Léonard Seydoux

We invert horizontal S-wave Fourier acceleration spectra between 0.5 and 25 Hz from 7,637 earthquakes in the 2010 Mw 8.8 Maule aftershock sequence to estimate source, path, and site terms in south-central Chile using a nonparametric generalized inversion technique. The dataset contains 95,515 event-station records from 117 stations during the first 10 months after the mainshock. To account for propagation heterogeneity, we divide crustal and slab-related earthquakes separately into spatially coherent clusters and invert each one independently. We use 100 randomized clustering iterations to evaluate the variance of the recovered terms. Relative site responses remain stable across these iterations, with site amplification above a factor of 10 at 2–4 Hz at some stations. Effective attenuation varies markedly across the rupture area, with Q0 values of 36–159 and frequency exponents of 0.35–1.4. Crustal clusters generally have lower Q0 and stronger frequency dependence than slab-related ones. We fit the recovered source spectra with an ω² model and obtain well-resolved source parameters for 6,279 events. Stress-drop distributions overlap between crustal and slab-related earthquakes, with median values of 2.93 and 2.76 MPa, respectively. For Mw ≥ 3, corner frequency scales approximately as M0^(−1/3), broadly consistent with self-similarity. Stress drops are systematically higher in a slab-related seismicity band between 42.5 and 52.5 km depth, with a median of 6.3 MPa. These results show that accounting for spatially variable propagation is important for interpreting source variability. The empirical model provides a basis for regional ground-motion characterization and for tracking postseismic fault evolution across the Maule rupture zone.

Volcanic eruptions are driven by magma supplied from the mantle, yet deep magmatic flux remains poorly constrained. Here we propose to quantify it based on deep seismic tremors. We adopt a model of tremor generation by rapid growth of gas bubbles in response to magma decompression and establish a physical relationship between the cumulative seismic moment and the volume of the ascending magma. By applying this approach to the Klyuchevskoy Volcanic Group in Kamchatka, we estimate the deep magma volumetric supply rate of the order of 1 km3 per year which is con-sistent with the very high eruptive activity of these volcanoes. The proposed framework provides a new way to probe magma transport at depth and improves constraints on future volcanic activity and associated hazards.

We invert horizontal 𝑺-wave Fourier acceleration spectra between 0.5 and 25 Hz from 7,452 earthquakes in the 2010 𝑴𝐰 8.8 Maule aftershock sequence to estimate source, path, and site terms in south-central Chile using a nonparametric generalized inversion technique. The dataset contains 101,566 event-station records from 2,257 crustal and 5,195 slab-related earthquakes recorded by 114 stations during the first 10 months after the mainshock. To account for propagation heterogeneity, we perform separate inversions for spatially coherent event clusters within both seismotectonic regimes, assign cluster-specific shear-wave velocity from a 3D velocity model, and use 100 randomized clustering iterations to evaluate the variance of the recovered terms. We find that relative site responses remain stable between inversion strategies. At some stations, relative amplification approaches a factor of 10 at 2–3 Hz. Effective attenuation varies markedly across the rupture area, with 𝑸𝟎 values of approximately 30–160 and frequency exponents of 0.4–1.5. Crustal clusters generally have lower 𝑸𝟎 and stronger frequency dependence than slab-related clusters. We fit the recovered source spectra with an 𝝎𝟐 model. Stress-drop distributions overlap between crustal and slab-related earthquakes, with median values of 0.95 and 0.78 MPa, respectively. Source scaling is close to self-similar above approximately 𝑴𝐰 3.5, where the weak dependence of stress drop on seismic moment becomes nearly flat. A slab-related seismicity band between 42.5 and 52.5 km depth contains systematically higher stress-drop values, with a median value of 1.75 MPa. These results show that resolving spatially variable propagation is necessary for interpreting source variability. The resulting empirical model provides a basis for regional ground-motion characterization and for tracking postseismic fault evolution across the Maule rupture zone.

We analyze four months of continuous seismic data recorded at two stations in the area of the 2019 Ridgecrest earthquake sequence. We identify seismic signal patterns using hierarchical clustering applied to deep scattering coefficients computed from continuous seismograms. To address the strong class imbalance between seismic events and noise, we propose three complementary refinement strategies. Two strategies operate at the data or feature level by reducing the influence of anthropogenic noise, whereas the third defines the latent-space geometry using nighttime data and applies it to the full dataset. Our analysis identifies several episodes of localized swarm-like activity preceding the Mw 6.4 foreshock on 4 July 2019. Comparison with the reference earthquake catalog shows no corresponding increase in cataloged seismicity, suggesting that the detected signals correspond to weak, localized, and possibly off-fault seismic activity. These results highlight the potential of unsupervised single-station analyses as an efficient complement to network-derived earthquake catalogs.

Recent volcanic activity in Iceland has attracted significant attention, threatening populations on the Reykjanes Peninsula and necessitating enhanced monitoring efforts. In this context, the use of Distributed Acoustic Sensing (DAS) in Iceland has significantly increased.In this work, we focus our attention on a DAS dataset collected during the 2021 Geldingadalir eruption, which took place in the Fagradalsfjall volcanic system. Our goal is to explore how DAS can be integrated into a seismic network to detect and locate volcanic tremor. To do so, we analyze the spatial coherence of both DAS and seismic stations records using CovSeisNet (Seydoux et al. 2016, Soubestre et al. 2019).First, we investigate the integration of DAS data into a seismic network for tremor source location. To this end, we quantify how differences between instruments and their associated measured physical quantities may affect the recorded phase. Specifically, when comparing a broadband seismometer measuring velocity with a DAS system measuring strain rate, the spatial derivative along the fiber axis induces a phase lag of pi/2 between the two sensors. We find that this is not the primary source of uncertainty in our current network configuration. We also observe that combining a dense DAS array (8 channels) located on one side of the volcano with only a few stations surrounding it results in a suboptimal network geometry. To address this, we explore scaling schemes designed to balance the relative contributions of each sensor and/or sensor pair. These approaches prove highly effective in reducing location uncertainty.Across the experiment, we observe an increase in spatial coherence at low frequency (0.1–1 Hz) that could have been interpreted as a signal. However, we could not find it in the seismic network and after changing the interrogator parameter it disappeared. We show that this band of more coherent energy can be explained by an interplay between array geometry artifacts in the coherence calculation (array aperture vs. wavelength) and the DAS self-noise.Overall, our work demonstrates that DAS technology can be effectively integrated into seismic networks, and provides approaches to manage the high measurement density inherent to DAS while distinguishing coherency caused by the interrogator.

The transition from stable to unstable states in geological systems, such as landslides and fault zones, remains poorly understood. Seismic precursors and foreshocks related to the transition, if present, are often difficult to observe, and their interpretation remains challenging. Here, we report observations consistent with a nucleation process preceding the glacier collapse on 28 May 2025 in the village of Blatten, Switzerland. We identify three main groups of events using an unsupervised machine learning approach applied to 20 days of continuous seismic data recorded before the main event. We separate rockfalls from the seismic signatures associated with sliding‐related processes in the glacier. The observations are consistent with slip‐weakening behavior, with seismic activity accelerating during the final 2 days before failure. These results highlight the potential of unsupervised learning to identify such seismic precursors prior to collapse.

We re‐examine the aftershock sequence of the M _w 8.8 Maule earthquake in south‐central Chile to understand how seismicity, magnitude‐frequency distribution, and fault structure vary along the rupture zone. Using the International Maule Aftershock Deployment (IMAD) data set, we analyze 10 months of continuous data from approximately 156 temporary stations and build a high‐resolution aftershock catalog for the Maule rupture zone. We apply the BackProjection and Matched‐Filtering (BPMF) workflow, which integrates a deep‐learning phase picker with backprojection‐based association, relative relocation, and template matching. We initially detect and relocate 130,575 earthquakes, then use a subset of high‐quality events as templates to identify smaller earthquakes missed by initial detection. The final catalog contains 537,387 earthquakes, nearly 13 times more events than in previous studies, with a completeness magnitude of ≈ M _w 1.8 and magnitudes ranging from M _w 0.2 to M _w 6.2. A local magnitude () calibration provides a homogeneous magnitude scale across the network. The dense catalog reveals detailed seismotectonic features along the rupture. In the Pichilemu region, aftershocks delineate a shallow normal fault system with L‐shaped geometry, whereas the Concepción area contains aseismic patches. Using the classical maximum likelihood and ‐more‐incomplete methods, we find that temporal ‐values range between 1.2 and 1.6 early in the sequence and converge toward ≈1.0. Meanwhile, ‐values vary strongly along strike, with higher values in the north and lower values in the south. These contrasts are consistent with along‐strike variations in effective stress and pore fluid pressure on the plate interface, in line with previous studies.

Iceberg calving and near-shore landslides in Greenland produces seiches in the fjords - standing tsunami-like waves on the order of minutes of period that resonate for hours (or even days in extreme cases, see Svennevig et al., 2024), and can pose danger to the population and cause damage to infrastructure in the villages. For more than a decade, it has been known that broadband seismic sensors on-shore are sensitive to the ground tilt induced by these waves (Amundsen et al., 2012). Seiches are typically seen in seismic data as very long period waves that last tens of minutes to hours. This means that a network of on-shore seismic sensors can be employed to provide a tsunami early warning (TEW) system for both on-site and network-wide TEW. A major advantage of seismic networks over pressure gauges is the fact that the sensors are not exposed to the destructive forces of sea ice and icebergs, which are abundant in many regions of Greenland. In this work we demonstrate how a seismic network in the Uummannaq fjord (Greenland) can be used to provide TEW to the villages in the fjord. We further demonstrate an algorithm that allows detecting seiches and discriminating them from other sources of long-period signals (mostly large teleseismic earthquakes). Such an algorithm, however, can be significantly affected by corrupted data (spikes, steps, etc.), which produce false alarms. We then demonstrate how we can remove these false triggers using a deep scattering network (Seydoux et al., 2020). Our results show that we can detect seiches with little false alarms and provide timely TEW in the Uummannaq fjord, including the 2017 Nuugatsiaq landslide. We also demonstrate our implementation of the developed TEW algorithm into a real-time system in SeisComP.

The ambient seismic field comprises waves generated by a wide range of tectonic, environmental, and anthropogenic processes, and they also encode information about the subsurface medium through which the waves propagate. Over the past two decades, ambient-field-based techniques have emerged as a powerful approach for monitoring temporal changes in the Earth’s subsurface properties. These methods exploit the statistical characteristics of continuous seismic records to detect subtle perturbations in the medium without relying on earthquake sources. Recently, Steinmann et al. (2022) proposed an alternative approach for monitoring the freezing of near-surface material. This approach is based on a statistical blind source separation and an unsupervised machine learning framework applied to continuous seismic data. We apply this method to groundwater monitoring in California using single-station seismic recording. The approach aims to disentangle overlapping seismic signatures from different sources and physical processes to isolate components related to hydrological variations. We will evaluate the performance and robustness of this method and discuss its potential for improved monitoring of groundwater-driven changes in subsurface seismic properties.

Ground motion records combine source, path, and site effects, and isolating them remains difficult, especially for small earthquakes. We apply a non-parametric generalized inversion technique (GIT) of S-wave spectra to the 2010 MW 8.8 Maule aftershock sequence in south-central Chile. The dataset includes about 7,000 events with ML 2.0–6.5 recorded over approximately ten months. To capture spatial variability across the broad rupture, we perform the inversion in local clusters of ~400 events. This strategy preserves lateral and depth heterogeneity and reduces bias from region-wide simplifications in the path and site terms. From the inverted source spectra we estimate seismic moment, corner frequency, stress drop, source kappa, and evaluate depth dependence and self-similarity. Preliminary results indicate an average stress drop of ~0.85 MPa, with weak depth dependence but higher values for larger events, suggesting a scaling with seismic moment. The mean source kappa is about 0.019 s. Path terms provide a frequency-dependent attenuation factor Q(f), while site terms yield frequency-dependent amplification functions that we compare with horizontal-to-vertical (H/V) spectral ratios. We invert clusters independently, then merge the recovered source, path, and site terms into a single region-wide ensemble to verify consistency across cluster boundaries.