Seismic velocities, particularly P-wave (Vp) and S-wave (Vs) are critical for imaging the Earth's interior and understanding geodynamic processes. While the basic spatial relationships between seismic velocity structures and seismogenic zones have been extensively discussed, they are often described in a simplistic and coarse manner. Systematic and statistical investigations of these relationships particularly within the shallow crust and uppermost mantle, remain scarce. Significant challenges for such analyses are that the shallow crust and upper mantle are geologically complex, characterized by heterogeneous lithologies, intricate thermal and mechanical properties, and variable fluid distributions. These factors lead to nonlinear and highly heterogeneous spatial distributions of the Vp and Vs, complicating the interpretation and modeling of their relationship with seismogenic zones. Moreover, multiple interpretive paths for the same phenomenon often introduce subjective biases, making objective quantification of these relationships challenging. This study aims to address these challenges by leveraging machine learning (ML) techniques to explore and quantify the non-linear and complex relationships between seismic velocity structures and seismogenic zones across the Japan Arc. Using two distinct three-dimensional seismic velocity datasets, we employed various ML models to enhance the robustness and reliability of our analyses. Our results demonstrate that variations in the spatial distribution of Vp and Vs—especially the Vp/Vs ratios, vertical gradients, and variance of Vp, Vs—serve as reliable indicators for distinguishing seismogenic zones from non-seismogenic zones across both depth and geographic space even though the tectonic settings vary significantly. To interpret the complex nonlinear patterns revealed by ML models, we employed Shapley Additive Explanations (SHAP), which elucidated the spatial relationship between seismic velocities and seismogenic zones. By examining local seismogenic zones, The results by SHAP found factors influencing seismogenic zones differ with depth: at shallow depths, Vp, Vs, Vp/Vs ratio, and variance of Vp, Vs are dominant, while at greater depths, gradient changes are primary. It may relate to the thermal structure and indicate different triggering mechanisms for earthquakes at various depths. These findings can provide deeper insights into the spatial coupling between seismic velocities and seismicity, thereby advancing our understanding of the factors controlling earthquake generation.

How to cite: Zhang, C., Yoshiya, U., Yeo, T., Kato, A., and Iwamori, H.: Relationships Between Seismic Velocity Structures and Seismogenic Zone Decoded by Interpretable Machine Learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-5005, https://doi.org/10.5194/egusphere-egu25-5005, 2025.

The aftershock sequence typically consists of numerous seismic events, with their distribution exhibiting clustering characteristics. In geologically complex areas, such as the convergence boundary between the Eurasian and Philippine Plate in eastern Taiwan, it is challenging to explain the relationship between seismic events and regional structures. This area has experienced several disastrous earthquakes in recent years, including the Hualien earthquake (M_w 6.4) in 2018, the Chihshang earthquake (M_w 7.0) in 2022, and the Hualien earthquake (M_w 7.3) in 2024. Here, we aim to explore the relationship between the aftershock sequences of three events and the known active faults. Firstly, we apply the algorithm to cluster aftershocks. We analyze aftershock sequences for three events with local magnitudes greater than 3, spanning 45 days after the mainshock. Secondly, we examine the relationship between these clustered sequences and the 3D fault models developed by National Science and Technology Center for Disaster Reduction (NCDR). The results reveal that the aftershock sequence of the 2018 Hualien earthquake can be divided into five clusters, while the 2022 Chihshang earthquake can be divided into seven clusters. The mainshocks are separately located at clusters which have the largest number of aftershocks within their respective sequences. The aftershock sequence of the 2024 Hualien earthquake can be divided into eight clusters. The mainshock is located at a cluster with minor number of aftershocks, which is distributed along the Lingding Fault. Additionally, 3D visualization is employed to better illustrate the relationship between earthquake sequences and active faults, as well as to study potential earthquake mechanisms.

How to cite: Tu, K.-T., Huang, M.-W., Yang, C.-Y., Ke, M.-C., and Ke, S.-S.: The Relationship between Clustered-Aftershocks and 3D-Fault Models in Eastern Taiwan, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-1767, https://doi.org/10.5194/egusphere-egu25-1767, 2025.

Spatiotemporal variations of the b-value (the slope of the Gutenberg-Richter relation) are commonly associated to many disparate factors, such as critical stress conditions, the tectonic regime, and the incompleteness of earthquake catalogs.

During the 2016/17 central Italy earthquake sequence, several studies reported notable b-value variations, including an unexpected increase prior to the Norcia event (Mw 6.5). Such observations highlighted the complex relationship between b-value changes and seismic activity.

To get a better understanding of this relation, we reanalyze this sequence with a focus on the spatiotemporal evolution of seismicity near the Norcia mainshock. We temporally divided the seismic sequence into subperiods separated by the largest events and employ a combination of three machine-learning algorithms: DBSCAN for performing event clustering, OPTICS for analyzing spatially nested dense zones within clusters and PCA for inferring the planar geometry of those zones as fault surfaces. We identified two specific zones and reconstruct two separate fault planes. Those two zones exhibited asynchronous activity before the mainshock. We used the two-sample Kolmogorov-Smirnov Test to investigate similarity between the magnitude-frequency distribution of earthquakes associated with these planes. The results show that the b-values associated with these fault planes remained stable over time. Yet, their temporal changes exhibited a correlation with spatial variations of seismicity. In particular, the analysis indicated a relationship between the b-value and the geometry of the active fault.  

These findings suggest that temporal variations of the b-value during an earthquake sequence may not necessarily reflect changes in underlying stress conditions, but rather the activation of different earthquake sources throughout the sequence, each with different lithological and geometrical properties. This finding highlights the importance of understanding the fine-scale structure of earthquake sources in a sequence for correctly interpreting b-value variations.

How to cite: Piegari, E., Corrado, P., Herrmann, M., and Marzocchi, W.: Investigating the complex relationship between b-value changes and seismic activity in Central Italy, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13370, https://doi.org/10.5194/egusphere-egu25-13370, 2025.

This study investigates the different phases of seismic cycles present in earthquake sequences recorded in Chile and Italy on the basis of variations in the probability model for the magnitude and for the spatial distribution of the epicenters. Chile and Italy are two countries that, despite having different tectonic characteristics, face significant challenges due to seismic activity.

The seismic records are analyzed on sliding time windows with a fixed number of events, which shift with each new earthquake. Two probabilistic models are proposed for earthquake magnitude: one is based on the q-exponential distribution, hereafter referred to as the “q-exponential” model for simplicity, and the other is the exponential distribution, which is well known to be consistent with the Gutenberg-Richter law (Rotondi et al., Geophys. J. Int., 2022). The q-exponential distribution is closely related to Tsallis entropy, a generalized form of entropy that accounts for non-extensive systems, and has been widely applied in statistical mechanics to study complex systems. It is characterized by the parameter q, and as q asymptotically approaches 1, it reduces to the exponential distribution. As for the spatial distribution of the earthquakes, we consider the cell areas of the Voronoi tessellation generated by epicenters and adopt four probability models: the q-exponential, exponential, generalized gamma, and tapered Pareto distributions (Rotondi & Varini, Front. Earth Sci., 2022).

By following the Bayesian approach, the posterior distribution of model parameters is estimated by a Markov chain Monte Carlo method based on the Metropolis-Hastings algorithm, and then the optimal distribution in each time window is selected by comparing the estimated values of the posterior marginal log-likelihood. Given the high computational cost, parallel programming has been chosen to drastically reduce the computational time from days to hours or even minutes.

The results obtained in the study areas agree in associating seismic phase changes to the variations of the estimated q-index in the magnitude case and of the best probability model as for the spatial distribution; this provides useful indications for the implementation of risk mitigation actions. In both study areas, despite their different tectonic behaviors, we obtain similar results for seismic sequences characterized by a strong earthquake preceded by foreshocks. During the foreshock activity (i.e., the preparatory phase leading to the strong event), we observe an increase in the q parameter of the magnitude distribution and a preference for the tapered Pareto model for the spatial distribution of the epicenters provided by the Voronoi cell areas.

This work 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; Chilean National Agency for Research and Development (ANID), Fondecyt grant ID 1241881; Research Center for Integrated Disaster Risk Management (CIGIDEN), ANID/FONDAP/1523A0009.

How to cite: González, A., Varini, E., Rotondi, R., Nicolis, O., and Ruggeri, F.: Exploring seismic cycle dynamics via variations in probability models: Chile and Italy case studies, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9765, https://doi.org/10.5194/egusphere-egu25-9765, 2025.

The field of earthquake prediction presents significant challenges, with effective prediction techniques are still elusive. Taiwan’s high seismicity is due to the collision of the Philippine Sea plate with the Eurasian plate, leading to frequent earthquakes every year. To address challenges in analyzing pre-earthquake strain transfer, number of radon monitoring stations were established across various tectonic zones. Using open-source software a Real-Time Database was developed for radon earthquake precursory research and tested for some major earthquakes occurred in Taiwan. This database enables faster processing of precursor data, hence, enhancing the efficiency related investigations.

Notably, large earthquakes, such as Meinong earthquake in Southern Taiwan, exhibited precursory signals in radon concentrations, with significant variations observed in soil radon concentrations about two weeks prior. The study suggests that variations in soil radon concentrations at different locations may help in predicting the general area of impending major earthquakes. Finding  from research indicates that soil-gas anomalies were associated with few earthquakes with a magnitude of 5 or more that were recorded during the study period. Stress accumulation and changes in strain fields prior to seismic events  are likely linked to these variations in soil radon levels.

Overall, soil radon measurements have emerged as a promising and practical tool  for investigating earthquake precursors in Taiwan. However, further research and validation are essential to refine these findings and advance the development of reliable earthquake prediction method.

How to cite: Walia, V., Fu, C.-C., Lin, S.-J., and Kamra, P.: Radon Monitoring Data in Taiwan: Statistical Perspective to Investigate Earthquake Precursory Studies, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-16828, https://doi.org/10.5194/egusphere-egu25-16828, 2025.

Analysis of seismic precursors is crucial for understanding the spatio-temporal evolution of seismicity and assessing whether a system is approaching an unstable state. Precursors are indicators that are deemed to be related to the processes leading to crustal rupture. Therefore, their real-time monitoring can provide insights into the imminent occurrence of earthquakes. Precursors yield robust results only when analyzed using appropriate techniques. Specifically, the measurement of real-time precursor parameters and the analysis of their temporal trends is highly sensitive to data processing and depends heavily on the characteristics of the seismic data under study. Therefore, careful data management is essential to avoid inappropriate conclusions.

This study examines two seismic precursors: (1) b-value (i.e., slope of the Gutenberg and Richter law), which characterizes the relative likelihood of small versus large earthquakes within a population of events; (2) Hurst exponent, an indicator of the "memory" in time series and, consequently, of the type of stochastic process underlying them (i.e., random, persistent, or anti persistent). While the b-value has paramount importance in earthquake forecasting since its variation (which is deemed to be related to stress conditions of faults) can act as a first-order discriminator between conventional aftershock sequences and sequences including multiplets (i.e., two or more mainshocks that are closely associated in time and space), the Hurst exponent is widely used in econometrics to detect trends and mean reversion in financial data.

The aim of this study is to determine the optimal data-windowing configuration (window size and overlapping percentages), within the framework of a moving window approach, that produces results in agreement with theoretical expectations and effectively captures the characteristics of the seismic data under study. The methodology involves analyzing these precursors (i.e., b-value and Hurst Exponent) along with additional seismic metrics such as: (1) number of earthquakes above a given magnitude threshold, (2) maximum magnitude, and (3) strain energy. Correlation coefficients (e.g., Pearson, Spearman, Kendall) are computed to evaluate the relationships between these parameters under various windowing configurations, including a novel adaptive window approach. The process can be summarized as follows:

• For a given configuration (e.g., window size = 500 years, overlap = 50%), the window slides over time, and parameters are calculated at each step within the time window.
• Correlation coefficients (between pairs of parameters) are computed using various statistical methods (e.g., Pearson, Spearman, Kendall).
• This procedure is repeated across all configurations to identify the setting that maximizes correlations (i.e., higher correlation coefficients and lower p-values).

Analyzing synthetic time series shows that correlations between parameters are sensitive to the adopted configuration, as different data-windowing configurations may capture distinct seismicity patterns. Therefore, the selection of the most effective configuration is strictly study-specific. To enhance the reliability of the results, application of this methodology to other seismic parameters (e.g., Vp/Vs ratio) requires future consideration, especially to validate results against known seismic sequences. The first step towards this direction is the application of the method to time series associated with natural or induced (or triggered) seismicity (e.g., seismic activity in the Geysers geothermal field). 

How to cite: Azhideh, S., Barani, S., Ferretti, G., Taroni, M., Resta, M., and Massa, M.: Optimization of rolling window approach to analyze earthquake time series and identify possible precursors, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17245, https://doi.org/10.5194/egusphere-egu25-17245, 2025.

How to cite: Zhuang, J.: Evaluating Earthquake Forecast with Likelihood-Based Marginal and Conditional Scores, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-21558, https://doi.org/10.5194/egusphere-egu25-21558, 2025.

Probabilistic earthquake forecasting has made significant strides in the past decades, to the degree that government agencies around the world have implemented public, real-time systems. The underlying models are largely parametric and statistical, and comprise variants of the self-exciting Hawkes point process, such as the popular Epidemic Type Aftershock Sequence (ETAS) model. ETAS models have gained trust also as a result of their good relative performance in prospective forecast experiments by the Collaboratory for the Study of Earthquake Predictability (CSEP, cseptesting.org) in various tectonic settings around the globe. In recent years, however, machine learning variants of point processes have become available that offer significant advantages: they are much more flexible in their probabilistic description of earthquake interaction, and they are much faster. In this talk, I will review recent applications of neural point processes to seismicity forecasting around the world, which demonstrate distinct advantages and some (moderate) improvement in predictive skill. I will also argue that a clear community benchmarking process is required to make transparent and robust progress. Finally, I will present ongoing model enhancements of neural point processes and preliminary results from benchmarking in California. Machine learning has the potential of transform earthquake forecasting, but progress must be demonstrated in a robust and transparent manner. 

How to cite: Werner, M., Stockman, S., and Lawson, D.: Advancing Earthquake Forecasting with Machine Learning , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19183, https://doi.org/10.5194/egusphere-egu25-19183, 2025.