Abstract: Today machine learning is being applied to solve various problems in seismology using large available datasets. In many cases these data have already been analyzed by humans and serve as training or testing datasets. However, for some important problems in seismology we are severely data-limited. In particular, the enormous numbers of free parameters in deep learning algorithms require amounts of data that are often far larger than what is available today. This data limitation is especially severe for very large events, which are rare but very important cases. Traditional data-augmentation methods include interpolation and adding random noise to existing data. Here we suggest a way to generate realistic synthetic seismic waveforms using generative adversarial networks (GANs). We train a GAN to capture the major features of a large data set of real seismic waveforms, and to synthesize new waveforms. We show that these synthetic waveforms are realistic enough to fool professionals and contain realistic physical features, e.g. major body wave packets, coda wave decay and frequency dispersion. Compared to simulated synthetic waveforms, they contain some realistic high-frequency features that are not easily modeled by more traditional waveform simulation methods. Beyond augmenting the test data, the GANs can have potentially wide applications in seismology, such as seismic event/phase discrimination, clipped waveform completion, and waveform up-sampling. Updated results will be presented in the meeting.
Media has not been submitted for this Presentation
Abstract: We are developing a new seismic hazard model for Mainland China by integrating historical earthquake catalogs, geological faults, geodetic GPS data, and geology maps. To build the model, we construct an Mw-based homogeneous historical earthquake catalog spanning from 780 B.C. to present, create fault models from active fault data, and derive a strain rate model based on the most complete GPS measurements and a new strain derivation algorithm. We divide China and the surrounding regions into about 20 large seismic source zones. For each zone, a tapered Gutenberg-Richter (TGR) magnitude-frequency distribution is used to model the seismic activity rates. The a- and b-values of the TGR distribution are calculated using observed earthquake data, while the corner magnitude is constrained independently using the seismic moment rate inferred from the geodetically-based strain rate model. Small and medium sized earthquakes are distributed within the source zones following the location and magnitude patterns of historical earthquakes. Some of the larger earthquakes are distributed onto active faults, based on their geological characteristics such as slip rate, fault length, down-dip width, and various paleoseismic data. The remaining larger earthquakes are then placed into the background. A new set of magnitude-rupture scaling relationships is developed based on earthquake data from China and vicinity. We evaluate and select appropriate ground motion prediction equations by comparing them with observed ground motion data and performing residual analysis. To implement the modeling workflow, we develop a tool that builds upon the functionalities of GEM’s Hazard Modeler’s Toolkit. The GEM OpenQuake software is used to calculate seismic hazard at various ground motion periods and various return periods. To account for site amplification, we construct a site condition map based on geology. The resulting new seismic hazard maps can be used for seismic risk analysis and management.
Media has not been submitted for this Presentation
Abstract: Eighty-two percent of sand tailings dam incidents in Chile since 1915 have been seismically induced, including that which led to the 1965 El Cobre disaster. Subduction earthquakes (magnitudes 7.5+, return periods 10s−100s of years) have been traditionally regarded as representing the principal seismic hazard in the country and characteristically produce moderate, but widespread, damage. In contrast, much shallower earthquakes may be generated by potentially seismogenic crustal faults (PSCFs); these have relatively lower maximum magnitudes (7.0−7.5) and longer return periods (100s−1000s of years), yet are capable of producing extensive local damage within 5 km of the fault rupture. To our knowledge, a systemic study of the exposure of tailings storage facilities (TSFs) to earthquakes that could be generated by PSCFs has not been undertaken in Chile. Our study has filled this gap through integrating two publically available data sets: the 2016 national tailings database of SERNAGEOMIN and the catalogue of PSCFs published by the South America Risk Assessment Project of the Global Earthquake Model. There are 696 registered TSFs of varying size, age, operational status and construction type distributed throughout central and northern Chile. Of these, 91 lie within 5 km of a PSCF and 17 have been authorized to store volumes of tailings >106 m3. In order to constrain the exposure of TSFs more robustly, a deterministic seismic hazard assessment along each major PSCF is now being undertaken. The completeness of the study is limited by the fact that it is anticipated that most PSCFs have poorly known fundamental fault parameters (e.g., slip rates and recurrence times), which have only been investigated systematically over the last 10 years. Despite that, recent large earthquakes nucleated on previously unidentified PSCFs have raised awareness of their potential hazard. Our main conclusion is that complete seismic risk assessments of TSFs in Chile must incorporate PSCFs.
Media has not been submitted for this Presentation
Abstract: The use of a constrained joint inversion approach reduces the effect of noisy data into the inversion, since it make use of physically meaningful bound constraints to regularize the inversion process. In this study, we analyze waveforms data collected from the Libyan seismic network (LNSN) to compute Rayleigh wave dispersion curves and Ps-receiver functions. Rayleigh dispersion curves were measured for Rayleigh waves over periods of 20-80 s, and we invert the dispersion curves to obtain group velocities on a grid of 2.0° x 2.0° at specific periods. We first calculate crustal thickness using a standard receiver functions stacking (RFS) methodology, and then jointly invert Rayleigh waves and Ps-receiver using a constrained joint inversion approach to produce 1-D shear wave velocity models beneath all the LNSN stations. The data is not optimal, and we are limited in the number of receiver functions and group velocity curves. Regardless, the constrained joint inversion approach allows us to produce 1-D shear wave velocity models beneath all the LNSN stations. The results indicated a well-constrained depth compared to those obtained by inverted RFS, even with low quality data present in some stations. The depth to the Moho beneath all stations includes stations (TBQ, GHD, and ZLA) not previously reported. In addition, we report differences in Moho`s depth for other stations from previous results. Our results show a geological structure that highlights the local tectonic provinces.
Media has not been submitted for this Presentation
Abstract: The Source Physics Experiment (SPE) is an ongoing effort to improve explosion monitoring by conducting a controlled series of chemical explosions at the Nevada National Security Site (NNSS) and using the resulting observations to improve and validate physics-based simulations of explosion phenomena. Phase I of SPE was conducted on the Climax Stock granite which contains a network of well-characterized joints. It has been shown that these pre-existing joints may be responsible for the tangential motion observed during SPE chemical explosions. Near-field motions generated with hydrodynamic non-linear source models have been coupled to elastic wave propagation codes to propagate these resulting motions into the far-field domain which is assumed to be elastic and isotropic. This is likely not the case as the pre-existing joints continue beyond the inelastic source region of the explosion. To alleviate this impediment, we extend the current near-field to far-field, hydrodynamic-to-elastic coupling, from anisotropic-isotropic to fully anisotropic-anisotropic coupling. Near source hydrodynamic motions are computed using GEODYN-L while anisotropic elastic wave propagation is modeled using SW4. Motions are coupled between the two codes by introducing hydrodynamic motions from GEODYN-L as an internal boundary source to SW4. The anisotropic material model employed in the SW4 domain is derived from the properties of an observed fracture network with relatively well-constrained joint size, density, orientation, and aperture. We show that consideration of anisotropic material in the elastic regime has an important effect on the propagation of tangential motion. Propagation of motions generated in an anisotropic source region into an isotropic far-field domain may hinder the continuity of the waves in general and may impede the shear motion generation. Prepared by LLNL under Contract DE-AC52-07NA27344. LLNL-ABS-744746.
Media has not been submitted for this Presentation
Abstract: Double Benioff Zones (DBZs) have been previously observed in the Hikurangi trench, New Zealand. The correlation between seismicity and velocity heterogeneities can help understanding of their occurrence. Therefore, seismic tomography is commonly applied to identify the position and extent of the down-going slabs at depth in addition to crustal and upper mantle velocity heterogeneities. Previous studies developed three-dimensional seismic velocity models based on local and regional data. The recent availability of multiscale seismic tomography and waveform cross-correlation can boost the structural and seismicity resolution especially in areas with sparse data. Here, we investigate the characteristics of the DBZs along the strike of Hikurangi. We obtain local and regional waveform data from Geonet and teleseismic arrival times reported by the International Seismological Center for events within magnitude range of 2-5 between January 2012 and July 2017. We apply the teletomoDD algorithm based on absolute and differential times in different scales. This algorithm uses nested regional-global model parameterization that builds a coarser global model encompassing a finely gridded regional one. A differential-time relocation method based on waveform cross-correlation data improves the relative locations with ~500 m uncertainties in both horizontal and vertical. Our preliminary relocation results show a clear DBZ in the north and south of the trench, but the DBZ disappears in the central North Island. We also observe a decrease in DBZ layer separation and bending of the slab from the north to the south of the trench. Seismicity in the upper and lower planes if correlated with velocity anomalies provides insights about the causes of the within-slab-deformation (e.g. slab dehydration). Tracking such correlation along the trench may explain the presence/absence of the DBZs in some cross-sections and thereafter the mechanism responsible for intermediate-depth earthquakes.
Media has not been submitted for this Presentation
