Compressive simultaneous full-waveform simulation

Transcription

1 Compressive simultaneous full-waveform simulation Felix J. Herrmann 1, Yogi Erlangga 1 and Tim T. Y. Lin 1 (August 14, 2008) Running head: compressive simulations ABSTRACT The fact that the numerical complexity of wavefield simulation is proportional to the size of the discretized model and acquisition geometry, and not to the complexity of the simulated wavefield, is the main impediment within seismic imaging. By turning simulation into a compressive sensing problem where simulated data is recovered from a relatively small number of independent simultaneous sources we remove this impediment by showing that compressively sampling a simulation is equivalent to compressively sampling the sources, followed by solving a reduced system. As in compressive sensing, this allows for a reduction in sampling rate and hence in simulation costs. We demonstrate this principle for the time-harmonic Helmholtz solver. The solution is computed by inverting the reduced system, followed by a recovery of the full wavefield with a sparsity promoting program. Depending on the wavefield s sparsity, this approach can lead to significant cost reductions, in particular when combined with the implicit preconditioned Helmholtz solver, which is known to converge even for decreasing mesh sizes and increasing angular frequencies. These properties make our scheme a viable alternative to explicit time-domain finite-differences. 1 Seismic Laboratory for Imaging and Modeling, Department of Earth and Ocean Sciences, University of British Columbia, 6339 Stores Road, Vancouver, V6T 1Z4, BC, Canada 1

2 INTRODUCTION Currently, simulations costs are largely dictated by the mesh size of the computational domain, the redundancy in the source-receiver acquisition grid, and the central frequency of the source function. As opposed to solutions based on certain approximations, e.g. asymptotic (high-frequency) solutions that typically involve diagonalization of solution operators (see e.g. ten Kroode et al., 1998), efficient computation of high-fidelity large-scale full-waveform simulations remains an elusive area of research. To overcome this impediment, we argue that future improvements in wave-equation imaging and inversion will depend on a problem formulation with a numerical complexity that is no longer strictly determined by the size of the discretization but by the transform-domain compressibility of its solution. In this new paradigm, we bring computational costs in par with our ability to compress seismic data and images. The above premise is related to recent developments in theoretical signal processing known as compressive sensing (CS, Candès et al., 2006; Donoho, 2006) where the argument is made, and rigorously proven, that compressible signals can be recovered from severely sub- Nyquist sampling by solving a sparsity promoting program. In this approach, sub-sampling interferences are removed by exploiting transform-domain sparsity, properties of certain subsampling schemes, and the existence of sparsity promoting solvers. Following earlier work by Lin and Herrmann (2007), we adapt the ideas from CS towards the problem of seismic waveform simulation. Our main contribution here lies in the design of a simultaneous acquisition methodology for numerical waveform simulation (albeit our ideas are extendable to field acquisition, see e.g. Krohn and Neelamani, 2008). Aside from using the timeharmonic Helmholtz equation, which is particularly conducive to selecting (= CS sampling) 2

3 subsets of angular frequencies (Lin et al., 2008), we show that simultaneous acquisition is, in fact, also equivalent to an instance of CS. If angular frequencies are not subsampled, this approach is similar to the work of (Neelamani et al., 2008) designed for time-domain finite differences. By designing the acquisition according to the CS principles, the inteference phenomenon experienced in this type of acquisition can be mitigated. CS is important because it provides a link between subsampling strategies and the quality of recovery, which is largely lacking in simultaneous acquisition (Beasley, 2008) and in cost-reduction efforts such as the selection of subsets of frequencies during waveform simulation, imaging and inversion (Sirgue and Pratt, 2004; Mulder and Plessix, 2004). MOTIVATION Even though numerical solutions to wave equations vary in complexity, they generally compress in transformed domains such as wavelets, as observed by Donoho et al. (1999). In principle, this finding allows for recovery through sparsity promotion, following the sampling strategies prescribed by CS (see e.g. Herrmann and Hennenfent, 2008; Hennenfent and Herrmann, 2008, where this approach was followed to regularize data with missing traces through curvelet-domain sparsity promotion). To illustrate how these guidelines translate to subsampling solutions of the wave equation, we first briefly outline the principles of compressive sampling and sparse recovery, followed by a brief discussion on the Helmholtz wave-equation solver. Because simultaneous acquisition involves CS along the source coordinate, we can reduce the system size in accordance to the complexity of the simulation. Compressive sampling: Recovering a sparse ( spiky ) signal from sub samplings is the modern equivalent of finding needles in a haystack. For instance, consider the problem of 3

4 locating k arbitrary non-zero entries from a sparse spike train of length N k. According to CS theory, m incoherent measurements with m k ( means proportional to within log N factors) suffice to recover these non-zero entries. Two basis are incoherent, when the inner products of the columns are small. For instance, Dirac and Fourier are incoherent and so are any orthonormal basis and a orthonormalized Gaussian random matrix. CS theory proves that recovery through sparsity promotion is possible from a sample size m that is proportional to the signal s sparsity (here, the number of non-zeros, k) as opposed to the signal length N. We use this important finding in our formulation of the compressive simulation problem. Feasible Helmholtz solver for full waveform simulation: Since their inception, iterative implicit matrix-free solutions to the time-harmonic Helmholtz equation have been plagued by lack of numerical convergence for decreasing mesh sizes and increasing angular frequencies (Riyanti et al., 2006). By including deflation, a way to handle small eigenvalues that lead to slow convergence, Erlangga and Nabben (2007); Erlangga and Herrmann (2008) successfully removed this impediment, bringing 2- and 3-D solvers for the time-harmonic Helmholtz into reach. For a given shot (right-hand side b), the solution u (we use san-serif symbols for quantities related to the Helmholtz system) is computed with a Krylov method that involves the following system of equations AM 1 Qû = b, u = M 1 Qû, (1) where A, M, and Q represent the discretized monochromatic Helmholtz equation, the preconditioner, and projection matrices, respectively. As shown by Erlangga et al. (2004, 2006), convergence is guaranteed by defining the preconditioning matrix M in terms of the discretized shifted or damped Helmholtz operator M := ω2 c(x) 2 (1 βi), i = 1, 4

5 with β > 0. With this preconditioning, the eigenvalues of AM 1 are clustered into a circle in the complex plane. By the action of the projector matrix Q, these eigenvalues move towards unity on the real axis as shown in Figure 1. These two operations lead to an improved condition number, which explains the superior performance of this time-harmonic solver. Because the Helmholtz equation is time harmonic, it is particularly conducive to CS since each angular frequency (as reported by Lin et al., 2008) and simultaneous shot can be treated independently. In this respect our approach differs from the one by Neelamani et al. (2008), where a single simultaneous shot is run with finite differences for long times. Source-receiver CS-sampling equivalence: Aside from the required number of angular frequencies, the computational complexity of wavefield simulations is determined by the number of shots i.e., the number of right-hand sides. In the current simulation paradigm, these shots determine the number of single-source simulations for which the Helmholtz system has to be solved. As prescribed by CS, these costs can be reduced by designing a survey that consists of a relatively small number of simultaneous shots with sources that contain subsets of angular frequencies. For this to work, we need to show that this type of simulation is equivalent to compressively sampling the full solution, computed for all desired angular frequencies and source positions. Mathematically, this corresponds to showing that y = y for the solutions of the full (left) and compressed (right, underlined) systems i.e., B = D s }{{} single shots B = D RMs }{{} simul. shots := D s LU = B LU = B (2) y = RMDU := RMd y = DU. 5

6 Here, L = diag(a[ω i ]), is the block-diagonal discretized Helmholtz equation for each angular frequency ω i := 2πi f, i = 1 n f, with n f the number of frequencies and f its sample interval. The vector s contains the multi-shot source wavefield that is injected into the right-hand side by the adjoint (denoted by ) of the detection matrix D. This detection matrix extracts data at the receiver positions, and consists of an inverse Fourier transform and a permutation matrix that reorganizes the data into a shot-receiver time data volume. Its adjoint inserts data at the source positions. Each column of U contains the wavefields for all frequencies induced by the shots located in each column of B. The full simulation involves the inversion of the block-diagonal system (for all shots), followed by a detection i.e., we have d = DL 1 B for a data volume with n d = n f n s n r samples and where n s,r are the number of shot/receiver positions. After CS sampling, this volume is reduced to y = RMd, where the CS data vector y C n d is given by applying the rectangular CS matrix RM (defined in the next section) to the full simulation. This yields a reduced data vector with n d = n f n sn r n d samples where n f n f and n s n s. Applying this CS matrix directly to the source distribution s leads to the compressed system L, which after inversion gives y. This latter inversion is easier to compute because it involves only a subset of angular frequencies and simultaneous shots i.e., {U, B} contain only n s columns with n f frequency components each. In this case, D is a detection-permutation matrix, which now lacks the inverse Fourier transform. To demonstrate that both sampling strategies are equivalent, it is sufficient to show that R Ω full system {}}{ L 1 B R Σ = reduced system ( {}}{ Rω LR ) 1B, Σ where the matrices R Ω and R Σ represent subsampling along the rows and columns, respectively. The above relation holds because subsampling the frequencies and shots after full 6

7 simulation is equivalent to compressed simulation i.e., we have R Ω L 1 = L 1 R Ω, leading to R Ω L 1 BR Σ = L 1 R Ω BR Σ with {U, B} = R Ω{U, B}R Σ, which in turn implies y = y. CS PROBLEM FORMULATION With the above identification in place, we can now remove the interferences, which means recovering the full data from the compressive simulation with the reduced system. Following Lustig et al. (2008); Hennenfent and Herrmann (2008); Lin et al. (2008), successful recovery depends on three key components that need to be implemented at minimal costs to avoid offsetting our computational gain through system reduction. For technicalities of CS sampling criteria, we refer to the rapidly expanding literature on this topic. The sparsfying transform: Multi-dimensional seismic simulations contain bandwithlimited transients that compress well in the discrete wavelet domain (Donoho et al., 1999). The fact that wavelets are O(n), instead of O(n log n) for most other transforms, and are reasonably incoherent with the Fourier basis further supports this choice. Using the Kronecker product, we define the sparsity-promoting transform as S := W W W, with W the 1-D discrete wavelet transform. The compressive-sampling matrix: The success of our compressive simulation scheme hinges on devising a subsampling scheme for the physically distinct source and angular frequency axes that turns coherent interferences into harmless random noise in the physical domain (Hennenfent and Herrmann, 2008). Along the frequency coordinate, we accomplish this by selecting a random subset of frequencies, whereas random projections by a rectangular Gaussian matrix along the source direction enforces independence amongst the 7

8 simultaneous sources. To maximize independence amongst the sources, we apply a different restriction for each n s simultaneous shot i.e., we have R Σ 1 RΩ 1 RM :=. ( N F ), (3) R Σ n s RΩ n s where F is the Fourier transform in the time direction, N the projection with respect to an orthonormal Gaussian matrix, and where R Σ,Ω are the restriction matrices along the source and frequency axes. Application of the matrix RM to the original single-shot source wavefield s gives a simultaneous source wavefield s that is random and consists of n s simultaneous shots with n f angular frequencies each (see Figure 2(a)). Because seismic data is bandwidth limited, we select random frequencies in this band with high probability. Recovery by sparsity promotion: We reconstruct the seismic wavefield by solving the following nonlinear optimization problem x = arg min x 1 subject to Ax = y, (4) x with d = S x the reconstruction, A := RMS the CS matrix, and y (= y) the compressively simulated data (cf. Equation 2-right). Equation 4 is solved by SPGl 1 (van den Berg and Friedlander, 2008), a gradient-projection algorithm with root finding. COMPLEXITY ANALYSIS According to Riyanti et al. (2006), the cost of the iterative Helmholtz solver equals n f n s n it O(n d ), typically with n it = O(n) the number of iterations. For d = 2 and assuming n s = n f = O(n), this cost becomes O(n 5 ). Under the same assumption, the cost of a time-domain 8

9 solver is O(n 4 ). The iterative Helmholtz solver can only become competitive if n it = O(1), yielding an O(n 4 ) complexity. Erlangga and Nabben (2007); Erlangga and Herrmann (2008) achieve this by the method explained earlier. Despite this improvement, this figure is still overly pessimistic for simulations that permit a sparse representation. As long as the simulation cost exceeds the l 1 -recovery cost (cf. Equation 4), CS will improve on this result. This cost reduction depends on the evaluation of A that is dominated by explicit matrix-matrix multiplies as part of the Gaussian projections. These costs are O(n 3 ) for each frequency, which offers no gain aside from possible improvements on constants. However, if we replace the Gaussian projections by fast O(n log n) projections, e.g. with randomly-restricted noiselets (Coifman et al., 2001) or with random convolutions (Romberg, 2008), we are able to reduce this cost to O(n 2 log n). Now, the leading order cost of the l 1 recovery is reduced to O(n 3 log n), which is significantly less than the cost of solving the full Helmholtz system, especially for large problems (n ) and for extensions to d = 3. COMPRESSIVE SIMULATION EXPERIMENT To illustrate CS-recovery quality, we conduct a series of experiments for two velocity models, namely the complex model used in Herrmann et al. (2007), and a simple single-layer model. These models generate seismic lines that vary in complexity. During these experiments, we vary the subsampling ratio and the frequency-to-shot subsampling ratio. All simulations are carried out with a fully parallel Helmholtz solver with β = 0.5, for a spread with 256 collocated shots and receivers sampled at a 15 m interval. The time sample interval is s and the source function is a Ricker wavelet with a central frequency of 10 Hz. An example of the CS sampled source wavefield is included in Figure 2(a), which contains random noise restricted to subsets of angular frequencies. The corresponding compressively 9

10 Subsample ratio 0.25 n d 0.15 n d 0.07 n d Speedup n f = 2n s n f = 1n s n f = 0.5n s Table 1: Signal-to-noise ratios, SNR = 20 log 10 ( d e d 2 d 2 ) for reconstructions with different subsample and frequency-to-shot ratios. Notice the speedup with decreasing subsampling ratio. simulated seismic line is plotted in Figure 2(b). This line clearly contains acquisitionrelated interferences and is equivalent to the compressive sampling of the full seismic line (not shown). By solving Equation 4, we recover the full simulation for the two datasets. Comparisons between the full and compressive simulations are included in Figure 3 and show remarkable high-fidelity results for the latter (complex model: 16.1 db, simple model: 19.2 db) for a subsampling ratio of 1:4. As expected, the SNR for the simple model is better because of the reduced complexity, whereas the numbers in Table 1 confirm increasing recovery errors for increasing subsampling ratios. Moreover, the bandwidth limitation of seismic data explains improved recovery with decreasing frequency-to-shot ratio for a fixed subsampling ratio. Because the speedup of the solution is roughly proportional to the subsampling ratio, we can conclude that speedups of four to six times are possible at the expense of a minor drop in SNR. 10

11 DISCUSSION Relation to existing work: There exists a parallel between compressive simultaneous simulations (cf. Equation 2-right) and recent developments in simultaneous acquisition, whether conducted in the computer (Morton and Ober, 1998; Romero et al., 2000), or in the field (Beasley, 2008). In both cases, researchers and acquisition specialists conduct simultaneous source experiments. To limit the interference, or in CS language to create favorable recovery conditions, the sources are made as independent as possible. In simultaneous acquisition, this is done with delays or phase encoding whereas in CS much effort is spent towards the design and analysis of sampling criteria and CS matrices. With the Kronecker product, we have shown that it is possible to design CS matrices that not only favor recovery but which are also well suited to compressively sample two physically distinct coordinate axes. In contrast with CS, however, simultaneous acquisition misses rigorous theoretical justification, and this paper is to our knowledge one of the first attempts to identify and exploit this type of acquisition as CS. Similar approaches have been reported for MRI, Radar imaging, and Fourier optics (see e.g. Lustig et al., 2008; Romberg, 2008). While completing this paper, we came across an SEG expanded abstract (Neelamani et al., 2008) in which a closely related methodology is presented for time-domain finite differences. However, our approach is more readily parallelized and addresses the issue of computational complexity. Finally, there is also an interesting connection with interferometric deconvolution where noise sources also play a key role (Vasconcelos and Snieder, 2008). Extensions: The implications of identifying simultaneous acquisition and CS go far beyond the simulation of seismic wavefields. As applications of CS to real physical systems become more prevalent, we can also expect major developments in seismic imaging. For 11

12 instance, our approach may be applicable for land acquisition as long as we can design simultaneous source experiments that are physically realizable (Krohn and Neelamani, 2008; Romberg, 2008). Because CS sampling is linear (Bobin et al., 2008), our simulation method is incremental i.e., adding more simultaneous shots improves the recovery. Moreover, this linearity allows us to conduct seismic data processing, imaging and inversion directly on compressively-sampled data, an observation made independently by Berkhout (2008). CONCLUSIONS We have identified simultaneous acquisition as CS, which allowed us to derive a rigorous and cost-effective simulation scheme based on its principles. According to these principles, wavefields can be reconstructed from subsamplings commensurate with their complexity. We arrive at this result by source-receiver CS-sampling equivalence, which states that CS sampling seismic simulations is the same as CS sampling the source wavefield, followed by simulation with a reduced system. CS predicts improved recovery for compressible signals for increasing number of samples. We verified this behavior experimentally and this, in conjunction with the intrinsic linearity of the CS sampling, opens a number of enticing new perspectives. First, CS applied to simulations (and possibly during acquisition) decouples simulation- and acquisition-related costs from the model size. Instead, these costs depend on sparsity. Second, CS predicts improved recovery for increasing sample sizes, which opens the possibility to improve recovery by adding samples. Third, because of this linearity, we envision a seamless incorporation of this paradigm into seismic exploration. 12

13 ACKNOWLEDGMENTS The authors would like to thank Scott Morton for bringing this topic to our attention. We would like to thank the authors of SPARCO ( SPGl 1 ( and the Rice Wavelet Toolbox. This paper was prepared with Madagascar (rsf.sourceforge.net/) and was in part financially supported by the NSERC Discovery Grant (22R81254) of F.J.H. and CRD Grant DNOISE ( ). 13

14 REFERENCES Beasley, C. J., 2008, A new look at marine simultaneous sources: The Leading Edge, 27, Berkhout, A. J., 2008, Changing the mindset in seismic data acquisition: The Leading Edge, 27, Bobin, J., J.-L. Starck, and R. Ottensamer, 2008, Compressed sensing in astronomy. Candès, E., J. Romberg, and T. Tao, 2006, Stable signal recovery from incomplete and inaccurate measurements: Comm. Pure Appl. Math., 59, Coifman, R., F. Geshwind, and Y. Meyer, 2001, Noiselets: Appl. Comput. Harmon. Anal., 10, Donoho, D. L., 2006, Compressed sensing: IEEE Trans. Inform. Theory, 52, Donoho, P., R. Ergas, and R. Polzer, 1999, Development of seismic data compression methods for reliable, low-noise performance: SEG International Exposition and 69th Annual Meeting, Erlangga, Y. and F. J. Herrmann, 2008, An iterative multilevel method for computing wavefields in frequency-domain seismic inversion: Presented at the SEG Technical Program Expanded Abstracts, SEG. (To appear.). Erlangga, Y. A. and R. Nabben, 2007, On multilevel projection Krylov method for the preconditioned Helmholtz system: Elec. Trans. Numer. Anal., submitted. Erlangga, Y. A., C. Vuik, and C. W. Oosterlee, 2004, On a class of preconditioners for solving the Helmholtz equation: Appl. Numer. Math., 50, , 2006, Comparison of multigrid and incomplete LU shifted-laplace preconditioners for the inhomogeneous Helmholtz equation: Appl. Numer. Math., 56, Hennenfent, G. and F. J. Herrmann, 2008, Simply denoise: wavefield reconstruction via 14

15 jittered undersampling: Geophysics, 73, no. 3. Herrmann, F. J., U. Boeniger, and D. J. Verschuur, 2007, Non-linear primary-multiple separation with directional curvelet frames: Geophysical Journal International, 170, no. 2, Herrmann, F. J. and G. Hennenfent, 2008, Non-parametric seismic data recovery with curvelet frames: Geophysical Journal International, 173, Krohn, C. and R. Neelamani, 2008, Simultaneous sourcing without compromise: Presented at the Rome 2008, 70th EAGE Conference & Exhibition incorporating SPE EUROPEC. Lin, T. T. Y. and F. J. Herrmann, 2007, Compressed wavefield extrapolation: Geophysics, 72, SM77 SM93. Lin, T. T. Y., E. Lebed, Y. Erlangga, and F. J. Herrmann, 2008, Interpolating solutions of the Helmholtz equation with compressed sensing: Presented at the SEG Technical Program Expanded Abstracts, SEG. (To appear.). Lustig, M., D. Donoho, J. Santos, and J. Pauly, 2008, Compressed sensing MRI [A look at how CS can improve on current imaging techniques]: Signal Processing Magazine, IEEE, 25, Morton, S. A. and C. C. Ober, 1998, Faster shot-record depth migrations using phase encoding: SEG Technical Program Expanded Abstracts, , SEG. Mulder, W. and R. Plessix, 2004, How to choose a subset of frequencies in frequency-domain finite-difference migration: Geophysical Journal International, 158, Neelamani, N., C. Krohn, J. Krebs, M. Deffenbaugh, and J. Romberg, 2008, Efficient seismic forward modeling using simultaneous random sources and sparsity: Presented at the SEG International Exposition and 78th Annual Meeting. (To appear.). Riyanti, C. D., Y. A. Erlangga, R.-E. Plessix, W. A. Mulder, C. Vuik, and C. Oosterlee, 15

16 2006, A new iterative solver for the time-harmonic wave equation: Geophysics, 71, E57 E63. Romberg, J., 2008, Compressive sensing by random convolution. (Preprint available at pdf). Romero, L. A., D. C. Ghiglia, C. C. Ober, and S. A. Morton, 2000, Phase encoding of shot records in prestack migration: Geophysics, 65, no. 2, Sirgue, L. and R. G. Pratt, 2004, Efficient waveform inversion and imaging: A strategy for selecting temporal frequencies: Geophysics, 69, ten Kroode, A., D.-J. Smit, and A. Verdel, 1998, A microlocal analysis of migration: Wave Motion, 28, van den Berg, E. and M. P. Friedlander, 2008, Probing the Pareto frontier for basis pursuit solutions: Technical Report TR , UBC Computer Science Department. Vasconcelos, I. and R. Snieder, 2008, Interferometry by deconvolution, part 1 theory for acoustic waves and numerical examples: Geophysics, 73, S115 S

17 LIST OF FIGURES 1 Eigenvalues (λ) of the 1-D Helmholtz operator before and after preconditioning and spectral shaping. (a) Original eigenvalue spectrum of A. (b) Spectrum after preconditioning with M 1. (c) Spectrum after deflation with Q. 2 Compressive sampling with simultaneous sources. (a) Real part of the compressivelysampled source in the frequency domain. (b) The compressively-simulated data in the time domain. 3 Comparison between conventional and compressive simulations for simple and complex velocity models. (a) Seismic line for the simple model. (b) The same for the complex model. (c). Recovered simulation (with a SNR of 19.2 db) for the simple model from 25 % of the samples with the l 1 -solver running to convergence. (d) The same but for the complex model now with a SNR of 16.1 db. 17

18 Im(!) 0!1000!2000!3000!4000!5000! Re(!) x 10 (a) Im(!) 0!0.5!1! Re(!) (b) Im(!) 0!0.5!1! Re(!) (c) Figure 1: Eigenvalues (λ) of the 1-D Helmholtz operator before and after preconditioning and spectral shaping. (a) Original eigenvalue spectrum of A. (b) Spectrum after preconditioning with M 1. (c) Spectrum after deflation with Q. Herrmann et. al. 18

19 (a) (b) Figure 2: Compressive sampling with simultaneous sources. (a) Real part of the compressively-sampled source in the frequency domain. (b) The compressively-simulated data in the time domain. Herrmann et. al. 19

20 (a) (b) (c) (d) Figure 3: Comparison between conventional and compressive simulations for simple and complex velocity models. (a) Seismic line for the simple model. (b) The same for the complex model. (c). Recovered simulation (with a SNR of 19.2 db) for the simple model from 25 % of the samples with the l 1 -solver running to convergence. (d) The same but for the complex model now with a SNR of 16.1 db. Herrmann et. al. 20