Abyssal marine turbidites provide some of the longest and most spatially extensive records of subduction zone earthquake recurrence globally; however, correlation of these deposits over long distances and interpretation of synchronous emplacement requires both an understanding of the turbidite generating systems and precise dating. Here, we present an integrated suite of high-resolution bathymetry, subbottom profiles, and sediment cores from combined autonomous underwater vehicle, remotely operated vehicle, and ship-based studies at a key paleoseismic site in the southern Cascadia subduction zone. We demonstrate how widespread, earthquake-triggered landslides on the lower slope deposit discrete, proximal mass transport deposits (MTDs) that grade offshore into complex, interfingered abyssal turbidites, which correspond to records of megathrust earthquake history. We propose accretion and oversteepening of thrust folds on the lower slope both preconditions the slope to fail and provides a perpetual source of unstable material to fail during every earthquake cycle. Furthermore, we suggest the periodic and pervasive landsliding indicates coseismic deformation of the outer accretionary wedge during megathrust rupture.

ULF signals as earthquake precursors in various tectonic regions

Abstract Despite the devastating impact of earthquakes, they offer potential for machine learning prediction to mitigate damage. This study explores the application of common algorithms like Random Forests, Support Vector Machines (SVMs), XGBoost, and Long Short-Term Memory (LSTM) networks alongside the Autoregressive Integrated Moving Average (ARIMA) framework for earthquake frequency forecasting in Indonesian regions. A novel hybrid model combining machine learning with ARIMA for multi-step forecasting is introduced. Surprisingly, the LSTM model, renowned for its strong predictive capability in nonlinear relationships, performed significantly lower than traditional machine learning methods based on metrics such as Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). The results highlight the superior predictive capability of the hybrid ARIMA-XGBoost and ARIMA-RandomForest models in multi-step forecasting. These findings underscore the continued relevance and effectiveness of traditional machine learning approaches in earthquake data prediction, suggesting avenues for future research to refine hybrid models and improve multi-step regression forecasting accuracy.

Indonesia is a country with high seismic activity due to its location at the convergence of three major tectonic plates. This condition creates a strong need for earthquake pattern analysis and magnitude prediction to support disaster mitigation. This study aims to cluster earthquake data using the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm and to predict earthquake magnitude using the Random Forest algorithm optimized through hyperparameter tuning. The Indonesian earthquake dataset was obtained from Kaggle with a total of 92,887 valid entries. The DBSCAN clustering results revealed several active seismic zones, particularly in Sumatra, Java, Sulawesi, and Papua. The comparison of R² between the Baseline Random Forest and the Tuned Random Forest shows a significant improvement after the parameter tuning process. The Tuned Random Forest model achieves an R² value of 0.478, which is higher than the Baseline Random Forest's 0.442. This indicates that the tuned model is better able to explain the variance in the data and provides more accurate predictions.

Numerical simulation of stress transfer and triggered earthquake prediction along the Longmenshan-Foreland Basin driven by the Wenchuan Ms 8.0 earthquake

Ranking Earthquake Prediction Algorithms: A Comprehensive Review of Machine Learning and Deep Learning Methods

ABSTRACT Multi-precursor analysis, along with multi-method analysis, has made it possible to detect a large number of LAI seismic anomalies in the study of strong earthquake-affected areas. It seems necessary to use fusion-based methods to combine the detected anomalies and earthquake parameters (magnitude, time and location) obtained from the analysis of different precursors and extract the final earthquake parameters along with their uncertainty. This study is carried out in two phases, which are: 1) the deviation values of about 1030 LAI anomalies detected using 20 implemented predictor algorithms around the time and location of 23 powerful earthquakes that occurred in recent years have been considered. Various functions were fitted on the collected data, including the anomaly intensity, and real magnitude of the earthquake, and a functional model was developed to estimate magnitude parameter with RMSE of about 0.53. 2) Using the Dempster-Shaffer (DS) Theory, the predicted earthquake magnitudes from different precursors for three strong earthquakes including Turkey (February 06, 2023), Myanmar (March 28, 2025) and Russia (July 29, 2025) were fused together and the probability of the estimated magnitude values were presented. The results show that DS was able to estimate the magnitude of the upcoming earthquake with a desirable uncertainty.

Earthquakes pose significant threats to human life and infrastructure worldwide, motivating researchers to investigate potential precursory signals that may indicate impending seismic events. This study focuses on evaluating the capability of Long Short-Term Memory (LSTM) neural networks for time series prediction of IONOLAB-Total Electron Content (TEC) variations using Global Navigation Satellite Systems (GNSS) measurements during major earthquake events. We analyze IONOLAB-TEC data from 19 GNSS stations across Japan and China for four significant earthquakes: the 2011 Tohoku earthquake (Mw 9.1), the 2008 Iwate-Miyagi Nairiku earthquake (Mw 6.9), the 2008 Sichuan earthquake (Mw 7.9) and the 2010 Yushu earthquake (Mw 6.9). The LSTM model is trained using 10 days of TEC observations with 2.5-minute temporal resolution (576 observations per day) to forecast TEC values on the 11th day (earthquake day). Model performance is evaluated using three complementary metrics: Mean Squared Error (MSE), Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE). Results demonstrate high prediction accuracy across most stations, with MAPE values below 2% for 14 out of 19 stations. The best performances are achieved at MIZU station (MAPE: 0.45%) for the Tohoku earthquake and XIAN station (MAPE: 0.62%) for the Sichuan earthquake. Overall MSE values range from 0.0036 to 0.4493 and MAE values range from 0.0420 to 0.5014. The findings demonstrate that LSTM networks can effectively learn and reproduce temporal patterns in IONOLAB-TEC time series, accurately capturing diurnal variations and magnitude fluctuations. This study demonstrates IONOLAB-TEC prediction capability rather than earthquake precursor detection or operational early warning capability.

New Zealand's vulnerability to seismic hazards highlights the need for systems capable of providing earthquake early warning (EEW) alerts or rapid notice of strong shaking. Large offshore earthquakes along the subduction zone east of the North Island could also trigger catastrophic tsunamis, inundating coastal communities in under an hour. Although New Zealand operates a robust seismic and geodetic network capable of monitoring moderate-to-large earthquakes, the limited observational record of large earthquakes poses challenges for EEW design and response. This study evaluates magnitude estimation from G-FAST, an early warning algorithm that uses Global Navigation Satellite System (GNSS) data to characterize earthquake sources. We analyze synthetic rupture scenarios along the Hikurangi subduction margin generated by the earthquake simulator RSQSim. For each rupture, GNSS displacements are generated at each site and compared with Peak Ground Displacement (PGD) scaling relationships to test whether they replicate real earthquakes. While we also assess PGD values from rupture scenarios produced with simpler semi-stochastic kinematic modeling, those from RSQSim yield ground motions more consistent with expected values. Given these results, synthetic displacement data from RSQSim ruptures were ingested into G-FAST to evaluate performance for rapid earthquake characterization, finding that PGD-based estimates capture moment magnitude in 90% of cases. This framework demonstrates the utility of synthetic catalogs for testing geodetic EEW performance in characterizing large subduction earthquakes in the North Island region and provides a path- way toward tsunami early warning procedures.

Western Java’s ongoing seismic hazard highlights the need for understanding earthquake precursor mechanisms. Recent studies have increasingly focused on ultra-low-frequency (ULF) signals that may carry information related to pre-seismic phases. However, a key difficulty persists: isolating faint, localized lithospheric signals from the stronger ionospheric activity. Most previous investigations in Western Java have relied on single-sensor measurements, a limitation that complicates the detection of true anomalies. This study addresses this limitation by examining daily ULF variations before the 2023 Banten earthquake sequence (M5.4 and M5.1), using a multi-point setup to distinguish lithospheric signals from stronger background ionospheric noise. Continuous three-component geomagnetic data from two primary stations near the epicenter, Serang (SRG) and Sukabumi (SKB), and a distant reference station (TRD) in East Kalimantan were analyzed. The Z/G spectral density ratio was calculated in the 0.01–0.09 Hz range, using only data from quiet nighttime intervals (15:00–21:00 UTC) and magnetic storm-free days (Dst > -50 nT). The results identified and filtered false positive anomalies by correlating them with signals at the TRD reference station. Two distinct, validated pre-seismic anomalies were identified, concentrated in the 0.04–0.08 Hz band: a multi-station anomaly at H-20 (at SRG and SKB) and a localized, broadband anomaly at H-15 (at SRG). Both emissions were absent at TRD, confirming their lithospheric origin. These results highlight the importance of a multi-station approach for reliably identifying lithospheric ULF anomalies. However, this study is limited to a specific event sequence. Future investigations should focus on integrating broader sensor networks and ionospheric models across multiple seismic events to validate these findings globally and enhance false positive rejection methods.

The article summarizes the experience of many years of research related to the creation of a geophysical system capable of providing sufficiently effective short-term prediction of seismic hazards for the Avacha Bay area, where the largest settlements of Kamchatka are located. The authors use an approach in which the main goal is not accurate short-term prediction of strong earthquakes, but a qualitative («low», «elevated», «significantly elevated») assessment of the current seismic hazard for the controlled area. Qualitative assessments of seismic hazard are based on the current results of continuous monitoring of the preparation of strong earthquakes in the Avacha Bay area, which can cause significant tremors in the Petropavlovsk-Kamchatsky area. This approach makes it possible to inform the authorities and structures of the Ministry of Emergency Situations in a timely manner if there are grounds for a significant probability of concussions of at least 5 points in the Petropavlovsk-Kamchatsky area for the next 2–4 weeks. Igor Dobrovolsky's consolidation model is used as a basic physical model for the preparation of strong Kamchatka earthquakes.

Spatiotemporal point processes model discrete events distributed in space and time, with applications in criminology, seismology, epidemiology, and social networks. Classical models rely on parametric kernels, limiting their ability to capture heterogeneous, nonstationary dynamics. Recent advances integrate deep neural architectures, either by modeling the conditional intensity directly or by learning flexible, data-driven influence kernels. This article reviews the deep influence kernel approach, which balances statistical interpretability by retaining explicit kernels to capture event propagation, with expressive power from neural architectures. We outline key components, including functional basis decomposition, graph neural networks for encoding spatial or network structures, and both likelihood-based and likelihood-free estimation methods, while addressing scalability for large data. We also highlight theoretical results on kernel identifiability. Applications in crime analysis, earthquake aftershock prediction, and sepsis modeling demonstrate the framework's effectiveness. We conclude with promising directions for developing explainable and scalable deep kernel point processes.

Summary We present a probabilistic framework for evaluating earthquake forecasting models that use an alarm-based approach. In this approach, alarms are triggered by specific precursor signals. In a previous paper we compared such models and two ensemble models combining them in additive and multiplicative mode, with the ETAS (Epidemic Type Aftershock Sequence) forecasting model, which is defined in a probability-based approach, by making the latter to issue an alarm when the expected rate exceeds a predefined threshold. In this work we compare the alarm-based models with the ETAS and with another probability-based model, EEPAS (Every Earthquake a Precursor According to Scale) previously applied to Italy, using the testing procedures developed for probability-based models within the Collaboratory Study for Earthquake Predictability (CSEP) initiative. To do that, for the four alarm-based models, we compute empirical probabilities (frequencies) of Mw ≥ 5.0 earthquakes in Italy, inside and outside alarm time intervals issued by such models from 1990 to 2011. We then compare pseudo-prospectively the forecasting ability of all six models, by applying the CSEP tests on the time interval from 2012 to 2023. We found that the evaluation method used has a strong impact on the ranking of model performance. Probabilistic models like ETAS and EEPAS tend to score better under the CSEP testing framework whereas alarm-based models generally outperform probability-based ones when assessed using alarm-based metrics.

Earthquake prediction remains one of the central unsolved problems in geophysics, and ionospheric variability offers a promising yet debated window into the earthquake preparation process through lithosphere–atmosphere–ionosphere coupling. Progress has been hindered by methodological limitations in prior studies, including the use of inappropriate performance metrics for highly imbalanced seismic data, the reliance on geographically and temporally narrow data, and inclusion of inherent spatial or temporal features that artificially inflate model performance while preventing the discovery of genuine ionospheric precursors. To address these challenges, we introduce a global, temporally validated machine learning framework grounded in thirty-eight years of ionospheric observations from more than a hundred ionosonde stations. We eliminate lookahead bias through strict temporal partitioning, prevent overlapping precursor windows across samples to eliminate autocorrelation artifacts and apply sophisticated feature selection to exclude spatial and temporal identifiers, enabling prevention of data leakage and coincidence effects. We investigate whether spatiotemporally invariant ionospheric precursors exist across diverse seismic regions, addressing the field’s reliance on geographically isolated case studies. Cross-regional validation shows that our models yield modest classification skill above chance levels, with our best-performing model achieving a weighted F1 score of 71% though performance exhibits pronounced sensitivity to temporal validation configuration, suggesting these results represent an upper bound on operational accuracy. While multimodal fusion with complementary precursor channels could possibly improve performance, our focus remains on establishing whether ionospheric observations alone contain learnable, region-independent seismic signatures. These findings suggest that ionospheric precursors, if they exist as universal phenomena, exhibit weaker cross-regional consistency than previously reported in case studies, raising questions about their standalone utility for earthquake prediction while indicating potential value as one component within multimodal observation systems.

The possibility of implementing a medium-term earthquake forecast using the LURR method in the southern part of Sakhalin Island in certain segments of the West Sakhalin and Central Sakhalin faults is considered in light of the medium-term strong earthquake forecast mode (M=5.5 (±0.5)) currently in effect for the central and southern parts of Sakhalin Island (within the latitude range from 47°N to 49.5°N, longitude range from 141.5°E to 143°E) (Protocol No. 2 of the Sakhalin Branch of the Russian Expert Council on Earthquake Prediction dated April 9, 2025). The assessments were carried out within the framework of a two-stage approach to seismic event forecasting, when previously identified medium-term forecast zones using the calculation method are refined using short-term geophysical methods. For the short-term assessment stage, the study used data from electrotelluric potential (vertical component) and subsurface radon volumetric activity measurement sites within the medium-term forecast coverage area. It was shown that in 2025, synchronous increases in electrotelluric potentials were observed at electrotelluric potential measurement sites in the village of Kolkhoznoye (Nevelsky District) and in Yuzhno-Sakhalinsk, which characterizes periods of high seismic activity. Furthermore, anomalies were detected based on subsurface radon volumetric activity measurements at sites in Yuzhno-Sakhalinsk and Firsovo before the earthquake activation in late June 2025 (a double earthquake near Aniva and Bykov). Updated LURR data (July 2025) showed that the only anomalies over the past five years were recorded in mid-2023. Although this study does not identify short-term precursors, it does provide additional data on the persistence of unstable conditions and confirms the relevance of the proposed medium-term forecast based on direct geophysical measurements.

ABSTRACT Earthquake prediction has long been a profound challenge for earthquake science. The devastating consequences of earthquakes on human lives and property necessitate the continuous pursuit of more accurate prediction methods. This study employs and compares three deep learning techniques, including Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Bidirectional Long Short-Term Memory (BiLSTM) to estimate earthquake magnitudes in the southeastern Tibetan Plateau. These models' performance was evaluated using Mean Squared Error (MSE), Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE). The results indicate that the BiLSTM model outperformed other two comparative models with the smallest prediction deviation. We then integrated multi-head attention mechanism and Bayesian self-optimization to enhance both the prediction accuracy and model generalization of the BiLSTM model. Retrospective accuracy of earthquake magnitude estimation using the seismic catalog (2014-2024) in the southeastern Tibetan Plateau achieves 92.1%, with MSE = 0.08, MAE = 0.22, and RMSE = 0.28, demonstrating high effectiveness of the model. The results highlight the effectiveness of deep learning techniques in earthquake magnitude estimation, and can promote the development of data-driven earthquake prediction.

Predicting future seismic trends and occurrence of earthquakes remains a long-standing challenge in seismology. Despite substantial efforts to unravel the physical mechanisms underlying earthquake occurrence, currently, no well-defined physical or statistical model is capable of reliably predicting major earthquakes. However, machine learning methods have demonstrated exceptional proficiency in identifying patterns within large-scale datasets, offering a promising avenue for enhancing earthquake prediction performance. Within the framework of machine learning, this study has developed a feature extraction method based on seismic prediction zoning, improving the effectiveness of machine learning-based earthquake prediction. The research findings indicate that the ensemble learning Stacking method, which is based on seismic prediction zoning, exhibits superior performance and high robustness in predicting the annual maximum earthquake magnitude. Additionally, the long short-term memory (LSTM) method demonstrates commendable performance within specific tectonic zones (e.g., the southwestern Yunnan region), providing valuable guidance for analyzing seismic trends in these regions.

Accurate predictions of earthquakes are crucial for disaster preparedness and risk mitigation. Conventional machine learning models like Random Forest, SVR, and XGBoost are frequently used for seismic forecasting; however, capturing the intricate spatiotemporal relationships in earthquake data remains a challenge. To overcome this issue, we propose SeismoQuakeGNN, a novel Graph Neural Network (GNN) and Transformer-based hybrid framework that integrates spatial and temporal learning for improved seismic forecasting. Unlike existing GNN-based models, SeismoQuakeGNN introduces an optimized spatial encoding mechanism to dynamically learn seismic interdependencies, coupled with a Transformer-driven attention module to capture long-range temporal correlations. Furthermore, initial experiments with XGBoost demonstrated its limitations in learning earthquake patterns, reinforcing the need for deep spatial–temporal modeling. The new SeismoQuakeGNN method is capable of substantial and efficient data processing of relationships in both space and time, as well as providing superior transfer to different seismic areas, thereby qualifying as a dependable starting point to extensive earthquake forecasting and hazard evaluation.

TEC variation as earthquake precursor: A statistical and SARIMA-based study from Northeast India

Abstract Whether and when earthquakes of different sizes can be distinguished early in their rupture process is critical to improving earthquake early warning (EEW) systems. To address this question, we develop a transformer‐based framework called source time function Magnitude Network (STF‐MgNet), which leverages earthquake STFs with M w 5.5 to investigate whether the initial stages of the rupture process can predict the earthquake's final magnitudes. The proposed STF‐MgNet utilizes transformer blocks in conjunction with a U‐Net architecture to effectively manage long‐range dependencies and capture essential information in data sets, boosting its performance. Training on 2,126 global M w ≥ 5.5 earthquakes reveals three distinct nucleation‐phase diagnostic regimes: (a) For 5.5 M w ≤ 7 events, subsecond analysis of the STF achieves 80.0% magnitude estimation accuracy; (b) For 7 < M w ≤ 8 earthquakes, 3–5 s of observations capture key features of STFs, yielding 86.9% accuracy; (c) For M w > 8 events, approximately 5 s of monitoring resolves the interplay between determinism and stochasticity, maintaining 83.7% accuracy despite greater complexity. Thus, no matter how an earthquake begins, earthquakes of different sizes can be distinguished at some point. Moreover, the earthquake magnitude prediction accuracy of STF‐MgNet improves with increasing input sequences of STFs. This study supports the hypothesis that rupture onset at nucleation differs for earthquakes of different final magnitudes.