- How can machine learning bridge different data modalities and capture the different scales of geophysical processes?
- How can we efficiently encode physics knowledge into machine learning models or extract physical insights from machine learning black boxes?
- What standardized datasets and evaluation benchmarks are needed to ensure fair comparison, reproducibility, and progress?
- How can simulated data help alleviate data-poor scenarios, such as rare extreme events?
- What is the impact of recent developments in artificial intelligence, such as the advent of large language models and foundation models on geophysics?
- How can we lower model complexity to reduce environmental impact and enable use in low-power contexts?
What are the best practices for integrating machine learning into mission-critical production systems, such as early warning applications?
In this session, we aim to address these questions and related active topics in the development and application of machine learning for geophysical data. We aim to bring together machine learning researchers and practitioners from different geophysical domains to identify common challenges and opportunities. We welcome contributions from all fields of geophysics, covering a wide range of data types and machine learning techniques. We also encourage contributions for machine learning adjacent tasks, such as big-data management, data visualization, or software development.
Orals: Mon, 4 May, 16:15–17:55 | Room K2
The exploration and characterization of complex continuous seismological datasets remain challenging, particularly for highly active and/or noisy environments. Recently, several artificial intelligence based approaches have been proposed to facilitate the analysis of seismological data, either by characterizing detected events or continuous streams. Among these, we introduced an image-based self-supervised learning framework to explore continuous seismic records without requiring prior labeling or event detection. However, image-based representations may result in a loss of information, as they are derived transformations of the original raw seismic time series and may impact the discrimination of seismic events.
In this study, we adapted a self-supervised learning based clustering workflow to operate directly on multichannel seismic time series. The main challenge when using self-supervised contrastive learning approaches with time series is adapting the data augmentation techniques to ensure sufficient transformation without losing the physics contained in the seismological records. We leveraged the contrastive learning framework to analyse two months of continuous records from Ocean Bottom Seismometers deployed near the Fani Maoré submarine volcano, using data augmentation strategies consistent with seismological records, such as channel masking and window masking in the time and frequency domains. The model was trained using four-channel time series derived from the raw data (three-component seismometer and one hydrophone) using 60 s sliding windows with a 50% overlap, enabling the network to learn meaningful latent representations of the data. Clustering was then performed directly within the learned latent space, allowing the identification of distinct signal groups. Applied to the Fani Maoré dataset, this approach revealed several families of clusters, including very rare and previously undocumented events likely associated with the activity of the Fani Maoré submarine volcano.
How to cite: Rimpot, J., Retailleau, L., Saurel, J.-M., Hibert, C., Malet, J.-P., Forestier, G., and Weber, J.: A Time-Series-Based Self-Supervised Learning Approach for the Exploration of Complex Seismological Datasets: Application to the Fani Maoré Submarine Volcano, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13291, https://doi.org/10.5194/egusphere-egu26-13291, 2026.
The continuous seismic monitoring provides numerous waveform signals that provide useful information about geological processes such as landslides and geohazard activities. In tectonically active zones like the Himalayan arc, mass-movement events often overlap with natural earthquakes, creating challenges for reliable event discrimination. This study presents a semi-supervised learning framework that detects landslides like debris flow, rockfall, as well as earthquakes, from continuous seismological data with the help of a small amount of labelled dataset, calculating physically interpretable attributes from waveforms. The waveform and spectrum-based 157 features were extracted from segmented seismic windows, representing temporal, spectral, energy, and morphological attributes. The workflow combines dimensionality reduction, density-based clustering, and graph-based label propagation to identify and classify seismic events. To evaluate methodological choices, two complementary studies were conducted. The model benchmarking study compared four combinations of embedding and clustering algorithms, and a parametric sensitivity analysis that investigated the influence of key hyperparameters of the embedding and clustering algorithms. Feature importance analysis using statistical and machine-learning-based techniques was integrated throughout the study to ensure physical interpretability and to identify attributes most relevant for source discrimination. In this study, three months of continuous seismological data of the Tehri region, Uttarakhand, were analysed and revealed many previously undetected events. Uniform Manifold Approximation and Projection (UMAP) for dimensionality reduction, followed by Hierarchical Density-Based Spatial Clustering (HDBSCAN) for unsupervised event grouping, provided the best performance for seismic event detection. This approach effectively identified seismic events that would be difficult to observe using conventional methods. The proposed approach is well-suited for large-scale seismic monitoring applications where labelled data are limited and provides a broad application for geohazard detection and operational seismic analysis.
How to cite: Chakraborty, A. and Sharma, M.: Detection of Seismic Events Using Semi-Supervised Learning , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5798, https://doi.org/10.5194/egusphere-egu26-5798, 2026.
Earthquake catalogues are derived from continuous seismic recordings through signal detection, phase picking, association, location, and magnitude estimation, but pervasive noise still limits reliable automation, especially for small events and noisy stations, and often requires manual review. Building on the demonstration that deep-learning denoising can be applied to continuous data to improve network-wide earthquake monitoring (Dahmen et al., 2026), we advance toward operational deployment by (i) implementing denoising within the SeisComP ecosystem (Helmholtz Centre Potsdam, 2008), (ii) evaluating on larger continuous datasets, and (iii) systematically comparing multiple denoising approaches and testing monitoring-driven methodological refinements.
We train and compare multiple denoising models using a dedicated, curated training and benchmarking dataset composed of earthquake signals and noise recordings from Switzerland and its border regions, with event waveforms pre-cleaned to enhance label quality. Denoised waveforms are then propagated through an end-to-end monitoring workflow spanning signal detection, continuous waveform denoising, phase picking with arrival-time uncertainty estimation and peak amplitude estimation, and final catalog generation. The performance is benchmarked with monitoring-relevant metrics such as signal detection capability, waveform fidelity, phase-pick quality, and the reliability of amplitude estimation, thereby quantifying, for each denoiser, the trade-offs and improvements relative to standard digital filters and relative to applying common phase pickers to raw versus denoised data.
A case study using continuous data in realistic settings shows that catalogues based on denoised data can contain significantly more detected events with more associated phase picks, improved location quality, and more reliable magnitude estimates than catalogs derived from raw data, ultimately extending catalogue depth toward smaller magnitudes while preserving reliability.
This work is carried out within TRANSFORM², funded by the European Commission under project number 101188365 within the HORIZON-INFRA-2024-DEV-01-01 call.
References:
Dahmen, N., J. Clinton, M.-A. Meier, and L. Scarabello, 2026, Toward Operational Earthquake Seismogram Denoising, Bull. Seismol. Soc. Am., XX, 1–23, doi: 10.1785/0120250198
Helmholtz Centre Potsdam (2008). The SeisComP seismological software package, GFZ Data Services.
How to cite: Dahmen, N., Clinton, J., Meier, M.-A., and Scarabello, L.: Towards Operational Earthquake Data Denoising, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17302, https://doi.org/10.5194/egusphere-egu26-17302, 2026.
The STAR-E Project that aims to develop state-of-the-art information science techniques, including artificial intelligence, applicable in seismology is going on in Japan, supported by the Ministry of Education, Culture, Sports, Science and Technology (MEXT). We introduce our various activities in the SYNTHA-Seis, which is one of the research teams in the STAR-E Project. We have been developing deep learning models to detect earthquake signals in seismic continuous waveforms, such as for phase-picking (Tokuda and Nagao, 2023; Katoh et al., 2025; Gerardo Mendo et al., submitted) and for P-wave polarity determination with UQ using the Monte Carlo dropout method (Katoh et al., 2025). We have also been developing methods to extract waveform features of low-frequency tremors (LFTs), such as a template matching technique to extract LFTs waveforms (Gerardo Mendo et al., 2025), a deep learning model to detect LFTs in historical paper records obtained by mechanical seismograms more than fifty years ago (Kaneko et al., 2023), and a deep learning technique to acquire a stochastic differential equation expression of LFTs (Kusui et al., 2025). We also discuss the future direction of AI seismology in Japan.
How to cite: Nagao, H., Mendo Pérez, G. M., Katoh, S., and Kusui, T.: Deep Learning Models for Seismic Continuous Data Analysis Developed in STAR-E Project, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18544, https://doi.org/10.5194/egusphere-egu26-18544, 2026.
Estimating ground‐motion intensity at individual seismic stations is a fundamental task in seismology, with direct implications for seismic hazard assessment.
Ground‐motion intensity measures (IMs) such as Peak Ground Acceleration (PGA), Peak Ground Velocity (PGV), and Spectral Acceleration (SA) at selected periods are commonly used to quantify shaking severity and relate it to building structural response and seismic hazard.
Here, we present an AI-driven framework for predicting multiple IMs at the station level, building on previous graph-based approaches for the Italian seismic network.
Starting from the INSTANCE dataset, we applied filters on event magnitude (≥ 3) and waveform quality, resulting in 3076 events recorded across 565 stations. For waveform analysis, we used a 10-second time window starting 1 second before the first P-wave arrival, balancing prediction speed and accuracy.
Seismic waveforms are first encoded using a pre-trained PhaseNet model to extract compact temporal representations. Spatial dependencies are modeled with a masked Graph Convolutional Network (GCN) based on Delaunay triangulation, which links each station to its nearest neighbors while avoiding long or crossing edges. This structure allows identification of border stations (at mesh edges) and coastal stations (within 15 km of the coastline). A binary mask distinguishes these nodes, helping the model account for areas with high azimuthal gap, which can make IMs estimation more challenging.
Before the final prediction layer, the model concatenates the maximum waveform amplitude across stations with event metadata predicted by a fine-tuned LLM for magnitude and location estimation.
This enables joint exploitation of temporal waveform features, network geometry, and global event information.
The framework predicts PGA, PGV, and SA at periods of 0.3, 1, and 3 s at all stations. We obtained preliminary results for multiple configurations. The baseline model served as reference, which included waveform representations and the GCN but not LLM metadata, border features, or weighted loss. Including LLM-derived metadata consistently improved performance across all regions and parameters, reducing relative errors by 14% in Southern Italy and over 27% in Northern Italy. The addition of border and coastal features provided only minor improvements, confirming that metadata was the main factor driving gains.
Applying a weighted loss emphasizing stations closer to the epicenter further improved predictions, particularly in Northern and Southern Italy. In Central Italy, where network coverage is denser, improvements were smaller, suggesting that local contributions were already well captured. For the best configuration (LLM metadata and weighted loss), global mean absolute errors across stations were 0.511 (PGA), 0.423 (PGV), 0.514 (SA0.3), 0.447 (SA1.0), and 0.397 (SA3.0), demonstrating the model’s predictive accuracy.
Overall, these preliminary results show that combining waveform features, network topology, and LLM-informed event metadata can substantially enhance station-level IMs estimation, achieving better results in challenging conditions (north or south Italy) with respect to previous similar approaches. This method has potential for rapid earthquake characterization and early warning, where timely and accurate ground-motion predictions are essential.
How to cite: Bassani, A., Trappolini, D., Scifoni, A., Poggiali, G., Tinti, E., Galasso, F., Michelini, A., and Marone, C.: AI Framework for Ground-Motion Prediction Across the Italian Seismic Network, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18748, https://doi.org/10.5194/egusphere-egu26-18748, 2026.
Laboratory, theoretical and field data suggest that fault-zone properties should evolve during the seismic cycle as stress rises prior to failure and drops during earthquake rupture. Lab work shows systematic changes in elastic properties during the seismic cycle and that these changes can be used to predict lab earthquakes. Recent work shows that in some cases these results are also applicable to tectonic faults. Seismic data show a clear distinction between fault zone properties pre and post mainshock, as well as post-seismic time-dependent changes in elastic properties. Here we extend these works by developing tools to distinguish seismic waves pre/post mainshock that are both predictive and physically interpretable. We train a convolutional neural network (CNN) on RGB spectrograms developed from three component seismograms recorded at seismic stations around the M6.5 2016 Norcia earthquake. Our model can accurately distinguish foreshocks from aftershocks of the sequence. We train and test models on individual stations and also on all stations and subsets of the stations based on source-station geometry of the Norcia fault. Models trained on the full set of stations achieve >99% accuracy for foreshock/aftershock classification and models trained on individual stations achieve higher accuracy in tests. For each set we also performed the SHapley Additive exPlanations (SHAP) technique. We find that specific time-frequency signatures in the RGB spectrograms identify each class. Here we extend that framework to a multi-station setting by training CNN models on spectrograms from several seismic stations surrounding the mainshock. While the multi-station model achieves high classification accuracy (about 97%), SHAP analysis reveals a substantial reorganization of feature importance, including strong station-dependent variability and a reduced contribution from aftershock-related regions. Even for data from the reference station (NRCA), SHAP patterns differ markedly from those obtained in the single-station case, suggesting that heterogeneous training distributions alter global attribution mechanisms. To disentangle these effects, we additionally train station-wise CNN models, which achieve very high accuracy and produce more stable and physically coherent SHAP explanations. These results indicate that station-specific propagation effects play a key role in model interpretability and that caution is required when applying SHAP to models trained on spatially heterogeneous seismic datasets. The findings motivate future work toward hierarchical, region-aware, or physics-constrained interpretability frameworks.
How to cite: Marrocco, F., Magrini, M., Laurenti, L., Paoletti, G., Tinti, E., and Marone, C.: Station-Level and Network-Wide SHAP Explanation of CNN Models for Seismic Cycle Monitoring: Evidence from Norcia 2016, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10105, https://doi.org/10.5194/egusphere-egu26-10105, 2026.
Microseismic source localization is a key diagnostic in stimulated reservoirs, supporting spa-
tiotemporal tracking of fracture activation, stress transfer, and operational risk. Distributed Acoustic
Sensing (DAS) provides dense strain-rate observations along fiber-optic cables, but the resulting data
volume and strong near-well heterogeneity motivate localization workflows that are both fast and
physically constrained. We present an Eikonal-regularized U-Net Fourier Neural Operator (U-FNO)
that predicts full first-arrival traveltime fields for a given source location in a known 2-D velocity
model. The architecture combines Fourier-domain operator learning to capture long-range kinematic
structure with a multiscale encoder–decoder to recover spatial detail. Training is guided by an
Eikonal-consistency loss, complemented by source anchoring and a non-negativity constraint to
encourage physically admissible solutions. We benchmark U-FNO against a vanilla FNO baseline
and fast-marching traveltime solutions across velocity models of increasing complexity (smooth
gradient, Marmousi, and Utah FORGE). In the FORGE model, U-FNO reduces traveltime RMSE
by up to 97% relative to the baseline and reaches comparable misfit in up to 50% fewer training time.
Field transferability is assessed using 15 DAS-recorded events from the Utah FORGE microseismic
catalogue. Models are fine-tuned on four events and evaluated on the remaining events. U-FNO converges
within 2 minutes (versus 45 minutes for FNO) and reduces the mean location error from 33.71 m to
29.28 m. These results indicate that physics-regularized neural operators with multiscale structure can
deliver accurate, scalable, near-real-time localization for high-volume DAS monitoring in complex
geothermal settings.
How to cite: Al-Qadasi, B. and Bin Waheed, U.: Physics-Informed U-Net Fourier Neural Operator for DAS Microseismic Localization at Utah FORGE, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20316, https://doi.org/10.5194/egusphere-egu26-20316, 2026.
Traditional volcano monitoring relies on dense ground-based instrumentation (e.g., seismic, gravity, and deformation measurements), which is available for only a small fraction of the world’s volcanoes. While satellite observations can complement these measurements, estimating total erupted mass—a primary metric of eruption magnitude—remains largely a post-eruption task. Consequently, erupted mass has been quantified for only ~100 of the 1,282 known volcanoes.
These limitations motivate the use of machine learning (ML) in volcanic prediction. Unlike traditional approaches, ML can integrate diverse, globally available datasets to generate predictive insights for volcanoes with limited or no time-dependent monitoring. This capability enables proactive risk assessment and hazard planning, offering a scalable, cost-effective, and globally applicable tool for volcanic risk mitigation. By capturing complex nonlinear relationships among static geophysical, petrological, and tectonic parameters, ML allows estimation of eruption magnitude prior to or early in eruptive activity, an outcome infeasible using classical approaches.
To demonstrate this potential, we present a ML framework to forecast a volcano’s potential erupted mass using static geophysical, petrological, and tectonic characteristics, together with eruption history. The model was trained on a dataset of 914 historical eruptions from 101 volcanoes and applied to estimate erupted mass for 135 globally distributed volcanoes active between 1982 and 2024, assuming a representative eruption duration of 225 days.
This approach provides the first global-scale erupted mass estimates that do not rely on high-resolution, time-dependent monitoring data (e.g., time-lapsed gravity, deformation, or seismicity). Feature importance and permutation analyses indicate that predictions are dominated by static geophysical parameters. The most influential predictors are eruption duration, elevation, gravity, and magnetic data. Parameters with intermediate influence include Moho depth, subsurface thermal indicators (e.g., depth to the Curie isotherm at ~580 °C and the lithosphere–asthenosphere boundary at ~1330 °C), dominant rock type, last eruption, and surface heat flow. Volcano landform, eruption type and start date, host crustal type, and tectonic setting exhibit relatively minor predictive influence.
Our results demonstrate that comprehensive, globally available geophysical datasets can robustly constrain erupted mass for medium- to long-duration eruptions (>3 months), while short-duration eruptive behavior may be better captured through detailed historical eruption records.
How to cite: Mousavi, N., Fullea, J., and Mousavi, S. M.: Volcanic Eruption Mass Estimation: A Machine Learning Approach, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4281, https://doi.org/10.5194/egusphere-egu26-4281, 2026.
High-energy surface waves (ground roll) are a major source of coherent noise in land seismic data, often overlapping with reflections and degrading subsurface imaging quality. We propose an intelligent surface-wave suppression method based on a dual-constraint framework that integrates data-driven supervision with a physics-guided prior. A composite loss is constructed with (1) a data constraint in the time–space (t–x) domain, implemented as a supervised loss that compares the network output with the labeled targets, and (2) a physics constraint in the frequency–velocity (f–v) domain, where the surface-wave dispersion curve is exploited to delineate the physically plausible ground-roll region and to penalize residual energy inconsistent with the dispersion curve. We train UNet and TransUNet architectures on field datasets using this composite objective. Compared with purely data-driven training, the proposed dual-constraint loss reduces the dependence on potentially imperfect labels by enforcing dispersion-consistent behavior in the f–v domain, leading to lower residual surface-wave energy while maintaining reflection continuity. These results demonstrate that incorporating physically meaningful constraints into modern network architectures can improve robustness under imperfect supervision and enhance intelligent seismic surface-wave suppression.
How to cite: Wang, B., Shi, Y., Wen, J., and Ning, J.: A Data-Physics Dual-Constraint Framework for Intelligent Surface Wave Suppression, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4447, https://doi.org/10.5194/egusphere-egu26-4447, 2026.
Accurate delineation of subsurface lithological and structural characteristics is essential for applications ranging from groundwater exploration to environmental geophysics. Traditional single-modality inversion techniques often suffer from resolution trade-offs and ambiguity in petrophysical interpretation. In this work, we propose a novel multimodal deep learning framework for the joint inversion of Ground Penetrating Radar (GPR) and Vertical Electrical Sounding (VES) data, enabled through physics-informed Siamese neural network architecture. This network is explicitly designed to learn shared subsurface representations by encoding signal-specific features via dual 1D convolutional pathway, which are subsequently fused into a common latent embedding. From this shared space, the network predicts three key geophysical outputs: (i) layer-wise resistivity, (ii) normalized thicknesses , and (iii) geological model classification from eight lithological types.
To ensure physical consistency between modalities, the architecture incorporates an empirical dielectric–resistivity relationship derived from soil physics literature as a physics-informed regularization loss, coupling the inferred resistivity profile with dielectric behavior. The resistivity head uses a Huber loss on log-transformed outputs to reduce the effect of noise and outliers, while the thickness head is stabilized with Batch Normalization and dropout layers to prevent over fitting. A multi-class cross-entropy loss is used for geological classification, and a joint loss function ensures simultaneous optimization across modalities.
Training is conducted on a synthetically generated dataset comprising 24,000 4-layer models, covering diverse resistivity-thickness scenarios and geological facies. A dedicated subset includes thin-layer configurations, simulating challenging cases where GPR contributes enhanced resolution beyond VES capabilities. The network achieves a classification accuracy of 95.4%, a resistivity RMSE of 1.76 Ω·m, and thickness RMSE of 1.85 m on unseen validation data, validating its predictive performance. An ablation study with three independent random seeds (42, 123, 2025) confirms the network’s stability and generalizability.
Visual comparison of predicted vs. true resistivity and thickness profiles exhibits strong structural alignment, even in geologically complex models. Further, input-output attention diagnostics and multimodal fusion behavior reveal interpretable latent correlations between GPR and VES responses.
This work introduces a scalable and domain-aware inversion framework that learns geophysical realism, respects petrophysical coupling laws, and demonstrates potential for field-deployable AI-assisted subsurface mapping. The integration of empirical physics, attention mechanisms, and synthetic realism places this methodology at the frontier of modern geophysical inversion strategies.
How to cite: Gogoi, H., Sengupta, P., Hazarika, C., and Dutta, A. D.: Joint Inversion of GPR and VES Data Using Physics-Guided Siamese Networks, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21987, https://doi.org/10.5194/egusphere-egu26-21987, 2026.
Posters on site: Thu, 7 May, 14:00–15:45 | Hall X1
Indonesia is located in one of the most seismically active regions in the world and is monitored by more than 550 broadband and short-period seismic sensors. On average, around 40,000 earthquakes with magnitudes greater than 2 occur each year. However, this number of earthquakes with magnitudes larger than 2 has only been observable in recent years due to the significant expansion of the seismic network, for example in 2025. In earlier years, the recorded number of earthquakes was significantly lower, with the magnitude of completeness (Mc) reaching only about 4. Completing earthquake catalogs in a region is extremely important for revealing detailed main and secondary fault structures, which is essential for improved earthquake and tsunami hazard assessment. Recently, earthquake event recognition and phase picking using machine learning (ML) have proven highly successful in detecting smaller-magnitude earthquakes that are often overlooked by conventional methods such as STA/LTA. This study presents high-resolution earthquake catalogs generated using ML-based earthquake detection. The ML algorithms were pre-trained across various regions, tectonic settings, and environmental conditions using data from the last decade of combined BMKG and temporary seismic networks across Indonesia. We compare a three-catalog framework consisting of ML-derived, real-time, and analyst-reviewed catalogs. The results show that ML-based detection identifies significantly more earthquakes at lower magnitudes, with Mc reaching approximately 2, compared to real-time processing and human-reviewed catalogs. Using a recently published ML-based focal mechanism model, our results also show a substantially larger focal mechanism catalog, including many events with magnitudes smaller than 5 that are often not reviewed in conventional processing. This study demonstrates the importance of ML-based earthquake detection in improving the efficiency and completeness of earthquake detection and highlights its strong potential for operational integration into Indonesia’s seismic monitoring systems.
How to cite: Simanjuntak, A. V. H., Warnana, D. D., Pranata, B., Supendi, P., Daryono, D., Mai, M., Riama, N. F., and Palgunadi, K. H.: Completing Indonesia Earthquake Catalog for Better Earthquake and Tsunami Hazard Assessments, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20854, https://doi.org/10.5194/egusphere-egu26-20854, 2026.
To address this limitation, we explore replacing such inversions with convolutional neural networks (CNNs) trained on physically informed synthetic waveforms to infer sea-ice thickness. The synthetic icequake waveforms are generated by a one-dimensional forward model for flexural waves in floating ice that accounts for ice thickness, mechanical properties, and source–receiver distance. Realistic source spectra are incorporated using a library of field data.
On synthetic waveforms, the networks recover ice thickness and source distance with low error, indicating that the learned relationship between waveform characteristics and physical parameters captures the dominant dispersive physics. However, when applied directly to field icequake waveforms, the accuracy decreases, reflecting the limitations of using synthetic waveforms alone for training due to idealized model assumptions. Based on additional tests, we outline strategies to improve CNN performance and robustness when applied to field data.
How to cite: Michael, J., Moreau, L., and Malfante, M.: Deep Learning Seismic Waveforms to Predict Sea Ice Properties, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4942, https://doi.org/10.5194/egusphere-egu26-4942, 2026.
Accurate and robust seismic phase picking remains a fundamental challenge in automated earthquake monitoring, particularly under complex waveform conditions. While recent deep learning models such as PhaseNet and EQTransformer have demonstrated strong performance on commonly used benchmark datasets, their architectural design choices introduce limitations in temporal resolution, sequence modeling, and generalization to more realistic seismic scenarios.
PhaseNet adopts a convolutional U-Net structure that is effective for waveform segmentation but is constrained by a fixed receptive field and limited dynamic temporal modeling. EQTransformer, in contrast, employs an encoder–attention–decoder architecture capable of capturing long-range dependencies, yet relies on aggressive temporal downsampling to alleviate the quadratic cost of self-attention. This heavy compression can discard fine-grained temporal information and degrade phase onset precision.
In this work, we present EQMamba, a sequence modeling framework for seismic phase picking that emphasizes temporal fidelity and efficient long-range dependency modeling. The proposed architecture integrates structured state-space models with efficient linear attention mechanisms, allowing long waveform sequences to be processed with minimal downsampling. By preserving high-resolution temporal information while maintaining computational tractability, EQMamba is designed to better reflect the continuous-time and dynamical nature of seismic signals.
Beyond controlled single-event settings, this study places particular emphasis on model behavior under more realistic waveform conditions, including event superposition, amplitude imbalance, and temporal interference between phases. We introduce a revised data construction and evaluation strategy to systematically probe robustness in multi-event and mainshock–aftershock scenarios, which are often underrepresented in standard benchmarks. Model performance is analyzed not only through conventional picking metrics, but also via error distributions, phase confusion patterns, and failure modes as waveform complexity increases.
How to cite: Wang, C.-H., Wang, W.-H., and Lu, H.-H.: Toward Robust Seismic Phase Picking in Realistic Multi-Event Scenarios, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6122, https://doi.org/10.5194/egusphere-egu26-6122, 2026.
Optical images acquired by satellite are widely used to measure the displacement field caused by earthquakes co-seismic ruptures both in the near and the far field. So far, this technique relies mostly on image correlation to locate pixels from an image acquired before the event on another acquired after. However, such approach suffers from several limitations inherent in the correlation method used, such as the need for sufficient texture to make objects « recognizable » from one image to the other, or limited changes in the landscape over time, for the same reason. These limitations can lead to noisy results and even prevent any measurement from being made.
To overcome such limitations, machine learning can be used instead of correlation to train a model to compute displacement maps from a pair of images. Steady and significant progress have been made in machine learning technics, and especially for image processing and computer vision, in recent years, and they need to be adapted to our case study. First, a training dataset is carefully designed, to enable the network to learn how to measure pixel displacements in satellite images, with sub-pixel accuracy, in the most realistic way possible. Since no ground truth is available, we build synthetic examples, where a realistic and known deformation is applied to one of the images in a pair of 10-m-resolution Sentinel-2 satellite images, which originally contains no displacement. This realistic synthetic dataset is then used to feed a model.
Our network is capable of estimating a displacement field from images whose resolution differs signifiantly from that of the training dataset (for example, from a 0.5-m-resolution Pléiades image pair) and achieves results comparable to those of state-of-the-art methods, with even finer details, at both pixel and sub-pixel resolution levels. However, the ability of machine learning to overcome limitations due to landscape changes caused by time remains to be proven.
How to cite: Delorme, A., Rupnik, E., Klinger, Y., and Pierrot-Deseilligny, M.: Machine learning to compute, from optical images, the horizontal ground displacement field caused by an earthquake, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6784, https://doi.org/10.5194/egusphere-egu26-6784, 2026.
Phase picking in seismology is the first step of signal processing and locating a seismic event. At the beginning of seismological research, it is straightforward to manually pick P- and S-wave arrival times because the number of seismic stations is relatively small. Recently, instrumentation has improved, and the gap is lower than before. This widespread instrumentation creates problems for users who still have to pick manually, especially in national networks such as those in Türkiye. We used pre-trained deep learning models, trained on different seismic datasets, to estimate P- and S-wave arrival times for earthquakes in the Marmara Region, NW Türkiye. We compared these times with manual readings collected from the Ministry of Interior, the Disaster and Emergency Management Presidency, and the Earthquake Department (AFAD). The results indicate that PhaseNet produces arrival-time estimates that are largely consistent with expert manual readings, demonstrating its potential to substantially reduce analyst workload in large-scale seismic monitoring systems. The mean absolute error ranges from 6 to 14 seconds, and the number of total picks varies between 75,000 and 140,000 for both the P-wave and S-wave. This project has been supported by the Scientific and Technological Research Council of Türkiye (TÜBİTAK) under grant 124E294, and the results have been shared on the website (https://quakemlab.sakarya.edu.tr)
How to cite: Tezel, T., Erden, C., Arıkan, H., Küçükkara, M. Y., Yanık, K., and Utkucu, M.: QuakeMLab Phase I: Deep Learning-Based Automated Seismic Phase Picking Using PhaseNet, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8055, https://doi.org/10.5194/egusphere-egu26-8055, 2026.
Seismic waveform tomography typically uses either traditional adjoint methods to compute gradients for model updates or neural-network-based (NN-based) methods to directly predict the model. However, adjoint methods require complex analytical derivations that must be reformulated for each combination of model parameters (velocity or attenuation), wave equations (elastic or viscoelastic), and misfit functions (waveform, travel time, differential time or amplitude). Here, we replace existing adjoint methods with automatic differentiation (AD), which computes accurate gradients of wave equation-based data misfits directly without any analytical derivations. Compared with NN-based methods (e.g. PINN or neural operator), our AD-based tomography framework is fully white-box and does not require any training datasets. We demonstrate both theoretically and numerically that gradients computed with AD are identical to those from adjoint methods, regardless of the domain, wave equation, or misfit function. For a field application, we apply ambient noise differential AD tomography to data from the Southeastern Suture of the Appalachian Margin Experiment (SESAME) and obtain three 2D Love-wave shear velocity (Vsh) models. The imaged Paleozoic suture zone, Mesozoic rift basins, and Moho interface are consistent with previous studies. Our results highlight the unifying role of AD in geophysical inverse problems beyond gradient computation, with promise for broader future applications across geoscience.
How to cite: Wang, X., Li, Z., and Liu, X.: A Unified Seismic Tomography Framework Using Automatic Differentiation Applied to the Southern Appalachian Array, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9203, https://doi.org/10.5194/egusphere-egu26-9203, 2026.
Deep learning based seismic elastic inversion workflows strongly benefit from realistic synthetic angle-stack seismic data, especially where field data are limited or unavailable. This study presents a stratigraphically constrained workflow for exact Zoeppritz-based amplitude variation with offset (AVO) forward modelling using augmented well-log data. The methodology integrates P-wave velocity (Vp), S-wave velocity (Vs), and density (ρ) logs from real wells with interpreted seismic horizons to generate geologically consistent synthetic elastic models and angle-dependent seismic responses. Within each stratigraphic interval defined by seismic horizons, multivariate statistics of Vp, Vs, and ρ are estimated across all available wells, preserving intrinsic elastic parameter correlations. Synthetic wells are then generated through multivariate data augmentation conditioned to these statistics and constrained by the stratigraphic framework. The resulting well-log data are used for AVO forward modelling based on the exact Zoeppritz equations, computing angle-dependent P-wave reflection coefficients at elastic interfaces. These reflection coefficients are subsequently convolved with a seismic wavelet to generate synthetic angle stacks. The proposed workflow produces consistent sets of synthetic elastic wells logs and exact Zoeppritz-based angle-stack data, providing realistic and physically grounded training datasets for seismic elastic inversion.
How to cite: Bokhonok, O., Paulo Marchezi, J., Alexandra Delgado Blanco, L., Pereira Leite, E. P., Hartman, G., and Batezelli, A.: Stratigraphically Constrained Well-Log Data Augmentation for Angle-Stack AVO Forward Modelling, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10510, https://doi.org/10.5194/egusphere-egu26-10510, 2026.
The characterization of reservoir properties from seismic data is often hindered by the scarcity and spatial bias of well-log data. To overcome these limitations, data augmentation (DA) has become essential for training robust Deep Learning models. This study presents a comparative analysis of three distinct DA approaches: Statistical Methods, Variational Autoencoders (VAE), and Generative Adversarial Networks (GAN. We synthesize well-log suites for training a UNet architecture dedicated to seismic-to-porosity prediction. Our workflow begins with real well logs, expanding the dataset through stochastic perturbations (Statistical), latent manifold sampling (VAE), and adversarial learning (GAN). To bridge the gap between 1-D well data and seismic volumes, we perform forward modeling on the augmented suites, generating synthetic seismograms via convolution with representative wavelets. A UNet-based convolutional neural network is then trained on these synthetic pairs to perform the non-linear mapping from seismic amplitudes to porosity. The performance of each method is evaluated through the geological plausibility of the generated logs and the inversion accuracy on a blind-test well. Preliminary results indicate that while statistical methods improve robustness against noise, generative models, particularly GANs, excel in capturing the multi-scale heterogeneity required for high-resolution reservoir characterization. This research demonstrates that the choice of DA is a critical geophysical decision; by integrating generative AI into the inversion workflow, we provide a scalable framework to improve porosity estimation in data-poor environments, ensuring that synthetic extensions remain grounded in petrophysical reality and stratigraphic consistency.
How to cite: Marchezi, J. P., Leite, E., Bokhnok, O., Delgado, L., Hartman, G., and Batezelli, A.: Impact of Generative AI and Statistical Data Augmentation in Synthetic Well-Log Generation for Seismic Porosity Inversion: A Comparative Study, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10548, https://doi.org/10.5194/egusphere-egu26-10548, 2026.
Deep learning has improved automated seismic phase picking in recent years. However, many pickers are trained on a single dataset and often fail to generalize when deployed on out-of-distribution (OOD) data. Variations in earthquake magnitude, propagation distance, sensor instrumentation, and ambient noise between training and target domains lead to significant performance degradation under OOD conditions. To address this challenge, we propose a new framework to reduce performance degradation under OOD conditions based on a pipeline of seismic simulation -- phase labeling -- site-conditions simulation -- transfer learning. First, we use the seismic simulation tool AxiSEM3D to establish a 3-D waveform simulation, covering local, regional, and teleseismic scales. Next, we obtain precise P- and S-phase labels by computing theoretical arrival times from velocity models and refining these onsets with traditional automatic picking algorithms, ensuring high-fidelity phase annotations. Then, based on actual site conditions, we simulate instrument responses and synthesize ambient noise constrained by PPSD analysis. This gives us station-specific, noisy waveforms that closely match real observational conditions. Finally, we employ transfer learning to fine-tune a phase picker on this specific synthetic dataset, thereby enhancing the picker's performance in new conditions. The proposed framework aims to improve the ability of deep learning models to pick phases under OOD conditions. It enables reliable performance across regional variability and instrumentation differences without large-scale manual relabeling. It also reduces the amount of real data needed for training, making it useful for small datasets and leaving more data to analyse. Overall, this work introduces a transferable methodology for seismic phase picking under distribution shift and shows that physics-informed data augmentation combined with targeted transfer learning can effectively decrease OOD performance degradation, thereby increasing the applicability of deep-learning-based phase picking.
How to cite: Jiang, P., de Laat, J. I., Wang, X., and Fadel, I.: Earthquake phase picking in out-of-distribution data conditions: Can synthetic data from earthquake simulations help?, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10617, https://doi.org/10.5194/egusphere-egu26-10617, 2026.
Mount Etna, unlike many volcanoes that experience prolonged calm intervals, exhibits persistent and continuous activity characterized by frequent strombolian bursts, lava fountains, and effusive events. This study aims to automatically identify recurrent and distinctive patterns in seismic signals by extracting clusters of waveforms with similar spectral characteristics using a fully data-driven, unsupervised machine learning framework, and to assess their correspondence with observed volcanic activity.
We analyzed daily seismic spectrograms from two summit seismic stations, ECPN and ECNE, spanning November 2020 to November 2021, a period encompassing both quiet intervals and two major lava fountain cycles. For dimensionality reduction and feature extraction, we employed AutoencoderZ, an encoder–decoder model with skip connections, convolutional and fully connected layers, a bottleneck latent space, and transposed convolutions. This architecture compresses inputs while preserving critical spectral features for unsupervised clustering. Extracted features are optimized using the Relative Bias metric and clustered via Deep Embedded Clustering (DEC), enabling data-driven anomaly detection and pattern recognition by clustering similar waveforms.
The resulting clusters were compared with independent observational datasets and seismic related metrics, including lava fountain records, volcano-tectonic and long-period (LP) event catalogs, and root mean square (RMS) amplitude of volcanic tremor. This comparison demonstrates the approach’s ability to uncover hidden structures in the seismic data and highlight key temporal transitions associated with underlying processes such as magma and fluid dynamics . To improve robustness and reduce potential spatial bias, analyses were conducted using both single-station and dual-station approaches, providing a more reliable characterization of seismic variability.
Overall, this study highlights AutoencoderZ’s versatility in revealing complex patterns in Etna’s seismic activity.
How to cite: Abed, W., Zali, Z., Sciotto, M., Cocina, O., Cannata, A., Picozzi, M., Martínez-Garzón, P., Vuan, A., Saraò, A., and Sugan, M.: Understanding Volcanic Seismic Patterns with Unsupervised ML Clustering: Application at Mount Etna volcano, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11584, https://doi.org/10.5194/egusphere-egu26-11584, 2026.
Understanding earthquake rupture directivity is crucial for constraining source physics and improving seismic hazard assessment. In an ideal setting, one would analyze far-field seismograms free from subsurface scattering, travel-time uncertainties, and ambient noise, enabling direct inference of rupture directivity from the observed waveforms. In practice, however, recorded seismograms are strongly influenced by path effects, site responses, and additive noise, while station azimuthal coverage is often sparse and uneven. These limitations significantly complicate directivity analysis, particularly for 1) low-magnitude noisy events and 2) earthquakes with complex rupture processes, for which simple source–path deconvolution models are inadequate.
We employ a conditional Diffusion Transformer (DiT) to learn the dependence of far-field seismograms on the P-ray take-off direction. The DiT is trained on measured far-field seismograms from multiple earthquakes with moment magnitudes Mw ≥ 6.0. Once trained, the model generates virtual far-field seismograms conditioned on specified ray azimuth and take-off angle, while holding an empirical realization of the path effects fixed by conditioning on a reference observed seismogram. This enables controlled experiments in which variations attributable to source directivity can be examined independently of path-induced variability. In this sense, our approach closely mimics the idealized setting in which far-field seismograms vary only with source directivity and are free from complex path effects. In other words, this generative framework enables us to isolate and examine variations in the wavefield that are attributable solely to source directivity, while holding path effects constant. We demonstrate that this approach is particularly effective for earthquakes with complex multi-episode moment release. For all events considered, the generated wavefields vary smoothly with take-off direction, indicating physical consistency. Importantly, the DiT training is self-supervised, requiring neither synthetic earthquake simulations nor explicit correction for path effects. The proposed framework provides a scalable and physically consistent tool for investigating earthquake directivity and rupture complexity across a wide range of magnitudes.
How to cite: Gupta, P. and Bharadwaj, P.: Understanding the directivity of earthquakes by generating virtual far-field seismograms conditioned on the P-ray take-off direction , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18002, https://doi.org/10.5194/egusphere-egu26-18002, 2026.
In modern seismology, the rapid and accurate characterization of seismic sources following an earthquake is a fundamental task. Most observatories currently compute moment tensor solutions for events exceeding specific magnitude thresholds to ensure reliability. For instance, the National Institute of Geophysics and Volcanology (INGV) provides routine solutions for moderate to large events (Mw≥3.5) using established methods (Scognamiglio et al., 2009). However, characterizing smaller events remains challenging due to low signal-to-noise ratios and the need of modeling high-frequency waveforms that requires detailed knowledge of the velocity model, which is rarely available.
Machine learning (ML) techniques have emerged as powerful tools to address these limitations, particularly in improving the prediction of first-arrival seismic wave polarities. These ML-derived polarities can be effectively integrated into traditional frameworks to compute robust focal mechanisms. In this study, we implement a workflow that bridges deep learning polarity predictions with standard focal mechanism estimation techniques, focusing on the tectonic setting of the Italian Peninsula.
Our methodology consists of two primary stages. First, we trained a Convolutional Neural Network (CNN) for polarity prediction using the INSTANCE catalog (Michelini et al., 2021), which provides the high-quality, manually reviewed data essential for supervised learning. Second, we validated the model’s performance by analyzing approximately 4,700 earthquakes that occurred in Italy between January 1 2021, and January 1 2025, with magnitudes below 4.5.
To benchmark our results, we selected a subset of earthquakes with existing Time Domain Moment Tensor (TDMT) solutions (Scognamiglio et al., 2006). Using the polarities predicted by the CNN, we computed focal mechanisms using the SKHASH code (Skoumal et al., 2024). The accuracy of these solutions was then evaluated against the TDMT catalog using Kagan angle analysis (Kagan, 1991) to quantify the rotation between double-couple sources.
How to cite: Tavani, F., Scognamiglio, L., Artale Harris, P., and Meier, M.-A.: Source Characterization via Deep Learning: Integrating Polarity Prediction and Focal Mechanism Estimation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18532, https://doi.org/10.5194/egusphere-egu26-18532, 2026.
In a previous work, we established the predictive power of seismic statistical catalogue features for whole-region temporal forecasting. We here extend this framework to a spatiotemporal approach to assess localised seismic hazard in Japan and Chile. Using ensemble learning, we predict the occurrence of M>=5 earthquakes within a 15-day horizon across varying radial distances (r=3 to 24 km) to benchmark the framework's sensitivity as a proof-of-concept prior to scaling for larger magnitude hazards.
Results indicate robust predictive power, though performance is sensitive to the prediction radius. The Japan catalogue yields an AUC of 0.76 for predictions within 24 km. However, when the prediction radius is tightened to 12 km, while the model retains predictive power (AUC 0.62), the reduced performance underscores the challenge of highly localised forecasting. Crucially, we observe a distinct shift in feature importance as the spatial scale changes: parameters that track local variations in seismicity—specifically the b-value, within our feature set—rank significantly higher in localised models compared to whole-region baselines. This suggests that machine learning models can produce forecasts that reflect underlying physical fault processes.
We further present ongoing work regarding spatiotemporally overlapping predictions, testing the hypothesis that multiple alerts intersecting in both space and time indicate a compounded hazard probability. Finally, responding to the challenges of localised prediction, we introduce a novel experimental framework that augments our current statistical features by exploring additional spatial descriptors, including both deep learning representations and hand-crafted spatial features, designed to capture aspects of fault dynamics beyond standard catalogue statistics.
How to cite: Quan, W. and Gorse, D.: Leveraging the value of seismic catalogue features in building a spatiotemporal system to assess localised seismic hazard, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7938, https://doi.org/10.5194/egusphere-egu26-7938, 2026.
Earthquake catalogs are essential for monitoring natural hazards and improving our understanding of seismic processes, which depend on accurate detection and classification of both natural earthquakes and anthropogenic signals. This is especially important in intraplate regions like Norway characterized by low to moderate earthquake activity, rare impactful earthquakes and widespread anthropogenic events such as quarry blasts. Nonetheless, traditional detection workflows have struggled to keep pace with the growing number of seismic stations and temporary deployments. Recent advances in machine learning offer promising solutions for efficient detection and discrimination tasks. Here, we present preliminary results toward a fully automated seismic detection and classification system for the Norwegian National Seismic Network (NNSN).
We first evaluate pre-trained deep learning-based phase detection models PhaseNet (Zhu & Beroza 2019) and Earthquake Transformer (EQT, Mousavi et al. 2020) using a catalog of 2567 events from the NNSN bulletin from 2008 to 2025, which includes 1144 earthquakes and 1423 blasts. We find the models can detect earthquake phases with F1-scores of 0.70 and 0.67, for the PhaseNet and EQT respectively, and 0.73 and 0.70 for blasts, revealing slightly higher sensitivity to blasts than earthquakes.
Building on this, and with the goal of developing a robust deep-learning earthquake detection workflow, we set out to quality-control our classification of earthquakes and blasts. We apply a self-supervised approach based on Bootstrap Your Own Latent (BYOL, Grill et al. 2020), which learns representations by aligning augmented views of the same signal, enabling learning without relying on potentially biased labels. The only information provided to the model are the training hyperparameters and the number of classes we aim to identify (two: earthquakes and blasts). Our method achieves an F1-score of 0.93 in distinguishing blasts from earthquakes using self-supervised representations, and up to 0.94 when incorporating an additional supervised layer. Analysis of the results reveals previously misclassified events, demonstrating the effectiveness of self-supervised methods even with limited or biased labeled datasets.
Future work will focus on retraining detection models using NNSN data after BYOL based classification. Additionally, we will analyze the BYOL learned features to gain insights on the physical differences between earthquake and blast signals.
How to cite: Neves, M., Ottemöller, L., and Rondenay, S.: Exploring Deep Learning Approaches for Seismic Detection and Discrimination in Norway, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10618, https://doi.org/10.5194/egusphere-egu26-10618, 2026.
Submarine volcanism represents one of the dominant forms of magmatic output on Earth, yet our understanding of the underlying eruptive processes remains limited compared to subaerial systems. The remote locations and lack of direct visual observations often obscure the dynamics of these eruptions, where the interaction with the hydrostatic load and phase changes (steam/water) creates distinct physical regimes. In the absence of near-field observations, far-field hydroacoustic records—propagated over thousands of kilometers in the SOFAR channel—provide a critical, high-temporal-resolution window into these deep-sea events. However, a significant challenge lies in processing the high volume of continuous data to categorize signals into physically meaningful regimes without relying on manual classification or subjective thresholds.
In this study, we present Spec2Vec, a novel framework for the unsupervised classification of hydroacoustic time series, applied to the major 2021 Fukutoku-Oka-no-Ba shallow submarine eruption. Current unsupervised approaches in geo-acoustics often rely on decomposition methods like Non-Negative Matrix Factorization (e.g., SPECUFEX), Independent Component Analysis (ICA), the scattering transform, or latent representations from neural network auto-encoders. While effective, these methods can be computationally intensive or result in "black box" features lacking direct physical intuition. In contrast, Spec2Vec utilizes Hilbert space-filling curves to topologically map 2D time-frequency representations into a 1D sequence, strictly preserving multi-scale locality. From this linearized stream, we extract a compact set of entropy and scaling features. This approach captures the "texture" of the spectrogram—the specific arrangement of energy in time and frequency—more uniquely and efficiently than standard modal image features. The resulting feature space is fast to compute and highly interpretable, bridging the gap between raw acoustic data and physical source mechanics.
We evaluate this feature set by applying it to ten days of continuous hydroacoustic data from the Comprehensive Nuclear-Test-Ban Treaty Organization (CTBTO) International Monitoring System (IMS), capturing the main eruption sequence as well as pre- and post-eruptive phases. Through unsupervised learning, Spec2Vec automatically organizes the complex acoustic stream into coherent clusters. To validate the physical interpretability of these clusters, we correlate the unsupervised classes with independent eruptive proxies, including satellite-derived lightning data, plume height, and mass eruption rates. Furthermore, we inject synthetic acoustic source models—simulating single bubble oscillations, turbulent jets, bubble plumes, hydroacoustic earthquakes, explosions, and volcanic tremor—into the dataset to map clusters to specific source mechanisms.
Our results offer a rare, data-driven characterization of the Fukutoku-Oka-no-Ba eruption, identifying distinct phases of jetting and tremor that align with atmospheric observations. This demonstrates that Spec2Vec serves not merely as a feature generation tool, but as a generalizable engine for the automated discovery of physical processes in complex geophysical time series. This approach holds significant potential for scaling the analysis of global hydrophone datasets, enabling the systematic distinction and quantification of eruption rates and processes across the global submarine volcanic inventory.
How to cite: Swar, S., Mittal, T., and Olugboji, T.: Decoding the Dynamics of the 2021 Fukutoku-Oka-no-Ba Submarine Eruption via Interpretable Spectral-Hilbert Representations (Spec2Vec), EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14499, https://doi.org/10.5194/egusphere-egu26-14499, 2026.
In this study we introduce PHASER, a seismic event picker developed specifically for downhole microseismic applications. In recent years, DL models have been extensively explored for automating phase-picking and/or event detection for seismic data, with most applications focusing on teleseismic/regional earthquake signals from surface arrays. Surface arrays require a generalizable solution across different array geometries. In contrast, downhole arrays used to monitor industrial activities such as geothermal, CCS and hydraulic fracturing are more standardized in receiver placement, but present unique challenges and opportunities. Seismic phases arrive coherently on closely-spaced downhole geophones. However, access to downhole data is also often limited, and such data are often available only as event-based traces, and labels are often incomplete or inaccurate. To address these constraints, PHASER is trained in a multi-stage framework that remains effective and generalizable even when catalogue labels are incomplete and limited. PHASER incorporates association filtering into its training; pick probabilities are matched with their respective association probabilities based on learned source-related feature embeddings for each P- and S- arrival. By using a learned extraction threshold, PHASER avoids the manual parameter tuning typically required for pick extraction. PHASER demonstrates better continuous monitoring performance on unseen sites than existing DL phase-pickers, achieving a 6-fold performance over PhaseNet in F1 score from 0.107 to 0.584 on an out of sample test dataset.
How to cite: Leung, K., Verdon, J., and Werner, M.: PHASER: A Deep Learning Model for Real-time Downhole Microseismic Event Picking, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3133, https://doi.org/10.5194/egusphere-egu26-3133, 2026.
