Data-adaptive global full-waveform inversion

Abstract

Summary We present a novel approach to global-scale full-waveform inversion (FWI) that can reduce computational cost by over an order of magnitude, compared to previously published methods, without sacrificing physical and mathematical rigor. This is based on data-adaptation, and thereby application-oriented specialization, on two complementary levels. On the simulation level, we exploit the approximate azimuthal symmetry of seismic wavefields by implementing wavefield-adapted meshes and discrete adjoints, thereby lowering numerical simulation cost. On the measurement level, we employ a quasi-stochastic approach where variable mini-batches of data are used during an iterative misfit minimisation in order to promote a parsimonious exploitation of data. In addition to the methodological developments, we present an inversion of long-period (100 - 200 s) seismic waveforms from 1179 earthquakes for 3-D whole-mantle structure. The computational cost of the 72 iterations in the inversion approximately equals one third of a single iteration using an FWI approach with widely-used cubed-sphere-based meshes and non-stochastic gradient optimization. The resulting LOng-Wavelength Earth model (LOWE) constitutes the first global FWI constructed entirely from a spherically symmetric initial mantle structure. While mostly serving as a showcase for the method, LOWE contains a wealth of regional-scale structures that compare well to earlier tomographic images. Being conservatively smooth and based on minimal assumptions, it may therefore serve as starting model for future inversions at shorter period or smaller scales.

Full-waveform tomography of the African Plate

“…The area beneath Tanzania and Kenya is thought to be home to a mantle superplume (e.g. Ebinger & Sleep, 1998;Thrastarson et al, 2022), and our model aligns with those findings. The model suggests that the low-velocity anomaly beneath the Tanzania region may become less pronounced below the 660 km discontinuity.…”

Full‐Waveform Tomography of the African Plate Using Dynamic Mini‐Batches

“…Low velocities extend deep downwards beneath the Afar region and Tanzania. The area beneath Tanzania and Kenya is thought to be home to a mantle superplume (e.g., Boyce et al., 2021; Ebinger & Sleep, 1998; Moucha & Forte, 2011; Ritsema et al., 1999; Thrastarson et al., 2022), and our model aligns with those findings.…”

“…In such a case, the features we introduce will likely generalize to unseen data, and we are probably further improving the model rather than just fitting noise. Thrastarson et al (2022) have introduced and explained this concept in more detail.…”

CANVAS: An Adjoint Waveform Tomography Model of California and Nevada

“…In this study, we used Salvus (Afanasiev et al., 2019), a FWI workflow manager that handles all steps from data acquisition to inversions. Salvus has successfully been used to compute AWT models at regional scales (Doody et al., 2023; Gao et al., 2021; Rodgers et al., 2022; Wehner, Blom, et al., 2022; Wehner, Rawlinson, et al., 2022) and global scales (Thrastarson et al., 2022). The description below provides a broad overview of the methodology used.…”