Laplace-domain Full Waveform Inversion Research Articles

Laplace-domain waveform inversion using the l-BFGS method

ABSTRACTWe examine the performance of the limited-memory Broyden–Fletcher–Goldfarb–Shanno (l-BFGS) optimization method for full waveform inversions in the Laplace domain. Full waveform inversion methods generally adopt gradient-based optimization schemes for efficiency. Most Laplace-domain full waveform inversion methods published to date use the preconditioned steepest descent (PSD) method with the diagonal elements of the pseudo-Hessian as the preconditioner. These methods use a small fixed step length for stable convergence. The l-BFGS method is a quasi-Newton optimization method that efficiently approximates the inverse of the Hessian. We compare Laplace-domain inversion results from the PSD method with those from the l-BFGS method. Numerical examples using the SEG/EAGE salt model and the Pluto model demonstrate that the l-BFGS method with the Wolfe line search yields smaller cost function values than the PSD method.

Application of full waveform inversion algorithms to seismic data lacking low-frequency information from a simple starting model

Full waveform inversion (FWI) is a method that is used to reconstruct velocity models of the subsurface. However, this approach suffers from the local minimum problem during optimisation procedures. The local minimum problem is caused by several issues (e.g. lack of low-frequency information and an inaccurate starting model), which can create obstacles to the practical application of FWI with real field data. We applied a 4-phase FWI in a sequential manner to obtain the correct velocity model when a dataset lacks low-frequency information and the starting velocity model is inaccurate. The first phase is Laplace-domain FWI, which inverts the large-scale velocity model. The second phase is Laplace-Fourier-domain FWI, which generates a large- to mid-scale velocity model. The third phase is a frequency-domain FWI that uses a logarithmic wavefield; the inverted velocity becomes more accurate during this step. The fourth phase is a conventional frequency-domain FWI, which generates an improved velocity model with correct values. The detailed methods of applying each FWI phase are explained, and the proposed method is validated via numerical tests with a SEG/EAGE salt synthetic dataset and Gulf of Mexico field dataset. The numerical tests show that the 4-phase FWI inverts the velocity correctly despite the lack of low-frequency information and an inaccurate starting velocity model both in synthetic data and field data.

3D Laplace-domain waveform inversion using a low-frequency time-domain modeling algorithm

We have developed a Laplace-domain full-waveform inversion technique based on a time-domain finite-difference modeling algorithm for efficient 3D inversions. Theoretically, the Laplace-domain Green’s function multiplied by a constant can be obtained regardless of the frequency content in the time-domain source wavelet. Therefore, we can use low-frequency sources and large grids for efficient modeling in the time domain. We Laplace-transform time-domain seismograms to the Laplace domain and calculate the residuals in the Laplace domain. Then, we back-propagate the Laplace-domain residuals in the time domain using a predefined time-domain source wavelet with the amplitude of the residuals. The back-propagated wavefields are transformed to the Laplace domain again to update the velocity model. The inversion results are long-wavelength velocity models on large grids similar to those obtained by the original approach based on Laplace-domain modeling. Inversion examples with 2D Gulf of Mexico field data revealed that the method yielded long-wavelength velocity models comparable with the results of the original Laplace-domain inversion methods. A 3D SEG/EAGE salt model example revealed that the 3D Laplace-domain inversion based on time-domain modeling method can be more efficient than the inversion based on Laplace-domain modeling using an iterative linear system solver.

Comparison of two schemes for Laplace-domain 2D scalar wave equation

Abstract Laplace-domain modeling plays an important role in Laplace-domain full waveform inversion. In order to provide efficient numerical schemes for Laplace-domain modeling, two 9-point schemes for Laplace-domain 2D scalar equation are compared in this paper. Compared to the finite-element 9-point scheme, the average-derivative optimal 9-point scheme reduces the number of grid points per pseudo-wavelength from 16 to 4 for equal directional sampling intervals. For unequal directional sampling intervals, the average-derivative optimal 9-point scheme reduces the number of grid points per pseudo-wavelength from 13 to 4. Numerical experiments demonstrate that the average-derivative optimal 9-point scheme is more accurate than the finite-element 9-point scheme for the same sampling intervals. By using smaller sampling intervals, the finite-element 9-point scheme can approach the accuracy of the average-derivative optimal 9-point scheme, but the corresponding computational cost and storage requirement are much higher.

Laplace-domain wave-equation modeling and full waveform inversion in 3D isotropic elastic media

Abstract The 3D elastic problem has not been widely studied because of the computational burden. Over the past few years, 3D elastic full waveform inversion (FWI) techniques in the time and frequency domains have been proposed by some researchers based on developments in computer science. However, these techniques still have the non-uniqueness and high nonlinearity problems. In this paper, we propose a 3D elastic FWI algorithm in the Laplace domain that can mitigate these problems. To efficiently solve the impedance matrix, we adopt a first-order absorbing boundary condition that results in a symmetric system. A conjugate gradient (CG) solver can be used because the Laplace-domain wave equation is naturally positive definite. We apply the Jacobi preconditioner to increase the convergence speed. We identify the permissible range of Laplace damping constants through dispersion analysis and accuracy tests. We perform the Laplace-domain FWI based on a logarithmic objective function, and the inversion examples are designed for a land setting, which means that the source is vertically excited and multi-component data are considered. The inversion results indicate that the inversion that uses only the vertical component performs slightly better than the multi-component inversion. This unexpected result is obtained partly because we use a vertically polarized source. We analyze the residuals and Frechet derivatives for each component to examine the characteristics of the Laplace-domain multi-component FWI. The results indicate that the residuals and Frechet derivatives for the horizontal component have a singularity problem. The numerical examples demonstrate that the singularity problem is related to the directivity of the displacement and to taking the logarithm of Laplace-domain wave fields. To avoid this singularity problem, we use a simple method that excludes the data near the singular region. Although we can use either simultaneous or sequential strategies to invert the Laplace-domain data, we apply a simultaneous inversion strategy in this paper. Nonetheless, the numerical examples demonstrate that our inversion algorithm yields reasonable long-wavelength velocity structures.

3D Laplace-domain full waveform inversion using a single GPU card

The Laplace-domain full waveform inversion is an efficient long-wavelength velocity estimation method for seismic datasets lacking low-frequency components. However, to invert a 3D velocity model, a large cluster of CPU cores have commonly been required to overcome the extremely long computing time caused by a large impedance matrix and a number of source positions. In this study, a workstation with a single GPU card (NVIDIA GTX 580) is successfully used for the 3D Laplace-domain full waveform inversion rather than a large cluster of CPU cores. To exploit a GPU for our inversion algorithm, the routine for the iterative matrix solver is ported to the CUDA programming language for forward and backward modeling parts with minimized modification of the remaining parts, which were originally written in Fortran 90. Using a uniformly structured grid set, nonzero values in the sparse impedance matrix can be arranged according to certain rules, which efficiently parallelize the preconditioned conjugate gradient method for a number of threads contained in the GPU card. We perform a numerical experiment to verify the accuracy of a floating point operation performed by a GPU to calculate the Laplace-domain wavefield. We also measure the efficiencies of the original CPU and modified GPU programs using a cluster of CPU cores and a workstation with a GPU card, respectively. Through the analysis, the parallelized inversion code for a GPU achieves the speedup of 14.7–24.6x compared to a CPU-based serial code depending on the degrees of freedom of the impedance matrix. Finally, the practicality of the proposed algorithm is examined by inverting a 3D long-wavelength velocity model using wide azimuth real datasets in 3.7 days.

Dispersion analysis of an average-derivative optimal scheme for Laplace-domain scalar wave equation

Laplace-domain modeling is an important foundation of Laplace-domain full-waveform inversion. However, dispersion analysis for Laplace-domain numerical schemes has not been completely established. This hampers the construction and optimization of Laplace-domain modeling schemes. By defining a pseudowavelength as a scaled skin depth, I establish a method for Laplace-domain numerical dispersion analysis that is parallel to its frequency-domain counterpart. This method is then applied to an average-derivative nine-point scheme for Laplace-domain scalar wave equation. Within the relative error of 1%, the Laplace-domain average-derivative optimal scheme requires four grid points per smallest pseudowavelength, whereas the classic five-point scheme requires 13 grid points per smallest pseudowavelength for general directional sampling intervals. The average-derivative optimal scheme is more accurate than the classic five-point scheme for the same sampling intervals. By using much smaller sampling intervals, the classic five-point scheme can approach the accuracy of the average-derivative optimal scheme, but the corresponding cost is much higher in terms of storage requirement and computational time.

Subsurface model estimation using Laplace-domain inversion methods

Laplace-domain full-waveform inversion is used to estimate subsurface macro-velocity models from seismic data. This method can be applied to marine or land data and combined with the frequency-domain technique for a higher resolution. Laplace-domain full-waveform inversion is robust with respect to a lack of low-frequency information in the seismic data. This method can be used as an alternative or complement to conventional full-waveform inversion methods; however, several important aspects must be considered when inverting data in the Laplace domain. This article provides an overview of the Laplace-domain full-waveform inversion method and related techniques that have been developed to date.

Laplace-domain full waveform inversion using irregular finite elements for complex foothill environments

Refraction-traveltime tomography is the most common approach and widely used for estimating velocity models with rugged topography and strongly variant near-surface geology. However, for complex geographical structures, there is often a restriction to the application of the conventional approach because the refracted energy can be trapped by the near-surface structure, which leads to limited depth penetration. To solve this problem, we propose a velocity estimation algorithm for foothill areas using Laplace-domain full waveform inversion (FWI) with irregular finite elements. Because the Laplace-domain FWI uses wavefields damped exponentially in time, the acoustic wave equation can be applied to foothill datasets without suppressing various types of elastic noise. In this study, irregular finite elements are generated to depict complicated surface topography using a Delaunay triangulation and tetrahedralization algorithm. Furthermore, adaptive mesh generation that formulates larger size elements with greater depth is used for minimizing the intensive computational costs in solving the full wave equation in the 2D and 3D domains. The validity of our proposed algorithm is demonstrated for 2D and 3D synthetic datasets and a 2D real exploration dataset acquired in the complex Aquio field foothill area in Bolivia.

Efficient Laplace-domain full waveform inversion using a cyclic shot subsampling method

Full waveform inversion is a method used to recover subsurface parameters, and it requires heavy computational resources. We present a cyclic shot subsampling method to make the full waveform inversion efficient while maintaining the quality of the inversion results. The cyclic method subsamples the shots at a regular interval and changes the shot subset at each iteration step. Using this method, we can suppress the aliasing noise present in regular-interval subsampling. We compared the cyclic method with divide-and-conquer, random, and random-in-each-subgroup subsampling methods using the Laplace-domain full waveform inversion. We found examples of a 2D marine field data set from the Gulf of Mexico and a 3D synthetic salt velocity model. In the inversion examples using the subsampling methods, we could reduce the computation time and obtain results comparable to that without a subsampling technique. The cyclic method and two random subsampling methods yielded similar results; however, the cyclic method generated the best results, especially when the number of shot subsamples was small, as expected. We also examined the effect of subsample updating frequency. The updating frequency does not have a significant effect on the results when the number of subsamples is large. In contrast, frequent subsample updating becomes important when the number of subsamples is small. The random-in-each-subgroup scheme showed the best results if we did not update the subsamples frequently, while the cyclic method suffers from aliasing. The results suggested that the cyclic subsampling scheme can be an alternative to the random schemes and the distributed subsampling schemes with a frequently changing subset are better than lumped subsampling schemes.

Laplace-domain full-waveform inversion of seismic data lacking low-frequency information

The lack of the low-frequency information in field data prohibits the time- or frequency-domain waveform inversions from recovering large-scale background velocity models. On the other hand, Laplace-domain waveform inversion is less sensitive to the lack of the low frequencies than conventional inversions. In theory, frequency filtering of the seismic signal in the time domain is equivalent to a constant multiplication of the wavefield in the Laplace domain. Because the constant can be retrieved using the source estimation process, the frequency content of the seismic data does not affect the gradient direction of the Laplace-domain waveform inversion. We obtained inversion results of the frequency-filtered field data acquired in the Gulf of Mexico and two synthetic data sets obtained using a first-derivative Gaussian source wavelet and a single-frequency causal sine function. They demonstrated that Laplace-domain inversion yielded consistent results regardless of the frequency content within the seismic data.

2-D acoustic Laplace-domain waveform inversion of marine field data

SUMMARY The Laplace-domain full waveform inversion method can build a macroscale subsurface velocity model that can be used as an accurate initial model for a conventional full waveform inversion. The acoustic Laplace-domain inversion produced is promising for marine field data examples. Although applying an acoustic inversion method to the field data generally requires several pre-processing steps, pre-processing for the Laplace-domain inversion has not been explained in detail. We provide a detailed explanation of how to apply the Laplace-domain waveform inversion to field data through numerical tests with Gulf of Mexico data sets. The pre-processing includes bandpass filtering, muting of the noise before the first arrival, and extraction of the water depth. We choose the range and the interval between the Laplace damping constants empirically by applying a threshold value to the damped time traces and the Laplace-domain wavefields. The observed data are transformed to the Laplace domain using the selected damping; this method yielded a long-wavelength inversion result. The damping constant and the maximum offset affect the penetration depth of the inversion result. The maximum recording time is important for a stable Laplace-transformation and affects the inversion result; however, the latter effect is not significant.

Equivalent source distribution for efficient 3-D acoustic wave equation modelling in the Laplace domain

SUMMARY Since the recent introduction of the Laplace-domain full waveform inversion, an efficient and accurate modelling technique for the 3-D Laplace-domain wave equation has been sought. The efficiency and accuracy of the 3-D acoustic wave equation modelling in the Laplace domain strongly depends on how to accurately account for free surface conditions and the actual source and receiver locations. In terms of efficiency, fortunately, the Laplace-domain wave equation can be solved on a coarse grid because the field is not propagating as if it were a potential field. However, it is not possible to accurately compute the Laplace-domain response by assuming that the source and the receivers are located at the grid nodes when we use a coarse grid. To resolve this problem, we propose an equivalent source distribution algorithm that allows us to simulate the free surface condition accurately using a coarse-grid finite-element or finite-difference method. It is shown that the equivalent source vector obtained from a homogeneous half-space model can be used for arbitrarily complex models. The extension of the equivalent source to complex heterogeneous media is explained by the approximation of the Dirac delta function. Numerical tests show that our algorithm is better than the Kaiser windowed sinc function method in the Laplace domain. Our technique for solving the 3-D Laplace-domain wave equation can significantly reduce the computational time required for the 3-D Laplace-domain acoustic full waveform inversion because we can use the coarse grid to accurately simulate conventional marine seismic exploration.