Accelerating numerical wave-propagation using wavefield adapted meshes, Part I: Forward and adjoint modelling

Driel, Martin van; Boehm, Christian; Krischer, Lion; Afanasiev, Michael

doi:10.31223/osf.io/43ydf

Cited by 2 publications

(2 citation statements)

References 26 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This however changed when the azimuthal smoothness of a global wavefield was discovered by Leng et al (2016), which naturally gave rise to a spectral-element pseudo-spectral method AxiSEM3D (Leng et al 2016(Leng et al , 2019 adapted to wavefield complexity: As the wavefield is sparsely represented on a Fourier basis, it leads to a drastic speedup in computational cost. Building on Leng et al's idea on azimuthal smoothness, this has also been identified by van Driel et al (2019), who showed that the discrete adjoint approach can obtain exact 2D Fréchet kernels at a much lower computational cost. They achieved such adaptability by an alternative implementation of exploiting azimuthal smoothness by azimuthally adaptive meshing instead of Fourier series, thereby retaining the conventional 3D SEM solver strategy at the expense of having to mesh in 3D.…”

Section: Introductionmentioning

confidence: 89%

A complexity-driven framework for waveform tomography with discrete adjoints

Szenicer

Leng

Nissen‐Meyer

2020

Geophysical Journal International

View full text Add to dashboard Cite

Summary We develop a new approach for computing Fréchet sensitivity kernels in full waveform inversion by using the discrete adjoint approach in addition to the widely used continuous adjoint approach for seismic waveform inversion. This method is particularly well suited for the forward solver AxiSEM3D, a combination of the spectral-element method (SEM) and a Fourier pseudo-spectral method, which allows for a sparse azimuthal wavefield parametrization adaptive to wavefield complexity, leading to lower computational costs and better frequency scaling than conventional 3D solvers. We implement the continuous adjoint method to serve as a benchmark, additionally allowing for simulating off-axis sources in axisymmetric or 3D models. The kernels generated by both methods are compared to each other, and benchmarked against theoretical predictions based on linearized Born theory, providing an excellent fit to this independent reference solution. Our verification benchmarks show that the discrete adjoint method can produce exact kernels, largely identical to continuous kernels. While using the continuous adjoint method we lose the computational advantage and fall back on a full-3D frequency scaling, using the discrete adjoint retains the speedup offered by AxiSEM3D. We also discuss the creation of a data-coverage based mesh to run the simulations on during the inversion process, which would allow to exploit the flexibility of the Fourier parametrization and thus the speedup offered by our method.

show abstract

Section: Introductionmentioning

confidence: 89%

A complexity-driven framework for waveform tomography with discrete adjoints

Szenicer

Leng

Nissen‐Meyer

2020

Geophysical Journal International

View full text Add to dashboard Cite

show abstract

“…Similarly, data storage of models, kernels, waveforms etc., necessary to solve the forward and inverse problems, push the limits of modern clusters, while complexity scales as tomography implements larger domains, higher resolutions, and increased data. Nevertheless, the trade-offs of full-waveform methods are attractive given the physical accuracy they provide, and much of the current research in the field is focused on optimizing the forward and inverse problems through novel approaches such as wavefield adapted meshes (Thrastarson et al, 2020), or dynamic batching (van Driel et al, 2020;van Herwaarden et al, 2020).…”

Section: Full-waveform Tomographymentioning

confidence: 99%

Adjoint tomography of the Hikurangi subduction zone and the North Island of New Zealand

Chow¹

2021

Preprint

View full text Add to dashboard Cite

Seismic tomography is a powerful tool for understanding Earth structure. In New Zealand, velocity models derived using ray-based tomography have been used extensively to characterize the complex plate boundary between the Australian and Pacific plates. Advances in computational capabilities now allow us to improve these velocity models using adjoint tomography, an imaging method which minimizes differences between observed and simulated seismic waveforms. We undertake the first application of adjoint tomography in New Zealand to improve a ray-based New Zealand velocity model containing the Hikurangi subduction zone and the North Island of New Zealand. In support of this work we deployed the Broadband East Coast Network (BEACON), a temporary seismic network aimed at improving coverage of the New Zealand permanent network, along the east coast of the North Island. We concurrently develop an automated, open-source workflow for full-waveform inversion using spectral element and adjoint methods. We employ this tool to assess a candidate velocity model’s suitability for adjoint tomography. Using a 3D ray-based traveltime tomography model of New Zealand, we generate synthetic seismic waveforms for more than 10 000 source–receiver pairs and evaluate waveform misfits. We subsequently perform synthetic checkerboard inversions with a realistic New Zealand source–receiver distribution. Reasonable systematic time shifts and satisfactory checkerboard resolution in synthetic inversions indicate that the candidate model is appropriate as an initial model for adjoint tomography. This assessment also demonstrates the relative ease of use and reliability of the automated tools. We then undertake a large-scale adjoint tomography inversion for the North Island of New Zealand using up to 1 800 unique source–receiver pairs to fit waveforms with periods 4–30 s, relating to minimum waveform sensitivities on the order of 5 km. Overall, 60 geographically well-distributed earthquakes and as many as 88 broadband station locations are included. Using a nonlinear optimization algorithm, we undertake 28 model updates of Vp and Vs over six distinct inversion legs which progressively increase resolution. The total inversion incurred a computational cost of approximately 500 000 CPU-hours. The overall time shift between observed and synthetic seismograms is reduced, and updated velocities show as much as ±30% change with respect to initial values. A formal resolution analysis using point spread tests highlights that velocity changes are strongly resolved onland and directly offshore, at depths above 30 km, with low-amplitude changes (> 1%) observed down to 100 km depth. The most striking velocity changes coincide with areas related to the active Hikurangi subduction zone. We interpret the updated velocity model in terms of New Zealand tectonics and geology, and observe good agreement with known basement terranes, and major structural elements such as faults, sedimentary basins, broad-scale subduction related features. We recover increased spatial heterogeneity in seismic velocities along the strike of the Hikurangi subduction zone with respect to the initial model. Below the East Coast, we interpret two localized high-velocity anomalies as previously unidentified subducted seamounts. We corroborate this interpretation with other work, and discuss the implications of deeply subducted seamounts on slip behavior along the Hikurangi margin. In the Cook Strait we observe a low-velocity zone that we interpret as a deep sedimentary basin. Strong velocity gradients bounding this low-velocity zone support hypotheses of a structural boundary here separating the North and South Islands of New Zealand. In the central North Island, low-velocity anomalies are linked to surface geology, and we relate seismic velocities at depth to crustal magmatic activity below the Taupo Volcanic Zone. This new velocity model provides more accurate synthetic seismograms and additional constraints on enigmatic tectonic processes related to the North Island of New Zealand. Both the velocity model itself, and the underpinning methodological contributions, improve our ever-expanding understanding of the North Island of New Zealand, the Hikurangi subduction zone, and the broader Australian-Pacific plate boundary.

show abstract

Accelerating numerical wave-propagation using wavefield adapted meshes, Part I: Forward and adjoint modelling

Cited by 2 publications

References 26 publications

A complexity-driven framework for waveform tomography with discrete adjoints

A complexity-driven framework for waveform tomography with discrete adjoints

Adjoint tomography of the Hikurangi subduction zone and the North Island of New Zealand

Contact Info

Product

Resources

About