveform Inversion with Gauss- Newton-Krylov Method

Geometric Multigrid Methods

ITERATIVE METHODS FOR NONLINEAR ELLIPTIC EQUATIONS

Stabilization and Acceleration of Algebraic Multigrid Method

Aspects of Multigrid

Multigrid absolute value preconditioning

A robust computational method for the Schrödinger equation cross sections using an MG-Krylov scheme

Iterative Methods and Multigrid

An improved two-grid preconditioner for the solution of three-dimensional Helmholtz problems in heterogeneous media

SUMMARY INTRODUCTION. f ad j (t) = 2 Es,r. The kernel

Source estimation for frequency-domain FWI with robust penalties

An Algebraic Multigrid Method for Eigenvalue Problems

Iterative Methods for Solving A x = b

An Efficient Two-Level Preconditioner for Multi-Frequency Wave Propagation Problems

arxiv: v1 [math.na] 11 Jul 2011

P276 Multiparameter Full-Waveform Inversion for Velocity and Attenuation Refining the Imaging of a Sedimentary Basin

A fast method for the solution of the Helmholtz equation

New Multigrid Solver Advances in TOPS

INTERGRID OPERATORS FOR THE CELL CENTERED FINITE DIFFERENCE MULTIGRID ALGORITHM ON RECTANGULAR GRIDS. 1. Introduction

Adaptive algebraic multigrid methods in lattice computations

A New Multilevel Smoothing Method for Wavelet-Based Algebraic Multigrid Poisson Problem Solver

Multigrid Method for 2D Helmholtz Equation using Higher Order Finite Difference Scheme Accelerated by Krylov Subspace

IDR(s) A family of simple and fast algorithms for solving large nonsymmetric systems of linear equations

Adaptive Multigrid for QCD. Lattice University of Regensburg

Spectral element agglomerate AMGe

A parameter tuning technique of a weighted Jacobi-type preconditioner and its application to supernova simulations

The Removal of Critical Slowing Down. Lattice College of William and Mary

DELFT UNIVERSITY OF TECHNOLOGY

Nonlinear seismic imaging via reduced order model backprojection

Preface to the Second Edition. Preface to the First Edition

Contents. Preface... xi. Introduction...

A projected Hessian for full waveform inversion

Compressive Sensing Applied to Full-wave Form Inversion

An Efficient Low Memory Implicit DG Algorithm for Time Dependent Problems

Iterative Methods and Multigrid

A Domain Decomposition Based Jacobi-Davidson Algorithm for Quantum Dot Simulation

Geophysical Journal International

Youzuo Lin and Lianjie Huang

THE EFFECT OF MULTIGRID PARAMETERS IN A 3D HEAT DIFFUSION EQUATION

A Novel Aggregation Method based on Graph Matching for Algebraic MultiGrid Preconditioning of Sparse Linear Systems

6. Multigrid & Krylov Methods. June 1, 2010

Von Neumann Analysis of Jacobi and Gauss-Seidel Iterations

Robust solution of Poisson-like problems with aggregation-based AMG

FWI with Compressive Updates Aleksandr Aravkin, Felix Herrmann, Tristan van Leeuwen, Xiang Li, James Burke

Solving Ax = b, an overview. Program

Multigrid and Domain Decomposition Methods for Electrostatics Problems

Multigrid finite element methods on semi-structured triangular grids

The Conjugate Gradient Method

A MULTIGRID ALGORITHM FOR. Richard E. Ewing and Jian Shen. Institute for Scientic Computation. Texas A&M University. College Station, Texas SUMMARY

AMS526: Numerical Analysis I (Numerical Linear Algebra) Lecture 23: GMRES and Other Krylov Subspace Methods; Preconditioning

Partial Differential Equations

A Robust Preconditioned Iterative Method for the Navier-Stokes Equations with High Reynolds Numbers

Discrete Projection Methods for Incompressible Fluid Flow Problems and Application to a Fluid-Structure Interaction

An advanced ILU preconditioner for the incompressible Navier-Stokes equations

Multigrid and Iterative Strategies for Optimal Control Problems

Comparison of V-cycle Multigrid Method for Cell-centered Finite Difference on Triangular Meshes

2012 SEG SEG Las Vegas 2012 Annual Meeting Page 1

MULTI-LEVEL TECHNIQUES FOR THE SOLUTION OF THE KINETIC EQUATIONS IN CONDENSING FLOWS SIMON GLAZENBORG

Preconditioners for the incompressible Navier Stokes equations

Introduction to Multigrid Methods

Vollständige Inversion seismischer Wellenfelder - Erderkundung im oberflächennahen Bereich

The amount of work to construct each new guess from the previous one should be a small multiple of the number of nonzeros in A.

Iterative Methods for Linear Systems of Equations

A SHORT NOTE COMPARING MULTIGRID AND DOMAIN DECOMPOSITION FOR PROTEIN MODELING EQUATIONS

Efficient smoothers for all-at-once multigrid methods for Poisson and Stokes control problems

Model reduction of wave propagation

A multigrid method for large scale inverse problems

SUMMARY ANGLE DECOMPOSITION INTRODUCTION. A conventional cross-correlation imaging condition for wave-equation migration is (Claerbout, 1985)

Notes for CS542G (Iterative Solvers for Linear Systems)

Nested Krylov methods for shifted linear systems

Scientific Computing II

Transcription:

European Conference on Computational Fluid Dynamics ECCOMAS CFD 2006 P. Wesseling, E. Oñate, J. Périaux (Eds) c TU Delft, Delft The Netherlands, 2006 AN HELMHOLTZ ITERATIVE SOLVER FOR THE THREE-DIMENSIONAL SEISMIC IMAGING PROBLEMS? René-Edouard Plessix Shell International E&P, Kesslerpark 1, 2288 GD, Rijswijk, The Netherlands e-mail: web page: http://www.shell.com Key words: iterative solver, wave equation, Helmholtz, seismic Abstract. A preconditioned iterative solver for the three-dimensional frequency-domain wave equation applied to seismic problem is evaluated. The preconditioner corresponds to the approximate inverse of a heavily damped wave equation deduced from the (undamped) wave equation. The approximate inverse is computed with one multigrid cycle. Numerical results show that the method is robust and that the number of iterations roughly linearly increases with frequency, when the grid spacing is adapted to keep a constant number of discretization points per wavelength. To evaluate the interest of this new iterative solver, the number of floating point operations required for the frequency-domain imaging algorithms and the time-domain imaging algorithms is roughly approximated. This rough estimate indicates that the time-domain migration approach is more than one order of magnitude faster. The velocity model building problem based on a least-squares formulation and a scale separation approach is about three time faster in the frequency domain. 1 INTRODUCTION Imaging the Earth from seismic data require many (thousands of) solutions of the wave equation because multi-shot and multi-receiver data are used to constrain the algorithm. To process large three-dimensional problems, high-frequency (ray) approximations are commonly made. With the increase of the computer power, finite-frequency solutions of the wave equation is now used when high-frequency approximations are not suitable. The finite-frequency solutions are mainly obtained with a finite-difference approach. In order to deal with three-dimensional problems on todays computers, paraxial (or oneway) approximations of the frequency-domain wave equation are used with the depth as the preferred propagation direction. This leads to efficient three-dimensional imaging algorithms. However the paraxial approximation is often inaccurate with large lateral velocity contrasts and with large distances between the source and the receiver. To more precisely handle complex Earth geometries and to improve the processing of long offset data, data with large source-receiver distances, solutions of the wave equation 1

without high-frequency or paraxial approximations are needed. The wave equation can be solved either in the frequency domain or in the time domain. The choice is mainly dictated by the computation time and memory requirements 5. The time-domain wave equation can be solved with a marching approach in time and an explicit scheme. Currently threedimensional imaging algorithms, so-called reverse time migrations, are developed based on time-domain solutions and start to be used on large computer clusters. The frequencydomain wave equation, namely the Helmholtz equation in the acoustic case with constant density, leads to a large sparse linear matrix. With two-dimensional problems a direct solver, based on an LU decomposition, can be used. This is very attractive because only one LU decomposition is required to compute the solutions at multiple source locations. This leads to attractive and efficient imaging methods that are faster than the timedomain imaging methods 5,7. In three-dimensional spaces, due to the size of the matrix, the situation is different and an iterative solver is required. In practice, three-dimensional frequency-domain solutions of the wave equation are not used because most of the iterative solvers are not robust at seismic frequency or are too expensive. The difficulty arises because the matrix of the linear system is not definite, namely it has positive and negative eigenvalues and because the model size corresponds to several hundreds of wavelength in the three space directions, which means that several hundreds of discretization points in each direction are needed. The efficiency of the iterative method depends on the preconditioner. Recently a new preconditioner based on a heavily damped wave equation and a multi-grid solver has been proposed by Erlangga et al. 4. The iterative solver is a preconditioned BI-CGSTAB. The preconditioner corresponds to the approximate inverse of a heavily damped wave-equation after one multigrid cycle. Preliminary results obtained with this solver on two-dimensional problems have been presented by Riyanti et al. 11. Here a three-dimension implementation of this new iterative solver of the wave equation is discussed together with its relevance for the imaging problem in exploration geophysics. Contrary to Erlangga et al. 4,11 the multigrid method is implemented with a standard full-weighting coarsening, a tri-linear interpolation, a matrix-free algorithm, and a 8 th order scheme in space. In this way realistic three-dimensional models up to about 80 million unknowns, for instance corresponding to a grid of 610 by 610 by 240 points, can be computed on a single cpu with 16 GB of RAM. The paper is organized as follows. After a review of the method, a three dimensional numerical example is shown and some convergence plots are presented. Then based on the analysis of numerical results and a rough estimate of the number of floating point operations required for the frequency domain and time domain imaging algorithms, the relevance for the three-dimensional seismic imaging problem is discussed. 2

2 Method 2.1 Iterative solver The acoustic constant density wave equation corresponds to the Helmholtz equation: ω2 u(x, ω) u(x, ω) = s(x, ω), (1) v 2 (x) with u the pressure field, s the source function, v the velocity, ω the angular frequency and x the spatial coordinates. k = ω is the wave number. v A finite-difference discretization with a J order scheme in space gives: Jj=1 [ a J j (u l j,m,p + u l+j,m,p ) + b J j (u l,m j,p + u l,m+j,p )+ ] c J j (u l,m,p j + u l,m,p+j ) + (d J kl,m,p 2 )u l,m,p = s l,m,p. a J j, bj j, cj j and dj are the coefficients of the scheme. u l,m,p, s l,m,p, and k l,m,p are the discrete values of the pressure field, the source function, and the wavenumber. Equation 2 is valid for i, j, k [J, n x J] [J, n y J] [J, n z J]. Outside, the field values are set to 0. In practice, absorbing boundary layers are added to the domain to minick an infinite domain. In the boundary layers, a complex part to the wavenumber is added; this means that the wave field is damped. Therefore the field can be assumed to be 0 close to the boundaries. In a matrix from, equation 2 reads (2) A(k)u = s, (3) where u and s are now the pressure and source vectors. A is generally non-definite, because the left-hand side of equation 1 is the sum of a positive operator ( ) and a negative operator ( k 2 I, with I the identity). With n the number of discretization points in one spatial direction, A is an n 3 n 3 sparse matrix. With realistic geophysical examples, n is between 200 and 1000. Due to the size of the matrix, a direct solver cannot be used. The logical approach to solve the linear system 3 is an iterative method based on a Krylov method, for instance the BI-CGSTAB method. The efficiency of the solver relies on the preconditioner. The preconditioner system, with the preconditioner M, is A(k)M 1 v = f, with v = Mu. (4) With the iterative solver the vector A(k)M 1 v needs to be evaluated. This is carried out in two steps; first the system v = Mu is solved, then the matrix-vector product A(k)u is computed. In order to efficiently solve the system v = Mu, Erlangga et al. 4 proposed to use one multigrid cycle and to consider M = A(k(β r + iβ i )), with i = 1 and β r and 3

β i two (positive) real numbers. The choice of β r and β i should be such that M is close enough to A and that the multigrid method is applicable to M. The first condition means that β r should be close to 1 and β i as small as possible. The second condition means that β i should be large enough to avoid the difficulties associated with the non-definiteness of A(k). It appears that β r close to 1 and β i between 0.25 and 1 generally lead to a correct choice. Physically M corresponds to the discretization of the wave equation with a wave number equal to k(β r +iβ i ), namely the discretization of a heavily damped wave equation. Indeed, assuming k constant, the solution of the damped wave equation is a e i k(βr+iβ i) x with a the amplitude (geometrical divergence) term and. the L2-norm. β i induces an extra damping equal to e kβ i x. Over a wavelength, i.e. for x = 2π, the extra damping k is 0.04, with β i = 0.5. Over a wavelength, 96% of the energy is lost due to this extra damping. The heavily damped wave equation almost behaves like a diffusive equation since the energy hardly propagates after a wavelength. This explains why the multigrid method can be used on M. 2.2 Implementation In order to handle large problems, the implementation of the method differs from the one of Erlangga et al. 4. A matrix-free approach is used. The boundary conditions are Dirichlet boundary conditions and absorbing layers are added to the model to mimic an infinite space. The idea of multigrid is to estimate the oscillatory part of the solution on the fine grid and to estimate the smooth part of a coarser grid. With h the grid spacing on the fine grid, the discretization of the damped wave equation, equation 2 after replacing k by k(β r + iβ i ), leads to the system: v h = M h u h. (5) The superscript h means that the quantities are related to the grid with the spacing h. An approximation, w h, of u h, with the smoother operator, S h (M h, v h, u h 0 ), is computed: w h = S h (M h, v h, u h 0 ). (6) u h 0 is an initial guess for uh, generally u h 0 = 0. In the implementation the smoother is one iteration of the symmetric Gauss-Seidel method. w h contains an approximation of the oscillatory part of the solution, but the smooth part of the solution is difficult to obtain from the Gauss-Seidel method. The residual, r h = v h M h w h, satisfies: r h = M h e h, with e h = u h w h. (7) When w h approximates the oscillatory part of u h, r h and e h have mainly smooth variations. r h and e h should then be correctly approximated on a coarser grid with a spacing 2h. With Ch 2h the coarsening operator from the fine grid to the coarse grid, and P 2h h the prolongation 4

operator from the coarse grid to the fine grid, the system r h = M h e h becomes: r 2h = C 2h h Mh P h 2h e2h with r 2h = C 2h h rh and e h = P h 2h e2h. (8) In the implementation, the coarsening operator is the full-weigthing operator and the prolongation operator is the tri-linear interpolation. The matrix-free implementation consists of approximating Ch 2h M h P2h h by M 2h, the discretization of the damped wave equation on the grid with a 2h spacing. Therefore the system on the coarse grid is r 2h = M 2h e 2h. (9) This equation can be solved with the same approach that the one for the system 5, namely by applying recursively the two-grid approach. The interpolation of e 2h can add suspicious oscillations in e h. To remove those oscillations, the smoother can be applied on the approximate solution w h + e h. A multigrid cycle defines the way to go from the fine grid to the coarsest grid and to return to the fine grid. This gives an approximation of u h. To completely solve the system 5 the cycle can be applied iteratively. In the implementation, the system u = Mv is approximated with one cycle. There are many possibilities to go back and forth from the fine grid to the coarsest grid. The full multigrid V-cycle is chosen 2. Notice that since the implementation is matrix-free, it is easy to implement high order schemes. 3 Three-dimensional numerical results 3.1 A salt dome example To illustrate the method, the wave propagation through a salt dome model of 8920 m by 4440 m by 5100 m is computed. The velocity model is displayed in Figure 1. The real part of the pressure wave field at 10.625 Hz and 21.25 Hz is shown in Figures 2 and 3. The source is an explosive source located in the middle top of the model, at the point (4000 m,2000 m,10 m). The grid spacing is 40 m at 10.625 Hz and 20 m at 21.25 Hz. The minimum wavelength being 160 m at 10.625 hz and 80 m at 21.25 Hz, this corresponds to a discretization with 4 points per minimum wavelength, which is normally sufficient with a 8 th order scheme. On each side, an absorbing layer with 35 points is added. Therefore the model has 294 182 199 points at 10.625 Hz and 517 293 326 points at 21.25 Hz. For the preconditioner the wavenumber has been multiplied by β r + iβ i = 1 + 0.5i. It is interesting to notice the effects of the salt dome on the pressure wave field, the reflection on the hard layer around 3000 m depth and the turning waves in the shallow part (which corresponds to the sediment part). 5

Figure 1: Salt velocity model. The velocity varies from 1700 m/s (blue) to 4900 m/s (brown). Figure 2: Real part of the pressure field at 10.625 Hz. (The source is located at (4000 m,2000 m,10 m).) 3.2 Convergence results The convergence of the BI-CGSTAB method is shown in Figure 4. The stopping criterion is the normalized residual norm, f Au. We can notice that the number of iterations f is double from the 10.625 Hz example to the 21.25 Hz example. The number of points 6

Figure 3: Real part of the pressure field at 21.25 Hz. (The source is located at (4000 m,2000 m,10 m).) 10 1 10 0 10 1 Normalised error 10 2 10 3 10 4 10 5 10 6 0 50 100 150 200 250 300 350 Number of bicgstab iterations Figure 4: Convergence of the BI-CGSTAB for the 10.625 Hz example, blue line, and the 21.25 Hz example, red line. per wavelength has been kept constant. This result seems to indicate that the number of iterations is proportional to the average number of discretization points in one direction. This average number, called n, is simply obtained by taking the cubic root of the total number of discretization points. To confirm this result, I have carried out more computations on the salt dome example, Figure 1, and I have computed the frequency pressure responses on two other geophysical 7

350 300 Number of bicgstab iterations 250 200 150 100 50 100 150 200 250 300 350 400 450 Average number of points in one direction Figure 5: Convergence results on three three-dimensional examples. The blue lines correspond to the model Figure 1, the black lines correspond to the SEG/EAGE salt model, and the red lines correspond to the Overthrust model. The thick lines correspond to the number of iterations versus the average number of discretization in one direction when the stopping criterion is 10 4 and the dashed lines when the stopping criterion is 10 5. models, the SEG/EAGE salt model and the Overthrust model 1 (not shown here). During all these computations the number of points per minimum wavelength is kept constant. For the SEG/EAGE salt model and the Overthrust model, a discretization of 5 points per wavelength was used. In Figure 5 the number of BI-CGSTAB iterations versus n, the average number of discretization in one direction, is plotted for the three different models and for two stopping criteria values, 10 4 and 10 5. In practice, it appears that 10 4 is sufficient. Figure 5 numerically shows that the number of iterations roughly increases with the frequency because n is proportional to the frequency when the number of points per wavelength is kept constant. Since the multigrid solver is in O(n 3 ), one BI-CGSTAB iteration requires O(n 3 ) iterations. This result then shows that the new iterative solver for the Helmholtz equation is in O(n 4 ). Notice that the solver has converged on all the examples, which indicates that it is robust. For the SEG/EAGE model and the Overthrust model, the largest model that fits on 16 GB of RAM corresponds to the 12 Hz case and a model of about 80 million unknowns. 4 The application The numerical results seem to indicate that the new iterative solver is more efficient and robust than the ones based on the ILU decomposition or separation of variables 4,8. However is it good enough for geophysical applications? The imaging problem consists of retrieving velocity field from seismic data. Multi-shot and multi-receiver data set are used. The classic approach is to build an image from the 8

cross-correlation of the incident field with the back-propagated field, obtained by backpropagating the shot data into the Earth. Mathematically, this means to compute the gradient of the least-squares misfit function between the computed synthetics and the data. In practice, the problem is split into two parts. The first problem is the mapping in depth of the events recorded in time, this is the so-called migration. The smooth part of the velocity, corresponding to the low spatial frequencies, is assumed known. The second problem, called velocity model building problem, is the determination of the smooth part of the velocity. The migration problem requires to process a sufficiently large frequency band to obtain an image with a good resolution. With offset data, it was nevertheless shown that the Nyquist criterion could be relaxed, meaning that the number of frequencies could be reduced 6. Generally the migration problem does not require an iterative approach because the least-squares functional is almost quadratic in the perturbations of the model around the known smooth part. The velocity model building problem does require an iterative approach. However, the minimization of the least-squares misfit functional with respect to the smooth part of the velocity may be done per frequency, or per small overlapping frequency band 7,10. The scale separation is more difficult to carry out when working in the time domain 3. It is then useful to approximately count the number of floating point operations to determine in which domain the migration and the model building problem should be done. 4.1 Rough estimation of the number of operations of the time-domain approach It is assumed that the time-domain wave equation is solved with an explicit scheme, with a 8 th order scheme in space and a 2 nd order scheme in time. The evaluation of a time step then requires about 70n 3 operations (a multiplication and a addition are counted for one operation and a division for 5.). The total number of operations is roughly 70 n t n 3. For stability reason, the time sampling, t, and the spacing, x, satisfy x t = α 1 v max, with v max the maximum velocity. α 1 is generally smaller than 1 depending on the particular implementation, here we assume α 1 = 0.2. By definition the length of the model is L = n x, and the recording time is T = n t t. In practice, the recording L time is also linked to the length of the model, for instance T = α 2 v min, with v min the minimum velocity and α 2 around 2. This gives n t = α 2 v max α 1 v min n. With vmax v min 3, n t = 30n. This crude calculation gives the total number of operations for the time-domain solver, N op t : N op t = αn 4, with α 2100. (10) 9

4.2 Rough estimation of the number of operations of the frequency-domain approach Whereas in the time-domain approach the number of grid points is fixed for a given frequency band (and related to the highest frequency modeled during the computation), the number of grid points in the frequency domain can be adapted to the current frequency to save computation time. Let us call n the (average) number of discretization points in one direction associated with a given frequency. The cost of the frequency-domain approach depends on the cost of one BI-CGSTAB iteration and the number of iterations. One BI-CGSTAB iteration mainly consists of 12 vector-vector operations and 2 matrix-vector operations. Since the calculus is carried out in the complex space, the 12 vector-vector operations represent about 85 n 3 (scalar) operations. A matrix-vector operation is constituted by one multigrid cycle, to solve u = Mv, and one wave-operator evaluation Au. The cost of a multigrid cycle is mainly due to the computation done on the fine grid, since the cost is divided by 8 at each coarsening. More precisely, the cost of the multigrid cycle is the cost on the fine grid multiplied by j 1 = 8 1.15. On the fine grid and with the current implementation, 8 j 7 the main operations are two Gauss-Seidel evaluations (corresponding to one post-smoother with a symmetric Gauss-Seidel), one wave-operator evaluation (related to the computation of the residual), 2 restrictions (one for the pressure field and one for the velocity), one prolongation (for the pressure field) and one vector-vector operation. A Gauss-Seidel evaluation is more or less similar to the wave-operator evaluation. With a 8 th order scheme, this means about 140n 3. A restriction or a prolongation take about 50n 3 operations. After adding all these operations the number of floating point operations of one BI-CGSTAB iteration is roughly γ 1 n 3, with γ 1 1500. From the numerical results and the convergence plot, Figure 5, we have n it = γ 2 n, with γ 2 0.5 and n it the number of BI-CGSTAB iterations. This crude calculation gives the total number of operations for the iterative solver, N op f : 4.3 The imaging problem N op f = γn 4, with γ 800. (11) From a computational point of view the synthetics depends on n. The number of discretization points per direction depends on the frequency in the frequency domain and on the maximum frequency in the time domain. We can then write with c(n(f), v), the synthetics: B(n(f), v)c(n(f), v) = s, (12) where B is the wave operator and f the current frequency in the frequency domain or the maximum frequency in the time domain. The least-squares misfit function is simply, with d the data: J(v) = 1 2 c(n(f), v) d 2, (13) 10

and the gradient with J(v) v =< λ(n(f), v), B(n(f), v) c(n(f), v) > (14) v B(n(f), v) λ(n(f), v) = (c d). (15) <, > is the scalar product; in the time domain, this would corresponds to the crosscorrelation. B(n(f), v) is the reverse time operator in the time domain and the adjoint operator in the frequency domain. The system 15 is solved as the system 12. Therefore if the memory or I/O requirements are not taken into account, the computational cost of the gradient is roughly equal to twice the computational cost of the wave operator. 4.3.1 The migration problem As mentioned before the migration problem requires to process a large enough frequency band. Let us call f max the maximum frequency. In the time domain, the least-squares misfit functional is 1 2 c(n(f max), v) d 2. (16) Therefore the number of operations for the time-domain migration is roughly N op t,mig 2N op t = 2αn 4 (f max ). (17). In the frequency-domain, the least-squares misfit functional is n f j=1 1 2 c(n(f j), v) d 2, (18) with n f the number of frequencies and f j [0, f max ]. When the number of grid points per wavelength is kept constant, we have n(f) = f j f max n(f max ). For the migration, the full frequency band needs to be used to obtain a good resolution. f j = j f, with f the frequency spacing and n f = fmax f. Using equation 11 and the fact that n f j=1 j 4 j5, the 5 number of operations for the frequency approach is γ n 5 fn 4 (f max ). The Nyquist theorem tells us that f = 1, with T the recording time. As mentioned previously T = α 2n(f max) x T v min. A discretization of 5 points per minimum wavelength gives 5 x = v min. Therefore n f f n f = α 2 5 n(f max ). The total number of operations for the migration problem with the Helmholtz solver is then roughly N op f,mig = 2βn(f max) 5 with β = γα 2 25 11 64. (19)

This rapid estimation of the number of operations required for the migration problem shows that the number of operations for the frequency-domain approach divided by the number of operations for the time-domain approach is about 0.04n. In practice, it is possible to reduce the number of frequencies (by a factor smaller than 2) in the frequencydomain approach if the data contains large offsets 6. This estimation indicates that the time-domain approach is roughly one order of magnitude faster for medium size problems and will be even more advantageous for larger problems. 4.4 The velocity model building problem The velocity model building problem requires the minimization of the least-squares misfit functional. Unfortunately the least-squares functional has many local minima. Since only a gradient type of optimization is generally affordable, the initial guess should lie in the domain of attraction of the global minimum. The size of the domain of attraction is roughly proportional to the inverse of the maximum frequency present in the data 3. To take advantage of this behaviour, an idea is to first work with the low frequencies of the data and then to progressively increase the frequency content. In the frequency domain this means that the minimization is carried out per frequency (or per narrow frequency band) 7,10. In the time domain, this means that the minimization is carried out per lowpass filtered data. Given a frequency set (f j ), the velocity model building problem then consists of a sequence of minimization problems: with ṽ(f j 1 ) as starting model. ṽ(f j ) = argmin v 1 2 c(n(f j), v) d 2 (20) The number of operations to compute c(n(f j ), v) is given by equation 10 for the time domain and equation 11 for the frequency domain. The two approaches have the same complexity. However, the number of operations of the frequency-domain approach is about a thrid of the one of the time-domain approach. 5 CONCLUSIONS A three-dimensional iterative solver for the wave equation with a preconditioned BI- CGSTAB method has been studied in a geophysical context. The preconditioner consists of a multigrid cycle applied to a heavily damped wave equation. This leads to a robust method even with a relatively high seismic frequency. A numerical study based on three three-dimensional examples shows that the number of iterations of the BI-CGSTAB methods roughly linearly increases with frequency. This confirms the results obtained on two-dimensional examples. Based on the numerical results, the relevance for the imaging problem has been studied. It was found that the migration problem is more efficiently solved in time domain than in frequency domain, when the frequency-domain approach is based on the new proposed 12

Helmholtz solver and the time-domain approach on an explicity scheme. Indeed, after a rough count of the number of floating point operations required for the computation of the wave-equation solution, it was notice that the time-domain approach will be at least one order of magnitude faster for realistic size problems. The situation is different for the velocity model building based on a least-squares formulation and a scale separation. For this problem, the two approaches have the same complexity, however the frequencydomain approach requires about a thrid of the number of floating point operations of the time-domain approach. The imaging problem requires the processing of multi-shot data. It may then still be possible to speed-up the frequency-domain approach since the wave equation needs to be solved for multiple right hand sides. This remains an open question. In this study, the memory and I/O requirements have not been taken into account. In a practical implementation of the imaging algorithm, those aspects are also crucial. REFERENCES [1] F. Aminzadeh, J. Brac, and T. Kunz. 3-D Salt and Overthrust Models. SEG/EAGE 3-D Modeling Series No.1, SEG, (1997). [2] W.L. Briggs, Van Emden Henson and S.F. McCormick. A multigrid tutorial - Second Edition, SIAM, (2000). [3] C. Bunks, F.M. Saleck, S. Zaleski and G. Chavent. Multiscale seismic waveform inversion. Geophysics, 60, 1457-1473, (1995). [4] Y.A. Erlangga, C.W. Oosterlee and C. Vuik. A novel multigrid based preconditioner for heterogeneous Helmholtz equation. SIAM J. Sci. Comp., Vol. 27, 1471 1492, (2006). [5] K. Marfurt. Accuracy of finite-difference and finite-element modelling of the scalar and elastic wave equation. Geophysics, 49, 533 549, (1984). [6] W.A. Mulder and R.-E. Plessix. How to choose a subset of frequencies in frequencydomain finite-difference migration. Geophysical Journal International, 158, 801-812 (2004). [7] S. Operto, J., Virieux, J.X., Dessa and G. Pascal. High-resolution crustal imaging from multifold ocean bottom seismometers data by frequency-domain full-waveform tomography: application to the eastern Nankai trough. submitted to Journal of Geophysical Research, (2005). [8] R.-E. Plessix and W.A. Mulder. Separation-of-variables as a preconditioner for an iterative Helmholtz solver. Applied Numerical Mathematics, 44, 385-400, (2003). 13

[9] R.-E. Plessix and W.A. Mulder. Frequency-domain finite-difference amplitudepreserving migration. Geophysical Journal International, 157, 957-987, (2004). [10] R.G. Pratt, Seismic waveform inversion in frequency domain. Part I: theory and verification in a physical scale domain. Geophysics, 64, 888-901, (1999). [11] C.D. Riyanti, Y.A. Erlangga, R.-E. Plessix, W.A. Mulder, C. Vuik and C.W. Oosterlee. A new iterative solver for the time-harmonic wave equation. Accepted in Geophysics, (2006). 14