Improvements in monitoring technologies now deliver seismic data of unprecedented quality and quantity. Earthquake catalogues are more complete and accurate than ever, and many are now publicly available, enabling analysing understudied regions and expanding global knowledge. New-generation catalogues, sometimes compiled with machine learning, reveal seismicity structures in ways not previously possible.
Additionaly, geodetic, geological, and geochemical data, fluid analyses, laboratory experiments, and earthquake simulators generating synthetic catalogues help refine models and test hypotheses. Integrating such multidisciplinary perspectives enhances our understanding of earthquake generation.
To exploit these datasets, statistical approaches and machine learning are essential. These tools uncover hidden relationships and clustering, and address challenges of data inhomogeneity, paving the way for deeper understanding and robust forecasting.
We invite contributions on developments in physical and statistical modelling and machine learning, including:
• Spatial, temporal, and magnitude properties of earthquake statistics
• Earthquake clustering analyses
• Effects of fluid diffusion and geodetic deformation on seismicity
• Physical and statistical models, including for understudied regions (e.g., Africa, Southeast Asia)
• Quantitative testing of models
• Data requirements and analyses for validation
• Machine learning applied to seismic data
• Uncertainty quantification in pattern recognition and machine learning
• Reliability and completeness of catalogues
• Time-dependent hazard assessment
• Software and methods for earthquake forecasting
Orals: Wed, 6 May, 14:00–18:00 | Room 1.15/16
State-of-the-art machine learning models for seismic event detection, such as EQTransformer and PhaseNet, use supervised learning, which requires labeled event catalogs and curated waveforms. This dependence creates two fundamental limitations: the cost of preparing high-quality datasets, and an annotation bias which limits the model training to event types well represented in existing catalogs. Unsupervised deep learning has the potential to overcome these limitations, but despite its prevalence in other domains, this approach remains rather under-explored for the problem of seismic event detection.
We present RECOVAR, an unsupervised deep learning method that trains directly on continuous waveform data, requiring no labeling or catalog preparation. The architecture consists of an ensemble of convolutional autoencoders, each trained independently. Detection exploits how these latent representations differ for signal versus noise: coherent seismic arrivals produce convergent representations with high cross-covariance, while stochastic noise produces uncorrelated representations.
Since continuous recordings are dominated by noise, a naive approach to training on continuous waveforms ends up creating a model that focuses excessively on representing noise, and results in quite good but suboptimal event detection. We introduce a dynamic training pipeline that preferentially resamples low-scoring segments using the model's own cross-covariance scores, which results in strong detection performance.
RECOVAR achieves event detection ROC AUC scores of 0.97-0.99 on the STEAD and INSTANCE benchmarks, comparable to PhaseNet and EQTransformer. We demonstrate a regional application to the 2019 Istanbul Silivri earthquake sequence, training directly on continuous waveforms without any catalog preparation. We show the utility of RECOVAR as a post processing tool that filters picks by supervised methods, retaining 99% of true picks by PhaseNet while filtering half of the false positives, and with less conservative settings, removing 83% of false positives while retaining 84% of true detections.
RECOVAR provides an unsupervised deep learning alternative for seismic detection. Training directly on continuous data without labels avoids the annotation bias that is inherent to supervised methods, which potentially opens the door to detecting rare event types absent from established catalogs. As demonstrated by its post-filter performance, RECOVAR also integrates naturally within existing detection pipelines.
How to cite: Adıgüzel, E., Efe, O., Ozakin, A., Konca, A. O., and Ergintav, S.: RECOVAR: An unsupervised deep learning approach to seismic event detection by training on continuous waveform data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18336, https://doi.org/10.5194/egusphere-egu26-18336, 2026.
1 Research Objectives and Methods
To clarify the sequence characteristics and post-earthquake trend of the Ms 7.3 earthquake in the sea area off Hualien County, Taiwan, on April 3, 2024 (focal depth 12 km, epicenter at 23.81°N latitude and 121.74°E longitude), this study is based on observational data from the China Earthquake Networks Center (CENC), combined with the regional geological and tectonic background. Selecting seismic catalog data from November 2023 to December 2024, systematic analysis was conducted on the spatiotemporal distribution, intensity variation, frequency characteristics and dynamic evolution law of the earthquake using analytical tools such as M-T diagrams, H-T diagrams, h-value curves, b-value curves and creep curves.
2 Study Area Characteristics and Seismic Sequence Analysis
2.1 Tectonic setting and spatial distribution of seismic activity: The study area is located at the subduction boundary between the Eurasian Plate and the Philippine Sea Plate, controlled by the tectonic background of the northern segment of the Huadong Valley Fault Zone. Seismic activity features a spatial pattern of "stronger and denser in the east, weaker and sparser in the west". Epicenters are concentrated within the range of 121-123°E longitude and 23-25°N latitude, showing an overall northeast-southwest trend.
2.2 Seismic sequence type and source characteristics: The earthquake sequence is a typical mainshock-aftershock type, with the mainshock releasing 98.2% of the total energy of the sequence. Aftershocks are active after the mainshock, and their attenuation follows the modified Omori formula. Shallow-focus earthquakes (0-50 km) dominate, which are highly destructive; a small number of deep-focus earthquakes also occur, reflecting stress adjustment processes at different crustal levels.
2.3 Seismic sequence parameter analysis: Analysis of the seismic sequence parameters reveals that the post-earthquake h-value is 1.1 (faster than the conventional attenuation rate), and the b-value is 1.0166 (with a higher proportion of small and medium-sized earthquakes). There is a significant linear correlation between magnitude and logarithmic frequency, consistent with the Gutenberg-Richter law. The creep curve clearly shows a three-stage evolutionary characteristic: "strain accumulation — mainshock release — post-earthquake adjustment".
3 Post-Earthquake Trend Determination and Research Significance
3.1 Post-earthquake trend judgment: The intensity and frequency of aftershock activity will continue to attenuate, and the probability of a magnitude 7.0 or above strong earthquake occurring in the short term (within several months) is extremely low. However, deep-seated stress adjustment in the region is not yet complete; special attention should be paid to stress transfer in the unruptured area of the northern segment of the Huadong Valley Fault Zone to prevent the occurrence of delayed strong aftershocks.
3.2 Research significance: The conclusions of this study provide scientific support for the research on earthquake mechanisms and disaster prevention and control work in eastern Taiwan.
How to cite: Wu, B.: Characteristics of the Seismic Sequence and Determination of Post-Earthquake Trends for the MS 7.3 Earthquake in the Sea Area Off Hualien County, Taiwan, April 3, 2024, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2107, https://doi.org/10.5194/egusphere-egu26-2107, 2026.
We applied the machine-learning–based probabilistic forecasting algorithm NESTORE (NExt STRong Related Earthquake) to the seismicity of New Zealand. NESTORE analyses nine features describing aftershock occurrence, source area evolution, and temporal trends in magnitude and radiated energy, computed over progressively increasing time windows following the mainshock. These features enable the algorithm to estimate the probability that a mainshock of magnitude Mm will be followed by a subsequent event of magnitude ≥ Mm–1 within the space-time domain of the associated eismic cluster. Clusters in which such a strong aftershock occurs are classified as “Type A,” indicating higher potential hazard, while others are classified as “Type B.” For each cluster, the algorithm outputs the corresponding probability of belonging to Type A.
New Zealand’s position along the boundary between the Australian and Pacific plates results in widespread, complex deformation and a relevant seismic activity, including major events up to magnitude 7.8. This setting makes the region an ideal testing ground for operational, data-driven forecasting tools such as NESTORE. Understanding and forecasting seismic activity is critical for rapid hazard assessment and mitigation efforts.
To evaluate NESTORE’s performance, we employed two testing strategies. The first was a chronological approach, in which the algorithm was trained using seismic clusters occurring before a chosen cutoff time and then used to retrospectively forecast cluster behaviour after that time. The second approach employed stratified k-fold cross-validation, allowing us to assess model generalization across multiple randomized data partitions. To further enhance training quality, we applied the outlier-detection procedure REPENESE (RElevant features, PErcentage class weighting, NEighborhood detection and SElection).
Our results show that the k-fold validation approach provides a more robust and stable performance evaluation than the chronological approach, although changes in the catalogue may make the more recent clusters a more reliable test set. NESTORE correctly classified 88% of seismic clusters 18 hours after the mainshock, including 77% of Type A clusters and 92% of Type B clusters. Notably, the Canterbury/Christchurch 2010–2011 sequence, a critical and highly destructive Type A cluster, was correctly classified by the algorithm.
Overall, the results of this work underscore the potential for use of NESTORE for short-term aftershock forecasting in New Zealand.
How to cite: Caravella, L. and Gentili, S.: Forecasting strong aftershocks in New Zealand with the machine-learning NESTORE algorithm: two different testing approaches, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-524, https://doi.org/10.5194/egusphere-egu26-524, 2026.
Foreshocks, aftershocks, and swarms are common types of seismicity clusters. Foreshock patterns are recognized as of special value for earthquake forecasting. Beforehand discrimination of foreshocks from other clusters and from background seismicity is of great importance for short-term hazard assessment, but it remains a challenge. A promising prospect is that different seismicity clusters are characterized by distinct patterns in space, time, and magnitude, thus reflecting different underlying geophysical processes. In foreshocks, the count number increases at the inverse of time, but usually an activity lull is observed a few days before the mainshock; the b-value drops, while epicenters usually move towards the mainshock epicenter. In aftershocks, the epicenters expand away from the mainshock epicenter, the event count decreases exponentially with time, and the b-value increases. Swarms are not associated with specific patterns of epicentral and temporal distributions, while the b-value usually increases. On-time identification of statistically significant seismicity changes could be supportive towards real-time discrimination between different types of clusters. This approach was tested with the seismic sequence of the Mw8.8 megathrust mainshock that ruptured the subduction interface off eastern Kamchatka on 29 July 2025, based on classic earthquake statistics and advanced complex network tools. On 20 July 2025 an earthquake of Mw7.4 occurred; many smaller shocks followed. However, the foreshock sequence was only recognized a posteriori. We investigated if the foreshock sequence could be detectable beforehand. To examine this crucial issue, we introduced the concept of Virtual Real-Time (VRT) analysis, which is different from usual retrospective analysis because VRT utilizes incomplete knowledge of the earthquake sequence, i.e., the catalogue and other data available only up to each point of time T of the ongoing seismic sequence. This means the analysis is performed as if we were in the actual conditions of the sequence. VRT analysis was combined with a decision matrix based on the different patterns of different clusters and on testing appropriate null hypotheses. Considering 20 July 2025 as Τf=1 day, the VRT analysis detected the transition from the state of background seismicity to that of foreshocks on Τf=3 (23 July), if not earlier, and persistently on every subsequent day prior to the mainshock up to Τf=9 days (29 July). The imminence of an even larger earthquake became evident from the foreshock lull in about Τf=7 days, while its magnitude was approximated by an empirical relationship between magnitude and the area covered by the foreshocks. Setting the mega earthquake at time Τa=1 day, the transition from the state of foreshocks to that of aftershocks was detectable at Ta=2 days and at every subsequent day, thus signifying that the mega earthquake was the mainshock. All seismicity changes from one state to the other were found to be highly significant. The results obtained underline the important capabilities for earthquake forecasting from the recognition of foreshocks beforehand. The data used in this work were obtained from the large-scale research facilities «Seismic infrasound array for monitoring Arctic cryolithozone and continuous seismic monitoring of the Russian Federation, neighboring territories, and the world» (https://ckp-rf.ru/usu/507436/).
How to cite: Triantafyllou, I., Zavyalov, A., Papadopoulos, G., Siettos, C., and Spiliotis, K.: Virtual Real-Time (VRT) forecasting of the Kamchatka 29 July 2025 mega earthquake (Mw8.8) based on foreshock activity, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16933, https://doi.org/10.5194/egusphere-egu26-16933, 2026.
Fault zone properties evolve throughout the seismic cycle, reflecting variations in stress conditions and progressive damage. Recent studies applying explainable machine learning to the 2016–2017 Norcia, central Italy, earthquake sequence have demonstrated that these temporal variations can be detected directly from seismic waveforms (Laurenti et al. 2024 doi.org/10.1038/s41467-024-54153-w). Here, we extend this approach to investigate whether similar signatures can be identified at a different spatial and magnitude scale.
In this work, we study a microseismic sequence close to the village Diemtigen in central Switzerland that occurred between April 2014 to September 2015. The dataset includes 4 main events with magnitudes between ML 2.7 and 3.2, along with the earthquakes recorded before and after each main event. The high-resolution dataset was assembled using template-matching analysis (Simon et al., 2021 doi.org/10.1029/2021GL093783).
We train a convolutional neural network (CNN) to classify foreshocks and aftershocks, and we use SHapley Additive exPlanations (SHAP) to interpret the results. The CNN is trained on spectrograms derived from raw waveforms. SHAP provides pixel-level attribution maps for each spectrogram, allowing us to identify which frequency-time components contribute most to the predictions. The CNN distinguishes between seismic traces before and after a main event, even if the waveform is pure seismic noise, without any earthquake recording. When classifying earthquake traces, SHAP analysis highlights key features in foreshocks in correspondence to the P-S arrival in the frequency range of 30-40 Hz. This observation is consistent with previous results from the Norcia earthquake sequence (Magrini et al. 2026 doi.org/10.1007/978-3-032-10185-3_25), where the same method identified comparable time-frequency features associated with foreshock activity.
This framework offers new physics-based insight into the evolution of fault zones. It demonstrates the potential of Explainable AI to complement classical earthquake sequence analysis by revealing subtle, physically meaningful signatures directly from seismic data, and thereby bridging data-driven approaches with seismological understanding.
How to cite: Laurenti, L., Simon, V. M., Toledo Zambrano, T. A., Kraft, T., Meier, M.-A., Magrini, M., Marrocco, F., Paoletti, G., Tinti, E., and Marone, C.: Analysis of Foreshocks and Aftershocks in a microseismic sequence in Switzerland using Explainable AI, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13268, https://doi.org/10.5194/egusphere-egu26-13268, 2026.
The statistical analysis of the recently published high-resolution seismic catalogue for Campi Flegrei (January 2022–March 2025, Tan et al. 2025) reveals that deep-sourced degassing controls Campi Flegrei seismicity, illuminating pathways to the surface along a subset of permeable structures and generating seismicity only in specific volumes. Analysing the catalogue using machine learning cluster analysis to identify objective volumetric seismicity sources, two main seismogenic volumes emerge: a deep ring of cigar-shaped 1D source volumes, and a cloud of shallower 1D/3D source volumes connecting the ring's northern sector to the surface. The found clusters were compared with other existing information about the caldera structure (e.g. known faults, deep and surficial tomography studies of different nature, geochemical data), showing that ring seismicity encircles a potential primary volcanic source (main degassing zone) and occurs at the intersection between pre-existing faults and a sub-horizontal south-dipping rheological interface, while the cloud track the main gas plumes detaching from the ring and infiltrating through faults into the shallowest seismic volumes below Accademia, Solfatara-Piscarelli and Rione Terra. Interesting spatio-temporal variations in the rate of activity of the different sources seem to track pressurization cycles, leading to the activation of new volumes during high activity periods.
Xing Tan, A. Tramelli, S. Gammaldi, G.C. Beroza, W.L. Ellsworth, W. Marzocchi, A clearer view of the current phase of unrest at Campi Flegrei caldera. Science 390, 70-75 (2025). doi:10.1126/science.adw9038
How to cite: Selva, J., Piegari, E., Natale, J., Vitale, S., Chiodini, G., Caliro, S., and Marzocchi, W.: Clues on the ongoing unrest at Campi Flegrei from the high-definition seismic 2022-2025 catalogue, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11349, https://doi.org/10.5194/egusphere-egu26-11349, 2026.
Earthquake sequences exhibit intricate space–time–magnitude patterns that have motivated the use of statistical methods to uncover properties and relationships that standard time-series analysis techniques are unable to capture. Although these methods have made it possible to highlight properties such as clustering, scaling, long-range dependencies and related features, appropriate analyses of the complexity of the seismic phenomenon have not yet been developed or applied.
Since Bandt and Pompe’s seminal work, the permutation entropy and statistical complexity form the basis for constructing the so-called complexity–entropy causality plane (CECP). Permutation entropy and statistical complexity provide insight into two different aspects of a dataset. Permutation entropy measures the level of intrinsic randomness: data that are more predictable and tend to repeat a limited number of ordinal patterns exhibit lower permutation entropy, whereas data with a greater variety of patterns and less predictability show higher values. For a fixed value of permutation entropy, statistical complexity indicates the extent to which certain ordinal patterns are favored over others. In other words, higher complexity—at a given entropy level—reflects a greater deviation from a uniform distribution, suggesting that some ordinal patterns occur more frequently than others. By computing both measures for a time series, one can simultaneously assess the randomness of the data and the degree of structural or correlational organization within its fluctuations.
While the CECP has been widely used to investigate the complex patterns of continuous time series, it has yet to be applied to analyze point processes, particularly in the context of seismic events. Thus, the present paper aims at analyzing the dynamics of seismic point processes in the CECP, offering new insights into their underlying patterns and behaviors.
We first analyzed in the CECP the magnitude series generated by the physics-based numerical model developed by Olami, Feder, and Christensen (OFC) in 1992. Although introduced several decades ago, the OFC model remains a robust framework, successfully reproducing key qualitative features of real-world seismicity, such as the Gutenberg-Richter law, the Omori law, and the Ruff–Kanamori diagram.
We further investigated magnitude sequences from Italian seismic regions affected by the strongest earthquakes since 1985. Our results indicate that these magnitude sequences display in the CECP a pattern that aligns very well with that observed in the OFC model and apparently correlated with the magnitude of the strongest events.
Although preliminary, these results underscore the potential of CECP analysis for seismicity studies, providing new and diverse ways to describe, interpret, and explore earthquake dynamics.
How to cite: Telesca, L.: Permutation Entropy and Statistical Complexity Analysis of earthquake sequences, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3734, https://doi.org/10.5194/egusphere-egu26-3734, 2026.
How to cite: Morales-Yañéz, C., Benavente, R., and Bonilla-Contreras, F.: b-transD: Spatial and temporal variation of b-value and their uncertainty using Bayesian trans-dimensional inference, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8613, https://doi.org/10.5194/egusphere-egu26-8613, 2026.
The b-value of the Gutenberg-Richter law is crucial for modern hazard models and seismicity forecasting. It quantifies the relative frequency of small earthquakes vs. infrequent large events. A growing number of studies suggest that the b-value changes with factors such as time (Gulia et al., 2018), differential stress (Scholz, 2015), and thermal regime (Nishikawa and Ide, 2014). However, translating the knowledge of such b-value variation into measurable improvements of earthquake forecasting capabilities has not been convincingly achieved yet (e.g., Iturrieta et al., 2024).
In this work, we investigate whether a temporally changing b-value can improve our ability to forecast future magnitudes. For this, we implement a method to estimate temporally and spatially changing b-values, with a given time- and length-scale, together with a measure of how strong the variation is (b-significant, Mirwald et al., 2024). Further, we develop a method to evaluate the information gain (IG) that is more robust in the presence of short-term aftershock incompleteness.
We apply these methods to the 2016-2017 central Italy earthquake sequence, using a machine-learning-enhanced earthquake catalog containing >900k events (Tan et al., 2021). Specifically, we first estimate the optimal temporal, spatial, and combined spatiotemporal scales for forecasting future seismicity using the first half of the dataset. Using the second half of the dataset, we then assess pseudoprospectively if a varying b-value, estimated with the parameters obtained in the first step , results in a positive information gain compared to a stationary reference model.
References
Gulia, L., Rinaldi, A.P., Tormann, T., Vannucci, G., Enescu, B., Wiemer, S., 2018. The Effect of a Mainshock on the Size Distribution of the Aftershocks. Geophysical Research Letters 45, 13,277-13,287. https://doi.org/10.1029/2018GL080619
Iturrieta, P., Bayona, J.A., Werner, M.J., Schorlemmer, D., Taroni, M., Falcone, G., Cotton, F., Khawaja, A.M., Savran, W.H., Marzocchi, W., 2024. Evaluation of a Decade-Long Prospective Earthquake Forecasting Experiment in Italy. Seismological Research Letters 95, 3174–3191. https://doi.org/10.1785/0220230247
Mirwald, A., Mizrahi, L., Wiemer, S., 2024. How to b -Significant When Analyzing b -Value Variations. Seismological Research Letters. https://doi.org/10.1785/0220240190
Nishikawa, T., Ide, S., 2014. Earthquake size distribution in subduction zones linked to slab buoyancy. Nature Geosci 7, 904–908. https://doi.org/10.1038/ngeo2279
Scholz, C.H., 2015. On the stress dependence of the earthquake b value. Geophysical Research Letters 42, 1399–1402. https://doi.org/10.1002/2014GL062863
Tan, Y.J., Waldhauser, F., Ellsworth, W.L., Zhang, M., Zhu, W., Michele, M., Chiaraluce, L., Beroza, G.C., Segou, M., 2021. Machine-Learning-Based High-Resolution Earthquake Catalog Reveals How Complex Fault Structures Were Activated during the 2016–2017 Central Italy Sequence. The Seismic Record 1, 11–19. https://doi.org/10.1785/0320210001
How to cite: Mirwald, A., Mizrahi, L., Meier, M.-A., and Wiemer, S.: Can temporally and spatially varying b-values improve earthquake forecasts? Insights from a machine-learning-enhanced catalog in central Italy., EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13557, https://doi.org/10.5194/egusphere-egu26-13557, 2026.
The Lilliefors test is commonly applied to assess the exponentiality of earthquake magnitudes and, consequently, to estimate the minimum threshold above which seismic events are completely recorded (the completeness magnitude). In theory, the test assumes continuously distributed exponential data; however, real earthquake catalogs typically report magnitudes with finite resolution, resulting in a discrete (geometric) distribution. To address this mismatch, standard practice adds uniform noise to the data prior to testing for exponentiality.
In this work, we analytically demonstrate that uniform dithering cannot recover the exponential distribution from its geometric counterpart. Instead, it produces a piecewise-constant residual lifetime distribution, whose deviation from the exponential model becomes increasingly detectable as the catalog size or bin width increases, as confirmed also by numerical experiments. We further prove that an exponential distribution truncated over the bin interval is the exact noise distribution required to correctly restore the continuous exponential distribution over the whole magnitude range. Numerical tests also show that this correction yields Lilliefors rejection rates consistent with the significance level for all bin widths and catalog sizes.
Correcting the exponentiality test for binned magnitudes according to these results ensures a more reliable estimation of the completeness threshold, particularly in the case of high-resolution earthquake catalogs.
How to cite: Spassiani, I. and Stallone, A.: Correcting the exponentiality test applied to binned earthquake magnitudes, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17786, https://doi.org/10.5194/egusphere-egu26-17786, 2026.
In previous studies (Rotondi et al., Geophys J Int 2022) we examined the seismic sequences related to the strong earthquakes that occurred in central Italy at L'Aquila in 2009 and at Amatrice-Norcia in 2016, estimating the q-exponential probability distribution of the magnitude. Specifically, we considered events with Mw 2+ recorded in the intervals (2005-2009) for the L’Aquila case and (2014-2018) for the Amatrice-Norcia case in order to explore the link between changes in magnitude distribution and various seismic phases.
The temporal variations were noted in the values of the Tsallis entropy and of the corresponding q entropic index estimate when we evaluated them on time windows with a fixed number of data, that shift at each new event, making inference according to Bayesian MCMC methods (Rotondi et al., Seismol Res Lett 2025). These analyses revealed a link between changes in q and different phases of seismic activity, with low q values potentially marking the preparatory phase preceding strong events.
In the present work, this approach is extended by analyzing all seismic events recorded in Central Italy between 2005 and 2024 as a single unified sequence, and drawing data both from the Italian Seismological Instrumental and Parametric Database (ISIDe) and from the HOmogenized instRUmental Seismic catalog (HORUS), which provides more accurate and homogeneous moment magnitude estimates.
Our goal is to determine whether the temporal variations in Tsallis entropy and its parameter q identified in our previous work truly act as both sufficient and necessary precursory signals of strong earthquakes. It turns out that variations in the q-index alone are not a sufficiently reliable seismic precursor, as low q values may not be followed by strong events.
However, a more reliable identification of periods of heightened seismic activity is achieved by jointly analyzing q and the parameter β, which is physically related to the expected released energy. In particular, the correlation between q and β evaluated through a moving correlation analysis allows the identification of periods of intense seismic activity. A persistent and significant decrease in q, combined with a positive correlation between q and β, suggests the onset of a preparatory phase for an impending seismic event. The use of the HORUS catalog has further strengthened the significance of these conclusions.
This research is supported by ICSC National Research Centre for High Performance Computing, Big Data and Quantum Computing (CN00000013, CUP B93C22000620006) within the European Union-NextGenerationEU program.
How to cite: Varini, E., Rotondi, R., and González Fuentes, A.: Tracking seismic regime changes in Central Italy (2005-2024) through variations in the parameters of the q-exponential magnitude distribution, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13822, https://doi.org/10.5194/egusphere-egu26-13822, 2026.
We propose a stochastic partial differential equation (SPDE) formulation of the Epidemic-Type Aftershock Sequence (ETAS) model for efficient Bayesian inference of spatially varying background seismicity. While recent Bayesian ETAS formulations already model the background rate using Gaussian Process priors, their application to large earthquake catalogs is limited by the associated dense covariance structure. Using synthetic earthquake catalogs, we demonstrate that the proposed SPDE–ETAS model accurately recovers both background and triggering parameters, achieving estimation performance comparable to previous Gaussian Process–based Bayesian ETAS models and superior stability relative to kernel-based approaches. The sparse precision matrix induced by the SPDE representation leads to substantial reductions in computational cost and memory usage, enabling scalable inference without compromising accuracy. Application to the Italian earthquake catalog (1960–2025) reveals spatially coherent background seismicity patterns aligned with major tectonic features, and provides robust and well-constrained Bayesian estimates of ETAS triggering parameters. These results establish the SPDE–ETAS framework as a computationally efficient and flexible alternative for Bayesian earthquake modeling, particularly suited for large and high-resolution seismic catalogs.
How to cite: Rahmani, S. T.-E., Zöller, G., and Hainzl, S.: SPDE–ETAS: Fast and Accurate Bayesian Inference for the Spatio-Temporal Epidemic-Type Aftershock Sequence (ETAS) Model, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10943, https://doi.org/10.5194/egusphere-egu26-10943, 2026.
Rigorous evaluation of earthquakes forecasts is a crucial step in understanding and improving the capabilities of earthquakes forecasting models. The UCERF3-ETAS model is currently the most advanced seismicity model combining a long-term seismicity model incorporating hypotheses of fault rupture dynamics and elastic rebounding with an Epidemic-Type Aftershock Sequence (ETAS) model for short-term seismicity. UCERF3-ETAS has also been used on demand for operational earthquake forecasting of important seismic sequences like the 2019 Ridgecrest one. Here, we have evaluated a very large database of UCERF3-ETAS next-day forecasts for California from 1 August 2008 to 31 August 2018. Each next-day forecast is composed of 100,000 synthetic catalogs generated by the model. The synthetic catalogs comprise events with magnitude $M_w \geq 2.5$, start at 00:00:00 UTC, last 24 hours, and include all events prior to midnight in the history for generating the next day’s forecasts. We evaluate the consistency of the model against 17,655 $M_w \geq 2.5$ earthquakes that occurred in California in the period 2007-2018 using the statistical tests for catalogue based forecasts developed by the Collaboratory Study of Earthquake Predictability. We find that the number of events provided by the forecast is generally consistent with the observations, especially during relevant seismic sequences such as the $7.2 M_w$ El-Mayor Cucapah, while swarm type sequences are more challenging. The magnitude distribution is also consistent overall. We also study the spatial evolution of the magnitude distribution to highlight regions where the model is expecting large earthquakes to happen and find that they are coherent with observed seismicity. Finally, we compare UCERF3-ETAS forecasts against fully prospective next-day forecasts produced by 27 different models operated by CSEP during between 2007 and 2018, and collected in a openly available database which constitutes a natural benchmark for the problem. We find that UCERF3-ETAS improves upon older models by providing positive information gains in most periods. The information gain tends to be zero or negative during swarms when UCERF3-ETAS is compared against models having a non-parametric component signaling possible benefits of including one to better describe this type of seismic sequences.
How to cite: Serafini, F., Bayona, J. A., Silva, F., Milner, K., Field, N., and Werner, M. J.: A decade-long pseudo-prospective evaluation of UCERF3-ETAS next-day seismicity forecasts, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21656, https://doi.org/10.5194/egusphere-egu26-21656, 2026.
Operational Earthquake Forecasting (OEF) predominantly relies on Epidemic-Type Aftershock Sequence (ETAS) models for short-term seismicity forecasts. We first develop and calibrate a baseline ETAS model for the European region, systematically exploring parameterisations that include alternative productivity laws and spatially variable background rates informed by the European Seismic Hazard Model (ESHM20). These extensions provide a consistent reference framework for regional-scale OEF. Building on this baseline, we improve the spatial triggering component by replacing isotropic kernels with event-specific elliptic kernels that incorporate directional information inferred from aftershock distributions. In near-real-time forecasting, the estimation of kernel orientations introduces a latency, as directional information becomes available only after sufficient aftershocks have occurred. However, our model leads to improved performance in pseudo-prospective forecasts, highlighting the relevance of spatial anisotropy in triggered seismicity. We also find reduced bias in ETAS parameters, primarily the productivity law.
We further investigate mismatches between expected and observed aftershock productivity by proposing simple productivity updates based on residuals between predicted and observed aftershock counts, yielding modest positive information gain on average. A sequence-by-sequence analysis reveals, however, that some sequences transition from early underestimation to later overestimation, or vice versa, limiting the effectiveness of uniform adaptive schemes. We therefore explore whether early sequence behaviour and covariates such as tectonic regime, location, and geophysical features can help anticipate subsequent productivity evolution. Finally, we assess the practical value of increasing model complexity for OEF, questioning whether statistically significant performance gains translate into meaningful improvements over simpler forecasting approaches.
How to cite: Han, M., Mizrahi, L., and Wiemer, S.: Elliptic Triggering Kernels and Adaptive Productivity in European OEF, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12197, https://doi.org/10.5194/egusphere-egu26-12197, 2026.
Sabah, the easternmost state of Malaysia, is the most tectonically active region of Borneo despite being distant from active plate boundaries. Global earthquake catalogues record recurrent earthquakes of Mw ≈ 5.0 at roughly five-year intervals, primarily concentrated along the northwestern and southeastern flanks of Sabah. Seismicity along the northwestern flank is particularly focused around Mount Kinabalu and the offshore Baram Delta, where normal faulting and half-graben basin development coexist with thrust faulting. In contrast, seismicity along the southeastern flank is dominated by thrust-faulting earthquakes at depths up to 30-40 km. Numerous studies suggest that this intraplate seismicity reflects the reactivation of post-subduction structures inherited from the Proto–South China Sea subduction and subsequent Celebes Sea subduction beneath Sabah in the Neogene. Despite this activity, detailed seismicity studies remain sparse due to historically limited seismic station coverage. Recent expansion of the Malaysian National Seismic Network and the temporary Northern Borneo Seismic Network (nBOSS) between 2018-202 provide new opportunities to develop an enhanced earthquake catalogue with improved source characterisation. This study aims to produce a refined earthquake catalogue for Sabah, with particular focus on the Mount Kinabalu and Darvel Bay regions. We integrate machine-learning-based tools, including PhaseNet, together with in-house software (QuakeMigrate and MTfit), to automatically analyse spatial and temporal patterns of seismicity and focal mechanisms. The goal is to improve our understanding of the active tectonics of northern Borneo and assess the implications for regional seismic hazard in this post-subduction setting.
How to cite: tonprasert, W. and Rawlinson, N.: Seismicity and active tectonics of northern Borneo , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11466, https://doi.org/10.5194/egusphere-egu26-11466, 2026.
The East African Rift System (EARS) is an active continental rift that experiences frequent earthquakes, yet seismic hazard assessment across the region remains difficult. Challenges stem from sparse monitoring networks, the absence of standardized guidelines, and numerous active faults that remain unmapped or poorly characterized. In addition, regional earthquake catalogs are incomplete and often depend on limited data, introducing considerable uncertainty into seismic hazard estimates. Despite these issues, conventional least-squares regression methods are still commonly used for magnitude conversion, even though they are sensitive to outliers, rely on assumptions that are made but rarely validated, and possess several limitations. These limitations constrain the generation of reliable homogenized earthquake catalogs essential for seismicity, seismotectonic, and hazard assessments.
This study proposes a robust statistical framework for deriving regional magnitude conversion relationships using the Restricted Maximum Likelihood (REML) estimation method. REML is particularly advantageous for data-scarce regions such as EARS, as it explicitly accounts for measurement uncertainties, non-constant variance, and the prevalence of outliers common in mixed-magnitude earthquake catalogs. The methodology incorporates rigorous statistical tests, including Box–Cox transformations for date normality, residual diagnostics, and variance stability evaluations.
To demonstrate its usefulness, the proposed framework is applied to catalogs from three regions along the EARS: (1) the Main Ethiopian Rift (Eastern Branch), (2) Sudan (a tectonically stable region), and (3) Malawi (Western Branch). The resulting magnitude conversion relationships exhibit significantly reduced uncertainty and provide confidence bounds, thereby enhancing the reliability of homogenized earthquake catalogs. The proposed approach strengthens the consistency of earthquake datasets across East Africa and offers a valuable tool for improving seismic hazard and risk assessments in similar data- limited regions worldwide.
How to cite: Al-Ajamee, M.: A Restricted Maximum Likelihood Framework for Earthquake Magnitude Conversion in Data-limited Regions of the East African Rift, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-664, https://doi.org/10.5194/egusphere-egu26-664, 2026.
The North-Eastern Himalaya, Indo-Burma Ranges, and Andaman-Nicobar together form one of the most seismically active and structurally intricate tectonic regions in the world, hosting numerous Mw ≥ 6.5 earthquakes throughout recorded history. Understanding the physical controls for the occurrence of such high-magnitude events is vital for improving hazard assessment and the prediction of possible regions of future earthquakes. Conventional methods often struggle to integrate a large number of geological, geodetic, and geophysical factors that influence earthquake generation, as all these factors together play a role in masking and amplifying the effects of one another. In this study, we address this challenge by developing a multi-parameter, explainable artificial intelligence (XAI)-based approach to identify the dominant factors influencing megathrust earthquakes in this region. We have used a clustering technique to compile 16 different parameters, like gravity anomalies, plate convergence rate, accumulated strain, sediment cover, slab geometry, crustal thickness, slab age, and seismic attenuation factor, to form a comprehensive input to the model. Since the study region represents two different tectonic setups- continent-continent collision zone in the Himalayan and the Andaman Arakan ocean-continent subduction zone in the Indo-Burmese ranges, therefore the dataset was separated based on their tectonic characteristics. A Fully Connected Neural Network (FCN) has been trained and deployed to classify earthquakes into Class 1 (Mw ≥ 6.5) and Class 0 (Mw < 6.5). An XAI technique, Layerwise Relevance Propagation (LRP), was applied to determine which of the parameters are heavily influencing the classification or model's predictions. LRP is an XAI method that traces a model’s prediction backward through the network and redistributes the output score with respect to input features to show which parts contributed the most.
LRP research reveals distinct and geologically consistent elements that determine the major players for the occurrence of earthquakes in the two tectonic regimes. In the continent–continent collision zone, composite strain, composite plate convergence velocity, gravity anomalies (Bouguer and free-air), and slab depth emerge as the dominant parameters influencing earthquake classification. Conversely, the oceanic subduction regime is primarily controlled by sediment thickness, gravity gradient, slab age, along with composite velocity and composite strain. Notably, higher values of composite velocity and composite strain are consistently associated with the occurrence of megathrust earthquakes in both tectonic settings, highlighting their fundamental role in strain accumulation and seismic rupture processes. The significance of sediment thickness may be understood by its influence on the roughness of the subduction interface. A thicker sediment cover makes subduction smoother by making the slab bathymetric imperfections less noticeable, whereas a thinner sediment cover makes the interface rougher, which causes more strain to build up along the megathrust. This process aligns with the frequent occurrence of megathrust earthquakes in the area, such as the 2004 Great Sumatra earthquake. The proposed model successfully captures this relationship between sediment thickness, strain accumulation, and seismic potential.
This first-order study demonstrates that combining XAI with multi-parameter tectonic datasets establishes a robust framework for identifying and understanding the primary causes of seismicity in complex orogenic/geodynamic systems.
How to cite: Ghosh, R., Pathak, P., and Mohanty, W. K.: Multi-Parameter Controls on Megathrust Earthquakes Revealed by Explainable Artificial Intelligence in a Complex Orogenic System, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16943, https://doi.org/10.5194/egusphere-egu26-16943, 2026.
The Czech Republic is a moderately active seismic region. Although most recorded earthquakes are weak, some events are strong enough to be felt by the population (e.g. Hlučín, December 2017, ML 3.5; West Bohemia, December 2025, ML 2.5 – 3.0). The majority of seismicity is mining-induced; however, areas of natural seismicity also exist, such as the Opava region and West Bohemia.
Seismic activity in the Czech Republic is monitored by several seismic networks, with the Czech Regional Seismic Network (CRSN) serving as the primary system. Seismic monitoring includes rigorous event classification, i.e. distinguishing between natural and induced seismicity as well as between earthquakes and surface explosions recorded by seismic stations.
With the growing volume of seismic data, semi-automated seismic event processing has become increasingly necessary. Automatic seismic event classification based on seismic signals represents a key step toward this goal. In previous work, we achieved promising results using machine learning (ML) techniques applied to data from the Ostrava-Krásné Pole station (OKC), which monitors the northeastern Czech Republic, an area with historically significant mining activity.
In this study, we extend seismic event classification to stations with a different instrumentation and apply newer ML approaches. Namely, we analyze data from the Moravský Beroun (MORC) and Vranov (VRAC) stations of the CRSN, both equipped with broadband STS-2 sensors with a lower corner period of 120 s and recording continuous seismic waveforms at 100 Hz. The studied dataset used for binary classification consists of records of mining-induced seismic events (8,338 from MORC, 4,085 from VRAC) and quarry blasts (4,193 from MORC, 3,041 from VRAC), which were localized in the Czech Republic and its neighbouring countries in 2023 – 2025. Induced events with known P- and S-wave arrivals and explosions with known P-wave arrivals were selected. The P-wave and S-wave arrival times were taken from bulletins provided by the Institute of Physics of the Earth.
Each processed event record includes 1 s before the P-wave arrival and either 20 s after the S-wave arrival (if available) or 30 s after the P-wave arrival. Data preprocessing included Z-score normalization and time-frequency transformation of the seismic signals.
We evaluated several models, including LSTM, LSTM-FCN, LSTM with an attention block, a hybrid CNN-Vision Transformer (CNN-ViT) neural networks, and XGBoost. The evaluated models achieved F1-scores of 0.92 (LSTM-based), 0.94 (XGBoost), and 0.96 (CNN-ViT), with comparable performance for MORC-only, VRAC-only, and combined datasets.
Furthermore, we combined data from the MORC and VRAC stations with records from the OKC station in a multimodal approach (37,561 events). Despite differences in instrumentation (e.g. lower corner periods of 120 s versus 30 s), the models achieved consistently high performance, with F1-scores ranging from 0.92 to 0.96 (CNN-ViT model yielding the best results).
These results demonstrate that machine learning models represent a promising step toward automated seismic event classification and more efficient seismic signal processing.
How to cite: Skotnica, M., Pecha, M., Pazdírková, J., Rušajová, J., and Rieznikov, B.: Machine learning-based seismic event classification at selected stations of the Czech Regional Seismic Network, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13017, https://doi.org/10.5194/egusphere-egu26-13017, 2026.
The territory of Caucasus is a seismo-active region affected by the tectonic interaction of Arabian and Eurasian plates. The strong deformation processes developed here cause the accumulation of tectonic energy - stress, which discharges by the occurrence of numerous earthquakes. The monitoring and study of earthquake precursors represent the task of global importance.
It is known that there are number of earthquakes’ precursors, which can be registered in various geophysical fields (geomagnetic, hydrogeodeformation), but in order to consider the precursors registered before the activation of tectonic processes as a reliable earthquake indicators, it is necessary to reveal the solid connection between the seismic activity and the variation of the parameters, characterizing various geophysical fields.
The existing modern multiparametric monitoring system in Georgia, allow us to conduct a probabilistic assessment of expected earthquake magnitudes in different locations across Georgia, using modern Machine Learning (ML) methods, namely deep neural networks (DNN) technology, applied to experimental monitoring data on water level in boreholes and geomagnetic data.
During observation we consider the earthquake forecast as a binary problem of machine learning on the imbalanced data base applied to five regions of Georgia. For the training we used the geophysical data base collected in 2020-2024, namely, variations of statistical characteristics of geomagnetic field components, seismic activity, water level in deep boreholes and tides.
In the present study, special attention is paid to the identification of stable precursor patterns by integrating multiple geophysical parameters within a unified analytical framework. Feature engineering and normalization techniques were applied to reduce noise and enhance the sensitivity of weak pre-seismic signals. The performance of the developed ML models was evaluated using standard classification metrics, including precision, recall, F1-score, and probability gain, demonstrating an improvement in detection capability compared to single-parameter approaches. The preliminary results indicate that joint analysis of geomagnetic, hydrogeological, and tidal data increases the reliability of probabilistic seismic forecasting and provides a promising basis for the development of an operational early-warning support system for seismically active regions of Georgia.
How to cite: Jimsheladze, T.: Preliminary results on variation of geophysical parameters during preparation of seismic events in Georgia using Machine Learning tools, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12516, https://doi.org/10.5194/egusphere-egu26-12516, 2026.
Situated along three major plate boundaries, Anatolian plate is represented by major destructive earthquakes exceeding Mw > 7. Accurate forecasting of earthquake epicenters is crucial for both structural resilience and efficient risk reduction. In this work, we develop a machine-learning based epicenter prediction framework covering the entire territory of Türkiye, using the national seismic catalogue provided by KOERİ and AFAD. The dataset in particular is partitioned into four consistent clusters derived from localized strain fields estimated through integrated InSAR and GNSS observations (e.g Weiss et al 2020). For training the models, we removed the background max shear strain < 50 nanostrain/year and consider, namely, the North Anatolian fault system, East Anatolian fault system, western Anatolian extensional region, and East Anatolian shortening zone. In addition, all earthquakes are classified as large or medium using a magnitude threshold of Mw ≥ 5, yielding eight distinct datasets. For each dataset, we train and evaluate seven machine-learning models—Linear Regression, Random Forest, XGBoost, Multilayer Perceptron (MLP), Support Vector Regression (SVR), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU)—to predict future epicenters (latitude and longitude) from historical spatiotemporal information. Comparing all models within each geodetic cluster allows us to identify which model families perform better under specific tectonic deformation regimes, while simultaneously revealing which regions exhibit higher predictability. This multi-model, multi-region evaluation provides new insights into data-driven seismic forecasting across the Anatolian plate where the role of various plate boundary scale faults (shear zones) are associated with destructive earthquakes.
How to cite: Gülümsek, A., Göğüş, O. H., Kılınçkaya, M. T., and Bodur, Ö.: Predicting Earthquakes Across The Anatolian Plate Through Machine Learning Algorithms, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1005, https://doi.org/10.5194/egusphere-egu26-1005, 2026.
The Western Anatolia-Aegean region is dominated by active lithospheric extension, magmatism and widespread seismicity (including Samos earthquake Mw=7.0, 30.10. 2020). However, its complex tectonic setting including continuum/distributed mode of deformation rather than block type (more localized), and uneven station coverage highlight the limitations of traditional centralized machine-learning approaches. Notably, owing to data-sharing restrictions and the lack of regionally representative training datasets there are substantial challenges for developing reliable short-term earthquake forecasting models. To address these issues, we develop a federated learning (FL) framework that enables multiple seismic agencies and stations to collaboratively train predictive models without exchanging raw waveform data. Our dataset integrates multi-station acceleration and INSAR-GPS based displacement time-series, regional geological parameters, and spatiotemporal feature windows derived from AFAD, KOERI, IRIS, and USGS archives. Within this framework, we formulate two forecasting tasks: (i) classification of the likelihood of an earthquake exceeding a magnitude threshold within 24–72 hours, and (ii) regression-based estimation of short-term seismic intensity. Several deep-learning architectures, including 1D-CNN, LSTM, and CNN–LSTM hybrids, are implemented under both centralized and federated training schemes to systematically evaluate the effect of non-IID data distributions, communication constraints, and regional variability on forecasting skill. Comparative experiments show that FL preserves most of the predictive performance of centralized models while providing critical advantages in data privacy, scalability, and institutional participation. These results highlight the potential role of federated machine learning to support next-generation seismic forecasting systems, foster cross-institutional collaboration, and facilitate operational earthquake preparedness across data-restricted regions.
How to cite: Kılınçkaya, M. T., Göğüş, O. H., Gülümsek, A., and Bodur, Ö.: Federated Learning–Based Earthquake Forecasting in the Western Anatolia-Aegean Extensional Province, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1006, https://doi.org/10.5194/egusphere-egu26-1006, 2026.
The Western Quebec Seismic Zone (WQSZ) is an intraplate region of Canada that experiences an unusual amount of earthquake activity far from plate boundaries. Over the past 40 years, more than 2000 earthquakes have been recorded in the WQSZ. Although the majority of earthquakes occur at relatively low magnitudes (between M 2-3), Canada’s national capital, Ottawa, and its second-largest city, Montréal, are both located within the WQSZ. As a result of their political and economic importance, the Canadian Government implemented an Earthquake Early Warning system to the region in late 2025.
Previous studies have primarily focused on potential faulting mechanisms and/or large earthquakes in the region (M > 5). However, the WQSZ is relatively understudied, with limited modern data science techniques applied to the seismic database. Given the wide surface area covered by the region and the regularity of events (i.e. an earthquake every 6.5 days), there is an urgent need to better understand the spatial and temporal patterns of seismicity across the WQSZ to further inform the hazards on a more local scale.
Clustering analysis is used to help group data into spatial patterns where the relationship is previously unknown. In this study, we apply an unsupervised machine learning clustering algorithm, Density-Based Spatial Clustering of Applications with Noise (DBSCAN), to analyze seismicity in the WQSZ and to delineate distinct spatial clusters. However, the output of the clustering analysis through DBSCAN is dependent on the choice of values for key parameters. To address this, a wide range of parameter values are tested to create a broad suite of cluster patterns and a statistical framework is developed to help identify the most robust patterns that best represent the geological and geophysical context for the region.
Our framework combines DBSCAN patterns with temporal, statistical and geological analysis to create a new high-resolution spatio-temporal characterization of seismicity in the WQSZ. These findings not only improve the understanding of localized seismic risk in Western Québec but also provide an application of cluster analysis to real-world logistical issues of seismic hazard analysis, including identifying areas of highest risk for earthquake preparedness and emergency planning in the region.
How to cite: Yasokaran, O. and Heron, P.: Determining the spatio-temporal patterns of intraplate earthquakes of the Western Quebec Seismic Zone using clustering analysis, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8153, https://doi.org/10.5194/egusphere-egu26-8153, 2026.
In January 2025, the area between Santorini and Amorgos experienced the onset of intense seismic activity, with more than 8,500 seismic events of local magnitude ML ≥ 1.5 recorded, and a maximum‑magnitude event of Mw = 5.2. The swarm developed within a complex seismotectonic regime and attracted significant scientific interest because of its proximity to densely populated islands.
In this study, we investigated the spatiotemporal evolution and statistical properties of the seismic swarm to better understand the underlying physical processes, using a dataset extracted from the high-resolution catalogue developed by Fountoulakis et al. (2025), covering the period from 27/01/2025 to 04/03/2025 and including seismic events of magnitude ML ≥ 1.5. The seismic activity initiated beneath the Santorini caldera and progressively migrated northeast towards the Kolumbo submarine volcano and the offshore region of Amorgos, following a NE–SW-trending extensional fault system approximately 60 km long.
The spatiotemporal analysis revealed two distinct phases of activity, separated by a short transition period (Zaccagnino et al., 2025). The primary phase, from 1 to 9 February 2025, is characterised by rapid spatial expansion and an abrupt increase in the seismicity rate. The secondary phase, from 11 February to 4 March 2025, shows a more coherent migration pattern and a normal decay in the seismicity rate. Using non-additive statistical physics, we estimated the entropic parameters of the inter-event times and distances for both phases and found that they were well described by q-exponential distributions, with entropic parameters qT=1.15, Tq=3.453sec (R2=0.953), qD=0.8 and Dq=4.225Km (R2=0.998) for the primary phase, and qT=1.54, Tq=4.357sec (R2=0.924) and qD=0.72, Dq=8.791Km (R2=0.999) for the secondary phase. These results demonstrate that the 2025 Santorini Amorgos seismic sequence was governed by a non‑additive dynamics, with distinct physical characteristics between the two phases of activity.
References
Fountoulakis, I., Evangelidis, C. P. (2025). The 2024–2025 seismic sequence in the Santorini-Amorgos region: Insights into volcano-tectonic activity through high-resolution seismic monitoring. Seismica, 4 (1). https://doi.org/10.26443/seismica.v4i1.1663
Zaccagnino, D., Michas, D., Telesca, L., Vallianatos, F. (2025). Precursory patterns, evolution and physical interpretation of the 2025 Santorini-Amorgos seismic sequence, Earth and Planetary Science Letters, 671, 119656. https://doi.org/10.1016/j.epsl.2025.119656.
How to cite: Vallianatos, F. and Pavlou, K.: Spatiotemporal pattern of the Santorini - Amorgos 2025 seismic sequence in terms of non additive statistical mechanics, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13885, https://doi.org/10.5194/egusphere-egu26-13885, 2026.
Since January 27, 2025, intense seismic activity has been recorded in the offshore area between Santorini and Amorgos, with more than 4,500 events. The sequence began inside the Santorini caldera and gradually migrated northeast. The strongest earthquake was an ML 5.3 event on February 10, 2025. In this study offers a seismological analysis that integrates patterns of seismic activity over space and time, static Coulomb failure stress changes, and shifts in seismic velocity structure to explore the mechanisms behind the swarm's development.
The analysis is based on seismological data from the NKUA monitoring network for the period from 1 January to 3 March 2025. Coulomb stress changes were computed for events with Mw ≥ 4.7 using elastic half-space modelling, while a modified Wadati method was applied to a subset of well-located events to estimate the regional average Vp/Vs ratio. The results reveal a northeastward migration of seismicity, closely aligned with NE–SW-oriented fault structures in the Santorini–Amorgos area.
Coulomb stress modelling for the events of magnitude Mw > 4.8 reveals predominantly positive stress changes at the hypocenters of subsequent earthquakes, suggesting that static stress transfer contributed significantly to the progressive activation of neighboring faults. At the same time, the estimated Vp/Vs ratio of approximately 1.75 is consistent with a fluid-influenced seismogenic environment, supporting the involvement of crustal heterogeneities and possible fluid-related processes during the swarm.
The combined observations suggest that the 2025 Santorini–Amorgos seismic sequence was controlled by the interaction between fault-driven stress redistribution and variations in crustal properties. This approach provides new insights into earthquake triggering mechanisms in complex volcanic–tectonic settings of the South Aegean and highlights the importance of multidisciplinary analyses for seismic hazard assessment.
How to cite: Pavlou, K., Sardeli, E., Karakonstantis, A., Pappas, S., Athanasopoulos, A., Tomaras, A., Kazakou, A.-A., Travlostathi, C., and Vallianatos, F.: Spatiotemporal distribution, Coulomb stress changes, and temporal variations in Vp/Vs ratio during the 2025 Santorini-Amorgos seismic swarm., EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10805, https://doi.org/10.5194/egusphere-egu26-10805, 2026.
The Epidemic Type Aftershock Sequence (ETAS) model, a widely used self-exciting, marked Hawkes process, has become a standard tool in statistical seismology. However, the standard ETAS formulation assumes a stationary background seismicity rate and therefore lacks the ability to capture the spatiotemporal structure of background seismicity. In this study, we extend the GP-ETAS model proposed by Molkenthin (2022) to incorporate a spatiotemporally varying background rate. We use nonparametric Gaussian process (GP) priors to describe spatiotemporal background seismicity and estimate them using a Bayesian inference framework with Markov chain Monte Carlo (MCMC) sampling techniques. We apply the extended GP-ETAS model to regions affected by Slow Slip Events (SSEs), which are known to generate stress changes that are both spatially and temporally heterogeneous, significantly influencing background seismicity patterns. The extended GP-ETAS model enables quantitative spatiotemporal analysis of SSE-driven variations in background seismicity.
How to cite: Niu, Y. and Zhuang, J.: Spatiotemporal Modeling of Background Seismicity Using Gaussian Processes, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21142, https://doi.org/10.5194/egusphere-egu26-21142, 2026.
Beno Gutenberg and Charles F. Richter (1941) already hypothesized that their exponential relation between the magnitude and occurrence frequency of earthquakes would not be valid for the largest ones, as there should be a maximum limit to the earthquake size. This departure would have profound implications for global seismic hazard assessment, as there would actually be fewer large earthquakes than extrapolated from the distribution of smaller ones.
But statistically proving or disproving this hypothesis has been elusive, and a debate is ongoing on whether a statistically significant departure can be observed in the available global data. It was first necessary to develop a magnitude scale reliable up to the largest earthquake sizes (moment magnitude Mw, in the 1970s) and gathering ever-increasing earthquake catalogues (especially since the 1980s).
Not all statistical tests may identify a given departure as significant, because the largest earthquakes are infrequent, so their sample size is small. Recently it has been proposed that the whole observed distribution is still a simple exponential (Taroni, 2025). But several earlier results already had already identified a significant departure by which the tail of the distribution decays faster (Yoder et al., 2012, Serra & Corral, 2017, Corral & González, 2019).
To settle this question, here we use the largest available dataset: the ISC-GEM catalogue (International Seismological Centre, 2026) since the early XX century. In the analysis, we explicitly account for the magnitude uncertainties (substantial before the advent of the World-Wide Standardized Seismograph Network in the late 1960s).
This approach allows us considering the largest earthquakes ever instrumentally recorded and about triples the number of large earthquakes (Mw ≥ 6.5) available for analysis, compared to considering only the seismicity since the 1980s as typically done.
Using robust statistical tests, we show that the observed departure from a single Gutenberg-Richter law (clearly visible for Mw larger than ~7.6) is statistically significant, and examine the shape of this tail and its persistence in time.
References cited
Corral, Á., González, Á. (2019). Power law size distributions in geoscience revisited. Earth and Space Science, 6, 673–697. https://doi.org/10.1029/2018ea000479
Gutenberg, B. & Richter, C. F. (1941). Seismicity of the Earth. Geological Society of America Special Papers, number 34. 131 p.
International Seismological Centre (2026). ISC-GEM Earthquake Catalogue, https://doi.org/10.31905/d808b825
Serra, I., & Corral, A. (2017). Deviation from power law of the global seismic moment distribution. Scientific Reports, 7, 40045. https://doi.org/10.1038/srep40045
Taroni, M. (2025). The Gutenberg–Richter law strikes back: the exponentiality of magnitudes is confirmed by worldwide seismicity. Geophysical Journal International, 243 (2), ggaf366, https://doi.org/10.1093/gji/ggaf366
Yoder, M. R., Holliday, J. R., Turcotte, D. L., & Rundle, J. B. (2012). A geometric frequency-magnitude scaling transition: Measuring b = 1.5 for large earthquakes. Tectonophysics, 532-535, 167–174. https://doi.org/10.1016/j.tecto.2012.01.034
How to cite: González, Á., Corral, Á., and Serra, I.: The largest earthquakes recorded for over a century significantly depart from a simple Gutenberg-Richter distribution, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11541, https://doi.org/10.5194/egusphere-egu26-11541, 2026.
Several machine learning algorithms have been developed for earthquake catalog declustering and have demonstrated high accuracy, particularly for the well-studied Southern California region. This study applies a machine learning–based Probabilistic Random Forest (PRF) approach to earthquake declustering and compares its performance with that of the Random Forest (RF) method in the Southern California region by introducing noise into the dataset. Although the Southern California dataset is of high quality due to a dense seismic network, inherent observational and instrumental noise can still affect model performance. Five features are considered, each describing a different aspect of the space–time–magnitude interactions inherent in seismicity. The rescaled time (T*) represents the temporal interval between consecutive seismic events, while the rescaled distance (R*) quantifies their spatial separation. The magnitude difference is expressed as Δmj = mi − mj, where i denotes the nearest neighbor, and generally attains larger values when event j is an aftershock of a stronger mainshock. The number of siblings refers to the count of events that share the same nearest neighbor as event j, with higher values indicating multiple aftershocks associated with a common parent event. The number of offspring denotes the number of subsequent events that identify event j as their nearest neighbor, thereby reflecting its triggering potential. For training and testing the RF and PRF algorithms, the original dataset was supplied to the epidemic type aftershock sequence (ETAS) model for parameter estimation using the maximum likelihood method. Based on the estimated parameters, 100 different realizations of the combined background–cluster labeled dataset were generated using the thinning algorithm. Background events were labeled as “0”, whereas clustered events were labeled as “1” in the synthetic dataset. Three types of feature noise are introduced to assess model robustness: Type-I applies uniform Gaussian noise across all objects and features, Type-II assigns different noise levels to randomly grouped objects and features, and Type-III applies independent noise levels to training and testing datasets. Noise magnitudes are controlled by feature-wise standard deviations and an overall noise factor, with noisy values sampled from Gaussian distributions. For the synthetic datasets, figure illustrates the difference in declustering accuracy between the Probabilistic Random Forest (PRF) and standard Random Forest (RF) models across the three types of noise. For Type I noise, the maximum accuracy improvement is approximately 2%, while Type II noise shows an increase of around 2.5%. Type III noise, which represents a more complex noise scenario, exhibits a moderate accuracy gain of about 1.5%. For the real seismic datasets, the accuracy differences between PRF and RF are generally higher. As shown in figure, Type I noise leads to an accuracy improvement of nearly 2%, Type II noise also shows an enhancement of about 2%, while Type III noise, representing the most complex scenario, exhibits a substantial improvement of nearly 6%. The results demonstrate that as noise complexity increases particularly when the correlation within the noise becomes weaker, the PRF model consistently outperforms the standard RF classification.
How to cite: Seal, A. and Jana, N.: Earthquake Catalog Declustering in Southern California Using a Probabilistic Random Forest Approach, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16116, https://doi.org/10.5194/egusphere-egu26-16116, 2026.
Seismic catalogs combine background seismicity driven by tectonic loading with clustered earthquakes that reveal stress transfer and fault interactions. Both components are essential for seismic hazard models, which require accurate declustering.
Traditionally, declustering relies on window-based methods widely used in operational seismology for their simplicity and real-time efficiency. However, these methods suffer from rigid geometric constraints, depend on mainshock identification, and are highly sensitive to parameter choices, which may lead to over- or underestimation of earthquake cluster size. Machine learning-based approaches can mitigate these limitations by adapting flexibly to data patterns without rigid geometric or mainshock assumptions. Density-based algorithms such as DBSCAN and OPTICS identify spatial clusters effectively but struggle with spatiotemporal aftershock sequences because they treat time independently from space. ST-DBSCAN addresses this by using separate spatial and temporal radii, enabling flexible space-time clustering critical for aftershock analysis.
In this comparative study, we applied both approaches – ST-DBSCAN and window-based methods – to the New Zealand earthquake catalog to highlight the strengths and limitations of each, analyzing 15 overlapping clusters (>100 events, centroids <10 km apart). We found that ST-DBSCAN better captures fine-scale structures, whereas window-based methods produce more compact large-scale groupings. We analyze in detail the 2010–2013 Canterbury–Christchurch sequence, validating cluster membership against an independent dataset of approximately 150 earthquakes (Mw > 3.5), which reveals methodological differences in spatiotemporal resolution.
How to cite: Piegari, E., Gentili, S., and Caravella, L.: ST-DBSCAN vs Window-Based Methods: A Comparative Cluster Analysis of the New Zealand Earthquake Catalog, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16529, https://doi.org/10.5194/egusphere-egu26-16529, 2026.
Posters virtual: Mon, 4 May, 14:00–18:00 | vPoster spot 3
EGU26-4745 | ECS | Posters virtual | VPS12
A New Statistical Method to distinguish Different Earthquake Cluster TypesMon, 04 May, 14:24–14:27 (CEST) vPoster spot 3
Earthquake clusters can be broadly classified into two types: swarm-like sequences and mainshock–aftershock sequences. The spatial organization of the two types provides important insights into underlying tectonic processes and fluid migration in earthquake source regions. In this study, we apply the nearest-neighbor distance approach on the Southern California focal-mechanism earthquake catalog (the CNN_SoCal catalog) and introduce two new statistical indicators-skewness and kurtosis to distinguish between these two classes of earthquake clusters. We find that the square root of kurtosis and skewness provide effective and interpretable indicators for clusters classification. In the kurtosis–skewness diagram, swarm-like sequences and mainshock–aftershock sequences tend to occupy distinct regions, enabling a practical distinction between the two sequence types without relying on subjective inspection of individual clusters. Overall, the proposed approach offers an efficient way to differentiate swarm-like and mainshock–aftershock seismicity in large catalogs. The method is computationally light, easy to implement, and suitable for rapid screening of earthquake sequence types in high-resolution regional datasets.
How to cite: Fan, Y. and Hu, F.: A New Statistical Method to distinguish Different Earthquake Cluster Types, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4745, https://doi.org/10.5194/egusphere-egu26-4745, 2026.
EGU26-17842 | ECS | Posters virtual | VPS12
Sensitivity of Time-Dependent Earthquake Conditional Probabilities to Catalogue Declustering in the HimalayasMon, 04 May, 14:30–14:33 (CEST) vPoster spot 3
Earthquake catalogue declustering is a critical preprocessing step in time-dependent seismicity analyses (Gardner and Knopoff, 1974; Reasenberg, 1985), yet its systematic influence on conditional earthquake probability estimates remains insufficiently quantified, particularly in tectonically complex continental collision zones such as the Himalayas (Bungum et al., 2017). Renewal-based recurrence models typically assume that declustered catalogues isolate tectonically driven mainshock recurrence by removing dependent events. However, recent advances in declustering theory demonstrate that methodological choices, ranging from fixed spatio-temporal windows to adaptive and stochastic approaches, can substantially modify inter-event time statistics and inferred recurrence memory (Zaliapin et al., 2008; Zaliapin & Ben-Zion, 2020; Teng & Baker, 2019). Despite these developments, the implications of declustering-induced variability for time-dependent conditional probabilities remain underexplored in active orogenic belts.
In this study, we explicitly quantify how alternative declustering strategies influence time-dependent recurrence behavior and conditional rupture probabilities across selected Himalayan seismic source zones. Inter-event time series were constructed for moderate-to-large earthquakes (M ≥ 4.0) using both raw (non-declustered) and declustered catalogues derived from regional earthquake compilations. Declustering was performed using commonly applied fixed-window and adaptive approaches to capture epistemic variability associated with catalogue preprocessing. The resulting inter-event times were analyzed within renewal process models, including Brownian Passage Time (BPT), Lognormal, Weibull, and Gamma distributions, to estimate conditional probabilities as functions of elapsed time since the most recent major event.
Results show that declustered catalogues consistently yield smoother initial probability gradients and delayed probability peaks relative to raw catalogues, reflecting reduced short-term temporal clustering in inter-event time distributions. These shifts correspond to systematic changes in inferred renewal memory parameters, with declustering suppressing short-term contagion effects while largely preserving long-term mean recurrence intervals. In the Himalayas, collision-driven aftershock swarms and spatially heterogeneous fault interactions amplify these effects, introducing substantial epistemic uncertainty in early-time conditional probabilities, which can locally exceed factors of two to three depending on the declustering strategy employed. In contrast, long-term probability remains comparatively robust across declustering scenarios, consistent with steady-state tectonic strain accumulation.
These findings identify catalogue declustering as a dominant and often underappreciated source of uncertainty in time-dependent seismic probability modelling, reinforcing recent calls for ensemble-based and transparent pre-processing strategies in probabilistic seismic hazard workflows. This study advances a methodological framework for interpreting renewal-based conditional probabilities in clustered tectonic regimes. The Himalayas emerge as a natural laboratory where combined raw and declustered analyses can yield more resilient probabilistic interpretations.
How to cite: Pratap, B. and Sharma, M. L.: Sensitivity of Time-Dependent Earthquake Conditional Probabilities to Catalogue Declustering in the Himalayas , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17842, https://doi.org/10.5194/egusphere-egu26-17842, 2026.
