Cutting-edge Methods for Seismic Imaging I

6 October 2020 at 10 AM Pacific

Featured Discussants: Jeroen Ritsema (University of Michigan) and Carl Tape (University of Alaska Fairbanks)

Time Presentation

10:00–10:05 AM

Opening Remarks from Clifford Thurber

10:05–10:25 AM

“Markov Chain Monte Carlo Regional Tomography Models of Western North America: Findings from Alaska & Southern California.” Elizabeth Berg, University of Utah, Utah, USA

10:25–10:45 AM

“Adjoint tomography of New Zealand’s North Island Using an Automated, Open-source Workflow.” Bryant Chow, Victoria University of Wellington, Wellington, New Zealand

10:45–11:05 AM

“Seismic Traveltime Tomography Based on Stochastic Voronoi Cells Parameterization: Applications from Local to Global Scales.” Hongjian Fang, Massachusetts Institute of Technology, Massachusetts, USA

11:05–11:10 AM

Break

11:10–11:30 AM

“Accelerating Global-Scale Full-Waveform Inversion.” Sölvi Thrastarson, ETH Zürich, Zürich, Switzerland

11:30–11:50 AM

“Uncertainty Quantification in Full Waveform Tomography with Ensemble Data Assimilation – Methods and Anisotropy.” Julien Thurin, ISTerre, Université Grenoble Alpes, France

11:50–12:20 PM

Question and Answer

12:20–12:50 PM

Breakout Rooms

Thank you to SmartSolo for sponsoring the Virtual
Tomography series!

smartsolo.com

 


Session Recording

Please make sure you are logged in to your SSA account, and click the button below to watch the session recording.

WATCH


Markov Chain Monte Carlo Regional Tomography Models of Western North America: Findings from Alaska & Southern California

Presenting Author: Elizabeth Berg, University of Utah, Utah, USA

Authors: Elizabeth Berg, University of Utah, Utah, USA; Fan-Chi Lin, University of Colorado Boulder, Colorado, USA;  Vera Schulte-Pelkum, University of Colorado Boulder, Colorado, USA; Weisen Shen, State University of New York at Stony Brook, New York, USA; Amir Allam, University of Utah, Utah, USA

We create isotropic shear velocity models of Southern California and Alaska with sensitivity from the near-surface into the upper mantle and resolve Moho depths. This is possible through data recorded on hundreds of stations provided by regional networks in Southern California in 2015, and from the USArray deployment in Alaska from 2014 to 2018. We use Rayleigh-wave measurements from earthquakes for long-periods (15s-100s), and from ambient noise cross-correlations for shorter periods (2s-24s periods). These measurements include Rayleigh-wave phase velocity via Helmholtz tomography and Eikonal tomography for earthquake and ambient noise data respectively, as well as Rayleigh-wave ellipticity. By combining surface wave measurements with azimuthally-corrected receiver functions we obtain sensitivity to near-station subsurface structure, including horizontal interfaces and sensitivity to regional 3-D subsurface structure. We leverage complementary sensitivity of these datasets in a Markov Chain Monte Carlo joint inversion and are able to quantify model uncertainty and gain understanding of sensitivity. In Southern California, we observe well-known large-scale features, including prominent sediment basins in the Central Valley, Los-Angeles, San Bernardino and Ventura basins. We also create self-consistent Moho depth maps across both regions. In Southern California, we observe thicker crust in the Peninsular and Sierra Nevada Ranges, and thinner crust in the Salton Trough and the Coast Ranges. In Alaska, we resolve every major basin, both shallow and deep, and we image the interaction of the Pacific slab and continental crust and its impact to create more or less prominent volcanism. In Alaska, we find thicker crust beneath the Brooks Range and thinner crust in the Yukon Composite Terrane and in interior Alaska.



Adjoint Tomography of New Zealand’s North Island Using an Automated, Open-source Workflow

Presenting Author: Bryant Chow, Victoria University of Wellington, Wellington, New Zealand

Authors: Bryant Chow, Victoria University of Wellington, Wellington, New Zealand; Yoshihiro Kaneko, GNS Science, Wellington, New Zealand, Kyoto University, Kyoto, Japan; Carl Tape, University of Alaska Fairbanks, Alaska, USA; Ryan Modrak, Los Alamos National Laboratory, New Mexico, USA; John Townend, Victoria University of Wellington, Wellington, New Zealand

We have developed an automated, open-source workflow for full-waveform inversion using spectral-element and adjoint methods (Chow et al., 2020). As a study area we choose the eastern North Island of New Zealand, which encompasses the Hikurangi subduction zone. The chosen domain offers a unique opportunity for imaging material properties near an active subduction zone, due to the availability of well-recorded earthquakes in close proximity to the plate interface. We have performed realistic synthetic inversions using New Zealand source and receiver distributions to determine a set of parameters usable in real-data inversions. We have then undertaken an iterative inversion using 5,000 measurements from a subset of long-period (10–30 s) earthquake waveform data to resolve broad-scale velocity changes with respect to an initial ray-based 3D velocity model of New Zealand. Velocity changes are resolved at shallow crustal depths in tectonically active areas, such as the Taupō Volcanic Zone, regions of geodetically detected slow slip, and in the vicinity of the locked-to-creeping transition within the subduction zone. Improved fits between data and synthetic waveforms motivate an ongoing large-scale inversion using 200 earthquakes recorded on as many as 100 broadband stations. Initial iterations are used to fit long-period waveforms (15–30 s), while progressively shorter periods are introduced to a target period of 3 s, corresponding to features on the order of several kilometers in size. We present ongoing efforts towards an accurate, high-resolution tomographic model of the North Island of New Zealand, which will be used to further understanding of enigmatic tectonic processes.



Seismic Traveltime Tomography Based on Stochastic Voronoi Cells Parameterization: Application from Local to Global Scales

Presenting Author: Hongjian Fang, Massachusetts Institute of Technology, Massachusetts, USA

Authors: Hongjian Fang, Massachusetts Institute of Technology, Massachusetts, USA; Malcolm White, University of Southern California, California, USA; Yehuda Ben-Zion, University of Southern California, California, USA; Rob van der Hilst, Massachusetts Institute of Technology, Massachusetts, USA

The amount of seismic arrival time data has increased dramatically in the past few decades. With the use of machine learning based automatic picking technique, the number is expected to grow more rapidly. However, current travel time tomography methods have to use a reduced data set to homogenize the data distribution and to decrease the computational costs. In this study, we propose a new tomography method that adopts a Poisson-Voronoi cells parameterization, which avoids the use of explicit regularization terms by decomposing the original high-dimensional problem into a series of low-dimensional sub-problems. Moreover, we use a subset of the whole dataset in each sub-problem, which is similar to the mini-batch technique in machine learning. The sub data set can be optimally selected in a way that can make the data sampling more homogeneous while making full use of the available data. We apply the method to seismic arrival times at different scales, from a local scale data set in the Ridgecrest area to a global scale in the Earth mantle. Detailed model features will be discussed in the talk.



Accelerating Global-Scale Full-Waveform Inversion

Presenting Author: Sölvi Thrastarson, ETH Zürich, Zürich, Switzerland

Authors: Sölvi Thrastarson, ETH Zürich, Zürich, Switzerland; Dirk-Philip van Herwaarden, ETH Zürich, Zürich, Switzerland; Lion Krischer, ETH Zürich, Zürich, Switzerland; Christian Boehm, ETH Zürich, Zürich, Switzerland; Martin van Driel, ETH Zürich, Zürich, Switzerland; Michael Afanasiev, ETH Zürich, Zürich, Switzerland; Andreas Fichtner, ETH Zürich, Zürich, Switzerland

Full-waveform inversion (FWI) has proven to produce subsurface images of previously unprecedented accuracies. However, with the requirement of multiple accurate seismic wavefield simulations, FWI is a computationally demanding procedure. Incorporating all available data into an FWI study is already a challenge. With the wealth of seismic data increasing exponentially, one has to think of ways to exploit the available data without hitting computational barriers. In this study, we present a novel approach to both numerical simulations and to the adjoint method which reduces the computational cost of FWI by an order of magnitude.

In smooth media it has been observed that wavefield complexity is highly anisotropic. The wavefield is relatively smooth in the lateral direction with respect to the source location, a feature which can be exploited in the meshing process. By adapting the meshing of the wavefield to the expected complexity, the number of required elements to fully resolve the wavefield can be reduced by an order of magnitude. This results in a mesh optimized for a single source location and medium. Accurate gradients can be computed via the discrete adjoint approach.

We will present a real data example of a global FWI where we were able to recover the large scale features of the Earth at the cost of less than 20 regular wavefield simulations. In the example, the wavefield adapted meshes are used in combination with a stochastic mini-batch optimization scheme where only a fraction of the dataset is used in each iteration. The example shows how global FWI can be done with a large dataset, in a computationally tractable manner.



Uncertainty Quantification in Full Waveform Tomography with Ensemble Data Assimilation – Methods and Anisotropy

Presenting Author: Julien Thurin, ISTerre, Université Grenoble Alpes, France

Authors: Julien Thurin, ISTerre, Université Grenoble Alpes, France; Romain Brossier, ISTerre, Université Grenoble Alpes, France; and Ludovic Métivier, LJK, CNRS, Université Grenoble Alpes, France

Uncertainty estimation and quality control are critically missing in most geophysical tomographic applications. The few solutions to cope with that issue are often left out in practical applications when these grow in scale and involve complex modeling. We present a joint full-waveform inversion and ensemble data assimilation scheme, allowing local Bayesian estimation of the solution that brings uncertainty estimation to the tomographic problem. This original methodology relies on a deterministic square root ensemble Kalman filter commonly used in the data assimilation community: the ensemble transform Kalman filter. Combined with a 2D visco-acoustic frequency-domain full-waveform inversion scheme, the resulting method allows access to a low-rank approximation of the posterior covariance matrix of the solution. It yields uncertainty information through an ensemble-representation, that can conveniently be mapped across the physical domain for visualization and interpretation. We present the combination of ensemble transform Kalman filter with full-waveform inversion along with the scheme design and algorithmic details that lead to our mixed application. Both synthetic and field-data results are considered, along with the biases that are associated with the limited rank ensemble representation. Finally, we review the open questions and development perspectives linked with data assimilation applications to the tomographic problem.