Direct inversion of data-space Hessian for efficient time-domain extended-source waveform inversion using the multiplier method

TL;DR

This work introduces a direct, frequency-domain inversion of the data-space Hessian within a time-domain extended-source FWI framework, using receiver-side Green functions computed in the time domain. By replacing Hessian blocks with circulant surrogates and exploiting FFT-diagonalization, the method forms a block-diagonal Hessian in the frequency domain and solves small $N_r\times N_r$ systems per frequency, enabling simultaneous use across multiple sources. The approach reduces computational complexity from $O(N_x N_r^2 N_t^3)$ to $O(N_x N_r^2 N_t^2)$ and memory from $O(N_r^2 N_t^2)$ to $O(N_r^2 N_\omega)$, with randomized receiver sampling further accelerating the workflow. Numerical experiments on Marmousi II and 2004 BP salt demonstrate substantial speedups and improved data fits, especially for late-arriving and deep structures, validating the method's practicality for large-scale extended-source FWI in complex media.

Abstract

The augmented Lagrangian (AL) method has been successfully applied for solving the full waveform inversion (FWI) problem. In AL-based FWI, the Lagrange multipliers serve as source extensions, offering several advantages to the inversion, such as improved robustness to cycle skipping, faster convergence, and simplified penalty parameter tuning. Time-domain applications of this method have been enabled by reformulating the optimization problem in the data space, significantly reducing memory requirements by projecting source-side multipliers into the data space. These data-side multipliers act as data extensions, effectively expanding the data space. A key challenge in these methods lies in computing the data-side multipliers, which involves solving a linear system to deblur the data residuals using the data-space Hessian matrix before it serves as the adjoint source. This Hessian matrix is prohibitively large to construct and invert explicitly. Iterative Krylov methods can be applied to solve this system as inner iterations, but they require two PDE solves per inner iteration per source, leading to significant computational costs. In this work, we present a key improvement to extended waveform inversion based on multiplier methods. We propose a novel approach that significantly reduces the computational cost of Hessian inversion. The method computes receiver-side Green functions in the time domain and directly constructs frequency-domain Hessian matrices for all required frequencies. These Hessian matrices, with dimensions equal to the number of receivers, can be computed, inverted, and stored in memory. Once constructed, they can be used simultaneously for all sources, further enhancing efficiency. Numerical experiments demonstrate the substantial computational gains achieved by the proposed method, highlighting its effectiveness for extended-source FWI in the time domain.

Figures (20)

• Figure 1: Multiplier-based time-domain extended FWI using data-space Hessian
• Figure 2: Fourier block-diagonal approximation of the data-space Hessian.
• Figure 3: Structure of the data-space Hessian matrix for $N_t=7$ and $N_r=5$. (a) Time domain matrix $\bold{Q}$. (b) Frequency domain matrix $\hat{\bold{Q}}$ (for $N_f=2N_t$). (c) Permutation matrix $\bold{\Pi}$. (d) Frequency domain matrix after permutation $\bold{\Pi}\hat{\bold{Q}}\bold{\Pi}^T$.
• Figure 4: (a) Marmousi II model. (b) Initial model. (c–e) Inversion results using (c) reduced FWI, (d) multiplier-based FWI with diagonal Hessian approximation, and (e) multiplier-based FWI with direct Hessian inversion.
• Figure 5: Shot gather comparisons for the Marmousi II model for a source at 8.5 km. (a) True data. (b) Data from initial model. (c) Relative error vs. iteration for direct inversion, CG, and GMRES. (d–f) Predicted data using direct inversion, CG, and GMRES, respectively. (g–i) Residuals relative to true data.