Frequency-domain Full Waveform Inversion Research Articles

Preconditioned non-linear conjugate gradient method for frequency domain full-waveform seismic inversion

We present preconditioned non-linear conjugate gradient algorithms as alternatives to the Gauss-Newton method for frequency domain full-waveform seismic inversion. We designed two preconditioning operators. For the first preconditioner, we introduce the inverse of an approximate sparse Hessian matrix. The approximate Hessian matrix, which is highly sparse, is constructed by judiciously truncating the Gauss-Newton Hessian matrix based on examining the auto-correlation and cross-correlation of the Jacobian matrix. As the second preconditioner, we employ the approximation of the inverse of the Gauss-Newton Hessian matrix. This preconditioner is constructed by terminating the iteration process of the conjugate gradient least-squares method, which is used for inverting the Hessian matrix before it converges. In our preconditioned non-linear conjugate gradient algorithms, the step-length along the search direction, which is a crucial factor for the convergence, is carefully chosen to maximize the reduction of the cost function after each iteration. The numerical simulation results show that by including a very limited number of non-zero elements in the approximate Hessian, the first preconditioned non-linear conjugate gradient algorithm is able to yield comparable inversion results to the Gauss-Newton method while maintaining the efficiency of the un-preconditioned non-linear conjugate gradient method. The only extra cost is the computation of the inverse of the approximate sparse Hessian matrix, which is less expensive than the computation of a forward simulation of one source at one frequency of operation. The second preconditioned non-linear conjugate gradient algorithm also significantly saves the computational expense in comparison with the Gauss-Newton method while maintaining the Gauss-Newton reconstruction quality. However, this second preconditioned non-linear conjugate gradient algorithm is more expensive than the first one.

Two-dimensional frequency-domain visco-elastic full waveform inversion: Parallel algorithms, optimization and performance

Full waveform inversion (FWI) is an appealing seismic data-fitting procedure for the derivation of high-resolution quantitative models of the subsurface at various scales. Full modelling and inversion of visco-elastic waves from multiple seismic sources allow for the recovering of different physical parameters, although they remain computationally challenging tasks. An efficient massively parallel, frequency-domain FWI algorithm is implemented here on large-scale distributed-memory platforms for imaging two-dimensional visco-elastic media. The resolution of the elastodynamic equations, as the forward problem of the inversion, is performed in the frequency domain on unstructured triangular meshes, using a low-order finite element discontinuous Galerkin method. The linear system resulting from discretization of the forward problem is solved with a parallel direct solver. The inverse problem, which is presented as a non-linear local optimization problem, is solved in parallel with a quasi-Newton method, and this allows for reliable estimation of multiple classes of visco-elastic parameters. Two levels of parallelism are implemented in the algorithm, based on message passing interfaces and multi-threading, for optimal use of computational time and the core-memory resources available on modern distributed-memory multi-core computational platforms. The algorithm allows for imaging of realistic targets at various scales, ranging from near-surface geotechnic applications to crustal-scale exploration.

Which data residual norm for robust elastic frequency-domain full waveform inversion?

Elastic full-waveform inversion is an ill-posed data-fitting procedure that is sensitive to noise, inaccuracies of the starting model, definition of multiparameter classes, and inaccurate modeling of wavefield amplitudes. We have investigated the performance of different minimization functionals as the least-squares norm [Formula: see text], the least-absolute-values norm [Formula: see text], and combinations of both (the Huber and so-called hybrid criteria) with reference to two noisy offshore (Valhall model) and onshore (overthrust model) synthetic data sets. The four minimization functionals were implemented in 2D elastic frequency-domain full-waveform inversion (FWI), where efficient multiscale strategies were designed by successive inversions of a few increasing frequencies. For the offshore and onshore case studies, the [Formula: see text]-norm provided the most reliable models for P- and S-wave velocities ([Formula: see text] and[Formula: see text]), even when strongly decimated data sets that correspond to few frequencies were used in the inversion and when outliers polluted the data. The [Formula: see text]-norm can provide reliable results in the presence of uniform white noise for [Formula: see text] and [Formula: see text] if the data redundancy is increased by refining the frequency sampling interval in the inversion at the expense of computational efficiency. The [Formula: see text]-norm and the Huber and hybrid criteria, unlike the [Formula: see text]-norm, allow for successful imaging of the [Formula: see text] model from noisy data in a soft-seabed environment, where the P-to-S-waves have a small footprint in the data. However, the Huber and hybrid criteria are sensitive to a threshold criterion that controls the transition between the criteria and that requires tedious trial-and-error investigations for reliable estimation. The [Formula: see text]-norm provides a robust alternative to the [Formula: see text]-norm for inverting decimated data sets in the framework of efficient frequency-domain FWI.

Three-dimensional frequency-domain full-waveform inversion with an iterative solver

With the acquisition of wide-aperture seismic data sets, full-waveform inversion is an attractive method for deriving velocity models. Three-dimensional implementations require an efficient solver for the wave equation. Computing 3D time-harmonic responses with a frequency-domain solver is complicated because a large linear system with negative and positive eigenvalues must be solved. Time-domain schemes are an alternative. Nevertheless, existing frequency-domain iterative solvers with an efficient preconditioner are a viable option when full-waveform inversion is formulated in the frequency domain. An iterative solver with a multigrid preconditioner is competitive because of a high-order spatial discretization. Numerical examples illustrated the efficiency of the iterative solvers. Three dimensional full-waveform inversion was then studied in the context of deep-water ocean-bottom seismometer acquisition. Three dimensional synthetic data inversion results showed the behavior of full-waveform inversion with respect to the initial model and the minimum frequency available in the data set. Results on a 3D real ocean-bottom seismometer data set demonstrated the relevance of full-waveform inversion, especially to image the shallow part of the model.

Seismic imaging of complex onshore structures by 2D elastic frequency-domain full-waveform inversion

Quantitative imaging of the elastic properties of the subsurface at depth is essential for civil engineering applications and oil- and gas-reservoir characterization. A realistic synthetic example provides for an assessment of the potential and limits of 2D elastic full-waveform inversion (FWI) of wide-aperture seismic data for recovering high-resolution P- and S-wave velocity models of complex onshore structures. FWI of land data is challenging because of the increased nonlinearity introduced by free-surface effects such as the propagation of surface waves in the heterogeneous near-surface. Moreover, the short wavelengths of the shear wavefield require an accurate S-wave velocity starting model if low frequencies are unavailable in the data. We evaluated different multiscale strategies with the aim of mitigating the nonlinearities. Massively parallel full-waveform inversion was implemented in the frequency domain. The numerical optimization relies on a limited-memory quasi-Newton algorithm thatoutperforms the more classic preconditioned conjugate-gradient algorithm. The forward problem is based upon a discontinuous Galerkin (DG) method on triangular mesh, which allows accurate modeling of free-surface effects. Sequential inversions of increasing frequencies define the most natural level of hierarchy in multiscale imaging. In the case of land data involving surface waves, the regularization introduced by hierarchical frequency inversions is not enough for adequate convergence of the inversion. A second level of hierarchy implemented with complex-valued frequencies is necessary and provides convergence of the inversion toward acceptable P- and S-wave velocity models. Among the possible strategies for sampling frequencies in the inversion, successive inversions of slightly overlapping frequency groups is the most reliable when compared to the more standard sequential inversion of single frequencies. This suggests that simultaneous inversion of multiple frequencies is critical when considering complex wave phenomena.

FWT2D: A massively parallel program for frequency-domain full-waveform tomography of wide-aperture seismic data—Part 1: Algorithm

This is the first paper in a two-part series that describes a massively parallel code that performs 2D frequency-domain full-waveform inversion of wide-aperture seismic data for imaging complex structures. Full-waveform inversion methods, namely quantitative seismic imaging methods based on the resolution of the full wave equation, are computationally expensive. Therefore, designing efficient algorithms which take advantage of parallel computing facilities is critical for the appraisal of these approaches when applied to representative case studies and for further improvements. Full-waveform modelling requires the resolution of a large sparse system of linear equations which is performed with the massively parallel direct solver MUMPS for efficient multiple-shot simulations. Efficiency of the multiple-shot solution phase (forward/backward substitutions) is improved by using the BLAS3 library. The inverse problem relies on a classic local optimization approach implemented with a gradient method. The direct solver returns the multiple-shot wavefield solutions distributed over the processors according to a domain decomposition driven by the distribution of the LU factors. The domain decomposition of the wavefield solutions is used to compute in parallel the gradient of the objective function and the diagonal Hessian, this latter providing a suitable scaling of the gradient. The algorithm allows one to test different strategies for multiscale frequency inversion ranging from successive mono-frequency inversion to simultaneous multifrequency inversion. These different inversion strategies will be illustrated in the following companion paper. The parallel efficiency and the scalability of the code will also be quantified.

Parsimonious finite-volume frequency-domain method for 2-DP-SV-wave modelling

SUMMARY A new numerical technique for solving 2-D elastodynamic equations based on a finite-volume frequency-domain approach is proposed. This method has been developed as a tool to perform 2-D elastic frequency-domain full-waveform inversion. In this context, the system of linear equations that results from the discretization of the elastodynamic equations is solved with a direct solver, allowing efficient multiple-source simulations at the partial expense of the memory requirement. The discretization of the finite-volume approach is through triangles. Only fluxes with the required quantities are shared between the cells, relaxing the meshing conditions, as compared to finite-element methods. The free surface is described along the edges of the triangles, which can have different slopes. By applying a parsimonious strategy, the stress components are eliminated from the discrete equations and only the velocities are left as unknowns in the triangles. Together with the local support of the P0 finite-volume stencil, the parsimonious approach allows the minimizing of core memory requirements for the simulation. Efficient perfectly matched layer absorbing conditions have been designed for damping the waves around the grid. The numerical dispersion of this FV formulation is similar to that of O(Δx2) staggered-grid finite-difference (FD) formulations when considering structured triangular meshes. The validation has been performed with analytical solutions of several canonical problems and with numerical solutions computed with a well-established FD time-domain method in heterogeneous media. In the presence of a free surface, the finite-volume method requires 10 triangles per wavelength for a flat topography, and fifteen triangles per wavelength for more complex shapes, well below the criteria required by the staircase approximation of O(Δx2) FD methods. Comparisons between the frequency-domain finite-volume and the O(Δx2) rotated FD methods also show that the former is faster and less memory demanding for a given accuracy level, an attractive feature for frequency-domain seismic inversion. We have thus developed an efficient method for 2-D P–SV-wave modelling on structured triangular meshes as a tool for frequency-domain full-waveform inversion. Further work is required to improve the accuracy of the method on unstructured meshes.

Velocity model building by 3D frequency-domain, full-waveform inversion of wide-aperture seismic data

We assessed 3D frequency-domain (FD) acoustic full-waveform inversion (FWI) data as a tool to develop high-resolution velocity models from low-frequency global-offset data. The inverse problem was posed as a classic least-squares optimization problem solved with a steepest-descent method. Inversion was applied to a few discrete frequencies, allowing management of a limited subset of the 3D data volume. The forward problem was solved with a finite-difference frequency-domain method based on a massively parallel direct solver, allowing efficient multiple-shot simulations. The inversion code was fully parallelized for distributed-memory platforms, taking advantage of a domain decomposition of the modeled wavefields performed by the direct solver. After validation on simple synthetic tests, FWI was applied to two targets (channel and thrust system) of the 3D SEG/EAGE overthrust model, corresponding to 3D domains of [Formula: see text] and [Formula: see text], respectively. The maximum inverted frequencies are 15 and [Formula: see text] for the two applications. A maximum of 30 dual-core biprocessor nodes with [Formula: see text] of shared memory per node were used for the second target. The main structures were imaged successfully at a resolution scale consistent with the inverted frequencies. Our study confirms the feasibility of 3D frequency-domain FWI of global-offset data on large distributed-memory platforms to develop high-resolution velocity models. These high-velocity models may provide accurate macromodels for wave-equation prestack depth migration.

3D finite-difference frequency-domain modeling of visco-acoustic wave propagation using a massively parallel direct solver: A feasibility study

We present a finite-difference frequency-domain method for 3D visco-acoustic wave propagation modeling. In the frequency domain, the underlying numerical problem is the resolution of a large sparse system of linear equations whose right-hand side term is the source. This system is solved with a massively parallel direct solver. We first present an optimal 3D finite-difference stencil for frequency-domain modeling. The method is based on a parsimonious staggered-grid method. Differential operators are discretized with second-order accurate staggered-grid stencils on different rotated coordinate systems to mitigate numerical anisotropy. An antilumped mass strategy is implemented to minimize numerical dispersion. The stencil incorporates 27 grid points and spans two grid intervals. Dispersion analysis showsthat four grid points per wavelength provide accurate simulations in the 3D domain. To assess the feasibility of the method for frequency-domain full-waveform inversion, we computed simulations in the 3D SEG/EAGE overthrust model for frequencies 5, 7, and [Formula: see text]. Results confirm the huge memory requirement of the factorization (several hundred Figabytes) but also the CPU efficiency of the resolution phase (few seconds per shot). Heuristic scalability analysis suggests that the memory complexity of the factorization is [Formula: see text] for a [Formula: see text] grid. Our method may provide a suitable tool to perform frequency-domain full-waveform inversion using a large distributed-memory platform. Further investigation is still necessary to assess more quantitatively the respective merits and drawbacks of time- and frequency-domain modeling of wave propagation to perform 3D full-waveform inversion.

An Adaptive Wavenumber Sampling Strategy for 2.5D Seismic-Wave Modeling in the Frequency Domain

Simulation of seismic waves from a 3D point-source in a 2D medium may be performed in the frequency-wavenumber domain (called 2.5D modelling). It involves computing the Fourier-transformed Green's function for a number of frequency (ω) and strike direction wavenumber (k y ) values and doubly inverse transforming to convert to the traveltime and distance space. Such modeling produces a wavefield with 3D features but the computation becomes pseudo 2D (i.e., in the xz-plane) rather than 3D (in the xyz-frame). The common sampling strategy for the wavenumber is inefficient for 2.5D wave modeling because it employs a large number of wavenumbers (k y ). This leads to a high cost of computer time in the linear-equation-solving processing, which detracts from the advantages of 2.5D modeling. In this paper, we use two analytic frequency-wavenumber-domain solutions for seismic waves in a homogeneous medium and an inhomogeneous media (two semi-infinite media in contact) to investigate the properties of the solutions and an efficient sampling strategy for choosing the wavenumbers. We have carried out analytic and numerical experiments with these solutions, and present adaptive Gauss–Legendre abscissae for the wavenumber sampling in terms of a modeling situation. We show that the effective range and the number of sampling points of the wavenumber define the adaptive sampling strategy, and they can be estimated in terms of the wavelength and the maximum source-receiver offset. We apply this sampling strategy to the finite-element method and demonstrate that the range and number of sampling points may be adapted for obtaining significant computational efficiency and satisfactory accuracy for every frequency component. Such 2.5D wave modeling can be readily applied for frequency-domain full-waveform inversion for seismic surface measurements and crosshole seismic waveform tomography.