ALBERT

Description: Seismic event detection and phase picking are the base of many seismological workflows. In recent years, several publications demonstrated that deep learning approaches significantly outperform classical approaches, achieving human-like performance under certain circumstances. However, as studies differ in the datasets and evaluation tasks, it is unclear how the different approaches compare to each other. Furthermore, there are no systematic studies about model performance in cross-domain scenarios, i.e., when applied to data with different characteristics. Here, we address these questions by conducting a large-scale benchmark. We compare six previously published deep learning models on eight datasets covering local to teleseismic distances and on three tasks: event detection, phase identification and onset time picking. Furthermore, we compare the results to a classical Baer-Kradolfer picker. Overall, we observe the best performance for EQTransformer, GPD and PhaseNet, with a small advantage for EQTransformer on teleseismic data. Furthermore, we conduct a cross-domain study, analyzing model performance on datasets they were not trained on. We show that trained models can be transferred between regions with only mild performance degradation, but models trained on regional data do not transfer well to teleseismic data. As deep learning for detection and picking is a rapidly evolving field, we ensured extensibility of our benchmark by building our code on standardized frameworks and making it openly accessible. This allows model developers to easily evaluate new models or performance on new datasets. Furthermore, we make all trained models available through the SeisBench framework, giving end-users an easy way to apply these models.

Description: The task of association of seismic phases into events is particularly challenging in every workflow dealing with seismically active regions where dense networks are usually deployed. Additionally, recents developments of picking algorithms based on deep learning frameworks provide an increased number of phase onsets with respect to standard approach based on the computation of characteristic functions of the seismic waveforms. Especially the number of S-picks has been increased by orders of magnitude, which now requires the simultaneous association of P- and S-wave arrivals or even detecting events only based on S-wave arrivals. We present an improved version of HEX (Hyperbolic Event eXtractor), a new technique based on the logic of Random Sample Consensus here applied to association of seismic phases. Since this algorithm is particularly effective in dealing with high noise in the input data, we show the benefits of HEX2.0 to analyze picks from deep learning methods. These datasets are characterized by either detections of small amplitude earthquakes that, according to network geometry, may not show at a sufficient number of seismic stations to declare an event or, sometimes, to a high rate of false positives. The application of HEX2.0 on real data from a seismic sequence of Sannio-Matese (Southern Apennines, Italy) occurred in 2013-2014 show that: i) resulting events show a high number of phases compared to previous catalogs; ii) few phases are discarded in the event location. HEX2.0 provides an accurate, easy-to-use and computationally effective solution to the seismic phase association problem.

Description: Machine-learning (ML) methods have seen widespread adoption in seismology in recent years. The ability of these techniques to efficiently infer the statistical properties of large datasets often provides significant improvements over traditional techniques when the number of data are large (millions of examples). With the entire spectrum of seismological tasks, for example, seismic picking and detection, magnitude and source property estimation, ground-motion prediction, hypocenter determination, among others, now incorporating ML approaches, numerous models are emerging as these techniques are further adopted within seismology. To evaluate these algorithms, quality-controlled benchmark datasets that contain representative class distributions are vital. In addition to this, models require implementation through a common framework to facilitate comparison. Accessing these various benchmark datasets for training and implementing the standardization of models is currently a time-consuming process, hindering further advancement of ML techniques within seismology. These development bottlenecks also affect “practitioners” seeking to deploy the latest models on seismic data, without having to necessarily learn entirely new ML frameworks to perform this task. We present SeisBench as a software package to tackle these issues. SeisBench is an open-source framework for deploying ML in seismology—available via GitHub. SeisBench standardizes access to both models and datasets, while also providing a range of common processing and data augmentation operations through the API. Through SeisBench, users can access several seismological ML models and benchmark datasets available in the literature via a single interface. SeisBench is built to be extensible, with community involvement encouraged to expand the package. Having such frameworks available for accessing leading ML models forms an essential tool for seismologists seeking to iterate and apply the next generation of ML techniques to seismic data.

Description: On June 21st, a Mw6.2 earthquake struck the Afghan-Pakistan-border-region, situated within the India-Asia collision. Thousand thirty-nine deaths were reported, making the earthquake the deadliest of 2022. We investigate the event's rupture processes by combining seismological and geodetic observations, aiming to understand what made it that fatal. Our Interferometric Synthetic Aperture Radar-constrained slip-model and regional moment-tensor inversion, confirmed through field observations, reveal a sinistral rupture with maximum slip of 1.8 m at 5 km depth on a N20°E striking, sub-vertical fault. We suggest that not only external factors (event-time, building stock) but fault-specific factors made the event excessively destructive. Surface rupture was favored by the rock foliation, coinciding with the fault strike. The distribution of Peak-Ground-Velocity was governed by the sub-vertical fault. Maximum slip was large compared to other events globally and might have resulted in peak-frequencies coinciding with resonance-frequencies of the local buildings and demonstrates the devastating impact of moderate-size earthquakes.

Description: Detecting phase arrivals and pinpointing the arrival times of seismic phases in seismograms is crucial for many seismological analysis workflows. For land station data, machine learning methods have already found widespread adoption. However, deep learning approaches are not yet commonly applied to ocean bottom data due to a lack of appropriate training data and models. Here, we compiled an extensive and labeled ocean bottom seismometer (OBS) data set from 15 deployments in different tectonic settings, comprising ∼90,000 P and ∼63,000 S manual picks from 13,190 events and 355 stations. We propose PickBlue, an adaptation of the two popular deep learning networks EQTransformer and PhaseNet. PickBlue joint processes three seismometer recordings in conjunction with a hydrophone component and is trained with the waveforms in the new database. The performance is enhanced by employing transfer learning, where initial weights are derived from models trained with land earthquake data. PickBlue significantly outperforms neural networks trained with land stations and models trained without hydrophone data. The model achieves a mean absolute deviation of 0.05 s for P-waves and 0.12 s for S-waves, and we apply the picker on the Hikurangi Ocean Bottom Tremor and Slow Slip OBS deployment offshore New Zealand. We integrate our data set and trained models into SeisBench to enable an easy and direct application in future deployments.

Description: Northern Chile sets an ideal laboratory to perform seismic experiments to better understand the subduction process in highly erosive margins. The Taltal segment (~22°S - ~26°S) has been subject of notable scientific attention; however, previous studies have mainly focused on the areas affected by large earthquakes, leaving unattended the processes in the overriding plate. By benefiting from a large temporary network of 84 short-period geophones and the high rate of seismicity in the region, we tectonically characterize the Taltal segment by deriving regional 3D Vp, Vs, and Vp/Vs velocity models, a seismic catalog of 23,000 earthquakes and focal mechanisms for events with M〉 4.0. The seismic velocity models illuminate first-order structures such as the Nazca plate, the upper crust of the South American plate, and highlight changes from high to low ratios in the overriding plate that correlate with large-scale structures such as the Atacama fault system (AFS) in the coastline and the West Fault System (WFS) towards the Andes. In terms of the seismicity distribution, we observe a dip change, at intermediate depth (150-200 km), that could be associated with slab-pull activity. Coastal area shows clustered seismicity that might indicate splay faulting reaching the plate interface. Upper-crust activity is mainly related to the AFS and WFS but also to the circulation of fluids around the Lascar volcano. Finally, to get a comprehensive picture of the seismotectonics of the Taltal segment we explore the solutions of focal mechanisms that better describe the type of sources of the earthquakes.

Description: In seismological full waveform inversion, the choice of misfit functions plays a critical role in quantifying the discrepancy between observed and synthetic data, affecting convergence rate and also final results. We revisit and compare six commonly used misfit functions, including cross-correlation time-shift (CC), least-square waveform difference (L2), multitaper time-shift (MT), exponentiated phase shift (EP), time–frequency phase shift (TF) and zero-lag cross-correlation coefficient (CCC), with respect to their definitions, adjoint sources and misfit kernels for velocity perturbations. Synthetic tests are performed for several canonical models. First, we simulated wave propagation in a model with a single rectangular anomaly with sharp boundaries and a smoothed variant of that model. We analysed the resulting misfit kernels first for the P-wave phase, which is highly distorted in the sharp model due to strong heterogeneities, and mostly experiences traveltime perturbations in the smooth model. Second, we considered a model where a laterally limited region is subject to layered anomalies (low velocity in the middle crust and high velocity in the lower crust) and determine misfit kernels for S and surface waves in this model. Based on these two simplified seismological scenarios, we further perform iterative test inversions using different misfit functions. Combining the features of misfit kernels and synthetic inversion results, we find that CCC, L2 and EP are the most effective at identifying the sharpness of velocity anomalies from the direct body waves and their scattered phases. Consequently, inversion based on those misfit measures yielded the best recovery in the inversion test. For surface and S waves from crustal sources, TF appears to be the most effective in constraining the heterogeneous structure in the crust but needs more iterations for convergence than other misfit functions.

Description: The recent rapid improvement of machine learning techniques had a large impact on the way seismological data can be processed. During the last years several machine learning algorithms determining seismic onset times have been published facilitating the automatic picking of large data sets. Here we apply the deep neural network PhaseNet to a network of over 900 permanent and temporal broad band stations that were deployed as part of the AlpArray research initiative in the Greater Alpine Region (GAR) during 2016-2020. We selected 384 well distributed earthquakes with ML ≥ 2.5 for our study and developed a purely data-driven pre-inversion pick selection method to consistently remove outliers from the automatic pick catalog. This allows us to include observations throughout the crustal triplication zone resulting in 39,599 P and 13,188 S observations. Using the established VELEST and the recently developed McMC codes we invert for the 1D P- and S-wave velocity structure including station correction terms while simultaneously relocating the events. As a result we present two separate models differing in the maximum included observation distance and therefore their suggested usage. The model AlpsLocPS is based on arrivals from ≤ 130 km and therefore should be used to consistently (re)-locate seismicity based on P & S observations. The model GAR1D_PS includes the entire observable distance range of up to 1000 km and for the first time provides consistent P- & S-phase synthetic travel times for the entire Alpine orogen. Comparing our relocated seismicity with hypocentral parameters from other studies in the area we quantify the absolute horizontal and vertical accuracy of event locations as ≈ 2.0 km and ≈ 6.0 km, respectively.