Earthquake Forecasting Research Articles

Application of Neural Network Technologies for Tectonic Earthquake Forecast

Successful earthquake prediction includes statistical, tectonic and physical forecasting. The main requirements for this are the establishment of the laws of earthquake mechanics and control of the geodynamic state in the region at the right times. However, resolving this issue faces difficulties of both theoretical and practical nature. Despite the fact that specialists all over the World have collected the fairly complete database on earthquakes, tectonic, electromagnetic, hydrological, etc. signs of earthquakes, the very nature of predicting the future source remains uncertain. The results obtained in the world on statistical forecasting using artificial intelligence give hope for the possibility of predicting earthquakes if we combine tectonic forecasting with the destruction of materials under experimental conditions and numerical modeling under the roof of deep learning neural network technologies. The paper provides the first results of predicting medium-term tectonic earthquakes using artificial intelligence for the Fergana depression in Uzbekistan.

A Novel Method for Evaluating Earthquake Forecast Model Performance and Its Implications for Refining Seismic Likelihood Model

Summary Utilizing statistical tests to evaluate earthquake forecasting models is crucial to improve forecasting strategies for seismic hazard assessment. We develop a novel evaluation method for alarm-based earthquake forecast, taking into account the magnitude of seismic energy and the impact area of earthquakes, instead of using solely seismic event number and epicenter locations in conventional approaches. First, we derive a scale-law of Seismic Area by statistically analyzing coseismal maps of past M ≥ 7.0 earthquakes. Second, we proportionally allocate Seismic Moment to surrounding cells based on corresponding seismic area within each cell (SASM-test). Compared to the Molchan test which is conventionally applied for models that forecast the epicenter location, our proposed SASM-test can be applied to the evaluation of forecasting models that focus on the whole earthquake rupture (source area). Third, we apply the SASM-test method to the time-independent probabilistic earthquake forecasting model for the southeastern Tibetan Plateau (RELM-TibetSE) and compare it with other evaluation methods. The retrospective testing shows that the SASM-test demonstrate relatively higher sensitivity, enabling to detect subtle differences between similar models that conventional methods may overlook. Additionally, retrospective test results indicate that: (a) Earthquake forecasting models using GNSS data performed better in forecasting the ‘source area’ than the ‘epicenter location’; (b) Forecasting models based on principal strain rate outperformed the models based on maximum shear strain rate in forecasting both the epicenter location and the source area; (c) Incorporating spatially varying seismogenic layer thickness and rigidity into seismic forecasting models could improve their ability to forecast the ‘source area’ compared to using uniform seismogenic layer properties. The newly proposed SASM-test method can provide a more sensitive and comprehensive approach for the evaluation of earthquake forecasting models, contributing to the refinement of seismic hazard assessments.

A model of the earthquake cycle along the Gofar oceanic transform faults

The Gofar oceanic transform fault at the East Pacific Rise has one of the best seismic cycles recorded by modern instruments. The timing, location, and magnitude of major earthquakes (Mw>5.5) have been well constrained by data from global seismic networks for the past 30 years. The earthquake interval is short, about 3-5 years. Several segments have already experienced 5 cycles since 1995, when the seismic network was good enough for surface wave relocation. Two ocean bottom seismometer deployments (2008-2009, 2021-2023) also provide constraints on the seismic properties on the fault. This makes Gofar an ideal place to study earthquake cycles. Here, we developed a model for the seismic cycle along the Gofar transform fault using a semi-analytical approach for rapidly calculating 3D time-dependent deformation and stress caused by screw dislocations embedded within an elastic layer overlying a Maxwell viscoelastic half-space. The 160-km long fault is divided into three major segments with six asperities. Our model simulates the earthquake pattern on this fault for the past 30 years. Most of the time, each asperity ruptured as a large earthquake every 3-5 years. Most segments have a nearly constant Coulomb stress threshold of 2-3 MPa, providing optimal conditions for the forecasting of future earthquakes along Gofar. For three cases that deviated from this simple regular pattern, a large earthquake occurred with a centroid location between two asperities. This is likely due to concurrent rupture that involved both asperities. We also modeled surface deformation with different elastic layer thicknesses and mantle viscosities. Even though most deformation is in the horizontal direction, the difference in both horizontal and vertical directions between models can be as large as a few centimeters per year. Several seafloor geodesy methods can be used to differentiate between models, and seafloor pressure might be the most appropriate one at this remote location.

Optimization of the Approach to Systematic Earthquake Forecasting

Optimization of the Approach to Systematic Earthquake Forecasting

MUDA: dynamic geophysical and geochemical MUltiparametric DAtabase

Abstract. In this paper, the new dynamic geophysical and geochemical MUltiparametric DAtabase (MUDA) is presented. MUDA is a new infrastructure of the National Institute of Geophysics and Volcanology (INGV), published online in December 2023, with the aim of archiving and disseminating multiparametric data collected by multidisciplinary monitoring networks. MUDA is a MySQL relational database with a web interface developed in PHP, aimed at investigating possible correlations between seismic phenomena and variations in endogenous and environmental parameters in quasi real time. At present, MUDA collects data from different types of sensors such as hydrogeochemical probes for physical–chemical parameters in waters, meteorological stations, detectors of air radon concentration, diffusive flux of carbon dioxide (CO2) and seismometers belonging both to the National Seismic Network of INGV and to temporary networks installed in the framework of multidisciplinary research projects. MUDA publishes data daily, updated to the previous day, and offers the chance to view and download multiparametric time series selected for different time periods. The resultant dataset provides broad perspectives in the framework of future high-frequency and continuous multiparametric monitoring as a starting point to identify possible seismic precursors for short-term earthquake forecasting. MUDA can be accessed at https://doi.org/10.13127/muda (Massa et al., 2023).

Earthquake Magnitude Correlations Expose Short-Term Catalog Incompleteness

Abstract The independence of earthquake magnitudes is a fundamental assumption and limitation in earthquake forecasting. To assess its validity, we examine correlations between the magnitude of successive earthquakes. We first investigate the 2019 Ridgecrest foreshock sequence and find a significant magnitude correlation as well as an unusually low Gutenberg–Richter b-value (≈0.7 using moment magnitudes). We demonstrate that these anomalous features are not indicative of a precursory phase, but a consequence of short-term incompleteness (STI) that is not detected by conventional methods to estimate catalog completeness. Synthetic simulations of this sequence support this explanation: imposing STI leads to significant magnitude correlation and biased b-value estimates. Expanding our investigation to seismicity across southern California reveals pervasive magnitude correlation due to STI, not limited to sequences with large earthquakes. Our findings suggest that magnitude correlation is the most effective indicator of STI, rather than a characteristic of the underlying earthquake-generating process.

Physics informed neural network can retrieve rate and state friction parameters from acoustic monitoring of laboratory stick-slip experiments

Various machine learning (ML) and deep learning (DL) techniques have been recently applied to the forecasting of laboratory earthquakes from friction experiments. The magnitude and timing of shear failures in stick-slip cycles are predicted using features extracted from the recorded ultrasonic or acoustic emission (AE) signals. In addition, the Rate and State Friction (RSF) constitutive laws are extensively used to model the frictional behavior of faults. In this work, we use data from shear experiments coupled with passive acoustic (variance, kurtosis, and AE rate) interleaved with active source ultrasonic monitoring (transmitted wave amplitude) to develop physics-informed neural network (PINN) models incorporating the RSF law and AE rate generation equation with wave amplitude serving as a proxy for friction state variable. This PINN framework allows learning RSF parameters from stick-slip experiments rather than measuring them through a series of velocity step experiments. We observe that when the stick-slip cycles are irregular, the PINN models outperform the data-driven DL models. Transfer learning (TL) PINN models are also developed by pre-training on data collected at one normal stress level followed by forecasting shear failures and retrieving RSF parameters at other stress levels (i.e., with different recurrence intervals) after retraining on a limited amount of new data. Our findings suggest that TL models perform better compared to standalone models. Both standalone and TL PINN-estimated RSF parameters and their ground truth values show excellent agreements thus demonstrating that RSF parameters can be retrieved from laboratory stick-slip experiments using the corresponding acoustic data and that the transmitted wave amplitude provides a good representation of the evolving frictional state during stick-slips.

Forecasting future earthquakes with deep neural networks: application to California

SUMMARY We use the spatial map of the logarithm of past estimated released earthquake energies as input of fully convolutional networks (FCN) to forecast future earthquakes. This model is applied to California and compared with an elaborated version of the epidemic type aftershock sequence (ETAS) model. Our long-term earthquake forecast simulations show that the FCN model is close to the ETAS model in forecasting earthquakes with $M \ge 3.0,\,\,4.0,\,\,{\rm{and\,\,}}5.0$ according to the Molchan diagram. Moreover, training and implementing the FCN model is 2000–4000 times faster than calibrating the ETAS model and generating its probabilistic forecasts. The FCN model is straightforward in terms of its neural network structure and feature engineering. It does not require extensive knowledge of statistical seismology or the analysis of earthquake catalogue completeness. Using the earthquake catalogue with $M \ge 0$ as FCN input can enhance the model's performance in some time–magnitude forecasting windows.

Electromagnetic and Radon Earthquake Precursors

Earthquake forecasting is arguably one of the most challenging tasks in Earth sciences owing to the high complexity of the earthquake process. Over the past 40 years, there has been a plethora of work on finding credible, consistent and accurate earthquake precursors. This paper is a cumulative survey on earthquake precursor research, arranged into two broad categories: electromagnetic precursors and radon precursors. In the first category, methods related to measuring electromagnetic radiation in a wide frequency range, i.e., from a few Hz to several MHz, are presented. Precursors based on optical and radar imaging acquired by spaceborne sensors are also considered, in the broad sense, as electromagnetic. In the second category, concentration measurements of radon gas found in soil and air, or even in ground water after being dissolved, form the basis of radon activity precursors. Well-established mathematical techniques for analysing data derived from electromagnetic radiation and radon concentration measurements are also described with an emphasis on fractal methods. Finally, physical models of earthquake generation and propagation aiming at interpreting the foundation of the aforementioned seismic precursors, are investigated.

Critical Questions About CSEP, in the Spirit of Dave, Yan, and Ilya

Abstract In honor of our dear departed friends Yan Kagan, Dave Jackson, and Ilya Zaliapin, we propose a selection of broad questions regarding earthquake forecasting and especially the Collaboratory for the Study of Earthquake Predictability (CSEP) in particular and give our thoughts on their answers. This article reflects our opinions, not necessarily those of Yan Kagan, Dave Jackson, and Ilya Zaliapin, and not necessarily those of the seismological community at large. Rather than provide definitive answers, we hope to provoke the reader to think further about these important topics. We feel that Dave Jackson in particular might have liked this approach and may have seen this as an appropriate goal.

Algorithmic Identification of the Precursory Scale Increase Phenomenon in Earthquake Catalogs

Abstract The precursory scale increase (Ψ) phenomenon describes the sudden increase in rate and magnitude in a precursory area AP, at precursor time TP, and with precursor magnitude MP prior to the upcoming large earthquake with magnitude Mm. Scaling relations between the Ψ variables form the basis of the “Every Earthquake a Precursor According to Scale” (EEPAS) earthquake forecasting model. EEPAS is a well-established space–time point process model that forecasts large earthquakes in the medium term, that is, the coming months to decades, depending on Mm. In Aotearoa New Zealand, EEPAS contributes to hybrid models for public earthquake forecasting and to the source model of time-varying seismic hazard models, including the latest revision of the National Seismic Hazard Model. The Ψ phenomenon was recently shown not to be unique for a given earthquake, with smaller precursory areas AP associated with larger precursor times TP and vice versa. This trade-off between AP and TP has also been found for the spatial and temporal distributions of the EEPAS models. Detailed analysis of the Ψ phenomenon has so far been limited by the manual and labor-intensive procedure of identifying Ψ in earthquake catalogs. Here, we introduce two algorithms to automatically detect Ψ and apply them to real and simulated earthquake catalog data. By randomizing the catalog and removing aftershocks, we confirm that the Ψ phenomenon is a feature of space–time earthquake clustering prior to major earthquakes. Multiple Ψ identifications confirm the trade-off between AP and TP, and the scaling relations for both real and simulated catalogs are consistent with the original scaling relations on which EEPAS is based. We identify opportunities for future work to refine the algorithms and apply them to physics-based simulated catalogs to enhance the understanding of Ψ. A better understanding of Ψ has the potential to improve forecasting of large upcoming earthquakes.

Earthquake Predictability and Forecast Evaluation Using Likelihood-Based Marginal and Conditional Scores

Abstract Earthquake probability forecasts are typically based on simulations of seismicity generated by statistical (point process) models or direct calculation when feasible. To systematically assess various aspects of such forecasts, the Collaborative Studies on Earthquake Predictability testing center has utilized N- (number), M- (magnitude), S- (space), conditional likelihood-, and T- (Student’s t) tests to evaluate earthquake forecasts in a gridded space–time range. This article demonstrates the correct use of point process likelihood to evaluate forecast performance covering marginal and conditional scores, such as numbers, occurrence times, locations, magnitudes, and correlations among space–time–magnitude cells. The results suggest that for models that only rely on the internal history but not on external observation to do simulation, such as the epidemic-type aftershock sequence model, test and scoring can be rigorously implemented via the likelihood function. Specifically, gridding the space domain unnecessarily complicates testing, and evaluating spatial forecasting directly via marginal likelihood might be more promising.

Bayesian Earthquake Forecasting Using Gaussian Process Modeling: GP-ETAS Applications

Abstract Numerous seismicity models are known to simulate different observed characteristics of earthquake occurrence successfully. However, their ability of prospective forecasting future events is a priori not always known. The recently proposed semiparametric model, Gaussian process epidemic-type aftershock sequence (GP-ETAS) model, which combines the ETAS model with GP modeling of the background activity, has led to promising results when applied to synthetic seismicity. In this study, we focus on the ability of GP-ETAS for different forecasting experiments in two case studies: first, the Amatrice, Italy, sequence during 2016 and 2017, and second, long-term seismicity in Southern California. The results indicate that GP-ETAS performs well compared with selected benchmark models. The advantages become particularly visible in cases with sparse data, in which GP-ETAS shows in general a more robust behavior compared to other approaches.

New Features in the pyCSEP Toolkit for Earthquake Forecast Development and Evaluation

Abstract The Collaboratory for the Study of Earthquake Predictability (CSEP) is a global community dedicated to advancing earthquake predictability research by rigorously testing probabilistic earthquake forecast models and prediction algorithms. At the heart of this mission is the recent introduction of pyCSEP, an open-source software tool designed to evaluate earthquake forecasts. pyCSEP integrates modules to access earthquake catalogs, visualize forecast models, and perform statistical tests. Contributions from the CSEP community have reinforced the role of pyCSEP in offering a comprehensive suite of tools to test earthquake forecast models. This article builds on Savran, Bayona, et al. (2022), in which pyCSEP was originally introduced, by describing new tests and recent updates that have significantly enhanced the functionality and user experience of pyCSEP. It showcases the integration of new features, including access to authoritative earthquake catalogs from Italy (Bolletino Seismico Italiano), New Zealand (GeoNet), and the world (Global Centroid Moment Tensor), the creation of multiresolution spatial forecast grids, the adoption of non-Poissonian testing methods, applying a global seismicity model to specific regions for benchmarking regional models and evaluating alarm-based models. We highlight the application of these recent advances in regional studies, specifically through the New Zealand case study, which showcases the ability of pyCSEP to evaluate detailed, region-specific seismic forecasts using statistical functions. The enhancements in pyCSEP also facilitate the standardization of how the CSEP forecast experiments are conducted, improving the reliability, and comparability of the earthquake forecasting models. As such, pyCSEP exemplifies collaborative research and innovation in earthquake predictability, supporting transparent scientific practices, and community-driven development approaches.

Predicting the Unpredictable: Advancements in Earthquake Forecasting Using Artificial Intelligence and LSTM Networks

Predicting the Unpredictable: Advancements in Earthquake Forecasting Using Artificial Intelligence and LSTM Networks

Exploiting the synergy of SARIMA and XGBoost for spatiotemporal earthquake time series forecasting

AbstractEarthquakes are vibrations that occur on the surface of earth, generating fires, ground shaking, tsunamis, landslides and cracks. These incidents can cause severe damage and loss of life. Accurate earthquake forecasts are critical for anticipating and mitigating these hazards, which can avoid damage to buildings and infrastructure and save lives. To address the challenges given by earthquakes probabilistic nature, this paper presents a hybrid SARIMA–XGBoost approach to earthquake magnitude prediction. The suggested technique consists of a two‐step process: an exploration phase that uses exploratory data analysis, which includes descriptive statistics and data visualisation, and a prediction phase that focusses on forecasting future earthquakes. Using a large significant earthquake dataset spanning 1965–2023, the study intends to gain insights and lessons for more effective earthquake prediction methods. Further, in a comparison analysis, the results of SARIMA‐XGBoost model are compared to those of traditional ARIMA and SARIMA models. The results highlight the superior performance of the hybrid SARIMA–XGBoost model, showcasing a mean absolute error (MAE) of 0.038, a mean squared error (MSE) of 0.0040, and a root mean squared error (RMSE) of 0.068. These metrics collectively underscore the model's enhanced accuracy in forecasting earthquake magnitudes. The notably low values of MAE, MSE and RMSE indicate that our hybrid approach significantly improves prediction accuracy compared to alternative models. By integrating SARIMA's time series (TS) analysis with XGBoost's machine learning (ML) capabilities, the hybrid model reduces forecasting errors more effectively, demonstrating its clear advantage in precision.

Electromagnetic Trigger Effects in the Ionosphere–Atmosphere–Lithosphere System and Their Possible Use for Short-Term Earthquake Forecasting

Electromagnetic Trigger Effects in the Ionosphere–Atmosphere–Lithosphere System and Their Possible Use for Short-Term Earthquake Forecasting

Catalog of focal mechanism solutions for the Sichuan and Yunnan region from 2012 to 2022 using the community velocity model of Southwest China

The focal mechanism solution is one of the important focal parameters for exploring fault activity and studying regional stress distribution and it has a wide range of applications. The geological structure of the Sichuan-Yunnan region in China is complex, with frequent earthquakes and abundant historical observation data, making it one of the popular areas of concern for scholars. This study utilizes the high-precision community velocity model v2.0 of southwest China, obtained through joint inversion based on multiple data methods. The Cut-And-Paste (CAP) method was employed to fit and invert the observed waveforms of 1475 events with ML ​≥ ​3.5 in the Sichuan-Yunnan region from January 2012 to December 2022, thereby constructing a catalog of double-couple focal mechanisms. By comparing the focal mechanism inversion results of small earthquakes with those from multiple one-dimensional velocity models and conducting comparative statistical analysis on events below magnitude 4, it has been demonstrated that the model used in this study provides a better fit than one-dimensional models. This contributes to establishing the lower magnitude limit for producing deeper focal mechanism solutions. This study compares the results of larger magnitude earthquakes in the catalog with those published by the Global Centroid-Moment Tensor (GCMT) project and smaller magnitude earthquakes with the catalog released by the Institute of Earthquake Forecasting, China Earthquake Administration. These comparisons serve to validate the accuracy of the catalog results. Leveraging the high-resolution velocity model, this catalog has re-examined the historical earthquake focal mechanism catalog of the Sichuan-Yunnan region. The inversion has yielded reliable results for smaller magnitudes and a greater number of events, providing additional data and support for understanding the regional stress field, active faults, the mechanisms of large earthquake genesis, and earthquake prediction efforts. Consequently, this enhances the depth of scientific research in the Sichuan-Yunnan region.

A combining earthquake forecasting model between deep learning and epidemic-type aftershock sequence (ETAS) model

SUMMARY The scientific process of earthquake forecasting involves estimating the probability and intensity of earthquakes in a specific area within a certain timeframe, based on seismic activity features and observational data. Among the various methodologies, epidemic-type aftershock sequence (ETAS) models, rooted in seismic empirical laws, stand as widely used tools for earthquake forecasting. In this study, we introduce the CL-ETAS model, a novel approach that integrates convolutional long short-term memory (ConvLSTM), a deep learning model, with the ETAS model. Specifically, we leverage the forecasting outputs of ETAS to enhance both the training and forecasting processes within the ConvLSTM framework. Through forecasting tests, our findings illustrate the effectiveness of the CL-ETAS model in capturing the trends observed in earthquake numbers ($M \ge 3$) in Southern California following three main shocks. Overall, our model outperforms both a simple ETAS model and ConvLSTM in this context.

A Review of 15 Years of Public Earthquake Forecasting in Aotearoa New Zealand

Abstract New Zealand entered a period of increased seismic activity in 2003, which, to date, has included around 20 large or damaging earthquakes. Building on decades of forecast model research and including research into formal model evaluation, GNS Science, the governmental institution tasked with providing natural hazards information, began issuing public earthquake forecasts following the 2009 Mw 7.2 Darfield earthquake. This article provides a review of the public earthquake forecasting methods and outcomes since that time. Initially the release of forecasts was motivated by the scientific team, but interest, use and understanding quickly rose in the public, government, and private sectors which led to regular refinement and additions to the type of forecast information provided. The basic tenet of forecasting has remained the same, with the forecasts being used to inform decisions requiring consideration of timeframes from days to decades. The operational models were all based on the previous 10–20 years of research in model development and testing; this included the development of methods for combining multiple models into a single hybrid model that typically provides increased statistical forecast skill. The hybrid models combine models from three classes of time horizons: (1) short-term aftershock clustering based on the epidemic-type aftershock sequence class of models, (2) medium-term clustering with time horizons of years to decades, and (3) long-term models with time horizons of decades to nominally time independent. Forecasts have been issued in response to 14 large or significant earthquakes with the forecasts and accompanying information regularly updated on the GeoNet webpages. An important component to the successful deployment and uptake of the models was the foundation laid in the previous decades in model development and testing. This provided understanding and confidence in the models’ ability to provide useful forecasts for response and recovery decisions.