Deep learning models have been widely applied in seismological signal processing, including phase picking, signal denoising, polarity determination, and phase association. But when it comes to the tasks to process phase information across a station network, most current workflows rely on either traditional physics informed optimization (e.g., hypocenter and grid search for source locations and focal mechanisms), or machine leaning models pre-trained on fixed station arrays, which limits the generality of the model to specific areas. We propose that by making proper use of transformer encoders (self-attention layers), in which we treat every station’s phase information as a “token vector” and append with the station locations/metadata as “positional encodings”, we can train a model capable of processing phase information from any general set of station networks.To demonstrate this idea, we built a transformer-based focal-mechanism determination model, named FOCONET, which directly solves for the strike, dip, and rake angles of a double couple focal mechanism, based on the locations, first-motion polarities, S/P amplitude ratios, and SNRs from a set of stations. FOCONET is trained on 440,000 noise-added synthetics generated with random source locations, focal mechanisms, and station distributions. We use the Kagan angle - the minimum rotation angle between the predicted and the (known) ground truth mechanism - to evaluate the prediction quality. Tested on the noised synthetics with known focal mechanisms, FOCONET reaches average Kagan angles of 29°, 16° and 12° when using data from 12, 24 or 32 stations. This is well within typical focal mechanism errors (25°–30°), and 3­º–10° lower than the predictions from traditional methods including: FPFIT and HASH (S/P included or excluded). We also tested our FOCONET on 200+ M>2.5 events of the 2016 Amatrice, Italy earthquake sequence with A-class HASH solutions for comparison, and achieved an average Kagan angle of 20º. Given its stronger performance on the synthetic test data, it is plausible that the FOCONET predictions may in fact be closer to the unknown ground truths. The success of FOCONET in focal mechanism determination from a network of stations suggests that similar joint-station seismological task would benefit from transformer-based models.

The use of seismic catalogs enhanced through advanced detection techniques improves the understanding of earthquake processes by illuminating the geometry and mechanics of fault systems. In this study, we performed accurate hypocentral locations, source parameters estimation and stress release modelling from deep catalogs of microseismic sequences nucleating in the complex normal fault system of the Southern Apennines (Italy). The application of advanced location techniques resulted in the relocation of ~ 30% of the earthquakes in the enhanced catalogs, with relocated hypocenters clearly identifying local patches on kilometer-scale structures that feature consistent orientation with the main faults of the area. When mapping the stress change on the fault plane, the inter-event distance compared to the size of the events suggests that the dominant triggering mechanism within the sequences is static stress transfer. The distribution of events is not isotropic but dominantly aligned along the dip direction. These slip-dominated lineations could be associated with striations related to fault roughness and could map the boundary between locked and creeping domains in Apulian platform and basement.

Understanding mechanical processes occurring on faults requires detailed information on the microseismicity that can be enhanced today by advanced techniques for earthquake detection. This problem is challenging when the seismicity rate is low and most of the earthquakes occur at depth. In this study, we compare three detection techniques, the autocorrelation FAST, the machine learning EQTransformer, and the template matching EQCorrScan, to assess their ability to improve catalogs associated with seismic sequences in the normal fault system of Southern Apennines (Italy) using data from the Irpinia Near Fault Observatory (INFO). We found that the integration of the machine learning and template matching detectors, the former providing templates for the cross-correlation, largely outperforms techniques based on autocorrelation and machine learning alone, featuring an enrichment of the automatic and manual catalogs of factors 21 and 7 respectively. Since output catalogs can be polluted by many false positives, we applied refined event selection based on the cumulative distribution of their similarity level. We can thus clean up the detection lists and analyze final subsets dominated by real events. The magnitude of completeness decreases by more than one unit compared to the reference value for the network. We report b-values associated with sequences smaller than the average, likely corresponding to larger differential stresses than for the background seismicity of the area. For all the analyzed sequences, we found that main events are anticipated by foreshocks, indicating a possible preparation process for mainshocks at sub-kilometric scales.

We develop a new approach based on the variance of noise cross-correlation to characterize the noise source and wave propagation under the influence of a non-diffuse noise field. Based on the random errors for the Fourier spectra of stacked noise cross-correlation (Liu et al. 2016) and the assumption of diffuse field, we derive an analytical expression for the variance of every time point in the stacked cross-correlation and validate the theory using synthetic diffuse noise data. The ambient seismic noise field in Southern California is, however, not fully diffuse. The observed correlated neighboring frequencies in the noise data – a definitive character of the non-diffuse field (Liu & Ben-Zion 2016), map into the noise cross-correlation, biasing its variances from theoretical predictions under fully diffuse field assumption. For the secondary ocean microseism, we find strong positive correlation between the correlated neighboring frequencies and the deviations from diffuse field theory predicted variances. The ballistic arrivals on average contain more significant bias than the coda segments of the noise cross-correlation, which agrees with previous time-lapse monitoring studies based on ambient seismic noise. In addition, the station pairs having significant bias between actual and diffuse field theory predicted variances are aligned with the strongest source direction and exhibit significant beamforming source fluctuation.

Small stress changes such as those from tidal loading can be enough to trigger earthquakes. If small and large earthquakes initiate similarly, high resolution catalogs with low detection thresholds are best suited to illuminate such processes. Below the Sea of Marmara section of the North Anatolian Fault, a segment of 150 km is late in its seismic cycle. We generated high-resolution seismicity catalogs for a hydrothermal region in the eastern Sea of Marmara employing both AI-based and template matching techniques to investigate a complex long-lasting sequence including seismicity up to MW 4.5. We document a strong effect of the Sea of Marmara level changes on the local seismicity. Both high resolution catalogs show that local seismicity rates are significantly larger during time periods shortly after local minima on sea level. Local strainmeters indicate that the associated strain changes, on the order of 30-300 nstrain, are sufficient to promote seismicity.

The successful application of deep learning for seismic phase arrival time picking has increased the efficacy of earthquake catalog development workflows. Earthquake catalogs with lower magnitude of completeness and better locational precision than current standard practice can now be generated with very limited need for human review and without the need for earthquake templates, which are not always available. Here, we report on a ‘Deep Earthquake Catalog’ with over 300,000 events from a geographically extensive region spanning Oklahoma and Southern Kansas from January 2010 to December 2020 developed using a workflow that leverages deep learning for phase picking. The increased number of events and improved spatial resolution compared to the previous statewide catalogs reveals numerous discrete faults and both broad trends and localized patterns of seismicity. This rich dataset provides new opportunities for data-driven analyses of induced earthquakes.

The important task of tracking seismic activity requires both sensitive detection and accurate earthquake location. Approximate earthquake locations can be estimated promptly and automatically; however, accurate locations depend on precise seismic phase picking, which is a laborious and time-consuming task. We adapted a deep neural network (DNN) phase picker trained on local seismic data to meso-scale hydraulic fracturing experiments. We designed a novel workflow, transfer-learning aided double-difference tomography, to overcome the three orders of magnitude difference in both spatial and temporal scales between our data and data used to train the original DNN. Only 3,500 seismograms (0.45% of the original DNN data) were needed to re-train the original DNN model successfully. The phase picks obtained with transfer-learned model are at least as accurate as the analyst’s, and lead to improved event locations. Moreover, the effort required for picking once the DNN is trained is a small fraction of the analyst’s.

▪ Our attenuation inversion method is based on Liu et al. (2015) using linear triplet of stations ▪ Application to the dense 3C linear array crossing the San Jacinto Fault in Ramona, CA ▪ Clear fundamental mode Rayleigh and Love wave phases are extracted ▪ Attenuation tomography maps reveal new information in addition to velocity tomography ▪ Shear velocity radial anisotropy and possible shear attenuation anisotropy

The 17 May 2012 M4.8 Timpson earthquake is the largest known earthquake in eastern Texas. It is suspected to have been induced by wastewater injection from two nearby, high-volume wells. Its cataloged aftershocks form a NW-SE trend, which unlike other induced earthquakes sequences, is unfavorably oriented for failure in the local stress field. To understand this, we enriched the catalog using PhaseNet, a deep-learning-based picker followed by double-difference relocation with cross-correlation-based differential traveltimes. We clustered the aftershocks based on waveform similarity. Most of the seismicity falls into two-clusters, which define a complex fault structure of two parallel subfaults that are more favorably oriented than the overall trend. We inferred from waveform similarity that the sequence initiated on the northern subfault with a M3.9 foreshock and M4.8 mainshock, then extended to the southern subfault with a M4.1 aftershock, and was finally reactivated on the northern subfault with two more M4 events.

Cross-correlation of fully diffuse wavefields averaged over time should converge to the Green’s function; however, the ambient seismic field in the real Earth is not fully diffuse, which interferes with that convergence. We apply blind signal separation to reduce the effect of spurious non-diffuse components on the cross-correlation tensor of the ambient seismic field. We describe the diffuse component as having uncorrelated neighboring frequencies and equal intensity at all azimuths, and an independent (i.e., statistically uncorrelated) non-diffuse component arising from a spatially isolated point source for which neighboring frequencies are correlated. Under the assumption of linear independence of the spurious non-diffuse wave outside the stationary phase zone and the constructive interference of noise waves within that zone, we can suppress the spurious non-diffuse component from the noise interferometry. Our numerical simulations show good separation of one spurious non-diffuse noise source component for either non-diffuse Rayleigh or Love waves. We apply this separation to the Rayleigh-wave component of the Green’s function for 136 cross-correlation pairs from 17 stations in Southern California. We perform beamforming over different frequency bands for the cross-correlations before and after the separation, and find that the reconstructed Rayleigh waves are more coherent. We also estimate the bias in Rayleigh wave phase velocity for each receiver pair due to the spurious non-diffuse contribution.

We revisited the June, 2010 - October, 2011 Guy-Greenbrier earthquake sequence in central Arkansas using PhaseNet, a deep neural network trained to pick P and S arrival times. We applied PhaseNet to continuous waveform data and used phase association and hypocenter relocation to locate nearly 90,000 events. Our catalog suggests that the sequence consists of two adjacent earthquake sequences on the same fault and that the second sequence may be associated with the wastewater disposal well to the west of the Guy-Greenbrier Fault, rather than the wells to the north and the east that were previously implicated. We find that each sequence is comprised of many small clusters that exhibit diffusion along the fault at shorter time scales. Our study demonstrates that machine learning based earthquake catalog development is now feasible and will yield new insights into earthquake behavior.