SM2.2 | Machine learning for time series in geophysics
Orals |
Thu, 16:15
Fri, 10:45
Mon, 14:00
EDI
Machine learning for time series in geophysics
Co-organized by ESSI1/NP4
Convener: Jannes MünchmeyerECSECS | Co-conveners: Josefine UmlauftECSECS, Rene Steinmann, Léonard Seydoux, Fabio Corbi
Orals
| Thu, 01 May, 16:15–17:55 (CEST)
 
Room D1
Posters on site
| Attendance Fri, 02 May, 10:45–12:30 (CEST) | Display Fri, 02 May, 08:30–12:30
 
Hall X1
Posters virtual
| Attendance Mon, 28 Apr, 14:00–15:45 (CEST) | Display Mon, 28 Apr, 08:30–18:00
 
vPoster spot 1
Orals |
Thu, 16:15
Fri, 10:45
Mon, 14:00

Orals: Thu, 1 May | Room D1

The oral presentations are given in a hybrid format supported by a Zoom meeting featuring on-site and virtual presentations. The button to access the Zoom meeting appears just before the time block starts.
16:15–16:25
|
EGU25-9276
|
solicited
|
On-site presentation
Clément Hibert, Joachim Rimpot, Camille Huynh, Charlotte Groult, Jean-Philippe Malet, Germain Forestier, Jonathan Weber, Camille Jestin, Vincent Lanticq, Floriane Provost, Antoine Turquet, and Tord Stangeland

Seismology, beyond the study of earthquakes, has become an indispensable tool for understanding environmental changes, offering unique insights into a wide range of phenomena and natural risks, from slope instabilities to glacial dynamics and hydrological hazards. However, the sheer volume and complexity of modern seismic datasets, amplified by the emergence of dense seismic networks and technologies such as Distributed Acoustic Sensing (DAS), pose significant challenges. Recent advances in Artificial Intelligence (AI) and Machine Learning (ML) have revolutionized our ability to analyze these datasets, enabling a deeper exploration of seismic data to find rare and exotic environmental seismic sources. 

Supervised learning approaches have been successfully applied to create large-scale instrumental catalogs of landslides and other environmental processes, at different  spatio-temporal scales, from short-term datasets recorded on dense local seismic stations networks, to chronicles spanning decades on seismic networks covering whole regions of the world (Alaska, Alps, Greenland). These techniques achieve high detection rates and robust classification of seismic events, even for low-magnitude or rare signals that traditional methods might overlook. Supervised learning approaches also allow us to advance our capability to estimate physical properties from seismic waves, such as the use of machine  learning to infer mass and kinematics of slope instabilities, which provide critical inputs for understanding the dynamics of these events and their associated hazards. These methodologies not only allow us to document environmental processes more exhaustively but also open up possibilities for studying poorly understood or previously undetectable seismic sources. Going beyond supervised learning, we have developed workflows based on self-supervised and unsupervised approaches to analyze continuous seismic data, uncovering unexpected patterns and revealing hidden environmental seismic sources recorded by dense seismic stations networks. Distributed Acoustic Sensing represents another frontier, turning fiber optic cables into dense seismic networks. By combining DAS with innovative AI-driven methods, we have demonstrated the potential to detect and classify low-magnitude earthquakes and anthropogenic sources, even in noisy environments, paving the way for real-time seismic monitoring on unprecedented scales.

By applying these AI-driven approaches, we are enhancing the field of environmental and exotic sources seismology, improving our ability to analyze vast seismic archives, and offering new tools to monitor, understand, and mitigate geohazards in a changing environment. This talk will highlight the latest methodological advances and showcase how they are applied to various geological and environmental contexts, from landslides, avalanches and glaciers in the Alps to fiber optic networks at different scales, underscoring the far-reaching implications of AI for seismological sources identification.

How to cite: Hibert, C., Rimpot, J., Huynh, C., Groult, C., Malet, J.-P., Forestier, G., Weber, J., Jestin, C., Lanticq, V., Provost, F., Turquet, A., and Stangeland, T.: Uncovering environmental and other exotic seismic sources with machine learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9276, https://doi.org/10.5194/egusphere-egu25-9276, 2025.

16:25–16:35
|
EGU25-11335
|
ECS
|
On-site presentation
Zahra Zali, Patricia Martínez-Garzón, Grzegorz Kwiatek, Gregory Beroza, Fabrice Cotton, and Marco Bohnhoff

We developed a deep learning model for automatic dimensionality reduction and feature extraction from time series. The model employs an encoder-decoder architecture with skip connections, enabling efficient compression and reconstruction of input data while preserving essential features. These features are used for unsupervised clustering enabling anomaly detection, and pattern recognition.

We initially developed the model to analyze seismic data from the 2021 Geldingadalir volcanic eruption in Iceland, successfully identifying a weak yet important pre-eruptive tremor that commenced three days before the eruption. Advancing the architecture with additional layers and skip connections allowed for highly accurate input reconstruction. The latter version, named AutoencoderZ, demonstrated its ability to process different data types. We applied AutoencoderZ to investigate low-frequency patterns preceding the 2023 MW 7.8 Kahramanmaraş Earthquake in Türkiye. The model identified tremor-like episodes linked to anthropogenic activities at cement plants near the earthquake’s epicenter. Additionally, we applied AutoencoderZ to strainmeter data from the Sea of Marmara, achieving accurate reconstructions and enabling the detection of distinct tectonic-related signals.

This study highlights AutoencoderZ’s potential as a powerful tool for uncovering patterns in continuous geophysical data, providing valuable insights for monitoring and interpreting seismic and strainmeter signals.

How to cite: Zali, Z., Martínez-Garzón, P., Kwiatek, G., Beroza, G., Cotton, F., and Bohnhoff, M.: Unsupervised Clustering and Pattern Identification from Continuous Seismic and Strainmeter Data in Tectonic and Volcanic Settings, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11335, https://doi.org/10.5194/egusphere-egu25-11335, 2025.

16:35–16:45
|
EGU25-2833
|
ECS
|
On-site presentation
Adèle Doucet, Léonard Seydoux, Nobuaki Fuji, Yosuke Aoki, and Jean-Philippe Métaxian

Mount Fuji volcano, located 100~km away from Tokyo, directly threatens over 30~million people. It last erupted in 1707, and has remained dormant since then. Seismicity --and particularly Low-Frequency Earthquakes (LFE)-- is to now the primary indicator of processes occurring beneath the volcano and is usually linked to fluid movement. Yet, these signals are usually manually picked and classified as such, without the ability to formally define them for automatic detection systems. 

Our goal is to develop an automatic method to detect and classify LFEs, among other seismic events at Mount Fuji using the continuous seismic records from 2008 at 11 stations. First, we use the CovSeisNet software to detect events by analyzing the wavefield coherence, derived from the network covariance matrix width. Over one year of continuous data, the wavefield coherence shows distinct patterns that correspond to various event types, including LFEs and tectonic earthquakes. To enable interpretation, we apply a manifold learning algorithm (UMAP) to reduce the dimension of the coherence patterns into two dimensions to ease the interpretation. We name this low-dimensional representation a "coherence atlas" where each point represents a time window of seismic data, grouped by similarity. This automatic approach enables not only the detection but also the classification of seismic events, as compared with the Japan Meteorological Agency catalog. Moreover, the atlas helps identify previously unrecorded events and facilitates the definition of new event classes. By autonomously mapping and classifying seismic activity beneath Mount Fuji, this method offers unprecedented insights into its activity and allows us to detect new events that had been hidden in the manually prepared catalog.

How to cite: Doucet, A., Seydoux, L., Fuji, N., Aoki, Y., and Métaxian, J.-P.: Unsupervised exploration of seismic activity at Mount Fuji, Japan, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-2833, https://doi.org/10.5194/egusphere-egu25-2833, 2025.

16:45–16:55
|
EGU25-453
|
ECS
|
On-site presentation
Jana Klinge, Sven Schippkus, and Céline Hadziioannou

Seismic arrays are essential for collecting and analyzing seismic data, significantly enhancing our understanding of geophysical processes such as the localization of seismic sources. We introduce the concept of Virtual Seismic Arrays, where the array recordings are predicted from a single reference station, removing the need for continuous deployment of all array stations. This work builds on the research by Klinge et al. (2025), which introduced a Deep Learning approach using encoder-decoder networks to learn and predict transfer properties between two seismic stations. By training the algorithm on data of the Gräfenberg array in the secondary microseism frequency band, we develop models that effectively capture the transfer characteristics between a chosen reference station and each of the other stations within the array. To evaluate how well the models represent the underlying wave propagation, we use beamforming and apply it to both the original data from all stations and the corresponding predictions generated by the models. We assess two scenarios: one where the dominant backazimuths and slownesses are consistent with the training dataset, and another where the models are applied to data from different conditions. Our results show strong agreement between the predicted and original beamforming results, demonstrating the potential of Virtual Seismic Arrays for future application.

How to cite: Klinge, J., Schippkus, S., and Hadziioannou, C.: Enhancing seismic monitoring with Virtual Seismic Arrays: Application of a deep learning framework, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-453, https://doi.org/10.5194/egusphere-egu25-453, 2025.

16:55–17:05
|
EGU25-21456
|
On-site presentation
Alexander Bauer, Jan Walda, and Conny Hammer

Single-station waveforms of teleseismic earthquakes are highly complex, because they are a superposition of numerous phases corresponding to different wave types and propagation paths. Moreover, data recorded at single stations is contaminated by noise, which often has similar or larger amplitudes than the arrivals of teleseismic earthquakes, especially in densely-populated areas. For high precision research facilities, for example in the field of particle physics or gravity wave detection, a precise knowledge of the seismic wavefield generated by teleseismic earthquakes can be critical for the calibration of experiments. However, the density of seismological stations is often sparse, particularly in regions with low seismic hazard such as Northern Germany.

To overcome this limitation, we introduce a deep learning scheme for the prediction of very-low-frequency earthquake waveforms from synthetic data at arbitrary locations within the metropolitan area of Hamburg, Germany. For this aim, we propose to train a convolutional neural network (CNN) to predict the measured earthquake waveforms from their synthetic counterparts. While synthetic earthquake waveforms can be conveniently generated for arbitrary coordinates and moment tensors with Instaseis and the IRIS synthetics engine (Syngine), the amount of available measured waveforms is constrained by the availability of seismological stations and their installation date. In this work, we use measured data from a station in Bad Segeberg, north of Hamburg, which has been measuring continuously since 1996. During first experiments, we trained a CNN on data from earthquakes larger than M6.0 and obtained reasonable initial results. However, the number of such earthquakes is limited and the measured waveforms used as labels partly contained noise of considerable amplitude, which caused the neural network to predict unwanted noise.

In order to increase the amount of earthquakes in the training data and mitigate their contamination with noise, we propose a two-step approach. In the first step, we generate a large number of noise-free synthetic waveforms and contaminate them with artificially generated noise that has the same characteristics as the noise measured at the station in Bad Segeberg. With this dataset, we train a first CNN to denoise the synthetic earthquake waveforms. In the second step, we apply the trained neural network to the actual earthquake waveforms measured in Bad Segeberg to denoise them. We then train a second CNN to translate synthetic earthquake waveforms to the denoised measured ones. Results for earthquakes not part of the training data demonstrate that the second CNN provides convincing estimates of measured earthquake waveforms, not only for the station in Bad Segeberg, but also for stations in Hamburg. This can be seen as a first step towards a three-dimensional prediction of the earthquake wavefield without the need for densely-distributed stations.

How to cite: Bauer, A., Walda, J., and Hammer, C.: Prediction of measured earthquake waveforms from synthetic data: a two-step deep learning approach, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-21456, https://doi.org/10.5194/egusphere-egu25-21456, 2025.

17:05–17:15
|
EGU25-19236
|
ECS
|
On-site presentation
Kadek Hendrawan Palgunadi, Andreas Bergmeister, Andrea Bosisio, Laura Ermert, Maria Koroni, Natanaël Perraudin, Simon Dirmeier, and Men-Andrin Meier

Reliable synthesis and prediction of seismic waveforms play an important role in evaluating seismic hazards and designing earthquake-resilient structures. However, current methods, such as ground motion models and physics-based simulations, are often limited in fully capturing the complexity of seismic wave propagation, at higher frequencies (>5 Hz). Some of these limitations can potentially be overcome through machine learning techniques. In earthquake engineering, machine learning models have been used for predicting peak ground accelerations and Fourier spectra responses. To model entire waveforms, extensive efforts to generate seismic waveforms have employed advanced machine learning techniques, such as generative models, with most previous approaches relying on generative adversarial networks (GANs). In contrast to these earlier models, this study presents an efficient and extensible generative framework to produce realistic high-frequency seismic waveforms, compared to GANs. Our approach encodes spectrograms of the waveform data into a lower-dimensional sub-manifold using an autoencoder, and a state-of-the-art diffusion model is subsequently trained to generate these latent embeddings. Conditioning is currently performed on key parameters: earthquake magnitude, recording distance, site conditions, and faulting style. The resulting generative model can synthesize waveforms with frequency content up to 50 Hz, from which several scalar ground motion statistics, such as peak ground motion amplitudes, spectral accelerations, or Arias intensity can be directly derived. We validate the quality of the generated waveforms using standard seismological benchmarks and performance metrics from image generation research. Our openly available model produces high-frequency waveforms that align with real data across a wide range of input parameters, including regions where observations are sparse, and accurately reproduces both median trends and variability of empirical ground motion statistics. Our generative waveform model can be potentially used to perform seismic hazard where broadband data are often required such as to train earthquake early warning model. Given the increasing number of generative waveform models, we emphasize that they should be openly accessible and included in community efforts for ground motion model evaluations.

How to cite: Palgunadi, K. H., Bergmeister, A., Bosisio, A., Ermert, L., Koroni, M., Perraudin, N., Dirmeier, S., and Meier, M.-A.: High Resolution Generative Waveform Modeling Using Denoising Diffusion, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19236, https://doi.org/10.5194/egusphere-egu25-19236, 2025.

17:15–17:25
|
EGU25-17400
|
ECS
|
On-site presentation
Gabriele Paoletti, Daniele Trappolini, Elisa Tinti, Fabio Galasso, Cristiano Collettini, and Chris Marone

Fault zone properties evolve dynamically during the seismic cycle due to stress changes, microcracking, and wall rock damage. Understanding these changes is vital to gaining insights into earthquake preparation and post-seismic processes. The latter include fault healing, which refers to the recovery of mechanical and elastic properties in fault zones after seismic and aseismic fault slip. Despite its importance, detecting and characterizing fault healing through seismic signals remains a challenge due to the subtle nature of these changes.

In this study, we investigate the potential of deep learning techniques, specifically a 4-layer Convolutional Neural Network (CNN), to characterize post-seismic evolution by analyzing raw seismic waveforms recorded after the largest event (Mw 6.5,  30 October) of the 2016 Central Italy seismic sequence. These data provide a unique opportunity to examine fault zone dynamics. A key aspect of our approach is the hypothesis that ray paths traversing highly impacted areas of the fault zone contain richer information about its temporal evolution. To test this hypothesis, we examined seismic waves from two clusters — DHwS, located in the hanging wall beneath the hypocentral region, and C1, situated in the footwall. They represent contrasting ray trajectories as recorded on seismic stations MC2 and MMO1. Seismic waves recorded at MC2 pass through heavily damaged fault regions, which are likely to reveal evolving fault properties, whereas MMO1 predominantly captures paths that skirt or in the case of C1 completely miss these impacted areas, serving as a comparative baseline.

We assessed temporal variations in elastic properties using binary classification tests on normalized, raw seismic waveforms of events before and after a reference date. This date was arbitrarily selected within the temporal range of the analyzed seismicity and serves solely as a neutral point of comparison. Our hypothesis is that if the CNN can achieve good classification performance, it implies the presence of time-evolving properties in the fault zone, potentially linked to healing processes or other time-dependent factors.

To further validate these findings, we employed adversarial training, a technique designed to disentangle time-dependent effects from structural changes. By introducing controlled label noise into one cluster during training, we isolated the influence of confounding factors such as seasonal variations. Preliminary results suggest that adversarial training enhances the model's robustness and provides valuable insights into the time-evolving properties of the fault zone.

Deep learning offers significant potential for analyzing spatiotemporal changes in elastic properties and thus the evolution of fault zone properties over the seismic cycle. By detecting subtle temporal and structural changes, we hope to gain a deeper understanding of fault dynamics and post-seismic processes.

How to cite: Paoletti, G., Trappolini, D., Tinti, E., Galasso, F., Collettini, C., and Marone, C.: Deep learning to investigate post-seismic evolution of fault zone elastic properties, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17400, https://doi.org/10.5194/egusphere-egu25-17400, 2025.

17:25–17:35
|
EGU25-2836
|
ECS
|
On-site presentation
Céline Hourcade, Kévin Juhel, and Quentin Bletery

State-of-the-art earthquake early warning systems use the early records of seismic waves to estimate the magnitude and location of the seismic source before the shaking and the tsunami strike. Because of the inherent properties of early seismic records, those systems systematically underestimate the magnitude of large events, which results in catastrophic underestimation of the subsequent tsunamis. Prompt elastogravity signals (PEGS) are low-amplitude, light-speed signals emitted by earthquakes, which are highly sensitive to both their magnitude and focal mechanism. Detected before traditional seismic waves, PEGS have the potential to produce unsaturated magnitude estimates faster than state-of-the-art systems. Accurate instantaneous tracking of large earthquake magnitude using PEGS has been proven possible through the use of a Convolutional Neural Network (CNN). However, the CNN architecture is sub-optimal as it does not allow to capture the geometry of the problem. To address this limitation, we design PEGSGraph, a novel deep learning model relying on a Graph Neural Network (GNN) architecture.
PEGSGraph accurately estimates the magnitude of synthetic earthquakes down to Mw~7.6-7.7 and determines their focal mechanisms (thrust, strike-slip or normal faulting) within 70 seconds of the event's onset, offering crucial information for predicting potential tsunami wave amplitudes. Our comparative analysis on Alaska and Western Canada data shows that PEGSGraph outperforms PEGSNet, providing more reliable rapid magnitude estimates and enhancing tsunami warning reliability.

How to cite: Hourcade, C., Juhel, K., and Bletery, Q.: PEGSGraph: A Graph Neural Network for Fast Earthquake Characterization Based on Prompt ElastoGravity Signals, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-2836, https://doi.org/10.5194/egusphere-egu25-2836, 2025.

17:35–17:45
|
EGU25-13238
|
ECS
|
On-site presentation
Matteo Bagagli, Francesco Grigoli, and Davide Bacciu

In this work, we introduce HEIMDALL, a grapH-based sEIsMic Detector And Locator specifically designed for microseismic applications. Building on recent progress in deep learning (DL), HEIMDALL employs spatiotemporal graph-neural networks to detect and locate seismic events in continuous waveforms. It simultaneously associates and provides preliminary locations by leveraging the output probability functions of the graph-neural network over a dense, three-dimensional grid (0.1 km spacing). By integrating detection and location within a single framework, HEIMDALL aims to address persistent challenges in microseismic data analysis, such as accurately associating wavefront arrivals and enabling consistent and robust event localization in complex geothermal regions. To train our models, we utilize synthetics generated using Green’s function available in the area, in combination with a small fine-tuning over a subset of real data. This approach allows us to achieve homogeneous coverage of the study area while addressing nuances that inevitably arise across synthetic and real domains.

Our evaluation focuses on data collected at the Hengill Geothermal Field in Iceland as part of the COSEIMIQ project (December 2018 to August 2021). Specifically, we analyzed one month of continuous seismic recordings from December 2018 and a brief sequence on February 3, 2019, which occurred in the middle of the geothermal plant. The dataset also features frequent burst sequences, providing an ideal testbed for advanced detection and location algorithms. By benchmarking HEIMDALL against multiple approaches, we reveal both the strengths and limitations inherent in our novel method and in more conventional workflows used in observational seismology.

Ultimately, we highlight the importance of continued innovation in ML-based workflows for induced seismicity monitoring at enhanced geothermal system (EGS) sites, where the capacity to detect and accurately locate a large number of microseismic events can be critical for operational safety and resource management.

How to cite: Bagagli, M., Grigoli, F., and Bacciu, D.: A Waveform-Based Graph Neural Network Approach for Microseismic Monitoring, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13238, https://doi.org/10.5194/egusphere-egu25-13238, 2025.

17:45–17:55
|
EGU25-14041
|
On-site presentation
Shinya Nakano, Ryuho Kataoka, and Masahito Nose

Substorms are a magnetospheric phenomenon which causes high geomagnetic and auroral disturbances. It is widely accepted that substorm activity is controlled by solar wind conditions. It is, however, difficult to predict substorms deterministically because of the complex physical processes underlying substorm occurrences. We propose a framework for modelling time series of event occurrences controlled by external forcing. In this framework, occurrences of external-driven events are represented with a non-stationary Poisson process, and its intensity, which corresponds to the occurrence rates per unit time, is described with a simple machine learning model, the echo state network, which is fed with forcing variables. The echo state network is trained by maxmising the likelihood given the event time series data. 

We apply this approach for analysing time series of substorm onsets identified from Pi2 pulsations, which are irregular geomagnetic oscillations associated with substorm onsets. We train the echo state network to well describe the response of substorm activity to solar-wind conditions. We then examine the characteristics of the substorm activity by feeding synthetic solar-wind data into the echo state network. The results show what solar wind variables effectively contribute to the substorm occurrence. 

Our echo state network model is also useful for examining the statistical properties of the substorm occurrence rate. For example, we can evaluate what mainly controls the seasonal and UT variations of substorm activity. There are two explanations for the seasonal and UT variations. One explanation is that the seasonal and UT variations is controlled by the inner product between the solar-wind magnetic field and the Earth's dipole axis. The other is that the variations are due to the angle between the solar-wind flow and the Earth's dipole axis. Since these two effects are related with different input variables in our echo state network model, we can examine the contribution of each effect to the substorm occurrence frequency. The result shows that the seasonal and UT variations are mostly dependent on the angle between the solar-wind flow and the Earth's dipole axis. 

How to cite: Nakano, S., Kataoka, R., and Nose, M.: Modelling of time series of external-driven events with echo state network and its application to substorm activity analysis, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14041, https://doi.org/10.5194/egusphere-egu25-14041, 2025.

Posters on site: Fri, 2 May, 10:45–12:30 | Hall X1

The posters scheduled for on-site presentation are only visible in the poster hall in Vienna. If authors uploaded their presentation files, these files are linked from the abstracts below.
Display time: Fri, 2 May, 08:30–12:30
X1.82
|
EGU25-739
|
ECS
Alicia Ximena Cortés Barajas, Marco Calò, Erik Molino Minero Re, and Francesca Di Luccio

Detecting seismic events is essential for monitoring tectonic and volcanic activity, especially in marine environments where noise makes analysis particularly challenging. This study introduces a method that combines Evolutionary Neural Architecture Search (ENAS) with the third generation of the Non-dominated Sorting Genetic Algorithm (NSGA-III) to design and optimize neural networks for seismic event detection using Ocean Bottom Seismometers (OBS) data.

In this work we developed a methodology to analyze heavily noisy data recorded by the TYDE OBS experiment in the southern Tyrrhenian Sea, Italy. In 2000, 14 seismic stations were deployed on the seafloor and around the Aeolian Islands recording data for about 6 months. Stations consisted of wide-band Ocean Bottom Seismometers (OBS) and Hydrophones (OBH).

The preprocessing pipeline includes feature extraction with Discrete Wavelet Transform (DWT) and dimensionality reduction using Principal Component Analysis (PCA), which reduces over 6000 coefficients to just 55 while preserving 95% of the variance. Applied to 90-second overlapping windows, this approach has achieved strong results, with F1 scores exceeding 90% in balanced noisy datasets.

Building on these results, this study explores unsupervised clustering to group similar seismic events and identify possible false positives through anomaly detection. By using adaptive clustering methods that determine the optimal number of clusters based on the data, this approach aims to enhance reliability while providing additional insights into the detected seismic events.

This work highlights the potential of automated tools to complement traditional seismic monitoring methods, balancing accuracy and model complexity while improving efficiency in noise-heavy environments.

How to cite: Cortés Barajas, A. X., Calò, M., Molino Minero Re, E., and Di Luccio, F.: Optimizing neural network architectures and using clustering to detect seismic events in noisy ocean bottom seismometer data, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-739, https://doi.org/10.5194/egusphere-egu25-739, 2025.

X1.83
|
EGU25-5384
|
ECS
Yuriko Iwasaki, Susan Schwartz, and Noah Finnegan

The mechanics of slow frictional creep in landslides remains debated and only a few detailed seismic studies have been conducted in landslide-prone areas. To illuminate basal slip processes for a slow moving landslide, we deployed a dense 80 node seismic array at Oak Ridge Earthflow in California’s Diablo Range for one to two months during the rainy seasons of 2023 and 2024, both winters where decimeters of landslide displacement occurred at Oak Ridge. Simultaneously, GNSS receivers, strain meters, and piezometers were deployed at the same site. During our deployments, various types of very small signals were recorded by the seismometers. These events were local, detected only by nearby stations sited within about 100 m of each other. The cause of these events remains unclear, whether due to shear slip at the base of the earthflow or other sources, such as water movement or animal activity. To investigate the cause of these signals, and evaluate the role of stick-slip motion and shear localization, we automatically detected the events and analyzed their spatiotemporal distribution. We used quakephase (Shi et al., 2024) to identify the phases of the very small signals. The primary challenge with automatic picking in our dataset is the long processing time due to high sampling rates. To address this issue, we applied array signal processing, covseisnet (Seydoux et al., 2016), to extract signal candidates based on the coherence of dominant frequencies across the seismic array, followed by automatic picking. This approach successfully and efficiently identified specific signals we believe are associated with earthflow motion. These signals are not continuously observed but concentrate within specific time periods. We focused on events in these time periods, utilizing scattering networks and matched-filter techniques for more detailed classification. By combining our results with other temporal data, such as pore fluid pressure, precipitation, temperature, and displacement, we will discuss the causes of these signals to better understand the mechanism of the earthflow motion.

How to cite: Iwasaki, Y., Schwartz, S., and Finnegan, N.: Classification of Small Seismic Signals Associated with the Oak Ridge Earthflow in California Using a Combination of Machine Learning and Array Signal Processing, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-5384, https://doi.org/10.5194/egusphere-egu25-5384, 2025.

X1.84
|
EGU25-5429
|
ECS
Jaesung Park, Jina Jeong, and Mbarki Sinda

Traditional well-log analysis often involves incomplete datasets, which reduces the accuracy of petrophysical assessments. This study thus introduces an innovative dual-model approach that integrates a conditional variational autoencoder (CVAE) with a long short-term memory (LSTM) to predict missing shear-slowness (DTS) data and other well-log data. Utilizing well-logs and the corresponding lithological sequence from the Volve oil field in the North Sea, the proposed model demonstrates excellent prediction capabilities when facing multiple types of missing well-logs. Our findings reveal that the CVAE-LSTM model not only enhances DTS prediction accuracy but also adapts to the inherent variability of geological formations. It outperforms traditional autoencoder and standalone LSTM models across a range of metrics, including correlation coefficients, the root mean squared error, and Kolmogorov–Smirnov statistics, validating the predictive accuracy of the proposed model and the alignment of the statistical distributions for predicted and actual logs. The robustness of the proposed model is further highlighted by its ability to maintain its high performance despite the absence of key well-log data such as compressional slowness and the neutron porosity index. This study demonstrates the effectiveness of advanced machine-learning techniques in overcoming the limitations associated with incomplete well-log datasets.

How to cite: Park, J., Jeong, J., and Sinda, M.: Deep-learning-based dual model with an iterative prediction process for the improvement of missing well-log predictions , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-5429, https://doi.org/10.5194/egusphere-egu25-5429, 2025.

X1.85
|
EGU25-6412
|
ECS
Elisa Caredda, Andrea Morelli, Maddalena Errico, Giampaolo Zerbinato, Marius Paul Isken, and Simone Cesca

Monitoring microseismicity is fundamental to advancing our understanding of fault mechanics under natural and anthropogenic influences. Recent advancements in seismological methodologies, particularly those employing deep learning techniques, have significantly improved the detection of weak earthquakes while preserving high levels of precision and reliability.

This study aims to enhance the detection and characterization of seismicity in the Val d’Agri region (Southern Italy) by implementing advanced deep learning-based methodologies, focusing on understanding the tectonic and anthropogenic influences driving seismic activity. The Val d’Agri region is a tectonically active area of considerable scientific and industrial relevance, hosting Europe’s largest onshore oil reservoir and an artificial lake. By employing state-of-the-art deep learning and full waveform earthquake detection methods we identified and located seismic events over a three-year period, achieving a twofold increase in detected events compared to the manually revised bulletin, with a recall rate of ~95%.

Spatial and temporal analyses, based on a density-based clustering approach, revealed distinct seismic clusters. The seismicity is mostly concentrated along the Monti della Maddalena fault system in the southwestern region, characterized by shallow earthquakes (5–7 km depth), while the northeastern and northwestern areas exhibit sparser and deeper activity (15–20 km depth). High-resolution event localization illuminated fault geometries and spatial distributions with high detail. Additionally, our dataset highlights a temporal correlation between seismicity rates and the filling and emptying phases of the Pertusillo artificial reservoir.

Our findings underscore the utility of automated workflows in improving seismic monitoring and fault characterization, providing critical insights into tectonic processes and reservoir-induced seismicity.

How to cite: Caredda, E., Morelli, A., Errico, M., Zerbinato, G., Isken, M. P., and Cesca, S.: Enhancing seismicity detection and characterization in the Val d’Agri region: insights into tectonic and induced processes using Deep Learning techniques, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-6412, https://doi.org/10.5194/egusphere-egu25-6412, 2025.

X1.86
|
EGU25-6730
Marco Maria Scuderi, Giulio Poggiali, Federico Pignalberi, and Giacomo Mastella

Laboratory acoustic emissions (AEs), resembling small-scale earthquakes, provide vital insights into frictional instability mechanics. Recent advancements in acoustic monitoring technology allow for the rapid collection of thousands of AE waveforms within minutes, highlighting the critical need for efficient detection and analysis methods. This study presents a deep learning model designed to automatically detect AEs in laboratory shear experiments.

Our dataset comprises approximately 30,000 manually identified AE waveforms obtained under varying experimental boundary conditions using two fault gouge materials: Min-U-Sil quartz gouge and glass beads. We modified the PhaseNet model, originally designed for detecting seismic phases in natural earthquakes, by optimizing its architecture and training process to develop AEsNet—an advanced AE detection model that consistently outperforms existing picking methods for Min-U-Sil quartz gouge and glass beads.

To assess the model's generalizability across different boundary conditions and materials, we employed transfer learning, examining performance relative to training dataset size and material diversity. Results indicate that while model performance remains consistent across varying boundary conditions, it is notably influenced by the specific material type due to distinct frequency characteristics inherent to each material. This sensitivity stems from the distinct frequency characteristics of AEs, reflecting the microphysical processes unique to each granular material. Consequently, models trained on one material do not transfer effectively to another.

However, rapid fine-tuning of AEsNet substantially improves its performance, outperforming a similarly fine-tuned PhaseNet model pre-trained on natural earthquakes. This highlights the necessity of tailoring models to the specific features of laboratory-generated AEs, aligning with observations in transfer learning applications for natural seismicity.

In summary, our deep learning approach effectively enhances AE detection across diverse laboratory settings, enabling the creation of reliable AE catalogs that deepen our understanding of fault mechanics. This advancement facilitates the development of reliable AE catalogs, significantly contributing to the understanding of fault mechanics in controlled experimental environments.

How to cite: Scuderi, M. M., Poggiali, G., Pignalberi, F., and Mastella, G.: Readapting PhaseNet to Laboratory Earthquakes: AEsNet, a Robust Acoustic Emission Picker Illuminating Seismic Signatures of Different Fault Gouge Materials, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-6730, https://doi.org/10.5194/egusphere-egu25-6730, 2025.

X1.87
|
EGU25-7736
|
ECS
Jorge Antonio Puente Huerta, Christian Sippl, and Vaclav Kuna

Earthquake Early Warning (EEW) systems are vital for providing timely alerts in seismically active regions, potentially reducing damage and saving lives. However, achieving both rapid and reliable alerts remains a significant challenge. Recent advances in deep learning (DL) and established workflows for picking, associating, and locating events offer complementary paths to improved performance. In this work, we propose to investigate a multi-station deep learning framework that can be integrated with existing event-location pipelines or used directly to estimate ground shaking (e.g., peak ground acceleration, PGA). By fusing raw seismic waveforms with station metadata (e.g., location, sensor characteristics) in an end-to-end manner, the approach aims to capture both local site conditions and regional propagation effects. As an initial step, we will establish baseline performance using simpler neural networks (e.g., CNNs, LSTMs), then expand to more advanced models to evaluate potential gains in accuracy and speed. Preliminary findings indicate that aggregating real-time signals from multiple stations can outperform single-station methods in both alert timing and predictive reliability. Ultimately, our goal is to develop an adaptable, data-driven EEW pipeline that accommodates either direct shaking forecasts or event-based parameter estimation, enabling seamless integration into larger-scale monitoring networks and enhancing the timeliness of earthquake alerts.

How to cite: Puente Huerta, J. A., Sippl, C., and Kuna, V.: Toward a Multi-Station Deep Learning Framework for Enhanced Earthquake Early Warning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7736, https://doi.org/10.5194/egusphere-egu25-7736, 2025.

X1.88
|
EGU25-9795
|
ECS
Foteini Dervisi, Margarita Segou, Brian Baptie, Piero Poli, Ian Main, and Andrew Curtis

The catastrophic nature of earthquakes drives the need for understanding seismic events, as well as for providing forecasts of when these are likely to occur. Due to the clustering nature of earthquakes, large magnitude events often trigger aftershocks that occur close to the mainshock in both space and time. In this study, we use a convolutional neural network to develop a data-driven spatiotemporal model to forecast next-day seismicity in an attempt to provide information that can contribute to answering one of the most pressing questions: whether a larger magnitude earthquake is to be expected after an intermediate magnitude event. We design our test to estimate expected seismicity within one day after earthquakes of magnitude four and above. We assemble a comprehensive dataset of earthquake catalogues from diverse tectonic regions to achieve a representative sample of input data and use it to create weekly spatiotemporal sequences of seismicity consisting of daily maps. Leveraging the predictive power of deep learning, our model uncovers complex patterns within this large dataset to produce next-day expected seismicity rate and magnitude forecasts in regions of interest. We use gradient-weighted class activation mapping (Grad-CAM) to provide visual explanations of the produced forecasts. We evaluate the performance of our forecasting model using data science and earthquake forecasting metrics and compare against persistence, which assumes no change between consecutive days, echoing typical experimental setups of forecasting models. Furthermore, we use a time series forecasting foundation model to generate next-day aftershock forecasts on the same dataset and compare these results against those produced by the convolutional neural network. We find that deep learning approaches are a promising solution for producing short-term aftershock forecasts, providing valuable insights that can contribute to better earthquake preparedness and response and be integrated with disciplinary statistics and physics-based forecasts.

How to cite: Dervisi, F., Segou, M., Baptie, B., Poli, P., Main, I., and Curtis, A.: Explainable artificial intelligence for short-term data-driven aftershock forecasts , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9795, https://doi.org/10.5194/egusphere-egu25-9795, 2025.

X1.89
|
EGU25-10208
Jindong Song, Yefei Ren, Hongwei Wang, and Ruizhi Wen

China is a country prone to earthquakes. Earthquake early warning (EEW) is one of the important means to mitigate earthquake disasters. To mitigate the damage caused by destructive earthquakes, we are currently developing an artificial intelligence (AI)-based EEW system in China. By utilizing emerging technologies such as artificial intelligence and big data analysis, we have developed a complete set of methods for continuous measurement of EEW parameters based on AI. This AI-based method achieves the research goal of improving the accuracy and timeliness of EEW parameters measurement in the entire workflow of EEW system, including waveform interference elimination, earthquake event recognition, P-wave picking, magnitude estimation, seismic damage zone prediction, and so on. Presently, the offline testing results of this AI-based method on earthquake data in the Sichuan-Yunnan region of China show that the AI-based magnitude estimation reaches ±1 magnitude estimation error 4 s earlier than the existing EEW system. Meanwhile, some modules such as AI-based magnitude estimation and waveform interference elimination have been running online in the Fujian Earthquake Agency in China. In the future, China is expected to establish and improve an AI-based EEW system, and further reduce the casualties and economic losses caused by earthquakes through AI-based EEW system.

How to cite: Song, J., Ren, Y., Wang, H., and Wen, R.: The development of AI-based earthquake early warning system in China, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10208, https://doi.org/10.5194/egusphere-egu25-10208, 2025.

X1.90
|
EGU25-13430
|
ECS
Antonia Kiel, Vera Schlindwein, and Conny Hammer

Ice mass loss in polar regions is a major contributor to sea level rise driven by climate change. To better predict ice mass loss due to calving and melting, it is essential to monitor ice dynamics by linking observed seismic signatures to physical processes such as meltwater infiltration into crevasses or crack formation caused by high tides at the grounding line. However, current knowledge about the distinct patterns of icequake types remains limited.

To address this gap, approximately 15 years of continuous seismic data from the Watzmann Array near the Neumayer Station in Antarctica are analyzed to automatically cluster seismic recordings. This analysis involves the automatic extraction of seismic events and the application of beamforming to each event. As a result, directional information is incorporated and the local noise is significantly reduced.

In the following, clustering methods, combined with techniques like dynamic time warping and feature extraction, are employed to categorize seismic events into distinct groups representing different icequake types. A key focus of this work is on leveraging dynamic time warping to cluster seismic waveforms directly, prioritizing the identification of physical properties inherent in the signals rather than relying solely on features extracted through machine learning. This approach ensures that the obtained clustering reflects the true underlying source processes rather than being limited to abstract feature representations.

In a follow-up study, these clusters can be related to environmental factors and directional information. Finally, with this we hope to shed some light on the hidden source processes of observed icequake types.

How to cite: Kiel, A., Schlindwein, V., and Hammer, C.: Towards the robust Clustering of various cryogenic signal types using Seismic Array Information, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13430, https://doi.org/10.5194/egusphere-egu25-13430, 2025.

X1.91
|
EGU25-13666
Yiming Sun, Hua-Liang Wei, Edward Hanna, and Linh Luu

Recent advances in machine learning (ML) have enabled significant progress in geoscience by capturing complex relationships and enhancing predictive skills. However, the success of many ML algorithms in data-rich settings does not seamlessly transfer to climate and atmospheric applications, where observational datasets are often limited. This underscores the need for methods that deliver high predictive accuracy under data-scarce conditions while retaining interpretability.

Here, we compare various ML approaches with the Nonlinear AutoRegressive Moving Average model with eXogenous inputs (NARMAX) in typical small-data climate applications, such as seasonal weather forecasting and Greenland Blocking Index (GBI) prediction. NARMAX, a transparent, interpretable, parsimonious and simulatable (TIPS) framework, demonstrates robust performance and avoids common pitfalls such as overfitting and unstable predictions when data are scarce. Notably, it achieves superior or competitive forecast accuracy for small or limited data conditions, underscoring its practical value in operational climate science. By adopting a sparse system identification approach, NARMAX yields model structures that readily reveal key predictors and their relative contributions, providing valuable physical and statistical insights into climate variability.

Our findings illustrate how NARMAX bridges the gap between purely data-driven modelling (focusing on prediction) and mechanistic modelling (focusing on physical insights), offering a clear pathway for refining model strategies and deepening our understanding of climate dynamics. We propose that NARMAX and similar methods play an inherently powerful role for both small and large data modelling problems and meanwhile serve as potent components to potentially improve the explainability of ML methods. By showcasing both interpretability and predictive efficacy, this work encourages the adoption of machine learning methods that best meet the needs for specific data modelling tasks in climate science and beyond.

How to cite: Sun, Y., Wei, H.-L., Hanna, E., and Luu, L.: Small Data for Big Tasks in Seasonal Weather Forecasting: A Balanced Perspective on Interpretability and Predictability of NARMAX and Machine Learning Methods, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13666, https://doi.org/10.5194/egusphere-egu25-13666, 2025.

X1.92
|
EGU25-14779
|
ECS
Hyeji Chae, Daeun Na, Seung-Goo Kang, and Wookeen Chung

Seismic data recorded in shallow water in the Arctic Ocean contain not only primary reflections but also surface-related multiples with strong amplitudes and short-period characteristics. These multiples generate false stratigraphic boundaries on stacked seismic sections, thereby reducing the accuracy of geological interpretation. Therefore, the attenuation of multiples is an essential step in seismic data processing for accurate geological interpretations. Recently, with the advancement of deep learning technology, research on suppressing surface-related multiples using deep learning networks (such as U-Net and stacked BiLSTM) has been actively proposed.

Firstly, surface-related multiple suppression algorithms using U-Net and stacked BiLSTM were applied to Arctic field data respectively. Each algorithm was designed to predict surface-related multiples by using input data that contained both primary reflections and surface-related multiples. Fractional Fourier transform (FrFT) and continuous wavelet transform (CWT), which represent time-series data in the time-frequency domain, were applied to synthetic data and used as input data feature for each network. Finally in order to suppress the surface-related multiples for seismic data in shallow depth Arctic Ocean, the proper methods (network architectures, input data feature) are suggested.

 

Acknowledgments

This research was supported by Korea Institute of Marine Science & Technology Promotion (KIMST) funded by the Ministry of Oceans and Fisheries, Korea (RS-2023-00259633).

 

How to cite: Chae, H., Na, D., Kang, S.-G., and Chung, W.: Deep Learning-Based Surface-Related Multiple Suppression in Shallow Arctic Seismic Data, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14779, https://doi.org/10.5194/egusphere-egu25-14779, 2025.

X1.93
|
EGU25-15039
|
ECS
Martín Sepúlveda, Marcos Moreno, and Matthew Miller

Advances in the processing of Global Navigation Satellite System (GNSS) positioning data and the increasing densification of geodetic networks have provided an unprecedented opportunity to detect and analyse transient deformation signals, including Slow Slip Events (SSE). These events, characterised by very slow rupture and durations of days to months, are often associated with areas of low coupling and sometimes show clear recurrence patterns. Despite their importance in subduction zones, reliable detection of SSEs remains an ongoing challenge. The sheer volume of GNSS data, combined with high noise levels and the subtle nature of these signals, requires efficient and robust methods capable of rapidly processing large datasets.

To overcome these challenges, we propose a method that relies on feature extraction techniques and machine learning to improve the detection and analysis of possible SSEs. Specifically, we use the TSFRESH algorithm to extract relevant features from GNSS time series, coupled with supervised machine learning classification techniques. Preliminary results of our current model, trained on synthetic data and validated through various performance tests, demonstrate high detection capabilities and accuracy. We further validated the model using a collection of GNSS time series from the Cascadia subduction zone with a single-station method scaled to the entire network, where the model showed satisfactory performance in detecting possible SSEs compared to similar work. Future efforts will focus on improving the robustness and generalisation of the model to new data, and refining methods for estimating the slip and duration of each possible SSE.

How to cite: Sepúlveda, M., Moreno, M., and Miller, M.: Machine learning and feature extraction for detecting transient signals in GNSS time series, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15039, https://doi.org/10.5194/egusphere-egu25-15039, 2025.

X1.94
|
EGU25-15129
|
ECS
Seoje Jeong, Hyeji Chae, Seung-Goo Kang, Sung-Ryul Shin, and Wookeen Chung

 In seismic exploration data from the Arctic Ocean, refractions are recorded earlier than direct waves due to the shallow depths and subsea permafrost with high velocity. These refraction signals could be utilized for estimating the velocity, thickness, and depth of the subsea permafrost. However, it is very challenging work to pick the accurate first arrivals of seismic data in the Arctic Ocean because of many factors such as ambient noise and etc. Therefore, identifying first-break refractions is crucial and can be performed by manual or automated picking methods. Various semi-automatic techniques have been developed to identify first-break refractions, but these methods are often sensitive to pulse variations and require parameter tuning. Recently, deep learning-based methods have also been explored, but their reliance on training data often results in inconsistent performance, making it essential to generate training data optimized for the target environment. 

This study presents a recurrent neural network-based algorithm optimized for Arctic Ocean environments to automatically identify first-break refractions. To effectively classify first-break refractions, a stacked bidirectional long short-term memory (BiLSTM) network was constructed to iteratively learn bidirectional long-term dependencies by utilizing the temporal patterns of time-series data. Additionally, the training data were generated by creating velocity models that reflect the subsurface properties of subsea permafrost, enabling the generation of first-break refraction label data. The proposed network demonstrated superior performance in identifying first-break refractions from noisy data, achieving over 95% accuracy in numerical experiments and field tests. Field data applications demonstrated that the proposed network achieves high accuracy in classifying first-break refractions, validating its robustness and adaptability.

 

 

Acknowledgments

This research was supported by Korea Institute of Marine Science & Technology Promotion (KIMST) funded by the Ministry of Oceans and Fisheries, Korea (RS-2023-00259633).

How to cite: Jeong, S., Chae, H., Kang, .-G., Shin, .-R., and Chung, W.: Auto-Picking of First-Break Refractions in Arctic Ocean Seismic Data Using Stacked BiLSTM Networks, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15129, https://doi.org/10.5194/egusphere-egu25-15129, 2025.

X1.95
|
EGU25-10901
Quentin Bletery, Gabriela Arias, Kévin Juhel, Céline Hourcade, Andrea Licciardi, Adolfo Inza, Martin Vallée, and Jean-Paul Ampuero

Earthquake early warning (EEW) systems implemented worldwide use early seismic records of P waves to rapidly detect, locate, and estimate the magnitude (Mw) of potentially damaging earthquakes. These systems are well known to saturate for large magnitude events, which results in dramatic underestimation of the subsequent tsunamis. Alternative approaches based on different signals have been proposed to rapidly estimate the magnitude of large events, but these approaches are much slower (taking 10 to 20 minutes for first warning). Prompt elasto-gravity signals (PEGS) are light-speed gravitational perturbations induced by large earthquakes that can be recorded by broadband seismometers. They have tremendous potential for early warning but their extremely small amplitudes (on the order of 1 nm/s2) have challenged their possible operational use. We designed a deep learning approach to rapidly estimate the magnitude of large earthquakes based on PEGS. We applied this approach to the seismic networks operating in Japan, Chile, Alaska and Peru. We will present the performances obtained in these different contexts. In Alaska, the approach has proven capable to reliably estimate the magnitude of Mw ≥ 7.6 earthquakes (without saturation) in less than 2 minutes, outperforming state-of-the-art tsunami early warning algorithms. Motivated by these performances, we initiated a first implementation of an operational tsunami warning system based on PEGS in Peru. We will present the simulated real-time performance of this system. 

How to cite: Bletery, Q., Arias, G., Juhel, K., Hourcade, C., Licciardi, A., Inza, A., Vallée, M., and Ampuero, J.-P.: Machine-learning-based operational tsunami warning from light-speed elasto-gravity signals, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10901, https://doi.org/10.5194/egusphere-egu25-10901, 2025.

X1.96
|
EGU25-11459
|
ECS
Valentin Kasburg, Marcel van Laaten, Markus Zehner, Jozef Müller, and Nina Kukowski

To date, the discrimination of seismic events recorded in seismic networks is often performed manually by experts, classifying events into categories such as earthquakes, quarry blasts, or mining events. While recent studies have shown that deep learning algorithms, particularly Convolutional Neural Networks (CNNs), can efficiently and accurately distinguish between different types of seismic events, their application for automated seismic event discrimination remains limited. This limitation arises from several factors, including the absence of globally applicable models that maintain high precision for local seismic networks, the scarcity of data required for fine-tuning Deep Learning (DL) models, and the lack of interpretability in the decision-making processes of these black-box models.

In this contribution, we explore the use of Vision Transformers (ViTs) as a novel approach for automating seismic event discrimination. To assess their potential for accuracy and explainability, we applied CNNs and ViTs to classify seismic events such as earthquakes, quarry blasts, and mining events. For this purpose, we pretrained the models on openly available seismic event data from Utah and Northern California and then fine-tuned and tested them on data from the Seismic Network of the Ruhr-University Bochum (RuhrNet) and the Thuringian Seismic Network (TSN).

Our findings reveal that ViTs can analyze the entire spectrogram of a seismic event in a coherent manner, offering superior generalizability in pattern recognition compared to CNNs. In addition to achieving high discrimination accuracy, the attention weights of ViTs provide insights into the models’ decision-making process, offering a transparent and interpretable explanation of the underlying mechanisms driving its classifications.

How to cite: Kasburg, V., van Laaten, M., Zehner, M., Müller, J., and Kukowski, N.: Automating Seismic Event Discrimination: A Comparative Study of Convolutional Neural Networks and Vision Transformers, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11459, https://doi.org/10.5194/egusphere-egu25-11459, 2025.

X1.97
|
EGU25-16157
|
ECS
Aurora Bassani, Daniele Trappolini, Giulio Poggiali, Elisa Tinti, Fabio Galasso, Chris Maron, and Alberto Michelini

Estimation of earthquake parameters has always been a focus for seismologists. Efficient and rapid determination of earthquake location and magnitude is essential for mitigating the potential hazards associated with seismic shaking. Nowadays, Earthquake Early Warning Systems (EEWS) are implemented in most earthquake-prone areas, with the system varying according to the specific needs. Although methods for their estimation exist, many still lack a fast enough process, which is crucial for reducing the waiting time before issuing a warning.

Here, we propose a novel model to enhance multi-station EEWS using Large Language Models (LLM). We adopt a pre-trained LLM and fine-tune it on a customized version of INSTANCE (The Italian Seismic Dataset for Machine Learning), thus eliminating the need to develop and train a tailor-made architecture. The model uses stations with P-wave arrival times up to 5 s apart from the first recorded one, and, for each seismic trace, it exploits a very small time window around the P-wave arrival time (0.21 s), thus effectively reducing warning latency.

Comparative analysis against the automatic method employed by the Italian National Institute of Geophysics and Volcanology (INGV) demonstrates that our model achieves comparable performance in magnitude estimation and superior accuracy in epicenter, hypocenter and origin time prediction. For instance, the LLM-based model achieves average errors of 6.3 km, 11.1 km, and 1.1 s for epicenter, hypocenter, and origin time estimation, respectively, in contrast to 8.6 km, 15.0 km, and 1.8 s for the INGV automatic solution resulting in an average improvement of more than 26% for all parameters.

We study the validity of our model by assessing its ability using P- and S-waves to predict magnitude, and show that in this case study the S-waves are not strictly necessary for accurate predictions.

How to cite: Bassani, A., Trappolini, D., Poggiali, G., Tinti, E., Galasso, F., Maron, C., and Michelini, A.: Real Time Estimation of Earthquake Location and Magnitude Using Large Language Models, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-16157, https://doi.org/10.5194/egusphere-egu25-16157, 2025.

Posters virtual: Mon, 28 Apr, 14:00–15:45 | vPoster spot 1

The posters scheduled for virtual presentation are visible in Gather.Town. Attendees are asked to meet the authors during the scheduled attendance time for live video chats. If authors uploaded their presentation files, these files are also linked from the abstracts below. The button to access Gather.Town appears just before the time block starts. Onsite attendees can also visit the virtual poster sessions at the vPoster spots (equal to PICO spots).
Display time: Mon, 28 Apr, 08:30–18:00
Chairpersons: Alice-Agnes Gabriel, Philippe Jousset

EGU25-1772 | Posters virtual | VPS21

Application of Optimized 3D U-Net Residual Network with CBAM and MEGA Modules in Seismic Fault Detection 

yu wang
Mon, 28 Apr, 14:00–15:45 (CEST) | vP1.3

With the continuous development of deep learning technologies, fault prediction techniques based on various neural networks have been evolving. The deep learning modules based on U-Net residual networks have shown significant advantages in both learning efficiency and effectiveness. In this paper, we propose a deep learning model that integrates a 3D U-Net residual architecture, Convolutional Block Attention Module (CBAM), and Multi-scale Enhanced Global Attention (MEGA) module for automatic seismic fault detection and segmentation. This model can effectively handle complex 3D seismic data, fully exploiting both spatial and channel information, significantly improving the prediction accuracy for small faults, while only slightly increasing the computational cost.

Firstly, the model uses the 3D U-Net as the backbone framework, where the residual blocks (BasicRes) extract features through multiple convolution layers. The CBAM module is incorporated to apply attention weighting, enhancing the model's ability to focus on critical information. The CBAM module combines channel attention and spatial attention, effectively adjusting the importance of feature maps from different dimensions, enabling the model to identify potential fault features in complex seismic data.

Secondly, the MEGA module is introduced into the model, which further improves the model's feature representation ability by fusing multi-scale features and applying a global attention mechanism. By weighting global information, the MEGA module helps the model better capture key seismic fault features during feature fusion. This design allows the model to focus not only on local details but also to fully utilize the global contextual information in 3D data, thereby enhancing the accuracy of fault detection.

After validation, the model achieved promising results in seismic fault detection tasks, automatically identifying and segmenting fault structures in seismic data. The accuracy was improved from 80% with the original 3D U-Net residual network to 85%-87%. This provides strong support for applications such as seismic exploration and subsurface imaging.

Keywords: Seismic Fault Detection, 3D U-Net, Convolutional Block Attention Module (CBAM), Multi-scale Enhanced Global Attention (MEGA), Deep Learning

How to cite: wang, Y.: Application of Optimized 3D U-Net Residual Network with CBAM and MEGA Modules in Seismic Fault Detection, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-1772, https://doi.org/10.5194/egusphere-egu25-1772, 2025.

EGU25-3391 | Posters virtual | VPS21

Study and Case Application of Fluvial Reservoir Prediction Based on the Fusion of Seismic Attribute Analysis and Machine Learning Technologies 

Zheng Huang and Junhua Zhang
Mon, 28 Apr, 14:00–15:45 (CEST) | vP1.4

Seismic attribute analysis technology has been widely used in the prediction of fluvial reservoir sand body, but the traditional seismic attribute fusion technology based on linear model has low prediction accuracy and limited application range. This study focused on the non-linear fitting between seismic attributes and reservoir thickness, and used a variety of machine learning technologies to predict the fluvidal reservoir in Chengdao area of Dongying Sag (China).The channel sand body in Chengdao area is deep buried, thin in thickness, fast in velocity and affected by gray matter, so it is difficult to predict, which greatly restricts the oil and gas exploration in this area. In this study, on the basis of fine well earthquake calibration, several seismic attributes such as amplitude, frequency, phase, waveform and correlation are extracted and correlation analysis is done to remove redundant attributes. Then model training and parameter set optimization are carried out, thickness prediction is carried out with verification set, and vertical resolution is improved by logging reconstruction and waveform indication inversion. The results show that compared with the conventional support vector machine and back propagation neural network, the prediction accuracy of echo state network optimized by Sparrow algorithm is greatly improved. Based on the comprehensive prediction method of fluvial reservoir, three large channels developed in the lower part of Chengdao area and several small channels developed in the upper part of Chengdao area are effectively described. The research method can be used for reference to the similar complicated river facies prediction.

How to cite: Huang, Z. and Zhang, J.: Study and Case Application of Fluvial Reservoir Prediction Based on the Fusion of Seismic Attribute Analysis and Machine Learning Technologies, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-3391, https://doi.org/10.5194/egusphere-egu25-3391, 2025.